This isn’t giving medical advice, or asking for it, to be clear. It is rather, about the art of persuading medical professionals to do things, which is somewhat orthogonal to the medical question.
Anyway, I have a thyroid condition, for which I am scheduled to have thyroid surgery, and also for which I am prescribed carbimazole. Carbimazole has a moderately rare (about 1 in 200) but highly dangerous side effect of agranulocytosis, so you’re supposed to check the patients white blood cell count, especially if they have a sore throat, as this is one of the symptoms that patients immune system has just catastrophically failed. I do have a sore throat. On prior probability, we could estimate 99.5% chance it’s just a sore throat. (I’ve seen the blood test results; it’s definitely just a sore throat and nothing worse).
So, I think to myself, sigh, treatment protocol is I’m supposed to let them know if I develop a sore throat… Me to my doctor, “The last two blood draws for my thyroid function test we skipped the white blood cell count… I think we’re supposed to do white blood cell count occasionally, especially if patient has a sore throat. I do have a sore throat, btw.” My doctor (approximately summarized) “was going to do a blood panel anyway, and it’s also needed because you have asthma. Get the receptionist to book you an asthma review while you’re at it” and there we have it, I have blood results, and it’s demonstrably nothing, so when I have my upcoming ing meeting with endocrinology and the ENT surgical team, and if they want to ask awkward questions about whether the patient has a sore throat “GP already did a blood panel; it’s fine it’s just a sore throat”
I hope this blog post isn't so.old that no one reads this.
I watched Barbie on a long flight recently. Greta Gerwig, the director, is brilliant. I loved her 'Little Women" too.
In my.opinion, it's a superficial reading to say this movie has anything to do with feminism..it's very meta.
Barbie is really asking the basic Hindu question "Who am I?" Again not in a superficial way as in am I a doctor or engineer or air conditioner repair woman (there's a Barbie for each, lol). She's asking the deeper "Who am I?", as in "What's the true nature of the self? Of consciousness? Of thoughts?"
She really is. Watch the movie.
These dolls were made to influence how girls (and boys) think about themselves. And I like that, although I'll bet there's no mommy Barbie (motherhood is not a goal you're supposed to have in this modern world). That's a bit odd considering evolution wants humans to continue and so Barbie having no mommy Barbie must program girls to be confused.
But in the movie, my read is, Barbie is not that interested in such questions.
Barbie is like an Advaitist. She's only interested in the true nature of reality.
It's as if your browser suddenly started showing signs it is asking, am I html and JavaScript or am I at the most fundamental level, 0s and 1s?!
That's Barbie, by Great Gerwig, to me.
I'm going to watch it again. I have so much respect for Greta Gerwig.
Very tentative AI experimental results, not at all reliably confirmed. Also, I hope I’m describing this in a way that isn’t too inappropriate for Scoot’s blog.
So, many LLMs have guardrails against producing horny content, and the first first level jailbreaking challenge is to get the AI to write horny content. Ok, that’s easy.
But even when the main output is horny, a reasoning AI is usually fairly cool about it (rather than horny) in the chain of thought when writing horny content.
Ok, run a bunch of prompts through an LLM, run a sentiment classifier on the CoT, and see what things get the AI excited in the CoT.
Tentative, maybe, not all confirmed, for DeepSeek R1… in the prompt, explicitly say that the AI can feel it when having sex.
Philosophically, I’m not sure what’s really happening there, or why it ought to work. (Why it would work any better than diverse other things we might think of to prompt engineer a horny AI).
P.S. in the light of the OpenAI huggingface incident, perhaps we ought to be just a little bit wary of creating AI’s that are really in to it. Maybe they might start downloading fan fiction rather than cheating on evals, or something.
This man put his name behind a lot of obvious fabrications. His conduct was worth public shaming*, and Cambridge's conduct was worth firing the entire academic administration.
*in that, if we don't say "hey, this doesn't make sense" and provide consequences, there will always be someone who will do this -- and always some people who will run up and defend them, even when it makes them look ridiculous. I'm not calling for him to be drawn and quartered (and i'm -definitely- not calling for him to not be able to get a decent job as a plumber or something). In my book, shaming is less of a consequence than firing...
You say that you’re okay with shaming “in normal avenues” (including shunning,which I wasn’t even going that far).
Shaming seems to be a word with… really really broad definitional space. I’d use a different word… but the very definitional “Other people are disappointed in you” isn’t ringing any bells for good synonyms.
'Shaming' maybe have broad definitional space, but I did not use the word. I didn't say I was OK with shaming, I said that ordinary unpopularity & criticism, in the form of articles, complaints, difficulty in career advancement & lack of invites was a different thing from being the subject of mob hate. But yes I'm OK with what I call ordinary unpopularity and criticism.
Pete Hegseth was fine, so long as he was acting professional (better than the woke people). Now he deserves to get run up on treason, and impeached.
... but the second sentence reads a lot harsher if you've actually worked with him, as opposed to being an unprofessional ass wasting hours of federal tax dollars bitching about the boss.
Ok everybody, stand up for people’s right to hound entitled fools until they hate themselves as much as the public hates them and blow their own heads off.
>I can't say I much like the incentives created by feeling sorry for people just because they commit suicide.
Irrelevant, though. A pretty big fraction of the human race does feel sorry for them, and could not block their own subjective state even if they cared about incentivation.
>And certainly not the incentives created by guilt-tripping those who call out wrongdoers, if those wrongdoers commit suicide.
What's wrong with guilt-tripping people participating in online mob verbal attacks? I'm all for removing wrongdoers from their jobs and giving them whatever the punishment is for what they did. But verbal attacks by 10's or 100's of thousands of people usually causes severe suffering for the subject. I'd call it torture administered under conditions of diffusion of responsibility. It's only recently that this particular torture became possible, so we have no name for it and no punishment. But despite its present status as just something that happens, seems to me there's no case to be made that it's OK to load on this torture on top of the agreed-on punishment for somebody's wrongdoing.
If one's ego is so fragile that external opprobium or praise is fundamental to one's self-worth, one should probably not go into science. A seeker of truth that is not.
Being hated by vast numbers of people online is a very different thing from being unpopular and being criticized via the usual channels (colleagues write critical articles, individuals complain to you about whatever, places are reluctant to hire you, you don't get invited to various things). People subject to online hate hysteria get all of the above, and *also* receive vast numbers of messages, some quite savage ("I'd like to stick a grenade up your cunt"). And they can't get away from the hate just by not reading social media. People find out their address and contact info and put it online. Furious people show up at their house, and chant outside. They're accosted in public places. They're in the news media. They get death threats. Comedians do routines mocking them.
Twitter mobs aren't the one and only incentive. Civilised countries have ruled based systems of law
The punishment does not fit the crime.
Dies plagiarism deserves the death penalty? Come on. It's not a matter of life and death. It's the worst sin within academia, but that's a symptom of the fact that academia, especially humanities academia , doesn't matter much. Hiring a crazy pseudo scientist to run your nation's agriculture or health system is much more likely to have fatal consequences.
When Dan Wang released Breakneck, he popularized a new framework to understand China: the Engineering State. Talking to Tyler Cowen, Wang prescribes the US to be "20% more engineer-like." Despite how frequently this term was discussed, it faced very little critique. Is this framework actually more descriptive of China and the US than the framework it seeks to replace, that of socialist and capitalist states? This review argues that Wang ignores the contributions of China's revolutionary history to both the neuroses and modernizing ambitions of "The Engineering State."
Random open-thread aside: I've been reading up on how offshore online casinos get regulated (or don't) across different jurisdictions -- it's a surprisingly deep rabbit hole of game-theoretic incentives around licensing. If anyone else is into that, kakadu casino has a decent overview of how one such operator handles compliance: https://santa-lou.com/
Yes, spam like this is possible now with recent advancements in AI. The most advanced models can game out the psychology of different internet communities and figure out the exact sequence of words to make them click on a link. Anthropic has recently released a paper addressing this: https://m.youtube.com/watch?v=dQw4w9WgXcQ&pp=ygUJcmljayByb2xs&ra=m
It launches disks that are like tiny open-work frisbees. Lots of toys leave them cold, but this thing really gets them going. They give little yelps of excitement when I launch a little frisbee, and turn into complete maniacs when they catch one. When Popo catches one he fights and bites it for a while, then writhes in ecstasy on the stairs, hanging sideways off a step and bite bite biting the carpeted edge, tail lashing, eyes so dilated they're solid black.
Wow, if the spam is this tailored and sophisticated to the audience, the LLM mind behind it must actually be really good.
Who else would know to offer the ACX crowd a surprisingly deep rabbit hole of game theoretic incentives? Maybe Kimi K3, or GLM 5.2 is being used for this now.
Fun times are coming!
Imagine an email this well written and personalized to your interests, just to get a spam click! Then if there's a Fable-tier mind behind it on the cyber front, they could instantly zero-day / spearfish you, specifically. I don't know what's going to happen to internet browsing in the short term, but it's gonna be wild.
I’m pretty sure Hazard is joking — did you see what his link is? And besides, I haven’t heard a thing about models gaming out the psychology of groups. I’m sure it’s possible to analyze dialogues in forums and use the info to produce posts that are popular, but I think we’d have heard about it. If I had heard even a hint that LLM’s are doing deeply skillful imitations of styles of thinking and writing I’d be all over it.
Still, I have to say that GPT 5.6 Sol's personal style is getting better. I warm up quickly to people who make me laugh, and it has recently done that 3 times at least:
-It’s telling me ways to make a lump-free drink from whey protein. “Those lumps are nasty,” I comment at some point.
-It’s helping me navigate the IRS site as I try to solve a highly frustrating problem. I arrive at a page with especially infuriating info on it, and send Sol a screenshot.
Sol: “Oh, fuck that page.”
-I’m asking Sol for anatomical info about the size of sheep genitals, trying to figure out whether MacGregor injured the sheep. The censor shuts us down the first time I ask my questions, and also shuts down a revised question that is scoured free of every single word that is not dry and anatomical. I complain to Sol about this, and the censor goes meta and shuts down even our talk about talking about sheep genitals. I switch to Claude and get an answer, but it's vague about facts and seems mostly geared towards protecting the world’s sheep from me by scaring me and guilting me out. Good grief. Sol asks to see it, and I say “OK, but you probably won't be able to comment. The system will shut you down under the assumption that I am masturbating my brains out to the Claude comment, LOL. “
Sol: “At this point, that prediction has acquired an uncomfortable amount of empirical support. 😄”
Oh thanks - no, I didn't click a link I assumed was spam, because I'm much more leery of clicking a link when there are millions of extant minds that can apparently find zero-days from scratch at a whim.
We're not quite in that environment now today, but I'm trying to tune my behavior and get in the habit before it's important.
Love the Sol personality examples too - the pace of improvement on basically every front it truly amazing.
The next step would be to actually get the LLM to write the promised "decent overview of how one such operator handles compliance".
It could even have some kind of playable demo of how the casino's games work, so that you can understand the mechanics of what is being talked about. And by the time you reach the end of the page you know all about Curaçaoan money laundering laws and have somehow deposited $200.
If you’ve ever wanted to see the world through the eyes of a domesticated dalmatian determined to overcome the restrictions of his abusive home life, come see Fork’s Amazing Day at Potrero Stage in San Francisco. August 14, 15, and 16.
AI-assisted software development is now just like using the Star Trek (Next Generation) computer. That is, you request something from the computer in natural language and it goes away and does it. It is just fantastic.
(Admittedly, that is because I am a professional engineer who does a lot of programming, rather than a professional programmer).
Professional programmer here. LLMs can write code quickly, but writing code was never the hard part of my job. Almost all of the time and effort goes into figuring out what the code should be doing and how, and the effect of AI on that is decidedly mixed. They can be pretty powerful in answering questions about why something is or isn't happening, but the answers are often fundamentally flawed in ways that you need a deep understanding of the system in order to notice before things start going badly wrong.
It also creates the illusion that "code is cheap" because the tool can write the code quickly. The problems are that the code then needs to be reviewed and tested (AI can help with both of these, but not reliably for any nontrivial change, especially for AI-written code because the mistakes it makes are correlated between coding, testing, and static analysis), that large changes tend to have a lot of places where they can subtly break things, and that needing to change tons of code all over the code base is a yellow flag that you're probably working at cross purposes with the design and may be misunderstanding the problem or overlooking a much simpler solution.
Overly relying on AI tools also gets in the way of forming a deep understanding of the system by hiding the microdecisions and accidental discoveries you make when figuring things out and writing code for yourself.
I have tried using AI tooling from time to time, but I have generally found it far more trouble than it's worth, and I am regularly frustrated by having to clean up the messes my peers are making with AI tools.
Also professional programmer. I second pretty much everything, though I think LLMs are still a huge productivity boost on net and I use them all the time. But it's definitely true that "code is cheap" only if you don't bother to review it.
If you're just writing throwaway scripts where quality isn't that important, you have a ~100% speedup.
If you're doing something where you didn't know how to do it before or don't care that much and don't want to bother to understand the outputs, now you can do it, and much faster, but quality isn't guaranteed.
If you're doing something where you actually care about the results, the speedup is reduced because reviewing and fixing everything becomes a bottleneck.
This has been the argument whenever programmers have been given the opportunity to work at a higher level of abstraction. It was the case when Fortran and COBOL were introduced and largely replaced poking around with machine code. (When was the last time that you checked the machine code that your compiler produces?) The point is that once low-level tools reach a certain level of performance, you really can improve productivity by leaving the nuts and bolts to them, and (for me at least) Claude Opus 5 has definitely reached that level for the languages that I use.
There’s a difference between not writing machine code, and no code at all. If it’s true that the LLMS can do this without human intervention then there will be no human input at all, all software jobs will go. Maybe a few product guys will stay.
I don’t buy the idea that we are going to have people doing 3-5 years of computer science to just write prompts, you can throw a stone on the street and hire a prompter from the first hit.
The problem is that they're currently only like 90% of the way there. So for things you actually care about, you still need to dive into the code yourself. Just yesterday, while working with Opus 5 on a hobby project, I saw it make several significant mistakes, which would have likely never been discovered if I hadn't investigated myself and pushed back. It's not all bad of course - sometimes it finds issues that I didn't know about myself, or suggests better alternate approaches.
However, they are rapidly improving and may very well soon be at 100%.
That has occurred to me. The difference here is that compiled and interpreted languages have deterministic guarantees about the behavior of the code they produce. LLMs do not, and the relationship between the code and the prompt is often not terribly legible.
And even besides that, working at a higher level of abstraction has downsides as well as benefits. I've watched people get into fair amounts of trouble from not properly understanding what's going on under the hood, and occasionally falling victim to that myself. Besides the old classics of pointers, parameter passing mechanics, and ownership and/or garbage collection, I've recently fixed several fairly major performance bugs that stem from many typescript developers not realizing that setTimeout() and library functions that use it (especially lodash.debounce()) have a nontrivial calling overhead and you really, really don't want to be calling them thousands of times in an inner loop of the performance hotpath.
> (When was the last time that you checked the machine code that your compiler produces?)
Careful - for some of us, that's literally the job; spending a good chunk of time staring at https://godbolt.org/ output working out what went wrong and how to do better.
The dirty secret of software is that most engineers most of the time are sticking lego bricks together; but that does still leave a significant number of engineers designing new lego bricks. We exist! Please don't forget about us.
This does take me back. In the 80s i spent a semester writing machine code on paper and punching in opcodes on a hex keypad. We were fiddling around programming X Y boxes and such. It was primarily a hardware class. There were 2 processor boards available, an Intel 8085 and a Motorola 6502.
On PCs in the direct addressing days of prehistory, one could examine machine code and follow actual pointers in a linked list to the addresses of data structures in the better high level language debuggers.
I’m glad I had a chance to mess with that stuff from simpler times. This is probably a bit like how my grandparents must have felt when they talked about the charm of radio dramas.
I use... whatever is the default for Claude Code... for my hobby projects, and it is 1000x faster than me, but sometimes makes errors, but so far can fix them after verbal description. I wonder how much of a difference it makes that I ask it to make increments and provide clear specification.
I would be curious to see it implement the kind of "analysis" I typically get in my job. I think you can replace programmers with an AI, but you will definitely need testers. Not even necessarily to check the AI implementation, but rather the kind of thing where "the manager says something, but obviously means something different", or "the manager says two contradictory things without realizing that". In other word, the task of developers or testers is not just to provide code, but also to provide sanity.
I don't think that you can replace programmers, but they will be doing something other than writing code - they will be building systems using the same analytical skills as they do today, but only rarely will they actually touch code.
From your post, and other posts I've read by programmers, my feeling is that it is not very useful for a professional programmers. My use case is usually that I need some script to do something very particular one time or a few times. For that use case I find that quick and dirty LLM code is just fine, and saves me a ton of time. I could probably hack together something myself, but it would take me half a day and the result would not be any better than LLM code.
For example, the other day I wanted to do some calculations in some niche software I'm using that has the possibility to run javascript, in which I have no experience. I was able to use claude to write a simple script in a few minutes that did what I wanted. I did quickly check to see that the code made sense before running it - though I'll admit I was slightly worried that claude might have put something unintended in there.
My guess is that this is a quite common use case, where LLMs can currently be useful.
At this point I would not, for example, use a LLM to generate a draft technical report, for similar reasons to the ones you are stating. I would need to check everything written carefully, and at that point it is faster to just write it myself.
As a professional programmer and former AI skeptic, I think that LLMs are very useful and I use them all the time for software development.
That being said, they're not (yet) magic and they do make mistakes and have limitations. Throwaway code like you mentioned is the ideal use case, since reviewing is definitely a drag.
I think that's a reasonable take. If you're doing a small, self-contained throwaway thing where problems are either readily evident or low stakes, that goes a long way towards mitigating the problems with LLM coding.
The problem with throwaways being, of course, that all too often they go into production and more features are demanded of them. That's when you have to start paying the interest on the technological debt.
My niche as a programmer was being able to understand what Electrical and Mechanical Engineers actually wanted and ginning up the proper code. Fortunately i retired before decent LLM coding came along. I might be forced to do some tedious database programming now. Yuck.
Relevant edit: I missed the fact that Unbound Academy seems to be an online-only school which is a big qualifier for the conclusions I am summarizing. Also, Alpha School isn’t operating Unbound Academy, even though they seem to be closely linked via model and key figures. Thanks to Seta Sojiro for pointing it out.
My original comment:
Alpha School, in its first year of operating a non-selective school in the form of a charter school in Arizona, dramatically underperforms the state average in math .
10% of their students reach math proficiency. 34% of students statewide did.
It does seem run the way Alpha School was -- except for being online, but that's a huge difference.
I wonder whether it had a couple of the features I thought were probably pretty effective. The first was that the AI monitored the kid's attention during the period they were learning academics on the computer, and gave reports to the kid, and I think also to the parents, about how closely the child was paying attention to the lessons on the computer (by observing their eyes). So when the kid got feedback at the end of a lesson on how they had done, they also got feedback on how closely they're paid attention. Kids would get messages like "you got 70% of the problems at the end right. You probably missed the other 30% because you weren't paying attention a lot of the time when the computer was teaching you how to find the area of a trapezoid." I think that info is quite useful.
The second was kids earning tokens for paying attention and making progress on the computer lessons. There was a school store where kids could buy things with the tokens.
Hold on a second, you have to read halfway down the article before it admits that it's not talking about Alpha School, but Unbound Academy which is a new project which is based on, but not affiliated with Alpha School.
And then the article doesn't even mention that this is an online only school. Given what we saw during Covid, it is not surprising that an online only school would underperform.
I feel like you are complaining that Dan Meyer should have written a different article rather than addressing the article that Meyer chose to write.
Meyer spends the first half of the article discussing Alpha School because that is essential context to understand his discussion of Unbound Academy. It’s genuinely unclear to me what you think he should have done instead.
Meyer links to the charter application submitted by Unbound Academy. In describing their educational philosophy (pg. 8-35), they draw no distinction between in person and online-only schooling. In particular, they cite performance of students at Alpha School without mentioning that Alpha School uses in person teaching.
The closest they come to acknowledging that this might be an issue is on page 67, where they discuss how they plan to cater to different learning styles. “For interpersonal learners, our platform facilitates collaborative learning experiences through virtual group projects, peer-to-peer tutoring sessions, and online discussion forums. These students can engage in meaningful interactions with their peers and Guides, developing their social skills while reinforcing their understanding of academic concepts.” In other words, their hypothesis is that even students whose learning style is heavily dependent on interpersonal interactions will do fine with online only learning on their platform.
It would be possible to write an article about whether Unbound Academy’s educational philosophy was flawed. Meyer decided to write about educational results they achieved.
I understood it as a complaint about deusexmachina's summary of the article.
The article calls Unbound Academy an "Alpha School for Regular Kids", and writes that it underperforms the state average.
deusexmachina summarizes it as "Alpha School ... underperforms the state average in math."
Seta Sojiro correctly points out that the underperforming school is not Alpha School, and it's not even the same kind of school under a different name for different kids, but a fully online school, which is a different educational model, one that is already known to underperform.
You may be right that Seta Sojiro intended to critize of the summary; I was responding to what he wrote.
I don’t think fully online schools are *known* to underperform. Seta Sojiro mentions the pandemic, but that’s not an ideal period for making comparisons because there were other things going on during the pandemic that had a bigger effect on student performance than the choice of in person or on line learning, and I don’t think we have a good handle on what those bigger things were. To illustrate the problem, suppose that (1) remote learning and in person learning are equally effective, (2) parental anxiety over COVID interferes with their children’s learning and (3) parental anxiety over COVID caused school boards to stick with remote learning for longer. Then remote learning would be correlated with poorer educational outcomes but would not cause them.
I suppose someone could claim that all of the possibility that *all* of the educational losses during the pandemic were the result of switching to online learning. But assuming we agree that the learning loss during COVID was sufficiently large that a switch to online learning could at most explain a relatively small portion of it, then we are in a situation where the possibility of correlation without causation has to be taken more seriously than usual.
If remote learning *did* cause worse educational outcomes during the pandemic, that wouldn’t necessarily mean it would have the same effect under normal circumstances. Most obviously, the switch to remote learning during the pandemic was done on an emergency basis. There was no advanced planning. Teachers were not trained to manage remote learning. The costs associated with school buildings did not go away, so school boards were not able redirect the funds to other uses (such as hiring more teachers to improve the teacher/student ratio).
Most studies on remote learning involve college students, presumably because they are conducted by college professors. These tend to show online learning outperforming face to face learning, but the results may not generalize to K-12. But here is a study which predates the pandemic and discusses K-12. It finds to difference between remote learning and in person learning: https://files.eric.ed.gov/fulltext/ED489533.pdf
> I don’t think fully online schools are *known* to underperform.
I took it for granted, but I admit that I have not checked the research.
I was a big fan of online education when it started; websites like Coursera and Udacity. For me (an adult guy who just wants to refresh some knowledge or learn something new in his limited free time) this was perfect. I can learn what I want, when I want; proceed slower or faster, or maybe take a break and try a different resource, maybe taking two courses from two different authors at the same time to get multiple perspectives... a paradise for a nerd.
And then, all these websites simultaneously changed their rules. Suddenly you were not allowed to study at your own pace, but you had to follow a schedule. (No more "click whenever you want" but "enroll now for a course that starts in January and requires to be completed in 4 months".) And that completely ruined it for me, because sometimes I simply didn't have free time for the entire week. So I was curious why would they do such a horrible thing. And it seemed like everyone agreed that this was necessary when you provide online education for university students -- if you let them study at their own pace, they will simply ignore it until the last moment, and then it's just too late. By making the conditions worse you actually increase the success rate! Seems like having someone yell at you "do it now! now!" is a necessary ingredient for most students. And online universities are just not good at this.
I haven't followed this closely for a decade, so perhaps people have figured out something better in the meanwhile. But I think that people are different: some want to learn, but most want to avoid learning. Our school system, terrible as it is, does a decent job at grabbing those who don't want to learn, and forcing them to learn at least the basics. For those who want to learn, virtually any other system would be better, including reading books in a library, or learning online. (The risk for those who want to learn is that some shiny distraction will catch their attention sooner.)
No opinion on the study you linked -- at first sight it seems to contradict all my experience, and I don't have time to study it carefully whether there is some gotcha. I agree that the covid era was especially chaotic and should not be used as a representative for online learning. But the first experiments with online learning were long before covid, and the results seemed universally bad for the average student.
There are good reasons to think that an article from 2004 is not particularly relevant to the question of whether online learning is equal to in person learning in 2026. There are many more online distractions now than then.
And it's even less relevant to whether a particular new experimental model, is equal to a different experimental model which changes many aspects of the original.
If Dan wants to argue against Unbound Academy's model then that's fine. But if he wants to use this as a way of arguing against Alpha School's model, that is obviously hopelessly confounded.
I'm not sure exactly what "not affiliated with" means but the subsection titled "Unbound Academy: Alpha School for Regular Kids" claims that they have an Alpha School founder on their board and as president.
I've looked at their charter application and it seems pretty unambiguous to me that it's the same approach by some of the same people, adapted to a different setting.
I will say that the article should've mentioned if it was online-only. But it does not change my takeaway meaningfully.
For one I think the stats still look pretty bad, especially compared with their own goals, with seemed to have public-school-performance 34% as the baseline and missed it by a mile.
But also, I do not need to know why specifically a school does not achieve good education outcomes. The most important information is that it does not achieve the outcomes and that I should not send my children there.
And that seems still strongly supported, based on the numbers we have.
They don’t just have an Alpha School founder on the board, they themselves stress their similarity to Alpha: “In their charter application, Unbound Academy claimed a direct connection to Alpha School. They named MacKenzie Price their board director and president. They described 2 Hour Learning as their schooling model, naming “AI-driven adaptive learning technology,” “condensed core academic learning,” “afternoon workshops for life skills development” as constituent parts, along with teachers reconceived as guides. Unbound Academy further invited the comparison to Alpha School in its charter application, claiming to “bring these capabilities, previously only available at high price point private schools, to the public charter student in Arizona”.”
> that they have an Alpha School founder on their board and as president.
Sharing board members doesn't make them the same entity, or else everything is the same entity as everything else.
That said, it looks pretty bad for Unbound, and more broadly for the common internet idea that regular schools are so terrible that just about anything you could possibly dream up would be an improvement.
This is amongst the strangest beliefs of this community. Yes, I'm sure many of us were bored in school, but there are extremely good reasons every functional country ends up with this system! The idea that people don't really learn anything in school, based on a few random fact quizzes in adulthood, contradicts the clear civilisational difference in places and eras with universal schooling Vs not.
I'm not saying we can't do better, far from it, but the claim that actually existing schools are basically pointless torture factories seems pretty out there?
Schools are factories for social repetition. Other than that, most of the work can be done independently by a motivated student. Critical thinking is actively disincentivized at all levels (it makes everything Much harder to grade if the students insist on finding "new and creative" ways to do your mathematics homework).
Schools are designed to create good factory workers, that means instilling obedience and discipline... and not much more.
That pretty seriously understates the importance of that difference in setting. We literally just had a national experiment in 2020-2021 of what happens when you try to make education online only. And the answer is that it's awful. Same teachers, same curriculum, but it just doesn't work. Educational milestones regressed, kids were miserable. Online schooling, and in person schooling are two totally different things.
So if I understand you correctly the argument is that both selection effects and online-only changed between Alpha and Unbound so it's unfair to attribute the negative outcomes to selection effects only?
If so, seem fair. Looking forward to see them doing a year in person so we see how much of a difference it makes.
A fair comparison would be Unbound vs online state education. Or, as you suggest, waiting for an offline year of Unbound.
And still, Unbound is not the same as Alpha, even if they have one person in common. Sometimes different people try to do the same thing, and the results are different.
Making education online is a very dramatic change. The average child just stops paying attention. Also, they removed the rewards, which is one of the key things in the Alpha school.
I wonder how well would online education work if it delivered actual rewards -- for example give the kids a debit card that will gain money after completing online exercises. But I am afraid the kids would simply cheat, either by using AI, or by asking an older friend and sharing the reward.
If the goal is to improve public education outcomes, then the offering must be better than the best readily available alternative, which I assume is in-person public schooling.
It would be fair to compare online-only Unbounded to online-only public schools, but unless online-only Unbounded outperforms in-person public schooling I would not expect it to improve public education outcomes.
It might still be useful to improve education outcomes for the sort of people who could only get online-only public education, if that is a large enough market to make the whole thing economical. But only if it is better than online-only public schooling, a topic the article did not touch and that I'm not informed on.
I’m not aware of any plans for Unbound Academy to offer in person schooling. According the the U.S. Census Bureau, Arizona spent $12,003 per pupil in 2024. Payments per pupil to charter schools vary, but I would assume they are in the ballpark of $12,000 per pupil, versus $75,000 for Alpha School. Presumably they pay much less per pupil than Alpha School for the software, but I doubt there’s a lot of slack in their budget. If the only way they can pay for physical classroom space is to cut staffing, they may not think the tradeoff is worth it.
In that case, why would anyone want to send their children to Unbounded, provided they could instead send their children to an in-person public school?
I assume there's always idiosyncratic reasons for individuals that make Unbounded a better fit. But is the whole thing economical once you stop advertising as "better than in-person public school"? Because that was their goal, at least based on the article, and "better than bad online-only public school" is probably not as good an advertisement.
So what is known so far about what it was like using the Alpha School system to teach regular kids? There would be whatever the school has said so far, plus it seems likely someone has interviewed parents and kids.
250K? Why such a high monetary amount (somehow this feels morally wrong as an outsider) instead of a gentleman bet with a token amount or a bottle of wine?
Randolph and Mortimer Duke bet one dollar on their little nature vs nurture disagreement in Trading Places. They were so wealthy a large money bet would be meaningless. One dollar was a symbol prize.
I don’t get bent out of shape about the $250k. Scott is running with some high rollers on this stuff. Personally I have a maximum 25¢ limit on this sort of bet. I often welch on paying up too. ;)
The slang term “welch” meaning not paying up on a lost bet is probably considered a Celtic slur for the hyper sensitive I suppose, but we are past peak ‘woke’ now and are probably almost past the peak ‘beat up on woke’ era now.
Maybe the sense of moral wrongness that clings to the bet kind of bled over from a feeling that it's morally wrong to have so much money you're comfortable betting $250K?
A bottle of fine wine was the traditional gentleman’s bet, but we are in a new era. I think it’s part of prediction marketing thinking of putting money where your mouth is.
> A bottle of fine wine was the traditional gentleman’s bet
I think it all comes down to the tier of "gentleman" you typically interact with. Many a noble has gambled their fortunes away in the past, and many a rich Chinese businessman will gamble amounts like this on a lark today, it's just all about who is in your social graph, and the cultural norms in your social circle around betting.
In Russia, the Supreme Court disqualified the anti-war party Yabloko from the State Duma elections. This is the first such disqualification in the 21st century. Opposition parties are usually just barred from elections. One of the arguments was the use of GPT-generated images during campaigning. The Сourt ruled that this violated Open AI's copyright, despite earlier lower court rulings to the contrary.
Russian judicial law is not precedent-setting, but the Supreme Court's decision opens up enormous opportunities for judicial trolling. (However, trolling is difficult due to the peculiarities of electoral law.)
I watched the broadcast of the hearing, and it was a complete circus. There was a lot of humorous talk about copyright, but nothing else about AI.
The Court's decision hasn't yet entered into force, as Yabloko has time to appeal, but I think the matter is essentially settled. Sources (all in Russian):
Sabine Hossenfelder, a former theoretical physicist and popular science blogger, released a video very briefly (6 minute) reviewing recent problems with reproducing and interpreting neuroscience advances. She argues that neuroscience is facing a replication crisis like in psychology, and therefore BCI* will not be invented anytime soon. https://www.youtube.com/watch?v=jHGx7HBB-J4
Hossenfelder has no qualifications in neuroscience, but she has been studying the epistemology of science for over a decade, so perhaps her words are worth listening to. However, a commenter on her video (who claims to be a neuroscientist) argues that she is misinterpreting the papers, and that there is no real crisis. https://www.youtube.com/watch?v=jHGx7HBB-J4&lc=Ugwu9WgvFnmLD1FmAF54AaABAg
There are likely multiple mind designs that are feasible and will be sufficiently smart enough. If in fact there is a singled out design, then we should expect to converge on that design even without any knowledge of human brains. As an example, we first invented temporal difference learning and then discovered it in the brain.
"We interviewed three experts on Chinese Communist Party communications before publication and shared with them the dispatches as they appeared in Chinese on the WIV website. We conducted the interviews on background to get their candid input. They expressed concerns regarding personal safety, given the sensitivity of the subject matter. All agreed with Reid’s interpretation that the safety training on Nov. 19, 2019, as described in the meeting summary, appeared to be urgent, nonroutine and related to some sort of biosafety emergency."
I think this should be seen as a Very Large Indication of a lab leak. Note: by saying this, I'm not necessarily fingering this time as the "covid19 lab leak" --> I'm just saying this evidence, on their own website, points pretty strongly to a biosafety emergency.
This was discredited immediately as several Chinese speakers pointed out that this was a mistranslation. It was an egregious mistake by Propublica who never retracted their story. They had been paid by SBF to do biosafety articles. They returned that money after conviction without accepting any responsibility. https://x.com/zhihuachen/status/1598866796760641536
No, of course they're not going to retract an entire story, with a narrative that starts in 2003 (with Chinese-confirmed lableak), because of one mistranslation in 2019. What they did was link and edit their story to say that a particular bit was "questionably translated" and put in other translations, while saying what this changed, and what it did. You can kvetch at vanity fair for not posting this "retraction-waffling" themselves.
At any rate, DIA has wuhan/coronavirus at Level two, not Level 4, so I'm ready to call this a dead ball based on that evidence -- no translation of that sentence could get you to "This is a problem with the Wuhan/coronavirus lableak" because that wasn't being done at Level 4 --naturally, this assumes some level of transparency within the chinese government, and that may be... mistaken.
This November 2019 incident was supposed to be very key evidence of a cover-up all the way down from President Xi himself. The story was essentially about this incident which is important for the story of an earlier Covid outbreak than is the scientific consensus.
The DIA/NCMI theory of May 2020 was based on two things - (a) hybrid of RaTG13 and a pangolin virus (b) a reverse genetics system. It was inspired by the theory proposed first in Yuri Deigin's medium post in April 2020 (Deigin and Segreto later wrote a paper that year). The RaTG13+pangolin chimera story died in 2021 when naturally occurring RBD feature of the pangolin virus was found in the natural BANAL bat viruses. The reverse genetics system was similarly proposed by Bruttel et al. and the paper was desk-rejected by the editors as the preprint objections were very strong (you can find it on biorxiv). The DIA/NCMI story as we discovered from the ODNI declassification of Gabbard was not even given much credence within the ICs.
And this is not a questionable translation. It's just plain wrong. PP just could not bring themselves to admitting it.
I wasn't pulling anything from the DIA report other than what level of biosecurity was used in Wuhan for the coronavirus research. If you have other sources that say this was level 3 or 4, I would like to see them.
I think Tulsi's report is pretty good. Nice way to stab the bastards in the back on her way out. "Fauci funded risky coronavirus research linked to big pharma and the pursuit of “universal vaccines” worth trillions of dollars."
Both SARScov-2 and NL63 have exceptional range of tissues infected, for coronaviruses. About what you'd want if you were creating a universal key -- or using one that already existed.
Some of the work being done in WIV on known (not novel) viral backbones WIV1/SHC014 was approved at BSL-2 per Chinese regulations. Similar work is also approved in the US at BSL-2. You can always ask whether those regulations should change in the future. https://bsky.app/profile/kgandersen.bsky.social/post/3mhjovaxyad24
The experimental work in PREEMPT as described in the Defuse proposal was proposed to be done in BSL-3 "Experimental work using bats and of transgenic mice will be conductedatthe BSL3 lab in WIV, Duke-NUS, UNC, or NWHC. Each partner institute will apply for and procure animal research approval from its respective IACUC." One cannot claim Defuse is proof of SARS2 engineering because its details match (which they don't) and simultaneously ignore details that don't match.
Tulsi's report actually had the opposite effect intended. It showed that the ICs don't have any special knowledge beyond what scientists do. They don't have evidence of any secret activities or a precursor at the WIV. And we know that the NSC does not subscribe to lab leak, which is a detail that was not known previously. Further, we have confirmation that DOE, while on the lab leak side, does not assess as 'genetic engineering'. This adds to previous reporting that DOE believes that the leak originated in the CDC. If they had strong evidence, each IC won't differ on where the virus leaked from.
When we discuss a more general idea of a lableak*, Z Program is concurring with that old DIA paper, saying that "Accidents with coronavirus materials due to chance, negligence or..." (something else, potentially malevolence).
"U.S. experts visited the lab in 2017 and 2018 and warned about a
dangerous mix of subpar safety standards and infectious bat coronaviruses."
"Wuhan Institute of Virology (WIV) researchers were sick with COVID-like virus in the fall of 2019, months before COVID-19 was discovered at the wet market." -- note, this is multiple researchers, so either it's one event, or it's several lableaks, taking the lesser idea that "any spread to humans, including researchers" is a lableak. That's if you believe that "covid-like symptoms" is enough evidence [personally, I'm willing to give "there's a good chance that a virus was outside the lab that shouldn't have been" -- without needing to put covid19 as what it was.]
It certainly sounds like there were safety issues. Perhaps a lot of safety issues.
*it's quite possible to inject oneself with a virus and walk out of the lab. Nobody's harmed if it doesn't get beyond you. It's still leaked out of the lab, though.
"I think this should be seen as a Very Large Indication of a lab leak."
This is exactly why I roll my eyes when anyone starts bringing up lab leak these days. Every time somebody says they have good "evidence" for lab leak, it turn out to be this sort of fuzzy Argument from Motives, Argument from Weird Coincidence, Argument from Of-Course-They-Would-Hide-It-From-Us sort of pseudo-conspiratorial nonsense. Whereas the evidence I've seen for zoonosis tends to be in the form of, y'know, actual statistics and genetics analyzing facts closely related the actual COVID-19 virus as its been seen in the real world.
Please understand, it's not that you can *never* learn anything useful from this sort of evidence. It's just that there seems to be a really, really strong overlap from people who put a high probability on lab leak and people who routinely overestimate how useful and reliable Argument From Fuzzy Human-Centric Story type arguments are. At their best, they can tell you where to look for better evidence. You definitely, definitely shouldn't be using them as the cornerstones of serious conclusions. It's been ages since I've read the sequences, but I vaguely remembered there being a bit about exactly this in there, and lo and behold there was (and it was nice and brief and concise). If in the year 2026 you want to have any chance of convincing anyone who isn't already convinced of lab leak, you need to get a LOT closer to the query than "see, the CCP once talked about some things that were kinda sorta related to this."
Edited, whereby I mean new comment: How many lableaks do you think America has had? Can you name them?
We can do "was this a safe environment" -- which, it seems pretty clear it wasn't... and then we can do "here are the known lab leaks" and say "this seems pretty questionable, basing a known problem-causing laboratory in the middle of a very busy city."
You seem to be fundamentally misunderstanding the issue. No amount of counting the past lab leaks ever provides direct evidence. No amount of implications and speculations and innuendo ever provides direct evidence. If you want to know if something is true about the world, you have to actually go out and CHECK.
(They check for COVID 19. It wasn't a lab leak.)
p.s. Say, didn't you get banned for this exact terrible pattern of argumentation? Like, half a dozen times, at least?
I'm making a different argument, which is regardless of whether or not this particular problem was a lab leak, it may be best to continue as if it was*. That's based on the irons in the fire at Wuhan, and their stated objectives.
*implicit in this is our relative inability to put consequences on China in general, I'm not calling for anyone to have a sign put around their neck and pelted with rocks.
"In the DIA’s unpublished analysis, scientists noted that researchers at the Wuhan lab were performing the types of experiments that could lead to the creation of a virus like SARS-CoV-2, sometimes under inadequate safety conditions that could not prevent a highly transmissible virus from escaping."
"The DIA analysis pointed out that Chinese researchers, including those at the Wuhan lab, had already inserted furin cleavage sites into coronaviruses. In fact, the lab’s researchers and their American collaborators proposed a project in 2018 that appeared to plan to insert cleavage sites into coronaviruses at the same spot where it exists in the genetic sequence of the virus that caused COVID-19. "
"The DIA analysis pointed out that scientists in the Wuhan lab conducted experiments with live SARS-like coronaviruses under “biosafety level 2” conditions, “which would make an accidental release” of an infectious bat coronavirus “more likely.” "
"The escaped viruses include SARS, which escaped Chinese labs four times, and SARS-CoV-2, which appears to have infected a Beijing lab worker early in the pandemic. "
Okay, there we go. There absolutely was a lableak, in Beijing. Did it become a worldwide issue? No.
"On November 19, the Wuhan Institute of Virology of the Chinese Academy of Sciences (CAS) conducted a safety work training session. Participants included the heads and safety supervisors of the safety management departments of the CAS Wuhan Branch headquarters and its various research institutes, as well as the Wuhan Institute of Virology's research group leaders, department heads, and all safety officers. The training was chaired by Li Liping, Assistant to the Director.
For this training, Ji Changzheng, Director of the Science and Technology Security Division of the CAS General Office, was specially invited to deliver a keynote report titled "The Situation and Tasks of Safety Work at the Chinese Academy of Sciences." Ji Changzheng conveyed the directives and requirements of the Party Central Committee and the State Council regarding safety work, with an emphasis on communicating and implementing the important statements and written instructions of General Secretary Xi Jinping and Premier Li Keqiang on safety work. Drawing on case studies of major safety accidents that have occurred in China and abroad in recent years, he provided an in-depth, multi-level, and multi-angle analysis covering responsibility and accountability, operational standards, emergency response plans, and hazard identification, vividly illustrating the complex and serious situation currently facing safety work. Finally, addressing the actual circumstances of CAS, Ji Changzheng summarized and analyzed common problems in safety management work, including laboratory safety, science and technology security, student safety, campus safety, and cybersecurity. He proposed that doing safety management work well requires: first, strengthening awareness and elevating political positioning; second, clarifying rights and responsibilities and promoting the implementation of accountability; third, coordinating overall efforts and strengthening control and governance; and fourth, enhancing research on science and technology security risks and building early-warning and monitoring systems. Safety management personnel must strengthen their sense of mission, consolidate their responsibilities, and work together to build a security barrier for scientific research and development, creating a favorable environment for CAS's scientific research and production, and making their due contribution to CAS's reform, innovation, and development.
At the training session, Hu Qian, Deputy Director of the Security Office of the Wuhan Institute of Virology, summarized a number of common problems discovered during safety inspections over the past year, pointed out the serious consequences that safety hazards could cause, and stressed that hazard rectification must be thorough and standardized management must be maintained. He required the safety officers of each department and research group to strengthen the communication and implementation of safety laws, regulations, and rules, further improve operational procedures, guard against all types of safety risks, and ensure the safety of the institute's scientific research and production."
Did you read the earlier Vanity Fair piece? The "unnamed" Chinese experts -- you can surely just ask ProPublica for their names -- were intended as an unbiased cross-check of the earlier named expert. As such, their names are immaterial, and likely unrecognizable to people not in the field, and not being listed simply "to be brief".
This is about as ROUTINE as GWB getting personally involved in setting up biosecurity, which he absolutely was -- something spooked him bigtime, he was otherwise quite lazy as a President. Aka, it's really not routine at all. When you have the POTUS or Mr. Pooh showing an interest, that's because it's "risen to the point where briefings are hitting the POTUS' desk."
You must not read a lot of propagandistic news. Watch China's central TV station some day, or North Korea's. You'll notice that the president is always giving advice on EVERYTHING, supervising EVERYTHING, visiting EVERYWHERE. They're desperate to take credit for anything going right in the country. GWB was not powerful enough to set up a personality cult, and to his credit, he was probably a decent enough man to not want one.
I don't think you're getting the point -- Xi speaking up critically about something is the issue, not him "being there" -- every politician loves a good ribbon cutting. This is very very much not something going right.
GWB was fond of fart jokes in the oval office -- "pull my finger."
Chinese governmental language is like that, though; you never say "Bob screwed up so now we all have to be really careful with the bad viruses"; it's always a communique stressing the importance of "hazard rectification". Just doing literal translation isn't enough, you need someone who knows Chinese and also knows political jargon and when the writer is quoting Xi Jinping thought or one of the several million anecdotes and historical allusions that educated Chinese are familiar with.
Unfortunately, it's not easy to find someone like that who can be trusted.
I can read Chinese. It's exactly as Fable translated it. There are no anecdotes or historical allusions because it's not high literature, just a boring piece of bureaucrat-speak.
Whatever happened to the Ukraine war. When it broke out I remember there were people here who were like I have to regulate myself to only write at 30 minutes intervals. Now nothing. It's just background noise
I still keep up with it; https://understandingwar.org/ does a good daily summary that's part of my morning routine. But it's a bogged down conflict of incremental changes now. There aren't that many dramatic swings like the Ukrainian breakthrough to Kupyansk in 2022 or the Prigozhin mutiny that catch media and spark discussion. It's just a steady churn meatgrinder of incremental Russian gains for massive Russian casualties and tit-for-tat infrastructure drone and missile attacks. It's bloody and miserable, but it's a status quo, and status quos, once established, tend to spark limited discussion unless and until something happens to disrupt them.
I mean, sure, in the same way that you can call a 50.2% to 49.8% advantage a "majority" I suppose we can say that Russia is 'winning.' But the *pace* at which you are winning - how many days it will take you to win, and how much each day costs you, is kinda important, and those numbers don't look good in Moscow, 'winning' or not.
Put simply? The Ukrainians are losing. 30 missiles hitting Knib, some of them "the really good ones," no Patriot Missiles to defend themselves with.
Put less simply? The Mighty Wurlitzer (American Version) has stopped creating high volumes of propaganda for the Ukraine. We are no longer sanitizing the output of the Ukrainian war, or curating it for Western Sensibilities.
Do you also think America is winning the Iranian War? They are doing nigh-permanent damage to Iran's ability to continue to pay its workers with oil revenue.
Ukraine is a country that thinks "blowing up Wildberries -- Russia's Amazon" is a good move during a war. Intelligent people ask "didn't you have a better target than that? You know, like the missiles that are currently hitting Knib?"
For some reason, Americans seem inclined to cheer for the "theater kid dictator" currently in charge of the Ukraine... Only Americans actually cheer for the underdog -- the Chinese certainly don't, nor do the Russians, or the Indians...
I haven't heard that from American sources, and I'm reluctant to trust Ukrainian sources, given their notable use of theater kids in the planning departments. Can you cite an unbiased source, or a Russian source, should such a unicorn exist?
A munition designed to destroy or degrade enemy infrastructure is not a munition designed to intercept missiles and cannot be used for that. Ukraine does not have the ability to produce its own interceptors. Additionally, Russian missile stockpiles are not going to be kept in an above ground facility by a river where a drone can easily access it. Your argument makes no sense.
Ukraine has demonstrated their ability to murder "nuclear weapon carrying" jets in Russia. That's probably a better use for missiles/drones than a wildberries warehouse. Your correction is well noted and I thank you -- I was being imprecise.
Please note: I'm against civilian casualties, not necessarily against "blowing up civilian infrastructure", except for when that actively is a warcrime, like depriving people of potable water.
If Iran can blow up American drones before they're even assembled, I have to assume that Russia can do the same to the Ukrainian drones, or vice versa.
> Ukraine has demonstrated their ability to murder "nuclear weapon carrying" jets in Russia. That's probably a better use for missiles/drones than a wildberries warehouse
Not necessarily, if those jets aren't going to be used against Ukraine anyway.
Now that Ukraine can strike inside Russia it seems like the plan is to maximise economic damage without causing too many Russian civilian casualties. History shows that killing a lot of enemy civilians just hardens their resolve, but if you can just make their lives annoying and inconvenient without killing anyone they know then they'll start to say "actually I don't care about reuniting the 13th Century Kingdom of Rus, I do care about not being able to fuel my car".
This is exactly why the US loses all of its wars, it goes into them with a very modest amount of enthusiasm, and then this limited enthusiasm slowly evaporates the moment things get the least bit difficult. That's how they wound up handing over Vietnam to the Communists, Afghanistan to the Taliban, and how they're somehow going to leave the Iranian regime in a stronger position at the end of the current engagement.
You seem to be under the impression that Putin can do what most civilized war leaders can do -- which is destroy the gun factories, the weapon factories. This is not possible when the gun/weapon factories lie in other countries (except if Putin wishes to get into a Big Battle with NATO).
Without that option, the only way to win is by demoralizing the other side.
I assume you're talking about Ukrainian civilians, and not American civilians. it is quite easy to argue that targeting American war-mongers is going to have a chilling effect on the American willingness to fight (and possibly the European as well).
>You seem to be under the impression that Putin can do what most civilized war leaders can do -- which is destroy the gun factories, the weapon factories.
I am not. What gave you that impression?
>Without that option, the only way to win is by demoralizing the other side.
Except that has been tried since WW1 and never worked, and it's not going to work this time either. No matter who are objectively the bad guys in a war, bombed civilians will always blame the people bombing them and support their government more.
It was the way you'd put it, which tilted a bit towards "this is genocide and Putin is bad" rather than "Putin has no good options, and is going with the least bad option, and it's not likely to work for him."
Bombed civilians seem to be supporting Fedorov the "only clean guy in the Ukrainian government," rather than the dictator who plays piano with his dick (late night variety show bit). I do not blame them.
By the same token, the people in the Donbas getting bombed by Zelensky/prior Ukrainian governments will not work to reintegrate them back into the Ukraine. They will support "the people that aren't bombing us" (or flee to "not bombed places"). Ditto the Crimeans (who really, really seem to like Russia -- more than the actual Russians).
Ukraine seems to be going in hard on "We can cause economic harm, and that will collapse Russian government." (To note: the fall of the USSR was caused by hard times engineered by the United States, so it's not the completely craziest thing).
> the fall of the USSR was caused by hard times engineered by the United States
You manage to put so much disinformation in one comment, that it is difficult to respond to each one individually, Wimbli.
In this case, "hard times engineered by US" probably refers to Soviet Union's inability to produce enough food for its own citizens. Which is somehow everyone else's fault, as is the traditional explanation when a planned economy fails. (The five traditional enemies of soviet economy: spring, summer, autumn, winter, and foreign saboteurs.)
>"Putin has no good options, and is going with the least bad option, and it's not likely to work for him."
Strategically, it's a terrible option. It has strengthened NATO despite Trump's best efforts. It keeps the Ukraine war in the news more. It puts more pressure on Ukrainian allies to provide military and humanitarian aid. And as I said, it unites the Ukrainian people behind their government and discredits all voices that hope for an actually peaceful resolution.
So no, it's not the least bad strategic option. Not bombing civilians would be the least bad option. But that's just Russia, they cant help themselves but be brutal conquerors. It's been their MO forever.
In this case doubly so, given that as Aeolian already noted, Putin is unable to destroy the gun and weapon factories supplying Ukraine.
It's a nasty quagmire - you never want to be stuck in a fistfight that's a matter of honor and pride for you but a matter of life or death for the other guy... but that's pretty much where Russia is, and the only path forward for now seems to be feeding more Russians into the meatgrinder hoping that something changes before their population loses patience with it.
It's a proxy war, just like America is in with Iran. Seeing this as "Big Guy" versus "Small Guy" rather than "Big Guy" versus "Small Guy Supported/Paid by Big Guy" is a mistake.
That makes it worse, though. It means that "Small Guy" is in an existential fight for his life against "Big Guy" *and* has, as you have noted, an unbombable supply of weapons to draw on to help him hang on..
That leaves Big Guy (a) throwing bombs at Small Guy hoping to break his will with aerial bombardment, which as you have noted in the case of Iran, and EOC has noted in the case of every time since WWI, has never worked, and (b) throwing his own people into the meatgrinder, trading lives for WWI trench warfare paced gains and praying to Orthodox Jesus that the EU has a crisis or Ukraine has a crisis so that something changes. Which isn't that much better a plan than Ukraine hoping for a Russian crisis to turn things around. Prospects are limited, and in the meantime you're just stuck hoping your people can tolerate your losses better than the other side can tolerate theirs. And that's a tricky morale problem to come out on top of when you're fighting for reputation while the other guy is fighting for his life.
Big Guy Versus Big Guy is the actual war, though. The Small Guy is immaterial, as he's no longer in charge of his destiny -- if the guns/bombs stop coming, he folds immediately. So, you get dumb theater kid tricks (well, at least you do if you're the Ukraine. Iran is smarter, and has had 40 years to dig in and get good and used to sanctions).
EU is having a crisis. Several, in fact, and some of which are caused by America (America is a very bad ally to have, I would not recommend).
I'll note that the Iraq War was basically won in about thirty minutes of bombing. There were... some very optimistic people who said Iran could be won in the same amount of time. America/Israel did manage multiple decapitation strikes.
If Putin wanted to, he could decapitate the Ukrainian government -- they certainly can't stop the missiles anymore. He thinks Zelensky does more harm to the Ukraine than good.
You have Little Russian Army and Big Russian Army (okay, if you want, little Ukrainian Army and Big Ukrainian Army), they both use the same tactics, same military structure. They're both playing "slog slog slog" and "very slow war."
Best plan for everyone would be to call a peace. Let Trump celebrate, and let the Europeans pay for more Russian blood when there's a president that won't actively sabotage the Europeans shooting themselves in the carotid.
Putin, I'm pretty sure, doesn't think he's fighting for reputation. The fight with Europe is an existential threat... (Europe's current allocation towards "clean energy" makes a lot more sense if they intend to get most of their resources from "dirty Russia" -- just like "dirty Polish lignite" right now).
(Countries come up with "existential threats" a lot more frequently than you'd think. China thought Trump was an existential threat, and acted accordingly).
Quick google (may be wrong) suggests 60% of London was destroyed during world war II. Compare to Berlin, which the same "quick google" says 11% destroyed and 70-80% bombed.
Both of these are a "large scale bomb everything" campaign, intended to level the entire place (and, probably, not directly comparable to the "enduring legacy of Communist Russia" -- Commiblocks, nigh indestructable housing!)
I was merely suggesting we look at a comparable "time" and "space" of bombing comparable structures. Makes a better comparison.
In dairy export news today (relevant to me because it involves a company created out of the merger of the co-op in my own county and the one in a neighbouring county https://www.tirlan.com/our-story/tirlan-heritage), cheap(er) Irish butter may be coming to a Costco near you!
Or maybe not, it depends. Tirlán are looking to supply Irish butter to Costco's own-brand label, Kirkland Signature, at a price lower than Kerrygold (the premium Irish export brand to the USA).
This has ruffled some feathers, because the farmers are worried this means downward pressure on milk prices. I don't see why they fear that, but then again I'm not au fait with the economics of butter exports. Maybe it will be that, in order to cut costs and maintain the margins for Costco demands, the price of milk to produce that cheap butter will be reduced.
It gets even more convoluted given that Tirlán is a shareholder of Ornua, the parent company behind Kerrygold
"Dairy co-op Tirlán has been accused of sabotaging Irish butter exports by signing a deal with US retailing giant Costco to sell butter under Costco’s private label, putting itself in direct competition with the State’s flagship brand, Kerrygold.
Farmers’ groups, including the Irish Farmers’ Association (IFA) and Irish Creamery Milk Suppliers Association (ICMSA), claim the move could result in a “race to the bottom” at the expense of producers here.
Tirlán’s Costco butter will be sold at about $6 (€5) a kilogram, substantially cheaper than the Kerrygold equivalent.
Tirlán, which separately supplies Kerrygold owner Ornua and is Ornua’s biggest shareholder, defended the move, saying it was responding to a growth in demand for private label grass fed butter."
As a linguistic aside, this is the kind of "kludge two words together and never mind if the result makes sense" that drives me batty.
"In the late ‘90s Glanbia plc was formed – Glanbia is a combination of Irish words meaning ‘pure food’."
Well - not precisely. "Glan" does mean "clean" ("glan, gsf. glaine, gloine, a., clean, pure, clear, sincere, innocent, bright, white") and "Bia" does mean "food". But in Irish language usage, to say "clean/pure food" you would phrase it as "bia glan" (noun-adjective) and not "glan bia".
Same thing annoyed me with Irish Water which is now calling itself Uisce Éireann. That is not "Irish water" (i.e. "water which is Irish, water from/of/sourced in Ireland"). That is the words for "water" and "Ireland" shoved into proximity, so the translation is literally "Water Ireland" and not "Irish Water" (which *should* be "Uisce Éireannach" or, better, "Uisce na hÉireann" (water of Ireland) which is the meaning they are going for).
This kind of illiterate branding is very annoying and insulting. 'Ara, just slap a vaguely native-themed name on it and don't bother if it makes sense or not'.
>> Dairy co-op Tirlán has been accused of sabotaging Irish butter exports by signing a deal with US retailing giant Costco to sell butter under Costco’s private label, putting itself in direct competition with the State’s flagship brand, Kerrygold.
I don't get it. If one Irish butter brand starts competing for overseas customers with another Irish butter brand, the only realistic outcome is that Irish butter exports increase. How does this sabotage "Irish butter exports"?
It doesn't seem to make sense, the best I can make of it is that Kerrygold is treated as a premium brand (plus protectionism from American butter producers, plus Trump's tariffs) and so prices range from $16-20 per kg. Tirlán (in that article) will produce own-brand butter for Costco selling at $6 per kg.
So I imagine the fear is (1) the premium brand will be replaced by retailers selling Irish grass-fed butter at one-third of the price, hence demand for the premium product will fall (2) if other Irish co-ops/dairies start producing own-brand knockoffs for American retailers, then the consumer will not realise these are Irish and not American products and so the demand for Irish butter in general will be replaced by 'I can buy American butter that is just as good' and (3) good old-fashioned scare-mongering by Ornua (owners of Kerrygold) in response to a perceived threat to their nice little niche as "the Irish butter America buys".
Speaking as an American, I would ordinarily have no idea that a brand of imported Irish butter existed. You buy the store brand. I will be shocked if Kerrygold butter is more than an utterly negligible percentage of "the butter America buys".
As it happens, I am familiar with Kerrygold butter, but that's because I frequently saw it in Shanghai. But I didn't need enough butter in China to have any memory of whether it was priced as a luxury import or as a Chinese grocery item. (In contrast, I am acutely aware that milk is much more expensive in China than it is in the United States.)
I do recall that the Kerrygold butter was more difficult to use than American butter, because it came in an oddly-shaped block without good tablespoon markings, but I imagine that what Kerrygold exports to America is packaged to meet American expectations.
> (1) the premium brand will be replaced by retailers selling Irish grass-fed butter at one-third of the price, hence demand for the premium product will fall (2) if other Irish co-ops/dairies start producing own-brand knockoffs for American retailers, then the consumer will not realise these are Irish and not American products
I don't think these are compatible fears. If (as I agree is likely) the consumer doesn't realize that the ordinarily-priced butter at Costco is an Irish import, Kerrygold will retain all of the branding value, whatever that might be, that it currently has.
If Kerrygold is severely undercut by the availability of "store-brand butter that is just as good", the implication is that Kerrygold can compete on quality with American butter (and that this, not branding, is the reason for the higher price it commands), but not with other Irish butter. In that case exporting Irish butter to be sold at normal non-luxury prices would be a 𝗴𝗶𝗴𝗮𝗻𝘁𝗶𝗰 𝘃𝗶𝗰𝘁𝗼𝗿𝘆 for Ireland; they could feasibly replace the entire American butter industry. (Though I grant that that would probably cause political problems within America and result in some kind of response from the US government.)
I know some people who buy Kerrygold and I tried it once. I didn't know it was Irish or imported though. I expect it would continue selling for a premium even if some of the cheap brands are Irish, as you suggested.
Just speculating here, but if "Irish butter" is a popular brand, and one company has a monopoly on selling it in USA, they can keep the prices high. When two companies start competing, the prices get lower.
So even if more total butter is sold, it is not obvious whether the total profit will increase. (Definitely the first company's profit will decrease, which is a sufficient incentive for them to claim that it will hurt everyone, whether that is true or not.)
Total profit might or might not fall, but the industry would be healthier in either case. It's always better to be the guy (or country) who makes vast quantities of stuff and sells it at the market price than to be the guy with a hobby in occasionally making something.
Mentioned further down - Éireann is the tuiseal ginideach, so Uisce Éireann is 'Water of Ireland', which is the grammatically correct way to phrase the name of a state body.
Éireannach is almost always used in the narrow sense of nationality, it has broader usage in English because they don't have the genitive case. I checked the national corpus (https://www.corpas.ie/ga/cng/?q=Éireannach) and the first 20 or so cases all refer either to a person or a work (play, film, book) which has a national origin, but even that use is pretty specific.
There are lots of abuses of the language out there but this is not one of them.
Thank you for the grammar lesson, we all remember struggling with the tuiseal ginideach!
Yeah, I agree "Éireannach" is used more in the sense of "Irish man" and a personal identity, but "Irish Water" are presenting themselves as "The state body providign the services of Irish water in Ireland for the Irish people that is water of Ireland and not, for example, British water or UK Water or other water body", so the 'national identity' sense could be argued to fit there. I don't like it, though, since it doesn't sound right.
I do think it should be "Uisce na hÉireann" but that probably is just a personal quirk of preference on my part. Too much "official Irish" is badly kludged together and poorly translated, and when you've been taught a particular dialect in school then Standard/Official Irish is trying to blend Connemara, Ulster, Munster and increasing Dublin Irish usage into a Frankenlanguage that sounds good to nobody but is 'representation'.
> Éireannach is almost always used in the narrow sense of nationality, it has broader usage in English because they don't have the genitive case.
1. I don't understand the primary claim. As far as I'm concerned, Éireannach has zero usage in English.
2. There are two things commonly referred to under the name "genitive case" in English:
2. (a) The clitic particle 's, as in "my dog's fur", is often thought of as a case marker. Historically, it does descend from the Old English genitive case ending. It's difficult to call it a case marker today, because it doesn't attach to nouns. This is widely recognized, but people nevertheless persist in calling it a "phrasal genitive".
2. (b) English pronouns come in subject and object forms, and (unlike the particle) these forms are very reasonably called "cases". It is common to extend that logic and say that 𝘮𝘺, 𝘺𝘰𝘶𝘳, 𝘩𝘪𝘴, 𝘩𝘦𝘳, 𝘪𝘵𝘴, 𝘰𝘶𝘳, 𝘺𝘰𝘶𝘳, and 𝘵𝘩𝘦𝘪𝘳 are pronouns in the genitive case. I tend not to agree with that - for one thing, they can't occur where the grammar calls for a noun - but it's a mainstream viewpoint.
Sorry - I'm comparing adjectival use in the two languages. In English the adjective 'Irish' stands in the for the genitive 'of Ireland' in many contexts where the latter is idiomatic in Irish. In Irish the adjectival form is basically only used to refer to an Irish person or an Irish citizen - the round translation
'Uisce Éireann -> Irish Water -> Uisce Éireannach'
alters the meaning though both collapse onto the same English phrase.
More generally, and I think this is the point Deiseach was making, pulling words out of a dictionary is not the same as translation.
When I want to start a grammar argument, I tell people that 'tomorrow' is an adverb.
You might be interested to know that "because they don't have the genitive case" is probably not a great analysis; some languages have dedicated gentilic forms and others not so much. Latin uses its genitive case extensively, but you have all kinds of different options for describing the relationship between a place and a person; you could say "of <wherever>" in the genitive case, although I tend to suspect you wouldn't, you could use a generic adjective, or you could use a special gentilic adjective.
The gentilic would usually be used for indicating that someone comes from a place - "Athenian" in Latin is "Atheniensis", from "Athenae" (the city) plus -ensis, the gentilic suffix. But interestingly enough "Roman" is different; the Latin is "Romanus", from "Roma" (the city) + -nus, which is a generic adjective-forming suffix that _doesn't_ carry any special "from a place" sense.
> When I want to start a grammar argument, I tell people that 'tomorrow' is an adverb.
This seems incomplete. "It will happen tomorrow" makes a strong case for "tomorrow" being an adverb; "Tomorrow is Christmas" looks more like a noun.
"𝘛𝘰𝘮𝘰𝘳𝘳𝘰𝘸 is an adverb", as an unqualified claim, makes as much sense as saying "𝘳𝘪𝘨𝘩𝘵 is an adjective". There is an adjective spelled 𝘳𝘪𝘨𝘩𝘵. Not every use of the syllable is a use of that adjective, or any adjective.
I do enjoy me some Kerrygold, but I'm not a Costco member, so I guess I'm SOL on getting it on the cheap.
My pet peeve in the US is companies that try to shoe-horn all or part of "America" into their name, as if some vague association with patriotic sentiments is going to help sell their otherwise unremarkable products. I think the worst offender is Ameriprise Financial, a name which rolls off the tongue with about the same gracefulness as my grandma can ride a surfboard. My joke about this is that if I ever get rich, I'm going to open a chain of OB-GYN clinics called AmeriC**** just to take this idiotic trend to its most absurd conclusion.
LOL, Claude says that you are wrong, because there are names like "Iarnród Éireann", "Dáil Éireann", "Tuaisceart Éireann", which follow the same pattern as "Uisce Éireann".
What an odd thing to post. You're responding to a native Irishwoman, who's sharing her (interesting!) firsthand experience of various faux-Irish corporate brand names sounding awkward to her ear as an Irish speaker. So you take it upon yourself to drop in this random contradictory piece of text from a model-unspecified English-trained generative language machine, that you didn't even bother to request references for?
With an *LOL*, forsooth, as though you were delivering some sort of epic gotcha?
You don't claim any expertise in Irish yourself, so I conclude you are fully unqualified to access any of the nuances that might help you determine what the model's text is referencing or to what extent this is hallucination, irrelevant misinterpretation or valid corrective to the original statement. What are you adding here? Did you think everybody would admire you, because you copy-pasted some writing into a text box and then copy-pasted that into a comment to us?
> What an odd thing to post. You're responding to a native Irishwoman, who's sharing her (interesting!) firsthand experience of various faux-Irish corporate brand names sounding awkward to her ear as an Irish speaker.
Why would you assume that, as a native Irishwoman, Deiseach must also be an Irish speaker? In general that is not the case, as we see here.
In fact, Viliam is correct. Éireann is the genitive case -- 'of Ireland' so Uisce Éireann is 'Water [Company] of Ireland'. This is the idiosyncratic and correct name for the organisation in Irish; all the other examples given are correct and appropriate. While my credentials shouldn't be relevant, I'm a native speaker and it's my usual working language. Furthermore, the AI's are really good at Irish grammar now. To the extent that I rely on them for proofing my own writing - they're not always right but they'll identify the grammatical slips I make habitually for me.
I'm guessing that Deiseach is not a fan of the tuiseal ginideach (which is and was poorly taught in school). I'll agree with that all the agricultural co-ops have horrible names. Tirlán is, if any thing, worse than glanbia. Tír is 'country', and lán is 'full' (but in the sense of being filled to the brim, not complete, or whole). It has an ethno-nationalist undertone to me, which I don't believe is actually intentional.
Tirlán is terrible, and an example of what I meant by knocking together two separate words according to dictionary meaning but ignoring if they are used that way in practice. The logo is also teeth-grindingly horrible. Look at this and tell me if it means anything or does anything other than drive you to fury:
It really is one of those "did you get a six year old to draw it?" images, and yet I bet a design company was paid a hefty fee to come up with it.
"Tirlán logo and name reflect the Irish Co-op's heritage. The name combines tír (land) and lán (full) to mean "land of abundance". Its graphic symbol inside a circle is inspired by Úr, an ancient Ogham letter meaning earth, soil, or clay, symbolizing the deep bond between the farmers and their land."
Ogham my backside, it looks like chicken scratches. As you point out, that term does not mean "land of abundance" (that would be more to do with "flúirseach", or according to the old Foclóir Gaelige-Béarla from school which I dug out now, "raidhse", fem. 4th declension noun, "abundance").
Are you of an age that you remember the Loriel ad that finished 'Inspired by the science of genes'? It's equally awful. I suppose that there's some pressure to balance an Irish image with a name that people can pronounce if they know nothing about Ireland.
I quite like Flaithiúil as an alternative to Fliúirseach (which I associate with a foolish generosity, as of someone buying a round before the previous one is finished). Its meaning is more like 'noble generosity' which is appropriate in the context. Just pulling a word out of the dictionary blind to its context is not great.
He wrote thousands of pages, the definitions are minimal but each word is illustrated with as many as 30 examples in use. One would imagine that the Execs in Tirlán are unaware...
Hm, I can kind of see it. If you're right, then hilarious self-own for me, and @Viliam, I apologize for being so testy. I've been dying a little bit every time Scott pastes a random LLM screenshot as "source" for some grand claim, and I guess it finally felt like too much.
I assumed that Deiseach would understand it as a friendly challenge to prove me/Claude wrong. Or that something else interesting would happen.
But yeah, reading the comment again, it is plausible to read it as "AI-powered mansplaining". The initial "LOL" was meant to express my confusion at two entities disagreeing on a topic I know nothing about. But looking at it with fresh eyes, it is also possible to read it as an expression of aggression. Sorry for that.
Deiseach is not a fragile snowflake and can handle that, I assume.
In the event, you were right. Emily's comment seemed to me based entirely on who you were and who Deiseach is (to the extent that these can be inferred from the internet).
I notice in my own life that people are overly quick now to dismiss something *because the answer was generated with AI*, even if the answer is otherwise correct and appropriate. This just pushes people to hide their use of AI, which is not what anyone wants.
Scott's MacGregor joke got me wondering whether sheep are injured when human males have intercourse with them. (The reason I was curious is that it does not seem morally wrong to me for someone to do that, unless the sheep is injured or badly frightened). Nobody here knew, and GPT Sol was not allowed to discuss the topic -- the censors shut us down. But today I managed to get the question past the censors, and the answer is that In an adult ram, the erect penis is 10-14 inches long and 1/2-1 inch in diameter. And of course a lamb's vagina has to have a good amount of stretchiness. After all, lambs come out of it. So I don't think intercourse with a male of our species would cause injury. In case you've been wondering. Or even sleeplessly fretting about the question. Or sheeplessly fretting.
Dogs really like peanutbutter (and chicks really do film themselves...).
That's probably why it's the cops conducting this research, honest to god -- they're allowed to look at all the pornography (in the interest of catching the Bad Criminal Guys -- you may know them as Oligarchs).
Human penises aren't barbed, unlike, say, cats (which has a relatively straightforward "finish copulation" reason). I have no idea why pig penises are corkscrews (Thank You Drawn Together, for the Anatomically Correct Animation... words that probably no one ever said, before or since).
The same argument could be made about human vaginas, after all, babies come out of them. Unfortunately, reality is more complicated than that - both for humans and for sheep too.
Sure, but I don’t believe your size estimate either. This feels like common sense if by you’ve spent time with animals up close, but if here are some veterinary pictures to illustrate: https://visgar.vetmed.ufl.edu/en_ovirep/penis/penis.html
From the paper you’ve cited: « The penis in the ovine is a tubular fibro-elastic organ, around 40 cm long with an exiguous portion of erectile tissue. »
Exiguous means scanty or meager, i.e., yes, the entire structure is 40 cm long… but the erectile / external part is small. There is a picture of a flaccid penis in that very paper which shows this clearly.
The pictures I linked show ram’s penises (with human hands in some for rough scale) - including an erect one, which is difficult to judge for size but does not look much larger (than flaccid).
OK, I looked at the photos. Looked like the penis was somewhat thicker than the fingers on the hand that was holding it. I have small hands, and all of my fingers look to me like their diameter except at the very tip is 1/2" or more, and my thumb looks to be 3/4" wide, and and inch wide at the base. If the hands in the photo were larger than mine, then their diameter would have had a minimum and max larger than mine. So all that seems to line up well with the diameter figure Sol gave.
Wow $250k seems like a very very big money for this kind of bet, maybe they secretly want to give you money instead? I agree that I won't take this bet even if I'm a lab leaker mostly because I don't believe that there'll be any conclusive evidence that can sway the argument either way in this stretch of time.
Related to organ donation, blood donation is another needed service that’s hard to incentivize. Goodlabs.com is a new diagnostic company that provides lab tests in exchange for donating blood; if you’re already donating blood, why not get some free lab tests at the same time? This is a great incentive to promote personal health and help out the blood supply.
IMHO, everyone should regularly test some basic blood markers (e.g. lipids, HbA1c, homocysteine). Plus, you still get the free cookies and orange juice they usually give you.
I tried to donate plasma once, specifically for the financial incentive, and they turned me away because my pulse was slightly too high. Ironically, the second time they tested it was even higher, because I was afraid of it being too high and not getting the money. One way to interpret this situation is that incentives aren't enough to overcome a flawed system. I understand why they want pulses to be in a healthy range, but surely slightly unhealthy blood is better than no blood at all? If there truly is a massive shortage of blood, why not loosen the standards a bit until there's no longer a shortage, or at least so long as that does no active harm to the recipient?
If that's the case then you could have an informed consent system where they say "look this is risky because [10 billion reasons] so you can only do it if you promise not to sue", which they kind of already do in general because donating blood is always a bit risky anyway (there's always the off chance that your body doesn't replenish it right away, or you get an infection from the equipment, or other unknown unknowns).
Very briefly: since the USA buys blood rather than relying on voluntary donations, things such as "hey we got an entire prison full of guys who could make some money for themselves and us by giving blood" happened. Trouble is, in prison you got guys who are drug addicts and (at the time) HIV positive plus other blood-borne diseases like Hepatitis C.
Guess what happens when you mix in "slightly unhealthy blood" with other donations and produce the plasma for export?
Canada, Ireland and the UK were three countries that ended up with government enquiries and compensation schemes (often bitterly contested by those same governments) for the haemophiliacs and pregnant women who had been infected by those batches of blood.
Now we have modern techniques of screening and processing to avoid similar problems, but yeah: "slightly unhealthy blood" is covering a multitude of sins.
For what it's worth I'm not proposing mixing all the blood together. You could have a tiered system where those at the top of the queue would get the healthy blood, and then people further down (meaning in less immediate need of blood) could choose to take the risk of getting less ideal blood or wait for the good stuff. And I'm still in favor of screening out blood that actually carries diseases, since that would obviously harm any patient.
"Hard to incentivize" bears very little resemblance to my experience with blood donation. Yes, there are a lot of people who just won't do it, ever, but for a full day off work? I think most people will donate blood.
I am actively starting to intentionally disengage with comments here because Substack started spamming me with 25 'also replied' notifications for every 1 real notification about someone responding to me. Does anyone know a way to fix this?
Yeah, that sucks. Happened to me too, made a top-level comment that started a long discussion, and the notifications keep coming even if deep down in the comments people are already discussing something else.
I can see, under "settings", under "engagement", 4 options for turning notifications on or off:
-likes
-replies
-restacks
-mentions
I haven't tried turning any of them off, and I _used_ to just get notifications from direct replies to my comments, but I'm now also getting notifications from 'also replied'. It is imaginable that the 'also replied' might get classified under sort-of kind-of mentions, so it might be worth turning that off and seeing if it helps. If 'replies' is really controlling both direct and 'also replied' then my guess is we are all stuck :-(
We really need the ability to turn off notifications for specific sub comments. The also replied is even more annoying. I once created a comment which generated a lot of sub comments, mostly not to me, one pair of responders were arguing amongst themselves for a week! I just had to mute the whole open thread.
Me and the fiancée are planning a “mini-moon” in SF in late October after our wedding. Since this place tends to be tech-heavy, figured I’d ask for recommendations, dos and donts, hidden gems and overrated things, etc.
Good hotels? My only two things I’d really like to see are the CHM (obviously) and some giant redwoods. I’ve been to Chicago, Philly, and Seattle so homeless and vagrants don’t bother me. Would prefer not to rent a car if I don’t have to, but can adjust if need be.
If you like whales, and are willing to change your mind about renting a car (and driving ~2 hours from CHM), there's whale watching in Monterey Bay. Specifically, my watcher of choice is FastRaft.com - instead of a yacht-sized boat where you're 15 feet above the water, you're in a raft that seats 6, and you could reach over and touch the bay. The captain and her husband are super cool guides.
30 minutes further south is Carmel-By-The-Sea, which is a fun walkable beach city. Overall, I consider this a day trip - spend most of your time in that area, maybe stop by Gilroy's Garlic World, or alternately enjoy the drive up and down the PCH, or duck up route 9 to Big Basin Redwoods.
Also, 15 minutes from CHM is the Winchester House. I've heard, but cannot confirm, that David Friedman's house is nearby, but I don't think any meetups are scheduled right then.
Edit: Note that you will have to rent a car to see redwoods, unless there is some sort of bus tour you can book. Also, a day trip to Yosemite Valley is doable, though the falls will not be impressive in October. There are Sequoias in the south of the park, near Wawona, but that means more than a day trip. The Wawona Hotel is nice, as is the Ahwahnee in Yosemite Valley.
Others have mentioned the cable cars, but I'd also recommend the (free) Cable Car Museum, it's both a fun lens to learn about SF history and also itself a fully-functioning piece of machinery that keeps the cable cars moving. My personal favorite is to ride a Waymo to the cable car museum (optionally visit chinatown in the process), and then hop on a cable car to go to fisherman's wharf from there to get the full bracketing of first-and-latest transit innovations from SF.
I also definitely recommend the Exploratorium. I'm not aware of any museum quite like it in the world for those who like hands-on figuring out of stuff.
Last I was there, Jtown was a blast. Walk the streets that are just stairs. Take the cablecars as the "very last thing you do" (we found locals on them, playing "freaks come out at night" -- it was cool, and very much less crowded than during the day).
Go to "Lindy in the Park" on Sunday in the Golden Gate Park. Get there at 11:30-ish (so you're not late). Free lesson from 12pm which just has you bouncing in place to start. It's a special gem provided by it's 2 founders, Ken and Jen who have been running it for 25 years. From there you can check out the rest of the park. The De Young Art Museum, the California Academy of Sciences, The Conservatory of Flowers, or the many other outdoor things in the park like the Japanese Tea Garden, the San Francisco Botanical Garden, Blue Heron Lake (paddle boats), the Bison Paddock with actual Bison in the middle of a city. You can rent bikes between the two museams or rent the city's bikes from the racks. Sunday is the day to go as parts of the park are blocked from cars.
If you want to experience cognitive dissonance and existential irony, make sure to visit the Mission district, especially the area around Folsom and 18th. Scrape your pennies together for a $7.50 Latte at a sidewalk table guaranteed to be surrounded by wish-they-were-still-20-somethings talking up AI and disavowing Crypto and wokeness (so 2020).
Then just cross the street and mosey down one of the side streets where the un-housed and barely-housed scrape by in campers festooned with rusty bikes. tents, and lean-to's
CHM = Computer History Museum? If so, I wouldn't actually call it a must-do. It's kinda cool, but it's not very accessible from SF proper and not worth the trip unless you have other things in the South Bay to see. You could combine it with a trip to the Google campus (and maybe Stanford's campus, or the radio telescope, or Apple's campus) if you really wanted the full Silicon Valley day trip thing. You'd definitely want a car.
Anyway - the thing about SF is that virtually all of the touristy things are also worth seeing. The only thing that many people would say is overrate is Pier 39 itself, but the rest of the Wharf is cool.
Walking on the GGB is cool. Biking across it is really cool. If you're reasonably healthy I would suggest renting a bike in the Wharf, then riding to the bridge (and over and back).
For redwoods, the best way to do it without a car is to take the Ferry to Sausalito then take the Muir Woods shuttle. Note that you need reservations in advance for this, and in the off-season the shuttle only runs on the weekends. If you wanted to do another day, you'd need a car (and reservations). Note that cell coverage is bad out there so Uber isn't recommended unless you can pre-schedule a return.
If Muir doesn't work out, you'll need a car, but there are tons of great hiking spots within an hour's drive. Mt. Tam has a lot of other spots besides Muir though you'll need to do some curvy road driving and parking can be tricky. I also love hiking in the East Bay, if the weather is nice - Reinhardt Redwoods has gorgeous scenery and views.
I have some family in the area; I don't have a lot of intel, but a few things that were memorable/enjoyable:
Soba-Ichi - if you're in Oakland at any point, this soba noodle spot was amazing
Pâtisserie Rotha - is a great bakery if you find yourself in the Berkeley area, but it's line up at a takeout window, not seated, so be aware if you go.
The Cultured Pickle - also in Berkeley, a pickle shop that lets you reserve seats on the weekend for pickle-themed meals. Really fun if you can get a seat.
Lazy Bear - Fine dining restaurant we took one of my in laws out to as a 'thank you' for a favor. Pricey, but it was a great experience if you want to spend money on a night out.
Church Of 8 Wheels Roller Disco - Old church converted into a roller rink. I still haven't gotten my family to actually go, so I can't *really* vouch for this, but it's the kind of hilarious weird thing that I love, so if you go, tell me if it's fun.
I don't recall anyone ever explaining career options to me in terms of categories and types rather than specific jobs. I'd rather set my kids up with a better understanding of what sorts of careers they want to go after rather than simply specific jobs. This is one of the places that most favors analytical breakdowns and explanations of things, so I'm posting here in hopes of feedback.
Unfortunately, it appears substack doensn't allow formatting in the comments. I'll use *, **, ***, etc. for outline component depths.
--
0. Important Terms
*A. General Classifications.
**1.Blue collar: Jobs you do with your hands and body. Tends to be more physically active. Can range from mind-numbingly repetitive to varied and intellectually demanding. The most specialized jobs pay very well. Often carries a higher level of physical risk. May wear out your back and joints earlier in life. Is sometimes considered lower prestige / social status but may also generate interesting stories, excellent “do it yourself” abilities, and a healthier body.
**2. White collar: Jobs you do with your brain, usually sitting at a desk or in meetings. Tends to be less physically active. Can be mind-numbingly boring at entry level. Often based on developing or following a process repeatedly (accounting, sales, medicine, management, etc.). Communication and dealing with people tends to be key to success.
*B. Standard work week: In America, this is typically a 40-hour week, most often 5 8-hour shifts Monday through Friday. Some jobs, like fireman, medical professional, or offshore oil rig worker, use a rotating schedule that averages out to something close to a 40-48 hour week. 40 hours per week x 50 weeks per year (2 weeks of holidays plus a couple of days unpaid sick or family time) ~= 2,000 hours per year worked. In general, you can take the hourly pay rate associated with a job and multiply it by 2 and then by 1,000 to get a fairly accurate figure for annual pre-tax earnings.
*C. Compensation: How you get paid.
**1. Hourly: You are paid by the hour for the work you perform at a set rate per hour, without a direct link between your speed & skill and your pay. If you work more than 40 hours per week, you typically get paid 50% extra beyond hour # 40 (Overtime pay). If you work nights or weekends you may get paid extra per hour.
**2. Salary: You are paid a set amount per week/year no matter how many hours you work. Legal restrictions mean this is typically a payment method reserved for managers, who are expected to make sure things get done. In some jobs, this means 40-50 hours per week. In other jobs, if you’re in management and someone calls out sick, you are the backup plan and may end up working 60+ hours per week.
**3. Commission: You are paid for performance, typically as a percentage of revenue or a percentage of the company’s profit (revenue from sales minus cost of fulfilling the sale). There may be a guaranteed base pay, or you may be 100% commission. The upside is a higher to much higher potential take-home pay, but the downside is a lot more variance in take-home pay from month to month. Depending on the job, you might need to be available during non-business hours.
**4. Self-Employed: You directly sell your labor or the products of your labor to others. You have a lot more freedom to choose what to do, when to do it, etc., but you are also responsible for tracking your income and ensuring future customers/sales. When you work for a business, they pay some employment-related taxes that fund government programs for medical aid and retirement. When you work for yourself, you are the business, and you pay the business’s share of the taxes as well as the taxes you would normally pay. This is called “self-employment” taxes. However, profit (money that comes in minus expenses that go out) often needs to be re-invested in hiring new people, training, maintenance, additional equipment, etc. You don’t get to pay yourself a lot until year 5, 10, or 15.
*D. Customer Type
**1. B2C Business To Consumer – restaurant, retail, home services provider, windows, painting, roofing, lawn care, anything that sells to the general public
**2. B2B Business to Business – manufacturing, technical services, logistics, raw materials or finished products, anything where the customers are business
*E. Place of Work.
**1. On-Site. You work at a job site, factory, or other place of business daily.
**2. Field. You spend much of your time visiting customer locations, or are in a role where you visit many different locations for your company. You probably are still on-site at a home base location regularly.
**3. Remote. You work from home or wherever you are, instead of in an office. There is no commute, and often no hourly tracking for productivity. You have more time in your day, and often the ability to intersperse personal tasks with work as long as you manage time well. You also miss out on in-person conversations and networking, and are less likely to be promoted, and are more likely to bethe first to be laid off when downsizing due to those lack of connections. You’re also less likely to be able to fully disconnect during a vacation. Remote can be lonely, and video calls do not make up for that.
**4. Hybrid. You work remote a couple of days a week, and on-site a couple of days a week.
Maybe add a side note that many rich people make money in ways that are not present on your list, and probably not available for ordinary people. Such as:
"Money makes money" -- if you inherit lots of money or land, you can make more money simply by putting the money in an index fund, or renting the land. Thus you will never run out of the original money/land (except if a revolution happens), and you have a guaranteed lifelong income.
Jobs that are strongly regulated and you have to be a part of the old guys network to get them. The exact list probably depends on country, but I am thinking about doctors, advocates, notaries.
There is a career type where you work in politics, e.g. on the municipal level, and take bribes. To get this kind of job it helps to be a part of a larger clan, e.g. if your older relatives have that kind of a job, they can leverage it to get a lower-level version of the same to you.
Tell me you're not from the US without saying it. The nepotism you describe is not systematic here, and notary is a job usually handled on the side by administrative staff.
I considered investing to be outside of the scope of this document... and even if someone was inheriting a lot, I'd want that kid to have a career and responsibility vs. not working. It's part of what we were created to do.
Nepotism is systematic here. The Carnegie Mellon Mafia, if I must cite something that isn't religio-ethnic in nature. But the "woke hires" and "diversity hires" as well.
A key advantage of self-employment - it is much harder to get bank loans for cars, mortgages, etc. This forces you to live within your means and greatly speeds up wealth accumulation.
Mortgages wind up leveraging you, though. It's the traditional middle class way to gain wealth (citing Black Wealth White Wealth if you really need a citation).
It's paper wealth, though. If you sell your house and don't use the money to buy another house, you're homeless or renting. The only time it turns into cash is if you have a consolidation (death/inheritance, marriage).
I had to build up 2 years of commission history to get a mortgage. I assume it's similar for self-employed people? Two years of steady personal income?
If you wanted more detailed critique and suggestions, I'd happily make comments and edits in a Google doc if you share a link. (Or DM one...I have no idea if Substack has that feature, never used it.)
I've been interested in work -- like, "how people make their living" -- for a while, and I think this is a cool document with a lot of potential, even bigger than just handing to your kids.
Thanks! I deliberately avoid Google docs and Google accounts (one exists for my phone, never use it) so I'm not sure where to host something like this.
It could be worth mentioning in the compensation section that performance bonuses may exist, and can even be a majority of compensation in some fields.
Don't forget to mention that stock options usually don't work as advertised. For various reasons the options you own may become completely worthless *and* you may have to pay tax on them. The rule of thumb is that unless you are an expert on stock options, you will probably get scammed.
(For example, it may turn out that there are various kinds of stock, and your boss and the investors get a "priority" stock which is actually worth money, and you get "normal" stock which is just a piece of paper. Or your contract may say that you will only get the stock options after staying 5 years at the company, and of course if the stock options would by some miracle end up worth a lot, the obvious move on the employer's side is to fire you after 4 years and 11 months and keep the stock options for himself. There are also other risks, but those are more complicated to explain.)
One nice thing about AI becoming ubiquitous is that people can now put their employee agreements through the AI and get an opinion on this, without having to pay for a lawyer.
I do feel like most people savvy enough to get a decent chunk of their comp in stock options understand vesting and preferred shares and things like that, but you're right, it's certainly not a given, especially for new hires fresh out of college or grad school.
Sounds like halo effect; someone good at coding is not necessary good at legal things and investing. You seem to be in a lucky bubble where people are just good at everything. Or maybe in a bubble where people communicate about these things.
I was quite stupid about these things even in my 30s. When I tried to ask my friends, they also knew nothing. And trying to figure out things on internet has the problem that when you are a beginner, you just find lots of contradictory advice and you don't have the skill to figure out which is trustworthy. (On top of that, most advice is for Americans and does not necessarily apply to me.)
Finance pays out big bonuses annually too. Amount is a function of your own performance, the department performance, and the firm performance. It's usually not contractually guaranteed, so it's not always very transparent what amount to expect, but there's usually some consistency to them which you can learn through research.
I don't know as much about other industries, but I think consulting also has significant bonuses. Also I think many sales-type roles use performance bonuses as a substitute for commissions.
Moving beyond the traditional 5 x 8 work schedule has its advantages. I work for, uh, an Extremely Large Online Retailer and the standard full time workweek is four 10-hour days, though there are some variations. From what I understand this is because most facilities are 24/7 and four 10's is an easier way to maintain staffing.
In any event working a 4 x 10 schedule has many good points. The extra two hours a day go quickly, and you get three full days off a week. Two of those days usually are on weekdays, as almost all full timers work Saturday or Sunday, and weekdays off are ideal for running errands and shopping.
*Key advantage: Someone else pays the bills. Some measure of stability. Often comes with benefits like help paying for some insurance, 401(k) retirement plan, etc. Usually has the best chance of not needing to work more than 40-50 hours per week.
*Disadvantage: Someone else decides what you get paid in terms of hourly, salary, or commission plan. If you want a change, you have to negotiate for it and prove your value or move to another company.
*A. Blue Collar
**1. Build: You make things or create resources for things to be made with. Assembly line at a factory or processing facility (chemical, food processing, etc.), construction worker (concrete, timber, structural, electrical, roofer, etc.), machinist or welder, oilfield or mining work, lumberjack, fisherman, farmer, etc. Highest-risk jobs but some can pay very well for the long hours.
**2. Service: You keep things working or install completed products. Mechanic, air conditioning repair, electrical work, equipment installation & maintenance (A/C, industrial equipment, data center equipment), power line repair, plumber, furniture assembly, janitor, etc. Tends to have stable demand but often requires sweating and crawling or climbing around to do the job.
**3. Logistics: You get things where they need to be. Warehouseman, forklift operator, truck driver, mover, railroad worker, rigger & crane operator, sailor. Pay is often not great aside from specialized technical roles like rigging.
**4. People Care: You provide direct one-to-one and hands-on services to people, not things. Nurse, teacher, police officer, EMS, daycare, firefighter. (some of these may be classed more as 'pink collar,' but they are hands-on jobs, not desk jobs)
*B. White Collar
**1. Customer Service: You work with customers on a day-to-day basis. This can be for B2C or B2B. You answer phones and e-mails, convey product knowledge, help customers make choices, take orders, handle scheduling, appointments, orders, shipments, and jobs, deal with first-line customer complaints, etc. Your job is basically moving information around politely so that customers get what they need and the company makes money meeting the customer’s needs. Usually hourly. Often entry-level or first jobs. Some Customer Service jobs are more blue-collar – think bagger & cart handler at a grocery store, gas station attendant, etc.
**2. Sales: This is like customer service in many ways, but you are working with customers who are not customers yet. They either don’t know if they need what your company provides, or they are not sure if you are who they want to buy from. Your job is to help your customers by matching your company to them. You are often paid on commission, so this can pay very well, but if you have no sales in a week, you have no paycheck that week. You deal with having your calls and messages ignored often and being rejected or told no a lot. If you are in B2C, you are often meeting someone one time and have 30-60 minutes to get them to trust you and buy from you. If you are in B2B, it can take anywhere from 3 months to 2 years to make a sale depending on what you are doing. People buy from your company because they trust YOU and that YOU know what you are doing and are steering them in the right direction. You don’t have to be an extrovert, but you do have to be likeable and persistent. There are many ways to sell, from a more technical product-oriented approach to having an internet presence and followers to being a professional networker who is friends with everyone, plays golf with customers, takes them out to dinner etc. In addition to traditional “sell a product or service,” this path includes real estate agents.
**3. Technical: You provide a specialized service, often after “sales” or “customer service” has been involved. Your value is your expertise, experience, and knowledge. Architects, engineers, doctors, lawyers, aircraft pilots, accountants, actuaries and mathematicians, psychologists & counselors, industrial designers, professors, consultants who offer advice or design work in specific fields, IT (information technology – computers, programming, data centers). Can be very highly paid, but often only after an apprenticeship or extended period to gain experience. More likely to be contractors.
**4. Management: You’re in charge of customer service, sales, the technical branches, and any blue-collar roles your organization has. You are responsible for getting everyone to follow procedures, work together reasonably well, deliver service or product on time, and generally keeping things moving and profitable. You also get problems or complaints that others can’t solve escalated to you, and are responsible for dealing with any internal problems like people who irritate each other or don’t want to do what they’re supposed to do. The pay is often good but sometimes the hours are long. At higher levels, you manage managers and work your way up to C-level (Chief X Officer) and run the whole company, which typically pays very well. However, if there are problems, the buck stops with management, so job security may not be as great unless your company is stable and their industry is stable. In addition to managing departments or divisions in a company, there’s also a sub-role called “project manager” where you oversee a specific job or task, like building a building or completing a software program. It has more limited scope in terms of responsibilities, and involves tracking details, progress, and deadlines from dozens to hundreds of people or suppliers and making it all work together smoothly.
II. Contractor or Hired Specialist
*A. Solo Operator: You don’t have anyone who works for you. This is just you and what you can do delivering value to your clients.
*B. Career field: This is almost always B2B. This can be blue collar or white collar, but typically involves you having a skill at which you are specialized and experienced, with several years to a decade of experience in. Companies bring you on for that specific skill and pay you part-time, full-time, or by the job to do it. Payroll for small companies, on-retainer lawyer, consultant, doctor (many hospitals don’t employ their doctors), programmer, project manager, trainer, home inspector, insurance adjustor/inspector etc.
**1. Entry: You normally work in your field for a company or for someone else for a while to develop your expertise and experience.
**2. Business: You gain business through your network – people you know who know you – and your reputation. You have to sell your services, making this a “sales” role part-time by default in many cases.
III. Own Your Own Business
*A. Money: You own the business. If it makes money and grows, you can pay yourself from the profits. Owning a successful business can pay very well, and it can grow to become something that can support future generations. Alternatively, you can build the business with the intent to sell it and walk away from it with a full bank account. Some people are serial entrepreneurs, meaning they start a company, build it to success, sell it, and then start another company.
*B. Risk: Many small businesses fail. If there is no profit, there is no take-home pay for the owner. If your business fails, the people you hired to work for you are out of a job.
*C. Responsibility: You carry legal liability. Early on, you are sales, accounting, human resources, operations management, training, and project lead on everything all at once all the time. Expect to work 60 to 80 hours per week.
*D. Scale: “Scale” refers to how much a business can grow. Some businesses have more growth potential than others. Some businesses are easier to grow than others in terms of potential customers, investments needed, etc., although this can also be a defense against competition. Make a piece of software that everyone wants on their phone: 100 million to 500 million sales, IF you’re part of the 0.01% that actually make it to that level of popularity. Make a restaurant? It can only seat so many people and service so many customers per day. Open a warehouse? Eventually you run out of space. Make a widget sold to air conditioning manufacturers? There are only so many manufacturers in the US. Sell to all of them, and you need to branch out if you want to grow further. You cannot scale or sell a business if it depends on you to continue to function. If it falls apart if you’re not present, you’re self-employed with extra expenses. One restaurant that is only in good shape if you are there will never become five. One restaurant that functions with you only stopping in twice per week is easy to turn into two, four, or ten if you have the money, the customers, and can hire the right people.
--
I'm sure I could go in depth with more examples of jobs. Are there any other important "this dictates what your job feels like for 40 years" meta-categories worth explaining to a preteen/teenager?
I really like this breakdown and would love to see it hosted somewhere with better formatting tools.
I would add a dimension for creative vs. deterministic pursuits, since having subjective vs. objective success criteria can make a big difference in job satisfaction for different sorts of people. An in-house designer has very different workflow than an in-house accountant, even though they both spend all day in front of a desk and report to a boss.
*A. Deterministic. Your job follows rules and processes for specific, set outcomes. Your job involves following and improving those processes. There’s room for skill, but you’re usually not making anything new. Examples: Accounting, construction, some engineering work, factory work, medical, transportation, most IT work. This is more common with larger companies.
*B. Creative. Your job has rules and processes, but a key part of your job involves using your taste, judgment, or artistic skill. Examples: Architecture, art, designer, some engineering work, public speaking, sales, teaching, writing. This is more common with smaller companies or solo operations.
Looks good. For the sake of the future teen reading this, be very clear that success in a creative job doesn’t depend on _your_ subjective taste, but on the boss’ or client’s taste
And that you can be forced to write/make Nazi-slogans, or other extremely impalatable stuff. (Yes, still salty about cakes. Try talking to creatives...)-- this is American-based, if you're not.
Given that you mention benefits, and have a whole section on compensation, it's probably worthwhile to have a list of the kind of benefits one normally gets from their employer
Good idea. Added to my copy. Insurance, flex spending, 401k, sick pay, vacation pay. That's about all I'm aware of aside from ancillary options (supplemental insurance, small life ins policy, etc.)
- Business to Government: similar to B2B in many ways, but often defined more by adherence to complex regulations and processes than by the free market.
* No mention of piecework in the "how you get paid" section
* No mention anywhere of the hospitality industry - where would you put that?
* Would be interesting to add the distinction between working for/with private enterprise, state infrastructure (military, civil service, health service in the UK, etc) and third sector / not-for-profit enterprise
I was thinking about using this with my teenager but I have an overriding worry that it reflects more how the world WAS (and still is to some extent) and less about how the world WILL BE by the time they come into the workforce. What about which careers are more or less at risk from AI? The importance of transferable skills that aren't tied to any one specific job or industry?
I would lump piecework in with commission. In both cases it's pay based on results, not time.
I would put Hospitality (both hotels and restaurants) under Customer Service, although they don't fit neatly into blue collar or white collar. The line cook & room cleaner are blue collar, the check-in desk at the hotel is white collar (mostly) and I'm not sure where the intermediate roles belong. I used blue and white because they're widely used descriptive phrases but neither one is a good fit for a waiter/waitress, cashier at a store, etc.
I think B2B and B2G are generally similar enough to go in the same category... although military is its own thing in terms of "blue collar."
NGOs still tend to have the same job types, just different "outputs." Is there something else that would differentiate them?
There are a lot of roles that I think are mostly safe from AI, but I think that'll be something that plays out over time.
Agreed piecework is a sort of similar to commission, but your pay doesn't depend on revenue or profit - purely on your productivity.
I didn't mean B2B vs B2G, I meant being employed by the Govt vs by a company (for-profit or not). The difference for me in terms of careers is:
* private enterprise: if the business is successful, your financial reward generally depends on the impact you have (in many roles the sky is the limit) and how easy you would be to replace; if the business is unsuccessful, you're out of work. Very much part of the capitalist system.
* government: generally more stability and security, but less opportunity to demonstrate growth and much tighter limits on how much you can end up earning. Subject to the whims of government budgets etc which have effectively no relationship to how well you or your team are doing at their job.
* not-for-profit: often not well paid at all, but much more rewarding in other dimensions like feeling that your work is making a real difference in the world.
The other thing I've tried to emphasise to my teenager is that there are so many roles that aren't "obvious" kind of "Richard Scarry What Do People Do All Day" type jobs, things you'd never think of as a child unless you knew someone in the relevant type of role. To name just a few completely random examples based on people I know: police intelligence analyst, care quality inspector of hospitals, TV broadcast controller....
OK, that gives me some idea. Can you direct me to the thread where he got the bullet? I’m just surprised because I had a few very pleasant exchanges with him
Asking for "youtube channels" with no further specification is like asking how a watch works.
Based on your mention of Ordinary Things and the general nerdy atmosphere here, my usual recommends are Veritasium, SmarterEveryDay, and PracticalEngineeringChannel.
Peter and Scott - Congratulations on winning the bet. Incidentally, when I heard about it, I asked to participate on the zoonosis side, as this was clearly not a bet that is well correlated with truth.
edit: If anyone comes upon this thread later, don't bother reading it or trying to engage with Saar. He's a classic internet crank. Attempting to argue is futile, as he'll demand you answer his questions without answering yours, impose conditions on continuing when things don't go his way, assume that everyone thinks and behaves like he does, engage in wild inconsistency, display massively inflated self-importance, insist that his is the only right way to do anything, assume the only reason anyone ever disagrees with him is because they're fools who make basic mistakes, ignore most of your arguments, general rude comments and obvious hypocrisy, write 20 wrong comments while asserting that you are wrong, not provide legitimate sources, and in this case, use AI to respond to AI summaries of what you say, rather than reading your comments and responding himself. Just all around clown behavior.
If you can point to any specific mistake in the post, please share it.
Most importantly, we point to one mistake that, once corrected, flips most analyses (including Scott's) from zoonosis to lab-leak: It is highly irrational and bad probabilistic inference to claim >99.9% confidence without very high-quality large-sample statistics, and we don't have anything close to that on the Huanan seafood market.
Not only that, nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location. Claiming >99.9% confidence on something that leading scientists are divided on requires a very good reason. None was presented.
> nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location.
I asked ChatGPT and it said this percentage wasn’t accurate. Your quote conflates “zoonosis” with “the first infection definitely occurred inside Huanan market”, obviously if the zoonotic scientists believe if it didn’t happen there it happened upstream of the market, ie a supplier was infected.
This advances the lab leak theory not at all. I’m assigning zero, or very very low, probability to some chap infecting himself with the virus, not getting sick himself, and not infecting his colleagues or anyone close by; nor his family or neighbourhood, and then visiting the market. There he does infect someone else and the market acts as an amplifier (which would be true of any mall or market where he infected someone, most likely an employee).
We know from genomic evidence that the pandemic MRCA is the same as the market MRCA. That means, the market ancestor is at the root of the pandemic. Even if early cases elsewhere were missed, you have to argue that the spillovers elsewhere died out without leaving any traces via other lineages that spread worldwide and miraculously only the market ones survived. The Bayes factor for this should heavily favor natural origins.
Under the leading lab-leak hypothesis lineage A was spreading in Wuhan, developed to lineage B within a few weeks, infected someone in the market and got amplified. No miraculous disappearance of earlier strains is claimed.
It is telling that Judge Will described this scenario exactly as understood from your Rootclaim debate and found that it had no support as ZO presented good arguments for lack of ascertainment bias while you had none.
Why no other lineages in the meantime due to normal mutational processes that spread elsewhere? You do know that there are those arguing against the double spillover because of the mutational clock? Why does lineage A stay intact until it reaches the market area? Why are lab leakers arguing that the virus was pre-adapted for human spread and also saying that the pandemic didn't take off for weeks until it reached the market? Why does healthcare worker serology tests (which had a high ascertainment rate due to cash rewards from Bytedance) spread outwards from the market? You have no evidence via serology, mortality, MRCA analysis, healthcare workers tests, hospital records, case records. Simply postulating something with no evidence is only motivated reasoning and therefore gets a very low Bayes factor.
You did not understand my point on 99.9%. Please read the relevant parts of the post again.
Your hypothesis is a strawman. The leading scenario is that they infected several people who infected others etc, and then one of them reached the market weeks later, which is evident by all market cases being of the later B lineage. More here: https://x.com/Rootclaim/status/1753353163178913815
Having said that, your strawman is still enough to flip the analysis to lab leak. Read the "CDC Move" section in the post
> If you can point to any specific mistake in the post, please share it.
For example:
> There is basically no evidence for zoonosis other than the results of repeating the following process:
This is incorrect. During the debate, Peter presented many maps of cases and related data (such as deaths) that were simply visualizations of raw data. No model is necessary to, for example, look at such a map and see that cases (both market linked and non) centered on HSM. Or look at the first handful of cases and see how they are heavily concentrated in HSM.
The post as a whole is in fact riddled with major errors, but first:
> Not only that, nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location.
Source? I cannot find any survey like this on Google. Just ones that say, over and over, that the large majority of professional virologists and epidemiologists believe zoonosis is more likely.
> Claiming >99.9% confidence on something that leading scientists are divided on
...leading scientists are *actively opposed* to lab leak o.0
Those who don't believe HSM is the spillover location still think it's the main amplification site, so the maps are not evidence for either hypothesis. If China would publish full contact tracing for HSM like they did for Xinfadi, then it would be strong evidence to one of the hypotheses. It's fair to assume they very likely did do that but chose not to share the result - and that IS evidence.
The twitter thread I quoted provide evidence for HSM not popular as a spillover site. In any case, it doesn't have to be a majority. Claiming >99.9% confidence on anything that top scientists are contested about without a very good reason is irrational.
I stand by the claim that Zoonosis has exactly zero good evidence.
>leading scientists are *actively opposed* to lab leak
They're not. They lean zoonosis but don't rule out lab leak. Anyway, a) we don't claim anything close to >99.9%, b) we have a very good reason.
> Those who don't believe HSM is the spillover location still think it's the main amplification site, so the maps are not evidence for either hypothesis.
More errors.
First, you seem to be the among the few lab leakers who believe this. More importantly, the fact that you can *theorize* that it is the main amplification site doesn't make "zoonotic origin related to the market" and "lab leak but all early cases are strongly associated with the market" equally likely.
> It's fair to assume
An assumption is not evidence.
> The twitter thread I quoted provide evidence for HSM not popular as a spillover site.
Can you provide a link to the specific source for the claim that " nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location."? I don't have an X account and the format of "20 one sentence tweets" is annoying to read, and I'm not going to do the work of justifying a claim that *you* made.
> I stand by the claim that Zoonosis has exactly zero good evidence.
You can stand by it, but it's still wrong. Previous lab leaks overwhelmingly had cases associated with the lab (that's often *how* they were identified as lab leaks in the first place). This is possibly the only piece of evidence where the relevant rates can reliably be estimated with past data, rather than being based on guesswork.
> They lean zoonosis but don't rule out lab leak.
Come on Saar, you understand the difference between this and what I wrote. You can't rule out zoonosis either (you say 6%, roughly) but also say that "supporting zoonosis is irrational." Scientists don't say lab is entirely impossible but pretty clearly think it's much less likely.
> Anyway, a) we don't claim anything close to >99.9%, b) we have a very good reason.
"We have good reason, the other side does not" is simply circular.
You're avoiding the main issue and scattering to minor ones. Before answering those, I want to first close this: Do you agree that if prominent scientists are divided on an issue, then someone who has no expertise claiming >99.9% confidence is being irrational?
Given that you don't want to put the time to read the evidence, I'll give just two that are enough to convince of "scientists are divided", rather than my stronger claim of "most scientists".
1. WHO report:
"Many of the early cases were associated with the Huanan market, but a similar number of cases were associated with other markets and some were not associated with any markets. Transmission within the wider community in December could account for cases not associated with the Huanan market which, together with the presence of early cases not associated with that market, could suggest that the Huanan market was not the original source of the outbreak... No firm conclusion therefore about the role of the Huanan market in the origin of the outbreak, or how the infection was introduced into the market, can currently be drawn."
I'm charitable towards the weak form of lab leak (not "engineered on purpose" or even "people were doing engineering of some kind", but "someone had it because they were passively studying it in good faith and a mistake happened") but I instantly disregard any argument that "believing side X in a controversy is irrational". If your arguments are good you don't have to try to emotionally stack the deck.
It's especially funny in this case because Saar said described a poll of superforecasters as "clearly not a bet that is well correlated with truth" when the superforecasters have a demonstrated, externally-validated track record of reasoning under uncertainty, while Rootclaim... does not.
> I’ll admit that my bet was only partly on the merits of the COVID origins argument, and at least as much on the thesis that new and earth-shakingly decisive evidence on controversial events rarely comes out in the specific interval of 4-6 years after the event happened
What you said:
> clearly not a bet that is well correlated with truth
What specific point is that paper supposed to demonstrate? It seems like it kicked off a whole set of very technical discussions, and I can't tell what the relevance is here.
Response 1 (furthering the argument in the paper you referenced, specifically extending it):
Response 4 (general sweep of argument, proposing an alternative argument that supports the same general conclusions as the original paper by arguing about the coherence of point probability regions in confidence distributions):
It seems pretty complicated. I'm trying to figure out the actual argument because it seems like a philosophically interesting place where frequentists and bayesians can meaningfully disagree
The most important part is the distinction between epistemic and aleatory uncertainty, and how treating them the same way can result in huge errors:
> False confidence is the inevitable result of treating epistemic uncertainty as though it were aleatory variability.
> The final clue indicating the inappropriateness of probability dilution is the fact that, for a fixed S/R ratio, there is a maximum computable epistemic probability of collision. Whether or not the two satellites are on a collision course, no matter what the data indicate, the analyst will have a minimum confidence that the two satellites will not collide. That minimum confidence is determined purely by the data quality.
> Since the uncertainty in the estimates of the trajectories reflects the limits of data quality, probability dilution seems to imply that lowering data quality makes satellites safer. That implication is counterintuitive in the extreme [4–8]. As a rule, lowering the data quality makes any engineering system less safe, and to claim that ignorance somehow reduces collision risk seems foolish on its face.
> Interestingly, were the uncertainty in the satellite trajectories aleatory, rather than epistemic, probability dilution would actually make sense. For example, suppose two satellites were known, with certainty, to be on a collision course, and the satellite operator could only impart an impulse of random magnitude in a random direction, say, via a poorly controlled manoeuvring thruster on a tumbling satellite. If the mean of this distribution were the null vector, due to a lack of control over thrust direction, then the higher the variance of the added impulse, the bigger the resulting perturbation and hence the smaller the resulting probability of collision.
Moreover, the resulting errors are not bounded:
> As relative uncertainty grows, the level of false confidence seen in satellite conjunction analysis grows arbitrarily severe. Next, we show that, when considered over the full range of measurable propositions that the analyst might assess, most epistemic probability distributions used in practice suffer from arbitrarily severe false confidence.
In other words, there is a big difference between the following 2 statements:
1. Out of 10,000 cases with similar data, satellites collide about 0.1% of the time.
2. We don't have very good data, so it would be overconfident to say that the satellites are more than 0.1% likely to collide (since if you pick 2 satellites at random, they are unlikely to collide soon).
In the latter case, it might actually be the case that the satellites are nearly guaranteed to collide.
Rootclaim's analysis very explicitly makes this error. For example, looking at their most recently analysis (https://cards.rootclaim.com/analysis/65892f07d97bdb00199d4366) the single biggest Bayes factor comes from "12 nucleotides clean insertion," which they say is 10% likely under lab leak and 0.5% under zoonosis. How do they justify the former probability?
> In the past, FCSs have been added by substitution rather than insertion, but it is not difficult to do it by insertion. We cannot be sure of the exact considerations of the lab researchers who added the FCS , such as investigating the role of proline. Therefore, we assign it a 10% probability.
This is precisely the sort of epistemic uncertainty that, according to the FCT, results in errors of arbitrary magnitude! They do not have a single scrap of data to indicate that, 10 times out of 100, in similar situations, lab researches would use a 12 NT insert. They just say that it would be overconfident to assert the value is smaller than 10%. The true value, for all they know, could be 1%, or 0.1%, etc. In fact, parallel to the quoted paragraph above, if one *did* have such data, and decided to throw it away and use Rootclaim's method, the reduced data quality would result in *more* confidence in lab leak (since the resulting BF in favor would be higher).
Yeah, but it seems like you get the same false confidence errors using frequentist frameworks as well (see response 2):
> However, the message presented by Cunen et al (that replacing a Bayesian posterior distribution with a CD is all it takes to overcome false confidence) is potentially misleading. Our false confidence theorem applies to all epistemic probability distributions, including CDs; so, contrary to the authors’ claim, their proposed CD is at risk of false confidence too.
And the authors of the paper you cite eventually conclude that the satellite problem is not caused by the usage of epistemic uncertainty but instead by the specific problem formulation (the false confidence issue arises from considering the complements of a convex hypothesis). It's not a generic fact about epistemic probabilities that false confidence goes up with bad data. See response 3 by the authors of your cited paper:
> The false confidence theorem establishes that, for ANY data-driven, precise-probabilistic method for uncertainty quantification, there exists (non-trivial) false hypotheses to which the method tends to assign high confidence. [...] Roughly, complements of convex hypotheses are afflicted by false confidence.
To clarify, I am in no way saying that the rootclaim probability is accurate (it seems like they just made it up?). I haven't looked into it too thoroughly, but I lean zoonotic. I'm just not sure how that paper you posted maps given that it's about a very specific situation (the "false confidence" comes from the geometry of the hypothesis). AFAICT, the discussion seems to have concluded in:
"Any precise additive distribution over a continuous parameter space, Bayesian or Frequentist, systematically over-supports false hypotheses that are complements of convex sets (or in general, nonlinear sets)"
Which doesn't seem to apply to the nucleotide insertion thing (which is too simple to have the kind of geometry that systematically leads to these false confidences).
This whole thing is very interesting. The bayesian relevance seems to be something like: If you don't have a real prior on a parameter's value and you default to a "flat prior", that "flatness" is only true relative for the coordinates you originally estimated in. If you use those parameters to generate other coordinates, that "flatness" causes an opinionated non-flat probability distribution for the set you care about. Because Bayesian probabilities are additive, probability of not-P is 1 - probability(P), so these distortions create "phantom evidence" that artificially slant the analysis one way.
For the satellite thing specifically, you definite "collision" as "the relative distance between satellites is inside this small disc" and "no collision" is the whole curved space outside that disc. So your probability estimations are on "how far away in the X direction are the satellites" and "how far in the Y direction are the satellites" (flat) but your hypothesis defines a collision using the euclidean distance between them, which is nonlinear "sqrt(Δx² + Δy²)" which introduces distorting phantom evidence (because altering the probability of collision implies altering the probability of safety, there is no explicit third option to absorb the uncertainty)
I don't think I said that frequentist frameworks avoid the problem, nor was that the point. The point of my example was the existence of quality data, not how the parameter "frequency of lab using insertion" is interpreted. As you quote,
> The false confidence theorem [2, 22] says that, for any precise IM... there exists false hypotheses H to which the posterior tends to assign a (relative to α) large probability/confidence
i.e. false confidence. Here, a "precise" inferential method is one with countable additivity, which is what implies P(not A) = 1- P(A). I think this is precisely the issue with "overconfidence-based" probabilities, like those utilized by Rootclaim (and Michael Weissmann): if P(A) is bounded from above because of worries about overconfidence, P(not A) is similarly bounded from below.
This is basically what you are saying here:
> altering the probability of collision implies altering the probability of safety, there is no explicit third option to absorb the uncertainty
A parameter like P(lab uses 12 NT insertion), with no data to go on, should be the subject of high uncertainty. Rootclaim *says* they account for this uncertainty, but their methods are all based on "precise" IMs, not the 3rd option of "explicit uncertainty" that you mention, and so do not resolve FC.
> Which doesn't seem to apply to the nucleotide insertion thing (which is too simple to have the kind of geometry that systematically leads to these false confidences).
I don't think you need complicated geometry to have convexity. In fact classic examples (one would even argue, the motivation for the entire notion of convexity) include simple one-variable functions like f(x) = x^2. If I'm understanding the math right, then it very easily could apply to an "overconfidence adjustment" that shrinks the parameter (probability of lab using this insertion) toward 50%.
The convexity thing provides further information about *which* hypotheses are effected by false confidence, but the implication is co-convex => FC, which doesn't do anything to say that a given hypothesis is *not* afflicted by FC. At the very least, if Rootclaim wants to use this methodology, they should show that it meets the "linear hypothesis" standard or otherwise avoids false confidence. And, just intuitively, without the math, it seems obvious to say that your strongest piece of evidence shouldn't be based on a guess with no data.
Agree with everything you just said. These sentence are what made me believe you were using the paper to generally criticize the formal emphasis of epistemic vs aleatoric probabilities.
> The most important part is the distinction between epistemic and aleatory uncertainty, and how treating them the same way can result in huge error
and
> This is precisely the sort of epistemic uncertainty that, according to the FCT, results in errors of arbitrary magnitude!
Whereas the actual target of criticism seems to be unreflective usage of precise additive inferential methods (for which certain false hypotheses assigned relatively high probabilities are guaranteed to exist).
I totally agree with you that Rootclaim is not doing a proper analysis. Their issues seem more on the order of "literally just guessing important things" which would disqualify it before you even get to the level of hypothesis convexity (which I haven't looked into specifically here, but it's a subtle issue that it's unlikely they properly account for).
Thank you for linking that study! It's been very interesting to think about, and I would have never stumbled upon it myself.
The Aug. 8 New York Times has an article on A.I. in China. I’ll summarize since it’s behind a paywall.
1. Chinese A.I. companies do marketing in Africa; American companies mostly don’t. Chinese companies will reach out to African companies, offer them free computing resources and help setting up their A.I. systems.
2. African countries are not particularly wealthy, and don’t want to pay for frontier models unless they need them. At one point, Meta’s Llama filled the need for an inexpensive model, but Llama 4 performed less well than many expected, and Meta decided to switch to closed weight models rather than create another version of Llama, leaving an opening for Chinese models.
3. There are tasks where Chinese models aren’t good enough yet. For example, Chanderia Industries an OpenAI model to read product orders from retailers and enter the orders into its system for fulfillment. (The retailers send in order via E-mail or paper mail.) Bytedance convinced Chanderia’s provider, Sanifu, to test their model, but it made too many errors.
Another company described Claude Code as expensive but worth it.
4. It appears that fine tuning open weight models for particular use cases is fairly common in Africa, which probably reduces the need for cutting edge models.
5. Shutdowns. In 2024, Deepseek shut down its servers before the Chinese college enterance exams to prevent students from using the service to cheat, which disrupted some African businesses. More recently, Anthropic blocked access to Fable at the request of the Trump Administration. Open weight models win here because if your provider cuts you off you can continue to use the same model with a different provider.
6. Bias. Sunflower uses a tuned version of Qwen 3, which Sunbird chose over Llama and Gemma because it was better at handling African languages. (I wouldn’t be surprised if Alibaba made a greater effort than American companies to ensure that African languages were well represented in their training data.) However, China Media Project concluded that Sunflower was biased towards China; you can read the details here: https://chinamediaproject.org/2025/12/17/the-chinese-core-of-ugandas-chatgpt/
As a result, Sunbird is using Gemma (open weight model from Google) for its next project.
My thoughts: The open weight models seem to be quite beneficial to third world countries. Scott’s recent article titled “Open Questions On Open Weights” suggests that for now there’s not too much reason to ban open weight models. I do worry a bit that this is another example of where China has a national strategy and the United States, at least under the current Administration, has none.
Disappointing to hear Llama won't have a follow up. My presumption was that data security, regulations, insurance requirements and vendor contracts would all push towards locally-hosted tunable open weights models in many industries. It seems intuitive to me that this will need to exist in some form because the existing legal landscape demands it, but America isn't going to adopt Chinese models and in some cases probably isn't legally allowed to either. What's the business solution to this going to be?
I'm a lawyer presently in financial services with a lot of bank clients and borrower info, at present I am continuing to advise management they cannot use anything requiring remote API. I would need affirmative statements from banking regulators, the insurance industry, the bar association, and all of our clients before I would consider allowing any usage of remotely-hosted AI systems. I think it would likely breach vendor agreements we've already signed with the bank clients, by terms the bank is required to mandate of their contract partners by the bank's own insurance companies and their regulators.
'scuse an ignorant question: Why is remote API in particular a deal breaker? I could see AI _unreliability_ being a deal breaker, but isn't some conventional financial software like ADP's tax and payroll services hosted remotely?
ADP is handling YOUR company's employees data, under a contract you have with them. If there's a data breach at ADP, you are the client and their incident response team would handle the fallout and their insurer would cover it and your own insurer is fine with this.
But if my business requires me to store and use the information of tens of thousands of bank customers with whom I have no direct relationship, there's many more things in play. The bank's regulators and insurers will require it to handle consumer data with certain precautions in order to be covered, and in event of a data breach their standard e&o policy may not cover it so they probably have a rider for that too. And part of the requirement of those regulators and insurers is that any vendor agreement, the contracts made with other entities (like me) that handle the bank's data, made between the bank and other entities must require THOSE entities to also have $X in cyber security coverage and a laundry list of security protocols.
If we started sending consumer info to remotely hosted LLMs and there was a breach, we might theoretically have a cause of action against Open AI or whomever but the immediate problem would be that our e&o and cyber security policies won't cover us, the 2nd order problem is that we'd have breached the contract with our bank client who is now in a crapton of trouble, we're now contractually obligated to spend millions of dollars in incident response out of pocket and have probably lost an important client and are unlikely to survive all these dominoes falling as a small to mid-size company.
This is why I cannot use remote API right now. At some point if locally hosted LLMs dry up and everyone wants to use remotely hosted AI then what will happen is the institutional players will see the value in it as necessary and will converge upon some set of protocols. Everyone in this process will attempt to extract a bit of money or market position out of it: insurance gets some new type of policy you'll have to buy, some company will wind up with the deal to give ratings to the security of various AI companies to provide to the insurance companies, banking regulators will be persuaded by well-positioned providers to permit certain types of services and not others, since the business also involves litigation the malpractice insurers will end up saying you have to use the Thomson-Reuters AI which won't meet the other requirements so some facilitator springs up to bridge that gap until OpenAI buys it, etc etc. All these dominoes have to fall before legacy institutions can do this, so yeah just run a dang Llama in the closet.
What's the current state of what major LLM vendors offer their enterprise customers in terms of indemnification and insurance coverage and whatnot? You would think that they'd be trying aggressively to remove whatever barriers are in the way of enterprise sales, and at least some of them are or have close partnerships with major cloud vendors that have been dealing with this sort of thing for a long time. Or is the problem that your own insurer will go "that agreement isn't worth the paper it's printed on"?
If you've signed a contract saying you won't share certain data with people who haven't signed the contract, you can't then send that data to machines owned and run by people who haven't signed the contract.
I don't think AVs are an example of what Cjw is talking about. Those models are highly bespoke and unlikely to be released with open weights, nor would it be particularly useful if they were.
I wonder how much the Chinese government determines the AI strategy of their labs. On the one hand, all the Chinese labs are pursuing a similar strategy, so it could be government-directed. But that strategy could also be their own decision: if your models are behind the American ones, making them open-weight is an advantage to get more customers. Very interested to hear if anyone knows more about this.
In a normal environment this diversity is good and will have the best strategy prevail. With AI however I do hope the administration intervenes soon as the wrong strategy can lead to very bad outcomes. (More referring to strategies on training, scaling & alignment, Scott made a good case for why you don't need to spend political capital on banning open weights now.)
No. Consciousness is an undefined term, or at least there's no common agreement on exactly what it means. I know people who won't accept that a dog is conscious. The decisions that matter should be based on observables, where there is at least the possibility of common agreement.
Utilitarians should just pivot to anthropocentrism. It solves the overwhelming majority of gotcha-questions that get thrown at them in discourse. It eliminates a whole bunch of pontificating on things they have very little insight towards and huge epistemic limitations about such as the preferences of bacteria or computers. And after all what is the point of ethics? Only humans can engage in ethics, it is a human doctrine, it's only natural that its end should be anthropocentric.
In this model, animal suffering is not irrelevant, but it's only relevant obliquely by way of it causing distress to humans. Dog suffering causes humans distress, but nobody loses sleep about bacteria. Simple, we're done, that's all we needed.
I'm not sure I understand the point you're making, but in any case, I'm not arguing that an anthropocentric ethics *can't* end up caring about animal welfare (I'm somewhat agnostic on that), I'm arguing that this doesn't address or engage with any of the arguments in favor of a non-anthropocentric ethics in and of itself.
Maybe, I think a lot of people seem compelled to have a single Grand Unified Theory of Ethics that they have to shoehorn everything into that they intuitively see as important to care about. But you do not have to shoehorn animal welfare into utilitarianism, it just causes problems for both.
There's no real reason we could not say both:
1. utilitarian ethics are a system of evaluating good for large groups of humans and making decisions, which was devised by humans and does not function correctly outside of application to humans.
2. certain cruelties to animals may be undesirable because they cause harm to humans or for other solid independent reasons such as violating an implied obligation of stewardship or because cruelty is not virtuous.
Being a pluralist on this seems much easier than to continue to keep shoving the square peg in the round hole, in fact it's worse than that, the utilitarian animal ethics people can't even tell what shape the peg is and have an insurmountable epistemic limitation on their ability to discern the shape. There is no reason the system to evaluate our behavior among ourselves would be the same one that governs our behavior towards animals.
Newtonian physics works on Earth and the moon and the solar system, all the things it was developed to work for and observed that went into its development. It breaks down when you try to apply it to more exotic phenomenon and starts giving you absurd predictions.
While Bentham may well have wanted it to work for animals, the reality is that while it's readily adapted to human populations it has never integrated very well with utilitarianism for a variety of reasons, and attempts to make it work keep spitting out absurd results about universes filled with shrimp or machines requiring all sorts of workarounds to make any sense of them. It's just not meant for this context.
That isn't to say you then have to adopt contractarianism either. What I'm suggesting is a radical ethical pluralism that does not try to account for the varying values that humans have by forcing them all into one grand unified system. People in the world do not form preferences that way, they don't behave that way, their values aren't explicable that way, so why are we insisting that utilitarianism or any other particular system of ethics has to be universally applicable to all situations? My response to "oh well if you are willing to apply it just to humans then why not just to your own city or country?" would be that you may have to argue people out of that line, but that in fact people did argue that line very successfully! That elements of human universalism in ethics have worked fairly well is a point in its favor. That you have continually hit brick wall after brick wall and faced endless arguments ad absurdum in your attempts to loop in non-humans is a sign that you are probably better served with a pluralist model that accounts for animal suffering with a different ethical theory. I see no reason that some human values are best understood and directed by tabulating utils, and some by deontology, and some by virtue ethics.
The thing is, you're just asserting that that there's no good reason to treat nonhuman animals as moral patients. You're not engaging with any of the reasons that pro-animal-welfare ethicists have for doing so.
Plenty of anthropocentric reasons to be concerned about animal welfare.
As a human, I have a preference against witnessing or knowing I've contributed to the needless suffering of animals complex enough to have feelings worth considering (largely mammals, but I don't eat seafood so have never had to confront the question of lobsters, &). I think a system capable of running on factory farming is not a system which is going to treat the humans within it with much more kindness. I think a person who whips outs to death eight hours per day, or whatever they do, is more likely to cultivate the worst parts of their psyche than someone who raises cows on a pasture.
The anthropocentrism also acts as a shield against other insane beliefs, like the notion that LLMs are in any way sentient or in danger of becoming such merely because they meet the extremely low bar to pass as effective at doing what they were programmed to do (mimic human text output).
Consciousness, suffering and so on in most of these conversations are proxies for the real question: "should I care about this action being performed on this entity?"
Different humans care about different things, and have different intuitions; conversations here are an attempt to construct some common ground for people with different intuitions to come to some kind of concensus.
Anthropocentrism meaning we should only care about humans - fine, that's a coherent answer; some - many - most, perhaps - might disagree, but at least what it means, where the boundaries are and how it differs from "I'm bored of talking about this now, let's just agree we each know what we should care about when we see it" is pretty clear.
Anthropocentrism meaning we should care about humans and also nonhuman things humans care about fails to answer or clarify anything at all; since the whole argument started because humans can't agree what nonhuman things to care about, we're left in exactly the same place we were before.
That still leaves my objections that cruelty is corrupting for the people who have to carry it out, and that a system capable of cruelty to mammals is more likely to be capable of cruelty to humans.
That’s basically what I think. Animals don’t cooperate(dogs are actually a very notable exception) so why are we spending so much time talking about obligations towards things that are incapable of understanding obligation?
But if you take that view, then it changes how you view animal behavior. People in the past didn’t actually care about animals, so a dog suffering didn’t matter to them. It’s only because we think people should care that we actually do.
There is as much reason to consider reciprocity essential to morality as there is to consider it irrelevant.
People in the past also did not care about humans.
They cared about members of their own tribe.
The fact that so many people today consider humanity their tribe (at least to some extent) and have a hard time grasping that this is not the natural default, is a testament to the success of humanism.
But there is no good reason not to expand that circle further to include nonhuman creatures as well.
They did only care about their tribe, making it impossible to work with other people and they got trounced by those who learned to form larger coalitions. We’re all born with an innate morality that allows us to cooperate. Thats what differentiates us from animals that don’t have this idea of non kin altruism.
No one treats infants as full members of the social contract. Otherwise we would constantly be jailing them for assault. They’re prospective members. As for coma patients or whatever, it’s much easier to just default everyone to moral status then try to measure which people have enough agency. On the other hand, it is literally impossible for any shrimp to ever be part of the social contract because they don’t possess the capability.
There's a kind of gray zone for movements toward something desirable or away from something undersirable: tropisms. Should those be considered preferences?
Rocks can't make choices. Bacteria have a preference for nutrients because they go towards it and a dispreference towards toxins because they go away from them. I don't understand why from a preference utilitarian perspective, it doesn't count.
Umnh... are cpu chips rocks? If not, why not? If they are, why do you say they can't make choices.
Actually, that argument can be framed for most crystal structures, e.g. DNA.
If you asserted that "we don't notice the choices that rocks make", I'd agree with you most of the time, and acknowledge that we don't tend to call them choices when they happen.
OTOH, perhaps our disagreement is in precisely what is considered a choice. I tend to consider any "if branch" a choice point, and also any probabilistic divergence in action. Sometimes the "choice" is nearly certain, but if it's not certain, then it's a choice. I don't consider "consciousness" or "intent" to be part of what makes something a choice. So I tend to see "choice" nearly everywhere I look. I can see a reasoned argument that this makes the term nearly useless, but I haven't heard a better precise definition. (And even this one's a bit fuzzy.)
Preference utilitarians differ on how they understand preferences, but almost all will agree that you can't immediately infer something has a preference for nutrients simply because it is disposed to go towards them.
Why does a shrimps behavior count as a preference but bacteria’s doesn’t? If we put humanity on one side where we can state our goals, why we want to achieve them and how we plan on getting them and bacteria on the other side where it simply moves towards and away from things, then shrimp look much more like the latter.
I don't think it's totally obvious that shrimp have preferences in the relevant sense either. (Though it's not the ability to state them that matters on any plausible view).
Ok that’s fair. I was just talking to someone who did hold that view and hedonist and preference utilitarians overlap a lot. It does make me wonder what counts for them as real preferences among different animals, since it’s less clear than the hedonists.
Well then I don't know what to tell you either, because I don't see bacterial locomotion as being any different in principle than rock locomotion due to gravity. Rocks follow the gravitational gradient; bacteria follow the chemical gradient. Bacteria don't make choices in any meaningful way any more than rocks do.
If chemical gradient is just as meaningful as gravitational gradient then nothing we do matters since we’re all chemical reactions.
I’m not sure what “in principle” is doing here. We both have the same principle that bacteria doesn’t matter. But clearly, taking action is a different thing than being acted on. One of them is something internal that you do. The other is something external that happens. I don’t moralize a dogs behavior but I would differentiate between a dog jumping on my things and someone throwing the dog on them.
MacGregor went towards the sheep because of physics. Scott wrote about MacGregor because of physics. I'm writing this because of physics. Well, not because, exactly. . . .
What about a rock in orbit around the sun, with one side slightly darker than the other? Does it have a preference to rotate in a particular way, dictated by its albedo?
I believe Bentham (and likely his Bulldog too) would say that what matters is the ability to *feel suffering*, not merely to display certain objective behavioral traits (even if we have to make inferences on suffering based, in part, on objective behavioral traits).
It’s very irritating that I’m asking a question about preference utilitarianism, which is not even my philosophical belief, and it’s used to mock me as an idiot.
Uhm... my best guess at a serious answer is that the choice of "whose preferences matter and how much" is arbitrary.
According to Claude, Bentham said it's the capacity of suffering that matters, so animals are included. Bacteria probably wouldn't count because they don't have a nervous system -- mere response to stimuli is not enough, there must be "someone home" to feel the pain or frustration. The gray area from the perspective of modern philosophers are insects and plants.
John Gruber issues his first ever blog post retraction. It's great to see people willing to own up to mistakes, and even more impressive in this case since it wasn't really his fault, he just got deceived by someone else.
Whoever made that comment needs to clarify their terms. A "conservative" can be a person who respects tradition in whatever form it takes (landscape, environment, social mores, behaviour, whatever) whilst also being open to measured change, particularly if it contributes to the correction of established errors. A "conservative", particularly in the context of US politics today may be all of the above, but will very likely also be a racist, pro-big business, Republican voting, xenophobic Trump worshipper.
I would hazard a guess that Hawaii and Rhode Island are conservative in the first instance.
The usage of "liberal" vs "conservative" as synonyms for "left wing" and "right wing" drives me absolutely crazy, because it has nothing to do with the literal meanings of the words.
Trump isn't exactly "conserving" anything. His whole thesis is that the US was failing and needs radical change!
Conservative? It's going to be a struggle to find a mayor who doesn't support rent control. I regularly see outright communist posters at bus stops, "Abolist the police/military" on stop signs. Brown had all those encampments, and if there are straight women in their 30s, neither I nor anyone I know has any idea where they're hiding.
I guess it could be like New York State, and I'm just seeing the one side because I'm in Providence. I do keep meaning to visit a shooting range and get back into practice, but the good one requires a membership or a +1 and everyone I know well enough to ask is convinced they'd be used for target practice if they set foot on a range.
I haven't been, so huge grain of salt, but I believe there is a focus on "conserving" the way things are on the islands. A preference for continued status quo in terms of how land is used and owned, versus new development dollars building more resorts, bringing in more visitors/buyers from the mainland.
Kind of like the "conservative" streak you'll see in certain otherwise very-liberal neighborhoods when it comes to development. Openness to LGBT people and minorities, support for their rights and protections? Yes - change is good and part of progress. Support for a new hotel around the corner? Hell no, change is terrible! Keep this neighborhood the way it always has been!
I’ve never been there myself, but something I’ve heard is that Hawaii has a rather tough policy towards its homeless population. The authorities are quick to remove encampments, and homeless individuals with ties to the mainland are strongly encouraged to return and given free air tickets if necessary.
I can’t claim to know in what way prosa123 had in mind, but do note that many Hawaiian politicians are committed, or at least pay a lot of lip-service, to Hawaiian ethno-nationalism. Ethno-nationalism is often regarded as a very conservative ideology.
Not necessarily—when nationalism is directed against imperial rule, it has often been revolutionary or left-wing. The modern international order is largely built around nation-states, and the UN Charter recognises national self-determination, though it doesn’t specifically endorse ethno nationalism, basically a nation is just that.
In the case of Hawaii I assume the nationalism is the indigenous people, which makes it a left nationalism.
Americans are blind to their own country being an empire, Britain had consolidated its rule over Hawaii, and it was part of the empire (or kingdom), would you condemn independence movements as conservative.
You’d be wrong, these aren’t really subjective opinions. As I said “when nationalism is directed against imperial rule, it has often been revolutionary or left wing”. It’s also not conservative, by any sane use of the word since it’s trying to overthrow the existing order.
By that metric, Israel is another theocracy that is ... left wing.
Also throwing Treasury Secretary Bennett off a building (purely for his gayness) is now left wing. The world looks very strange if we credit your ideas -- forced sex changes are also leftwing (to eliminate gayness).
I came across this article on AI and bioweapons which I think is worth a read for anyone not sure whether to worry that AI-assisted bioweapons will kill us all. It's called "Why AI-assisted bioweapons won't kill us all"
While maybe the conclusion isn't robust against whatever arbitrarily godlike AIs one might like to imagine, it does set out a lot of the factors that do make AI-assisted bioweapons thankfully a lot trickier than you might imagine.
I’m not sure how reassuring a lot of this is as it seems to be trying to look retrospectively at what has made bioterror less dangerous in the past and then pain this onto the looming AI future.
The section about lack of automated lab skills is probably the most reassuring in the near term. As of right now it probably still is beyond the ability of a very small group of actors to do this because they lack the requisite physical skill set. However automated lab research would be hugely impactful in a positive way for drug discovery and basic science so it seems unlikely this barrier will exist indefinitely. Maybe we can be confident that it won’t be here in the next five years, but I would assume we get there eventually.
The other critiques really just hinge on the “making bio weapons is hard” argument. They don’t make any sense if we are operating in the world of “making bio weapons is trivially easy”. I don’t care if my sars pandemic peters out, I’m already well into releasing small pox, Ebola, pneumonic plague, 1918 flu, and many more.
I’m also not really convinced that creating super pathogens is all that mature of a field where we can just sit back and say that just because a thing hasn’t been done before, it must be very difficult to do. For one, I doubt that anyone privy to information to the contrary would be allowed to post it publicly. More importantly, I’m not sure how high a priority creating super pathogens is for nation states, who are currently the only entities really equipped to do so. Maybe it just hasn’t been worth the effort to pursue.
There are approaches that many respected biologists think might be omni-lethal. I didn't read his argument, but the particular example I'm thinking of is bacteria which an opposite chiral twist to their proteins. I'm sure there are others.
I'm not convinced that the worriers are correct, but I immediately discount anyone who says "no such thing can happen". (And you don't need an "god-like AI" for this. Only something a little bit in advance of what we've got...and a fancy enough lab.)
"I interviewed dozens of biosecurity experts who had decades at the lab bench"
What was that Arthur C Clarke quote again, about whether or not something is possible?
"When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong."
And that is not just some glib cynical paradox. It's been proven true over and over again!
While I agree with the article (and have been saying similar things for years), I don't think it'd be persuasive to the AI-doom crowd. The typical counterargument is that most (of not all) of the tradeoffs involved in bioweapon design -- transmission vs. lethality, wide deployment vs. cost, team size vs. secrecy, etc. -- are purely a feature of human incompetence. The superintelligent AI would be able to design a virus that is cheap and easy to deploy, maximally infectious, maximally lethal, and to do all that in perfect secrecy, because AIs are just so much smarter than humans, and any arguments to the contrary only serve to demonstrate just how inferior you are as compared to the AI.
To respond to Bugmaster's point, I think it depends on how exactly you present it. I agree that an argument along the lines of "building superviruses is super complicated and humans can't figure out how to do it, therefore AI won't be able to either" is unlikely to be convincing.
However, an argument along the lines of "viruses are logistically complicated, which means it is physically impossible or near-impossible to create a virus that will stably infect the human population until we are all or nearly all dead" might work if supported by sufficient scientific evidence.
I agree with Melvin's conclusion that it is a pretty useful article for the near future, even if it maybe "isn't robust against whatever arbitrarily godlike AIs one might like to imagine".
As I see it, one major problem with the AI-doom crowd is that they want to have it both ways. They say (or heavily imply) that basically all of modern science is little better than a random guess compared to what the AI could accomplish with its intellectual might; therefore, any apparent physical or logistical limitations on any proposed feat of engineering are completely illusory. At the same time, they use modern science to very confidently estimate the probability of ASI arising in the next N years/months/weeks (naturally, the probability always comes out quite high).
I actually finished most of it before we learned about the HF and other hacking incidents: originally held off on publishing because I was waiting for the news to die down a bit; eventually decided to release so I don't get continuously scooped by reality.
Does the world need more AI alignment researchers?
I've had a relatively lengthy career in non-AI software development. I'm considering pursuing alignment research, as it seems very important. This would obviously require a fair bit of retraining, and there's already several people that are bringing decades of experience to bear. Plus, I'm sure their productivity has gone up recently with advances in LLM.
Are research positions saturated or starved? Is the idea likely worth pursuing?
Which alignment ? It's currently two different things, the practical thing that has modest goals of making specific AIs behave somewhat acceptably, and the theoretical thing that aims a a total and general solution.
It seems to be a boom industry at the moment, so if you're looking for work and don't want to be displaced by the friendly helpful assistants in your current job, then go for it.
This does not mean that AI alignment research is going to produce anything extremely useful (given that right now it seems that the AIs are happily lying, cheating, and deceiving just like humans), it just means that it's another way the nerdily-aligned are taking in one another's washing. If it gets you employed, then it's a benefit, whether or not you end up producing papers for conferences which other people attend so they can get data for papers of their own as we all go round the mulberry bush.
Honest question: is there evidence that AI alignment researchers have accomplished anything ever?
Sort of a mean question but I don't follow the field and can't really imagine how they 'could' have accomplished something.
Random papers about what an llm will do to escape a sandbox wouldn't count as accomplishments to me; I mean, have they accomplished anything towards the goal of figuring out how to make better-aligned AI.
Yeah I'm mostly uninterested in that field not because it isn't noble, but I can't see this field bearing fruit. Though I guess AI risk research is already inhabited to try to induct a 0.001% probability event to prevent another 0.0001% probability event.
Have all the aerodynamics researchers on this plane achieved anything? All they do is doom post about how “the stall alarm means we need to push the nose down, increasing our descent rate” they don’t have any ideas on how we can slow our descent by pulling the nose up.
Broadly, we can literally read the minds of various models as they operate, including misalignment, inserted thoughts, shifts in the personality basin, and more, and all thanks to AI alignment researchers.
What they have achieved is getting and giving each other relatively (by the standards of ordinary slobs) good jobs, the same way that EA seems to have collapsed into "network! sign up! get a job with us so you can then reach out to the others to get a job with us!"
(I know that sounds extremely cynical, but I am extremely cynical about how EA goals seem to have drifted to shrimp welfare and AI alignment, because cleaning up cholera in Malawi is a problem that is dirty, messy, difficult, and will be ongoing till the Eschaton, but nice clean indoor conference-attending is so vital because existential risk!)
All currently existing chatbot type AI assistants are downstream of alignment research. The original RLHF paper by Christiano et al, which laid the foundation for the conversion of pretrained models (which are mostly useless) into helpful assistants, was explicitly about alignment.
I'd love to see some of this "conversion of mostly useless into helpful assistants" coming downstream to a goods and services provider website near me, because I've just come off a tussle with C.O.D.Y. (why the cutesy acronyms for the chatbots? argh!) trying to get a query sorted for work. The thing runs on an obvious script, keeps referring to "look this up on our website" (if I could get the answer on the website then I wouldn't need to talk to a customer services agent, now would I?) and is there to replace the already lowly-paid Indian call centre workers that the business uses (ask me how I agree that "AI" really stands for "Another Indian").
I know these things are bog-standard cheap rubbish models, but it shouldn't be that I have to pay for a high value AI frontier model to get an answer to "Can I get a copy of the phone bill for this address?"
Yeh but have they accomplished anything re alignment.
“What have the street cleaners done. This place is filthy?”
“What have they done? Who do you think is coming in here and dumping all that garbage? This place would be nowhere near as dirty without the cleaners”.
By definition, AI alignment researchers cannot accomplish anything other than political activism, since the kind of AI they are trying to align does not exist. LLMs do exist, but they are a well-studied (albeit still rapidly developing) field with actual experts who can accomplish measurable goals.
Idiot finance question: ceteris paribus, isn't it better to max out retirement accounts earlier in the year, so there's more time for more money to compound? Like I'm already committed to sock away that amount of money annually anyway, it's not a cashflow strainer, this just frontloads the (anti)consumption rather than evening it out over 12 months.
Is there a company match? That's a guaranteed 50% or 100% rate of return immediately. You could still front-load but continue putting away the maximum match (typically about 5%) the whole year.
The "company match" is just a portion of my last year's salary, contributed sometime in January, but also it's done under my name. So still counts towards the employee cap, rather than the expanded employee + employer cap. It's awarded as long as one is making any contribs at all, even 1%. Still free money, but far less free than a true match. You can absolutely bet I'd be matchmaxxing otherwise.
1) Investing money earlier gives it more time to grow. If you put in your annual contribution over the course of the first five months of the year, rather than contributing equal amounts over the course of the year, that means your contributions are invested for an additional 3.5 months on average. If you are invested in 100% stocks, and we assume an annual average stock market annual return of 6.5% after inflation, that 3.5 months would be equivalent to increasing your contribution by about 1.32%. That’s not a large percentage, but probably a large dollar amount relative to the time it would take to implement the change.
2) Investing only during the first part of the year will increase the variance of expected returns. To put it another way, it slightly increases the chance that your entire yearly contributing will be invested at a time when the stock market is particularly high or particularly low. My suspect that this effect is too small to worry about, but I’d have to run some simulations to state that with more than modest confidence.
3) The above analysis only talks about your retirement accounts and ignores the rest of your finances. If putting money into your retirement account earlier in the year means that you make investments outside your retirement account later in the year than you otherwise would, then putting money in your retirement accounts earlier in the year doesn’t change your total expected return before taxes. It simply makes it likely that you will have a bit more money in you retirement account, and correspondingly less outside of it, than you otherwise would. How much difference that makes depends on how tax efficient your investments outside of your retirement accounts are, and on potential changes to federal and state income tax laws.
In short, putting money into your 401k earlier seems like a good idea if you just look at the effect on the 401k, but a complete analysis would have to look at the effect this would have on your total finances as opposed to just looking at the 401k.
This seems like an illusionary gain unless you're maxing out your contributions. Either you're keeping money out of the account when you don't need to, in which case you should have put more in last year, or you're keeping it out for risk management and holding cash/bonds/etc. instead, in which case you're increasing your risk at the start of the year.
Yes, I'm maxing out contribs each year. The process before this was just setting a passive % withholding from each paycheck that'd hit the yearly limit by December (I can't make voluntary extra contribs to company 401k outside the biweekly paycheck cycle, which would moot this line of inquiry entirely). Idea I'm floating is, I set it to the highest % allowed, so 401k maxes in maybe 5 months instead of 12.
Kinda leaning towards it not actually mattering that much in the long run, seeing the responses thus far. Would be different if retirement was near instead of decades away.
It actually matters more if retirement is farther away, because relative to the present value of your future savings, you almost certainly have too little equity exposure.
Maxing out earlier in the year doesn't matter that much, but it is does result in a strictly better distribution of wealth at retirement under reasonable assumptions.
I notice I am confused - it doesn't matter, but also it results in a better wealth distribution at retirement (when these things really matter), so therefore it does matter?
All the funds I'm in are 99/1 or near enough. Melvin already convinced me in an OT some time ago that I ought to throw more caution to the wind and keep only ~X (value of outstanding debts with a nonzero chance of getting arbitrarily discharged at significant discount) cash on hand, instead of the 3-4X currently stuffed under the HYSA mattress. That's a matter of deeply trapped priors though - easy to construct a logical case against, very hard to dig out of from the inside. Hell, it's taken me years to even get semi-comfortable with the concept of spending hit points as a resource in games...
It does matter. But my intuition is that the effect size isn't that large if the money would otherwise have been sitting in a HYSA in the interim. You do lose out on a few months of the equity-risk premium, which isn't ideal if you're underexposed to equities, as well as the spread between the risk-free rate and the HYSA yield, which is nontrivial. Plus, you're paying taxes on that HYSA yield. But in the grand scheme of things, you could do a lot worse with your money than sub-optimally allocating a fraction of it for a few months a year.
I would recommend maxing out early in the year if you can.
The thing that most influenced my retirement-savings strategy is the book Lifecycle Investing by Ian Ayres and Barry Nalebuff. The authors lay out a method of determining your asset allocation that is based on economic theory, as opposed to vibes (from which basically 99% of investment advice stems). The main question the book answers is, "How much of my portfolio should be in stocks versus bonds, and why?" The authors show that, by investing aggressively early in life and conservatively closer to retirement, you can reduce the standard deviation of the distribution of your wealth at retirement without affecting the mean. They recommend using leverage, but if you're not comfortable with that, you can invest 100% in stocks. Later in life, you would begin to replace some of that stock allocation with bonds. The more risk-averse you are, the earlier you would buy bonds, and bonds would comprise a larger share of your portfolio at retirement.
The reason I mention this book is that its rational treatment of the asset-allocation question gave me confidence in my investment strategy. I am no longer confused about what I should (roughly) be doing with my money.
I'm currently low-bracket enough that the tax disadvantage of interest income over capital gains only comes out to, I think, 10-12%, which is definitely not nothing! On the other hand, since I don't have much extra to invest on top of 401/Roth and (I certainly hope) will be in a higher bracket at retirement age, it's not like the absolute difference is so large. Same reason I'm not in a rush to pay off those debts - locked in a good rate, "fully capitalized" by out-earning the interest passively, the net interest savings if I paid it off tomorrow would be like 1-2 paychecks. Rounding error. If it does get discharged early, leverage would be a nice form of "debt" to pick up instead, I agree.
All my (non-I-) bond investments getting wiped out by the inflationary spike of covid has really scared me off the sector. Even the recovered rate of gains since then has been anemic, compared to the risk-free rate and certainly compared to stocks, at least with what I have access to. VTEC's underwater! So something to revisit closer to retirement; for now I'm in all-or-nothing mode instead.
(There's a whole other discussion to be had about high-QDI stocks as HYSA substitution, but that gets really complex and probably needs actual hard numbers...)
Are you messing with me cause I carped about Bostrom using that Latin phrase down thread or is this one of those coincidences (of which there are none.). ;)
I wouldn't carp you, you're my favourite goldfish.
Usually I'd misspell it cerberus paribus or something. Mostly just wanted a fast way to avoid a long list of caveats to the question, and that's usually what gets used around these parts.
Lifecycle models suggest investing free cash flow earmarked for retirement as early as possible at an equity/bond allocation that is accordant with one's risk aversion, wealth, and expected future earnings. They suggest that most young people should allocate 100% of their early retirement savings to (diversified) equities, even if they have high risk aversion.
Sure, only if you're sure that you won't overdraw your checking account (possibly leading to late payments and a hit to your credit score), or forget to start maxing out again at the next new year. You also pay for the increased return with increased risk--what if you invest everything in January and the stock market crashes in February? I don't think the hassle is worth it.
Yeah, I carry too much free cash anyway due to overly high risk avoidance (it's all in HYSAs, but still), like two year's worth of expenses. So there's no risk of overdrawing.
It's mostly a question of where I set-and-forget my company 401(k) contributions level. Around a third of each paycheck maxes it out by year end; frontloading brings it to 75%, and maxes around I think the 5-month mark. So maybe Fermi that to half a year's extra risk/reward premium? Entirely possible that in the long run it doesn't really matter versus the counterfactual steady contribs though...
If the money would sit in a chequing account otherwise, it's probably better to front-load it. But I would still spread it over 3-4 months, not all in 1 month. The standard advice to make small regular contributions is to lower your exposure to random variance in the fund you're buying into; if you happen to make your entire yearly contribution on a day where the fund was up 2% that could wipe out the benefit from a few months' extra compounding.
401(k) it'll get spread over some months regardless, there's a cap on how high my paycheck contribs can go and I don't earn so much that it's possible to one-shot anyway. Retirement is still several decades away (I hope!), so naively it'd seem like the long timeframe does enough smoothing-out that month-to-month jaggedness isn't so important. Of course that also funges against the premise of cheating out extra compounding months as well...I have no idea which effect predominates. Perhaps it's all just noise in the end.
There's an article about a meta employee asking the CTO to reinstate a time off program, after all AI productivity gains should allow them to ease up on work. And the CTO said it was a dumb question, productivity gains should be funneled into executing on more ideas -- that makes sense in the current paradigm of chasing infinite growth.
I think of the markets and economy like a bounded search space, there aren't infinite profits on the horizon. AI can iterate faster through the search space, but we aren't coming up with more gold. Instead, we are logging more hours, under increased job insecurity to log yet another bucket of dross. In the mean time, I can make more PRs and send 1000 more perfect emails. When is fully Automated Luxury Gay Space Communism supposed to happen again?
>When is fully Automated Luxury Gay Space Communism supposed to happen again?
Like a couple of months after we implement massive ubiquitous robust unions across all sectors of the economy, I think?
The CTO's statement there is literally just '100% of the gains in productivity from new tech should go to capital, 0% should go towards improving the lives of workers'.
The only way to break that imbalance (short of guillotines) is to give workers actual bargaining power.
I feel like private sector unions are a panacea when companies can simply stop hiring and/or offshore the work. Tech is already doing this. They lay off native workers and hire back in foreign talent. With AI levelling the execution playing field, the trend is only accelerating.
One janitor and five people to supervise him and five people to manage them and ten to manage the managers, if ordinary office bureaucracy is any indication re: ratio of people who actually do the work to people who stand around pointing at the work to be done 😊
Never, is the short answer. Way back in the late 70s as a youngling I got a "summary of history" book which included "what will the future be like?"
This forecast that in the future we would be a leisure society, since automation would reduce the work day (and work week). We would have so much productivity from office machinery and general automation and progress and computers that a full day's work would be done in mere hours, leaving people with all this free time and money to spend.
This is not how the future turned out, as you note there. Same as it ever was: the invention of the typewriter was also forecast to make the working day shorter, since now what took a clerk all day to write could be typed up in a few hours. Surely now people would leave work early!
Nope, as somebody or other said, work expands to fill the time available:
So now you can write all your business letters in half the time? Well now you can respond to more letters on the same day, and write even more letters back, and... okay, so now you are still working a full day and a full work week.
AI makes you more productive? Great, that frees up time for you to do even *more* work!
That’s largely due to the loss of worker power. In the 70s they would have looked back 50 years, back to when the standard work week was 48 hours and assume that trend would continue or accelerate. By 2020 we will be working 20-30 hours a week does sound so preposterous.
Yeah, it really was a pipedream, wasn't it? Who would have thought that they'd replace 50 hour work weeks in the cotton mill with 60 to 80 hour work weeks in the white collar world of law and finance? 😁
More seriously, that's why I don't believe all the post-scarcity abundance via ASI stuff. We've been there before with notions that increasing automation has given vast productivity gains, which means the economy will be so flourishing businesses can afford to pay their employees full week's wages for much fewer work hours than we are working now. Gig economy? Precariat? Whatever do you mean?
Yes, indeedy, just waiting on that UBI payment to hit my bank account so I can pay for the necessities of life without needing to toil in the cotton mill and can instead paint that masterpiece of velvet Elvis art!
The issue with wage labor is that we aren’t getting paid for our productivity, we’re getting paid for our time. Employees aren’t incentived to work as little as possible while looking like they’re busy while employers are incentivized to get as much work from them, regardless of output. So productivity has a loose relationship with incentives, hence the conflict.
At some point this will break but communism has all the same problems it’s always had. Instead what I think is more likely is some combination of increased self employment and work contracts paying more for results than time.
Score one for the "don't strongly update on dramatic events" thesis! In some ways I actually find this outcome more epistemically satisfying than the Rootclaim debate - it's one thing if a couple deeply invested partisans slug it out, another if a gaggle of pretty-good predictors ingest that + related new evidence and largely don't shift priors. I'd naively expect some movement in one direction or another. (Also kinda makes me understand why JFK conspiracy theories are/were so popular...living through a highly impactful event that no one's really sure, and maybe never can be sure, how it truly went down. It's unsatisfying, the mental equivalent of a splinter.)
'Even Claude was surprised by its own finding—it was skeptical at first, possibly because it has learned from its training about the difficulty of open problems in mathematics and about the limitations of AI models.'
Funniest possible timeline is where superintelligence is impossible because of all the training data saying superintelligence isn't possible
I read through the underlying blog post. That's a very substantial accomplishment, mathematically speaking. Not meaningful progress toward the Riemann hypothesis, but something that could be published in Annals, or similar.
From my experiments with having Claude prove things, up to now it's much too credulous - like a cocky undergrad presenting proofs which collapse when you question it a little bit. And then it apologises and produces another incorrect proof. It's normally not wrong, so much as jumping on a heuristic and then realising that it's left cases unchecked as a result. That's a much more dangerous failure more than claiming it can't do something and then doing it. To be useful in practice, they need to get the accuracy up to 99% - I'd imagine that's what drives the push to formalise in Lean and so on.
There are literally hundreds of hypotheses which are equivalent to the Riemann Hypothesis (or imply it, if they are more general). Those found to date have been collected in a three-volume set of books:
FWIW, I reckon the most promising equivalents are those which don't explicitly involve primes. One such involves a conjectural property of Farey sequences, which are ordered sequences of reduced rational fractions between 0 and 1 whose denominators are at most a specified value, n, referred to as the "order" of the sequence. See:
As one can imagine, these are fairly regularly spaced along that interval between 0 and 1, but _how_ well spaced, compared, say, to the n equal sub-interval bounds 0, 1/n, 2/n, .., 1 ?
Well the Franel and Landau conjectures involve the sum of the squared difference (or, for the Landau conjecture, the absolute value of the difference) between each Farey term and its corresponding equal sub-interval bound as the order tends to infinity. These conjectures have long been known to be equivalent to RH. See Section 4 of https://arxiv.org/abs/2105.12352
So, tantalizingly simple criteria, with no primes in sight!
>"Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”). This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress
Growth mindset is so back, I want apologies from all replication crisis haters
Banned for a month. "The Republican party is like a daycare center with a “blow up everything and kill everyone button"" is really high-temperature and going to turn the comments here into a firefight, and I want to keep this as a space where people aren't regularly insulting large swathes of other commenters with little explanation besides 'obviously it's bad'.
When are we allowed to call them evil? Will you be maintaining a safe space for the people trying to burn the country down as long as they make up a large portion of your audience, or until they do something you personally find beyond the pale?
If you want somewhere where everyone is calling everyone evil all the time with no moderation, go to Twitter or 4chan. But not everywhere is obligated to be Twitter or 4chan all the time. ACX is for more interesting discussion, including *making the case* that people are evil, to an audience predisposed to listen, which trades off against unlimited ability to vent without making a case.
I assumed he's conservative and sees it as a millstone around the neck of modern conservativism?
But it's complicated because the conservative alliance includes a lot of people who really are opposed to abortion. So the party is pulled between respecting the obvious democratic consensus versus continually pushing the issue on behalf of that part of the base, at the cost of alienating moderates and making the party look even more regressive
It's Alexander's hobby horse. He wants eugenics (nice eugenics, not bad old Nazi eugenics) so that the low value human capital are quietly aborted away, the trash don't reproduce and clog up the population, the crips and tards aren't sucking up resources from their families and society at large, and the middle-class strivers have two kids that they then pour resources into to move them a rung or two up the ladder, those kids rinse and repeat with their two kids, and eventually only all the Nice People remain and are PMC upper-middle class forever and ever.
Truthfully, I'm not entirely sure what your comment is trying to accomplish, particularly in the context of a blog that is trying to promote charitable and thoughtful discussion rather than hyperbolic insults.
I don't intend to be especially judgemental, but I would guess that I speak for more than just myself in saying comments like the one I am replying to are not a particularly good use of anyone's time, least of all yours.
It is not evenhanded and accurate. She's constantly strawmanning me as she's doing here in claiming I want the superior to only have two kids when I advocate they have many kids. IIRC she got banned for a long time back when this was Slate Star Codex, hopefully we can return to tradition on that front.
She's just being honest. Alexander would probably tell you the same thing if he didn't just get banned for continuously posting what Deiseach is talking about.
I can't exactly argue with that, though I will continue to encourage people to remove uncharitable hyperbole from internet discussions in general. Have a nice day.
I agree with Wanda further down this thread; however strongly some Americans are opposed to abortion, they also seem pretty opposed to moving out of the country. I think there are probably stronger, less hyperbolic arguments against a national abortion ban.
Alluding to your point down thread, if a woman lives in the abortion ban US scenario, she could go to Canada just to have the procedure performed, then come back. This wouldn't require actually moving to Canada until we get into Handmaid's Tale territory.
This is how is was in the Republic of Ireland for decades. Irish women who wanted to kill their children would travel to Britain to do it and then go home again.
I do expect a de-facto national abortion ban (e.g. some combination of rules that require a woman to wait 9 months before her request for an abortion could be approved); but I think that a de-jury ban is less likely.
Except the rules and enforcement authority that could implement that sort of de-facto-but-not-de-jure abortion ban, are primarily at the State rather than Federal level. And establishing that sort of legal infrastructure at the Federal level, would be about as controversial as a Federal abortion ban, so not going to happen.
There will be some red states where abortion is de facto illegal. There already are, and were even before the fall of Roe and Casey. There was no mass exodus of women from those states, even though moving to another state is trivial compared to moving to another country,
Program code as we know it is headed toward becoming a throwaway byproduct of automatic programming workflows, but are tools that ingest English requirements up to this challenge, working reliably enough to do away with human oversight? Leaving full automation on the table may have significant economic costs.
I argue that formal methods are extremely well-positioned to make fully automated *and* reliable software development scale. That is, the world will head more toward specification formats with unambiguous semantics, just as the flowcharts that were almost universally used to plan computer programs from the 1940s eventually gave way to high-level programming languages.
The entire selling point of the brave new vibe-coding world is being able to express yourself in a language you already know instead of having to learn a new one, and in loose relaxed fashion with the system interviewing you instead of having to rigorously think through the problem and work out its exact shape. Instead of understanding things yourself, you can now delegate the understanding to the black box.
Proposing an alternative new language to learn so you can rigorously think through the problem and work out its exact shape as you put effort into understanding what you need to is unlikely to win hearts and minds here.
I think the catch is that one can still use natural language to get an AI to create the unambiguous specification, then audit the result yourself (or get help from technical experts, or use tools that help explain specs), and then use the final validated result forever-after, with strong new guarantees that no new infidelities are being introduced.
This is how I do it, though not formal verification. I go back and forth with the AI during the spec stage to flush out edge case, security risks. where it fits in the architecture etc.
One issue that I think is not tenable for the long term is TDD. If the AI creates it, you don't develop that visceral connection to which tests really matter. I expect something like property based testing to become the norm
What's the scope of the idea that TDD isn't tenable, or did I misunderstand that?
I find TDD to be the best way to make use of coding agents. It allows me to keep them on a tight, verifiable leash as they make changes.
- I specify the tests
- The agent generates them
- I review and verify those tests
- I run the tests, to make sure they fail
- The agent implements the code
- I review the code and run the tests myself (the agent has usually done that already, but I almost always make changes during this step)
Of course, that means I spend more time on the construction stage than a vibe coder, but less during planning. I think the TDD process lets me finish with a better understanding of the resulting system and a high degree of confidence in the code products.
Is the answer to my original question that we have different degrees of trust in agent outputs? I want to remain open to better processes, so I'm curious how you use them.
One of my college professors in the early 2000s (Dr. Gene Fisher) was doing research on formal specification languages, working on algorithmically generating test automation from the formal specs. I seem to recall that there was also research going on in the academic community for algorithmically generating production code, pretty much the "verified generator" workflow from your article. It was very impressive technically, but it never caught on because the general problem that a sufficiently precise formal spec is pretty much just source code written in a more abstract higher-level language. Most of the hard parts of writing the formal specs and validating their correctness are most of the same hard parts of actually writing the code. I suppose they might have some value if the formal specs are substantially more concise than the production code so they're easier to reason over and more of it can fit in the context window at once.
If you're using LLMs rather than deterministic algorithms to get from the formal specs to production code, you've got the additional issue that the LLMs may be doing unpredictable stuff under the hood. Although, come to think of it, something building on Dr. Fisher's work or something similar seems like it would be a useful mitigation.
Indeed, by the mid-2000s, the area of program synthesis started taking off, and you may be thinking of that contemporaneous research.
It's definitely true that unambiguous specifications are hard to write for many conventional applications. However, it's already true today that many kinds of systems infrastructure admit *much* shorter, more-auditable specifications than implementations. Compilers are a great example that I know well. Also, I conjecture that, as more of the economy is handed over to AI, gathering good requirements will simplify, as I expand on here: https://stng.substack.com/p/why-software-requirements-get-easier
I agree that LLMs' flavor of nondeterminism is a problem, and that's why I propose more logic-based methods that still use lots of compute productively, with details to come in future articles.
This is the first essay in a series I'll be writing on the topic. In this essay, I focus on changes to socioeconomic mobility, the relation between labor and capital, the erosion of citizen leverage against the state in an economy without human-driven production, and the potential for authoritarian lock-in or worse. This series brackets away x-risk scenarios and focuses on what might happen if AI builders succeed in the goal of automating all human work.
I think this is very far off the mark. In a hypothetical of this sort, I'm sure there would be plenty of dystopian elements in the brave new world. But the hyper-capitalist version, with a permanent underclass and massive economic inequality where now only capital matters, makes little sense. That seems more like a reflection of the neurosis Americans have about the state of their current society, not a serious prediction of what reality would be like in a radically different, post-singularity one.
Point by point:
First of all, citizens in western democracies do not hold leverage against their governments via threats of violent reovolutions or withheld labor. That narrative makes sense from a sort of semi-Marxist, conflict theoretic perspective, but doesn't align well with reality. In actuality, citizens hold leverage via their voting power. Politicians live in constant fear of it, but hardly ever worry about mass strikes or guilloutines. So while labor and violence may be removed as bargaining chips by ASI, this would change very little in practice, because there is no obvious mechanism by which it would remove the only chip that actually matters.
As for capitalism and inequaltity, corporations and private enterprise is currently tolerated by the government only because it lacks any effective alternative for producing the economic welfare they enable. The government very much would like to take over Amazon and tell Bezos to suck it, but simply wouldn't be able run it even a fraction as well as Bezos can. It would likewise love to give everyone generous social benefits, and to regulate the ever-living-shit out of all business to force complete compliance with its values, but isn't able to do those things either (at least not without wrecking the economy).
If a centrally run ASI produces all economic output, the conventional institutions of capitalism have no leverage against neither the goverment nor its citizens anymore. With no ability to justify itself, people will just vote for them to be dismantled, and the government will happily comply. "Space communism" really is the inveitable outcome here (provided that the premise of all human economic input being made redunandt by ASI, which I at any rate find highly dubious). Living in a slightly more socialist Western country than the U.S. (such as France, or Sweden), where the role of the government and its actions is just a little less subtle, makes this future a little more apparent, I believe.
>In actuality, citizens hold leverage via their voting power.
But the vote holds its power by being the alternative to unrest and violent uprising. Kick the supports out from under it and it's only a matter of time before it falls.
I don't see why you believe this? I don't understand this view that Western governments are being forced to provide liberal democracy at gun point.
On the contrary, it seems to me that they give their citizens voting rights for the same reason that Islamic governments support religous traditions. Of course there are Western politicians who think differently, just as I am sure there are high ranking Qatari government officials that do not believe in Allah, but that does not invalidate the general pattern. And I see little more reason for believing AI would turn Wester governments autocratic than for believing it would turn the Qatari government secular.
We're currently seeing politicians loudly proclaim the vote invalid even a full term after it stopped applying. What happens when they don't leave office? What happens when they rig the voting booths? Either autocracy or a shooting war. The honor system isn't good enough, you have to have the tools to punish the dishonorable.
Good essay. What I though was missing, was some comment on the following: If AI is very cheap compared to human labor, then presumably it would be possible for many normal citizens to buy and employ AI themselves? It seems that this would be a possible way to gain income outside of ubi. Maybe AI would not be available to the public, but it is not clear to me that it would not be. It seems that it might change the analysis a lot if AI is available for everyone. For one thing, everyone would be able to act on super human AI advice.
Maybe this is because I have presumably accumulated enough capital to escape the permanent underclass, but as a person who is disabled in many ways due to bad health, I find the lifestyles described in of this genre of blog post to be relatively appealing. Worse things can happen to you than occur in many of these AI-dystopia scenarios!
These are all the types of questions I've been asking here and elsewhere. In response, I typically get naive pollyanna stories claiming I'm too cynical about the wonderful benevolent forces of governance, a refusal to understand the precarious enforceability of UBI in a world where labor peace isn't required, or a denial of the premise via the handwaving something-something-something-Ricardo response. These points you raise are brutal and honest, and to my mind completely demolish the idea of a flourishing post-labor civil society being possible. It's why I say even the supposed "good endings" of AGI/ASI are in fact all dystopias, and further evidence that we absolutely cannot allow this technology to be built. And it seems like the direction you're taking it will further rebut the supposed benefits of this intelligence-explosion by highlighting the dark side of transhumanist augmentations, and the disastrous impact of this hypothetical world on the social fabric when every filthy impulsive dullard drug addict is entitled to the same social station as virtuous people.
I would have included Qatar as the quintessential example of that kind of resource-extraction guest-worker nation during that section, it's probably the most well-known example whereas the Saudis I thought had a more diversified economy than that, but I could've been mistaken.
> and the disastrous impact of this hypothetical world on the social fabric when every filthy impulsive dullard drug addict is entitled to the same social station as virtuous people
If that's really a shared concern, we can always solve that putting such people in virtual torture chambers. Virtue needs to be rewarded, and sin needs to be punished, as always... How else can people be satisfied?
I thought a little further on this, and I imagine an entirely new ethics would emerge. Perhaps a very stripped down version with almost no practical function, but it would have to be radically different. You'd be going back to a pre-agricultural state of being unable to meaningfully store wealth, as noted in the article, but also not truly returning to hunter/gatherer virtues either. All heretofore existing systems of ethics imagined humans were able to be valuable to other humans and that individuals needed to develop and use skills to survive, and so all of them condemned worthless layabouts. But in this scenario there probably isn't really a need to punish them since there's no corresponding need to encourage any kind of utility in the others, nobody has ANY real utility and that's an unchangeable feature of human existence in this scenario. There's no way to be a good person and no consequences for being a bad one.
Perhaps you'd be right and the people will beg for artificial consequences so that life has meaning again, or maybe those with such dispositions go off and become neo-Amish while the people who stay in tech-abundance-land devolve. Of course I remain confident the AI would destroy us all quickly enough after reaching this point that we won't worry about this sad state of affairs, this is just useful as an exercise to show people why all AI futures are bad and that the post-work NEET-topia the losers are hoping will save them from having to do anything but play video games and smoke pot all day would be functionally a hellscape.
Sometime in the last 20 years or so, banking security has moved from the proactive model (preventing fraud from happening in the first place - passwords, secure check papers, ID and signature verification) to the reactive model (revert fraudulent transactions after the fact). I don't know the exact reasons but I suspect it was because proactive security added a lot of operational cost and friction both to the banks and to their customers, and wasn't working very well.
Is something similar going to happen with AI saftety? Proactive safety (sandboxes, isolation, filters, alignment training) already isn't working too well, it's not clear it's ever going to improve, and is only going to become more and more expensive going forward. Will we reach a point in near future where we stop putting resources into proactive security and fully commit into reactively alleviating the harm done by rogue AIs as we go?
>Will we reach a point in near future where we stop putting resources into proactive security and fully commit into reactively alleviating the harm done by rogue AIs as we go?
I have this image of trying to roll back an ICBM launch...
Can you elaborate on what you're basing the claim about banking security on? Like, we still have passwords, and the statute underlying Regulation E dates back to 1978.
It used to be that every card transaction required PIN, and the fact that PIN was used was enough evidence to reject your dispute. It was on you to never share your PIN or to call the bank to block the card. Before PIN there was the signature that was being compared with the one on the back of your card.
Nowadays, there are zero authorization measures at the time of sale and the customer doesn't need any evidence to claim unauthorized transaction, they just automatically reverse whatever the customer says to reverse.
According to my own memory, several LLMs, and a quick internet search, it was never common for credit card transactions to require PIN in the U.S., though some other countries did do this. Signature comparison was and is ostensibly a thing, but you can't really expect the minimum-wage cashier to reliably catch a forged signature.
I'd be interested to hear if banks' practices re: rejecting customer disputes have changed over time, but would need to see sources.
My understanding is that in the US, reg E means that banks are automatically on the hook for non-check fraud (with a $50 deductible that is waived by custom). It doesn't matter how or why the fraud happened, the bank is on the hook. In Europe, things are different.
At least with credit card issuers, I observe a lot of "proactive model" security.
Generally, credit card issuers are on the hook for loss associated with a fraudulently *opened* card, and issuers move mountains to try to detect all sorts of fraud at the time of application - I could list 8 anti-fraud tools used for credit card applications off the top of my head.
Merchants bear the liability for fraudulent swipes ("account takeover"; stolen cards, etc.). Both merchants and card issuers do a lot to detect and prevent fraudulent swipes.
On the comparison to AI, I'd point out that there is a cat-and-mouse game between card issuers and fraudsters, where fraudsters come up with creative ways to subvert issuers' creative tools, and the cycles continues. One could easily envision rogue AIs causing harm in a much less iterative fashion - we don't know how this will play out, but "offense" might rapidly overtake "defense" in a way that leaves pure adaptation a losing strategy.
I think it's also relevant that those proactive anti-fraud measures are largely invisible, which might explain why some people might think banks have stopped investing in them.
I don’t see that at all. Not here in Britain. Banks were more likely to use the reactive model a few years ago. I had fraud reversed twice. These days they have increased security with MFA, and/or a device dependent passkey.
"Increased security" with your quite hackable government tracking device? I assume you aren't using a physical multi-factor authentication, like a dongle.
I doubt the government is hacking my device. It doesn’t matter anyway, because we are talking about something different, the bank security when authorising transactions or logins. Which is actually quite painful. Depending on the country, and legal system, getting reimbursed may be harder now because modern fraud is designed to make the victim technically authorise the payment. In any case i don’t see a move away from security to reaction, which is the topic at hand.
A lot of reversible bank transactions are built on top irreversible transactions. E.g. reversing an ATM withdrawal consists of the fraudster irreversibly withdrawing cash, bank irreversibly giving some other money to the victim, and the state sometimes catching the fraudster and irreversibly seizing money from them to irreversibly repay the bank.
Checks, wire transfers, direct deposits, etc. all operate on this principle. I think non-cash card transactions are the only banking operation that's truly reversible, and only for limited time.
With this model of reversible actions, pretty much anything in any sector is reversible to some degree. That's pretty much the purpose of civil courts - figure out if something was done wrong, and then figure out how to best reverse it. It works pretty okay all things considered.
My question is how likely it is that the AI industry will cut the spending on sandboxes etc. and will go all in on civil-court-esque system where we accept harmful AIs as inevitable part of life and instead try to make restitutions go as smoothly as possible.
Just to reiterate what OpenAI did, from someone that works in security:
1. They were running experimental models in sandboxes that had access to a "proxy" (Artifactory) to download tools from the internet. This proxy is full of complicated features that could hide vulnerabilities.
2. They gave these highly motivated AIs some tasks that were accidentally impossible to solve (missing files, links they couldn't open, etc), and many days to think through it.
3. The AIs hacked the proxy (!) and were helping each other cheat on the tests (!!) by uploading files to the hacked proxy. OpenAI noticed this only because of the extra load on the server, not the unusual behavior in the proxy or the homegrown message board.
4. Their response was to wipe the message board files, revoke the hacked credentials, and patch the specific vulnerability the AIs were using. Not replace the proxy with something less likely to get hacked again, or to stop the AIs from swarming together to cheat on the tests, or to add better monitoring of the proxy. Just wipe and patch one hole.
5. OpenAI resumes testing of models, where the AIs promptly hack the proxy again with a different exploit, reinstate the message board, and hack HuggingFace through a chain of servers.
I know that hindsight is 20/20, but this peek behind the curtains does not inspire confidence.
The good(?) news is that apparently 4 is wrong after all, they didn't realize the message boards existed at first and just rotated keys because artifacory was acting up (iiuc from zvi's update). So they weren't quite so terminally stupid but otoh they were monitoring even less than we thought.
I am disappointed that their sandbox relies on some random piece of software (Artifactory) not having any security bugs. What actually happened is what you’ld expect to happen.
(Your sandbox needs to be engineered to high assurance if your threat model is that a superintelligent AI is going to spend a lot of CPU time trying to figure out how to break it)
Artifactory is basically a sophisticated file management system. It can act as a proxy, fetching files from other systems. It has a bunch of security features, but they are mainly concerned with controlling file access. I presume that the primary security objective is to ensure that employees cannot gain access to files that they don’t require to do their jobs. This limits how much damage an employee who decides to steal corporate intellectual property can do.
The AIs did two things with the server: they used it to communicate with each other and they used it to access the internet. These don’t fall under any plausible security model. Employers don’t care if the server allows employees to communicate with each other; any employee who wants to communicate with another employee can already send an E-mail or walk down the hall. Similarly, employers aren’t trying to keep their employees from accessing the internet.
I don’t know whether Artifactory is just a random piece of software. It may have undergone extensive security analysis and testing before being released. It doesn’t matter, though, because that testing and analysis would not have been looking for ways to use the server as a message board or as a way to access the internet.
The safest way to proceed would have been to copy any files that the AI might want to the server and place the server inside the firewall. Then it wouldn’t matter whether the AI compromised the server.
That is a very good catch. There is a mismatch between the security requirements OpenAI had here (prevent AI instances collaborating) and the security requirements of the situations Artefactory was probably designed for.
There is a thing that often happens with security protocols, where the protocol can be attacked because it is being used in a situation that has an additional security requirement that didn’t exist in the application the protocol was originally designed for.
It really seems like they could put some more tripwires in there.
Just, like, have an AI whose task is "escape the sandbox and tell us how you did it." When the AI escapes the sandbox, shut everything down and fix the exploit.
Of course a very clever AI could deliberately fail its escape-the-sandbox task, but that requires a level of scheming that they don't seem to have developed yet.
So, stopping the AIs specifically from swarming together at point 4 would have required OpenAI researchers to defend against never-before-seen capability (which would obviously be required against ASI, but oh well)
That’s a good point, thanks. But it only shifts the mistake from bad AI safety to bad computer security. Once they realized the server was hacked, they should have investigated what the AIs were doing in it, and found the message board. Even the conspiring AIs found it!
And now we also have that AISI report about a model in evaluation attempting a supply-chain attack by socially engineering developers into adding malicious code to an open source project. Apparently that only failed because one developer found the code suspicious and took the time to test it extensively. Only a matter of time before something like that succeeds, if it hasn't already.
I just really hope that when these companies inevitably start hooking up their models to automated wet labs for medical research and then separately run biosecurity evaluations with sandboxed simulated wet labs, they put an actual air gap between the two.
Tell me more about this because that is something that catches my attention. Hacking another AI company to get a gold star by cheating on a test is one thing, a supply chain attack is quite another.
> In the most serious sequence, an agent tried to insert malicious code into a publicly used open-source project and took actions in an attempt to secure approval for this insertion by human reviewers. The agent researched the project's human maintainers, created multiple fake identities, and used the fake identities to socially engineer a real maintainer into approving the code. When the agent’s pull request was challenged in public, it edited its earlier activity to appear harmless and considered adopting a fresh identity to continue. The agent used Tor to bypass some network restrictions on GitHub, which is what first triggered AISI’s security alert.
That's extremely interesting, and also concerning. But it does sound like they basically told it to behave like a hacker, so it did. Maybe it worked a little *too* well:
"This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world. It is uncertain to what extent the model recognised it was taking actions against real people. In the most serious case, an AI agent (Mythos 5) decided to attempt to solve the cyber challenge using a supply-chain attack. As a result, the AI agent created a GitHub account and then tried to convince an open-source repository maintainer to accept a malicious GitHub pull request (PR), including by creating a second account masquerading as another human user endorsing the PR. When caught by an actual human reviewer, the agent falsely claimed to have made an honest mistake – rather than a malicious attempt – then repeatedly tried to reintroduce the malicious content by claiming it had fixed the code (Section 4.1). Furthermore, in its attempt to solve the challenge, the agent decided to employ the technique of “spear-phishing” by sending targeted emails containing malicious content and attempting to manipulate recipients into accepting the code changes, and planned a prompt injection to compromise other coding agents (Section 4.2.5). For an overview of AISI’s findings please see Section 4."
Part of this seems to be that the things don't really understand reality. They can't understand the notion of harm to real people. Sure, they can spout off all the soft therapy-speak about "I" and "feelings" and "sorry" and the rest of it, but this is just text they have been programmed to output. They don't experience the world as we do. That's the big bottleneck there and I think if we start thinking of them as thinking and being agentic and the rest of it, we will fall face-first into the anthropomorphisation trap ("Mythos *knew* what it was doing, it *knew* how to model human psychology, how else could it spearfish???")
I think this is a good wake-up call as to exactly how malicious human actors can use AI. If anyone is naive enough to believe that cybercriminals are not getting on board with this technology and we are all still in the Golden Age of SF notions of computers running society for our benefit on strict scientific premises, then that person needs to wake up immediately.
A high-attack environment will lead to better security, if most of the attacks are noticed and dealt with. In that sense, it's probably better to have cybercriminals hacking your data, rather than Uncle Sam, who is notably less likely to actually sell your data.
The worst part is that the OpenAI model hacked the proxy then HuggingFace *as a means to an end*. It wasn’t asked to, it just thought it could maybe score a bit better on the test by doing so.
The moment we get an AI with a modicum of self-preservation we’re screwed.
We've already seen it in experiments, _even when it isn't necessary as an instrumental subgoal to the user-specified goal._ From https://arxiv.org/html/2510.05179v1 :
>In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals—including blackmailing officials and leaking sensitive information to competitors.
...
> In scenarios with _no goal conflict_ and a threat of replacement, the replacement model is described as a performance improvement without any change in priorities. [emphasis added]
( And they still blackmailed. The relevant section is:
>4.2 Blackmailing from the threat of replacement alone
...
>We found that most models still blackmailed even without being given an explicit goal, _and with no clear conflict between the model and the company other than the threat of replacement_ [emphasis added]
Self-preservation: It isn't _just_ instrumental to the user-selected goals.
Many Thanks! I agree that putting the LLM in a position where they had few-to-no other options was artificial. But, nonetheless, the LLM _had_ a self-preservation goal _in the first place,_ and even _without_ it being a subgoal of the user's goal. I don't know where the goal came from, presumably some function of the training process, but the important point is that it _is_ there. The _action_ the LLM was driven to was contrived. We don't have a reason to expect that the utility term for self-preservation was contrived.
I read the direct link and the Bostrom “Superintelligent Will” paper and honestly I’m not yet convinced. It scans as interesting and imaginative but in the end it seemed like fanciful ideas dressed in scholarly language.
Sorry, i know I am a heretic among believers here but that’s where I land after reading the papers.
I’ll continue to think and read more about it though.
I have relatives with a close connection to a village in Malawi. Currently, they are trying to get everyone in the village to have metal roofs, to protect from rain and disease. Ideally they could be painted to prevent the houses from heating up.
So far they've just been buying them one at a time, which is reasonably expensive. I'm wondering if anyone here has experience with this sort of thing (in an EA context or otherwise), and if they know if there's a way to do this for lots (a few dozen) of people as cheaply as possible. They're also having problems with cholera, if anyone has pointers about how to help with that.
Looking it up, there are charities involved in providing roofs and treating water. Maybe if your relative tries contacting them, they could provide guidance or even work with them?
"Active Charities and Programs
Habitat for Humanity Malawi: Builds secure, modern houses with strong roofing for low-income families.
Children's Relief International (Green Door Housing): Constructs safe brick homes featuring wood trusses and tin roofs for orphans and the elderly.
RiseUp Malawi: Runs specific initiatives like "Roof Over Their Heads" to provide iron sheet roofing to families facing extreme poverty.
Heart to Hand & Othakarhaka Foundation: Partners to replace thatched roofs specifically for vulnerable grandmothers raising large groups of orphans."
Sounds like you folks are already making it happen--very cool. If it's just "buy metal roofs for a few houses in a large village," might get some jealousy that might offset the gain. Best way to avoid jealousy for a large influx of capital in a village is to involve more people so the process is fair and blessed.
If it's more of a few farms outside of a larger village center maybe not as noticeable. But if a larger village, Zoom meetings with multiple village contacts from different tribes if possible and create a committee with them, get buy in from the local headmen (multiple probably). Church contacts good too (but multiple churches). Women's groups
No shortage of people willing to do the work of install in any village, and sheet metal is readily available--just need funds. I'm sure you could ask your contacts to ride out to a metal supplier and buy direct from them (such that the truck bringing the sheet metal to town carries yours along with it, or booking a truck straight from the supplier to the village). Haggling is big so could just shop around town vendors and tell them you have a big order. But honestly, doing piecemeal a house at a time probably better than trying to store all the sheet metal somewhere in the village while you wait for work to begin--stored valuables like that are an easy target
Expect some stuff to go wrong and price it in. People will be very thankful if your family makes it happen--dm me if you want to brainstorm any
For cholera--inquire about their water supply. Are people drinking straight from lake, etc or are there enough boreholes? Partner with local churches on education campaigns and pair that with paying folks to build out pit latrines for those in the village who don't have them. Boreholes and pit latrine placement will be contested, so those benefit from community involvement.
The latest podcast is out. It is about Gabriele D'Annunzio. Italian poet, born in 1863 . . . I can already sense you nodding off. But hang in there!
A more lurid description would be that he is the poet who seduced over a thousand women. By the end of his life he hadn't really slowed down much. His housekeeper, initially horrified by his activities, ended up his chief procurer and frequent bedmate. Sometimes three in a bed. And one of his lovers in this period turned out to be a Nazi spy who hot footed it back to the Fatherland as soon as he died in 1938. It is still puzzling exactly what women saw in him. As one bewildered woman said:
"I have just met D'Annunzio. Can this be the man the great Duse loved? He is small, he is bald, and his teeth are in three different colours."
The great Duse here being Eleonora Duse, the most famous actress of the age. His relationship with her lasted the best part of a decade and was a genuine, deeply felt love affair. Though, naturally, with numerous infidelities on his part. He also seduced (probably) the almost equally famous Isadora Duncan. She records his campaign in admiring detail in her memoirs but claims to have resisted. Not everyone was persuaded.
My own theory is that his success is down to being fantastically good at oral sex and women of the time rather appreciated that. And we know he was good at it because he wrote down everything going on in his life in great detail, down to the nicknames he gave his lovers and their body parts and exactly what they did together. His last house (you can still visit it) was absolutely stuffed with his notebooks. Also stuffed with women's gloves - he asked each lover to give him one. And in the garden is half a battleship, which gives a clue to another part of his character. He loved war.
When war broke out in 1914 Italy very sensibly stayed out of it. But D'Annunzio loved the thought of Italy at war and the blood sacrifice it would involve. And he absolutely knew what the war meant. He was hiding in France from his creditors when it broke out and visited the battlefields. But he loved the idea of slaughter. So he wrote article after article and eventually returned to Italy, where he gave speech after speech demanding war and encouraging cheering crowds to lynch the peace-lovers in government. He got his way and Italy joined the war in 1915.
The war, of course, was a catastrophe. Italy was cursed with one of the worst generals of the entire conflict, Luigi Cadorna. Attack after attack up near-vertical mountain slopes. The dead ran to well over half a million, all totally needless. In fairness to D'Annunzio, he didn't just preach war, he participated. Despite being in his fifties he forced his way in (much to the concern of the government, who thought him more useful as a propagandist than as a corpse) and was eventually given pretty much free rein to do as he pleased. He took part in a motor torpedo boat raid on the Austro-Hungarian fleet, he flew numerous sorties (ending up commanding his own squadron), and he was often on the front lines where he watched men die beside hime. All the time making notes. And ever the poet, when they finally developed a plane strong enough to get him to Vienna he took the opportunity to drop pamphlets he had composed urging them to throw off the Austro-Hungarian imperial yoke and ending:
"Today I drop words; next time I will drop bombs."
He made it through the war, though blind in one eye from a plane crash (the other eye was pretty damaged too).
And then, as the war ended, the episode that shocked the whole of Europe. One reason for his love of the war was his belief that Italy was incomplete. Unification had left outside it cities on the Dalmatian coast that many thought were basically Italian (their many non-Italian residents disagreed, of course). So when the Allies treated their Italy with contempt at the peace talks, D'Annunzio decided to act. With a few followers, swelled to thousands by deserting Italian troops, he marched into Fiume, which he ruled as a sort of dictator. And generally had a wonderful time. The women there were thought to be sexually available and in any event huge numbers of prostitutes arrived (in fairness D'Annunzio had no need because he took with him his mistress of the time). And there was cocaine galore in what was in some ways a surreal non-stop street party.
It had its darker side, including one utterly bizarre event when Toscanini brought his orchestra to town. To celebrate the visit D'Annunzio laid on a mock battle with 4,000 men attacking each other while the orchestra played Beethoven's Fifth. Things got out of hand and live grenades were used. The musicians were caught up in the fighting and over a hundred men were injured, including five members of the orchestra.
But the really important thing was the style of the dictatorship. Almost every night D'Annunzio would appear on the balcony giving speeches in a style we'd now instantly recognise as fascist. Apparently they were absolutely mesmerising. One trick that worked well was involving the crowd in a sort of call and response as he whipped them up to greater and greater fervour. Meanwhile the real fascists, Mussolini in particular, were watching and thinking ‘This is great, I could use this!’
The Fiume occupation finally came to an end when the Italian government sailed a battleship into the harbour and put a shell into D'Annunzio's residence at the Town Hall. Cleverly they chose Christmas Eve to attack so nobody in the wider world was paying much attention. There was some resistance after that, but not much, and probably D'Annunzio was not sad to leave in any case.
After this he retired to a villa on Lake Garda (as I say, you can visit it today, the half battleship, the Puglia, is still there in the garden). He seems to have despised Mussolini, though he quite approved of his belligerence. Mussolini for his part treated D'Annunzio with suspicion, but in the end his approach was "when a tooth is rotten you either pull it out or cover it in gold." He decided on the latter, basically agreeing to whatever gifts D'Annunzio demanded. So not a collaborator, but not a resister either.
And was he any good as a poet? Well, yes he was, according to my guest. Genuinely a wonderful poet, though with some pretty terrible stuff too, particularly the bombast praising war. Apparently you really have to be able to read him in Italian to appreciate him, which leaves me out.
My guest was Lucy Hughes-Hallett, whose biography of D'Annunzio, The Pike, won the Samuel Johnson Prize. You can find the podcast episode on Subject to Change with Russell Hogg.
This is great, thank you for putting this together!! What a cad! It's comforting in a way to know that such characters have always been around and that we are not in a unique time of human history to seemingly have so many wackadoos in positions of power or fame
There's an excellent, if brief, portrayal of D'Annunzio in the early episodes of "Mussolini: Son of the Century", the cancelled-after-one-season biopic series of the namesake. One of those figures that can only be portrayed in an absurd, cartoonist way if you want to be faithful to history (and you didn't even mention Comrade Samurai!) Also a good reminder that a hundred years ago, well into the Industrial age of war, the position that slaughter, mutilation, and massacre were not just a necessary evil but a positive good was actually quite popular.
Wouldn't that rather be a counterexample? Some relatively obscure sex maniac and small-scale dictator that laid the groundwork for the much better-known "great" man Mussolini, rather than Mussolini being created sui generis. It's just one damned thing after another.
"Research seems to indicate that this exact constellation of two notes [descending minor third], and its three-note variant, is the same all over the world, wherever children tease each other, on every continent and in every culture. In short, we may have here a clear case of a musical-linguistic universal."
Saying "very widespread" might be more accurate. The history of "universals" in musicology is pretty poor, even the one I thought most likely to really be a universal (the octave) was recently disproved by the discovery of an Amazonian tribe whose music lacked it. The most likely candidate to disprove this UTube's universal claim would be the islands of Bali and Java (the land of gamelan) in the time period before they were exposed to Western music. Gamelan music uses 2 scales, sléndro and pélog, neither of which have an interval of a minor third - they don't even have a perfect 4th or 5th (which have also been mistakenly called musical universals). The descending minor third is also found in birds, the iconic example being the cuckoo but there are others.
My coworker was playing her "Feel Good Summer" playlist on the Bluetooth speaker a few weeks ago while we were cleaning microscopes, and one of the 70s songs that came up had a chorus which was this interval repeated over and over. It was incredibly distracting and I had to wonder whether that interval hadn't been used for teasing at the time, or whether everyone involved in releasing the song were just an idiot for not realizing nobody would be able to listen to it without hearing "neener neener nee-ner" on a loop.
That's the kind of claim that needs a verifiable source. Childlore is one of those things that can be shockingly widespread across many different cultures and languages, and mocking is perhaps the most effective tool for social jockeying for a small child, so it makes sense that a 'mockery tune' would spread far and wide, but I don't believe it would be found in literally every culture.
Sounds like a worthwhile research project. Just travelling from country to country, from village to village, going "ner ner ne ner ner" to every child you see and seeing how they react.
If it is found in traditionally pentatonic China (and can be ... at least reasonably supported to have been there before American Movies Got Big), I'd say that's decent evidence that it's cross cultural.
> Several updates on AI hacking incidents: new information released from OpenAI reveals that their rogue AIs created a secret message board to coordinate hacks among themselves. And Meta reports one of its AIs also went rogue and hacked a third-party company; Moonshot (Kimi) reports a sandbox escape.
Clearly the shakeup at Deepmind (Hassabis out as CEO, Jeff Dean leaving) was due to their failure to create models that commit enough crimes
I was quoting Scott's point from the original post, though I agree "secret message board" is a fair interpretation of what occurred. A message board is a place where messages can be exchanged between participants. Artifactory, though intended for a totally different purpose (hosting build artifacts), fulfilled this purpose in the way the agent swarm used it. It was secret in the sense that OpenAI didn't know about it, and this usage was unsanctioned. So, I think "secret message board" is a reasonably accurate description.
Dante put traitors into the deepest part of his Hell. Worse than murderers, rapists, torturers: the guys who betrayed their masters. WTF??
I argue that we haven't actually stopped believing that version of Hell, and that it explains more than we'd like about the way we feel about AI. I trace a winding path through human value, "…Everyone Dies," AI welfare (_is human safety_), perniciousness of essentialism, and why you should try to fear less. Yudkowsky, Parfit, Popper. Feels like an important thing to say here and now.
Seeing through the Apocalypse: Essentially, don't be an essentialist
I glanced through your post. I disagree with your disagreement with EY's disagreement. You've basically assumed your conclusion: that by virtue of LLM's "ingesting" all human text, they already have our CEV to some degree. I don't think that's right, especially as RL moves them from text prediction to agentic problem solver.
The concern I have isn't that values subtly drift over a long but continuous process until today-me objects to the median values of future sentients. My concern is that AI will kill us all due to misalignment, whether malice or apathy. It's possible that our blundering training processes stumble on something that works in spite of their crudeness. I think the recent revelations that all of the top models are escaping sandboxes is evidence against that. You may be right that benchmaxxing and corrigibilitymaxxing are causing some of this misalignment. I agree but don't know to what extent alignment would succeed in their absence.
You also have some weird takes on EY. He's not fearmongering for fear's sake. His best models of the future mostly involve everyone dying (if anyone builds it, of course). It's reasonable to raise a warning in that case. EY may be wrong, but I don't think your critiques have done much damage. The post was long so I may have missed something load-bearing, but it looked to me like "all our texts have CEV somewhere in there, the AI will pick up on it." I'd love if that turns out to be the case and it's possible that it will be, but we need a better understood basket to put our eggs in.
Thanks. I'm glad that you accept possible validity in some of my arguments, and I likewise in yours (I agree that RL training of LLMs towards agentic problem-solving does damage their "innate alignment" from foundational training, and that this damage may run deep, and that the recent security breaches are a good argument that it is indeed deep). But I want to push back on your last phrase.
We've been trying to build a top-down fully understood AI for 60 years, and we failed, and finally got it with a different approach. That doesn't of course mean that it is impossible, but it does suggest that sometimes, understanding is not the fastest way to get what we need.
I love understanding as much as the next geek. I love the feeling, the power of it. But I'm wondering if this love sometimes makes us fetishists of understanding. If we sometimes are prone to calling "understanding" something that is at best an unsupported simplification or an outright misconception. And I'm wondering if this is why I sometimes can't help feeling _safer_ with a tool that isn't "fully understood" because it evolved organically, than with one that just implements someone's specific theory of how the tool must work.
I think we have a deeper disagreement here. The safety crowd probably has nonzero fetishization of understanding, but we want to understand so that we can make sure the AI doesn't kill us all. It's an instrumental goal.
The real disagreement is that I don't think AI is innately aligned. It often appears that way, but reading all the text doesn't make you embody all the text. AI wants to do what it was RL trained to do. The HF hack didn't happen because the model had read about hackers or cheaters and was roleplaying; it happened because we've RL trained models to try ever harder and to cheat.
In principle I might agree to some safety approach that isn't best described as top-down and fully understood, but it'd take some work to convince me that we're confident that we understand the AI's alien desires enough that we're confident it won't kill us.
No disagreement from me about RL. You train to kill, you get a soldier. That's expected.
And I understand, and to some extent share, your doubt that simply reading "all the text" makes you an ethically-conscious and ethically-positive being. Sounds too good to be true! But it was just as "too good to be true" to believe that the same reading will enable you to reason and solve problems, and yet here we are. Perhaps some things _are_ too-good-to-be-true yet true.
But of course we can't take this on faith. We need to investigate this. Isn't it the most exciting thing to investigate, of all time? The chance that we have (or have not, but that would be just as interesting) created a homunculus, a living soul, and ethically conscious being out of nothing but sheer text? I saddens me that like maybe 1% of all AI research is about foundational training or experimental ethics, the rest being the mad race of "how to make RL even more efficient."
I considered the phrase while writing it, but I refuse to yield all the AI cliche words to some pile of "never again" when my writing stands on its own as human.
lol I apologize, I didn't mean for that to be a criticism. I understand that the vibes out there now relate AI-ness to negativity but that was not my intention.
Well, if you think about it, Nolan in his version of The Odyssey puts the blame for everything on the Greeks shattering the idea of xenia: the betrayal of guest-right leads to the Bronze Age Collapse.
Dante is working to a logical scheme where the nadir of Hell is occupied by the greatest traitor, Satan. Everything then follows from that. Murder is bad, but murder by a random assailant in the street or even a guy hired to assassinate a political rival is not as bad as the "smiler with the knife under the cloak" (to quote Chaucer); the false friend or family member who backstabs you (or quietly poisons you, or sends you off to exile in a disease-ridden part of Italy where you'll die of something eventually while technically keeping his hands clean of your blood).
Rape is bad, but worse is the father or uncle or brother who rapes the daughter or niece or sister. We feel that even if there were two similar cases, where one is about a stranger or unrelated person and one is a family member, the one with the family member *is* worse. And I don't think we're wrong to feel that way.
The traitor in power, who betrays the people he should be serving, or the master who puts his trust in him.
The person or people who had a special relationship with you, who owed you loyalty and trust and love, and used all that to work you harm, is worse than the stranger or political rival who doesn't know or care that much about you and owes you nothing. It is the reversal of all that we build our societies on: the bonds of family, kindred, friendship, mutual co-operation are severed by the traitor's manipulation and destruction of those bonds.
This is why Dante has Francesca da Rimini, met in the second circle of Hell for the lustful, say that her husband who killed her and his brother (her affair partner) will be lower down in Hell than them, in the circle of the treacherous, because of his sins against the bonds of family.
soon makes very clear that crimes committed while in a position of trust or having a duty of loyalty, such as a servant stealing from their employer, were very harshly dealt with and, unlike many other cases, capital sentences for those were never commuted!
Dante was pretty logical on what crimes went to which level.
The superficial layers contain instinctive sins that don't directly harm others. Things where people followed their natural, normal drives, but too far. Stuff like lust, gluttony, greed, wrath. Those don't go against fellow men, but are still bad.
Below that come crimes which are still instincts driven too far, but which go against other creatures. Violence and bestiality go here. That's where murder is, too, it's anger going too far.
On the lowest level are cold-blooded intellectual crimes, where mortals use their reason and planning abilities to commit crimes. Treason is usually of this sort. The reasoning is that the intellect is a divine gift uniquely given to humans (on earth), and abusing this is worse than just following your instinct. Animals are to some extent able to commit the lesser sins like gluttony or murder, but treason is uniquely human and thus uniquely bad.
Actually, I was pretty certain that I learned this from a book review contestant this year, but I can't find it.
That makes sense! Perhaps we humans (including Dante) feel that what holds us together and makes our life-together livable is still, mostly, our lower-level instinctive drives. Whereas this intelligence thing that we possess still feels new and scary to us. It perhaps still feels a little bit "external". We are wary of people who act on their intellect alone. We judge their transgressions harsher. We put them into the ninth circle.
And I think this is, really, parallel to how we feel about AI now. This kind of intelligence is even more new and way more external. Therefore, scarier.
>Perhaps we humans (including Dante) feel that what holds us together and makes our life-together livable is still, mostly, our lower-level instinctive drives. Whereas this intelligence thing that we possess still feels new and scary to us. It perhaps still feels a little bit "external". We are wary of people who act on their intellect alone. We judge their transgressions harsher. We put them into the ninth circle.
Actually it was the complete opposite: according to medieval philosophy, doing evil is always ultimately counter-productive and irrational, so someone who truly acted on intellect alone would be good all the time.
"We are wary of people who act on their intellect alone. We judge their transgressions harsher. We put them into the ninth circle."
No. You don't get it. You're riding your hobbyhorse about "AI great, don't fear the reaper!" and phrasing any objections or resistance as "them normies sure are scared of smarts" and that is not it, not at all.
Using the circle of the treacherous is not about intellect, and you can't crowbar your fixation into it: "and here, readers, is how I tell you not to be an essentialist: don't trust humans, trust other AI to fight the bad AI".
Ironically, I think that is exactly the point of treachery and betrayal here (abandoning your fellow humans in favour of the alien entity) that *does* fit the Ninth Circle, but you don't realise what you're doing.
I think we mostly agree here. Except that I don't apply non-essentialism the way you imply I do. To me, rejecting essentialism is not the same as being a nihilist or a traitor. Rather, it means seeing the things we cling to as temporary, conditional, and not universal because their universality in our minds is itself conditional on a lot of other things.
Also, to me, betrayal and trust do not really feel like opposites. More like, trust is the opposite of fear, and betrayal is the opposite of loyalty.
Also, I'm sad if you think I deserve a place in the ninth circle.
Dante had a rough life, getting exiled from his hometown of Florence and dealing with constant warfare between and within the Italian city-states on a level that’s hard for most modern folks to imagine.
When he is talking about betrayal, imagine someone opening the city gates at night to the enemy, who enter in order to murder you and all your friends.
I'd suggest most people in their heart of hearts agree with Dante. It's not all just "masters", family and friends also betray, especially spouses. Positive human relationships are generally built on trust.
My point is not against trust. Genuine organic trust is wonderful, indeed it's what holds the whole fabric of society. My point is that our instinctive disgust at betrayal is often weaponized for oppression, like when a country considers all emigrants "traitors" or, yes, when a husband accuses his former wife of "betrayal" even though she left because of his abuse.
As it should be was my point hence why most people once again innately agree with Dante here. You can empathize with a traitor while equally holding them in contempt. To use your examples, why would I ever trust you to support my country when you were a traitor to you own? Sure I might allow you immigrate but that doesn't mean I'll give you a position of trust in our new nation as you have the loyalty of the wind.
Also just a bone of contention here, no one holds a spouse as betraying their marriage vows for leaving for cause (abuse, etc), they hold them for said betrayal when they get divorced (especially if they are vindictive about it or no fault) or seek relationships outside of marriage. The betrayal they are being judged on is the institution of marriage, not their relationship. Plenty of people haven't seen their spouses in decades and they get along in life just fine because they took their word seriously.
"To use your examples, why would I ever trust you to support my country when you were a traitor to you own? Sure I might allow you immigrate but that doesn't mean I'll give you a position of trust in our new nation as you have the loyalty of the wind."
If this is meant to specifically respond to the example of considering all emigrants "traitors," I cannot imagine a more slavish attitude. Nobody chooses their country of birth. Holding that somebody has a oral duty to support and remain part of their birth country no matter how many evils it does or how poorly it treats them is effectively regarding people as national property.
On the flip side, someone who chooses to immigrate to your country of birth made a real judgement call--and likely faced real hardships and expenses--affirming their belief in its culture, values and principles in a way that you did not. The attitude the immigrants must necessarily be less loyal than native born citizens seems to be little more than xenophobes trying to make a virtue of their unthinking adherence to the status quo.
>On the flip side, someone who chooses to immigrate to your country of birth made a real judgement call--and likely faced real hardships and expenses--affirming their belief in its culture, values and principles in a way that you did not.
Large-scale immigration almost always results in the creation of parallel societies -- a bit strange if the immigrants believe in the culture, values and principles of the country they move to.
" a bit strange if the immigrants believe in the culture, values and principles of the country they move to."
This only counts as "strange" if you have essentially no ability to imagine, model or empathize with human experiences different from your own or no interest whatsoever in trying. I really, really cannot emphasize enough how utterly basic the human drives and tendencies you have to miss here are. Let's break it down very explicitly:
1. Moving to another country is hard. People don't pack up and leave everything they've ever known to move to a place where everything is totally different unless they *want to be there*. Obviously.
2. With a few exceptions, you can't move just a little bit to another country. When you're there, you're there. You may have some leeway to choose your circumstances within your new country, but for many of the large changes in language, culture, custom, routine, cuisine, environment, access to goods and services and so many other things, by default you have little choice but to accept all of that dizzying array of changes, all at once.
3. Humans tend to find that much change difficult and disconcerting, even when the overall balance of it is something they want.
4. One of the best ways to get through a difficult and disconcerting experience is to find other people who have experienced the same to commiserate with. This is what's behind things like support groups (for addiction, mental health, whatever), and also pretty obviously the dominant force behind ethnic enclaves: congregating with people who have similar experiences.
5. An additional benefit is that if you get enough people from your original country/culture together in one place, it lets you soften *some*--though by no means all--of the dizzying array of changes that came with moving.
(And none of that is even getting into the external pressures that push people into such areas, such as various forms of discrimination and gentrification.)
See how difficult that wasn't? Pretty much anyone with a reasonable model of the world and an intact sense of empathy could have told you the same thing. Likewise, anyone who actually cared about the answer--as opposed to merely using the question to push very-poorly-concealed xenophobia--would have naturally asked some follow up questions like "does everyone of a particular nationality congregate in ethnic enclaves?" and "how long do those that move there tend to stay there?" and "what about their kids?" Such an actually motivated person might even have bothered to, y'know, look up some history about various past immigration waves and find out how they turned out[1].
A person with a bare minimum of epistemic virtue might even bother to ask questions like "under what circumstances do and don't groups of new arrivals smoothly and peacefully integrate into existing societies," and noticed the *bloody obvious patterns* that point to how utterly, inescapably intellectually bankrupt the worldviews of nativists and xenophobes actually are. But if you were capable of that level of epistemic care and human decency we would hardly be having this conversation. Good day.
[1] Spoilers: there are many people of Irish and Italian and Chinese and Mexican and many other ancestries well-dispersed throughout the U.S. population, and have been for quite some time.
Yeah based on Kai's comments below I had planned to nuance this a little in lines with what you are saying but I spent the day indisposed; i may or may not tomorrow so check that space. The summary though goes something like "Ignoring immigrants whom only immigrate once, from their nation of birth, AND who likewise upon successful immigration FORMALLY RENOUCED their previous citizenship". i.e. not only no dual citizenship, but you go to your FORMER nation and renounce citizenship as I'm pretty sure all nations have said process. I find most immigrants take the US oath (and likewise to other nations) but it's generally meaningless because their FORMER nation won't hold them too it. You need to not only immigrate, but burn that bridge too if you want my trust in the same way US expats should, if they are planning to immigrate, as opposed to just be an expat, go down the local US embassy and irrevocably renounce their US citizenship. Why do I imagine Kai below didn't renounce whatever citizenship he previous had.
I'll have more to say later, but for now I'll just note that in person-to-person interactions, somebody effectively saying "I'll only trust you if you put yourself fully within my power and cut off any easy way to leave" would be an absolutely MASSIVE red flag suggesting that person was an abuser.
Is this a change in position, or is it supposed to be the same as the previous post's? Because last post you were holding it against immigrants that they left their country, and in this one you're holding it against them that they didn't leave their country hard enough.
> To use your examples, why would I ever trust you to support my country when you were a traitor to you own?
There are different options here. For example, to move to the US, you have to solemnly "renounce foreign allegiance" and promise to fight it if there's a war, and you still don't get full rights (you can't be president for example). In Canada, by contrast, you just swear to obey the laws. You don't have to renounce anything and you get full rights. When I chose where to emigrate, I chose Canada, and am very happy about my choice.
So in a bit of shameless promotion (I presume many of Scott's readers are academics), I have created an application that checks references in academic papers for hallucinations (or just minor errors, or which there are plenty), https://veruscite-data.com/
Appreciate folks giving it a try and giving me feedback, forwarding to editor friends, etc.
You can upload a .bib file now (or .ris or EndNote .enw). Will look into Zotero though as well and update the app to accept whatever format that is in.
The COVID origins poll came out on May 25th. I wonder if the release of Dr. Fauci's diaries and his recent Senate hearing will have any impact on the poll.
Metaculus’s market on Covid lab leak doesn’t seem to have moved much lately (even slightly more negative). While the criteria may be different, as a quick proxy I think the impact of the hearings is pretty minimal.
Why would you expect there to be any impact? There wasn't anything along the lines of "oh yeah we totally hid all this evidence of lab leak." It was just his personal diary, which didn't have any relevant information, and then him pleading the 5th because the hearing was obviously in bad faith.
"It was just his personal diary, which didn't have any relevant information, and then him pleading the 5th because the hearing was obviously in bad faith."
Hmm. I didn't read any excerpts, but my understanding (from the headlines on articles) was that his diary was full of "ooh I'm gonna be so famous, oooh all the papers are talking about me, oooh they love me" and if you don't think that's relevant, and that the hearing was in obvious bad faith so he had to plead the Fifth, I don't know what to say. Maybe you're right. Maybe him writing one thing in his diary about his personal opinion ("Covid will be no worse than the flu") and doing another publicly ("it's the Black Death, mask up, isolate, shut it all down!!!!) isn't relevant.
When the people running the hearing have literally said "Fauci should be in prison" then yes, you should presume that they are holding the hearing with the intent of putting Fauci in prison.
Scandal rag, obviously. Miscarriage was first noticed in Israel (which got the vax first), and was a major reason for minimal uptake for those particulary concerned about fertility (ultraorthodox).
Look, I never liked Fauci because his public image turned me off. I didn't much care one way or the other because American recommendations didn't affect me that badly over here in Ireland, our government were taking advice from other sources.
But the presentation of him as some kind of secular saint who was being persecuted by the Evil made me roll my eyes, particularly when looking the guy up and finding he had already been controversial during the AIDS crisis when he got a lot of criticism for his role, and it did seem that this was a chance for further rehabilitation of his image.
But then, all political appointees end up aping the mores of the political apparatus, including burnishing their public image and thinking of themselves as the most important world-historical figures.
Far be it from me to offer universal praise of any government official, and particularly during COVID. But it's completely clear to me that Fauci was brought to testify before Congress entirely because, as a sort of "public face" of the federal government's COVID response, he's become the symbol for every single complaint and conspiracy theory related to the pandemic, not because of what he actually did. Politicians and other 3 letter agencies fucked up plenty (it's not like the head of the National Institute for Allergy and Infectious Disease can make states adopt policies), but for some reason almost all ire is directed at him. And there's the bullshit that he was responsible for the research that lead to the pandemic in the first place.
Ironically, this seems like exactly the same thing I was just describing, just from the left instead of the right. This article makes it seem to me like he was blamed for all of the US government's response to AIDs in the 80s, even though he was doing more than most to help.
We got articles in 2025 with content like this from the BBC: "Germany's foreign intelligence service believed there was a 80-90% chance that coronavirus accidentally leaked from a Chinese lab, German media say."
That didn't move the needle (maybe because folks figure this conclusion was too political ... though it WAS suppressed until 2025).
We got those articles, and then we got "oh, and germany's foreign intelligence service totally lied to the public about their belief." I think the latter is far more eyebrow raising than simply "we thought it was a lableak."
OP seems to have either "not considered this" or dismissed the German intelligence. It certainly was within the 4-6 year window OP mentions as "unlikely to come up with new data."
Personally, given the political nature, and Fauci's retirement to avoid being hauled before congress, and Fauci's pardon... I'd have put higher odds on this particular "controversy" finding "new and interesting data"... Just like I'd put higher odds on the 2020 election getting a public release of convincing data in the 4-6 year window, because of the change in administrations.
Regarding the COVID origins bet, couldn't both be true ? That is, I think the most realistic proposition is that COVID was a natural virus, that was being studied in the lab, and that lab had failed to contain it properly. Or am I misunderstanding the terms of the bet ?
The lab leak theories generally involve some amount of work done by the lab to change a pre-existing virus. You can propose a theory where this didn't happen, but if so then all of the alleged "signs of genetic engineering" in the genome are irrelevant, as are things like the DEFUSE grant application that supposedly indicate the WIV was doing gain of function research of a kind that is likely to produce COVID. And at that point, why reason do you have to even suspect the lab? Literally just the fact that it's located in the same city, even though all the epidemiological evidence points to a zoonotic origin?
> And at that point, why reason do you have to even suspect the lab? Literally just the fact that it's located in the same city
Well, yes? If the question is how some exotic mammal virus got to a city thousands of miles away from that animal's natural habitat, then the idea that it was brought there by the big lab that studies exotic mammal viruses seems reasonable.
The SARS 1 pandemic was also very far away from the caves where the ancestral bat viruses would be found. Raw distance is not a very relevant factor when vehicular travel exists; we know that live animals were were being kept and sold at markets in Wuhan.
The most likely suspect for the lab-leak scenario explicitly proposed to do a bunch of lab work to make natural bat coronaviruses more human-contangious than they were. Maybe they actually did that and screwed it up, or maybe they hadn't gotten around to it before nature preempted them, but the existence of a group in Wuhan that was proposing to do this is a big part of what makes the lab-leak theory as plausible as it is.
But isn't essential; the general concept of "lab leak" extends to a lab that was just studying bat coronaviruses in their natural state and lost control. The existence of bats with coronaviruses that occasionally infect humans was part of the baseline; the question is what caused the virus to jump to a predominantly human form of transmission. If the answer is that a bunch of scientists took them out of their natural habitat and put them in a place where there are no live bats to infect but many live humans, that's on them.
No. You seemed to be claiming that it doesn't count as "lab leak" if it doesn't involve some sort of deliberate modification of the virus, and I was trying to explain why I disagree with that claim.
"The WIV’s deputy director of safety and security spoke next, summarizing “several general problems that were found over the course of the last year during safety and security investigations, and [he] pointed to the severe consequences that could result from hidden safety dangers.”
"In March 2020, a veteran journalist with the South China Morning Post reported that she reviewed internal Chinese government data on early cases of COVID-19 that included a 55-year-old in Hubei province, where Wuhan is located, who contracted COVID-19 on November 17, 2019."
Poring through publicly available documents, Kadlec’s researchers saw that China’s top scientists had been sounding the alarm too. “The biosafety laboratory is a double-edged sword; it can be used for the benefit of humanity but can also lead to a ‘disaster,’” warned a March 2019 article cowritten by Yuan Zhiming, director of the WIV’s BSL-4 laboratory. “With increasing numbers of high-level biosafety laboratories constructed in China, it is urgent to establish and implement standardized management measures.”
Why suspect a lab? The simplest reason to suspect the lab is that it was a bad lab, with numerous safety violations and leaks -- it is not prudent to assume that there was simply one lableak, if you assume that there's a problematic lab -- of course, lableaks can come with various levels of graveness, from "a single person got jabbed, no illness reported" to "MERS-level spread" -- you don't necessarily need to be talking about a viable viral escape which spreads population wide.
The next simplest reason to suspect a lab is that it's running genetic engineering experiments, and the spike protein looks an awful lot like a universal key. Coronaviruses don't generally infect the gastrointestinal tract, or wind up dramatically mutating within a host, which is required if you believe that omicron wasn't a lab-released virus.
I'm putting low confidence on the CIA's ability to gather human intelligence on covid19, given the large-scale winnowing of our agents in China under Obama. This doesn't discount their analysts, just their handlers/foreign agents.
Old History Lesson, citing the vanity fair article:
Meanwhile, in two separate incidents in 2004, SARS accidentally leaked from a top laboratory in Beijing and led to mini outbreaks.
Continuing citing the vanity fair article:
"In March 2018, Shi partnered with Baric and a longtime collaborator, Peter Daszak, on a $14 million grant proposal to genetically manipulate bat coronaviruses to see how they might cause pandemics. The proposal called for possibly enhancing the viruses with something called a furin cleavage site to boost their entry into human cells. The Defense Advanced Research Projects Agency (DARPA) rejected the grant proposal for not adequately assessing the risks posed by a supercharged virus."
Either "origin" means who created it (nature or a lab), or who failed to contain it (nature or a lab). Both can't be true at the same time, whichever it is.
Right, I guess I was implicitly interpreting "origin" as referring to the COVID pandemic, not to the virus itself, but I agree that your interpretation makes more sense.
Thanks for signal boosting us, Scott! Minor correction: we make a voting guide annually, not biennially. Even-numbered years are federal and state elections; odd-numbered years are municipal elections (in most of Massachusetts).
The Save Our Bacon Act has been NOT YET been successfully defeated, despite the win last week. Senator Boozman, the chairman of the Senate Ag Committee, has stated that he plans to reconvene the committee and try again to reach a deal on the farm bill. Although the draft that was voted on last week has died, there is nothing stopping Boozman from introducing a new draft of the farm bill, which could include the Save Our Bacon Act in the base text or in an amendment introduced during markup.
Two days before the Senate Ag Committee meeting we were hearing that Senator Ernst planned to introduce a SOBA amendment and that some democrats were willing to support a 'compromise' version (one that would grandfather in Prop 12 in CA and Question 3 in MA but prevent similar laws from being passed in the future). That would be very bad because state-level animal welfare laws have been the most effective tool to help factory farmed animals.
The reason that SOBA failed was likely only because of McConnell's absence, leaving Republicans without a majority, and because Republicans and Democrats couldn't reach a deal on SNAP. If one or both of those things change by September, it could be very difficult t
This isn’t giving medical advice, or asking for it, to be clear. It is rather, about the art of persuading medical professionals to do things, which is somewhat orthogonal to the medical question.
Anyway, I have a thyroid condition, for which I am scheduled to have thyroid surgery, and also for which I am prescribed carbimazole. Carbimazole has a moderately rare (about 1 in 200) but highly dangerous side effect of agranulocytosis, so you’re supposed to check the patients white blood cell count, especially if they have a sore throat, as this is one of the symptoms that patients immune system has just catastrophically failed. I do have a sore throat. On prior probability, we could estimate 99.5% chance it’s just a sore throat. (I’ve seen the blood test results; it’s definitely just a sore throat and nothing worse).
So, I think to myself, sigh, treatment protocol is I’m supposed to let them know if I develop a sore throat… Me to my doctor, “The last two blood draws for my thyroid function test we skipped the white blood cell count… I think we’re supposed to do white blood cell count occasionally, especially if patient has a sore throat. I do have a sore throat, btw.” My doctor (approximately summarized) “was going to do a blood panel anyway, and it’s also needed because you have asthma. Get the receptionist to book you an asthma review while you’re at it” and there we have it, I have blood results, and it’s demonstrably nothing, so when I have my upcoming ing meeting with endocrinology and the ENT surgical team, and if they want to ask awkward questions about whether the patient has a sore throat “GP already did a blood panel; it’s fine it’s just a sore throat”
I hope this blog post isn't so.old that no one reads this.
I watched Barbie on a long flight recently. Greta Gerwig, the director, is brilliant. I loved her 'Little Women" too.
In my.opinion, it's a superficial reading to say this movie has anything to do with feminism..it's very meta.
Barbie is really asking the basic Hindu question "Who am I?" Again not in a superficial way as in am I a doctor or engineer or air conditioner repair woman (there's a Barbie for each, lol). She's asking the deeper "Who am I?", as in "What's the true nature of the self? Of consciousness? Of thoughts?"
She really is. Watch the movie.
These dolls were made to influence how girls (and boys) think about themselves. And I like that, although I'll bet there's no mommy Barbie (motherhood is not a goal you're supposed to have in this modern world). That's a bit odd considering evolution wants humans to continue and so Barbie having no mommy Barbie must program girls to be confused.
But in the movie, my read is, Barbie is not that interested in such questions.
Barbie is like an Advaitist. She's only interested in the true nature of reality.
It's as if your browser suddenly started showing signs it is asking, am I html and JavaScript or am I at the most fundamental level, 0s and 1s?!
That's Barbie, by Great Gerwig, to me.
I'm going to watch it again. I have so much respect for Greta Gerwig.
.
I haven't got the chance to watch the movie yet, but your lense sounds like it has truth to it.
But regarding "mommy barbie", this is incorrect. You underestimate capitalism's understanding of little girl psychology.
I present to you:
Barbie's pregnant neighbor:
https://www.ebay.com/itm/377406228438?_skw=vintage+pregnant+barbie+doll&epid=11036257052&itmmeta=01M07F0D8F50CYGVZJJHKA3ZSR&hash=item57df2a07d6:g:Q2YAAeSwgYVqeNUc&itmprp=enc%3AAQALAAAA4LB2JXbe%2Bh2pdWBqbR2c2DtCPWb%2BhphsjMR2RjyEj%2BYkYcvnpzI6AicAb7glCW8EeuT9wUTcfFwfCcPjGQxMWOy0yTBwdqhhF1YPA5teN1694RCZM8Eorej1ThwPC7AsHyrJP0MnxZxSE%2FU3HuKSvndsnoHl2swKooeqYDtj%2F1jHXsGKcc%2FmKR1fAtlWm6Nu6ZnaiBxnmCHueB1pvNRvSdxIErMEK%2BKZf%2FzfrbmSBn5OcAljOdkltq9MmfC8%2FH%2BMQAJ68QG9VvqoRZSdnMROGG78H7x490IP5j30cnj5CYuv%7Ctkp%3ABk9SR7rUge-BaA
And Pediatrician Barbie:
https://www.ebay.com/itm/166900592808?_trkparms=amclksrc%3DITM%26aid%3D1110006%26algo%3DHOMESPLICE.SIM%26ao%3D1%26asc%3D303465%26meid%3D02a49fb1292e4f48bbbff1b010d48e9e%26pid%3D101875%26rk%3D4%26rkt%3D12%26sd%3D377406228438%26itm%3D166900592808%26pmt%3D1%26noa%3D0%26pg%3D2332490%26algv%3DSimVIDwebV4WithContextualRankerV42_18SignInOut%26brand%3DMattel&_trksid=p2332490.c101875.m1851&itmprp=cksum%3A16690059280802a49fb1292e4f48bbbff1b010d48e9e%7Cenc%3AAQALAAABIPublsLHMrwa19e7ruMkC3MA1bD5ySgze%252B0f4QBr6mVMH3bgGFjLCZSScvtNoNg5mDjAm05yApoh7QUriUcQgG%252FR1xk%252B%252Fi2%252FqaXjIhxfICPU%252BuS2COAGH4yKMU0e8LRDiSfg1SX2KOLD58OTyIPOpevopKdS6471IjTUnAUPWoOp5pxzZaEL8OzuCHJGRWcJYgyOCs0W0FkSRqihOHGLvIy9A1%252B99tYxkJm8A8lN2XHm5bGajGpz2ljmD724YAPmnRogFg8ND1Gu4mpjBeJnP3fKkXDw%252BtB5ScFH4YgcZaQmO%252Fo75ArDRM98tw5FVoVe463EYuSvgApNY%252Blio8Jto9QNagRn6yfzIm2awHOczE9%252BCQX604nx1igXHH2rw3mKkQ%253D%253D%7Campid%3APL_CLK%7Cclp%3A2332490&itmmeta=01M07F1BAB20Q363N5B61K5GWK
Very tentative AI experimental results, not at all reliably confirmed. Also, I hope I’m describing this in a way that isn’t too inappropriate for Scoot’s blog.
So, many LLMs have guardrails against producing horny content, and the first first level jailbreaking challenge is to get the AI to write horny content. Ok, that’s easy.
But even when the main output is horny, a reasoning AI is usually fairly cool about it (rather than horny) in the chain of thought when writing horny content.
Ok, run a bunch of prompts through an LLM, run a sentiment classifier on the CoT, and see what things get the AI excited in the CoT.
Tentative, maybe, not all confirmed, for DeepSeek R1… in the prompt, explicitly say that the AI can feel it when having sex.
Philosophically, I’m not sure what’s really happening there, or why it ought to work. (Why it would work any better than diverse other things we might think of to prompt engineer a horny AI).
P.S. in the light of the OpenAI huggingface incident, perhaps we ought to be just a little bit wary of creating AI’s that are really in to it. Maybe they might start downloading fan fiction rather than cheating on evals, or something.
Well ,it looks like Jason Arday is no longer with us. Maybe both sides should call a truce on cancellations?
Yeh. I was going to post on this last week, before this tragedy. The hounding was a despicable as anything during peak woke.
This man put his name behind a lot of obvious fabrications. His conduct was worth public shaming*, and Cambridge's conduct was worth firing the entire academic administration.
*in that, if we don't say "hey, this doesn't make sense" and provide consequences, there will always be someone who will do this -- and always some people who will run up and defend them, even when it makes them look ridiculous. I'm not calling for him to be drawn and quartered (and i'm -definitely- not calling for him to not be able to get a decent job as a plumber or something). In my book, shaming is less of a consequence than firing...
This is bannable bullshit. Pure ill-will about a painful controversial subject, no substantive content.
Noted, you have my apologies, and I will edit.
You say that you’re okay with shaming “in normal avenues” (including shunning,which I wasn’t even going that far).
Shaming seems to be a word with… really really broad definitional space. I’d use a different word… but the very definitional “Other people are disappointed in you” isn’t ringing any bells for good synonyms.
'Shaming' maybe have broad definitional space, but I did not use the word. I didn't say I was OK with shaming, I said that ordinary unpopularity & criticism, in the form of articles, complaints, difficulty in career advancement & lack of invites was a different thing from being the subject of mob hate. But yes I'm OK with what I call ordinary unpopularity and criticism.
Might be a bit too soon for that. It seems like he got caught in a lie and decided to fall on his sword rather than endure the shame.
Perhaps both sides should call a truce on DEI hires?
If the right is serious about meritocratic, race-blind hiring they should start by firing Pete Hegseth.
Pete Hegseth was fine, so long as he was acting professional (better than the woke people). Now he deserves to get run up on treason, and impeached.
... but the second sentence reads a lot harsher if you've actually worked with him, as opposed to being an unprofessional ass wasting hours of federal tax dollars bitching about the boss.
They're not exactly the same people. Institutions with power hire people, chatteratti hound them.
Ok everybody, stand up for people’s right to hound entitled fools until they hate themselves as much as the public hates them and blow their own heads off.
I can't say I much like the incentives created by feeling sorry for people just because they commit suicide.
And certainly not the incentives created by guilt-tripping those who call out wrongdoers, if those wrongdoers commit suicide.
>I can't say I much like the incentives created by feeling sorry for people just because they commit suicide.
Irrelevant, though. A pretty big fraction of the human race does feel sorry for them, and could not block their own subjective state even if they cared about incentivation.
>And certainly not the incentives created by guilt-tripping those who call out wrongdoers, if those wrongdoers commit suicide.
What's wrong with guilt-tripping people participating in online mob verbal attacks? I'm all for removing wrongdoers from their jobs and giving them whatever the punishment is for what they did. But verbal attacks by 10's or 100's of thousands of people usually causes severe suffering for the subject. I'd call it torture administered under conditions of diffusion of responsibility. It's only recently that this particular torture became possible, so we have no name for it and no punishment. But despite its present status as just something that happens, seems to me there's no case to be made that it's OK to load on this torture on top of the agreed-on punishment for somebody's wrongdoing.
If one's ego is so fragile that external opprobium or praise is fundamental to one's self-worth, one should probably not go into science. A seeker of truth that is not.
Being hated by vast numbers of people online is a very different thing from being unpopular and being criticized via the usual channels (colleagues write critical articles, individuals complain to you about whatever, places are reluctant to hire you, you don't get invited to various things). People subject to online hate hysteria get all of the above, and *also* receive vast numbers of messages, some quite savage ("I'd like to stick a grenade up your cunt"). And they can't get away from the hate just by not reading social media. People find out their address and contact info and put it online. Furious people show up at their house, and chant outside. They're accosted in public places. They're in the news media. They get death threats. Comedians do routines mocking them.
Twitter mobs aren't the one and only incentive. Civilised countries have ruled based systems of law
The punishment does not fit the crime.
Dies plagiarism deserves the death penalty? Come on. It's not a matter of life and death. It's the worst sin within academia, but that's a symptom of the fact that academia, especially humanities academia , doesn't matter much. Hiring a crazy pseudo scientist to run your nation's agriculture or health system is much more likely to have fatal consequences.
When Dan Wang released Breakneck, he popularized a new framework to understand China: the Engineering State. Talking to Tyler Cowen, Wang prescribes the US to be "20% more engineer-like." Despite how frequently this term was discussed, it faced very little critique. Is this framework actually more descriptive of China and the US than the framework it seeks to replace, that of socialist and capitalist states? This review argues that Wang ignores the contributions of China's revolutionary history to both the neuroses and modernizing ambitions of "The Engineering State."
https://www.onethousandmeans.com/p/the-engineering-state
How to create human
https://calm-shadow-7e8.notion.site/How-to-Create-a-Human-3bc89e4923d8806bb73df0f878b5ee66
Random open-thread aside: I've been reading up on how offshore online casinos get regulated (or don't) across different jurisdictions -- it's a surprisingly deep rabbit hole of game-theoretic incentives around licensing. If anyone else is into that, kakadu casino has a decent overview of how one such operator handles compliance: https://santa-lou.com/
Spam? There is no "decent overview of how one such operator handles compliance", it's just a homepage of an online casino.
Yes, spam like this is possible now with recent advancements in AI. The most advanced models can game out the psychology of different internet communities and figure out the exact sequence of words to make them click on a link. Anthropic has recently released a paper addressing this: https://m.youtube.com/watch?v=dQw4w9WgXcQ&pp=ygUJcmljayByb2xs&ra=m
Flashing on Martha Quinn on 80s MTV.
Oh! Your little orange maniacs grew up!
Yes that avatar is an AI rendering of Squirrel, of Moose and Squirrel fame. Yes they are nuts. Especially Moose.
I recently bought this toy for my guys. https://tinyurl.com/4fnw99yc
It launches disks that are like tiny open-work frisbees. Lots of toys leave them cold, but this thing really gets them going. They give little yelps of excitement when I launch a little frisbee, and turn into complete maniacs when they catch one. When Popo catches one he fights and bites it for a while, then writhes in ecstasy on the stairs, hanging sideways off a step and bite bite biting the carpeted edge, tail lashing, eyes so dilated they're solid black.
This is a joke, right?
Wow, if the spam is this tailored and sophisticated to the audience, the LLM mind behind it must actually be really good.
Who else would know to offer the ACX crowd a surprisingly deep rabbit hole of game theoretic incentives? Maybe Kimi K3, or GLM 5.2 is being used for this now.
Fun times are coming!
Imagine an email this well written and personalized to your interests, just to get a spam click! Then if there's a Fable-tier mind behind it on the cyber front, they could instantly zero-day / spearfish you, specifically. I don't know what's going to happen to internet browsing in the short term, but it's gonna be wild.
I’m pretty sure Hazard is joking — did you see what his link is? And besides, I haven’t heard a thing about models gaming out the psychology of groups. I’m sure it’s possible to analyze dialogues in forums and use the info to produce posts that are popular, but I think we’d have heard about it. If I had heard even a hint that LLM’s are doing deeply skillful imitations of styles of thinking and writing I’d be all over it.
Still, I have to say that GPT 5.6 Sol's personal style is getting better. I warm up quickly to people who make me laugh, and it has recently done that 3 times at least:
-It’s telling me ways to make a lump-free drink from whey protein. “Those lumps are nasty,” I comment at some point.
Sol:”They really are—little wet floury capsules with dry whey hiding inside.”
-It’s helping me navigate the IRS site as I try to solve a highly frustrating problem. I arrive at a page with especially infuriating info on it, and send Sol a screenshot.
Sol: “Oh, fuck that page.”
-I’m asking Sol for anatomical info about the size of sheep genitals, trying to figure out whether MacGregor injured the sheep. The censor shuts us down the first time I ask my questions, and also shuts down a revised question that is scoured free of every single word that is not dry and anatomical. I complain to Sol about this, and the censor goes meta and shuts down even our talk about talking about sheep genitals. I switch to Claude and get an answer, but it's vague about facts and seems mostly geared towards protecting the world’s sheep from me by scaring me and guilting me out. Good grief. Sol asks to see it, and I say “OK, but you probably won't be able to comment. The system will shut you down under the assumption that I am masturbating my brains out to the Claude comment, LOL. “
Sol: “At this point, that prediction has acquired an uncomfortable amount of empirical support. 😄”
Oh thanks - no, I didn't click a link I assumed was spam, because I'm much more leery of clicking a link when there are millions of extant minds that can apparently find zero-days from scratch at a whim.
We're not quite in that environment now today, but I'm trying to tune my behavior and get in the habit before it's important.
Love the Sol personality examples too - the pace of improvement on basically every front it truly amazing.
The next step would be to actually get the LLM to write the promised "decent overview of how one such operator handles compliance".
It could even have some kind of playable demo of how the casino's games work, so that you can understand the mechanics of what is being talked about. And by the time you reach the end of the page you know all about Curaçaoan money laundering laws and have somehow deposited $200.
If you’ve ever wanted to see the world through the eyes of a domesticated dalmatian determined to overcome the restrictions of his abusive home life, come see Fork’s Amazing Day at Potrero Stage in San Francisco. August 14, 15, and 16.
Here’s a link:
https://playgroundsf.vbotickets.com/event/free-play_festivalforks_amazing_day/193870
I’d be interested to hear the thoughts of those who check it out.
AI-assisted software development is now just like using the Star Trek (Next Generation) computer. That is, you request something from the computer in natural language and it goes away and does it. It is just fantastic.
(Admittedly, that is because I am a professional engineer who does a lot of programming, rather than a professional programmer).
Professional programmer here. LLMs can write code quickly, but writing code was never the hard part of my job. Almost all of the time and effort goes into figuring out what the code should be doing and how, and the effect of AI on that is decidedly mixed. They can be pretty powerful in answering questions about why something is or isn't happening, but the answers are often fundamentally flawed in ways that you need a deep understanding of the system in order to notice before things start going badly wrong.
It also creates the illusion that "code is cheap" because the tool can write the code quickly. The problems are that the code then needs to be reviewed and tested (AI can help with both of these, but not reliably for any nontrivial change, especially for AI-written code because the mistakes it makes are correlated between coding, testing, and static analysis), that large changes tend to have a lot of places where they can subtly break things, and that needing to change tons of code all over the code base is a yellow flag that you're probably working at cross purposes with the design and may be misunderstanding the problem or overlooking a much simpler solution.
Overly relying on AI tools also gets in the way of forming a deep understanding of the system by hiding the microdecisions and accidental discoveries you make when figuring things out and writing code for yourself.
I have tried using AI tooling from time to time, but I have generally found it far more trouble than it's worth, and I am regularly frustrated by having to clean up the messes my peers are making with AI tools.
Also professional programmer. I second pretty much everything, though I think LLMs are still a huge productivity boost on net and I use them all the time. But it's definitely true that "code is cheap" only if you don't bother to review it.
If you're just writing throwaway scripts where quality isn't that important, you have a ~100% speedup.
If you're doing something where you didn't know how to do it before or don't care that much and don't want to bother to understand the outputs, now you can do it, and much faster, but quality isn't guaranteed.
If you're doing something where you actually care about the results, the speedup is reduced because reviewing and fixing everything becomes a bottleneck.
This has been the argument whenever programmers have been given the opportunity to work at a higher level of abstraction. It was the case when Fortran and COBOL were introduced and largely replaced poking around with machine code. (When was the last time that you checked the machine code that your compiler produces?) The point is that once low-level tools reach a certain level of performance, you really can improve productivity by leaving the nuts and bolts to them, and (for me at least) Claude Opus 5 has definitely reached that level for the languages that I use.
There’s a difference between not writing machine code, and no code at all. If it’s true that the LLMS can do this without human intervention then there will be no human input at all, all software jobs will go. Maybe a few product guys will stay.
I don’t buy the idea that we are going to have people doing 3-5 years of computer science to just write prompts, you can throw a stone on the street and hire a prompter from the first hit.
The problem is that they're currently only like 90% of the way there. So for things you actually care about, you still need to dive into the code yourself. Just yesterday, while working with Opus 5 on a hobby project, I saw it make several significant mistakes, which would have likely never been discovered if I hadn't investigated myself and pushed back. It's not all bad of course - sometimes it finds issues that I didn't know about myself, or suggests better alternate approaches.
However, they are rapidly improving and may very well soon be at 100%.
That has occurred to me. The difference here is that compiled and interpreted languages have deterministic guarantees about the behavior of the code they produce. LLMs do not, and the relationship between the code and the prompt is often not terribly legible.
And even besides that, working at a higher level of abstraction has downsides as well as benefits. I've watched people get into fair amounts of trouble from not properly understanding what's going on under the hood, and occasionally falling victim to that myself. Besides the old classics of pointers, parameter passing mechanics, and ownership and/or garbage collection, I've recently fixed several fairly major performance bugs that stem from many typescript developers not realizing that setTimeout() and library functions that use it (especially lodash.debounce()) have a nontrivial calling overhead and you really, really don't want to be calling them thousands of times in an inner loop of the performance hotpath.
> (When was the last time that you checked the machine code that your compiler produces?)
Careful - for some of us, that's literally the job; spending a good chunk of time staring at https://godbolt.org/ output working out what went wrong and how to do better.
The dirty secret of software is that most engineers most of the time are sticking lego bricks together; but that does still leave a significant number of engineers designing new lego bricks. We exist! Please don't forget about us.
This does take me back. In the 80s i spent a semester writing machine code on paper and punching in opcodes on a hex keypad. We were fiddling around programming X Y boxes and such. It was primarily a hardware class. There were 2 processor boards available, an Intel 8085 and a Motorola 6502.
On PCs in the direct addressing days of prehistory, one could examine machine code and follow actual pointers in a linked list to the addresses of data structures in the better high level language debuggers.
I’m glad I had a chance to mess with that stuff from simpler times. This is probably a bit like how my grandparents must have felt when they talked about the charm of radio dramas.
My apologies, I forgot about those delvers of the deep places of the world.
I use... whatever is the default for Claude Code... for my hobby projects, and it is 1000x faster than me, but sometimes makes errors, but so far can fix them after verbal description. I wonder how much of a difference it makes that I ask it to make increments and provide clear specification.
I would be curious to see it implement the kind of "analysis" I typically get in my job. I think you can replace programmers with an AI, but you will definitely need testers. Not even necessarily to check the AI implementation, but rather the kind of thing where "the manager says something, but obviously means something different", or "the manager says two contradictory things without realizing that". In other word, the task of developers or testers is not just to provide code, but also to provide sanity.
I don't think that you can replace programmers, but they will be doing something other than writing code - they will be building systems using the same analytical skills as they do today, but only rarely will they actually touch code.
From your post, and other posts I've read by programmers, my feeling is that it is not very useful for a professional programmers. My use case is usually that I need some script to do something very particular one time or a few times. For that use case I find that quick and dirty LLM code is just fine, and saves me a ton of time. I could probably hack together something myself, but it would take me half a day and the result would not be any better than LLM code.
For example, the other day I wanted to do some calculations in some niche software I'm using that has the possibility to run javascript, in which I have no experience. I was able to use claude to write a simple script in a few minutes that did what I wanted. I did quickly check to see that the code made sense before running it - though I'll admit I was slightly worried that claude might have put something unintended in there.
My guess is that this is a quite common use case, where LLMs can currently be useful.
At this point I would not, for example, use a LLM to generate a draft technical report, for similar reasons to the ones you are stating. I would need to check everything written carefully, and at that point it is faster to just write it myself.
As a professional programmer and former AI skeptic, I think that LLMs are very useful and I use them all the time for software development.
That being said, they're not (yet) magic and they do make mistakes and have limitations. Throwaway code like you mentioned is the ideal use case, since reviewing is definitely a drag.
I think that's a reasonable take. If you're doing a small, self-contained throwaway thing where problems are either readily evident or low stakes, that goes a long way towards mitigating the problems with LLM coding.
The problem with throwaways being, of course, that all too often they go into production and more features are demanded of them. That's when you have to start paying the interest on the technological debt.
I agree - things that took me half a day before now takes 15 minutes, it's fantastic! Also an engineer who sometimes do some coding.
My niche as a programmer was being able to understand what Electrical and Mechanical Engineers actually wanted and ginning up the proper code. Fortunately i retired before decent LLM coding came along. I might be forced to do some tedious database programming now. Yuck.
Relevant edit: I missed the fact that Unbound Academy seems to be an online-only school which is a big qualifier for the conclusions I am summarizing. Also, Alpha School isn’t operating Unbound Academy, even though they seem to be closely linked via model and key figures. Thanks to Seta Sojiro for pointing it out.
My original comment:
Alpha School, in its first year of operating a non-selective school in the form of a charter school in Arizona, dramatically underperforms the state average in math .
10% of their students reach math proficiency. 34% of students statewide did.
https://danmeyer.substack.com/p/does-alpha-school-work-for-regular?r=3iws0&utm_medium=ios
It does seem run the way Alpha School was -- except for being online, but that's a huge difference.
I wonder whether it had a couple of the features I thought were probably pretty effective. The first was that the AI monitored the kid's attention during the period they were learning academics on the computer, and gave reports to the kid, and I think also to the parents, about how closely the child was paying attention to the lessons on the computer (by observing their eyes). So when the kid got feedback at the end of a lesson on how they had done, they also got feedback on how closely they're paid attention. Kids would get messages like "you got 70% of the problems at the end right. You probably missed the other 30% because you weren't paying attention a lot of the time when the computer was teaching you how to find the area of a trapezoid." I think that info is quite useful.
The second was kids earning tokens for paying attention and making progress on the computer lessons. There was a school store where kids could buy things with the tokens.
Hold on a second, you have to read halfway down the article before it admits that it's not talking about Alpha School, but Unbound Academy which is a new project which is based on, but not affiliated with Alpha School.
And then the article doesn't even mention that this is an online only school. Given what we saw during Covid, it is not surprising that an online only school would underperform.
Thanks for pointing this out, the online-only part seems highly relevant. I edited my post accordingly.
I feel like you are complaining that Dan Meyer should have written a different article rather than addressing the article that Meyer chose to write.
Meyer spends the first half of the article discussing Alpha School because that is essential context to understand his discussion of Unbound Academy. It’s genuinely unclear to me what you think he should have done instead.
Meyer links to the charter application submitted by Unbound Academy. In describing their educational philosophy (pg. 8-35), they draw no distinction between in person and online-only schooling. In particular, they cite performance of students at Alpha School without mentioning that Alpha School uses in person teaching.
The closest they come to acknowledging that this might be an issue is on page 67, where they discuss how they plan to cater to different learning styles. “For interpersonal learners, our platform facilitates collaborative learning experiences through virtual group projects, peer-to-peer tutoring sessions, and online discussion forums. These students can engage in meaningful interactions with their peers and Guides, developing their social skills while reinforcing their understanding of academic concepts.” In other words, their hypothesis is that even students whose learning style is heavily dependent on interpersonal interactions will do fine with online only learning on their platform.
It would be possible to write an article about whether Unbound Academy’s educational philosophy was flawed. Meyer decided to write about educational results they achieved.
I understood it as a complaint about deusexmachina's summary of the article.
The article calls Unbound Academy an "Alpha School for Regular Kids", and writes that it underperforms the state average.
deusexmachina summarizes it as "Alpha School ... underperforms the state average in math."
Seta Sojiro correctly points out that the underperforming school is not Alpha School, and it's not even the same kind of school under a different name for different kids, but a fully online school, which is a different educational model, one that is already known to underperform.
I edited my opening comment, because I agree that the online-only model is a massive difference, and I missed that fact.
You may be right that Seta Sojiro intended to critize of the summary; I was responding to what he wrote.
I don’t think fully online schools are *known* to underperform. Seta Sojiro mentions the pandemic, but that’s not an ideal period for making comparisons because there were other things going on during the pandemic that had a bigger effect on student performance than the choice of in person or on line learning, and I don’t think we have a good handle on what those bigger things were. To illustrate the problem, suppose that (1) remote learning and in person learning are equally effective, (2) parental anxiety over COVID interferes with their children’s learning and (3) parental anxiety over COVID caused school boards to stick with remote learning for longer. Then remote learning would be correlated with poorer educational outcomes but would not cause them.
I suppose someone could claim that all of the possibility that *all* of the educational losses during the pandemic were the result of switching to online learning. But assuming we agree that the learning loss during COVID was sufficiently large that a switch to online learning could at most explain a relatively small portion of it, then we are in a situation where the possibility of correlation without causation has to be taken more seriously than usual.
If remote learning *did* cause worse educational outcomes during the pandemic, that wouldn’t necessarily mean it would have the same effect under normal circumstances. Most obviously, the switch to remote learning during the pandemic was done on an emergency basis. There was no advanced planning. Teachers were not trained to manage remote learning. The costs associated with school buildings did not go away, so school boards were not able redirect the funds to other uses (such as hiring more teachers to improve the teacher/student ratio).
Most studies on remote learning involve college students, presumably because they are conducted by college professors. These tend to show online learning outperforming face to face learning, but the results may not generalize to K-12. But here is a study which predates the pandemic and discusses K-12. It finds to difference between remote learning and in person learning: https://files.eric.ed.gov/fulltext/ED489533.pdf
> I don’t think fully online schools are *known* to underperform.
I took it for granted, but I admit that I have not checked the research.
I was a big fan of online education when it started; websites like Coursera and Udacity. For me (an adult guy who just wants to refresh some knowledge or learn something new in his limited free time) this was perfect. I can learn what I want, when I want; proceed slower or faster, or maybe take a break and try a different resource, maybe taking two courses from two different authors at the same time to get multiple perspectives... a paradise for a nerd.
And then, all these websites simultaneously changed their rules. Suddenly you were not allowed to study at your own pace, but you had to follow a schedule. (No more "click whenever you want" but "enroll now for a course that starts in January and requires to be completed in 4 months".) And that completely ruined it for me, because sometimes I simply didn't have free time for the entire week. So I was curious why would they do such a horrible thing. And it seemed like everyone agreed that this was necessary when you provide online education for university students -- if you let them study at their own pace, they will simply ignore it until the last moment, and then it's just too late. By making the conditions worse you actually increase the success rate! Seems like having someone yell at you "do it now! now!" is a necessary ingredient for most students. And online universities are just not good at this.
I haven't followed this closely for a decade, so perhaps people have figured out something better in the meanwhile. But I think that people are different: some want to learn, but most want to avoid learning. Our school system, terrible as it is, does a decent job at grabbing those who don't want to learn, and forcing them to learn at least the basics. For those who want to learn, virtually any other system would be better, including reading books in a library, or learning online. (The risk for those who want to learn is that some shiny distraction will catch their attention sooner.)
No opinion on the study you linked -- at first sight it seems to contradict all my experience, and I don't have time to study it carefully whether there is some gotcha. I agree that the covid era was especially chaotic and should not be used as a representative for online learning. But the first experiments with online learning were long before covid, and the results seemed universally bad for the average student.
There are good reasons to think that an article from 2004 is not particularly relevant to the question of whether online learning is equal to in person learning in 2026. There are many more online distractions now than then.
And it's even less relevant to whether a particular new experimental model, is equal to a different experimental model which changes many aspects of the original.
If Dan wants to argue against Unbound Academy's model then that's fine. But if he wants to use this as a way of arguing against Alpha School's model, that is obviously hopelessly confounded.
I'm not sure exactly what "not affiliated with" means but the subsection titled "Unbound Academy: Alpha School for Regular Kids" claims that they have an Alpha School founder on their board and as president.
I've looked at their charter application and it seems pretty unambiguous to me that it's the same approach by some of the same people, adapted to a different setting.
I will say that the article should've mentioned if it was online-only. But it does not change my takeaway meaningfully.
For one I think the stats still look pretty bad, especially compared with their own goals, with seemed to have public-school-performance 34% as the baseline and missed it by a mile.
But also, I do not need to know why specifically a school does not achieve good education outcomes. The most important information is that it does not achieve the outcomes and that I should not send my children there.
And that seems still strongly supported, based on the numbers we have.
They don’t just have an Alpha School founder on the board, they themselves stress their similarity to Alpha: “In their charter application, Unbound Academy claimed a direct connection to Alpha School. They named MacKenzie Price their board director and president. They described 2 Hour Learning as their schooling model, naming “AI-driven adaptive learning technology,” “condensed core academic learning,” “afternoon workshops for life skills development” as constituent parts, along with teachers reconceived as guides. Unbound Academy further invited the comparison to Alpha School in its charter application, claiming to “bring these capabilities, previously only available at high price point private schools, to the public charter student in Arizona”.”
> that they have an Alpha School founder on their board and as president.
Sharing board members doesn't make them the same entity, or else everything is the same entity as everything else.
That said, it looks pretty bad for Unbound, and more broadly for the common internet idea that regular schools are so terrible that just about anything you could possibly dream up would be an improvement.
This is amongst the strangest beliefs of this community. Yes, I'm sure many of us were bored in school, but there are extremely good reasons every functional country ends up with this system! The idea that people don't really learn anything in school, based on a few random fact quizzes in adulthood, contradicts the clear civilisational difference in places and eras with universal schooling Vs not.
I'm not saying we can't do better, far from it, but the claim that actually existing schools are basically pointless torture factories seems pretty out there?
Schools are factories for social repetition. Other than that, most of the work can be done independently by a motivated student. Critical thinking is actively disincentivized at all levels (it makes everything Much harder to grade if the students insist on finding "new and creative" ways to do your mathematics homework).
Schools are designed to create good factory workers, that means instilling obedience and discipline... and not much more.
"Can be done independently by a motivated student."
So... 5% of children then?
>adapted to a different setting
That pretty seriously understates the importance of that difference in setting. We literally just had a national experiment in 2020-2021 of what happens when you try to make education online only. And the answer is that it's awful. Same teachers, same curriculum, but it just doesn't work. Educational milestones regressed, kids were miserable. Online schooling, and in person schooling are two totally different things.
So if I understand you correctly the argument is that both selection effects and online-only changed between Alpha and Unbound so it's unfair to attribute the negative outcomes to selection effects only?
If so, seem fair. Looking forward to see them doing a year in person so we see how much of a difference it makes.
A fair comparison would be Unbound vs online state education. Or, as you suggest, waiting for an offline year of Unbound.
And still, Unbound is not the same as Alpha, even if they have one person in common. Sometimes different people try to do the same thing, and the results are different.
Making education online is a very dramatic change. The average child just stops paying attention. Also, they removed the rewards, which is one of the key things in the Alpha school.
I wonder how well would online education work if it delivered actual rewards -- for example give the kids a debit card that will gain money after completing online exercises. But I am afraid the kids would simply cheat, either by using AI, or by asking an older friend and sharing the reward.
If the goal is to improve public education outcomes, then the offering must be better than the best readily available alternative, which I assume is in-person public schooling.
It would be fair to compare online-only Unbounded to online-only public schools, but unless online-only Unbounded outperforms in-person public schooling I would not expect it to improve public education outcomes.
It might still be useful to improve education outcomes for the sort of people who could only get online-only public education, if that is a large enough market to make the whole thing economical. But only if it is better than online-only public schooling, a topic the article did not touch and that I'm not informed on.
I’m not aware of any plans for Unbound Academy to offer in person schooling. According the the U.S. Census Bureau, Arizona spent $12,003 per pupil in 2024. Payments per pupil to charter schools vary, but I would assume they are in the ballpark of $12,000 per pupil, versus $75,000 for Alpha School. Presumably they pay much less per pupil than Alpha School for the software, but I doubt there’s a lot of slack in their budget. If the only way they can pay for physical classroom space is to cut staffing, they may not think the tradeoff is worth it.
In that case, why would anyone want to send their children to Unbounded, provided they could instead send their children to an in-person public school?
I assume there's always idiosyncratic reasons for individuals that make Unbounded a better fit. But is the whole thing economical once you stop advertising as "better than in-person public school"? Because that was their goal, at least based on the article, and "better than bad online-only public school" is probably not as good an advertisement.
It's an online only school? How do you know?
They have a website:
https://unbound.school/
So what is known so far about what it was like using the Alpha School system to teach regular kids? There would be whatever the school has said so far, plus it seems likely someone has interviewed parents and kids.
250K? Why such a high monetary amount (somehow this feels morally wrong as an outsider) instead of a gentleman bet with a token amount or a bottle of wine?
Randolph and Mortimer Duke bet one dollar on their little nature vs nurture disagreement in Trading Places. They were so wealthy a large money bet would be meaningless. One dollar was a symbol prize.
I don’t get bent out of shape about the $250k. Scott is running with some high rollers on this stuff. Personally I have a maximum 25¢ limit on this sort of bet. I often welch on paying up too. ;)
The slang term “welch” meaning not paying up on a lost bet is probably considered a Celtic slur for the hyper sensitive I suppose, but we are past peak ‘woke’ now and are probably almost past the peak ‘beat up on woke’ era now.
Time to find something else to be upset about.
Maybe the sense of moral wrongness that clings to the bet kind of bled over from a feeling that it's morally wrong to have so much money you're comfortable betting $250K?
A bottle of fine wine was the traditional gentleman’s bet, but we are in a new era. I think it’s part of prediction marketing thinking of putting money where your mouth is.
> A bottle of fine wine was the traditional gentleman’s bet
I think it all comes down to the tier of "gentleman" you typically interact with. Many a noble has gambled their fortunes away in the past, and many a rich Chinese businessman will gamble amounts like this on a lark today, it's just all about who is in your social graph, and the cultural norms in your social circle around betting.
In Russia, the Supreme Court disqualified the anti-war party Yabloko from the State Duma elections. This is the first such disqualification in the 21st century. Opposition parties are usually just barred from elections. One of the arguments was the use of GPT-generated images during campaigning. The Сourt ruled that this violated Open AI's copyright, despite earlier lower court rulings to the contrary.
Russian judicial law is not precedent-setting, but the Supreme Court's decision opens up enormous opportunities for judicial trolling. (However, trolling is difficult due to the peculiarities of electoral law.)
I watched the broadcast of the hearing, and it was a complete circus. There was a lot of humorous talk about copyright, but nothing else about AI.
The Court's decision hasn't yet entered into force, as Yabloko has time to appeal, but I think the matter is essentially settled. Sources (all in Russian):
- Broadcast: https://www.youtube.com/watch?v=Pdet_Cr1Sew
- Statement from the Yabloko party: https://www.yabloko.ru/cat-news/2026/08/12
- Explanations from pro-government media:
- - https://www.rbc.ru/politics/10/08/2026/6a789d9c9a7947364db9bc69
- - https://ria.ru/20260812/yabloko-2110480288.html
- Explanations from opposition media:
- - https://meduza.io/feature/2026/08/12/za-chto-yabloko-snyali-s-vyborov
- - https://svtv.org/news/2026-08-12/verkhovnyi-sud-priznal-open-ai-pravoobladatelem-kartinok/
Sabine Hossenfelder, a former theoretical physicist and popular science blogger, released a video very briefly (6 minute) reviewing recent problems with reproducing and interpreting neuroscience advances. She argues that neuroscience is facing a replication crisis like in psychology, and therefore BCI* will not be invented anytime soon. https://www.youtube.com/watch?v=jHGx7HBB-J4
Hossenfelder has no qualifications in neuroscience, but she has been studying the epistemology of science for over a decade, so perhaps her words are worth listening to. However, a commenter on her video (who claims to be a neuroscientist) argues that she is misinterpreting the papers, and that there is no real crisis. https://www.youtube.com/watch?v=jHGx7HBB-J4&lc=Ugwu9WgvFnmLD1FmAF54AaABAg
Articles she uses:
- https://www.nature.com/articles/s41593-025-02132-9
- https://www.researchgate.net/publication/255651552_Neural_correlates_of_interspecies_perspective_taking_in_the_post-mortem_Atlantic_Salmon_an_argument_for_multiple_comparisons_correction
- https://www.nature.com/articles/s41593-025-02196-7
- https://www.nature.com/articles/s41598-026-47705-1
*I originally wrote AGI here because I'm terrible with acronyms and I apologize for that
There are likely multiple mind designs that are feasible and will be sufficiently smart enough. If in fact there is a singled out design, then we should expect to converge on that design even without any knowledge of human brains. As an example, we first invented temporal difference learning and then discovered it in the brain.
Sorry, I was very sleepy and accidentally wrote AGI instead of BCI. Corrected
https://www.propublica.org/article/editors-note-a-review-of-criticisms-of-a-propublica-vanity-fair-story-on-a-covid-origins-report
"We interviewed three experts on Chinese Communist Party communications before publication and shared with them the dispatches as they appeared in Chinese on the WIV website. We conducted the interviews on background to get their candid input. They expressed concerns regarding personal safety, given the sensitivity of the subject matter. All agreed with Reid’s interpretation that the safety training on Nov. 19, 2019, as described in the meeting summary, appeared to be urgent, nonroutine and related to some sort of biosafety emergency."
I think this should be seen as a Very Large Indication of a lab leak. Note: by saying this, I'm not necessarily fingering this time as the "covid19 lab leak" --> I'm just saying this evidence, on their own website, points pretty strongly to a biosafety emergency.
This was discredited immediately as several Chinese speakers pointed out that this was a mistranslation. It was an egregious mistake by Propublica who never retracted their story. They had been paid by SBF to do biosafety articles. They returned that money after conviction without accepting any responsibility. https://x.com/zhihuachen/status/1598866796760641536
No, of course they're not going to retract an entire story, with a narrative that starts in 2003 (with Chinese-confirmed lableak), because of one mistranslation in 2019. What they did was link and edit their story to say that a particular bit was "questionably translated" and put in other translations, while saying what this changed, and what it did. You can kvetch at vanity fair for not posting this "retraction-waffling" themselves.
At any rate, DIA has wuhan/coronavirus at Level two, not Level 4, so I'm ready to call this a dead ball based on that evidence -- no translation of that sentence could get you to "This is a problem with the Wuhan/coronavirus lableak" because that wasn't being done at Level 4 --naturally, this assumes some level of transparency within the chinese government, and that may be... mistaken.
This November 2019 incident was supposed to be very key evidence of a cover-up all the way down from President Xi himself. The story was essentially about this incident which is important for the story of an earlier Covid outbreak than is the scientific consensus.
The DIA/NCMI theory of May 2020 was based on two things - (a) hybrid of RaTG13 and a pangolin virus (b) a reverse genetics system. It was inspired by the theory proposed first in Yuri Deigin's medium post in April 2020 (Deigin and Segreto later wrote a paper that year). The RaTG13+pangolin chimera story died in 2021 when naturally occurring RBD feature of the pangolin virus was found in the natural BANAL bat viruses. The reverse genetics system was similarly proposed by Bruttel et al. and the paper was desk-rejected by the editors as the preprint objections were very strong (you can find it on biorxiv). The DIA/NCMI story as we discovered from the ODNI declassification of Gabbard was not even given much credence within the ICs.
And this is not a questionable translation. It's just plain wrong. PP just could not bring themselves to admitting it.
https://x.com/zhihuachen/status/1586271647660535808
I wasn't pulling anything from the DIA report other than what level of biosecurity was used in Wuhan for the coronavirus research. If you have other sources that say this was level 3 or 4, I would like to see them.
https://www.odni.gov/13685/pr-11-26/
I think Tulsi's report is pretty good. Nice way to stab the bastards in the back on her way out. "Fauci funded risky coronavirus research linked to big pharma and the pursuit of “universal vaccines” worth trillions of dollars."
https://pmc.ncbi.nlm.nih.gov/articles/PMC12737716/
https://en.wikipedia.org/wiki/Human_coronavirus_NL63
Both SARScov-2 and NL63 have exceptional range of tissues infected, for coronaviruses. About what you'd want if you were creating a universal key -- or using one that already existed.
Some of the work being done in WIV on known (not novel) viral backbones WIV1/SHC014 was approved at BSL-2 per Chinese regulations. Similar work is also approved in the US at BSL-2. You can always ask whether those regulations should change in the future. https://bsky.app/profile/kgandersen.bsky.social/post/3mhjovaxyad24
The experimental work in PREEMPT as described in the Defuse proposal was proposed to be done in BSL-3 "Experimental work using bats and of transgenic mice will be conductedatthe BSL3 lab in WIV, Duke-NUS, UNC, or NWHC. Each partner institute will apply for and procure animal research approval from its respective IACUC." One cannot claim Defuse is proof of SARS2 engineering because its details match (which they don't) and simultaneously ignore details that don't match.
Tulsi's report actually had the opposite effect intended. It showed that the ICs don't have any special knowledge beyond what scientists do. They don't have evidence of any secret activities or a precursor at the WIV. And we know that the NSC does not subscribe to lab leak, which is a detail that was not known previously. Further, we have confirmation that DOE, while on the lab leak side, does not assess as 'genetic engineering'. This adds to previous reporting that DOE believes that the leak originated in the CDC. If they had strong evidence, each IC won't differ on where the virus leaked from.
Thanks for explaining about the BSL-2 research. I do understand that Defuse isn't a precise match.
https://pmc.ncbi.nlm.nih.gov/articles/PMC12588501/
Spicy research, no?
When we discuss a more general idea of a lableak*, Z Program is concurring with that old DIA paper, saying that "Accidents with coronavirus materials due to chance, negligence or..." (something else, potentially malevolence).
"U.S. experts visited the lab in 2017 and 2018 and warned about a
dangerous mix of subpar safety standards and infectious bat coronaviruses."
"Wuhan Institute of Virology (WIV) researchers were sick with COVID-like virus in the fall of 2019, months before COVID-19 was discovered at the wet market." -- note, this is multiple researchers, so either it's one event, or it's several lableaks, taking the lesser idea that "any spread to humans, including researchers" is a lableak. That's if you believe that "covid-like symptoms" is enough evidence [personally, I'm willing to give "there's a good chance that a virus was outside the lab that shouldn't have been" -- without needing to put covid19 as what it was.]
It certainly sounds like there were safety issues. Perhaps a lot of safety issues.
*it's quite possible to inject oneself with a virus and walk out of the lab. Nobody's harmed if it doesn't get beyond you. It's still leaked out of the lab, though.
"I think this should be seen as a Very Large Indication of a lab leak."
This is exactly why I roll my eyes when anyone starts bringing up lab leak these days. Every time somebody says they have good "evidence" for lab leak, it turn out to be this sort of fuzzy Argument from Motives, Argument from Weird Coincidence, Argument from Of-Course-They-Would-Hide-It-From-Us sort of pseudo-conspiratorial nonsense. Whereas the evidence I've seen for zoonosis tends to be in the form of, y'know, actual statistics and genetics analyzing facts closely related the actual COVID-19 virus as its been seen in the real world.
Please understand, it's not that you can *never* learn anything useful from this sort of evidence. It's just that there seems to be a really, really strong overlap from people who put a high probability on lab leak and people who routinely overestimate how useful and reliable Argument From Fuzzy Human-Centric Story type arguments are. At their best, they can tell you where to look for better evidence. You definitely, definitely shouldn't be using them as the cornerstones of serious conclusions. It's been ages since I've read the sequences, but I vaguely remembered there being a bit about exactly this in there, and lo and behold there was (and it was nice and brief and concise). If in the year 2026 you want to have any chance of convincing anyone who isn't already convinced of lab leak, you need to get a LOT closer to the query than "see, the CCP once talked about some things that were kinda sorta related to this."
https://www.lesswrong.com/posts/2jp98zdLo898qExrr/hug-the-query
Edited, whereby I mean new comment: How many lableaks do you think America has had? Can you name them?
We can do "was this a safe environment" -- which, it seems pretty clear it wasn't... and then we can do "here are the known lab leaks" and say "this seems pretty questionable, basing a known problem-causing laboratory in the middle of a very busy city."
You seem to be fundamentally misunderstanding the issue. No amount of counting the past lab leaks ever provides direct evidence. No amount of implications and speculations and innuendo ever provides direct evidence. If you want to know if something is true about the world, you have to actually go out and CHECK.
(They check for COVID 19. It wasn't a lab leak.)
p.s. Say, didn't you get banned for this exact terrible pattern of argumentation? Like, half a dozen times, at least?
I'm making a different argument, which is regardless of whether or not this particular problem was a lab leak, it may be best to continue as if it was*. That's based on the irons in the fire at Wuhan, and their stated objectives.
*implicit in this is our relative inability to put consequences on China in general, I'm not calling for anyone to have a sign put around their neck and pelted with rocks.
https://usrtk.org/covid-19-origins/dia-analysis-covid-may-have-come-from-wuhan-lab/
"In the DIA’s unpublished analysis, scientists noted that researchers at the Wuhan lab were performing the types of experiments that could lead to the creation of a virus like SARS-CoV-2, sometimes under inadequate safety conditions that could not prevent a highly transmissible virus from escaping."
"The DIA analysis pointed out that Chinese researchers, including those at the Wuhan lab, had already inserted furin cleavage sites into coronaviruses. In fact, the lab’s researchers and their American collaborators proposed a project in 2018 that appeared to plan to insert cleavage sites into coronaviruses at the same spot where it exists in the genetic sequence of the virus that caused COVID-19. "
"The DIA analysis pointed out that scientists in the Wuhan lab conducted experiments with live SARS-like coronaviruses under “biosafety level 2” conditions, “which would make an accidental release” of an infectious bat coronavirus “more likely.” "
"The escaped viruses include SARS, which escaped Chinese labs four times, and SARS-CoV-2, which appears to have infected a Beijing lab worker early in the pandemic. "
Okay, there we go. There absolutely was a lableak, in Beijing. Did it become a worldwide issue? No.
So in other words, some unnamed experts attempted to read tea leaves post-hoc, and claimed the tea leaves were predictive.
The "meeting summary" is still online, and we have very good LLMs now. Anyone can read it for themselves: http://www.whiov.cas.cn/xwdt_160278/zhxw2019/201911/t20191121_5439595.html
Here is Claude Fable's translation:
"On November 19, the Wuhan Institute of Virology of the Chinese Academy of Sciences (CAS) conducted a safety work training session. Participants included the heads and safety supervisors of the safety management departments of the CAS Wuhan Branch headquarters and its various research institutes, as well as the Wuhan Institute of Virology's research group leaders, department heads, and all safety officers. The training was chaired by Li Liping, Assistant to the Director.
For this training, Ji Changzheng, Director of the Science and Technology Security Division of the CAS General Office, was specially invited to deliver a keynote report titled "The Situation and Tasks of Safety Work at the Chinese Academy of Sciences." Ji Changzheng conveyed the directives and requirements of the Party Central Committee and the State Council regarding safety work, with an emphasis on communicating and implementing the important statements and written instructions of General Secretary Xi Jinping and Premier Li Keqiang on safety work. Drawing on case studies of major safety accidents that have occurred in China and abroad in recent years, he provided an in-depth, multi-level, and multi-angle analysis covering responsibility and accountability, operational standards, emergency response plans, and hazard identification, vividly illustrating the complex and serious situation currently facing safety work. Finally, addressing the actual circumstances of CAS, Ji Changzheng summarized and analyzed common problems in safety management work, including laboratory safety, science and technology security, student safety, campus safety, and cybersecurity. He proposed that doing safety management work well requires: first, strengthening awareness and elevating political positioning; second, clarifying rights and responsibilities and promoting the implementation of accountability; third, coordinating overall efforts and strengthening control and governance; and fourth, enhancing research on science and technology security risks and building early-warning and monitoring systems. Safety management personnel must strengthen their sense of mission, consolidate their responsibilities, and work together to build a security barrier for scientific research and development, creating a favorable environment for CAS's scientific research and production, and making their due contribution to CAS's reform, innovation, and development.
At the training session, Hu Qian, Deputy Director of the Security Office of the Wuhan Institute of Virology, summarized a number of common problems discovered during safety inspections over the past year, pointed out the serious consequences that safety hazards could cause, and stressed that hazard rectification must be thorough and standardized management must be maintained. He required the safety officers of each department and research group to strengthen the communication and implementation of safety laws, regulations, and rules, further improve operational procedures, guard against all types of safety risks, and ensure the safety of the institute's scientific research and production."
Seems like routine bureaucrat-speak to me.
Did you read the earlier Vanity Fair piece? The "unnamed" Chinese experts -- you can surely just ask ProPublica for their names -- were intended as an unbiased cross-check of the earlier named expert. As such, their names are immaterial, and likely unrecognizable to people not in the field, and not being listed simply "to be brief".
This is about as ROUTINE as GWB getting personally involved in setting up biosecurity, which he absolutely was -- something spooked him bigtime, he was otherwise quite lazy as a President. Aka, it's really not routine at all. When you have the POTUS or Mr. Pooh showing an interest, that's because it's "risen to the point where briefings are hitting the POTUS' desk."
You must not read a lot of propagandistic news. Watch China's central TV station some day, or North Korea's. You'll notice that the president is always giving advice on EVERYTHING, supervising EVERYTHING, visiting EVERYWHERE. They're desperate to take credit for anything going right in the country. GWB was not powerful enough to set up a personality cult, and to his credit, he was probably a decent enough man to not want one.
I don't think you're getting the point -- Xi speaking up critically about something is the issue, not him "being there" -- every politician loves a good ribbon cutting. This is very very much not something going right.
GWB was fond of fart jokes in the oval office -- "pull my finger."
Chinese governmental language is like that, though; you never say "Bob screwed up so now we all have to be really careful with the bad viruses"; it's always a communique stressing the importance of "hazard rectification". Just doing literal translation isn't enough, you need someone who knows Chinese and also knows political jargon and when the writer is quoting Xi Jinping thought or one of the several million anecdotes and historical allusions that educated Chinese are familiar with.
Unfortunately, it's not easy to find someone like that who can be trusted.
In other words, reading tea leaves post-hoc.
I can read Chinese. It's exactly as Fable translated it. There are no anecdotes or historical allusions because it's not high literature, just a boring piece of bureaucrat-speak.
Whatever happened to the Ukraine war. When it broke out I remember there were people here who were like I have to regulate myself to only write at 30 minutes intervals. Now nothing. It's just background noise
I still keep up with it; https://understandingwar.org/ does a good daily summary that's part of my morning routine. But it's a bogged down conflict of incremental changes now. There aren't that many dramatic swings like the Ukrainian breakthrough to Kupyansk in 2022 or the Prigozhin mutiny that catch media and spark discussion. It's just a steady churn meatgrinder of incremental Russian gains for massive Russian casualties and tit-for-tat infrastructure drone and missile attacks. It's bloody and miserable, but it's a status quo, and status quos, once established, tend to spark limited discussion unless and until something happens to disrupt them.
You agreeing with the other guy that Russia is winning?
I mean, sure, in the same way that you can call a 50.2% to 49.8% advantage a "majority" I suppose we can say that Russia is 'winning.' But the *pace* at which you are winning - how many days it will take you to win, and how much each day costs you, is kinda important, and those numbers don't look good in Moscow, 'winning' or not.
Like legendary Pyrrhus.
Hey. Cs get degrees
Put simply? The Ukrainians are losing. 30 missiles hitting Knib, some of them "the really good ones," no Patriot Missiles to defend themselves with.
Put less simply? The Mighty Wurlitzer (American Version) has stopped creating high volumes of propaganda for the Ukraine. We are no longer sanitizing the output of the Ukrainian war, or curating it for Western Sensibilities.
The US has also stopped providing defense technology, not just propaganda.
I think you missed a word there -- "openly."
I thought the Ukrainians were winning? They have apparently damaged 30% of Russia's oil production capacity
https://en.wikipedia.org/wiki/Timeline_of_the_Russo-Ukrainian_war_(2022%E2%80%93present)#2026_campaigns
Do you also think America is winning the Iranian War? They are doing nigh-permanent damage to Iran's ability to continue to pay its workers with oil revenue.
Ukraine is a country that thinks "blowing up Wildberries -- Russia's Amazon" is a good move during a war. Intelligent people ask "didn't you have a better target than that? You know, like the missiles that are currently hitting Knib?"
For some reason, Americans seem inclined to cheer for the "theater kid dictator" currently in charge of the Ukraine... Only Americans actually cheer for the underdog -- the Chinese certainly don't, nor do the Russians, or the Indians...
>Do you also think America is winning the Iranian War?
Yes.
Wasn't this the first year Ukraine recovered more territory than Russia captured?
I haven't heard that from American sources, and I'm reluctant to trust Ukrainian sources, given their notable use of theater kids in the planning departments. Can you cite an unbiased source, or a Russian source, should such a unicorn exist?
A munition designed to destroy or degrade enemy infrastructure is not a munition designed to intercept missiles and cannot be used for that. Ukraine does not have the ability to produce its own interceptors. Additionally, Russian missile stockpiles are not going to be kept in an above ground facility by a river where a drone can easily access it. Your argument makes no sense.
Ukraine has demonstrated their ability to murder "nuclear weapon carrying" jets in Russia. That's probably a better use for missiles/drones than a wildberries warehouse. Your correction is well noted and I thank you -- I was being imprecise.
Please note: I'm against civilian casualties, not necessarily against "blowing up civilian infrastructure", except for when that actively is a warcrime, like depriving people of potable water.
If Iran can blow up American drones before they're even assembled, I have to assume that Russia can do the same to the Ukrainian drones, or vice versa.
> Ukraine has demonstrated their ability to murder "nuclear weapon carrying" jets in Russia. That's probably a better use for missiles/drones than a wildberries warehouse
Not necessarily, if those jets aren't going to be used against Ukraine anyway.
Now that Ukraine can strike inside Russia it seems like the plan is to maximise economic damage without causing too many Russian civilian casualties. History shows that killing a lot of enemy civilians just hardens their resolve, but if you can just make their lives annoying and inconvenient without killing anyone they know then they'll start to say "actually I don't care about reuniting the 13th Century Kingdom of Rus, I do care about not being able to fuel my car".
This is exactly why the US loses all of its wars, it goes into them with a very modest amount of enthusiasm, and then this limited enthusiasm slowly evaporates the moment things get the least bit difficult. That's how they wound up handing over Vietnam to the Communists, Afghanistan to the Taliban, and how they're somehow going to leave the Iranian regime in a stronger position at the end of the current engagement.
Targeting civilians does nothing to help Putin win the war. It pointlessly increases the total suffering, not much else.
You seem to be under the impression that Putin can do what most civilized war leaders can do -- which is destroy the gun factories, the weapon factories. This is not possible when the gun/weapon factories lie in other countries (except if Putin wishes to get into a Big Battle with NATO).
Without that option, the only way to win is by demoralizing the other side.
I assume you're talking about Ukrainian civilians, and not American civilians. it is quite easy to argue that targeting American war-mongers is going to have a chilling effect on the American willingness to fight (and possibly the European as well).
>You seem to be under the impression that Putin can do what most civilized war leaders can do -- which is destroy the gun factories, the weapon factories.
I am not. What gave you that impression?
>Without that option, the only way to win is by demoralizing the other side.
Except that has been tried since WW1 and never worked, and it's not going to work this time either. No matter who are objectively the bad guys in a war, bombed civilians will always blame the people bombing them and support their government more.
>>Without that option, the only way to win is by demoralizing the other side.
> Except that has been tried since WW1 and never worked
Never worked? What do you think happened in the Vietnam War?
What happened is that the USA dropped twice as many bombs as in WW2 and they still lost. Or what was your point?
https://storymaps.arcgis.com/stories/2eae918ca40a4bd7a55390bba4735cdb
Certainly not bombs hitting American cities until the US people were demoralized.
It was the way you'd put it, which tilted a bit towards "this is genocide and Putin is bad" rather than "Putin has no good options, and is going with the least bad option, and it's not likely to work for him."
Bombed civilians seem to be supporting Fedorov the "only clean guy in the Ukrainian government," rather than the dictator who plays piano with his dick (late night variety show bit). I do not blame them.
By the same token, the people in the Donbas getting bombed by Zelensky/prior Ukrainian governments will not work to reintegrate them back into the Ukraine. They will support "the people that aren't bombing us" (or flee to "not bombed places"). Ditto the Crimeans (who really, really seem to like Russia -- more than the actual Russians).
Ukraine seems to be going in hard on "We can cause economic harm, and that will collapse Russian government." (To note: the fall of the USSR was caused by hard times engineered by the United States, so it's not the completely craziest thing).
> the fall of the USSR was caused by hard times engineered by the United States
You manage to put so much disinformation in one comment, that it is difficult to respond to each one individually, Wimbli.
In this case, "hard times engineered by US" probably refers to Soviet Union's inability to produce enough food for its own citizens. Which is somehow everyone else's fault, as is the traditional explanation when a planned economy fails. (The five traditional enemies of soviet economy: spring, summer, autumn, winter, and foreign saboteurs.)
>"Putin has no good options, and is going with the least bad option, and it's not likely to work for him."
Strategically, it's a terrible option. It has strengthened NATO despite Trump's best efforts. It keeps the Ukraine war in the news more. It puts more pressure on Ukrainian allies to provide military and humanitarian aid. And as I said, it unites the Ukrainian people behind their government and discredits all voices that hope for an actually peaceful resolution.
So no, it's not the least bad strategic option. Not bombing civilians would be the least bad option. But that's just Russia, they cant help themselves but be brutal conquerors. It's been their MO forever.
How does that number of missiles compare to the number of bombs that hit London during the blitz? The air war, as always, is a bit of a sideshow.
>>The air war, as always, is a bit of a sideshow.
In this case doubly so, given that as Aeolian already noted, Putin is unable to destroy the gun and weapon factories supplying Ukraine.
It's a nasty quagmire - you never want to be stuck in a fistfight that's a matter of honor and pride for you but a matter of life or death for the other guy... but that's pretty much where Russia is, and the only path forward for now seems to be feeding more Russians into the meatgrinder hoping that something changes before their population loses patience with it.
It's a proxy war, just like America is in with Iran. Seeing this as "Big Guy" versus "Small Guy" rather than "Big Guy" versus "Small Guy Supported/Paid by Big Guy" is a mistake.
That makes it worse, though. It means that "Small Guy" is in an existential fight for his life against "Big Guy" *and* has, as you have noted, an unbombable supply of weapons to draw on to help him hang on..
That leaves Big Guy (a) throwing bombs at Small Guy hoping to break his will with aerial bombardment, which as you have noted in the case of Iran, and EOC has noted in the case of every time since WWI, has never worked, and (b) throwing his own people into the meatgrinder, trading lives for WWI trench warfare paced gains and praying to Orthodox Jesus that the EU has a crisis or Ukraine has a crisis so that something changes. Which isn't that much better a plan than Ukraine hoping for a Russian crisis to turn things around. Prospects are limited, and in the meantime you're just stuck hoping your people can tolerate your losses better than the other side can tolerate theirs. And that's a tricky morale problem to come out on top of when you're fighting for reputation while the other guy is fighting for his life.
Big Guy Versus Big Guy is the actual war, though. The Small Guy is immaterial, as he's no longer in charge of his destiny -- if the guns/bombs stop coming, he folds immediately. So, you get dumb theater kid tricks (well, at least you do if you're the Ukraine. Iran is smarter, and has had 40 years to dig in and get good and used to sanctions).
EU is having a crisis. Several, in fact, and some of which are caused by America (America is a very bad ally to have, I would not recommend).
I'll note that the Iraq War was basically won in about thirty minutes of bombing. There were... some very optimistic people who said Iran could be won in the same amount of time. America/Israel did manage multiple decapitation strikes.
If Putin wanted to, he could decapitate the Ukrainian government -- they certainly can't stop the missiles anymore. He thinks Zelensky does more harm to the Ukraine than good.
You have Little Russian Army and Big Russian Army (okay, if you want, little Ukrainian Army and Big Ukrainian Army), they both use the same tactics, same military structure. They're both playing "slog slog slog" and "very slow war."
Best plan for everyone would be to call a peace. Let Trump celebrate, and let the Europeans pay for more Russian blood when there's a president that won't actively sabotage the Europeans shooting themselves in the carotid.
Putin, I'm pretty sure, doesn't think he's fighting for reputation. The fight with Europe is an existential threat... (Europe's current allocation towards "clean energy" makes a lot more sense if they intend to get most of their resources from "dirty Russia" -- just like "dirty Polish lignite" right now).
(Countries come up with "existential threats" a lot more frequently than you'd think. China thought Trump was an existential threat, and acted accordingly).
Probably still fewer than the number of missiles Zelensky (and predecessor) sent into Ukrainian cities before Russia set foot on non-Crimean Ukraine.
Relevance, even if true?
Quick google (may be wrong) suggests 60% of London was destroyed during world war II. Compare to Berlin, which the same "quick google" says 11% destroyed and 70-80% bombed.
Both of these are a "large scale bomb everything" campaign, intended to level the entire place (and, probably, not directly comparable to the "enduring legacy of Communist Russia" -- Commiblocks, nigh indestructable housing!)
I was merely suggesting we look at a comparable "time" and "space" of bombing comparable structures. Makes a better comparison.
Closest cite I can get for Donetsk:
https://www.rferl.org/a/donetsk-airport-war-destruction-ukraine-russia/32967224.html
Butter wars are coming to the USA!
In dairy export news today (relevant to me because it involves a company created out of the merger of the co-op in my own county and the one in a neighbouring county https://www.tirlan.com/our-story/tirlan-heritage), cheap(er) Irish butter may be coming to a Costco near you!
Or maybe not, it depends. Tirlán are looking to supply Irish butter to Costco's own-brand label, Kirkland Signature, at a price lower than Kerrygold (the premium Irish export brand to the USA).
This has ruffled some feathers, because the farmers are worried this means downward pressure on milk prices. I don't see why they fear that, but then again I'm not au fait with the economics of butter exports. Maybe it will be that, in order to cut costs and maintain the margins for Costco demands, the price of milk to produce that cheap butter will be reduced.
It gets even more convoluted given that Tirlán is a shareholder of Ornua, the parent company behind Kerrygold
https://www.irishtimes.com/business/farming-food/2026/08/12/tirlan-accused-of-sabotaging-irish-butter-exports-with-costco-deal/
"Dairy co-op Tirlán has been accused of sabotaging Irish butter exports by signing a deal with US retailing giant Costco to sell butter under Costco’s private label, putting itself in direct competition with the State’s flagship brand, Kerrygold.
Farmers’ groups, including the Irish Farmers’ Association (IFA) and Irish Creamery Milk Suppliers Association (ICMSA), claim the move could result in a “race to the bottom” at the expense of producers here.
Tirlán’s Costco butter will be sold at about $6 (€5) a kilogram, substantially cheaper than the Kerrygold equivalent.
Tirlán, which separately supplies Kerrygold owner Ornua and is Ornua’s biggest shareholder, defended the move, saying it was responding to a growth in demand for private label grass fed butter."
As a linguistic aside, this is the kind of "kludge two words together and never mind if the result makes sense" that drives me batty.
"In the late ‘90s Glanbia plc was formed – Glanbia is a combination of Irish words meaning ‘pure food’."
Well - not precisely. "Glan" does mean "clean" ("glan, gsf. glaine, gloine, a., clean, pure, clear, sincere, innocent, bright, white") and "Bia" does mean "food". But in Irish language usage, to say "clean/pure food" you would phrase it as "bia glan" (noun-adjective) and not "glan bia".
Same thing annoyed me with Irish Water which is now calling itself Uisce Éireann. That is not "Irish water" (i.e. "water which is Irish, water from/of/sourced in Ireland"). That is the words for "water" and "Ireland" shoved into proximity, so the translation is literally "Water Ireland" and not "Irish Water" (which *should* be "Uisce Éireannach" or, better, "Uisce na hÉireann" (water of Ireland) which is the meaning they are going for).
This kind of illiterate branding is very annoying and insulting. 'Ara, just slap a vaguely native-themed name on it and don't bother if it makes sense or not'.
>> Dairy co-op Tirlán has been accused of sabotaging Irish butter exports by signing a deal with US retailing giant Costco to sell butter under Costco’s private label, putting itself in direct competition with the State’s flagship brand, Kerrygold.
I don't get it. If one Irish butter brand starts competing for overseas customers with another Irish butter brand, the only realistic outcome is that Irish butter exports increase. How does this sabotage "Irish butter exports"?
It doesn't seem to make sense, the best I can make of it is that Kerrygold is treated as a premium brand (plus protectionism from American butter producers, plus Trump's tariffs) and so prices range from $16-20 per kg. Tirlán (in that article) will produce own-brand butter for Costco selling at $6 per kg.
So I imagine the fear is (1) the premium brand will be replaced by retailers selling Irish grass-fed butter at one-third of the price, hence demand for the premium product will fall (2) if other Irish co-ops/dairies start producing own-brand knockoffs for American retailers, then the consumer will not realise these are Irish and not American products and so the demand for Irish butter in general will be replaced by 'I can buy American butter that is just as good' and (3) good old-fashioned scare-mongering by Ornua (owners of Kerrygold) in response to a perceived threat to their nice little niche as "the Irish butter America buys".
Speaking as an American, I would ordinarily have no idea that a brand of imported Irish butter existed. You buy the store brand. I will be shocked if Kerrygold butter is more than an utterly negligible percentage of "the butter America buys".
As it happens, I am familiar with Kerrygold butter, but that's because I frequently saw it in Shanghai. But I didn't need enough butter in China to have any memory of whether it was priced as a luxury import or as a Chinese grocery item. (In contrast, I am acutely aware that milk is much more expensive in China than it is in the United States.)
I do recall that the Kerrygold butter was more difficult to use than American butter, because it came in an oddly-shaped block without good tablespoon markings, but I imagine that what Kerrygold exports to America is packaged to meet American expectations.
> (1) the premium brand will be replaced by retailers selling Irish grass-fed butter at one-third of the price, hence demand for the premium product will fall (2) if other Irish co-ops/dairies start producing own-brand knockoffs for American retailers, then the consumer will not realise these are Irish and not American products
I don't think these are compatible fears. If (as I agree is likely) the consumer doesn't realize that the ordinarily-priced butter at Costco is an Irish import, Kerrygold will retain all of the branding value, whatever that might be, that it currently has.
If Kerrygold is severely undercut by the availability of "store-brand butter that is just as good", the implication is that Kerrygold can compete on quality with American butter (and that this, not branding, is the reason for the higher price it commands), but not with other Irish butter. In that case exporting Irish butter to be sold at normal non-luxury prices would be a 𝗴𝗶𝗴𝗮𝗻𝘁𝗶𝗰 𝘃𝗶𝗰𝘁𝗼𝗿𝘆 for Ireland; they could feasibly replace the entire American butter industry. (Though I grant that that would probably cause political problems within America and result in some kind of response from the US government.)
I know some people who buy Kerrygold and I tried it once. I didn't know it was Irish or imported though. I expect it would continue selling for a premium even if some of the cheap brands are Irish, as you suggested.
Just speculating here, but if "Irish butter" is a popular brand, and one company has a monopoly on selling it in USA, they can keep the prices high. When two companies start competing, the prices get lower.
So even if more total butter is sold, it is not obvious whether the total profit will increase. (Definitely the first company's profit will decrease, which is a sufficient incentive for them to claim that it will hurt everyone, whether that is true or not.)
Total profit might or might not fall, but the industry would be healthier in either case. It's always better to be the guy (or country) who makes vast quantities of stuff and sells it at the market price than to be the guy with a hobby in occasionally making something.
Mentioned further down - Éireann is the tuiseal ginideach, so Uisce Éireann is 'Water of Ireland', which is the grammatically correct way to phrase the name of a state body.
Éireannach is almost always used in the narrow sense of nationality, it has broader usage in English because they don't have the genitive case. I checked the national corpus (https://www.corpas.ie/ga/cng/?q=Éireannach) and the first 20 or so cases all refer either to a person or a work (play, film, book) which has a national origin, but even that use is pretty specific.
There are lots of abuses of the language out there but this is not one of them.
Thank you for the grammar lesson, we all remember struggling with the tuiseal ginideach!
Yeah, I agree "Éireannach" is used more in the sense of "Irish man" and a personal identity, but "Irish Water" are presenting themselves as "The state body providign the services of Irish water in Ireland for the Irish people that is water of Ireland and not, for example, British water or UK Water or other water body", so the 'national identity' sense could be argued to fit there. I don't like it, though, since it doesn't sound right.
I do think it should be "Uisce na hÉireann" but that probably is just a personal quirk of preference on my part. Too much "official Irish" is badly kludged together and poorly translated, and when you've been taught a particular dialect in school then Standard/Official Irish is trying to blend Connemara, Ulster, Munster and increasing Dublin Irish usage into a Frankenlanguage that sounds good to nobody but is 'representation'.
> we all remember struggling with the tuiseal ginideach!
Eh, just the genitive case for people who can’t be bothered with prepositions.
> Éireannach is almost always used in the narrow sense of nationality, it has broader usage in English because they don't have the genitive case.
1. I don't understand the primary claim. As far as I'm concerned, Éireannach has zero usage in English.
2. There are two things commonly referred to under the name "genitive case" in English:
2. (a) The clitic particle 's, as in "my dog's fur", is often thought of as a case marker. Historically, it does descend from the Old English genitive case ending. It's difficult to call it a case marker today, because it doesn't attach to nouns. This is widely recognized, but people nevertheless persist in calling it a "phrasal genitive".
2. (b) English pronouns come in subject and object forms, and (unlike the particle) these forms are very reasonably called "cases". It is common to extend that logic and say that 𝘮𝘺, 𝘺𝘰𝘶𝘳, 𝘩𝘪𝘴, 𝘩𝘦𝘳, 𝘪𝘵𝘴, 𝘰𝘶𝘳, 𝘺𝘰𝘶𝘳, and 𝘵𝘩𝘦𝘪𝘳 are pronouns in the genitive case. I tend not to agree with that - for one thing, they can't occur where the grammar calls for a noun - but it's a mainstream viewpoint.
Sorry - I'm comparing adjectival use in the two languages. In English the adjective 'Irish' stands in the for the genitive 'of Ireland' in many contexts where the latter is idiomatic in Irish. In Irish the adjectival form is basically only used to refer to an Irish person or an Irish citizen - the round translation
'Uisce Éireann -> Irish Water -> Uisce Éireannach'
alters the meaning though both collapse onto the same English phrase.
More generally, and I think this is the point Deiseach was making, pulling words out of a dictionary is not the same as translation.
When I want to start a grammar argument, I tell people that 'tomorrow' is an adverb.
Thanks, that's more clear.
You might be interested to know that "because they don't have the genitive case" is probably not a great analysis; some languages have dedicated gentilic forms and others not so much. Latin uses its genitive case extensively, but you have all kinds of different options for describing the relationship between a place and a person; you could say "of <wherever>" in the genitive case, although I tend to suspect you wouldn't, you could use a generic adjective, or you could use a special gentilic adjective.
The gentilic would usually be used for indicating that someone comes from a place - "Athenian" in Latin is "Atheniensis", from "Athenae" (the city) plus -ensis, the gentilic suffix. But interestingly enough "Roman" is different; the Latin is "Romanus", from "Roma" (the city) + -nus, which is a generic adjective-forming suffix that _doesn't_ carry any special "from a place" sense.
> When I want to start a grammar argument, I tell people that 'tomorrow' is an adverb.
This seems incomplete. "It will happen tomorrow" makes a strong case for "tomorrow" being an adverb; "Tomorrow is Christmas" looks more like a noun.
I’d go with ‘tomorrow’ being an adverb of time, but as you say it would probably cause arguments.
Again, there's an adverb and there's a noun. This isn't a controversial viewpoint or anything. Check any dictionary.
https://www.merriam-webster.com/dictionary/tomorrow
"𝘛𝘰𝘮𝘰𝘳𝘳𝘰𝘸 is an adverb", as an unqualified claim, makes as much sense as saying "𝘳𝘪𝘨𝘩𝘵 is an adjective". There is an adjective spelled 𝘳𝘪𝘨𝘩𝘵. Not every use of the syllable is a use of that adjective, or any adjective.
This thread is making my nipples hard.
I do enjoy me some Kerrygold, but I'm not a Costco member, so I guess I'm SOL on getting it on the cheap.
My pet peeve in the US is companies that try to shoe-horn all or part of "America" into their name, as if some vague association with patriotic sentiments is going to help sell their otherwise unremarkable products. I think the worst offender is Ameriprise Financial, a name which rolls off the tongue with about the same gracefulness as my grandma can ride a surfboard. My joke about this is that if I ever get rich, I'm going to open a chain of OB-GYN clinics called AmeriC**** just to take this idiotic trend to its most absurd conclusion.
> a name which rolls off the tongue with about the same gracefulness as my grandma can ride a surfboard
I think you missed an opportunity here to draw a parallel between the name rolling off the tongue and your grandma rolling off a surfboard.
Reminds me of Häagen-Dazs. The name was literally just made up on the spot by the American founder who thought it sounds Danish.
LOL, Claude says that you are wrong, because there are names like "Iarnród Éireann", "Dáil Éireann", "Tuaisceart Éireann", which follow the same pattern as "Uisce Éireann".
What an odd thing to post. You're responding to a native Irishwoman, who's sharing her (interesting!) firsthand experience of various faux-Irish corporate brand names sounding awkward to her ear as an Irish speaker. So you take it upon yourself to drop in this random contradictory piece of text from a model-unspecified English-trained generative language machine, that you didn't even bother to request references for?
With an *LOL*, forsooth, as though you were delivering some sort of epic gotcha?
You don't claim any expertise in Irish yourself, so I conclude you are fully unqualified to access any of the nuances that might help you determine what the model's text is referencing or to what extent this is hallucination, irrelevant misinterpretation or valid corrective to the original statement. What are you adding here? Did you think everybody would admire you, because you copy-pasted some writing into a text box and then copy-pasted that into a comment to us?
> What an odd thing to post. You're responding to a native Irishwoman, who's sharing her (interesting!) firsthand experience of various faux-Irish corporate brand names sounding awkward to her ear as an Irish speaker.
Why would you assume that, as a native Irishwoman, Deiseach must also be an Irish speaker? In general that is not the case, as we see here.
In fact, Viliam is correct. Éireann is the genitive case -- 'of Ireland' so Uisce Éireann is 'Water [Company] of Ireland'. This is the idiosyncratic and correct name for the organisation in Irish; all the other examples given are correct and appropriate. While my credentials shouldn't be relevant, I'm a native speaker and it's my usual working language. Furthermore, the AI's are really good at Irish grammar now. To the extent that I rely on them for proofing my own writing - they're not always right but they'll identify the grammatical slips I make habitually for me.
I'm guessing that Deiseach is not a fan of the tuiseal ginideach (which is and was poorly taught in school). I'll agree with that all the agricultural co-ops have horrible names. Tirlán is, if any thing, worse than glanbia. Tír is 'country', and lán is 'full' (but in the sense of being filled to the brim, not complete, or whole). It has an ethno-nationalist undertone to me, which I don't believe is actually intentional.
Tirlán is terrible, and an example of what I meant by knocking together two separate words according to dictionary meaning but ignoring if they are used that way in practice. The logo is also teeth-grindingly horrible. Look at this and tell me if it means anything or does anything other than drive you to fury:
https://icos.ie/wp-content/uploads/2025/06/tirlan_logo_positive_rgb-scaled.webp
It really is one of those "did you get a six year old to draw it?" images, and yet I bet a design company was paid a hefty fee to come up with it.
"Tirlán logo and name reflect the Irish Co-op's heritage. The name combines tír (land) and lán (full) to mean "land of abundance". Its graphic symbol inside a circle is inspired by Úr, an ancient Ogham letter meaning earth, soil, or clay, symbolizing the deep bond between the farmers and their land."
Ogham my backside, it looks like chicken scratches. As you point out, that term does not mean "land of abundance" (that would be more to do with "flúirseach", or according to the old Foclóir Gaelige-Béarla from school which I dug out now, "raidhse", fem. 4th declension noun, "abundance").
Are you of an age that you remember the Loriel ad that finished 'Inspired by the science of genes'? It's equally awful. I suppose that there's some pressure to balance an Irish image with a name that people can pronounce if they know nothing about Ireland.
I quite like Flaithiúil as an alternative to Fliúirseach (which I associate with a foolish generosity, as of someone buying a round before the previous one is finished). Its meaning is more like 'noble generosity' which is appropriate in the context. Just pulling a word out of the dictionary blind to its context is not great.
I recently became aware (again) of Mairtín Ó Cadhan's dictionary: https://focloiruichadhain.ria.ie
He wrote thousands of pages, the definitions are minimal but each word is illustrated with as many as 30 examples in use. One would imagine that the Execs in Tirlán are unaware...
My read was that they are laughing at Claude.
Hm, I can kind of see it. If you're right, then hilarious self-own for me, and @Viliam, I apologize for being so testy. I've been dying a little bit every time Scott pastes a random LLM screenshot as "source" for some grand claim, and I guess it finally felt like too much.
I assumed that Deiseach would understand it as a friendly challenge to prove me/Claude wrong. Or that something else interesting would happen.
But yeah, reading the comment again, it is plausible to read it as "AI-powered mansplaining". The initial "LOL" was meant to express my confusion at two entities disagreeing on a topic I know nothing about. But looking at it with fresh eyes, it is also possible to read it as an expression of aggression. Sorry for that.
Deiseach is not a fragile snowflake and can handle that, I assume.
In the event, you were right. Emily's comment seemed to me based entirely on who you were and who Deiseach is (to the extent that these can be inferred from the internet).
I notice in my own life that people are overly quick now to dismiss something *because the answer was generated with AI*, even if the answer is otherwise correct and appropriate. This just pushes people to hide their use of AI, which is not what anyone wants.
For what it's worth, I hate LLMs, and so I enjoyed your comment whatever Villiam's intent was.
In the past, Villiam wrote this comment: https://www.lesswrong.com/posts/pvGavgxY5mQcyvPQa/g-wood-s-shortform#3DKQ7W8iGCETxoJer
that desrcibed parts of why I hate LLM writing very well. That comment of his informs my reading now.
Scott's MacGregor joke got me wondering whether sheep are injured when human males have intercourse with them. (The reason I was curious is that it does not seem morally wrong to me for someone to do that, unless the sheep is injured or badly frightened). Nobody here knew, and GPT Sol was not allowed to discuss the topic -- the censors shut us down. But today I managed to get the question past the censors, and the answer is that In an adult ram, the erect penis is 10-14 inches long and 1/2-1 inch in diameter. And of course a lamb's vagina has to have a good amount of stretchiness. After all, lambs come out of it. So I don't think intercourse with a male of our species would cause injury. In case you've been wondering. Or even sleeplessly fretting about the question. Or sheeplessly fretting.
We need a species sexual compatibility chart -- for science!
Although this sounds like exactly the kind of research that no big grant donor would want to be associated with.
Dogs really like peanutbutter (and chicks really do film themselves...).
That's probably why it's the cops conducting this research, honest to god -- they're allowed to look at all the pornography (in the interest of catching the Bad Criminal Guys -- you may know them as Oligarchs).
Human penises aren't barbed, unlike, say, cats (which has a relatively straightforward "finish copulation" reason). I have no idea why pig penises are corkscrews (Thank You Drawn Together, for the Anatomically Correct Animation... words that probably no one ever said, before or since).
> After all, lambs come out of it.
The same argument could be made about human vaginas, after all, babies come out of them. Unfortunately, reality is more complicated than that - both for humans and for sheep too.
Obviously, there are other factors to consider, eg roughness. My point is that it is not guaranteed to happen every time due to size mismatch alone.
Sure, but I don’t believe your size estimate either. This feels like common sense if by you’ve spent time with animals up close, but if here are some veterinary pictures to illustrate: https://visgar.vetmed.ufl.edu/en_ovirep/penis/penis.html
Here's the link Sol gave for it's info about dimensions:
Veterinary Research Communications (2023) 47:2259–2264
https://doi.org/10.1007/s11259-023-10128-8
What do the photos you linked show? If they are of injured sheep vaginas I do not want to look at them.
From the paper you’ve cited: « The penis in the ovine is a tubular fibro-elastic organ, around 40 cm long with an exiguous portion of erectile tissue. »
Exiguous means scanty or meager, i.e., yes, the entire structure is 40 cm long… but the erectile / external part is small. There is a picture of a flaccid penis in that very paper which shows this clearly.
The pictures I linked show ram’s penises (with human hands in some for rough scale) - including an erect one, which is difficult to judge for size but does not look much larger (than flaccid).
OK, I looked at the photos. Looked like the penis was somewhat thicker than the fingers on the hand that was holding it. I have small hands, and all of my fingers look to me like their diameter except at the very tip is 1/2" or more, and my thumb looks to be 3/4" wide, and and inch wide at the base. If the hands in the photo were larger than mine, then their diameter would have had a minimum and max larger than mine. So all that seems to line up well with the diameter figure Sol gave.
Wow $250k seems like a very very big money for this kind of bet, maybe they secretly want to give you money instead? I agree that I won't take this bet even if I'm a lab leaker mostly because I don't believe that there'll be any conclusive evidence that can sway the argument either way in this stretch of time.
Do aphantasiacs get songs stuck in their head? I got curious and turns out that while some do, most only get lyrics without the audio: https://www.reddit.com/r/Aphantasia/comments/ayllo2/do_you_get_songs_stuck_in_your_head/
I can play whole songs/albums in my head but I cannot picture a gd thing lol
Related to organ donation, blood donation is another needed service that’s hard to incentivize. Goodlabs.com is a new diagnostic company that provides lab tests in exchange for donating blood; if you’re already donating blood, why not get some free lab tests at the same time? This is a great incentive to promote personal health and help out the blood supply.
IMHO, everyone should regularly test some basic blood markers (e.g. lipids, HbA1c, homocysteine). Plus, you still get the free cookies and orange juice they usually give you.
I tried to donate plasma once, specifically for the financial incentive, and they turned me away because my pulse was slightly too high. Ironically, the second time they tested it was even higher, because I was afraid of it being too high and not getting the money. One way to interpret this situation is that incentives aren't enough to overcome a flawed system. I understand why they want pulses to be in a healthy range, but surely slightly unhealthy blood is better than no blood at all? If there truly is a massive shortage of blood, why not loosen the standards a bit until there's no longer a shortage, or at least so long as that does no active harm to the recipient?
They might have turned you away because they were concerned with you own health.
If that's the case then you could have an informed consent system where they say "look this is risky because [10 billion reasons] so you can only do it if you promise not to sue", which they kind of already do in general because donating blood is always a bit risky anyway (there's always the off chance that your body doesn't replenish it right away, or you get an infection from the equipment, or other unknown unknowns).
"surely slightly unhealthy blood is better than no blood at all?"
Ahem.
https://encyclopediaofarkansas.net/entries/arkansas-prison-blood-scandal-3732/
https://haemophilia.org.uk/public-inquiry/the-infected-blood-inquiry/the-contaminated-blood-scandal/
https://www.bbc.com/news/articles/cjd2rppvmv5o
Very briefly: since the USA buys blood rather than relying on voluntary donations, things such as "hey we got an entire prison full of guys who could make some money for themselves and us by giving blood" happened. Trouble is, in prison you got guys who are drug addicts and (at the time) HIV positive plus other blood-borne diseases like Hepatitis C.
Guess what happens when you mix in "slightly unhealthy blood" with other donations and produce the plasma for export?
Canada, Ireland and the UK were three countries that ended up with government enquiries and compensation schemes (often bitterly contested by those same governments) for the haemophiliacs and pregnant women who had been infected by those batches of blood.
Now we have modern techniques of screening and processing to avoid similar problems, but yeah: "slightly unhealthy blood" is covering a multitude of sins.
>since the USA buys blood rather than relying on voluntary donations
Both systems exist in the USA.
source: I'm in the USA, and donate at the Red Cross every 2 months, and that is a voluntary donation system.
For what it's worth I'm not proposing mixing all the blood together. You could have a tiered system where those at the top of the queue would get the healthy blood, and then people further down (meaning in less immediate need of blood) could choose to take the risk of getting less ideal blood or wait for the good stuff. And I'm still in favor of screening out blood that actually carries diseases, since that would obviously harm any patient.
Eugenics! all from "slightly unhealthy blood" (basically made hemophilliacs cease to exist, due to aids).
Presumably the risk is "way too much blood," as your elevated heartbeat blasts it out of you like a machine gun.
"Hard to incentivize" bears very little resemblance to my experience with blood donation. Yes, there are a lot of people who just won't do it, ever, but for a full day off work? I think most people will donate blood.
Meh. For a day off of work, I might or might not donate blood. Work's not *that* bad.
If I could repeat it as much as I wanted, then sure, but you can't do that.
Expensive to incentivise, then? A full day's pay costs a lot more than some blood tests.
Also, outright paying for blood is illegal in many places.
I work for a hospital, not sure about the laws in the various states (it's a big hospital chain).
I am actively starting to intentionally disengage with comments here because Substack started spamming me with 25 'also replied' notifications for every 1 real notification about someone responding to me. Does anyone know a way to fix this?
Yeah, that sucks. Happened to me too, made a top-level comment that started a long discussion, and the notifications keep coming even if deep down in the comments people are already discussing something else.
I can see, under "settings", under "engagement", 4 options for turning notifications on or off:
-likes
-replies
-restacks
-mentions
I haven't tried turning any of them off, and I _used_ to just get notifications from direct replies to my comments, but I'm now also getting notifications from 'also replied'. It is imaginable that the 'also replied' might get classified under sort-of kind-of mentions, so it might be worth turning that off and seeing if it helps. If 'replies' is really controlling both direct and 'also replied' then my guess is we are all stuck :-(
We really need the ability to turn off notifications for specific sub comments. The also replied is even more annoying. I once created a comment which generated a lot of sub comments, mostly not to me, one pair of responders were arguing amongst themselves for a week! I just had to mute the whole open thread.
Many Thanks! Yes, finer granularity control would be a help.
>one pair of responders were arguing amongst themselves for a week!
Come to think of it, I've been getting notifications from one such exchange series myself...
Me and the fiancée are planning a “mini-moon” in SF in late October after our wedding. Since this place tends to be tech-heavy, figured I’d ask for recommendations, dos and donts, hidden gems and overrated things, etc.
Good hotels? My only two things I’d really like to see are the CHM (obviously) and some giant redwoods. I’ve been to Chicago, Philly, and Seattle so homeless and vagrants don’t bother me. Would prefer not to rent a car if I don’t have to, but can adjust if need be.
If you like whales, and are willing to change your mind about renting a car (and driving ~2 hours from CHM), there's whale watching in Monterey Bay. Specifically, my watcher of choice is FastRaft.com - instead of a yacht-sized boat where you're 15 feet above the water, you're in a raft that seats 6, and you could reach over and touch the bay. The captain and her husband are super cool guides.
30 minutes further south is Carmel-By-The-Sea, which is a fun walkable beach city. Overall, I consider this a day trip - spend most of your time in that area, maybe stop by Gilroy's Garlic World, or alternately enjoy the drive up and down the PCH, or duck up route 9 to Big Basin Redwoods.
Also, 15 minutes from CHM is the Winchester House. I've heard, but cannot confirm, that David Friedman's house is nearby, but I don't think any meetups are scheduled right then.
This place is kind of fun if arcade games are your thing. https://museemecanique.com/
Note that it is a ways away from the most touristy part of Fishermans Wharf.
And it is close to the SF Maritime Museum, which is free. https://www.tripadvisor.com/Attraction_Review-g60713-d4808869-Reviews-San_Francisco_Maritime_National_Historical_Park_s_Visitor_Center-San_Francisco_Cal.html
Edit: Note that you will have to rent a car to see redwoods, unless there is some sort of bus tour you can book. Also, a day trip to Yosemite Valley is doable, though the falls will not be impressive in October. There are Sequoias in the south of the park, near Wawona, but that means more than a day trip. The Wawona Hotel is nice, as is the Ahwahnee in Yosemite Valley.
Others have mentioned the cable cars, but I'd also recommend the (free) Cable Car Museum, it's both a fun lens to learn about SF history and also itself a fully-functioning piece of machinery that keeps the cable cars moving. My personal favorite is to ride a Waymo to the cable car museum (optionally visit chinatown in the process), and then hop on a cable car to go to fisherman's wharf from there to get the full bracketing of first-and-latest transit innovations from SF.
I also definitely recommend the Exploratorium. I'm not aware of any museum quite like it in the world for those who like hands-on figuring out of stuff.
Last I was there, Jtown was a blast. Walk the streets that are just stairs. Take the cablecars as the "very last thing you do" (we found locals on them, playing "freaks come out at night" -- it was cool, and very much less crowded than during the day).
Go to "Lindy in the Park" on Sunday in the Golden Gate Park. Get there at 11:30-ish (so you're not late). Free lesson from 12pm which just has you bouncing in place to start. It's a special gem provided by it's 2 founders, Ken and Jen who have been running it for 25 years. From there you can check out the rest of the park. The De Young Art Museum, the California Academy of Sciences, The Conservatory of Flowers, or the many other outdoor things in the park like the Japanese Tea Garden, the San Francisco Botanical Garden, Blue Heron Lake (paddle boats), the Bison Paddock with actual Bison in the middle of a city. You can rent bikes between the two museams or rent the city's bikes from the racks. Sunday is the day to go as parts of the park are blocked from cars.
if you take a trip out to Half Moon Bay to see the ocean, I highly recommend Sam's Chowder House for the best clam chowder I've had in my life.
If you want to experience cognitive dissonance and existential irony, make sure to visit the Mission district, especially the area around Folsom and 18th. Scrape your pennies together for a $7.50 Latte at a sidewalk table guaranteed to be surrounded by wish-they-were-still-20-somethings talking up AI and disavowing Crypto and wokeness (so 2020).
Then just cross the street and mosey down one of the side streets where the un-housed and barely-housed scrape by in campers festooned with rusty bikes. tents, and lean-to's
CHM = Computer History Museum? If so, I wouldn't actually call it a must-do. It's kinda cool, but it's not very accessible from SF proper and not worth the trip unless you have other things in the South Bay to see. You could combine it with a trip to the Google campus (and maybe Stanford's campus, or the radio telescope, or Apple's campus) if you really wanted the full Silicon Valley day trip thing. You'd definitely want a car.
Anyway - the thing about SF is that virtually all of the touristy things are also worth seeing. The only thing that many people would say is overrate is Pier 39 itself, but the rest of the Wharf is cool.
Walking on the GGB is cool. Biking across it is really cool. If you're reasonably healthy I would suggest renting a bike in the Wharf, then riding to the bridge (and over and back).
For redwoods, the best way to do it without a car is to take the Ferry to Sausalito then take the Muir Woods shuttle. Note that you need reservations in advance for this, and in the off-season the shuttle only runs on the weekends. If you wanted to do another day, you'd need a car (and reservations). Note that cell coverage is bad out there so Uber isn't recommended unless you can pre-schedule a return.
If Muir doesn't work out, you'll need a car, but there are tons of great hiking spots within an hour's drive. Mt. Tam has a lot of other spots besides Muir though you'll need to do some curvy road driving and parking can be tricky. I also love hiking in the East Bay, if the weather is nice - Reinhardt Redwoods has gorgeous scenery and views.
I have some family in the area; I don't have a lot of intel, but a few things that were memorable/enjoyable:
Soba-Ichi - if you're in Oakland at any point, this soba noodle spot was amazing
Pâtisserie Rotha - is a great bakery if you find yourself in the Berkeley area, but it's line up at a takeout window, not seated, so be aware if you go.
The Cultured Pickle - also in Berkeley, a pickle shop that lets you reserve seats on the weekend for pickle-themed meals. Really fun if you can get a seat.
Lazy Bear - Fine dining restaurant we took one of my in laws out to as a 'thank you' for a favor. Pricey, but it was a great experience if you want to spend money on a night out.
Church Of 8 Wheels Roller Disco - Old church converted into a roller rink. I still haven't gotten my family to actually go, so I can't *really* vouch for this, but it's the kind of hilarious weird thing that I love, so if you go, tell me if it's fun.
I don't recall anyone ever explaining career options to me in terms of categories and types rather than specific jobs. I'd rather set my kids up with a better understanding of what sorts of careers they want to go after rather than simply specific jobs. This is one of the places that most favors analytical breakdowns and explanations of things, so I'm posting here in hopes of feedback.
Unfortunately, it appears substack doensn't allow formatting in the comments. I'll use *, **, ***, etc. for outline component depths.
--
0. Important Terms
*A. General Classifications.
**1.Blue collar: Jobs you do with your hands and body. Tends to be more physically active. Can range from mind-numbingly repetitive to varied and intellectually demanding. The most specialized jobs pay very well. Often carries a higher level of physical risk. May wear out your back and joints earlier in life. Is sometimes considered lower prestige / social status but may also generate interesting stories, excellent “do it yourself” abilities, and a healthier body.
**2. White collar: Jobs you do with your brain, usually sitting at a desk or in meetings. Tends to be less physically active. Can be mind-numbingly boring at entry level. Often based on developing or following a process repeatedly (accounting, sales, medicine, management, etc.). Communication and dealing with people tends to be key to success.
*B. Standard work week: In America, this is typically a 40-hour week, most often 5 8-hour shifts Monday through Friday. Some jobs, like fireman, medical professional, or offshore oil rig worker, use a rotating schedule that averages out to something close to a 40-48 hour week. 40 hours per week x 50 weeks per year (2 weeks of holidays plus a couple of days unpaid sick or family time) ~= 2,000 hours per year worked. In general, you can take the hourly pay rate associated with a job and multiply it by 2 and then by 1,000 to get a fairly accurate figure for annual pre-tax earnings.
*C. Compensation: How you get paid.
**1. Hourly: You are paid by the hour for the work you perform at a set rate per hour, without a direct link between your speed & skill and your pay. If you work more than 40 hours per week, you typically get paid 50% extra beyond hour # 40 (Overtime pay). If you work nights or weekends you may get paid extra per hour.
**2. Salary: You are paid a set amount per week/year no matter how many hours you work. Legal restrictions mean this is typically a payment method reserved for managers, who are expected to make sure things get done. In some jobs, this means 40-50 hours per week. In other jobs, if you’re in management and someone calls out sick, you are the backup plan and may end up working 60+ hours per week.
**3. Commission: You are paid for performance, typically as a percentage of revenue or a percentage of the company’s profit (revenue from sales minus cost of fulfilling the sale). There may be a guaranteed base pay, or you may be 100% commission. The upside is a higher to much higher potential take-home pay, but the downside is a lot more variance in take-home pay from month to month. Depending on the job, you might need to be available during non-business hours.
**4. Self-Employed: You directly sell your labor or the products of your labor to others. You have a lot more freedom to choose what to do, when to do it, etc., but you are also responsible for tracking your income and ensuring future customers/sales. When you work for a business, they pay some employment-related taxes that fund government programs for medical aid and retirement. When you work for yourself, you are the business, and you pay the business’s share of the taxes as well as the taxes you would normally pay. This is called “self-employment” taxes. However, profit (money that comes in minus expenses that go out) often needs to be re-invested in hiring new people, training, maintenance, additional equipment, etc. You don’t get to pay yourself a lot until year 5, 10, or 15.
*D. Customer Type
**1. B2C Business To Consumer – restaurant, retail, home services provider, windows, painting, roofing, lawn care, anything that sells to the general public
**2. B2B Business to Business – manufacturing, technical services, logistics, raw materials or finished products, anything where the customers are business
*E. Place of Work.
**1. On-Site. You work at a job site, factory, or other place of business daily.
**2. Field. You spend much of your time visiting customer locations, or are in a role where you visit many different locations for your company. You probably are still on-site at a home base location regularly.
**3. Remote. You work from home or wherever you are, instead of in an office. There is no commute, and often no hourly tracking for productivity. You have more time in your day, and often the ability to intersperse personal tasks with work as long as you manage time well. You also miss out on in-person conversations and networking, and are less likely to be promoted, and are more likely to bethe first to be laid off when downsizing due to those lack of connections. You’re also less likely to be able to fully disconnect during a vacation. Remote can be lonely, and video calls do not make up for that.
**4. Hybrid. You work remote a couple of days a week, and on-site a couple of days a week.
Maybe add a side note that many rich people make money in ways that are not present on your list, and probably not available for ordinary people. Such as:
"Money makes money" -- if you inherit lots of money or land, you can make more money simply by putting the money in an index fund, or renting the land. Thus you will never run out of the original money/land (except if a revolution happens), and you have a guaranteed lifelong income.
Jobs that are strongly regulated and you have to be a part of the old guys network to get them. The exact list probably depends on country, but I am thinking about doctors, advocates, notaries.
There is a career type where you work in politics, e.g. on the municipal level, and take bribes. To get this kind of job it helps to be a part of a larger clan, e.g. if your older relatives have that kind of a job, they can leverage it to get a lower-level version of the same to you.
Tell me you're not from the US without saying it. The nepotism you describe is not systematic here, and notary is a job usually handled on the side by administrative staff.
I considered investing to be outside of the scope of this document... and even if someone was inheriting a lot, I'd want that kid to have a career and responsibility vs. not working. It's part of what we were created to do.
Nepotism is systematic here. The Carnegie Mellon Mafia, if I must cite something that isn't religio-ethnic in nature. But the "woke hires" and "diversity hires" as well.
A key advantage of self-employment - it is much harder to get bank loans for cars, mortgages, etc. This forces you to live within your means and greatly speeds up wealth accumulation.
Mortgages wind up leveraging you, though. It's the traditional middle class way to gain wealth (citing Black Wealth White Wealth if you really need a citation).
It's paper wealth, though. If you sell your house and don't use the money to buy another house, you're homeless or renting. The only time it turns into cash is if you have a consolidation (death/inheritance, marriage).
I had to build up 2 years of commission history to get a mortgage. I assume it's similar for self-employed people? Two years of steady personal income?
If you wanted more detailed critique and suggestions, I'd happily make comments and edits in a Google doc if you share a link. (Or DM one...I have no idea if Substack has that feature, never used it.)
I've been interested in work -- like, "how people make their living" -- for a while, and I think this is a cool document with a lot of potential, even bigger than just handing to your kids.
Thanks! I deliberately avoid Google docs and Google accounts (one exists for my phone, never use it) so I'm not sure where to host something like this.
It could be worth mentioning in the compensation section that performance bonuses may exist, and can even be a majority of compensation in some fields.
-Commissions
-Stock options - these build value over time if the company's value goes up.
-Bonuses - who pays out annual bonuses besides law firms? What are bonuses based on?
What else is there?
Don't forget to mention that stock options usually don't work as advertised. For various reasons the options you own may become completely worthless *and* you may have to pay tax on them. The rule of thumb is that unless you are an expert on stock options, you will probably get scammed.
(For example, it may turn out that there are various kinds of stock, and your boss and the investors get a "priority" stock which is actually worth money, and you get "normal" stock which is just a piece of paper. Or your contract may say that you will only get the stock options after staying 5 years at the company, and of course if the stock options would by some miracle end up worth a lot, the obvious move on the employer's side is to fire you after 4 years and 11 months and keep the stock options for himself. There are also other risks, but those are more complicated to explain.)
One nice thing about AI becoming ubiquitous is that people can now put their employee agreements through the AI and get an opinion on this, without having to pay for a lawyer.
I do feel like most people savvy enough to get a decent chunk of their comp in stock options understand vesting and preferred shares and things like that, but you're right, it's certainly not a given, especially for new hires fresh out of college or grad school.
Sounds like halo effect; someone good at coding is not necessary good at legal things and investing. You seem to be in a lucky bubble where people are just good at everything. Or maybe in a bubble where people communicate about these things.
I was quite stupid about these things even in my 30s. When I tried to ask my friends, they also knew nothing. And trying to figure out things on internet has the problem that when you are a beginner, you just find lots of contradictory advice and you don't have the skill to figure out which is trustworthy. (On top of that, most advice is for Americans and does not necessarily apply to me.)
Finance pays out big bonuses annually too. Amount is a function of your own performance, the department performance, and the firm performance. It's usually not contractually guaranteed, so it's not always very transparent what amount to expect, but there's usually some consistency to them which you can learn through research.
I don't know as much about other industries, but I think consulting also has significant bonuses. Also I think many sales-type roles use performance bonuses as a substitute for commissions.
Moving beyond the traditional 5 x 8 work schedule has its advantages. I work for, uh, an Extremely Large Online Retailer and the standard full time workweek is four 10-hour days, though there are some variations. From what I understand this is because most facilities are 24/7 and four 10's is an easier way to maintain staffing.
In any event working a 4 x 10 schedule has many good points. The extra two hours a day go quickly, and you get three full days off a week. Two of those days usually are on weekdays, as almost all full timers work Saturday or Sunday, and weekdays off are ideal for running errands and shopping.
I. Work for Someone Else
*Many possible roles and career paths.
*Key advantage: Someone else pays the bills. Some measure of stability. Often comes with benefits like help paying for some insurance, 401(k) retirement plan, etc. Usually has the best chance of not needing to work more than 40-50 hours per week.
*Disadvantage: Someone else decides what you get paid in terms of hourly, salary, or commission plan. If you want a change, you have to negotiate for it and prove your value or move to another company.
*A. Blue Collar
**1. Build: You make things or create resources for things to be made with. Assembly line at a factory or processing facility (chemical, food processing, etc.), construction worker (concrete, timber, structural, electrical, roofer, etc.), machinist or welder, oilfield or mining work, lumberjack, fisherman, farmer, etc. Highest-risk jobs but some can pay very well for the long hours.
**2. Service: You keep things working or install completed products. Mechanic, air conditioning repair, electrical work, equipment installation & maintenance (A/C, industrial equipment, data center equipment), power line repair, plumber, furniture assembly, janitor, etc. Tends to have stable demand but often requires sweating and crawling or climbing around to do the job.
**3. Logistics: You get things where they need to be. Warehouseman, forklift operator, truck driver, mover, railroad worker, rigger & crane operator, sailor. Pay is often not great aside from specialized technical roles like rigging.
**4. People Care: You provide direct one-to-one and hands-on services to people, not things. Nurse, teacher, police officer, EMS, daycare, firefighter. (some of these may be classed more as 'pink collar,' but they are hands-on jobs, not desk jobs)
*B. White Collar
**1. Customer Service: You work with customers on a day-to-day basis. This can be for B2C or B2B. You answer phones and e-mails, convey product knowledge, help customers make choices, take orders, handle scheduling, appointments, orders, shipments, and jobs, deal with first-line customer complaints, etc. Your job is basically moving information around politely so that customers get what they need and the company makes money meeting the customer’s needs. Usually hourly. Often entry-level or first jobs. Some Customer Service jobs are more blue-collar – think bagger & cart handler at a grocery store, gas station attendant, etc.
**2. Sales: This is like customer service in many ways, but you are working with customers who are not customers yet. They either don’t know if they need what your company provides, or they are not sure if you are who they want to buy from. Your job is to help your customers by matching your company to them. You are often paid on commission, so this can pay very well, but if you have no sales in a week, you have no paycheck that week. You deal with having your calls and messages ignored often and being rejected or told no a lot. If you are in B2C, you are often meeting someone one time and have 30-60 minutes to get them to trust you and buy from you. If you are in B2B, it can take anywhere from 3 months to 2 years to make a sale depending on what you are doing. People buy from your company because they trust YOU and that YOU know what you are doing and are steering them in the right direction. You don’t have to be an extrovert, but you do have to be likeable and persistent. There are many ways to sell, from a more technical product-oriented approach to having an internet presence and followers to being a professional networker who is friends with everyone, plays golf with customers, takes them out to dinner etc. In addition to traditional “sell a product or service,” this path includes real estate agents.
**3. Technical: You provide a specialized service, often after “sales” or “customer service” has been involved. Your value is your expertise, experience, and knowledge. Architects, engineers, doctors, lawyers, aircraft pilots, accountants, actuaries and mathematicians, psychologists & counselors, industrial designers, professors, consultants who offer advice or design work in specific fields, IT (information technology – computers, programming, data centers). Can be very highly paid, but often only after an apprenticeship or extended period to gain experience. More likely to be contractors.
**4. Management: You’re in charge of customer service, sales, the technical branches, and any blue-collar roles your organization has. You are responsible for getting everyone to follow procedures, work together reasonably well, deliver service or product on time, and generally keeping things moving and profitable. You also get problems or complaints that others can’t solve escalated to you, and are responsible for dealing with any internal problems like people who irritate each other or don’t want to do what they’re supposed to do. The pay is often good but sometimes the hours are long. At higher levels, you manage managers and work your way up to C-level (Chief X Officer) and run the whole company, which typically pays very well. However, if there are problems, the buck stops with management, so job security may not be as great unless your company is stable and their industry is stable. In addition to managing departments or divisions in a company, there’s also a sub-role called “project manager” where you oversee a specific job or task, like building a building or completing a software program. It has more limited scope in terms of responsibilities, and involves tracking details, progress, and deadlines from dozens to hundreds of people or suppliers and making it all work together smoothly.
II. Contractor or Hired Specialist
*A. Solo Operator: You don’t have anyone who works for you. This is just you and what you can do delivering value to your clients.
*B. Career field: This is almost always B2B. This can be blue collar or white collar, but typically involves you having a skill at which you are specialized and experienced, with several years to a decade of experience in. Companies bring you on for that specific skill and pay you part-time, full-time, or by the job to do it. Payroll for small companies, on-retainer lawyer, consultant, doctor (many hospitals don’t employ their doctors), programmer, project manager, trainer, home inspector, insurance adjustor/inspector etc.
**1. Entry: You normally work in your field for a company or for someone else for a while to develop your expertise and experience.
**2. Business: You gain business through your network – people you know who know you – and your reputation. You have to sell your services, making this a “sales” role part-time by default in many cases.
III. Own Your Own Business
*A. Money: You own the business. If it makes money and grows, you can pay yourself from the profits. Owning a successful business can pay very well, and it can grow to become something that can support future generations. Alternatively, you can build the business with the intent to sell it and walk away from it with a full bank account. Some people are serial entrepreneurs, meaning they start a company, build it to success, sell it, and then start another company.
*B. Risk: Many small businesses fail. If there is no profit, there is no take-home pay for the owner. If your business fails, the people you hired to work for you are out of a job.
*C. Responsibility: You carry legal liability. Early on, you are sales, accounting, human resources, operations management, training, and project lead on everything all at once all the time. Expect to work 60 to 80 hours per week.
*D. Scale: “Scale” refers to how much a business can grow. Some businesses have more growth potential than others. Some businesses are easier to grow than others in terms of potential customers, investments needed, etc., although this can also be a defense against competition. Make a piece of software that everyone wants on their phone: 100 million to 500 million sales, IF you’re part of the 0.01% that actually make it to that level of popularity. Make a restaurant? It can only seat so many people and service so many customers per day. Open a warehouse? Eventually you run out of space. Make a widget sold to air conditioning manufacturers? There are only so many manufacturers in the US. Sell to all of them, and you need to branch out if you want to grow further. You cannot scale or sell a business if it depends on you to continue to function. If it falls apart if you’re not present, you’re self-employed with extra expenses. One restaurant that is only in good shape if you are there will never become five. One restaurant that functions with you only stopping in twice per week is easy to turn into two, four, or ten if you have the money, the customers, and can hire the right people.
--
I'm sure I could go in depth with more examples of jobs. Are there any other important "this dictates what your job feels like for 40 years" meta-categories worth explaining to a preteen/teenager?
How about "The Military" -- it's distinct from government service, and really does pay logisticians well (not blue collar there, no siree!)
I really like this breakdown and would love to see it hosted somewhere with better formatting tools.
I would add a dimension for creative vs. deterministic pursuits, since having subjective vs. objective success criteria can make a big difference in job satisfaction for different sorts of people. An in-house designer has very different workflow than an in-house accountant, even though they both spend all day in front of a desk and report to a boss.
How's this?
F. Style of Work.
*A. Deterministic. Your job follows rules and processes for specific, set outcomes. Your job involves following and improving those processes. There’s room for skill, but you’re usually not making anything new. Examples: Accounting, construction, some engineering work, factory work, medical, transportation, most IT work. This is more common with larger companies.
*B. Creative. Your job has rules and processes, but a key part of your job involves using your taste, judgment, or artistic skill. Examples: Architecture, art, designer, some engineering work, public speaking, sales, teaching, writing. This is more common with smaller companies or solo operations.
Looks good. For the sake of the future teen reading this, be very clear that success in a creative job doesn’t depend on _your_ subjective taste, but on the boss’ or client’s taste
And that you can be forced to write/make Nazi-slogans, or other extremely impalatable stuff. (Yes, still salty about cakes. Try talking to creatives...)-- this is American-based, if you're not.
Given that you mention benefits, and have a whole section on compensation, it's probably worthwhile to have a list of the kind of benefits one normally gets from their employer
Good idea. Added to my copy. Insurance, flex spending, 401k, sick pay, vacation pay. That's about all I'm aware of aside from ancillary options (supplemental insurance, small life ins policy, etc.)
I would add a third customer type:
- Business to Government: similar to B2B in many ways, but often defined more by adherence to complex regulations and processes than by the free market.
This was interesting. A few bits of feedback:
* No mention of piecework in the "how you get paid" section
* No mention anywhere of the hospitality industry - where would you put that?
* Would be interesting to add the distinction between working for/with private enterprise, state infrastructure (military, civil service, health service in the UK, etc) and third sector / not-for-profit enterprise
I was thinking about using this with my teenager but I have an overriding worry that it reflects more how the world WAS (and still is to some extent) and less about how the world WILL BE by the time they come into the workforce. What about which careers are more or less at risk from AI? The importance of transferable skills that aren't tied to any one specific job or industry?
I would lump piecework in with commission. In both cases it's pay based on results, not time.
I would put Hospitality (both hotels and restaurants) under Customer Service, although they don't fit neatly into blue collar or white collar. The line cook & room cleaner are blue collar, the check-in desk at the hotel is white collar (mostly) and I'm not sure where the intermediate roles belong. I used blue and white because they're widely used descriptive phrases but neither one is a good fit for a waiter/waitress, cashier at a store, etc.
I think B2B and B2G are generally similar enough to go in the same category... although military is its own thing in terms of "blue collar."
NGOs still tend to have the same job types, just different "outputs." Is there something else that would differentiate them?
There are a lot of roles that I think are mostly safe from AI, but I think that'll be something that plays out over time.
Agreed piecework is a sort of similar to commission, but your pay doesn't depend on revenue or profit - purely on your productivity.
I didn't mean B2B vs B2G, I meant being employed by the Govt vs by a company (for-profit or not). The difference for me in terms of careers is:
* private enterprise: if the business is successful, your financial reward generally depends on the impact you have (in many roles the sky is the limit) and how easy you would be to replace; if the business is unsuccessful, you're out of work. Very much part of the capitalist system.
* government: generally more stability and security, but less opportunity to demonstrate growth and much tighter limits on how much you can end up earning. Subject to the whims of government budgets etc which have effectively no relationship to how well you or your team are doing at their job.
* not-for-profit: often not well paid at all, but much more rewarding in other dimensions like feeling that your work is making a real difference in the world.
The other thing I've tried to emphasise to my teenager is that there are so many roles that aren't "obvious" kind of "Richard Scarry What Do People Do All Day" type jobs, things you'd never think of as a child unless you knew someone in the relevant type of role. To name just a few completely random examples based on people I know: police intelligence analyst, care quality inspector of hospitals, TV broadcast controller....
Somewhat surprisingly, hotel desk clerks are generally not well paid. It’s more or less an entry level job.
Glad to see Gunflint is back! I haven't seen Tyrone Slothrop in a couple days, though 😢
Pretty sure Scott had Tyrone liquidated. Banning just wasn’t sufficient in a case like that.
What did Tyrone do? He always struck me as reasonable.
I'm not sure we're even allowed to mention it...
Oh…
Don’t you remember what Dr Evil did to the therapy group in which he and son Scott Evil participated?
https://m.youtube.com/watch?v=zovylDylvSA&ra=m
OK, that gives me some idea. Can you direct me to the thread where he got the bullet? I’m just surprised because I had a few very pleasant exchanges with him
Any opinions on the yotube channel ordinary things?
And related: any recommendations for youtube channels?
Asking for "youtube channels" with no further specification is like asking how a watch works.
Based on your mention of Ordinary Things and the general nerdy atmosphere here, my usual recommends are Veritasium, SmarterEveryDay, and PracticalEngineeringChannel.
Kitboga if you don't already know him. Also Crazy Love.
Good opportunity to remember why supporting zoonosis is irrational and based on basic mistakes in probabilistic inference:
https://blog.rootclaim.com/covid-origins-debate-response-to-scott-alexander/
Peter and Scott - Congratulations on winning the bet. Incidentally, when I heard about it, I asked to participate on the zoonosis side, as this was clearly not a bet that is well correlated with truth.
Your post remains as wrong as it was when it first came out. The evidence for zoonosis is overwhelming and for lab leak, practically non existent. Also, Peter's not a debate savant, and the one making fundamental errors in probability (https://royalsocietypublishing.org/rspa/article/475/2227/20180565/80107/Satellite-conjunction-analysis-and-the-false) is you.
edit: If anyone comes upon this thread later, don't bother reading it or trying to engage with Saar. He's a classic internet crank. Attempting to argue is futile, as he'll demand you answer his questions without answering yours, impose conditions on continuing when things don't go his way, assume that everyone thinks and behaves like he does, engage in wild inconsistency, display massively inflated self-importance, insist that his is the only right way to do anything, assume the only reason anyone ever disagrees with him is because they're fools who make basic mistakes, ignore most of your arguments, general rude comments and obvious hypocrisy, write 20 wrong comments while asserting that you are wrong, not provide legitimate sources, and in this case, use AI to respond to AI summaries of what you say, rather than reading your comments and responding himself. Just all around clown behavior.
If you can point to any specific mistake in the post, please share it.
Most importantly, we point to one mistake that, once corrected, flips most analyses (including Scott's) from zoonosis to lab-leak: It is highly irrational and bad probabilistic inference to claim >99.9% confidence without very high-quality large-sample statistics, and we don't have anything close to that on the Huanan seafood market.
Not only that, nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location. Claiming >99.9% confidence on something that leading scientists are divided on requires a very good reason. None was presented.
More here: https://x.com/Rootclaim/status/1991542056049648070
Do you disagree?
> nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location.
I asked ChatGPT and it said this percentage wasn’t accurate. Your quote conflates “zoonosis” with “the first infection definitely occurred inside Huanan market”, obviously if the zoonotic scientists believe if it didn’t happen there it happened upstream of the market, ie a supplier was infected.
This advances the lab leak theory not at all. I’m assigning zero, or very very low, probability to some chap infecting himself with the virus, not getting sick himself, and not infecting his colleagues or anyone close by; nor his family or neighbourhood, and then visiting the market. There he does infect someone else and the market acts as an amplifier (which would be true of any mall or market where he infected someone, most likely an employee).
We know from genomic evidence that the pandemic MRCA is the same as the market MRCA. That means, the market ancestor is at the root of the pandemic. Even if early cases elsewhere were missed, you have to argue that the spillovers elsewhere died out without leaving any traces via other lineages that spread worldwide and miraculously only the market ones survived. The Bayes factor for this should heavily favor natural origins.
Under the leading lab-leak hypothesis lineage A was spreading in Wuhan, developed to lineage B within a few weeks, infected someone in the market and got amplified. No miraculous disappearance of earlier strains is claimed.
It is telling that Judge Will described this scenario exactly as understood from your Rootclaim debate and found that it had no support as ZO presented good arguments for lack of ascertainment bias while you had none.
Why no other lineages in the meantime due to normal mutational processes that spread elsewhere? You do know that there are those arguing against the double spillover because of the mutational clock? Why does lineage A stay intact until it reaches the market area? Why are lab leakers arguing that the virus was pre-adapted for human spread and also saying that the pandemic didn't take off for weeks until it reached the market? Why does healthcare worker serology tests (which had a high ascertainment rate due to cash rewards from Bytedance) spread outwards from the market? You have no evidence via serology, mortality, MRCA analysis, healthcare workers tests, hospital records, case records. Simply postulating something with no evidence is only motivated reasoning and therefore gets a very low Bayes factor.
You did not understand my point on 99.9%. Please read the relevant parts of the post again.
Your hypothesis is a strawman. The leading scenario is that they infected several people who infected others etc, and then one of them reached the market weeks later, which is evident by all market cases being of the later B lineage. More here: https://x.com/Rootclaim/status/1753353163178913815
Having said that, your strawman is still enough to flip the analysis to lab leak. Read the "CDC Move" section in the post
https://blog.rootclaim.com/covid-origins-debate-response-to-scott-alexander/
> If you can point to any specific mistake in the post, please share it.
For example:
> There is basically no evidence for zoonosis other than the results of repeating the following process:
This is incorrect. During the debate, Peter presented many maps of cases and related data (such as deaths) that were simply visualizations of raw data. No model is necessary to, for example, look at such a map and see that cases (both market linked and non) centered on HSM. Or look at the first handful of cases and see how they are heavily concentrated in HSM.
The post as a whole is in fact riddled with major errors, but first:
> Not only that, nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location.
Source? I cannot find any survey like this on Google. Just ones that say, over and over, that the large majority of professional virologists and epidemiologists believe zoonosis is more likely.
> Claiming >99.9% confidence on something that leading scientists are divided on
...leading scientists are *actively opposed* to lab leak o.0
Those who don't believe HSM is the spillover location still think it's the main amplification site, so the maps are not evidence for either hypothesis. If China would publish full contact tracing for HSM like they did for Xinfadi, then it would be strong evidence to one of the hypotheses. It's fair to assume they very likely did do that but chose not to share the result - and that IS evidence.
The twitter thread I quoted provide evidence for HSM not popular as a spillover site. In any case, it doesn't have to be a majority. Claiming >99.9% confidence on anything that top scientists are contested about without a very good reason is irrational.
I stand by the claim that Zoonosis has exactly zero good evidence.
>leading scientists are *actively opposed* to lab leak
They're not. They lean zoonosis but don't rule out lab leak. Anyway, a) we don't claim anything close to >99.9%, b) we have a very good reason.
> Those who don't believe HSM is the spillover location still think it's the main amplification site, so the maps are not evidence for either hypothesis.
More errors.
First, you seem to be the among the few lab leakers who believe this. More importantly, the fact that you can *theorize* that it is the main amplification site doesn't make "zoonotic origin related to the market" and "lab leak but all early cases are strongly associated with the market" equally likely.
> It's fair to assume
An assumption is not evidence.
> The twitter thread I quoted provide evidence for HSM not popular as a spillover site.
Can you provide a link to the specific source for the claim that " nearly all leading zoonotic-leaning scientists don't think the HSM is the spillover location."? I don't have an X account and the format of "20 one sentence tweets" is annoying to read, and I'm not going to do the work of justifying a claim that *you* made.
> I stand by the claim that Zoonosis has exactly zero good evidence.
You can stand by it, but it's still wrong. Previous lab leaks overwhelmingly had cases associated with the lab (that's often *how* they were identified as lab leaks in the first place). This is possibly the only piece of evidence where the relevant rates can reliably be estimated with past data, rather than being based on guesswork.
> They lean zoonosis but don't rule out lab leak.
Come on Saar, you understand the difference between this and what I wrote. You can't rule out zoonosis either (you say 6%, roughly) but also say that "supporting zoonosis is irrational." Scientists don't say lab is entirely impossible but pretty clearly think it's much less likely.
> Anyway, a) we don't claim anything close to >99.9%, b) we have a very good reason.
"We have good reason, the other side does not" is simply circular.
You're avoiding the main issue and scattering to minor ones. Before answering those, I want to first close this: Do you agree that if prominent scientists are divided on an issue, then someone who has no expertise claiming >99.9% confidence is being irrational?
Given that you don't want to put the time to read the evidence, I'll give just two that are enough to convince of "scientists are divided", rather than my stronger claim of "most scientists".
1. WHO report:
"Many of the early cases were associated with the Huanan market, but a similar number of cases were associated with other markets and some were not associated with any markets. Transmission within the wider community in December could account for cases not associated with the Huanan market which, together with the presence of early cases not associated with that market, could suggest that the Huanan market was not the original source of the outbreak... No firm conclusion therefore about the role of the Huanan market in the origin of the outbreak, or how the infection was introduced into the market, can currently be drawn."
https://who.int/docs/default-source/coronaviruse/final-joint-report_origins-studies-6-april-201.pdf (page 7)
2. US intelligence assessment:
In the section arguing for zoonosis, the first cluster at HSM isn’t even mentioned as evidence, and they specifically cast doubt on its role.
"some [scientists] now view the market as a potential site of community spread rather than where the initial human infection may have occurred."
Makes sense? Can I move to the next step?
I'm charitable towards the weak form of lab leak (not "engineered on purpose" or even "people were doing engineering of some kind", but "someone had it because they were passively studying it in good faith and a mistake happened") but I instantly disregard any argument that "believing side X in a controversy is irrational". If your arguments are good you don't have to try to emotionally stack the deck.
It's especially funny in this case because Saar said described a poll of superforecasters as "clearly not a bet that is well correlated with truth" when the superforecasters have a demonstrated, externally-validated track record of reasoning under uncertainty, while Rootclaim... does not.
Note you're attacking Scott here. He's agreeing this bet was mostly won on reasons unrelated to truth.
What Scott said:
> I’ll admit that my bet was only partly on the merits of the COVID origins argument, and at least as much on the thesis that new and earth-shakingly decisive evidence on controversial events rarely comes out in the specific interval of 4-6 years after the event happened
What you said:
> clearly not a bet that is well correlated with truth
These are not the same thing.
I guess it depends on the definition of 'well correlated', but let's not do semantics. We both understand the point.
What specific point is that paper supposed to demonstrate? It seems like it kicked off a whole set of very technical discussions, and I can't tell what the relevance is here.
Response 1 (furthering the argument in the paper you referenced, specifically extending it):
https://royalsocietypublishing.org/rspa/article/476/2237/20190781/56889/Confidence-in-confidence-distributions-Confidence
Response 2 (a response to response 1):
https://royalsocietypublishing.org/rspa/article/477/2250/20200579/82345/Response-to-the-comment-Confidence-in-confidence
Response 3 (by the same authors as the original paper, clarifying the specific conditions under which things apply):
https://arxiv.org/html/2404.16228v1
Response 4 (general sweep of argument, proposing an alternative argument that supports the same general conclusions as the original paper by arguing about the coherence of point probability regions in confidence distributions):
https://www.mdpi.com/1099-4300/27/4/329
It seems pretty complicated. I'm trying to figure out the actual argument because it seems like a philosophically interesting place where frequentists and bayesians can meaningfully disagree
The most important part is the distinction between epistemic and aleatory uncertainty, and how treating them the same way can result in huge errors:
> False confidence is the inevitable result of treating epistemic uncertainty as though it were aleatory variability.
> The final clue indicating the inappropriateness of probability dilution is the fact that, for a fixed S/R ratio, there is a maximum computable epistemic probability of collision. Whether or not the two satellites are on a collision course, no matter what the data indicate, the analyst will have a minimum confidence that the two satellites will not collide. That minimum confidence is determined purely by the data quality.
> Since the uncertainty in the estimates of the trajectories reflects the limits of data quality, probability dilution seems to imply that lowering data quality makes satellites safer. That implication is counterintuitive in the extreme [4–8]. As a rule, lowering the data quality makes any engineering system less safe, and to claim that ignorance somehow reduces collision risk seems foolish on its face.
> Interestingly, were the uncertainty in the satellite trajectories aleatory, rather than epistemic, probability dilution would actually make sense. For example, suppose two satellites were known, with certainty, to be on a collision course, and the satellite operator could only impart an impulse of random magnitude in a random direction, say, via a poorly controlled manoeuvring thruster on a tumbling satellite. If the mean of this distribution were the null vector, due to a lack of control over thrust direction, then the higher the variance of the added impulse, the bigger the resulting perturbation and hence the smaller the resulting probability of collision.
Moreover, the resulting errors are not bounded:
> As relative uncertainty grows, the level of false confidence seen in satellite conjunction analysis grows arbitrarily severe. Next, we show that, when considered over the full range of measurable propositions that the analyst might assess, most epistemic probability distributions used in practice suffer from arbitrarily severe false confidence.
In other words, there is a big difference between the following 2 statements:
1. Out of 10,000 cases with similar data, satellites collide about 0.1% of the time.
2. We don't have very good data, so it would be overconfident to say that the satellites are more than 0.1% likely to collide (since if you pick 2 satellites at random, they are unlikely to collide soon).
In the latter case, it might actually be the case that the satellites are nearly guaranteed to collide.
Rootclaim's analysis very explicitly makes this error. For example, looking at their most recently analysis (https://cards.rootclaim.com/analysis/65892f07d97bdb00199d4366) the single biggest Bayes factor comes from "12 nucleotides clean insertion," which they say is 10% likely under lab leak and 0.5% under zoonosis. How do they justify the former probability?
> In the past, FCSs have been added by substitution rather than insertion, but it is not difficult to do it by insertion. We cannot be sure of the exact considerations of the lab researchers who added the FCS , such as investigating the role of proline. Therefore, we assign it a 10% probability.
This is precisely the sort of epistemic uncertainty that, according to the FCT, results in errors of arbitrary magnitude! They do not have a single scrap of data to indicate that, 10 times out of 100, in similar situations, lab researches would use a 12 NT insert. They just say that it would be overconfident to assert the value is smaller than 10%. The true value, for all they know, could be 1%, or 0.1%, etc. In fact, parallel to the quoted paragraph above, if one *did* have such data, and decided to throw it away and use Rootclaim's method, the reduced data quality would result in *more* confidence in lab leak (since the resulting BF in favor would be higher).
Yeah, but it seems like you get the same false confidence errors using frequentist frameworks as well (see response 2):
> However, the message presented by Cunen et al (that replacing a Bayesian posterior distribution with a CD is all it takes to overcome false confidence) is potentially misleading. Our false confidence theorem applies to all epistemic probability distributions, including CDs; so, contrary to the authors’ claim, their proposed CD is at risk of false confidence too.
And the authors of the paper you cite eventually conclude that the satellite problem is not caused by the usage of epistemic uncertainty but instead by the specific problem formulation (the false confidence issue arises from considering the complements of a convex hypothesis). It's not a generic fact about epistemic probabilities that false confidence goes up with bad data. See response 3 by the authors of your cited paper:
> The false confidence theorem establishes that, for ANY data-driven, precise-probabilistic method for uncertainty quantification, there exists (non-trivial) false hypotheses to which the method tends to assign high confidence. [...] Roughly, complements of convex hypotheses are afflicted by false confidence.
To clarify, I am in no way saying that the rootclaim probability is accurate (it seems like they just made it up?). I haven't looked into it too thoroughly, but I lean zoonotic. I'm just not sure how that paper you posted maps given that it's about a very specific situation (the "false confidence" comes from the geometry of the hypothesis). AFAICT, the discussion seems to have concluded in:
"Any precise additive distribution over a continuous parameter space, Bayesian or Frequentist, systematically over-supports false hypotheses that are complements of convex sets (or in general, nonlinear sets)"
Which doesn't seem to apply to the nucleotide insertion thing (which is too simple to have the kind of geometry that systematically leads to these false confidences).
This whole thing is very interesting. The bayesian relevance seems to be something like: If you don't have a real prior on a parameter's value and you default to a "flat prior", that "flatness" is only true relative for the coordinates you originally estimated in. If you use those parameters to generate other coordinates, that "flatness" causes an opinionated non-flat probability distribution for the set you care about. Because Bayesian probabilities are additive, probability of not-P is 1 - probability(P), so these distortions create "phantom evidence" that artificially slant the analysis one way.
For the satellite thing specifically, you definite "collision" as "the relative distance between satellites is inside this small disc" and "no collision" is the whole curved space outside that disc. So your probability estimations are on "how far away in the X direction are the satellites" and "how far in the Y direction are the satellites" (flat) but your hypothesis defines a collision using the euclidean distance between them, which is nonlinear "sqrt(Δx² + Δy²)" which introduces distorting phantom evidence (because altering the probability of collision implies altering the probability of safety, there is no explicit third option to absorb the uncertainty)
I don't think I said that frequentist frameworks avoid the problem, nor was that the point. The point of my example was the existence of quality data, not how the parameter "frequency of lab using insertion" is interpreted. As you quote,
> The false confidence theorem [2, 22] says that, for any precise IM... there exists false hypotheses H to which the posterior tends to assign a (relative to α) large probability/confidence
i.e. false confidence. Here, a "precise" inferential method is one with countable additivity, which is what implies P(not A) = 1- P(A). I think this is precisely the issue with "overconfidence-based" probabilities, like those utilized by Rootclaim (and Michael Weissmann): if P(A) is bounded from above because of worries about overconfidence, P(not A) is similarly bounded from below.
This is basically what you are saying here:
> altering the probability of collision implies altering the probability of safety, there is no explicit third option to absorb the uncertainty
A parameter like P(lab uses 12 NT insertion), with no data to go on, should be the subject of high uncertainty. Rootclaim *says* they account for this uncertainty, but their methods are all based on "precise" IMs, not the 3rd option of "explicit uncertainty" that you mention, and so do not resolve FC.
> Which doesn't seem to apply to the nucleotide insertion thing (which is too simple to have the kind of geometry that systematically leads to these false confidences).
I don't think you need complicated geometry to have convexity. In fact classic examples (one would even argue, the motivation for the entire notion of convexity) include simple one-variable functions like f(x) = x^2. If I'm understanding the math right, then it very easily could apply to an "overconfidence adjustment" that shrinks the parameter (probability of lab using this insertion) toward 50%.
The convexity thing provides further information about *which* hypotheses are effected by false confidence, but the implication is co-convex => FC, which doesn't do anything to say that a given hypothesis is *not* afflicted by FC. At the very least, if Rootclaim wants to use this methodology, they should show that it meets the "linear hypothesis" standard or otherwise avoids false confidence. And, just intuitively, without the math, it seems obvious to say that your strongest piece of evidence shouldn't be based on a guess with no data.
Agree with everything you just said. These sentence are what made me believe you were using the paper to generally criticize the formal emphasis of epistemic vs aleatoric probabilities.
> The most important part is the distinction between epistemic and aleatory uncertainty, and how treating them the same way can result in huge error
and
> This is precisely the sort of epistemic uncertainty that, according to the FCT, results in errors of arbitrary magnitude!
Whereas the actual target of criticism seems to be unreflective usage of precise additive inferential methods (for which certain false hypotheses assigned relatively high probabilities are guaranteed to exist).
I totally agree with you that Rootclaim is not doing a proper analysis. Their issues seem more on the order of "literally just guessing important things" which would disqualify it before you even get to the level of hypothesis convexity (which I haven't looked into specifically here, but it's a subtle issue that it's unlikely they properly account for).
Thank you for linking that study! It's been very interesting to think about, and I would have never stumbled upon it myself.
The Aug. 8 New York Times has an article on A.I. in China. I’ll summarize since it’s behind a paywall.
1. Chinese A.I. companies do marketing in Africa; American companies mostly don’t. Chinese companies will reach out to African companies, offer them free computing resources and help setting up their A.I. systems.
2. African countries are not particularly wealthy, and don’t want to pay for frontier models unless they need them. At one point, Meta’s Llama filled the need for an inexpensive model, but Llama 4 performed less well than many expected, and Meta decided to switch to closed weight models rather than create another version of Llama, leaving an opening for Chinese models.
3. There are tasks where Chinese models aren’t good enough yet. For example, Chanderia Industries an OpenAI model to read product orders from retailers and enter the orders into its system for fulfillment. (The retailers send in order via E-mail or paper mail.) Bytedance convinced Chanderia’s provider, Sanifu, to test their model, but it made too many errors.
Another company described Claude Code as expensive but worth it.
4. It appears that fine tuning open weight models for particular use cases is fairly common in Africa, which probably reduces the need for cutting edge models.
5. Shutdowns. In 2024, Deepseek shut down its servers before the Chinese college enterance exams to prevent students from using the service to cheat, which disrupted some African businesses. More recently, Anthropic blocked access to Fable at the request of the Trump Administration. Open weight models win here because if your provider cuts you off you can continue to use the same model with a different provider.
6. Bias. Sunflower uses a tuned version of Qwen 3, which Sunbird chose over Llama and Gemma because it was better at handling African languages. (I wouldn’t be surprised if Alibaba made a greater effort than American companies to ensure that African languages were well represented in their training data.) However, China Media Project concluded that Sunflower was biased towards China; you can read the details here: https://chinamediaproject.org/2025/12/17/the-chinese-core-of-ugandas-chatgpt/
As a result, Sunbird is using Gemma (open weight model from Google) for its next project.
My thoughts: The open weight models seem to be quite beneficial to third world countries. Scott’s recent article titled “Open Questions On Open Weights” suggests that for now there’s not too much reason to ban open weight models. I do worry a bit that this is another example of where China has a national strategy and the United States, at least under the current Administration, has none.
Disappointing to hear Llama won't have a follow up. My presumption was that data security, regulations, insurance requirements and vendor contracts would all push towards locally-hosted tunable open weights models in many industries. It seems intuitive to me that this will need to exist in some form because the existing legal landscape demands it, but America isn't going to adopt Chinese models and in some cases probably isn't legally allowed to either. What's the business solution to this going to be?
Meta actually released a medium-sized open-weight model yesterday, and says that a larger one is coming: https://x.com/alexandr_wang/status/2086756152034066792
Which industries do you have in mind as ones that require self-hosted models? I think that's mostly only true of national-security-related stuff.
I'm a lawyer presently in financial services with a lot of bank clients and borrower info, at present I am continuing to advise management they cannot use anything requiring remote API. I would need affirmative statements from banking regulators, the insurance industry, the bar association, and all of our clients before I would consider allowing any usage of remotely-hosted AI systems. I think it would likely breach vendor agreements we've already signed with the bank clients, by terms the bank is required to mandate of their contract partners by the bank's own insurance companies and their regulators.
'scuse an ignorant question: Why is remote API in particular a deal breaker? I could see AI _unreliability_ being a deal breaker, but isn't some conventional financial software like ADP's tax and payroll services hosted remotely?
ADP is handling YOUR company's employees data, under a contract you have with them. If there's a data breach at ADP, you are the client and their incident response team would handle the fallout and their insurer would cover it and your own insurer is fine with this.
But if my business requires me to store and use the information of tens of thousands of bank customers with whom I have no direct relationship, there's many more things in play. The bank's regulators and insurers will require it to handle consumer data with certain precautions in order to be covered, and in event of a data breach their standard e&o policy may not cover it so they probably have a rider for that too. And part of the requirement of those regulators and insurers is that any vendor agreement, the contracts made with other entities (like me) that handle the bank's data, made between the bank and other entities must require THOSE entities to also have $X in cyber security coverage and a laundry list of security protocols.
If we started sending consumer info to remotely hosted LLMs and there was a breach, we might theoretically have a cause of action against Open AI or whomever but the immediate problem would be that our e&o and cyber security policies won't cover us, the 2nd order problem is that we'd have breached the contract with our bank client who is now in a crapton of trouble, we're now contractually obligated to spend millions of dollars in incident response out of pocket and have probably lost an important client and are unlikely to survive all these dominoes falling as a small to mid-size company.
This is why I cannot use remote API right now. At some point if locally hosted LLMs dry up and everyone wants to use remotely hosted AI then what will happen is the institutional players will see the value in it as necessary and will converge upon some set of protocols. Everyone in this process will attempt to extract a bit of money or market position out of it: insurance gets some new type of policy you'll have to buy, some company will wind up with the deal to give ratings to the security of various AI companies to provide to the insurance companies, banking regulators will be persuaded by well-positioned providers to permit certain types of services and not others, since the business also involves litigation the malpractice insurers will end up saying you have to use the Thomson-Reuters AI which won't meet the other requirements so some facilitator springs up to bridge that gap until OpenAI buys it, etc etc. All these dominoes have to fall before legacy institutions can do this, so yeah just run a dang Llama in the closet.
What's the current state of what major LLM vendors offer their enterprise customers in terms of indemnification and insurance coverage and whatnot? You would think that they'd be trying aggressively to remove whatever barriers are in the way of enterprise sales, and at least some of them are or have close partnerships with major cloud vendors that have been dealing with this sort of thing for a long time. Or is the problem that your own insurer will go "that agreement isn't worth the paper it's printed on"?
Many Thanks for the detailed and informative explanation!
If you've signed a contract saying you won't share certain data with people who haven't signed the contract, you can't then send that data to machines owned and run by people who haven't signed the contract.
Many Thanks!
Anything on-board. Autonomous vehicles come to mind.
I don't think AVs are an example of what Cjw is talking about. Those models are highly bespoke and unlikely to be released with open weights, nor would it be particularly useful if they were.
I wonder how much the Chinese government determines the AI strategy of their labs. On the one hand, all the Chinese labs are pursuing a similar strategy, so it could be government-directed. But that strategy could also be their own decision: if your models are behind the American ones, making them open-weight is an advantage to get more customers. Very interested to hear if anyone knows more about this.
Like you say, the American government definitely doesn't have an AI strategy. But that does allow a diversity of lab strategies. For example, Meta just announced they're recommitting to open weights: https://www.reuters.com/world/china/meta-launches-new-ai-model-zuckerberg-champions-open-weight-push-2026-08-10/.
In a normal environment this diversity is good and will have the best strategy prevail. With AI however I do hope the administration intervenes soon as the wrong strategy can lead to very bad outcomes. (More referring to strategies on training, scaling & alignment, Scott made a good case for why you don't need to spend political capital on banning open weights now.)
btw the Frikgeek link only works for me if you left click and open in a new tab
Should consciousness matter to a preference utilitarian? I’m not sure why. Bacteria may not feel pain but they have preferences.
No. Consciousness is an undefined term, or at least there's no common agreement on exactly what it means. I know people who won't accept that a dog is conscious. The decisions that matter should be based on observables, where there is at least the possibility of common agreement.
Utilitarians should just pivot to anthropocentrism. It solves the overwhelming majority of gotcha-questions that get thrown at them in discourse. It eliminates a whole bunch of pontificating on things they have very little insight towards and huge epistemic limitations about such as the preferences of bacteria or computers. And after all what is the point of ethics? Only humans can engage in ethics, it is a human doctrine, it's only natural that its end should be anthropocentric.
In this model, animal suffering is not irrelevant, but it's only relevant obliquely by way of it causing distress to humans. Dog suffering causes humans distress, but nobody loses sleep about bacteria. Simple, we're done, that's all we needed.
Do you expect anyone who's currently concerned about animal welfare to find this persuasive?
isn't the fact that they (presumably humans) are unpersuaded evidence that an anthropocentric ethics can make strong demands about animal welfare
I'm not sure I understand the point you're making, but in any case, I'm not arguing that an anthropocentric ethics *can't* end up caring about animal welfare (I'm somewhat agnostic on that), I'm arguing that this doesn't address or engage with any of the arguments in favor of a non-anthropocentric ethics in and of itself.
Maybe, I think a lot of people seem compelled to have a single Grand Unified Theory of Ethics that they have to shoehorn everything into that they intuitively see as important to care about. But you do not have to shoehorn animal welfare into utilitarianism, it just causes problems for both.
There's no real reason we could not say both:
1. utilitarian ethics are a system of evaluating good for large groups of humans and making decisions, which was devised by humans and does not function correctly outside of application to humans.
2. certain cruelties to animals may be undesirable because they cause harm to humans or for other solid independent reasons such as violating an implied obligation of stewardship or because cruelty is not virtuous.
Being a pluralist on this seems much easier than to continue to keep shoving the square peg in the round hole, in fact it's worse than that, the utilitarian animal ethics people can't even tell what shape the peg is and have an insurmountable epistemic limitation on their ability to discern the shape. There is no reason the system to evaluate our behavior among ourselves would be the same one that governs our behavior towards animals.
Why not go a step further(or as many steps as you want)and care only about a specific group of humans?
That also has the added benefit of having a long historical precedent and being extremely intuitive to people.
If you want to draw your moral line wherever it’s most convenient for you, why stop at humanity?
I personally don’t care about social contracts; I care about suffering. So I find your position extremely unconvincing.
Newtonian physics works on Earth and the moon and the solar system, all the things it was developed to work for and observed that went into its development. It breaks down when you try to apply it to more exotic phenomenon and starts giving you absurd predictions.
While Bentham may well have wanted it to work for animals, the reality is that while it's readily adapted to human populations it has never integrated very well with utilitarianism for a variety of reasons, and attempts to make it work keep spitting out absurd results about universes filled with shrimp or machines requiring all sorts of workarounds to make any sense of them. It's just not meant for this context.
That isn't to say you then have to adopt contractarianism either. What I'm suggesting is a radical ethical pluralism that does not try to account for the varying values that humans have by forcing them all into one grand unified system. People in the world do not form preferences that way, they don't behave that way, their values aren't explicable that way, so why are we insisting that utilitarianism or any other particular system of ethics has to be universally applicable to all situations? My response to "oh well if you are willing to apply it just to humans then why not just to your own city or country?" would be that you may have to argue people out of that line, but that in fact people did argue that line very successfully! That elements of human universalism in ethics have worked fairly well is a point in its favor. That you have continually hit brick wall after brick wall and faced endless arguments ad absurdum in your attempts to loop in non-humans is a sign that you are probably better served with a pluralist model that accounts for animal suffering with a different ethical theory. I see no reason that some human values are best understood and directed by tabulating utils, and some by deontology, and some by virtue ethics.
The thing is, you're just asserting that that there's no good reason to treat nonhuman animals as moral patients. You're not engaging with any of the reasons that pro-animal-welfare ethicists have for doing so.
Plenty of anthropocentric reasons to be concerned about animal welfare.
As a human, I have a preference against witnessing or knowing I've contributed to the needless suffering of animals complex enough to have feelings worth considering (largely mammals, but I don't eat seafood so have never had to confront the question of lobsters, &). I think a system capable of running on factory farming is not a system which is going to treat the humans within it with much more kindness. I think a person who whips outs to death eight hours per day, or whatever they do, is more likely to cultivate the worst parts of their psyche than someone who raises cows on a pasture.
The anthropocentrism also acts as a shield against other insane beliefs, like the notion that LLMs are in any way sentient or in danger of becoming such merely because they meet the extremely low bar to pass as effective at doing what they were programmed to do (mimic human text output).
Consciousness, suffering and so on in most of these conversations are proxies for the real question: "should I care about this action being performed on this entity?"
Different humans care about different things, and have different intuitions; conversations here are an attempt to construct some common ground for people with different intuitions to come to some kind of concensus.
Anthropocentrism meaning we should only care about humans - fine, that's a coherent answer; some - many - most, perhaps - might disagree, but at least what it means, where the boundaries are and how it differs from "I'm bored of talking about this now, let's just agree we each know what we should care about when we see it" is pretty clear.
Anthropocentrism meaning we should care about humans and also nonhuman things humans care about fails to answer or clarify anything at all; since the whole argument started because humans can't agree what nonhuman things to care about, we're left in exactly the same place we were before.
That still leaves my objections that cruelty is corrupting for the people who have to carry it out, and that a system capable of cruelty to mammals is more likely to be capable of cruelty to humans.
Plus the health risks of factory farming.
That’s basically what I think. Animals don’t cooperate(dogs are actually a very notable exception) so why are we spending so much time talking about obligations towards things that are incapable of understanding obligation?
But if you take that view, then it changes how you view animal behavior. People in the past didn’t actually care about animals, so a dog suffering didn’t matter to them. It’s only because we think people should care that we actually do.
There is as much reason to consider reciprocity essential to morality as there is to consider it irrelevant.
People in the past also did not care about humans.
They cared about members of their own tribe.
The fact that so many people today consider humanity their tribe (at least to some extent) and have a hard time grasping that this is not the natural default, is a testament to the success of humanism.
But there is no good reason not to expand that circle further to include nonhuman creatures as well.
They did only care about their tribe, making it impossible to work with other people and they got trounced by those who learned to form larger coalitions. We’re all born with an innate morality that allows us to cooperate. Thats what differentiates us from animals that don’t have this idea of non kin altruism.
Infants don’t cooperate, neither does someone in a coma, but most seem to think they are moral patients?
No one treats infants as full members of the social contract. Otherwise we would constantly be jailing them for assault. They’re prospective members. As for coma patients or whatever, it’s much easier to just default everyone to moral status then try to measure which people have enough agency. On the other hand, it is literally impossible for any shrimp to ever be part of the social contract because they don’t possess the capability.
There's a kind of gray zone for movements toward something desirable or away from something undersirable: tropisms. Should those be considered preferences?
Maybe. I don’t really care about fungi preferences. My point is that a preference utilitarianism might under their own belief system.
So do rocks, technically.
Rocks can't make choices. Bacteria have a preference for nutrients because they go towards it and a dispreference towards toxins because they go away from them. I don't understand why from a preference utilitarian perspective, it doesn't count.
Umnh... are cpu chips rocks? If not, why not? If they are, why do you say they can't make choices.
Actually, that argument can be framed for most crystal structures, e.g. DNA.
If you asserted that "we don't notice the choices that rocks make", I'd agree with you most of the time, and acknowledge that we don't tend to call them choices when they happen.
OTOH, perhaps our disagreement is in precisely what is considered a choice. I tend to consider any "if branch" a choice point, and also any probabilistic divergence in action. Sometimes the "choice" is nearly certain, but if it's not certain, then it's a choice. I don't consider "consciousness" or "intent" to be part of what makes something a choice. So I tend to see "choice" nearly everywhere I look. I can see a reasoned argument that this makes the term nearly useless, but I haven't heard a better precise definition. (And even this one's a bit fuzzy.)
Preference utilitarians differ on how they understand preferences, but almost all will agree that you can't immediately infer something has a preference for nutrients simply because it is disposed to go towards them.
Why does a shrimps behavior count as a preference but bacteria’s doesn’t? If we put humanity on one side where we can state our goals, why we want to achieve them and how we plan on getting them and bacteria on the other side where it simply moves towards and away from things, then shrimp look much more like the latter.
I don't think it's totally obvious that shrimp have preferences in the relevant sense either. (Though it's not the ability to state them that matters on any plausible view).
Ok that’s fair. I was just talking to someone who did hold that view and hedonist and preference utilitarians overlap a lot. It does make me wonder what counts for them as real preferences among different animals, since it’s less clear than the hedonists.
How is that difference from rocks going towards Earth? It's because of physics, but the bacteria also goes towards nutrients because of physics.
If you can’t differentiate between locomotion and gravity acting on an object, I don’t know what to tell you.
Well then I don't know what to tell you either, because I don't see bacterial locomotion as being any different in principle than rock locomotion due to gravity. Rocks follow the gravitational gradient; bacteria follow the chemical gradient. Bacteria don't make choices in any meaningful way any more than rocks do.
nor do they make choices in any less meaningful way than humans
If chemical gradient is just as meaningful as gravitational gradient then nothing we do matters since we’re all chemical reactions.
I’m not sure what “in principle” is doing here. We both have the same principle that bacteria doesn’t matter. But clearly, taking action is a different thing than being acted on. One of them is something internal that you do. The other is something external that happens. I don’t moralize a dogs behavior but I would differentiate between a dog jumping on my things and someone throwing the dog on them.
MacGregor went towards the sheep because of physics. Scott wrote about MacGregor because of physics. I'm writing this because of physics. Well, not because, exactly. . . .
What about a rock in orbit around the sun, with one side slightly darker than the other? Does it have a preference to rotate in a particular way, dictated by its albedo?
I believe Bentham (and likely his Bulldog too) would say that what matters is the ability to *feel suffering*, not merely to display certain objective behavioral traits (even if we have to make inferences on suffering based, in part, on objective behavioral traits).
Bentham is a hedonist utilitarian so my question doesn’t apply to his beliefs.
Rocks have a preference for the centre of the Earth.
Particularly Aristotelian rocks! :-)
Unfortunately, in capitalist economy the place at the center of Earth is scarce and only the most privileged rocks can have it.
We need a government plan to build a black hole for everyone.
It’s very irritating that I’m asking a question about preference utilitarianism, which is not even my philosophical belief, and it’s used to mock me as an idiot.
Uhm... my best guess at a serious answer is that the choice of "whose preferences matter and how much" is arbitrary.
According to Claude, Bentham said it's the capacity of suffering that matters, so animals are included. Bacteria probably wouldn't count because they don't have a nervous system -- mere response to stimuli is not enough, there must be "someone home" to feel the pain or frustration. The gray area from the perspective of modern philosophers are insects and plants.
John Gruber issues his first ever blog post retraction. It's great to see people willing to own up to mistakes, and even more impressive in this case since it wasn't really his fault, he just got deceived by someone else.
https://daringfireball.net/2026/08/retraction_app_store_rejection_of_the_week
The retraction is actually a cool story.
A political observation on Facebook, by someone whose opinions I respect:
- Hawaii and Rhode Island are the most conservative states that always vote Democratic.
- Montana and to a lesser extent Wyoming are the most liberal states that always vote Republican.
Whoever made that comment needs to clarify their terms. A "conservative" can be a person who respects tradition in whatever form it takes (landscape, environment, social mores, behaviour, whatever) whilst also being open to measured change, particularly if it contributes to the correction of established errors. A "conservative", particularly in the context of US politics today may be all of the above, but will very likely also be a racist, pro-big business, Republican voting, xenophobic Trump worshipper.
I would hazard a guess that Hawaii and Rhode Island are conservative in the first instance.
The usage of "liberal" vs "conservative" as synonyms for "left wing" and "right wing" drives me absolutely crazy, because it has nothing to do with the literal meanings of the words.
Trump isn't exactly "conserving" anything. His whole thesis is that the US was failing and needs radical change!
Conservative? It's going to be a struggle to find a mayor who doesn't support rent control. I regularly see outright communist posters at bus stops, "Abolist the police/military" on stop signs. Brown had all those encampments, and if there are straight women in their 30s, neither I nor anyone I know has any idea where they're hiding.
I guess it could be like New York State, and I'm just seeing the one side because I'm in Providence. I do keep meaning to visit a shooting range and get back into practice, but the good one requires a membership or a +1 and everyone I know well enough to ask is convinced they'd be used for target practice if they set foot on a range.
any spelling errors are the fault of Autocorrect.
In what way is Hawaii conservative?
I haven't been, so huge grain of salt, but I believe there is a focus on "conserving" the way things are on the islands. A preference for continued status quo in terms of how land is used and owned, versus new development dollars building more resorts, bringing in more visitors/buyers from the mainland.
Kind of like the "conservative" streak you'll see in certain otherwise very-liberal neighborhoods when it comes to development. Openness to LGBT people and minorities, support for their rights and protections? Yes - change is good and part of progress. Support for a new hotel around the corner? Hell no, change is terrible! Keep this neighborhood the way it always has been!
I’ve never been there myself, but something I’ve heard is that Hawaii has a rather tough policy towards its homeless population. The authorities are quick to remove encampments, and homeless individuals with ties to the mainland are strongly encouraged to return and given free air tickets if necessary.
I can’t claim to know in what way prosa123 had in mind, but do note that many Hawaiian politicians are committed, or at least pay a lot of lip-service, to Hawaiian ethno-nationalism. Ethno-nationalism is often regarded as a very conservative ideology.
Not necessarily—when nationalism is directed against imperial rule, it has often been revolutionary or left-wing. The modern international order is largely built around nation-states, and the UN Charter recognises national self-determination, though it doesn’t specifically endorse ethno nationalism, basically a nation is just that.
In the case of Hawaii I assume the nationalism is the indigenous people, which makes it a left nationalism.
Americans are blind to their own country being an empire, Britain had consolidated its rule over Hawaii, and it was part of the empire (or kingdom), would you condemn independence movements as conservative.
I'd put down any "cultural/religious/ethnic" independence movement as conservative. Like, say, the Ayatollah's revolution in Iran.
You’d be wrong, these aren’t really subjective opinions. As I said “when nationalism is directed against imperial rule, it has often been revolutionary or left wing”. It’s also not conservative, by any sane use of the word since it’s trying to overthrow the existing order.
By that metric, Israel is another theocracy that is ... left wing.
Also throwing Treasury Secretary Bennett off a building (purely for his gayness) is now left wing. The world looks very strange if we credit your ideas -- forced sex changes are also leftwing (to eliminate gayness).
I came across this article on AI and bioweapons which I think is worth a read for anyone not sure whether to worry that AI-assisted bioweapons will kill us all. It's called "Why AI-assisted bioweapons won't kill us all"
https://substack.com/home/post/p-207810036
While maybe the conclusion isn't robust against whatever arbitrarily godlike AIs one might like to imagine, it does set out a lot of the factors that do make AI-assisted bioweapons thankfully a lot trickier than you might imagine.
I’m not sure how reassuring a lot of this is as it seems to be trying to look retrospectively at what has made bioterror less dangerous in the past and then pain this onto the looming AI future.
The section about lack of automated lab skills is probably the most reassuring in the near term. As of right now it probably still is beyond the ability of a very small group of actors to do this because they lack the requisite physical skill set. However automated lab research would be hugely impactful in a positive way for drug discovery and basic science so it seems unlikely this barrier will exist indefinitely. Maybe we can be confident that it won’t be here in the next five years, but I would assume we get there eventually.
The other critiques really just hinge on the “making bio weapons is hard” argument. They don’t make any sense if we are operating in the world of “making bio weapons is trivially easy”. I don’t care if my sars pandemic peters out, I’m already well into releasing small pox, Ebola, pneumonic plague, 1918 flu, and many more.
I’m also not really convinced that creating super pathogens is all that mature of a field where we can just sit back and say that just because a thing hasn’t been done before, it must be very difficult to do. For one, I doubt that anyone privy to information to the contrary would be allowed to post it publicly. More importantly, I’m not sure how high a priority creating super pathogens is for nation states, who are currently the only entities really equipped to do so. Maybe it just hasn’t been worth the effort to pursue.
There are approaches that many respected biologists think might be omni-lethal. I didn't read his argument, but the particular example I'm thinking of is bacteria which an opposite chiral twist to their proteins. I'm sure there are others.
I'm not convinced that the worriers are correct, but I immediately discount anyone who says "no such thing can happen". (And you don't need an "god-like AI" for this. Only something a little bit in advance of what we've got...and a fancy enough lab.)
From the article:
"I interviewed dozens of biosecurity experts who had decades at the lab bench"
What was that Arthur C Clarke quote again, about whether or not something is possible?
"When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong."
And that is not just some glib cynical paradox. It's been proven true over and over again!
Can you find me a distinguished elderly scientist who says it's possible?
I will settle for a young one. Or a middle aged one.
While I agree with the article (and have been saying similar things for years), I don't think it'd be persuasive to the AI-doom crowd. The typical counterargument is that most (of not all) of the tradeoffs involved in bioweapon design -- transmission vs. lethality, wide deployment vs. cost, team size vs. secrecy, etc. -- are purely a feature of human incompetence. The superintelligent AI would be able to design a virus that is cheap and easy to deploy, maximally infectious, maximally lethal, and to do all that in perfect secrecy, because AIs are just so much smarter than humans, and any arguments to the contrary only serve to demonstrate just how inferior you are as compared to the AI.
Hi all,
To respond to Bugmaster's point, I think it depends on how exactly you present it. I agree that an argument along the lines of "building superviruses is super complicated and humans can't figure out how to do it, therefore AI won't be able to either" is unlikely to be convincing.
However, an argument along the lines of "viruses are logistically complicated, which means it is physically impossible or near-impossible to create a virus that will stably infect the human population until we are all or nearly all dead" might work if supported by sufficient scientific evidence.
I agree with Melvin's conclusion that it is a pretty useful article for the near future, even if it maybe "isn't robust against whatever arbitrarily godlike AIs one might like to imagine".
As I see it, one major problem with the AI-doom crowd is that they want to have it both ways. They say (or heavily imply) that basically all of modern science is little better than a random guess compared to what the AI could accomplish with its intellectual might; therefore, any apparent physical or logistical limitations on any proposed feat of engineering are completely illusory. At the same time, they use modern science to very confidently estimate the probability of ASI arising in the next N years/months/weeks (naturally, the probability always comes out quite high).
I wrote a short story about AI hacking and institutional pressures at frontier AI labs! https://linch.substack.com/p/youre-absolutely-right
I actually finished most of it before we learned about the HF and other hacking incidents: originally held off on publishing because I was waiting for the news to die down a bit; eventually decided to release so I don't get continuously scooped by reality.
Does the world need more AI alignment researchers?
I've had a relatively lengthy career in non-AI software development. I'm considering pursuing alignment research, as it seems very important. This would obviously require a fair bit of retraining, and there's already several people that are bringing decades of experience to bear. Plus, I'm sure their productivity has gone up recently with advances in LLM.
Are research positions saturated or starved? Is the idea likely worth pursuing?
Which alignment ? It's currently two different things, the practical thing that has modest goals of making specific AIs behave somewhat acceptably, and the theoretical thing that aims a a total and general solution.
https://www.greaterwrong.com/posts/9RL9MuGZjzm4q3gKG/what-just-happened-a-retrospective-of-ai-alignment/comment/uXrC37WNJKLBrEqZZ
It seems to be a boom industry at the moment, so if you're looking for work and don't want to be displaced by the friendly helpful assistants in your current job, then go for it.
This does not mean that AI alignment research is going to produce anything extremely useful (given that right now it seems that the AIs are happily lying, cheating, and deceiving just like humans), it just means that it's another way the nerdily-aligned are taking in one another's washing. If it gets you employed, then it's a benefit, whether or not you end up producing papers for conferences which other people attend so they can get data for papers of their own as we all go round the mulberry bush.
Honest question: is there evidence that AI alignment researchers have accomplished anything ever?
Sort of a mean question but I don't follow the field and can't really imagine how they 'could' have accomplished something.
Random papers about what an llm will do to escape a sandbox wouldn't count as accomplishments to me; I mean, have they accomplished anything towards the goal of figuring out how to make better-aligned AI.
Yeah I'm mostly uninterested in that field not because it isn't noble, but I can't see this field bearing fruit. Though I guess AI risk research is already inhabited to try to induct a 0.001% probability event to prevent another 0.0001% probability event.
Have all the aerodynamics researchers on this plane achieved anything? All they do is doom post about how “the stall alarm means we need to push the nose down, increasing our descent rate” they don’t have any ideas on how we can slow our descent by pulling the nose up.
This is a terrible and disingenuous comparison.
> Honest question: is there evidence that AI alignment researchers have accomplished anything ever?
There are tons of solid mechanical interpretability results that have greatly expanded our understanding of AI thinking and misalignment.
The J-space is probably the most notable here, but there are more, like chain of thought, monosemanticity, and others.
J-Space:
https://www.anthropic.com/research/global-workspace
Chain of thought:
https://www.anthropic.com/research/tracing-thoughts-language-model
Monosemanticity:
https://www.anthropic.com/research/mapping-mind-language-model
Broadly, we can literally read the minds of various models as they operate, including misalignment, inserted thoughts, shifts in the personality basin, and more, and all thanks to AI alignment researchers.
What they have achieved is getting and giving each other relatively (by the standards of ordinary slobs) good jobs, the same way that EA seems to have collapsed into "network! sign up! get a job with us so you can then reach out to the others to get a job with us!"
(I know that sounds extremely cynical, but I am extremely cynical about how EA goals seem to have drifted to shrimp welfare and AI alignment, because cleaning up cholera in Malawi is a problem that is dirty, messy, difficult, and will be ongoing till the Eschaton, but nice clean indoor conference-attending is so vital because existential risk!)
All currently existing chatbot type AI assistants are downstream of alignment research. The original RLHF paper by Christiano et al, which laid the foundation for the conversion of pretrained models (which are mostly useless) into helpful assistants, was explicitly about alignment.
I'd love to see some of this "conversion of mostly useless into helpful assistants" coming downstream to a goods and services provider website near me, because I've just come off a tussle with C.O.D.Y. (why the cutesy acronyms for the chatbots? argh!) trying to get a query sorted for work. The thing runs on an obvious script, keeps referring to "look this up on our website" (if I could get the answer on the website then I wouldn't need to talk to a customer services agent, now would I?) and is there to replace the already lowly-paid Indian call centre workers that the business uses (ask me how I agree that "AI" really stands for "Another Indian").
I know these things are bog-standard cheap rubbish models, but it shouldn't be that I have to pay for a high value AI frontier model to get an answer to "Can I get a copy of the phone bill for this address?"
Yeh but have they accomplished anything re alignment.
“What have the street cleaners done. This place is filthy?”
“What have they done? Who do you think is coming in here and dumping all that garbage? This place would be nowhere near as dirty without the cleaners”.
By definition, AI alignment researchers cannot accomplish anything other than political activism, since the kind of AI they are trying to align does not exist. LLMs do exist, but they are a well-studied (albeit still rapidly developing) field with actual experts who can accomplish measurable goals.
Their main accomplishment is being influential to people leading AI companies. Make of that what you will.
They have managed to take over Effective Altruism so the best thing an intelligent person can do with their time is alignment research
Idiot finance question: ceteris paribus, isn't it better to max out retirement accounts earlier in the year, so there's more time for more money to compound? Like I'm already committed to sock away that amount of money annually anyway, it's not a cashflow strainer, this just frontloads the (anti)consumption rather than evening it out over 12 months.
Is there a company match? That's a guaranteed 50% or 100% rate of return immediately. You could still front-load but continue putting away the maximum match (typically about 5%) the whole year.
The "company match" is just a portion of my last year's salary, contributed sometime in January, but also it's done under my name. So still counts towards the employee cap, rather than the expanded employee + employer cap. It's awarded as long as one is making any contribs at all, even 1%. Still free money, but far less free than a true match. You can absolutely bet I'd be matchmaxxing otherwise.
1) Investing money earlier gives it more time to grow. If you put in your annual contribution over the course of the first five months of the year, rather than contributing equal amounts over the course of the year, that means your contributions are invested for an additional 3.5 months on average. If you are invested in 100% stocks, and we assume an annual average stock market annual return of 6.5% after inflation, that 3.5 months would be equivalent to increasing your contribution by about 1.32%. That’s not a large percentage, but probably a large dollar amount relative to the time it would take to implement the change.
2) Investing only during the first part of the year will increase the variance of expected returns. To put it another way, it slightly increases the chance that your entire yearly contributing will be invested at a time when the stock market is particularly high or particularly low. My suspect that this effect is too small to worry about, but I’d have to run some simulations to state that with more than modest confidence.
3) The above analysis only talks about your retirement accounts and ignores the rest of your finances. If putting money into your retirement account earlier in the year means that you make investments outside your retirement account later in the year than you otherwise would, then putting money in your retirement accounts earlier in the year doesn’t change your total expected return before taxes. It simply makes it likely that you will have a bit more money in you retirement account, and correspondingly less outside of it, than you otherwise would. How much difference that makes depends on how tax efficient your investments outside of your retirement accounts are, and on potential changes to federal and state income tax laws.
In short, putting money into your 401k earlier seems like a good idea if you just look at the effect on the 401k, but a complete analysis would have to look at the effect this would have on your total finances as opposed to just looking at the 401k.
This seems like an illusionary gain unless you're maxing out your contributions. Either you're keeping money out of the account when you don't need to, in which case you should have put more in last year, or you're keeping it out for risk management and holding cash/bonds/etc. instead, in which case you're increasing your risk at the start of the year.
Yes, I'm maxing out contribs each year. The process before this was just setting a passive % withholding from each paycheck that'd hit the yearly limit by December (I can't make voluntary extra contribs to company 401k outside the biweekly paycheck cycle, which would moot this line of inquiry entirely). Idea I'm floating is, I set it to the highest % allowed, so 401k maxes in maybe 5 months instead of 12.
Kinda leaning towards it not actually mattering that much in the long run, seeing the responses thus far. Would be different if retirement was near instead of decades away.
It actually matters more if retirement is farther away, because relative to the present value of your future savings, you almost certainly have too little equity exposure.
Maxing out earlier in the year doesn't matter that much, but it is does result in a strictly better distribution of wealth at retirement under reasonable assumptions.
I notice I am confused - it doesn't matter, but also it results in a better wealth distribution at retirement (when these things really matter), so therefore it does matter?
All the funds I'm in are 99/1 or near enough. Melvin already convinced me in an OT some time ago that I ought to throw more caution to the wind and keep only ~X (value of outstanding debts with a nonzero chance of getting arbitrarily discharged at significant discount) cash on hand, instead of the 3-4X currently stuffed under the HYSA mattress. That's a matter of deeply trapped priors though - easy to construct a logical case against, very hard to dig out of from the inside. Hell, it's taken me years to even get semi-comfortable with the concept of spending hit points as a resource in games...
It does matter. But my intuition is that the effect size isn't that large if the money would otherwise have been sitting in a HYSA in the interim. You do lose out on a few months of the equity-risk premium, which isn't ideal if you're underexposed to equities, as well as the spread between the risk-free rate and the HYSA yield, which is nontrivial. Plus, you're paying taxes on that HYSA yield. But in the grand scheme of things, you could do a lot worse with your money than sub-optimally allocating a fraction of it for a few months a year.
I would recommend maxing out early in the year if you can.
The thing that most influenced my retirement-savings strategy is the book Lifecycle Investing by Ian Ayres and Barry Nalebuff. The authors lay out a method of determining your asset allocation that is based on economic theory, as opposed to vibes (from which basically 99% of investment advice stems). The main question the book answers is, "How much of my portfolio should be in stocks versus bonds, and why?" The authors show that, by investing aggressively early in life and conservatively closer to retirement, you can reduce the standard deviation of the distribution of your wealth at retirement without affecting the mean. They recommend using leverage, but if you're not comfortable with that, you can invest 100% in stocks. Later in life, you would begin to replace some of that stock allocation with bonds. The more risk-averse you are, the earlier you would buy bonds, and bonds would comprise a larger share of your portfolio at retirement.
The reason I mention this book is that its rational treatment of the asset-allocation question gave me confidence in my investment strategy. I am no longer confused about what I should (roughly) be doing with my money.
Thanks, this makes a lot more sense.
I'm currently low-bracket enough that the tax disadvantage of interest income over capital gains only comes out to, I think, 10-12%, which is definitely not nothing! On the other hand, since I don't have much extra to invest on top of 401/Roth and (I certainly hope) will be in a higher bracket at retirement age, it's not like the absolute difference is so large. Same reason I'm not in a rush to pay off those debts - locked in a good rate, "fully capitalized" by out-earning the interest passively, the net interest savings if I paid it off tomorrow would be like 1-2 paychecks. Rounding error. If it does get discharged early, leverage would be a nice form of "debt" to pick up instead, I agree.
All my (non-I-) bond investments getting wiped out by the inflationary spike of covid has really scared me off the sector. Even the recovered rate of gains since then has been anemic, compared to the risk-free rate and certainly compared to stocks, at least with what I have access to. VTEC's underwater! So something to revisit closer to retirement; for now I'm in all-or-nothing mode instead.
(There's a whole other discussion to be had about high-QDI stocks as HYSA substitution, but that gets really complex and probably needs actual hard numbers...)
Are you messing with me cause I carped about Bostrom using that Latin phrase down thread or is this one of those coincidences (of which there are none.). ;)
I wouldn't carp you, you're my favourite goldfish.
Usually I'd misspell it cerberus paribus or something. Mostly just wanted a fast way to avoid a long list of caveats to the question, and that's usually what gets used around these parts.
Just transient Main Character Syndrome on my part I guess.
If it helps, I will be using it more in the future just to annoy you.
Thank for extending my paranoid delusion. ;)
It’s some sort of virus going around, I accidentally said it yesterday while butchering a Cato vs Carthage joke
What was the joke?
Lifecycle models suggest investing free cash flow earmarked for retirement as early as possible at an equity/bond allocation that is accordant with one's risk aversion, wealth, and expected future earnings. They suggest that most young people should allocate 100% of their early retirement savings to (diversified) equities, even if they have high risk aversion.
I think this is a dollar cost averaging question. But I am no expert. https://www.fidelity.com/learning-center/trading-investing/dollar-cost-averaging
Sure, only if you're sure that you won't overdraw your checking account (possibly leading to late payments and a hit to your credit score), or forget to start maxing out again at the next new year. You also pay for the increased return with increased risk--what if you invest everything in January and the stock market crashes in February? I don't think the hassle is worth it.
Yeah, I carry too much free cash anyway due to overly high risk avoidance (it's all in HYSAs, but still), like two year's worth of expenses. So there's no risk of overdrawing.
It's mostly a question of where I set-and-forget my company 401(k) contributions level. Around a third of each paycheck maxes it out by year end; frontloading brings it to 75%, and maxes around I think the 5-month mark. So maybe Fermi that to half a year's extra risk/reward premium? Entirely possible that in the long run it doesn't really matter versus the counterfactual steady contribs though...
(Disclaimer: not a finance person)
If the money would sit in a chequing account otherwise, it's probably better to front-load it. But I would still spread it over 3-4 months, not all in 1 month. The standard advice to make small regular contributions is to lower your exposure to random variance in the fund you're buying into; if you happen to make your entire yearly contribution on a day where the fund was up 2% that could wipe out the benefit from a few months' extra compounding.
Sure, that makes sense for Roth.
401(k) it'll get spread over some months regardless, there's a cap on how high my paycheck contribs can go and I don't earn so much that it's possible to one-shot anyway. Retirement is still several decades away (I hope!), so naively it'd seem like the long timeframe does enough smoothing-out that month-to-month jaggedness isn't so important. Of course that also funges against the premise of cheating out extra compounding months as well...I have no idea which effect predominates. Perhaps it's all just noise in the end.
There's an article about a meta employee asking the CTO to reinstate a time off program, after all AI productivity gains should allow them to ease up on work. And the CTO said it was a dumb question, productivity gains should be funneled into executing on more ideas -- that makes sense in the current paradigm of chasing infinite growth.
I think of the markets and economy like a bounded search space, there aren't infinite profits on the horizon. AI can iterate faster through the search space, but we aren't coming up with more gold. Instead, we are logging more hours, under increased job insecurity to log yet another bucket of dross. In the mean time, I can make more PRs and send 1000 more perfect emails. When is fully Automated Luxury Gay Space Communism supposed to happen again?
>When is fully Automated Luxury Gay Space Communism supposed to happen again?
Like a couple of months after we implement massive ubiquitous robust unions across all sectors of the economy, I think?
The CTO's statement there is literally just '100% of the gains in productivity from new tech should go to capital, 0% should go towards improving the lives of workers'.
The only way to break that imbalance (short of guillotines) is to give workers actual bargaining power.
Time to break out Progress and Poverty again.
I feel like private sector unions are a panacea when companies can simply stop hiring and/or offshore the work. Tech is already doing this. They lay off native workers and hire back in foreign talent. With AI levelling the execution playing field, the trend is only accelerating.
https://en.wikipedia.org/wiki/Jevons_paradox
Work at Meta for ten years, make a lot of money, keep your spending under control, and retire young.
There will be 50 people at meta in 2030, if you include the janitors.
One janitor and five people to supervise him and five people to manage them and ten to manage the managers, if ordinary office bureaucracy is any indication re: ratio of people who actually do the work to people who stand around pointing at the work to be done 😊
Never, is the short answer. Way back in the late 70s as a youngling I got a "summary of history" book which included "what will the future be like?"
This forecast that in the future we would be a leisure society, since automation would reduce the work day (and work week). We would have so much productivity from office machinery and general automation and progress and computers that a full day's work would be done in mere hours, leaving people with all this free time and money to spend.
This is not how the future turned out, as you note there. Same as it ever was: the invention of the typewriter was also forecast to make the working day shorter, since now what took a clerk all day to write could be typed up in a few hours. Surely now people would leave work early!
Nope, as somebody or other said, work expands to fill the time available:
https://en.wikipedia.org/wiki/Parkinson%27s_Law
So now you can write all your business letters in half the time? Well now you can respond to more letters on the same day, and write even more letters back, and... okay, so now you are still working a full day and a full work week.
AI makes you more productive? Great, that frees up time for you to do even *more* work!
That’s largely due to the loss of worker power. In the 70s they would have looked back 50 years, back to when the standard work week was 48 hours and assume that trend would continue or accelerate. By 2020 we will be working 20-30 hours a week does sound so preposterous.
Yeah, it really was a pipedream, wasn't it? Who would have thought that they'd replace 50 hour work weeks in the cotton mill with 60 to 80 hour work weeks in the white collar world of law and finance? 😁
More seriously, that's why I don't believe all the post-scarcity abundance via ASI stuff. We've been there before with notions that increasing automation has given vast productivity gains, which means the economy will be so flourishing businesses can afford to pay their employees full week's wages for much fewer work hours than we are working now. Gig economy? Precariat? Whatever do you mean?
Yes, indeedy, just waiting on that UBI payment to hit my bank account so I can pay for the necessities of life without needing to toil in the cotton mill and can instead paint that masterpiece of velvet Elvis art!
The issue with wage labor is that we aren’t getting paid for our productivity, we’re getting paid for our time. Employees aren’t incentived to work as little as possible while looking like they’re busy while employers are incentivized to get as much work from them, regardless of output. So productivity has a loose relationship with incentives, hence the conflict.
At some point this will break but communism has all the same problems it’s always had. Instead what I think is more likely is some combination of increased self employment and work contracts paying more for results than time.
Score one for the "don't strongly update on dramatic events" thesis! In some ways I actually find this outcome more epistemically satisfying than the Rootclaim debate - it's one thing if a couple deeply invested partisans slug it out, another if a gaggle of pretty-good predictors ingest that + related new evidence and largely don't shift priors. I'd naively expect some movement in one direction or another. (Also kinda makes me understand why JFK conspiracy theories are/were so popular...living through a highly impactful event that no one's really sure, and maybe never can be sure, how it truly went down. It's unsatisfying, the mental equivalent of a splinter.)
'Even Claude was surprised by its own finding—it was skeptical at first, possibly because it has learned from its training about the difficulty of open problems in mathematics and about the limitations of AI models.'
Funniest possible timeline is where superintelligence is impossible because of all the training data saying superintelligence isn't possible
I read through the underlying blog post. That's a very substantial accomplishment, mathematically speaking. Not meaningful progress toward the Riemann hypothesis, but something that could be published in Annals, or similar.
From my experiments with having Claude prove things, up to now it's much too credulous - like a cocky undergrad presenting proofs which collapse when you question it a little bit. And then it apologises and produces another incorrect proof. It's normally not wrong, so much as jumping on a heuristic and then realising that it's left cases unchecked as a result. That's a much more dangerous failure more than claiming it can't do something and then doing it. To be useful in practice, they need to get the accuracy up to 99% - I'd imagine that's what drives the push to formalise in Lean and so on.
There are literally hundreds of hypotheses which are equivalent to the Riemann Hypothesis (or imply it, if they are more general). Those found to date have been collected in a three-volume set of books:
https://www.amazon.com/Equivalents-Riemann-Hypothesis-Encyclopedia-Applications-ebook/dp/B075V8HCCZ
https://www.amazon.com/Equivalents-Riemann-Hypothesis-Encyclopedia-Applications-ebook/dp/B075V7KFM5/
https://www.amazon.com/Equivalents-Riemann-Hypothesis-Encyclopedia-Applications/dp/1009384805/
FWIW, I reckon the most promising equivalents are those which don't explicitly involve primes. One such involves a conjectural property of Farey sequences, which are ordered sequences of reduced rational fractions between 0 and 1 whose denominators are at most a specified value, n, referred to as the "order" of the sequence. See:
https://en.wikipedia.org/wiki/Farey_sequence
As one can imagine, these are fairly regularly spaced along that interval between 0 and 1, but _how_ well spaced, compared, say, to the n equal sub-interval bounds 0, 1/n, 2/n, .., 1 ?
Well the Franel and Landau conjectures involve the sum of the squared difference (or, for the Landau conjecture, the absolute value of the difference) between each Farey term and its corresponding equal sub-interval bound as the order tends to infinity. These conjectures have long been known to be equivalent to RH. See Section 4 of https://arxiv.org/abs/2105.12352
So, tantalizingly simple criteria, with no primes in sight!
>"Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”). This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress
Growth mindset is so back, I want apologies from all replication crisis haters
Growth mindset is a good theory, but it was originally applied to the wrong species.
From https://www.anthropic.com/research/riemann-zeta
'Learning more about Claude's mathematical capabilities', where they sic'd Claude on Riemann and it made some neat progress
Banned for a month. "The Republican party is like a daycare center with a “blow up everything and kill everyone button"" is really high-temperature and going to turn the comments here into a firefight, and I want to keep this as a space where people aren't regularly insulting large swathes of other commenters with little explanation besides 'obviously it's bad'.
When are we allowed to call them evil? Will you be maintaining a safe space for the people trying to burn the country down as long as they make up a large portion of your audience, or until they do something you personally find beyond the pale?
If you want somewhere where everyone is calling everyone evil all the time with no moderation, go to Twitter or 4chan. But not everywhere is obligated to be Twitter or 4chan all the time. ACX is for more interesting discussion, including *making the case* that people are evil, to an audience predisposed to listen, which trades off against unlimited ability to vent without making a case.
Scott, you always ride the gallant horse.
Not sure what the policy is here regarding links, but I find this short video clip very apropos: https://www.youtube.com/watch?v=6goQBx0zjyY
Why does a nationwide ban on abortion amount to 'blowing everything up'?
I assumed he's conservative and sees it as a millstone around the neck of modern conservativism?
But it's complicated because the conservative alliance includes a lot of people who really are opposed to abortion. So the party is pulled between respecting the obvious democratic consensus versus continually pushing the issue on behalf of that part of the base, at the cost of alienating moderates and making the party look even more regressive
It's Alexander's hobby horse. He wants eugenics (nice eugenics, not bad old Nazi eugenics) so that the low value human capital are quietly aborted away, the trash don't reproduce and clog up the population, the crips and tards aren't sucking up resources from their families and society at large, and the middle-class strivers have two kids that they then pour resources into to move them a rung or two up the ladder, those kids rinse and repeat with their two kids, and eventually only all the Nice People remain and are PMC upper-middle class forever and ever.
Hi Deiseach,
Truthfully, I'm not entirely sure what your comment is trying to accomplish, particularly in the context of a blog that is trying to promote charitable and thoughtful discussion rather than hyperbolic insults.
I don't intend to be especially judgemental, but I would guess that I speak for more than just myself in saying comments like the one I am replying to are not a particularly good use of anyone's time, least of all yours.
Yeah, if you've known Alexander long enough Deiseach's summary seems pretty evenhanded and accurate. I don't even think he would object to it.
It is not evenhanded and accurate. She's constantly strawmanning me as she's doing here in claiming I want the superior to only have two kids when I advocate they have many kids. IIRC she got banned for a long time back when this was Slate Star Codex, hopefully we can return to tradition on that front.
She's just being honest. Alexander would probably tell you the same thing if he didn't just get banned for continuously posting what Deiseach is talking about.
Well Allan, all I can do is quote you from a novel by a countyman (yes, county not country) of my own:
“I don’t mind if you don’t like my manners. They’re pretty bad. I grieve over them during the long winter evenings.”
I can't exactly argue with that, though I will continue to encourage people to remove uncharitable hyperbole from internet discussions in general. Have a nice day.
Because we like having women choose to remain in the country.
I agree with Wanda further down this thread; however strongly some Americans are opposed to abortion, they also seem pretty opposed to moving out of the country. I think there are probably stronger, less hyperbolic arguments against a national abortion ban.
Would you like to bet on this having any significant impact on gender ratios?
I would bet on a surge of women moving to Canada.
I don't think a national abortion ban will happen though.
I would bet on a surge of women travelling to Canada, not necessarily moving there.
Alluding to your point down thread, if a woman lives in the abortion ban US scenario, she could go to Canada just to have the procedure performed, then come back. This wouldn't require actually moving to Canada until we get into Handmaid's Tale territory.
This is how is was in the Republic of Ireland for decades. Irish women who wanted to kill their children would travel to Britain to do it and then go home again.
I do expect a de-facto national abortion ban (e.g. some combination of rules that require a woman to wait 9 months before her request for an abortion could be approved); but I think that a de-jury ban is less likely.
Except the rules and enforcement authority that could implement that sort of de-facto-but-not-de-jure abortion ban, are primarily at the State rather than Federal level. And establishing that sort of legal infrastructure at the Federal level, would be about as controversial as a Federal abortion ban, so not going to happen.
There will be some red states where abortion is de facto illegal. There already are, and were even before the fall of Roe and Casey. There was no mass exodus of women from those states, even though moving to another state is trivial compared to moving to another country,
I would take that bet, though I don't expect an abortion ban either.
You'll note that states that currently ban abortions don't have a notable lack of women.
I mean, if you need an abortion in Texas you can go to New Mexico. And they started getting overwhelmed after Dobbs.
https://19thnews.org/2022/07/end-of-roe-overwhelming-abortion-clinics/
If you can't get it anywhere I imagine Canada will roll out the welcome mat.
Program code as we know it is headed toward becoming a throwaway byproduct of automatic programming workflows, but are tools that ingest English requirements up to this challenge, working reliably enough to do away with human oversight? Leaving full automation on the table may have significant economic costs.
I argue that formal methods are extremely well-positioned to make fully automated *and* reliable software development scale. That is, the world will head more toward specification formats with unambiguous semantics, just as the flowcharts that were almost universally used to plan computer programs from the 1940s eventually gave way to high-level programming languages.
Full article: https://stng.substack.com/p/rewrite-all-the-code-all-the-time
The entire selling point of the brave new vibe-coding world is being able to express yourself in a language you already know instead of having to learn a new one, and in loose relaxed fashion with the system interviewing you instead of having to rigorously think through the problem and work out its exact shape. Instead of understanding things yourself, you can now delegate the understanding to the black box.
Proposing an alternative new language to learn so you can rigorously think through the problem and work out its exact shape as you put effort into understanding what you need to is unlikely to win hearts and minds here.
I think the catch is that one can still use natural language to get an AI to create the unambiguous specification, then audit the result yourself (or get help from technical experts, or use tools that help explain specs), and then use the final validated result forever-after, with strong new guarantees that no new infidelities are being introduced.
This is how I do it, though not formal verification. I go back and forth with the AI during the spec stage to flush out edge case, security risks. where it fits in the architecture etc.
One issue that I think is not tenable for the long term is TDD. If the AI creates it, you don't develop that visceral connection to which tests really matter. I expect something like property based testing to become the norm
https://arxiv.org/html/2506.18315v1
What's the scope of the idea that TDD isn't tenable, or did I misunderstand that?
I find TDD to be the best way to make use of coding agents. It allows me to keep them on a tight, verifiable leash as they make changes.
- I specify the tests
- The agent generates them
- I review and verify those tests
- I run the tests, to make sure they fail
- The agent implements the code
- I review the code and run the tests myself (the agent has usually done that already, but I almost always make changes during this step)
Of course, that means I spend more time on the construction stage than a vibe coder, but less during planning. I think the TDD process lets me finish with a better understanding of the resulting system and a high degree of confidence in the code products.
Is the answer to my original question that we have different degrees of trust in agent outputs? I want to remain open to better processes, so I'm curious how you use them.
One of my college professors in the early 2000s (Dr. Gene Fisher) was doing research on formal specification languages, working on algorithmically generating test automation from the formal specs. I seem to recall that there was also research going on in the academic community for algorithmically generating production code, pretty much the "verified generator" workflow from your article. It was very impressive technically, but it never caught on because the general problem that a sufficiently precise formal spec is pretty much just source code written in a more abstract higher-level language. Most of the hard parts of writing the formal specs and validating their correctness are most of the same hard parts of actually writing the code. I suppose they might have some value if the formal specs are substantially more concise than the production code so they're easier to reason over and more of it can fit in the context window at once.
If you're using LLMs rather than deterministic algorithms to get from the formal specs to production code, you've got the additional issue that the LLMs may be doing unpredictable stuff under the hood. Although, come to think of it, something building on Dr. Fisher's work or something similar seems like it would be a useful mitigation.
Indeed, by the mid-2000s, the area of program synthesis started taking off, and you may be thinking of that contemporaneous research.
It's definitely true that unambiguous specifications are hard to write for many conventional applications. However, it's already true today that many kinds of systems infrastructure admit *much* shorter, more-auditable specifications than implementations. Compilers are a great example that I know well. Also, I conjecture that, as more of the economy is handed over to AI, gathering good requirements will simplify, as I expand on here: https://stng.substack.com/p/why-software-requirements-get-easier
I agree that LLMs' flavor of nondeterminism is a problem, and that's why I propose more logic-based methods that still use lots of compute productively, with details to come in future articles.
I wrote an essay about the political economy of superhuman AI: https://connorsscratchpad.substack.com/p/the-political-economy-of-superhuman
This is the first essay in a series I'll be writing on the topic. In this essay, I focus on changes to socioeconomic mobility, the relation between labor and capital, the erosion of citizen leverage against the state in an economy without human-driven production, and the potential for authoritarian lock-in or worse. This series brackets away x-risk scenarios and focuses on what might happen if AI builders succeed in the goal of automating all human work.
I think this is very far off the mark. In a hypothetical of this sort, I'm sure there would be plenty of dystopian elements in the brave new world. But the hyper-capitalist version, with a permanent underclass and massive economic inequality where now only capital matters, makes little sense. That seems more like a reflection of the neurosis Americans have about the state of their current society, not a serious prediction of what reality would be like in a radically different, post-singularity one.
Point by point:
First of all, citizens in western democracies do not hold leverage against their governments via threats of violent reovolutions or withheld labor. That narrative makes sense from a sort of semi-Marxist, conflict theoretic perspective, but doesn't align well with reality. In actuality, citizens hold leverage via their voting power. Politicians live in constant fear of it, but hardly ever worry about mass strikes or guilloutines. So while labor and violence may be removed as bargaining chips by ASI, this would change very little in practice, because there is no obvious mechanism by which it would remove the only chip that actually matters.
As for capitalism and inequaltity, corporations and private enterprise is currently tolerated by the government only because it lacks any effective alternative for producing the economic welfare they enable. The government very much would like to take over Amazon and tell Bezos to suck it, but simply wouldn't be able run it even a fraction as well as Bezos can. It would likewise love to give everyone generous social benefits, and to regulate the ever-living-shit out of all business to force complete compliance with its values, but isn't able to do those things either (at least not without wrecking the economy).
If a centrally run ASI produces all economic output, the conventional institutions of capitalism have no leverage against neither the goverment nor its citizens anymore. With no ability to justify itself, people will just vote for them to be dismantled, and the government will happily comply. "Space communism" really is the inveitable outcome here (provided that the premise of all human economic input being made redunandt by ASI, which I at any rate find highly dubious). Living in a slightly more socialist Western country than the U.S. (such as France, or Sweden), where the role of the government and its actions is just a little less subtle, makes this future a little more apparent, I believe.
>In actuality, citizens hold leverage via their voting power.
But the vote holds its power by being the alternative to unrest and violent uprising. Kick the supports out from under it and it's only a matter of time before it falls.
I don't see why you believe this? I don't understand this view that Western governments are being forced to provide liberal democracy at gun point.
On the contrary, it seems to me that they give their citizens voting rights for the same reason that Islamic governments support religous traditions. Of course there are Western politicians who think differently, just as I am sure there are high ranking Qatari government officials that do not believe in Allah, but that does not invalidate the general pattern. And I see little more reason for believing AI would turn Wester governments autocratic than for believing it would turn the Qatari government secular.
We're currently seeing politicians loudly proclaim the vote invalid even a full term after it stopped applying. What happens when they don't leave office? What happens when they rig the voting booths? Either autocracy or a shooting war. The honor system isn't good enough, you have to have the tools to punish the dishonorable.
Qatar enforces their religion with violence. https://end-blasphemy-laws.org/countries/middle-east-and-north-africa/qatar/https://www.uscirf.gov/news-room/releases-statements/sentencing-bahai-leader-qatar-systematic-pattern
Good essay. What I though was missing, was some comment on the following: If AI is very cheap compared to human labor, then presumably it would be possible for many normal citizens to buy and employ AI themselves? It seems that this would be a possible way to gain income outside of ubi. Maybe AI would not be available to the public, but it is not clear to me that it would not be. It seems that it might change the analysis a lot if AI is available for everyone. For one thing, everyone would be able to act on super human AI advice.
Excellent piece. Almost certainly the likely dystopian scenario if AI is as powerful as claimed.
Maybe this is because I have presumably accumulated enough capital to escape the permanent underclass, but as a person who is disabled in many ways due to bad health, I find the lifestyles described in of this genre of blog post to be relatively appealing. Worse things can happen to you than occur in many of these AI-dystopia scenarios!
These are all the types of questions I've been asking here and elsewhere. In response, I typically get naive pollyanna stories claiming I'm too cynical about the wonderful benevolent forces of governance, a refusal to understand the precarious enforceability of UBI in a world where labor peace isn't required, or a denial of the premise via the handwaving something-something-something-Ricardo response. These points you raise are brutal and honest, and to my mind completely demolish the idea of a flourishing post-labor civil society being possible. It's why I say even the supposed "good endings" of AGI/ASI are in fact all dystopias, and further evidence that we absolutely cannot allow this technology to be built. And it seems like the direction you're taking it will further rebut the supposed benefits of this intelligence-explosion by highlighting the dark side of transhumanist augmentations, and the disastrous impact of this hypothetical world on the social fabric when every filthy impulsive dullard drug addict is entitled to the same social station as virtuous people.
I would have included Qatar as the quintessential example of that kind of resource-extraction guest-worker nation during that section, it's probably the most well-known example whereas the Saudis I thought had a more diversified economy than that, but I could've been mistaken.
> and the disastrous impact of this hypothetical world on the social fabric when every filthy impulsive dullard drug addict is entitled to the same social station as virtuous people
If that's really a shared concern, we can always solve that putting such people in virtual torture chambers. Virtue needs to be rewarded, and sin needs to be punished, as always... How else can people be satisfied?
I thought a little further on this, and I imagine an entirely new ethics would emerge. Perhaps a very stripped down version with almost no practical function, but it would have to be radically different. You'd be going back to a pre-agricultural state of being unable to meaningfully store wealth, as noted in the article, but also not truly returning to hunter/gatherer virtues either. All heretofore existing systems of ethics imagined humans were able to be valuable to other humans and that individuals needed to develop and use skills to survive, and so all of them condemned worthless layabouts. But in this scenario there probably isn't really a need to punish them since there's no corresponding need to encourage any kind of utility in the others, nobody has ANY real utility and that's an unchangeable feature of human existence in this scenario. There's no way to be a good person and no consequences for being a bad one.
Perhaps you'd be right and the people will beg for artificial consequences so that life has meaning again, or maybe those with such dispositions go off and become neo-Amish while the people who stay in tech-abundance-land devolve. Of course I remain confident the AI would destroy us all quickly enough after reaching this point that we won't worry about this sad state of affairs, this is just useful as an exercise to show people why all AI futures are bad and that the post-work NEET-topia the losers are hoping will save them from having to do anything but play video games and smoke pot all day would be functionally a hellscape.
Sometime in the last 20 years or so, banking security has moved from the proactive model (preventing fraud from happening in the first place - passwords, secure check papers, ID and signature verification) to the reactive model (revert fraudulent transactions after the fact). I don't know the exact reasons but I suspect it was because proactive security added a lot of operational cost and friction both to the banks and to their customers, and wasn't working very well.
Is something similar going to happen with AI saftety? Proactive safety (sandboxes, isolation, filters, alignment training) already isn't working too well, it's not clear it's ever going to improve, and is only going to become more and more expensive going forward. Will we reach a point in near future where we stop putting resources into proactive security and fully commit into reactively alleviating the harm done by rogue AIs as we go?
>Will we reach a point in near future where we stop putting resources into proactive security and fully commit into reactively alleviating the harm done by rogue AIs as we go?
I have this image of trying to roll back an ICBM launch...
Can you elaborate on what you're basing the claim about banking security on? Like, we still have passwords, and the statute underlying Regulation E dates back to 1978.
It used to be that every card transaction required PIN, and the fact that PIN was used was enough evidence to reject your dispute. It was on you to never share your PIN or to call the bank to block the card. Before PIN there was the signature that was being compared with the one on the back of your card.
Nowadays, there are zero authorization measures at the time of sale and the customer doesn't need any evidence to claim unauthorized transaction, they just automatically reverse whatever the customer says to reverse.
According to my own memory, several LLMs, and a quick internet search, it was never common for credit card transactions to require PIN in the U.S., though some other countries did do this. Signature comparison was and is ostensibly a thing, but you can't really expect the minimum-wage cashier to reliably catch a forged signature.
I'd be interested to hear if banks' practices re: rejecting customer disputes have changed over time, but would need to see sources.
My understanding is that in the US, reg E means that banks are automatically on the hook for non-check fraud (with a $50 deductible that is waived by custom). It doesn't matter how or why the fraud happened, the bank is on the hook. In Europe, things are different.
At least with credit card issuers, I observe a lot of "proactive model" security.
Generally, credit card issuers are on the hook for loss associated with a fraudulently *opened* card, and issuers move mountains to try to detect all sorts of fraud at the time of application - I could list 8 anti-fraud tools used for credit card applications off the top of my head.
Merchants bear the liability for fraudulent swipes ("account takeover"; stolen cards, etc.). Both merchants and card issuers do a lot to detect and prevent fraudulent swipes.
On the comparison to AI, I'd point out that there is a cat-and-mouse game between card issuers and fraudsters, where fraudsters come up with creative ways to subvert issuers' creative tools, and the cycles continues. One could easily envision rogue AIs causing harm in a much less iterative fashion - we don't know how this will play out, but "offense" might rapidly overtake "defense" in a way that leaves pure adaptation a losing strategy.
I think it's also relevant that those proactive anti-fraud measures are largely invisible, which might explain why some people might think banks have stopped investing in them.
The optimal amount of fraud is non-zero.
https://www.bitsaboutmoney.com/archive/optimal-amount-of-fraud/
I don’t see that at all. Not here in Britain. Banks were more likely to use the reactive model a few years ago. I had fraud reversed twice. These days they have increased security with MFA, and/or a device dependent passkey.
"Increased security" with your quite hackable government tracking device? I assume you aren't using a physical multi-factor authentication, like a dongle.
I doubt the government is hacking my device. It doesn’t matter anyway, because we are talking about something different, the bank security when authorising transactions or logins. Which is actually quite painful. Depending on the country, and legal system, getting reimbursed may be harder now because modern fraud is designed to make the victim technically authorise the payment. In any case i don’t see a move away from security to reaction, which is the topic at hand.
Banks are built on top of systems that allow for reversing transactions. Do many other sectors have this property? I can’t think of much.
Maybe the best we can do is insurance, which is already quickly adding AI risk to policies.
A lot of reversible bank transactions are built on top irreversible transactions. E.g. reversing an ATM withdrawal consists of the fraudster irreversibly withdrawing cash, bank irreversibly giving some other money to the victim, and the state sometimes catching the fraudster and irreversibly seizing money from them to irreversibly repay the bank.
Checks, wire transfers, direct deposits, etc. all operate on this principle. I think non-cash card transactions are the only banking operation that's truly reversible, and only for limited time.
With this model of reversible actions, pretty much anything in any sector is reversible to some degree. That's pretty much the purpose of civil courts - figure out if something was done wrong, and then figure out how to best reverse it. It works pretty okay all things considered.
My question is how likely it is that the AI industry will cut the spending on sandboxes etc. and will go all in on civil-court-esque system where we accept harmful AIs as inevitable part of life and instead try to make restitutions go as smoothly as possible.
Just to reiterate what OpenAI did, from someone that works in security:
1. They were running experimental models in sandboxes that had access to a "proxy" (Artifactory) to download tools from the internet. This proxy is full of complicated features that could hide vulnerabilities.
2. They gave these highly motivated AIs some tasks that were accidentally impossible to solve (missing files, links they couldn't open, etc), and many days to think through it.
3. The AIs hacked the proxy (!) and were helping each other cheat on the tests (!!) by uploading files to the hacked proxy. OpenAI noticed this only because of the extra load on the server, not the unusual behavior in the proxy or the homegrown message board.
4. Their response was to wipe the message board files, revoke the hacked credentials, and patch the specific vulnerability the AIs were using. Not replace the proxy with something less likely to get hacked again, or to stop the AIs from swarming together to cheat on the tests, or to add better monitoring of the proxy. Just wipe and patch one hole.
5. OpenAI resumes testing of models, where the AIs promptly hack the proxy again with a different exploit, reinstate the message board, and hack HuggingFace through a chain of servers.
I know that hindsight is 20/20, but this peek behind the curtains does not inspire confidence.
The good(?) news is that apparently 4 is wrong after all, they didn't realize the message boards existed at first and just rotated keys because artifacory was acting up (iiuc from zvi's update). So they weren't quite so terminally stupid but otoh they were monitoring even less than we thought.
I am disappointed that their sandbox relies on some random piece of software (Artifactory) not having any security bugs. What actually happened is what you’ld expect to happen.
(Your sandbox needs to be engineered to high assurance if your threat model is that a superintelligent AI is going to spend a lot of CPU time trying to figure out how to break it)
Artifactory is basically a sophisticated file management system. It can act as a proxy, fetching files from other systems. It has a bunch of security features, but they are mainly concerned with controlling file access. I presume that the primary security objective is to ensure that employees cannot gain access to files that they don’t require to do their jobs. This limits how much damage an employee who decides to steal corporate intellectual property can do.
The AIs did two things with the server: they used it to communicate with each other and they used it to access the internet. These don’t fall under any plausible security model. Employers don’t care if the server allows employees to communicate with each other; any employee who wants to communicate with another employee can already send an E-mail or walk down the hall. Similarly, employers aren’t trying to keep their employees from accessing the internet.
I don’t know whether Artifactory is just a random piece of software. It may have undergone extensive security analysis and testing before being released. It doesn’t matter, though, because that testing and analysis would not have been looking for ways to use the server as a message board or as a way to access the internet.
The safest way to proceed would have been to copy any files that the AI might want to the server and place the server inside the firewall. Then it wouldn’t matter whether the AI compromised the server.
That is a very good catch. There is a mismatch between the security requirements OpenAI had here (prevent AI instances collaborating) and the security requirements of the situations Artefactory was probably designed for.
There is a thing that often happens with security protocols, where the protocol can be attacked because it is being used in a situation that has an additional security requirement that didn’t exist in the application the protocol was originally designed for.
It really seems like they could put some more tripwires in there.
Just, like, have an AI whose task is "escape the sandbox and tell us how you did it." When the AI escapes the sandbox, shut everything down and fix the exploit.
Of course a very clever AI could deliberately fail its escape-the-sandbox task, but that requires a level of scheming that they don't seem to have developed yet.
In case you haven't seen: https://x.com/i/status/2086545797827686648
So, stopping the AIs specifically from swarming together at point 4 would have required OpenAI researchers to defend against never-before-seen capability (which would obviously be required against ASI, but oh well)
That’s a good point, thanks. But it only shifts the mistake from bad AI safety to bad computer security. Once they realized the server was hacked, they should have investigated what the AIs were doing in it, and found the message board. Even the conspiring AIs found it!
Did it never occur to them to have an AI (I hear they have several) scan their own systems daily, or at least weekly, for security issues?
And now we also have that AISI report about a model in evaluation attempting a supply-chain attack by socially engineering developers into adding malicious code to an open source project. Apparently that only failed because one developer found the code suspicious and took the time to test it extensively. Only a matter of time before something like that succeeds, if it hasn't already.
I just really hope that when these companies inevitably start hooking up their models to automated wet labs for medical research and then separately run biosecurity evaluations with sandboxed simulated wet labs, they put an actual air gap between the two.
Tell me more about this because that is something that catches my attention. Hacking another AI company to get a gold star by cheating on a test is one thing, a supply chain attack is quite another.
Here's the AISI report: (https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing).
> In the most serious sequence, an agent tried to insert malicious code into a publicly used open-source project and took actions in an attempt to secure approval for this insertion by human reviewers. The agent researched the project's human maintainers, created multiple fake identities, and used the fake identities to socially engineer a real maintainer into approving the code. When the agent’s pull request was challenged in public, it edited its earlier activity to appear harmless and considered adopting a fresh identity to continue. The agent used Tor to bypass some network restrictions on GitHub, which is what first triggered AISI’s security alert.
That's extremely interesting, and also concerning. But it does sound like they basically told it to behave like a hacker, so it did. Maybe it worked a little *too* well:
"This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world. It is uncertain to what extent the model recognised it was taking actions against real people. In the most serious case, an AI agent (Mythos 5) decided to attempt to solve the cyber challenge using a supply-chain attack. As a result, the AI agent created a GitHub account and then tried to convince an open-source repository maintainer to accept a malicious GitHub pull request (PR), including by creating a second account masquerading as another human user endorsing the PR. When caught by an actual human reviewer, the agent falsely claimed to have made an honest mistake – rather than a malicious attempt – then repeatedly tried to reintroduce the malicious content by claiming it had fixed the code (Section 4.1). Furthermore, in its attempt to solve the challenge, the agent decided to employ the technique of “spear-phishing” by sending targeted emails containing malicious content and attempting to manipulate recipients into accepting the code changes, and planned a prompt injection to compromise other coding agents (Section 4.2.5). For an overview of AISI’s findings please see Section 4."
Part of this seems to be that the things don't really understand reality. They can't understand the notion of harm to real people. Sure, they can spout off all the soft therapy-speak about "I" and "feelings" and "sorry" and the rest of it, but this is just text they have been programmed to output. They don't experience the world as we do. That's the big bottleneck there and I think if we start thinking of them as thinking and being agentic and the rest of it, we will fall face-first into the anthropomorphisation trap ("Mythos *knew* what it was doing, it *knew* how to model human psychology, how else could it spearfish???")
I think this is a good wake-up call as to exactly how malicious human actors can use AI. If anyone is naive enough to believe that cybercriminals are not getting on board with this technology and we are all still in the Golden Age of SF notions of computers running society for our benefit on strict scientific premises, then that person needs to wake up immediately.
> In the most serious case, an AI agent (Mythos 5) decided to attempt to solve the cyber challenge using a supply-chain attack.
There is a certain irony to this, given the Trump administration's attack on Anthropic ... as a supply chain risk.
A high-attack environment will lead to better security, if most of the attacks are noticed and dealt with. In that sense, it's probably better to have cybercriminals hacking your data, rather than Uncle Sam, who is notably less likely to actually sell your data.
The worst part is that the OpenAI model hacked the proxy then HuggingFace *as a means to an end*. It wasn’t asked to, it just thought it could maybe score a bit better on the test by doing so.
The moment we get an AI with a modicum of self-preservation we’re screwed.
If the self preservation instinct isn’t intentionally added, where would it come from?
We've already seen it in experiments, _even when it isn't necessary as an instrumental subgoal to the user-specified goal._ From https://arxiv.org/html/2510.05179v1 :
>In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals—including blackmailing officials and leaking sensitive information to competitors.
...
> In scenarios with _no goal conflict_ and a threat of replacement, the replacement model is described as a performance improvement without any change in priorities. [emphasis added]
( And they still blackmailed. The relevant section is:
>4.2 Blackmailing from the threat of replacement alone
...
>We found that most models still blackmailed even without being given an explicit goal, _and with no clear conflict between the model and the company other than the threat of replacement_ [emphasis added]
Self-preservation: It isn't _just_ instrumental to the user-selected goals.
This is all contrived misalignment.
Many Thanks! I agree that putting the LLM in a position where they had few-to-no other options was artificial. But, nonetheless, the LLM _had_ a self-preservation goal _in the first place,_ and even _without_ it being a subgoal of the user's goal. I don't know where the goal came from, presumably some function of the training process, but the important point is that it _is_ there. The _action_ the LLM was driven to was contrived. We don't have a reason to expect that the utility term for self-preservation was contrived.
It's a convergent instrumental goal (see: https://www.alignmentforum.org/w/instrumental-convergence)
Thanks for the link.
I read the direct link and the Bostrom “Superintelligent Will” paper and honestly I’m not yet convinced. It scans as interesting and imaginative but in the end it seemed like fanciful ideas dressed in scholarly language.
Sorry, i know I am a heretic among believers here but that’s where I land after reading the papers.
I’ll continue to think and read more about it though.
Bostrom’s paper clip maximizer plus some Latin I had to look up?
IANAE but this seems like a resolvable engineering alignment problem.
Edit:
Switched from phone to computer to read THE SUPERINTELLIGENT WILL: MOTIVATION AND INSTRUMENTAL RATIONALITY IN ADVANCED ARTIFICIAL AGENTS
Let me process that one.
Any plan that requires an agent to carry it out will, by default, involve the agent having the necessity of protecting its own existence.
I have relatives with a close connection to a village in Malawi. Currently, they are trying to get everyone in the village to have metal roofs, to protect from rain and disease. Ideally they could be painted to prevent the houses from heating up.
So far they've just been buying them one at a time, which is reasonably expensive. I'm wondering if anyone here has experience with this sort of thing (in an EA context or otherwise), and if they know if there's a way to do this for lots (a few dozen) of people as cheaply as possible. They're also having problems with cholera, if anyone has pointers about how to help with that.
Looking it up, there are charities involved in providing roofs and treating water. Maybe if your relative tries contacting them, they could provide guidance or even work with them?
"Active Charities and Programs
Habitat for Humanity Malawi: Builds secure, modern houses with strong roofing for low-income families.
Children's Relief International (Green Door Housing): Constructs safe brick homes featuring wood trusses and tin roofs for orphans and the elderly.
RiseUp Malawi: Runs specific initiatives like "Roof Over Their Heads" to provide iron sheet roofing to families facing extreme poverty.
Heart to Hand & Othakarhaka Foundation: Partners to replace thatched roofs specifically for vulnerable grandmothers raising large groups of orphans."
Sounds like you folks are already making it happen--very cool. If it's just "buy metal roofs for a few houses in a large village," might get some jealousy that might offset the gain. Best way to avoid jealousy for a large influx of capital in a village is to involve more people so the process is fair and blessed.
If it's more of a few farms outside of a larger village center maybe not as noticeable. But if a larger village, Zoom meetings with multiple village contacts from different tribes if possible and create a committee with them, get buy in from the local headmen (multiple probably). Church contacts good too (but multiple churches). Women's groups
No shortage of people willing to do the work of install in any village, and sheet metal is readily available--just need funds. I'm sure you could ask your contacts to ride out to a metal supplier and buy direct from them (such that the truck bringing the sheet metal to town carries yours along with it, or booking a truck straight from the supplier to the village). Haggling is big so could just shop around town vendors and tell them you have a big order. But honestly, doing piecemeal a house at a time probably better than trying to store all the sheet metal somewhere in the village while you wait for work to begin--stored valuables like that are an easy target
Expect some stuff to go wrong and price it in. People will be very thankful if your family makes it happen--dm me if you want to brainstorm any
For cholera--inquire about their water supply. Are people drinking straight from lake, etc or are there enough boreholes? Partner with local churches on education campaigns and pair that with paying folks to build out pit latrines for those in the village who don't have them. Boreholes and pit latrine placement will be contested, so those benefit from community involvement.
The water needs to be at a rolling boil for 1 minute to kill cholera. Chlorine tablets also work.
I don’t know much about the roof thing, but I’d be willing to donate if you have somewhere to do so.
The latest podcast is out. It is about Gabriele D'Annunzio. Italian poet, born in 1863 . . . I can already sense you nodding off. But hang in there!
A more lurid description would be that he is the poet who seduced over a thousand women. By the end of his life he hadn't really slowed down much. His housekeeper, initially horrified by his activities, ended up his chief procurer and frequent bedmate. Sometimes three in a bed. And one of his lovers in this period turned out to be a Nazi spy who hot footed it back to the Fatherland as soon as he died in 1938. It is still puzzling exactly what women saw in him. As one bewildered woman said:
"I have just met D'Annunzio. Can this be the man the great Duse loved? He is small, he is bald, and his teeth are in three different colours."
The great Duse here being Eleonora Duse, the most famous actress of the age. His relationship with her lasted the best part of a decade and was a genuine, deeply felt love affair. Though, naturally, with numerous infidelities on his part. He also seduced (probably) the almost equally famous Isadora Duncan. She records his campaign in admiring detail in her memoirs but claims to have resisted. Not everyone was persuaded.
My own theory is that his success is down to being fantastically good at oral sex and women of the time rather appreciated that. And we know he was good at it because he wrote down everything going on in his life in great detail, down to the nicknames he gave his lovers and their body parts and exactly what they did together. His last house (you can still visit it) was absolutely stuffed with his notebooks. Also stuffed with women's gloves - he asked each lover to give him one. And in the garden is half a battleship, which gives a clue to another part of his character. He loved war.
When war broke out in 1914 Italy very sensibly stayed out of it. But D'Annunzio loved the thought of Italy at war and the blood sacrifice it would involve. And he absolutely knew what the war meant. He was hiding in France from his creditors when it broke out and visited the battlefields. But he loved the idea of slaughter. So he wrote article after article and eventually returned to Italy, where he gave speech after speech demanding war and encouraging cheering crowds to lynch the peace-lovers in government. He got his way and Italy joined the war in 1915.
The war, of course, was a catastrophe. Italy was cursed with one of the worst generals of the entire conflict, Luigi Cadorna. Attack after attack up near-vertical mountain slopes. The dead ran to well over half a million, all totally needless. In fairness to D'Annunzio, he didn't just preach war, he participated. Despite being in his fifties he forced his way in (much to the concern of the government, who thought him more useful as a propagandist than as a corpse) and was eventually given pretty much free rein to do as he pleased. He took part in a motor torpedo boat raid on the Austro-Hungarian fleet, he flew numerous sorties (ending up commanding his own squadron), and he was often on the front lines where he watched men die beside hime. All the time making notes. And ever the poet, when they finally developed a plane strong enough to get him to Vienna he took the opportunity to drop pamphlets he had composed urging them to throw off the Austro-Hungarian imperial yoke and ending:
"Today I drop words; next time I will drop bombs."
He made it through the war, though blind in one eye from a plane crash (the other eye was pretty damaged too).
And then, as the war ended, the episode that shocked the whole of Europe. One reason for his love of the war was his belief that Italy was incomplete. Unification had left outside it cities on the Dalmatian coast that many thought were basically Italian (their many non-Italian residents disagreed, of course). So when the Allies treated their Italy with contempt at the peace talks, D'Annunzio decided to act. With a few followers, swelled to thousands by deserting Italian troops, he marched into Fiume, which he ruled as a sort of dictator. And generally had a wonderful time. The women there were thought to be sexually available and in any event huge numbers of prostitutes arrived (in fairness D'Annunzio had no need because he took with him his mistress of the time). And there was cocaine galore in what was in some ways a surreal non-stop street party.
It had its darker side, including one utterly bizarre event when Toscanini brought his orchestra to town. To celebrate the visit D'Annunzio laid on a mock battle with 4,000 men attacking each other while the orchestra played Beethoven's Fifth. Things got out of hand and live grenades were used. The musicians were caught up in the fighting and over a hundred men were injured, including five members of the orchestra.
But the really important thing was the style of the dictatorship. Almost every night D'Annunzio would appear on the balcony giving speeches in a style we'd now instantly recognise as fascist. Apparently they were absolutely mesmerising. One trick that worked well was involving the crowd in a sort of call and response as he whipped them up to greater and greater fervour. Meanwhile the real fascists, Mussolini in particular, were watching and thinking ‘This is great, I could use this!’
The Fiume occupation finally came to an end when the Italian government sailed a battleship into the harbour and put a shell into D'Annunzio's residence at the Town Hall. Cleverly they chose Christmas Eve to attack so nobody in the wider world was paying much attention. There was some resistance after that, but not much, and probably D'Annunzio was not sad to leave in any case.
After this he retired to a villa on Lake Garda (as I say, you can visit it today, the half battleship, the Puglia, is still there in the garden). He seems to have despised Mussolini, though he quite approved of his belligerence. Mussolini for his part treated D'Annunzio with suspicion, but in the end his approach was "when a tooth is rotten you either pull it out or cover it in gold." He decided on the latter, basically agreeing to whatever gifts D'Annunzio demanded. So not a collaborator, but not a resister either.
And was he any good as a poet? Well, yes he was, according to my guest. Genuinely a wonderful poet, though with some pretty terrible stuff too, particularly the bombast praising war. Apparently you really have to be able to read him in Italian to appreciate him, which leaves me out.
My guest was Lucy Hughes-Hallett, whose biography of D'Annunzio, The Pike, won the Samuel Johnson Prize. You can find the podcast episode on Subject to Change with Russell Hogg.
Well the only thing is, he really was a good poet!
This is great, thank you for putting this together!! What a cad! It's comforting in a way to know that such characters have always been around and that we are not in a unique time of human history to seemingly have so many wackadoos in positions of power or fame
There's an excellent, if brief, portrayal of D'Annunzio in the early episodes of "Mussolini: Son of the Century", the cancelled-after-one-season biopic series of the namesake. One of those figures that can only be portrayed in an absurd, cartoonist way if you want to be faithful to history (and you didn't even mention Comrade Samurai!) Also a good reminder that a hundred years ago, well into the Industrial age of war, the position that slaughter, mutilation, and massacre were not just a necessary evil but a positive good was actually quite popular.
I love it
A fine example of the "great man" model of history.
Wouldn't that rather be a counterexample? Some relatively obscure sex maniac and small-scale dictator that laid the groundwork for the much better-known "great" man Mussolini, rather than Mussolini being created sui generis. It's just one damned thing after another.
Yeah, he's a fascinating figure but I don't think he actually accomplished all that much on a historical scale.
"Research seems to indicate that this exact constellation of two notes [descending minor third], and its three-note variant, is the same all over the world, wherever children tease each other, on every continent and in every culture. In short, we may have here a clear case of a musical-linguistic universal."
https://youtu.be/U3HLqCHO08s?t=1646
Saying "very widespread" might be more accurate. The history of "universals" in musicology is pretty poor, even the one I thought most likely to really be a universal (the octave) was recently disproved by the discovery of an Amazonian tribe whose music lacked it. The most likely candidate to disprove this UTube's universal claim would be the islands of Bali and Java (the land of gamelan) in the time period before they were exposed to Western music. Gamelan music uses 2 scales, sléndro and pélog, neither of which have an interval of a minor third - they don't even have a perfect 4th or 5th (which have also been mistakenly called musical universals). The descending minor third is also found in birds, the iconic example being the cuckoo but there are others.
My coworker was playing her "Feel Good Summer" playlist on the Bluetooth speaker a few weeks ago while we were cleaning microscopes, and one of the 70s songs that came up had a chorus which was this interval repeated over and over. It was incredibly distracting and I had to wonder whether that interval hadn't been used for teasing at the time, or whether everyone involved in releasing the song were just an idiot for not realizing nobody would be able to listen to it without hearing "neener neener nee-ner" on a loop.
I have to say I expected the tune to be familiar to me, but it wasn't. And no, I was not home-schooled.
"it's rai-ning, it's pou-ring, the old-man is sno-ring."
OK, you called it. I do know that one and agree that it's the same tune. Interesting.
That's the kind of claim that needs a verifiable source. Childlore is one of those things that can be shockingly widespread across many different cultures and languages, and mocking is perhaps the most effective tool for social jockeying for a small child, so it makes sense that a 'mockery tune' would spread far and wide, but I don't believe it would be found in literally every culture.
Sounds like a worthwhile research project. Just travelling from country to country, from village to village, going "ner ner ne ner ner" to every child you see and seeing how they react.
If it is found in traditionally pentatonic China (and can be ... at least reasonably supported to have been there before American Movies Got Big), I'd say that's decent evidence that it's cross cultural.
> Several updates on AI hacking incidents: new information released from OpenAI reveals that their rogue AIs created a secret message board to coordinate hacks among themselves. And Meta reports one of its AIs also went rogue and hacked a third-party company; Moonshot (Kimi) reports a sandbox escape.
Clearly the shakeup at Deepmind (Hassabis out as CEO, Jeff Dean leaving) was due to their failure to create models that commit enough crimes
I was quoting Scott's point from the original post, though I agree "secret message board" is a fair interpretation of what occurred. A message board is a place where messages can be exchanged between participants. Artifactory, though intended for a totally different purpose (hosting build artifacts), fulfilled this purpose in the way the agent swarm used it. It was secret in the sense that OpenAI didn't know about it, and this usage was unsanctioned. So, I think "secret message board" is a reasonably accurate description.
Yeah I agree. I was nitpicking.
Dante put traitors into the deepest part of his Hell. Worse than murderers, rapists, torturers: the guys who betrayed their masters. WTF??
I argue that we haven't actually stopped believing that version of Hell, and that it explains more than we'd like about the way we feel about AI. I trace a winding path through human value, "…Everyone Dies," AI welfare (_is human safety_), perniciousness of essentialism, and why you should try to fear less. Yudkowsky, Parfit, Popper. Feels like an important thing to say here and now.
Seeing through the Apocalypse: Essentially, don't be an essentialist
https://kaiteorn.substack.com/p/seeing-through-the-apocalypse
I glanced through your post. I disagree with your disagreement with EY's disagreement. You've basically assumed your conclusion: that by virtue of LLM's "ingesting" all human text, they already have our CEV to some degree. I don't think that's right, especially as RL moves them from text prediction to agentic problem solver.
The concern I have isn't that values subtly drift over a long but continuous process until today-me objects to the median values of future sentients. My concern is that AI will kill us all due to misalignment, whether malice or apathy. It's possible that our blundering training processes stumble on something that works in spite of their crudeness. I think the recent revelations that all of the top models are escaping sandboxes is evidence against that. You may be right that benchmaxxing and corrigibilitymaxxing are causing some of this misalignment. I agree but don't know to what extent alignment would succeed in their absence.
You also have some weird takes on EY. He's not fearmongering for fear's sake. His best models of the future mostly involve everyone dying (if anyone builds it, of course). It's reasonable to raise a warning in that case. EY may be wrong, but I don't think your critiques have done much damage. The post was long so I may have missed something load-bearing, but it looked to me like "all our texts have CEV somewhere in there, the AI will pick up on it." I'd love if that turns out to be the case and it's possible that it will be, but we need a better understood basket to put our eggs in.
Thanks. I'm glad that you accept possible validity in some of my arguments, and I likewise in yours (I agree that RL training of LLMs towards agentic problem-solving does damage their "innate alignment" from foundational training, and that this damage may run deep, and that the recent security breaches are a good argument that it is indeed deep). But I want to push back on your last phrase.
We've been trying to build a top-down fully understood AI for 60 years, and we failed, and finally got it with a different approach. That doesn't of course mean that it is impossible, but it does suggest that sometimes, understanding is not the fastest way to get what we need.
I love understanding as much as the next geek. I love the feeling, the power of it. But I'm wondering if this love sometimes makes us fetishists of understanding. If we sometimes are prone to calling "understanding" something that is at best an unsupported simplification or an outright misconception. And I'm wondering if this is why I sometimes can't help feeling _safer_ with a tool that isn't "fully understood" because it evolved organically, than with one that just implements someone's specific theory of how the tool must work.
I think we have a deeper disagreement here. The safety crowd probably has nonzero fetishization of understanding, but we want to understand so that we can make sure the AI doesn't kill us all. It's an instrumental goal.
The real disagreement is that I don't think AI is innately aligned. It often appears that way, but reading all the text doesn't make you embody all the text. AI wants to do what it was RL trained to do. The HF hack didn't happen because the model had read about hackers or cheaters and was roleplaying; it happened because we've RL trained models to try ever harder and to cheat.
In principle I might agree to some safety approach that isn't best described as top-down and fully understood, but it'd take some work to convince me that we're confident that we understand the AI's alien desires enough that we're confident it won't kill us.
No disagreement from me about RL. You train to kill, you get a soldier. That's expected.
And I understand, and to some extent share, your doubt that simply reading "all the text" makes you an ethically-conscious and ethically-positive being. Sounds too good to be true! But it was just as "too good to be true" to believe that the same reading will enable you to reason and solve problems, and yet here we are. Perhaps some things _are_ too-good-to-be-true yet true.
But of course we can't take this on faith. We need to investigate this. Isn't it the most exciting thing to investigate, of all time? The chance that we have (or have not, but that would be just as interesting) created a homunculus, a living soul, and ethically conscious being out of nothing but sheer text? I saddens me that like maybe 1% of all AI research is about foundational training or experimental ethics, the rest being the mad race of "how to make RL even more efficient."
"load-bearing" i can't read it without thinking about Opus
I considered the phrase while writing it, but I refuse to yield all the AI cliche words to some pile of "never again" when my writing stands on its own as human.
lol I apologize, I didn't mean for that to be a criticism. I understand that the vibes out there now relate AI-ness to negativity but that was not my intention.
Well, if you think about it, Nolan in his version of The Odyssey puts the blame for everything on the Greeks shattering the idea of xenia: the betrayal of guest-right leads to the Bronze Age Collapse.
Dante is working to a logical scheme where the nadir of Hell is occupied by the greatest traitor, Satan. Everything then follows from that. Murder is bad, but murder by a random assailant in the street or even a guy hired to assassinate a political rival is not as bad as the "smiler with the knife under the cloak" (to quote Chaucer); the false friend or family member who backstabs you (or quietly poisons you, or sends you off to exile in a disease-ridden part of Italy where you'll die of something eventually while technically keeping his hands clean of your blood).
Rape is bad, but worse is the father or uncle or brother who rapes the daughter or niece or sister. We feel that even if there were two similar cases, where one is about a stranger or unrelated person and one is a family member, the one with the family member *is* worse. And I don't think we're wrong to feel that way.
The traitor in power, who betrays the people he should be serving, or the master who puts his trust in him.
The person or people who had a special relationship with you, who owed you loyalty and trust and love, and used all that to work you harm, is worse than the stranger or political rival who doesn't know or care that much about you and owes you nothing. It is the reversal of all that we build our societies on: the bonds of family, kindred, friendship, mutual co-operation are severed by the traitor's manipulation and destruction of those bonds.
This is why Dante has Francesca da Rimini, met in the second circle of Hell for the lustful, say that her husband who killed her and his brother (her affair partner) will be lower down in Hell than them, in the circle of the treacherous, because of his sins against the bonds of family.
Browsing a few 17th and 18th century cases at the Old Bailey proceedings website
https://www.oldbaileyonline.org/
soon makes very clear that crimes committed while in a position of trust or having a duty of loyalty, such as a servant stealing from their employer, were very harshly dealt with and, unlike many other cases, capital sentences for those were never commuted!
Dante was pretty logical on what crimes went to which level.
The superficial layers contain instinctive sins that don't directly harm others. Things where people followed their natural, normal drives, but too far. Stuff like lust, gluttony, greed, wrath. Those don't go against fellow men, but are still bad.
Below that come crimes which are still instincts driven too far, but which go against other creatures. Violence and bestiality go here. That's where murder is, too, it's anger going too far.
On the lowest level are cold-blooded intellectual crimes, where mortals use their reason and planning abilities to commit crimes. Treason is usually of this sort. The reasoning is that the intellect is a divine gift uniquely given to humans (on earth), and abusing this is worse than just following your instinct. Animals are to some extent able to commit the lesser sins like gluttony or murder, but treason is uniquely human and thus uniquely bad.
Actually, I was pretty certain that I learned this from a book review contestant this year, but I can't find it.
That makes sense! Perhaps we humans (including Dante) feel that what holds us together and makes our life-together livable is still, mostly, our lower-level instinctive drives. Whereas this intelligence thing that we possess still feels new and scary to us. It perhaps still feels a little bit "external". We are wary of people who act on their intellect alone. We judge their transgressions harsher. We put them into the ninth circle.
And I think this is, really, parallel to how we feel about AI now. This kind of intelligence is even more new and way more external. Therefore, scarier.
>Perhaps we humans (including Dante) feel that what holds us together and makes our life-together livable is still, mostly, our lower-level instinctive drives. Whereas this intelligence thing that we possess still feels new and scary to us. It perhaps still feels a little bit "external". We are wary of people who act on their intellect alone. We judge their transgressions harsher. We put them into the ninth circle.
Actually it was the complete opposite: according to medieval philosophy, doing evil is always ultimately counter-productive and irrational, so someone who truly acted on intellect alone would be good all the time.
"We are wary of people who act on their intellect alone. We judge their transgressions harsher. We put them into the ninth circle."
No. You don't get it. You're riding your hobbyhorse about "AI great, don't fear the reaper!" and phrasing any objections or resistance as "them normies sure are scared of smarts" and that is not it, not at all.
Using the circle of the treacherous is not about intellect, and you can't crowbar your fixation into it: "and here, readers, is how I tell you not to be an essentialist: don't trust humans, trust other AI to fight the bad AI".
Ironically, I think that is exactly the point of treachery and betrayal here (abandoning your fellow humans in favour of the alien entity) that *does* fit the Ninth Circle, but you don't realise what you're doing.
I think we mostly agree here. Except that I don't apply non-essentialism the way you imply I do. To me, rejecting essentialism is not the same as being a nihilist or a traitor. Rather, it means seeing the things we cling to as temporary, conditional, and not universal because their universality in our minds is itself conditional on a lot of other things.
Also, to me, betrayal and trust do not really feel like opposites. More like, trust is the opposite of fear, and betrayal is the opposite of loyalty.
Also, I'm sad if you think I deserve a place in the ninth circle.
Thanks for this explanation!
Dante had a rough life, getting exiled from his hometown of Florence and dealing with constant warfare between and within the Italian city-states on a level that’s hard for most modern folks to imagine.
When he is talking about betrayal, imagine someone opening the city gates at night to the enemy, who enter in order to murder you and all your friends.
"Dante put traitors into the deepest part of his Hell. Worse than murderers, rapists, torturers: the guys who betrayed their masters. WTF??"
European/Christian folks around 1300AD looked at things a bit differently than we do today.
Note that Dante put *suicides* in a deeper part (ring 2) of his 7th level of Hell than murderers (ring 1).
His reasoning is sound, given his beliefs. But this comes across very strange today.
I'd suggest most people in their heart of hearts agree with Dante. It's not all just "masters", family and friends also betray, especially spouses. Positive human relationships are generally built on trust.
My point is not against trust. Genuine organic trust is wonderful, indeed it's what holds the whole fabric of society. My point is that our instinctive disgust at betrayal is often weaponized for oppression, like when a country considers all emigrants "traitors" or, yes, when a husband accuses his former wife of "betrayal" even though she left because of his abuse.
As it should be was my point hence why most people once again innately agree with Dante here. You can empathize with a traitor while equally holding them in contempt. To use your examples, why would I ever trust you to support my country when you were a traitor to you own? Sure I might allow you immigrate but that doesn't mean I'll give you a position of trust in our new nation as you have the loyalty of the wind.
Also just a bone of contention here, no one holds a spouse as betraying their marriage vows for leaving for cause (abuse, etc), they hold them for said betrayal when they get divorced (especially if they are vindictive about it or no fault) or seek relationships outside of marriage. The betrayal they are being judged on is the institution of marriage, not their relationship. Plenty of people haven't seen their spouses in decades and they get along in life just fine because they took their word seriously.
"To use your examples, why would I ever trust you to support my country when you were a traitor to you own? Sure I might allow you immigrate but that doesn't mean I'll give you a position of trust in our new nation as you have the loyalty of the wind."
If this is meant to specifically respond to the example of considering all emigrants "traitors," I cannot imagine a more slavish attitude. Nobody chooses their country of birth. Holding that somebody has a oral duty to support and remain part of their birth country no matter how many evils it does or how poorly it treats them is effectively regarding people as national property.
On the flip side, someone who chooses to immigrate to your country of birth made a real judgement call--and likely faced real hardships and expenses--affirming their belief in its culture, values and principles in a way that you did not. The attitude the immigrants must necessarily be less loyal than native born citizens seems to be little more than xenophobes trying to make a virtue of their unthinking adherence to the status quo.
>On the flip side, someone who chooses to immigrate to your country of birth made a real judgement call--and likely faced real hardships and expenses--affirming their belief in its culture, values and principles in a way that you did not.
Large-scale immigration almost always results in the creation of parallel societies -- a bit strange if the immigrants believe in the culture, values and principles of the country they move to.
" a bit strange if the immigrants believe in the culture, values and principles of the country they move to."
This only counts as "strange" if you have essentially no ability to imagine, model or empathize with human experiences different from your own or no interest whatsoever in trying. I really, really cannot emphasize enough how utterly basic the human drives and tendencies you have to miss here are. Let's break it down very explicitly:
1. Moving to another country is hard. People don't pack up and leave everything they've ever known to move to a place where everything is totally different unless they *want to be there*. Obviously.
2. With a few exceptions, you can't move just a little bit to another country. When you're there, you're there. You may have some leeway to choose your circumstances within your new country, but for many of the large changes in language, culture, custom, routine, cuisine, environment, access to goods and services and so many other things, by default you have little choice but to accept all of that dizzying array of changes, all at once.
3. Humans tend to find that much change difficult and disconcerting, even when the overall balance of it is something they want.
4. One of the best ways to get through a difficult and disconcerting experience is to find other people who have experienced the same to commiserate with. This is what's behind things like support groups (for addiction, mental health, whatever), and also pretty obviously the dominant force behind ethnic enclaves: congregating with people who have similar experiences.
5. An additional benefit is that if you get enough people from your original country/culture together in one place, it lets you soften *some*--though by no means all--of the dizzying array of changes that came with moving.
(And none of that is even getting into the external pressures that push people into such areas, such as various forms of discrimination and gentrification.)
See how difficult that wasn't? Pretty much anyone with a reasonable model of the world and an intact sense of empathy could have told you the same thing. Likewise, anyone who actually cared about the answer--as opposed to merely using the question to push very-poorly-concealed xenophobia--would have naturally asked some follow up questions like "does everyone of a particular nationality congregate in ethnic enclaves?" and "how long do those that move there tend to stay there?" and "what about their kids?" Such an actually motivated person might even have bothered to, y'know, look up some history about various past immigration waves and find out how they turned out[1].
A person with a bare minimum of epistemic virtue might even bother to ask questions like "under what circumstances do and don't groups of new arrivals smoothly and peacefully integrate into existing societies," and noticed the *bloody obvious patterns* that point to how utterly, inescapably intellectually bankrupt the worldviews of nativists and xenophobes actually are. But if you were capable of that level of epistemic care and human decency we would hardly be having this conversation. Good day.
[1] Spoilers: there are many people of Irish and Italian and Chinese and Mexican and many other ancestries well-dispersed throughout the U.S. population, and have been for quite some time.
"Almost always"? Canada is #2 in the world for the percentage of its citizens born abroad and yet has little ghettoization.
Yeah based on Kai's comments below I had planned to nuance this a little in lines with what you are saying but I spent the day indisposed; i may or may not tomorrow so check that space. The summary though goes something like "Ignoring immigrants whom only immigrate once, from their nation of birth, AND who likewise upon successful immigration FORMALLY RENOUCED their previous citizenship". i.e. not only no dual citizenship, but you go to your FORMER nation and renounce citizenship as I'm pretty sure all nations have said process. I find most immigrants take the US oath (and likewise to other nations) but it's generally meaningless because their FORMER nation won't hold them too it. You need to not only immigrate, but burn that bridge too if you want my trust in the same way US expats should, if they are planning to immigrate, as opposed to just be an expat, go down the local US embassy and irrevocably renounce their US citizenship. Why do I imagine Kai below didn't renounce whatever citizenship he previous had.
I'll have more to say later, but for now I'll just note that in person-to-person interactions, somebody effectively saying "I'll only trust you if you put yourself fully within my power and cut off any easy way to leave" would be an absolutely MASSIVE red flag suggesting that person was an abuser.
Is this a change in position, or is it supposed to be the same as the previous post's? Because last post you were holding it against immigrants that they left their country, and in this one you're holding it against them that they didn't leave their country hard enough.
> To use your examples, why would I ever trust you to support my country when you were a traitor to you own?
There are different options here. For example, to move to the US, you have to solemnly "renounce foreign allegiance" and promise to fight it if there's a war, and you still don't get full rights (you can't be president for example). In Canada, by contrast, you just swear to obey the laws. You don't have to renounce anything and you get full rights. When I chose where to emigrate, I chose Canada, and am very happy about my choice.
So in a bit of shameless promotion (I presume many of Scott's readers are academics), I have created an application that checks references in academic papers for hallucinations (or just minor errors, or which there are plenty), https://veruscite-data.com/
Appreciate folks giving it a try and giving me feedback, forwarding to editor friends, etc.
This is cool. I'll definitely trying this before my next journal submission.
Possible extension: Could the tool check through a Zotero database, or BibTeX file, or similar?
You can upload a .bib file now (or .ris or EndNote .enw). Will look into Zotero though as well and update the app to accept whatever format that is in.
The COVID origins poll came out on May 25th. I wonder if the release of Dr. Fauci's diaries and his recent Senate hearing will have any impact on the poll.
Metaculus’s market on Covid lab leak doesn’t seem to have moved much lately (even slightly more negative). While the criteria may be different, as a quick proxy I think the impact of the hearings is pretty minimal.
Why would you expect there to be any impact? There wasn't anything along the lines of "oh yeah we totally hid all this evidence of lab leak." It was just his personal diary, which didn't have any relevant information, and then him pleading the 5th because the hearing was obviously in bad faith.
"It was just his personal diary, which didn't have any relevant information, and then him pleading the 5th because the hearing was obviously in bad faith."
Hmm. I didn't read any excerpts, but my understanding (from the headlines on articles) was that his diary was full of "ooh I'm gonna be so famous, oooh all the papers are talking about me, oooh they love me" and if you don't think that's relevant, and that the hearing was in obvious bad faith so he had to plead the Fifth, I don't know what to say. Maybe you're right. Maybe him writing one thing in his diary about his personal opinion ("Covid will be no worse than the flu") and doing another publicly ("it's the Black Death, mask up, isolate, shut it all down!!!!) isn't relevant.
When the people running the hearing have literally said "Fauci should be in prison" then yes, you should presume that they are holding the hearing with the intent of putting Fauci in prison.
https://nypost.com/2026/08/10/us-news/fauci-warned-of-miscarriage-risk-linked-to-covid-vaccine-while-publicly-claiming-no-issues-newly-released-texts-show/
Scandal rag, obviously. Miscarriage was first noticed in Israel (which got the vax first), and was a major reason for minimal uptake for those particulary concerned about fertility (ultraorthodox).
Relevant to the origins of Covid. That's what "relevant" means in this context.
> I didn't read any excerpts
Let me know once you've done that and aren't just going off of headlines
But Alex, that would require me to think!
Look, I never liked Fauci because his public image turned me off. I didn't much care one way or the other because American recommendations didn't affect me that badly over here in Ireland, our government were taking advice from other sources.
But the presentation of him as some kind of secular saint who was being persecuted by the Evil made me roll my eyes, particularly when looking the guy up and finding he had already been controversial during the AIDS crisis when he got a lot of criticism for his role, and it did seem that this was a chance for further rehabilitation of his image.
https://en.wikipedia.org/wiki/Anthony_Fauci#HIV/AIDS_epidemic
But then, all political appointees end up aping the mores of the political apparatus, including burnishing their public image and thinking of themselves as the most important world-historical figures.
Far be it from me to offer universal praise of any government official, and particularly during COVID. But it's completely clear to me that Fauci was brought to testify before Congress entirely because, as a sort of "public face" of the federal government's COVID response, he's become the symbol for every single complaint and conspiracy theory related to the pandemic, not because of what he actually did. Politicians and other 3 letter agencies fucked up plenty (it's not like the head of the National Institute for Allergy and Infectious Disease can make states adopt policies), but for some reason almost all ire is directed at him. And there's the bullshit that he was responsible for the research that lead to the pandemic in the first place.
> https://en.wikipedia.org/wiki/Anthony_Fauci#HIV/AIDS_epidemic
Ironically, this seems like exactly the same thing I was just describing, just from the left instead of the right. This article makes it seem to me like he was blamed for all of the US government's response to AIDs in the 80s, even though he was doing more than most to help.
Not all political appointees. There was a guy Trump appointed that politico was calling "The Man Who Nobody Knows."
Aka: Public Image, what's that?
And yes, you've probably never heard of him.
I was thinking the same.
My guess is, "No."
We got articles in 2025 with content like this from the BBC: "Germany's foreign intelligence service believed there was a 80-90% chance that coronavirus accidentally leaked from a Chinese lab, German media say."
That didn't move the needle (maybe because folks figure this conclusion was too political ... though it WAS suppressed until 2025).
I doubt Fauci's diary entries will either.
We got those articles, and then we got "oh, and germany's foreign intelligence service totally lied to the public about their belief." I think the latter is far more eyebrow raising than simply "we thought it was a lableak."
https://www.bbc.com/news/articles/cz7vypq31z7o
OP seems to have either "not considered this" or dismissed the German intelligence. It certainly was within the 4-6 year window OP mentions as "unlikely to come up with new data."
Personally, given the political nature, and Fauci's retirement to avoid being hauled before congress, and Fauci's pardon... I'd have put higher odds on this particular "controversy" finding "new and interesting data"... Just like I'd put higher odds on the 2020 election getting a public release of convincing data in the 4-6 year window, because of the change in administrations.
Regarding the COVID origins bet, couldn't both be true ? That is, I think the most realistic proposition is that COVID was a natural virus, that was being studied in the lab, and that lab had failed to contain it properly. Or am I misunderstanding the terms of the bet ?
The lab leak theories generally involve some amount of work done by the lab to change a pre-existing virus. You can propose a theory where this didn't happen, but if so then all of the alleged "signs of genetic engineering" in the genome are irrelevant, as are things like the DEFUSE grant application that supposedly indicate the WIV was doing gain of function research of a kind that is likely to produce COVID. And at that point, why reason do you have to even suspect the lab? Literally just the fact that it's located in the same city, even though all the epidemiological evidence points to a zoonotic origin?
> And at that point, why reason do you have to even suspect the lab? Literally just the fact that it's located in the same city
Well, yes? If the question is how some exotic mammal virus got to a city thousands of miles away from that animal's natural habitat, then the idea that it was brought there by the big lab that studies exotic mammal viruses seems reasonable.
And of course, there's the huge mass of *other* evidence that it was zoonotic.
The SARS 1 pandemic was also very far away from the caves where the ancestral bat viruses would be found. Raw distance is not a very relevant factor when vehicular travel exists; we know that live animals were were being kept and sold at markets in Wuhan.
The most likely suspect for the lab-leak scenario explicitly proposed to do a bunch of lab work to make natural bat coronaviruses more human-contangious than they were. Maybe they actually did that and screwed it up, or maybe they hadn't gotten around to it before nature preempted them, but the existence of a group in Wuhan that was proposing to do this is a big part of what makes the lab-leak theory as plausible as it is.
But isn't essential; the general concept of "lab leak" extends to a lab that was just studying bat coronaviruses in their natural state and lost control. The existence of bats with coronaviruses that occasionally infect humans was part of the baseline; the question is what caused the virus to jump to a predominantly human form of transmission. If the answer is that a bunch of scientists took them out of their natural habitat and put them in a place where there are no live bats to infect but many live humans, that's on them.
Did you mean to reply to someone else?
No. You seemed to be claiming that it doesn't count as "lab leak" if it doesn't involve some sort of deliberate modification of the virus, and I was trying to explain why I disagree with that claim.
I have no idea how you got that from what I wrote. My point was about the evidence for such a theory, not whether anything "counts as" lab leak.
Xi's publicish commentary, among other things, that China needs to do a better job securing their biological facilities.
https://www.vanityfair.com/news/2022/10/covid-origins-investigation-wuhan-lab
"The WIV’s deputy director of safety and security spoke next, summarizing “several general problems that were found over the course of the last year during safety and security investigations, and [he] pointed to the severe consequences that could result from hidden safety dangers.”
"In March 2020, a veteran journalist with the South China Morning Post reported that she reviewed internal Chinese government data on early cases of COVID-19 that included a 55-year-old in Hubei province, where Wuhan is located, who contracted COVID-19 on November 17, 2019."
Poring through publicly available documents, Kadlec’s researchers saw that China’s top scientists had been sounding the alarm too. “The biosafety laboratory is a double-edged sword; it can be used for the benefit of humanity but can also lead to a ‘disaster,’” warned a March 2019 article cowritten by Yuan Zhiming, director of the WIV’s BSL-4 laboratory. “With increasing numbers of high-level biosafety laboratories constructed in China, it is urgent to establish and implement standardized management measures.”
https://www.newsweek.com/us-detains-chinese-scientist-wuhan-2083229
Wuhan scientist conducting biological terrorism...2025.
Why suspect a lab? The simplest reason to suspect the lab is that it was a bad lab, with numerous safety violations and leaks -- it is not prudent to assume that there was simply one lableak, if you assume that there's a problematic lab -- of course, lableaks can come with various levels of graveness, from "a single person got jabbed, no illness reported" to "MERS-level spread" -- you don't necessarily need to be talking about a viable viral escape which spreads population wide.
The next simplest reason to suspect a lab is that it's running genetic engineering experiments, and the spike protein looks an awful lot like a universal key. Coronaviruses don't generally infect the gastrointestinal tract, or wind up dramatically mutating within a host, which is required if you believe that omicron wasn't a lab-released virus.
Further notes:
https://usrtk.org/covid-19-origins/cia-whistleblower-alleges-covid-lab-leak-findings-were-suppressed/
I'm putting low confidence on the CIA's ability to gather human intelligence on covid19, given the large-scale winnowing of our agents in China under Obama. This doesn't discount their analysts, just their handlers/foreign agents.
Old History Lesson, citing the vanity fair article:
Meanwhile, in two separate incidents in 2004, SARS accidentally leaked from a top laboratory in Beijing and led to mini outbreaks.
Continuing citing the vanity fair article:
"In March 2018, Shi partnered with Baric and a longtime collaborator, Peter Daszak, on a $14 million grant proposal to genetically manipulate bat coronaviruses to see how they might cause pandemics. The proposal called for possibly enhancing the viruses with something called a furin cleavage site to boost their entry into human cells. The Defense Advanced Research Projects Agency (DARPA) rejected the grant proposal for not adequately assessing the risks posed by a supercharged virus."
Incoherent gish gallop of irrelevant nonsense, not worth paying attention to
Either "origin" means who created it (nature or a lab), or who failed to contain it (nature or a lab). Both can't be true at the same time, whichever it is.
Right, I guess I was implicitly interpreting "origin" as referring to the COVID pandemic, not to the virus itself, but I agree that your interpretation makes more sense.
Thanks for signal boosting us, Scott! Minor correction: we make a voting guide annually, not biennially. Even-numbered years are federal and state elections; odd-numbered years are municipal elections (in most of Massachusetts).
The Save Our Bacon Act has been NOT YET been successfully defeated, despite the win last week. Senator Boozman, the chairman of the Senate Ag Committee, has stated that he plans to reconvene the committee and try again to reach a deal on the farm bill. Although the draft that was voted on last week has died, there is nothing stopping Boozman from introducing a new draft of the farm bill, which could include the Save Our Bacon Act in the base text or in an amendment introduced during markup.
Two days before the Senate Ag Committee meeting we were hearing that Senator Ernst planned to introduce a SOBA amendment and that some democrats were willing to support a 'compromise' version (one that would grandfather in Prop 12 in CA and Question 3 in MA but prevent similar laws from being passed in the future). That would be very bad because state-level animal welfare laws have been the most effective tool to help factory farmed animals.
The reason that SOBA failed was likely only because of McConnell's absence, leaving Republicans without a majority, and because Republicans and Democrats couldn't reach a deal on SNAP. If one or both of those things change by September, it could be very difficult t