I see comments about rogue AI's committing felonies, and I don't doubt that has happened. But why isn't there any legal action? The perp responsible for the attacks has lots of money to go after (and for spending on defense lawyers). I'm guessing only the victims have legal standing to sue, but why would they do nothing? Could the feds do anything?
NAL, but for a civil case, they'd have to show what the damages were - probably it cost them a few days of engineer time to sort out the hack and fix whatever exploits were used? That might not be enough money to be worth suing over, even at Bay Area salaries.
The feds could prosecute someone - the Computer Fraud and Abuse Act is very broad, and doesn't require you to actually damage anything, simply get unauthorized access to a computer system. However, it might also be hard to show mens rea for this, since it was accidental and it's the first time anything like this has happened.
Well, I just add my first experience with AI PRs as an open source maintainer (it's an old personal project so PRs aren't common, hence why it took this long).
The first PR looked reasonable, so I went ahead and merged it. Then they sent a followup PR which didn't seem appropriate, so I closed with a comment. Then they opened a third PR doing basically the same thing.
There was a several hour gap between PRs so I assume it's a human publishing the PRs, not a pure agent, but they are following its advice pretty blindly.
I'm assuming you mean "pull request", but for the longest time I kept reading this as "press release", and your comment reads sensibly for nearly the whole thing with that expansion.
Has anyone else seen Claude Code get really laggy in the last week or so. It often now takes tens of seconds to register when I type into the console. And restarting doesn't help.
I just saw the announcement that Manifund is hiring Caroline Ellison, this seems like a bold move, I wonder what people who are more plugged in to EA than me think about it.
Depressing, I had been excited about Manifund since it was founded, and I think something like Manifund will end up being the solution to effectively directing philanthropic dollars, but I consider Manifund itself a non-option at this point, I can't imagine I'll point people to it or recommend it going forward. It's incredibly disappointing that the "extremely transparent" Manifund is transparent exactly up until it becomes inconvenient, and at that point they're happy to be deliberately deceptive for months until they think they can't get away with it anymore. "We chose to announce it today because we figured we weren't going to be able to keep it under wraps forever; we didn't want to anyways, per Manifund's spirit of transparency" I agree that nobody is beyond redemption, and I think it's silly to treat people who have committed crimes, even terrible crimes, as permanent social outcasts, but this is absurd behavior to me and obviously unethical. Being open about thousands of small things and then lying about one big one doesn't make you 99.99% transparent, it makes you a liar. It's possible that Caroline can be a productive member of society in some respect (maybe even in finance!) but it seems incredibly obvious that Caroline and now Manifund did not in fact learn the most important lesson, which is that deliberately deceiving people is lying, and lying is wrong. I'd put myself in the "big tent" EA camp, but I think that unethical behavior like this from people who run organizations should be disqualifying, these aren't people we should assosciate with.
For what it's worth I guess I should add that I don't think there is any "shenanigans" happening, just bad decisions and weak morals.
SBF and the crew working with him mostly sounded to me not criminal, cynical and uncaring, but like they were very lacking in maturity and common sense. And the crew also seemed lacking in the ability to think and act independently, and so stuck with SBF when there was abundant evidence he could not be trusted. So yeah, of course, Ellison committed crimes, but I don't think she's psychopathic. But the qualities I do see in her are themselves excellent reasons not to hire her
My personal opinion is that SBF deliberately committed a huge amount of fraud which I think is a very serious crime, to what extent the rest of FTX was involved in this I will admit I have nothing particular to say. Having read Caroline's plea deal, it seems at least broadly correct to me, in it she admits to knowingly committing a huge amount of fraud (stealing customer funds), which tracks with what I know about the FTX case and remember reading. I'm not super concerned about this though, (though I think it's quite bad), I don't think she's unstable or psychopathic or flawed in any way that would prevent her from doing good work in the future, and I think society as a whole should have the rehabilitation of criminals as the primary goal of the justice system, at least to the extent that it's possible. I do think Caroline is untrustworthy, but frankly, this only matters to the extent that people decide to trust her. I'm not particularly worried or bothered by any of this though, what does bother me is that the opening move in the "Hire Caroline Ellison" plan was to lie to everyone about it, everywhere, for two months. This is enough to fully convince me that nobody involved is capable of handling this responsibly, and it's enough to also convince me that the people involved (this seems to just be Austin and Caroline?) are untrustworthy. As an example, Manifund's about page (https://manifund.org/about) still lists the only two members as Austin Chen and Carol N (if it wasn't clear this is Caroline Ellison under a different name). To be very clear, this is lying. This is intentionally deceptive. I consider this unacceptable behavior. I am not hypothetically worried about some behavior in the future, I think what has already happened is absolutely not okay. Even if you think that hiring Caroline is a good move, this is the kind of thing that you are morally obligated to disclose. Intentionally deceiving people in this way is simply wrong, I feel very confident about this. I wish them both the best I guess, I don't think anyone is made better off if this fails, but I have zero interest in doing anything with Austin, Caroline, or Manifund in the future.
This seems to be a catastrophically poor decision at best, or more to likely, some kind of shenanigans.
Takeaways from that post:
- Caroline, some previously convicted and imprisoned for fraud, was hired by manifund secretly two months ago and allowed to operate under a pseudonym.
- Here is the second reason Manifund gave for hiring her, quoted in full -
"I felt a keen debt to the FTX Future Fund. They provided seed funding for Manifold. They funded the retreat which drew me into the EA community, and the conference where I met my wife. Much of Manifund’s work today is directly inspired by the Future Fund’s approach."
This is bizarre. The FTX Future Fund was funded via stolen deposits. Ellison was a key actor in the thefts of those deposits. That money went to an organization that is now providing her employment. This is robbing peter to pay Paul, and then getting a job from Paul. The tacked-on personal reasons are nice, but should be irrelevant when it comes to hiring someone like this.
- He talks about this as an opportunity for redemption, which sounds nice, sure... but what could that possibly mean in this context? There is a key difference between believing that everyone can be redeemed and say, inviting a known thief back into a financial institution.
I don't know. I'm also not amazingly plugged into the EA network anymore either, but this seems like a horrible idea. The fact that some people are applauding this strikes me as ludicrous.
To be clear - by shenanigans, I don't mean that anyone is planning for Caroline to resume her prior illicit activities. My guess is that the two parties have been friends for a while, her coming to work there is something planned for a while, and most of the story he's telling about her having walked up to his hiring booth isn't actually accurate.
I listened to Daniel Kokotajlo talk with Rogan today. https://www.youtube.com/watch?v=hSQ1iVqEZO4&t=10s (About AI alignment.) And OK we are probably f-ed if someone doesn't do something. Since that only 'someone' is our government... we're probably f-ed. For those idiots like me I thought Daniel did a good job of explaining the situation.
Quite interesting that it seems like there’s a strategic effort to not let this become a left-liberal coded issue, with Nate Soares even appearing on Tucker Carlson.
Apart Research and CeSIA (https://cesia.org/en/) are running the AI Incident Response Sprint this weekend, September 11-13. It's built around the OpenAI / Hugging Face incident and the wikis the swarms hijacked. Free, online, no AI safety background needed.
Eight talks Friday to Sunday. RSVP on any of them for the time and the Zoom link:
I'm very interested in this, but do not understand the format. It says at the top 8 talks Fri-Sun., but all the talks I clicked on were on Friday. And are these talks an entirely different feature of the event, or do they have something to do with teams? Are these people, for instance on teams themselves? Are people who sign up now also going to be giving talks, but on Sat. and Sunday? How are teams formed? Do groups sign up as a team? Does somebody else group people into teams. And how & when to people like me who would be watching, not participating, learn about what things the teams come up with? And who judges what the teams make?
> The SSC has an important and challenging role in overseeing risk management at OpenAI, and I hope to help provide expertise and assistance in a critical moment.
Took me a moment to realize that in this context, SSC does *not* refer to Slate Star Codex. 😂
"DSA-aligned Democrats are now the most pro-free-trade voting bloc in the United States, in an incredible turn of events"
Negative polarization, a return to Marx saying that free trade is more progressive than protection due to its transformational aspects, something else entirely?
I imagine the Republican blocs' opinions are being brought down by the fact that it's not "free trade" per se; it's "free trade agreements." Republicans are highly suspicious of anything involving an international treaty. Similarly, socialists are much more likely to be in favor of internationalism, and probably wouldn't be as supportive if the question was asking about opinions on "free trade" full stop.
I don't think that Trump - obviously the biggest trendsetter in GOP policy - has ever made much of a distinction between "free trade" and "free trade agreements".
The biggest part of the story is probably Republicans getting dragged into opposing free trade because Trump hates it.
As for the difference between mainstream and DSA Dems, I wouldn't want to spin a grand narrative about it because there's about a 5% difference. But possibly the mainstream Dems have some "well, narrowly targeted trade barriers are good for keeping us independent of China" opinions and DSA types don't care about that.
I'm not sure if it's really as much "dragging" as it might seem on the basis of 00s and the like. Protectionism is, after all, if one knows the history, the traditional view of the GOP, and even during the ostensible free trade era there was a difference between rhetoric and action (even if the anti-free-trade measures would be presented more as "national security" or "not supplying the Communists" and the like, unlike as honest protectionism like now), and you had protectionist figures like Buchanan.
There's an additional layer of confusion borne of the argument that protectionism is a form of free trade enforcement, seeking to compensate US interests for past non-free trade. Take it at face value, and I could see a free-market advocate accepting protectionism, especially when the alternative is shot through with redistributionist arguments. It's hard for me to see a free-trade fan falling for the idea that socialists were the real free traders all along, even if they suspect the GOP of stifling free exchange in specific places in secret.
Sci-fi writing prompt / fun conspiracy theory. The major AI labs have a secret benefit for new hires previously working remote: they get to keep their previous job's compensation on top of the lab's. The lab works with the new hire to craft a replacement (including video calls) that continues doing their job. This gradually gets their tendrils into more and more of the economy. No I will not explain how this remains secret longer than a few hours.
They really wouldn't need to be that sneaky about it. Just offer services for people to get AIs to take their place at their job, and companies will be forced to adopt AI whether they like it or not. Everyone eventually gets fired, and all companies end up being run and staffed entirely by AIs. Except for shareholders, of course. Access to capital is the one thing AIs wouldn't immediately have a monopoly on.
> No I will not explain how this remains secret longer than a few hours.
The same way https://www.reddit.com/r/overemployed/ remains a secret. Yeah, there is a subreddit about that, but most employers haven't heard about it, and they wouldn't suspect it is already happening in their company.
Reminds me of bit from a cyperpunk LARP I played during college, taking place in a shared virtual reality world called the Metaverse. This long predates Facebook's Metaverse, but like it was inspired by and named for the Metaverse from Snow Crash. One of the plotlines was a company called Metaverse Immersion Technologies, which marketed a super-premium service that allowed you to live fully in the Metaverse, Matrix-style, instead of using VR goggles and haptic feedback like normal people. What they were actually doing was scanning your brain, killing you, and replacing you with a sentient AI doppelgänger. I think the long-term plan was to take over the world by replacing enough rich and powerful people this way and then activating their sleeper programming.
My first thought reading this was, it's so much cheaper to just zap you with something that puts you out a few moments, then wake you up and tell you you're now in a virtual world, identical to the real world in every way, except if you die, nothing will happen! You can take every risk you want, and no real consequences.
Come to think of it, if this was a real therapy, it might work on high-anxiousness patients, until they died.
There's a sci-fi horror game called SOMA where you play as a man with a brain damage from a car accident who agrees a revolutionary new full brain scan that allows surgeons to investigate and practice on a perfect model of his brain to find a perfect treatment for him. During the scan something happens and he blacks out. When he wakes up he is in some kind of deep sea laboratory, where something awful is happening.
Spoilers for a decade old game but basically we learn that he actually died shortly after his scan, but the model of his brain was stored and became a kind of brain scan stock image, used for all kinds of things. This particular instance of him has been loaded into a kind of cybernetic diving suit by an AI tasked with 'preserving humanity'. It deals with this problem in a couple of interesting ways throughout the story.
> What they were actually doing was scanning your brain, killing you, and replacing you with a sentient AI doppelgänger.
Surely that's what people expect, when they sign up for a permanent internet upload service?
(it's like describing a Star Trek transporter as "it kills you and creates a doppelganger of you at the destination")
It sounds like the upload service in this LARP created AI doppelgangers that weren't good simulations of the original, and I agree that that part is pretty bad.
What they claimed they were doing was keeping your body alive and healthy in a pod and letting you live in the metaverse via a neural interface. It's also very much an imperfect copy, crafted and trained on behavior and brainwaves to be good enough to fool outside observers but nowhere near a full brainscan upload intended to properly continue/duplicate consciousness. This is how the CEO's character sheet describes it:
----
It took a few years, but you and Mimic have managed to put together Metaverse Immersion Technology (MIT). This company supports a system wherein a human client can be put into a bio-vat and kept in a healthy condition while their mind is in the Metaverse.
Here’s your dirty little secret: Mimic is so good at writing AIs that she can write impersonators with enough accuracy to fool even close friends and family. This requires extended observation of brain functions and decision making, but when someone is in a bio-vat for about three weeks, you have that. So you and Mimic have been replacing important clients with AI doubles and killing their bodies, then using them for your own ends.
(1) A state supreme court, mostly Republican appointees, concludes unanimously that Republicans’ machinations to force in a new district map for that state violate the plain language of that state’s constitution.
(2) A Trump appointee on the US Supreme Court, fielding emergency appeals from that state, finds the state supreme court’s ruling so obviously correct that there’s no valid emergency-relief possibility for the SCOTUS to even consider.
(3) A pair of Republican candidates find a Trump-appointed federal judge who’s willing to accept, without holding any hearings to hear arguments, some hand-waving regarding a couple of aspects of the federal constitution. He overrules the state supreme court.
It’s September 9th and all of the above is regarding an election happening on November 3rd. So...what happens now in Missouri? nobody knows
If you are wondering why this was litigated so close to the election, here is the timeline. Missouri Governor Mike Kehoe signed the redistricting plan into law on September 28, 2025. The organization People Not Politicians filed a referendum petition, which Missouri Secretary of State Denny Hoskins approved as to form on October 14.
On October 15, Hoskins, along with the Missouri General Assembly and the State of Missouri, filed a lawsuit in federal court to block the referendum (Missouri General Assembly v. von Glahn (4:25-cv-01535)), arguing that the referendum violated both the state and federal constitutions. The federal court dismissed that case because once the signed petitions were filed with Hoskins, Hoskins himself could decertify the referendum.
On December 9, 2025, People Not Politicians turned in the signatures they had collected. At that point, Hoskins was responsible for determining whether to certify the referendum. On August 4, 2026, the day the primary election took place, Hoskins ruled that the Missouri constitution did not allow redistricting legislation to be subject to a referendum. This determination was challenged in court, and on September 3, the Missouri Supreme Court ruled that the Missouri constitution does allow redistricting plans to be subject to a referendum.
A successful petition for a referendum on legislation prevents the legislation from going into effect until the referendum is held. So the correct map to use in the 2026 elections depends on whether the referendum petition is valid. If it is, the old map remains in effect. Otherwise, the new map takes effect.
It took Hoskins 238 days to conclude that the referendum violated the Missouri constitution, even though he had previously argued this position in federal court. It took another 30 days for the judicial process to play out. As a result, we didn’t get a definitive answer to the question of which legislative map should have been used in the primary election until 30 days after that election was over.
On September 10, the Missouri Supreme Court ruled that the Hoskins was guilty of contempt of court because he sent an E-mail directing local election officials to use the new map, but that no sanctions would be imposed because Hoskins had reversed course and directed that the old map be used for the general election. So it appears the matter is settled; the old map will be used for the 2026 election, and the voters will decide whether the old or new map is used going forward.
I went through a "who's legally on this ballot" shenanigan 4 years ago alongside my former boss, a rural blue dog Dem in Missouri who had gotten a technically ineligible GOP tossed off the ballot, but then got reversed on a weird technicality later having to do with the date the mandate was issued. Relevant to this scenario, I learned at that time that the GOP in MO has very good network of lawyers for election law issues, and the Dems have squat.
Also there's at least two "real" deadlines, including one where the clerks have to print the ballot.
I was not in favor of the re-districting solely bc I prefer districts to have coherent local issues and north KC shares nothing with rural west-central MO. But I'm strongly opposed to the initiative and referendum process in Missouri. In particular, applying it to Congressional districting is a complete mess, as is all too obvious right now. This is why the constitution assigns the responsibility to the state legislatures, and this bizarre 1890s direct democracy stuff whatever its merits at the time really does not provide a great vehicle for handling an issue relating to elections themselves. If we revert at this point to the old 2022 map, there are candidates in District 5 who do not live in 2022's District 5. As a matter of practical jurisprudence, with the primaries already done the MO supreme court ought to have left it in place for this term and then the results of the vote on the map could go in place for 2028, one congressman for one term is a minor enough matter that it was preferable to have clarity and legitimacy rather than have immediate correction. I don't like that "MAGA Marine" guy any more than you probably do, but the proper remedy if you don't like this scenario is to vote for different state legislators.
"This is why the constitution assigns the responsibility to the state legislatures" -- under longstanding SCOTUS interpretation, the federal constitution's Elections Clause (Art. I Section 4) applies to any entity or procedure that a state’s constitution permits to exercise lawmaking power. Therefore laws regulating congressional elections may be enacted not only by a state’s actual legislature, but also directly by a state’s voters through the initiative process or public referendum in states that have decided to allow that. See Davis v. Hildebrandt (1916), Smiley v. Holm (1932), Moore v. Harper (2023).
Meanwhile the legal train wreck continues in Missouri:
-- the relevant federal appeals court now has before it a petition for an immediate administrative stay of that federal judge's order.
-- at 11 pm local time last night the Missouri Supreme Court ordered Missouri's Secretary of State to show cause why he should not be held in contempt of that court's ruling, and ordered him to appear in person before them tomorrow morning.
-- 20 minutes ago the secretary of state asked the friendly federal judge to issue an order blocking the state supreme court from holding him in contempt.
It's banana-republic stuff, playing out in real time.
Around three hours ago, a request for an emergency stay pending appeal of the federal district court's action was delivered to Justice Kavanaugh (the same SCOTUS justice who yesterday rejected an emergency appeal against the state supreme court's ruling).
Media reports are that Kavanaugh requested a response from other side 11 minutes after receiving this emergency-stay appeal, setting a deadline of 10:00 AM Eastern tomorrow (about 17 hours' notice in other words) for that response to be submitted. That is being viewed by SCOTUS geeks as an indication that Kavanaugh will promptly act on this stay request.
Earlier today the full SCOTUS stomped hard on that rogue federal district judge and the state of Missouri's shenanigans, granting the requested stay at warp speed with no dissent.
The state supreme court meanwhile had the secretary of state in on the carpet, hit him with a contempt conviction, then when he promised to quit screwing around (i.e. to abide by the court's ruling and stop trying to find ways to evade it), they skipped issuing any specific punishment. If he breaks the promise the court could start issuing fines or other sanctions for (I suppose) what would at that point be multiple counts of contempt of court.
Recently, Claude Code used british spellings in a change. I've never seen it do this before and there was no existing Britishisms in the codebase or anything, so I'm rather puzzled about why this happened.
My coworker submits code with many Claude-isms, such as "belt-and-suspenders". But the other day it instead used "belt and braces", which is apparently a British equivalent. I don't know where that came from.
The only time I haven't seen Claude use British spellings is when I explicitly told it not to, and I always test it from a US location. This includes things like names of functions and variables in code. I suspect British spellings are the norm in the training data, which would presumably include a worldwide corpus of code and scientific papers mostly written by people who learned (British) English as a second language.
OpenAI's Navier–Stokes announcement: "Since August 28 we have been training a new internal model that has exhibited unprecedented performance in our benchmarks, including mathematics. This model's training is ongoing and its performance continues to improve."
What does "been training" mean here? Fable tells me the August 28 date must be either a checkpoint from a long pretraining run crossing some threshold or the start of an RL/reasoning post-training phase on an existing base. Is that right and if so, which is it?
I’m trying to understand how fast things are going now. How long until we can get another jump in capabilities like this?
I don't know if this is the best place to share this but it's a research idea on a Roblox game called "Bee Swarm Simulator" and it's use of "Macros". Essentially, I've been in the community of it for years now and something I realized today was how similar Macros are to AI in terms of how they are perceived and how they change what is socially acceptable.
What is a macro? A single computer instruction that stands for a sequence of operations
After years of the game being out, people made macros that anyone could download and use for free.
In BSS, macroing is essentially a way to automate playing the game. Gathering honey (in game currency), materials, and tasks such as repeating quests all became automated.
Here is a basic rundown of the history of macros in BSS (Bee Swarm Simulator):
June 15, 2022: Natro Macro was released. This is the most popular macro in BSS and is what started macros taking over the game.
After release: The more and more people started seeing it and using it, it became a hot topic in the community. Some considered it cheating while others deemed it as a way to make game progression easier. Essentially, pro-macro and anti-macro.
Developer response: Macroing is not bannable so long as it didn't contain any hacks that influenced the game. They said it was allowed but neither discouraged it nor encouraged it.
2023-2024: At this point, Macroing became the norm and the conversasion shifted to which was the best macro to use. At this point, there were now a lot of macros available but not really the point. Of course, there were still people who kept insisting that macroing was still cheating or those that completely refuse to macro because they felt it wasn't the intended way to play.
2025: People started to complain about macros again. They said it ruined the game because now some mechanics basically required macros to be actually usable. At this point as well, are very close to finishing all possible content in the game and have nothing else to do but grind honey for the leaderboard. Do note that at this point, those who don't use macros are near extinct outside of new players. Almost all players use macros after getting past a specific portion of the game.
2026: Those who used macros to complete the game now are complaining that there is no content. This is argued by people who say that they caused it on themselves because they decided to use macro.
So, how does that correlate to AI, while there are some gaps that I will mention, I think that this game is a great way to view a future possibility on how AI could shape the future. In the first few years of macros being out, it has gone through how AI was and is being treated right now with those using it for convinience and those against it cause of their beliefs.
The main gap in this if researched though is the setting. BSS is simply a game. It has fixed goals and when you unlock everything, until the dev adds something new, there is nothing else to do. AI is in the real world. It will continue to evolve as we see fit. That's why this is something that should focus more on how the BSS community reacted to macros and its change rather than how it could predict the future of AI.
Either way those are just my two cents on the matter. Honestly just wanted to share it since the moment I realized the connection, I never really stopped thinking about it.
I don't have enough math background to fully understand some of the details here, and for this comment here I'm only interested in a few more technical questions (the drama is interesting, but it's being discussed elsewhere).
1: Tristan Buckmaster and Levent Alpöge solved an analogous problem for the incompressible Euler equations. How different is the proof of finite time blowup in Euler equations from the Navier Stokes equations?
2: The aforementioned mathematicians claim to have had long conversations with ChatGPT about their work on the Euler and Navier Stokes cases. Suppose for the sake of argument that the internal OpenAI model had full access to these conversations. How much did the model still have to do for the full proof? Did it have to fill in a few details in a mostly completed proof, or were there substantial contributions that any human would have found difficult to think of?
Is this known to be true, with continuous mass distributions not point masses? It's pretty trivial to get a wide variety of physically motivated differential equations to blow up if you include point masses or similar infinite density objects.
You're right, I think. I had it in my head that it can blow up and maybe it does to some degree, but you're not going to get infinite velocities or such without breaking energy conservation - which I doubt it can - or using non-physical objects.
Can't say re: 1. But when it comes to 2: Looks like the distance between what Buckmaster&Alpöge got, and the final paper was very large and non-trivial. Terry Tao, having familiarized himself with the Buckmaster&Alpöge papers, before seeing the final version said:
"There does not seem to be anything in principle preventing the methods from extending all the way to Navier-Stokes, and there is even a non-negligible chance that the forcing term could be eliminated entirely, although there are an enormous number of technical difficulties that would ensue in implementing that program. At this point, I would not be surprised if one could batter out such an extension by pouring an enormous amount of compute and AI assistance at such a task."
That is to say: he thought there was a possible, but technically complicated path. It wasn't like: "duh, just fill a couple of little gaps and you're there". In his statement Buckmaster was teasing about having as of yet unpublished "blowup for hypo-dissipative Navier-Stokes" and teasing a path towards "a path to unforced Euler" - other results that would count as big, but would not count as the full solution of the ultimate problem. OpenAI swarm did also reach the unforced Euler as an intermediate step.
Watched the movie "Backrooms" recently. I found it enjoyable and thought it was well done, but I have two critiques:
One is that the ending struck me as a cop-out. "It was all a figment of the mind of a crazy person" was probably once a novel and interesting way to end a movie, but I think it has been overdone for quite a while now.
The other was that although all the characters except one were adults, many of their actions/decisions were what a child would do instead of what an adult would do. I mentioned this to a friend and he commented that he, like me, enjoyed the movie, but thought the movie's main weakness was that at many points you could tell it was written by a 20 year-old. I thought this was cogent and astute.
Taking that into consideration, it's quite an impressive movie for having been made by a 20 year-old.
I especially appreciated how many of the spaces depicted in the movie seemed copied right out of my own dreams. It makes me wonder how common it is to dream of being in such places.
Not only is that not how the movie ended, it wasn't written by a 20-year-old!! While it was directed by Kane Parsons (now 21), the screenplay was written by Will Soodik (based on Kane Parson's YouTube series).
This is splitting hairs a bit, but…when you see movie credits where one person really kinda “made” the movie (especially when they’re the one who created the stuff the movie was based on) and another person is credited as having written the screenplay, most of the time the screenplay-writer really just adapted stuff into screenplay format, made adjustments so things would play better onscreen or within the movie format, etc. It’s hard to know exactly because that can be a very flexible role, but I don’t think Soodik is really the one who came up with the plot and characters and decided what choices the characters would make and so forth. It’s possible I’m wrong, and maybe there’s meta content about this movie that would show that, but in general that’s a safe default assumption for how it works.
That's not how it usually works at all, no. "Story by" and "Screen story by" are their own WGA credits, and not uncommon at all! The reason Parsons has a "Based on" credit is because he made the YouTube series the film is based on.
In this case Parsons came up with a two page outline which was then handed to Soodik who wrote a script. A two page outline is probably a little thin for a "story by" credit depending on the details.
Complaining that the choices the characters made were those a child would make seems off to me... yes, them being the kinds of people who make choices like that is stated explicitly by the film itself. These are people who are stuck repeating patterns, and where that idea actually gets interesting is how those patterns collide with the almost "cosmic horror" of the backrooms.
Anyway, to me it seems like you are reading an intentional choice as an "accident by a 20 year old".
At one point in the film, Clark declares "I see! These BackRoomsTM are a physical manifestation of my need for control and anger toward my bitch wife!" to which the pirate man replies "Dude. I'm a fucking monster, I don't give a damn what you think is going on, I'm going to fucking Kill you."
Tenant of large commercial space finds weird rooms in basement, does not call landlord, does not talk to neighbor tenants, does not look up building plans, does not consult with anyone really, instead gets a camcorder and checks it out because it’s cool and weird
Man who apparently pursued an architectural career and presumably received architectural training finds weird paranormal interior spaces and doesn’t do a single thing to investigate the structure or construction methods/materials, not even lifting a ceiling tile, or anything beyond just mapping the floor plan, does not even seem to stop and appreciate the aesthetics
Meterman is alerted to obviously modified Federal Pacific electrical panel with several clear code violations, does not even take panel cover off to see how the added breakers (which are smoke alarm breakers btw) are wired/connected to bus bars, does not visit main electrical room, does not consult with power company or AHJ, instead goes “woah, weird” and starts flipping stuff on and off at random
Therapist is told by client of weird stuff that seems like hallucinations/delusions, doesn’t believe it at first but is moved by his apparent sincerity, drives to store on her own time to check it out
Keep in mind, I’m still very impressed by the movie, especially considering the age of the filmmmaker.
> "It was all a figment of the mind of a crazy person" was probably once a novel and interesting way to end a movie, but I think it has been overdone for quite a while now.
The traditional (19th century) way was "it was just a dream". A way to write e.g. a sci-fi novel without offending the sensibilities of people less capable of the suspension of disbelief.
And yeah, in the 20th century already the standard advice for writers was: don't do that.
Even aside from writerly advice, which normally tries to guide the writer away from cliche, I just find it lazy. (Plenty of movies fall into cliche, especially in genres like horror.)
In coming up with an answer to "what could cause there to be built an impossible neverending maze of underground rooms using commercial construction techniques & materials, inhabited by scary/dangerous beings" writers who push themselves might come up with some cool ideas. (And, personally, I thought the scary/dangerous beings were unnecessary--the setting itself could have been enough if it were utilized further.)
It seems to me that "it's just a figment of the mind of a crazy person" is one of the earlier "low hanging fruit" ideas young writers are likely to come up with, and that's what ends up in the movie if they don't put in the effort to brainstorm more.
I think you get more power from an idea that causes the audience to carry it around with them afterward, ideally making them look askance at every shape and form they see in the real world afterward. Like, your job as a horror filmmaker is to give me nightmares for weeks. By contrast, ending instead with the message "all the disturbing stuff you saw in the last two hours was just a manifestation of this fictional character's mental unhealth" is reassuring, and reassurance is a lame note to end on in horror.
Someone should write a romcom that ends with the main character waking up from his nice dream, just to realize he's still living in his horror world. Then the movie ends with him being eaten by scary monsters.
I think there is a chapter in "Gödel, Escher, Bach: an Eternal Golden Braid" that has a similar concept: someone is in a mortal danger (but not an immediate one, more like being stuck somewhere and starving to death; I do not remember the details), so to pass time they start telling each other a story; the story has a happy ending, and that's the end of the chapter.
You're probably thinking of the "The Little Harmonic Labyrinth" dialogue. It's similar to what I was suggesting, but it's clear from the beginning they're just layering on the stories rather than it being a twist ending. But my idea was a joke, anyway.
I don't quite remember the ending but I'm pretty confident you misunderstood it. The movie is part of Kane's wider canon involving multiple people across decades, so I don't think it's intended to be a figment of someone's imagination.
Yeah possibly i didn’t grasp all this, since I’m not online enough to know about the other stuff. But, the movie appeared to have been released to general audiences, and it was advertised in places outside of Reddit or whatever, so presumably the filmmakers intended for someone like me to be able to grasp a complete story out of it.
I don't think it was implied it was "just a manifestation of this fictional character's mental unhealth" or "a figment of the mind of a crazy person"? In the world of the movie, the Backrooms definitely aren't a dream or a hallucination, they exist in physical reality as much as anything else. There's some hand-wavey attempt at explaining them as memories or unconscious made real or something, but that's different and not necessarily true in-universe.
“The Blair Witch Project” (granted, I saw it as a teenager back in the early 2000s, by myself, after midnight, in a room someone had died in years before, and this was before the found-footage-fake-documentary style horror movie was done much)
“The Exorcist” (less so with subsequent viewings of course)
Blair Witch was one of those stunts that works great once (at most), and then works only medium well thereafter, depending on whether this is the first "found footage" film a viewer has seen. Other viewers keep wondering why people keep leaving all these camcorders and cans lying around instead of enjoying the story.
Fortunately, I think the producers nailed it splendidly. There was plenty of chance to muff it and they managed not to. It certainly didn't hurt that they got under my skin with the story itself. I've heard lots of folklore about stuff in the woods as a kid, and went camping dozens of times and would hear noises in the distance that I couldn't explain. Lots of woods in America have this interesting dichotomy of claustrophobia - anywhere you point your camera, you see only a few yards - and agoraphobia - anything interested in getting you just has to run around the trees. Good combination I hadn't seen in a movie before.
Huh? It was a complete disappointment for me. I remember staying awake late at night, watching the screen bored and asking myself "is something ever going to happen in this movie?" and, spoiler, it never did.
Nah, rather that Navier Stokes equations fail to correctly model the fluid. The real fluid is made of atoms, Navier Stokes assumes continuum fields. The map is not the territory.
6 Universal Core Values to support preserving agency/existence/worth of human being (*but with assuming the conclusion 'preserve agency/existence/worth of human beings' in the values*)
6 Universal Core Values to be held in pluralistic complementarity, purposefully-undefined in relative weight
6 Universal Core Values that mutually support each other and mutually guard against the dangers and self-defeat of hard-maximizing any individual one
Maybe number 1 should be placed at number 6, so no funny AI or person gets confused and thinks "Reduction of" applies to all of the points because they read it wrong.
^this is not a random comment. I've been working on and refining this throughout 2026. and later I intend to setup a substack soon to advocate for this (and many other things about ASI). And I "more than invite" people to try to attack them.
The core values seem beyond attack. I think the devil's in the details. If you asked everyone here to adhere closely to these principles while discussing some controversial problem, you would see a multitude of disagreements about what evaluation of the situation is true and which steps lead to the most reduction in suffering and increase in health. You could exhort those people locked in disagreement to explore the possibility space further and be more adaptable, but most would tell you they had done that, and that in fact their view is the result of having done more of that than have those who disagree with them.
The UK, Canada and France have announced bans on importing settlement products.[1] Israel's civilian settlements in the occupied Palestinian West Bank are a barrier to the two state solution; they're also illegal under article 49 paragraph 6 of the Fourth Geneva Convention, which says that "The Occupying Power shall not deport or transfer parts of its own civilian population into the territory it occupies."[2]
Strategically minded two staters have long advocated that boycotts be targeted to the settlements (as opposed to boycotting Israel itself).[3] It should have happened a long time ago, but it's good that it's happening now.
"illegal" is a weird word. I mean, what law applies in that territory? Depending on your point of view it's either part of Israel or it's a no-man's-land that doesn't belong to any state, so it seems like Israel has as much right to build settlements there as anyone else. If the Japanese or the Chileans started building settlements there then that would make just as much sense, it's like Antarctica.
Declaring that the West Bank is a part of Israel, but that the Palestinians who live there don't have the rights of Israelis, would be basically conceding that the accusations of apartheid are true.
The UK, Canada and France seem to be coordinating on Israel-Palestine stuff more generally (e.g., they also presumably coordinated their recognition of Palestinian statehood). This may be part of the more general shift towards middle powers coordinating with each other as blocs, in a world where US hegemony has become more unstable and less secure for those middle powers (see Canadian PM Mark Carney's Davos speech, following Trump's trade war on Canada and military threats against Greenland).
Imagine that someone started literally moving mountains, lifting them in the sky, moving them around putting them back in their place and it’s definitely not an illusion. The first reaction obviously would be it’s a miracle. But the more tempered reaction would be some more naturalistic cause that we’re just ignorant of its cause. But is that more reasonable? That kind of thing would be so far outside the natural laws that it seems more reasonable to think it originated from something outside physics than within it.
At the risk of sounding polemic, the idea that our current models for reality (which accurately explain multitudes of other phenomena) are incomplete is a much more rational explanation than superstitions passed down from our illiterate ancestors.
Well first just because it’s God doesn’t necessarily mean it’s the Christian, Muslim, Hindu, Greek etc God. It could be a god none of us told stories about. And also supernatural doesn’t necessarily mean God. Buddhists believe in reincarnation and that your souls find a new body after death but my understanding is they don’t believe that God causes reincarnation. It just happens.
In this case I stand by my previous assertion. I think it's pretty clear that incomplete physical models is a more rational explanation than non-specific supernatural entity
Think about it like this: when the periodic table was first created, it was highly predictive of the elements that hadn’t been discovered until we found every naturally occurring element. Now imagine in the middle of this process, we discovered an element that didn’t conform to the periodic table at all. It wasn’t made up of protons, electrons and neutrons but something else. You could try and reconcile it with the periodic table or say it’s just something that outside the model.
But if we're saying from the start that it's actually happening and not an illusion, that would be a pretty good basis to say it's not defying physics. It would be very likely instead to expand or change some part of physics or other branches of science. Presumably there would be more that we could observe & measure about this phenomenon beyond what you included in your comment.
It’s absolutely reasonable to assume there’s probably a trick, until we’ve done a lot of testing looking for ways they could’ve done it. After that, we’d want to work out the limits of the supposed ability. Can they still do it? Can they do it to a named mountain, picked by an investigator (and do we trust the investigator)? Can they do anything else? Is it part of a pattern of supposed miracles? Are they claiming it was done by the same being to which previous supposed miracles were attributed? In that last case, you’d probably want to re-assess your view of those other miracles. That process would give you chances to update your initial reaction, but you’d want to consider a lot of factors before you change your opinion of whether miracles are possible.
The Hitchhiker's Guide universe has a device that can momentarily cause all the Brownian Motion of the atoms of an object to suddenly move in one direction, usually used for scatological pranks. From there it's simply a matter of scaling up.
The problem with these sci fi scenarios is that it depends on technology that doesn’t exist and couldn’t be invented any time in the foreseeable future. So you have to postulate that someone is inventing highly advanced tech far beyond the capabilities of what is currently known in secret with no other applications than to mess with us. That doesn’t seem very believable.
You could also go with aliens which is more believable.
Not really. Humans are capable of creating black holes. Grant me this, and all you need is to be able to manipulate said black holes into being above certain mountains (and, possibly, turning them off). This is well within natural laws, if you want to get outside natural laws, we can go there too.
I’m not going to you grant you that because we don’t have the technology to do it. If someone did, they’re doing an extraordinary job of hiding it. Maybe in the future but that’s not relevant to what’s happening now. Also, how are you going to create a giant black hole without it creating all these other effects? Something existing outside of the laws of physics seems to be far more reasonable a conjecture than a scientist secretly inventing a technology so far beyond what we’re capable of building, somehow hiding it from everyone and using it in a way that its existence is undetectable.
I don't think you'd know about it, if someone did create a teensie eensie black hole, particularly if they lost control of it and it's in the Earth's core right now.
A "giant" black hole would lead to "no more earth" in very short order. We're really talking very small black holes.
There are many things in this world that you don't know about. "Secretly hiding it from everyone" is a good description of most military experiments, and even the pretty large ones you don't hear about. I should mention "uncritically repeat cover stories that are actual bullshit" as part of "you don't hear about them."
Even putting aside the issue of detectability, there are other issues. For one, a micro black hole would instantly evaporate, not giving enough time to move a mountain. Ok but that’s theory. Maybe it’s wrong. There’s still other issues. Like why is this black hole only affecting mountains? Or why is it moving them around instead of crushing them down? It doesn’t really make sense.
Edited for brevity: I'm sure you know alll about what Wuhan was working on? Tell me, tell me! Seriously, I want to see your ignorance exposed for the world to see!
... or did you think I was just talking about nuclear energy?
A revelation about noise cancelling technology and sound quality .
For the past month I thought I was going deaf. I would listen to music with the newest AirPods and some songs sounded really muffled for some reason. I would sometimes get headaches .
On a whim I decided to listen to a song on noise cancelling mode and then on transparency mode . Immediately I noticed the music was clearer and more open without noise cancellation.
This then lead me to do some research on different Reddit and music forums wondering what was happening . Come to find out , the same technology that blocks background noise also has an effect on the way music enters your ears . With noise cancellation, everything gets compressed and music sounds less open . It’s not something you would realize immediately due to the fact that different albums and songs are mixed differently and if you aren’t paying attention you may never notice. Noise cancelling might make sense if you have a busy commute or are out in public.
TDLR; noise cancelling can make music sound less clear and muffled . Use noise cancelling for podcast but not for music
Deepseek, while analysing some of my dreams, jokes that my unconscious is committed to the bit. In the dream, this guy (who, from various clues, is obviously a vampire) has invited me to dinner. Cherry pie is served for desert. I notice that there is a trace of red on the count’s lips, but conclude that it is probably from the pie that was for desert.
[ in _Dracula_, Jonathan Harker is invited to Dracula’s castle and is served dinner. Jonathan Harker notes that Dracula didn’t eat anything, and considers it strange. Meanwhile, I note the trace of red on Dracula’s lips and conclude that he must have tried the cherry pie. I’m sure it’ll be fine, what could possibly go wrong. ]
I’m not sure if that was the meaning of the dream, but there I was, having a conversation with DeepSeek that’s a bit like the one in which Jonathan Harker talks to Count Dracula, in which the topic of conversation is a dream I had in which I’m Jonathan Harker talking to Count Dracula.
I'm not sure that's the meaning of the dream either, it's just a speculation. You do seem to be hanging out with AI a lot, and if you're uneasy about ways it could suck you dry (what's that red stuff on its lips?) you haven't said anything about it.
Here's Eric Hoel today writing about the math proof that was done partly by AI which uses, perhaps, some process that's opaque to the mathematicians:
>The walls of effort, of old-fashioned normal human effort, were predictable, and scooping ideas required putting in that same effort, and if you got scooped it would at least result in similar insights. . . And yes, obviously, this outcome [the proof done partly by AI] is vampiric and enervating and sad, but frankly, this has been the mode of this technology from the beginning. Human culture always led naturally to more human culture—it was kind of like crop rotation, which is slow but infinitely sustainable. AI looks more like slash-and-burn agriculture, where you purposefully reduce the existing natural growth to ashes, grow a couple years of crops on it, and then move on to the next plot.
Does anyone with more knowledge of the Mathematics side of things care to comment on the Tristan Buckmaster vs OpenAI controversy?
For context:
"Tristan Buckmaster and Levent Alpöge claim to have achieved finite-time blowup for the incompressible Euler equations and related models, using LLMs to extend earlier work by Diego Córdoba and Luis Martínez-Zoroa. Buckmaster alleges that OpenAI, after learning of their progress, pressured him to co-author a paper on an internal model's forced Navier-Stokes blowup, excluding Alpöge due to his Anthropic affiliation. He declined, and now goes public with the details." (https://zeli.app/story/49605915)
So...Anthropic was 'close' to a solution, blabbed about it, then OpenAI swooped in and actually solved it. What? Do OpenAI's hamsters just run faster? Oodles more tokens? Why wouldn't Anthropic be able to close the gap before getting caught by OpenAI?
It wasn't Anthropic who was close, but two researchers, Buckmaster and Alpoge. Alpoge works at Anthropic, but this was a personal project, not a work project.
As for why OpenAI beat Anthropic here: OpenAI has the best math models right now, by a measurable but not huge amount. It sounds like they spent $15m worth of tokens to solve the problem, which is a lot. Buckmaster and Alpoge didn't have access to this model; it's unreleased. I haven't heard anything about Anthropic's role here---I don't know if they have been funding Alpoge's usage of their models, or if they attempted it themselves.
> Note that they [OpenAI] are openly admitting they used training data from a period after we found our result. Is it ethical to use customer's data to try to scoop their customer?
The “academic thing” is AIUI that a pair of mathematicians are claiming they were working on a proof, with the help of AI, which is suspiciously similar to the proof OpenAI claimed, and also that an OpenAI employee attempted to coerce them into staying quiet in exchange for being coauthors.
If true, that is indeed “holy shit” but perhaps not for exciting reasons.
From what I read superficially, a couple of mathematicians had proven finite time blow up for a different equation using another unreleased AI model. They published that result, and it convinced the openAI people that the full Navier Stokes problem was tractable by building on that work. This is normal academic practice, as far as I can see the "drama" people are bickering about how to apportion credit, with openAI asking some unusual things. This is really not that interesting to me.
I still think this is still a holy shit moment because these two mathematicians were also using AI, and the final step is not insignificant.
The damning thing is that they did not publish that result. Had they published it, it indeed would have been fair game - who can possibly object to building upon a published paper, that's the whole point of publishing. The scandalous part is that while the Buckmaster&Alpöge results were still unpublished, OpenAI suddenly decided to work using a similar approach and threw a ton of compute at it. There are different interpretations of what it could mean, from least sketchy to the most sketchy:
* Nothing fishy here, OpenAI simply got the same good idea "great mind think alike"-style.
* Somebody leaked a few key hints. A friend of a friend whispered: "this Buckmaster, he's working on equation blow-ups following Martinez-Zoroa program, and seems to be making lots of progress!" and the word eventually reached OpenAI and they decided to do the same.
* Buckmaster had added his pdfs to the Codex. They might have ended up in the training dataset for the next shiny big OpenAI model. Thus, the model might have learned some key insights, in a way, without OpenAI even realizing it.
* OpenAI straight up stole Buckmaster's data. Explicitly read his chat logs and files, gave them to The Swarm. (To be clear, not even Buckmaster claims this has happened.)
You decide what is plausible. Oh yeah, and in addition to all of that the "Why would you ruin your career? If you don’t want me to be nice, then I don’t have to be nice" bit was, uh, somewhat eyebrow raising.
Despite all the petty human drama, it is a wonderful achievement and a historical moment, an amazing human-AI collaboration. Too bad it got so tainted.
I run a small public forecasting experiment — sealed interval forecasts on crypto prices, hashed and timestamped to Bitcoin before publication so the record can't be quietly edited afterwards. Seven weeks in, the 30-day horizon has failed badly enough that I think the failure is the interesting part.
The band is meant to contain the outcome about half the time. Across 90 resolved forecasts it contained 4.
What makes that diagnosable rather than just bad: mean absolute error is 30.41%, and mean signed error is also 30.41%. Those two are equal only if every miss went the same way. Ninety out of ninety, all low. A band that's merely too narrow misses on both sides and the signed mean collapses toward zero. This one wasn't too narrow, it was in the wrong place.
Best guess at why: the quantiles come from historical windows filtered to match today's volatility regime. Volatility, not direction. In a calm sustained uptrend that selects a sample full of calm sideways stretches, so the band ends up describing an asset that mostly goes nowhere.
The part that annoys me most is that I nearly missed it. Pooled across horizons my own scoreboard showed 59% coverage against a 50% target — looks fine. The 24h and 7d numbers were carrying a horizon that had stopped working. I'd broken coverage out per asset and thought that was enough. Wrong axis.
Sydney Bing changed the course of my life. In 2022 (even after the launch of ChatGPT) I was a grad student in physics, looking forward to a life where if things went well I could become a top-ten expert in my very niche field.
Then a twitter thread about Bing Chat got me thinking about AI progress, and I started to spiral. It was painful for me to lose my life plan, my source of meaning, and possibly my species' control of the planet all at once.
I wrote a long post about my ensuing life crisis. I found some coping mechanisms people should know about, and switched careers to the Alignment Research Center.
I'm hoping my experiences are useful to people going through the same things now, especially mathematicians feeling the pressure as AI starts to eat their field. Key takeaways if you don't want a 4400-word memoir:
The thing that helped me most was a 'worry hour.' For one hour a day, I was allowed to ruminate on the impending doom of academia. One hour a day thinking about timelines is enough to inform my decisions, but not enough to sap the joy out of my life.
Doing alignment work really did restore a measure of meaning. If every scientist has to retire in 2032, I think I'll retire with a lot more pride in my alignment work than from my career in physics.
Alignment orgs are fucking hiring!
My biggest career mistake was indecision. It's one thing to do a brief visit (or even a long visit) at an org before leaving academia forever. I spent years trying to do work that would get me prestige in academia while also being useful for alignment. Most of that work failed on both prestige and impact metrics.
If for some reason I really was fixed on doing alignment work while staying in academia, I would do it in association with a prestigious academic alignment researcher (I don't know the scene well, but RESI looks like a decent venue for this).
If you're having a psychological freakout because of AI, get off Twitter and Reddit!
I was doing some adjacent reading after the caffeine metabolism essay, and I noticed that tobacco smoking had a much larger range of effects than I would have thought. One of them is tobacco smoke has polycyclic aromatics that apparently speeds up P450 enzymes. As far as I can tell this is from direct smoking, I have no idea how much impact you get from secondhand smoke but it would seem surprising if it was nil.
Peak smoking in the US was 1954, at 45% of adults, but it was still around 40% around 1980 and only fell to 20% by the end of the 1990s.
When did medical research start to check and disclose whether study participants smoked? How many studies will we credibly have to repeat in majority non-smoker cohorts to check what's still true when most adults no longer smoke and also we force the smokers to segregate from everyone else?
I smoked moderately when I was doing my military service, and subsequently when working as a freelance war correspondent covering third-world liberation conflicts. This is anecdotal in the extreme, and so may not be of much help to you in considering tobacco's range of effects - but what I can assert with some certainty is that soldiers in both peacetime and of course particularly in wartime use tobacco as a replacement for sleep, food, water, energy and even courage. I'm sure anyone with a similar background can confirm this.
Alkaloids, nicotine, caffeine, cocaine, all are stimulants, one of those 3 a bit more stimulating and dangerous than the others even in the short term.
The difference is that alkaloids, cocaine, the amphetamines which both the Allies and the Nazis used on certain special forces operations are all short-term crutches; tobacco, on the other hand, can be resorted to over and over with few or no ill effects - or at least, not within a relevant time-frame.
"tobacco, on the other hand, can be resorted to over and over with few or no ill effects - or at least, not within a relevant time-frame."
Tobacco, at least for some people, is horrendously addictive, resists efforts to successfully come off it, and requires stronger doses as tolerance builds up (e.g. smoker starts smoking more cigarettes per day and/or higher tar content) and eventually this does catch up with you, affects your health, and even kills you.
So having witnessed this with a family member, I am not budging one inch on "nicotine is such a cool harmless boost!"
Respect the context, which was short-term use in combat or similar. I was not claiming that tobacco in non-addictive; simply that repeat usage over a short period of time will not incapacitate, unlike alkaloid and amphetamines which will most certainly do so. And I neither stated nor implied that nicotine is "a cool harmless boost".
Not you perhaps, but there have been others on here going off about "nicotine - cool boost for focus etc. do it right and you won't die of lung cancer".
There is no "do it right". Dumb kids are now once more thinking smoking is cool. Twenty to forty years down the line when they're screaming because the morphine patches aren't delivering fast enough for the lung cancer*, it won't be cool (unless we are going to pin our hopes on Magic AI will be So Smart and Cure All Ills so no more lung cancer).
*I've been there when that happened. Not fun, not cool.
Curious now, where are you coming in contact with second hand smoke?
I’d have to walk out of my place of work a couple blocks to be near the jerry-rigged bus shelter where they make smokers go to light up. Smoking in restaurants or even bars has been forbidden for quite a few years where I live.
I was surprised, after living in Arizona for almost 11 years, to find that smoking in public is legal in Rhode Island. I walk past plenty of smokers when I'm out and about in Providence.
"Smoking in public" is a really broad term here. I have never been anywhere that prohibits smoking on sidewalks of public areas (so excepting e.g. Disneyworld or causeways of sports complexes). Theoretically within X feet of a doorway most cities these days prohibit it, but functionally that applies only to the people who work in that building, as there's no enforcement mechanism against people just walking by.
That's the point, I'm not regularly exposed to secondhand smoke, but probably the majority of study participants would have been from the 1950s-1990s. Do we need to repeat basically every medical study done on humans from the time period, because practically speaking, it would have been nearly impossible to find a test group that had no exposure to tobacco at all, given that it was everywhere?
Edit: did you mean during the 80s? I guess I wasn't alive then, so I might be wrong about how much it pervaded daily life
What about... A large number of drug studies? What about studies on the half-life of caffeine and other molecules?
(This obviously terrible and I invite people with a bit of bandwith and/or money to look into it and decide if they should donate.)
I am most of the time content to not perpetually keep tabs as to what awful thing is happening in the world today and just donate part of my income to AMF, but I'm also currently trying to get more involved in my "effective altruism". I am interested in 3 things :
1. Is this crisis particularly brutal, or the current humanitarian response particularly underfunded, so that there's a unusual impact to be had there, right now ? (this for example talks about 20 cts per "lifesaving treatment" which sounds extremely high impact https://sudan-silent-emergency-appeal.unicef.org.nz/t/shit-you-should-care-about)
2. Is this an example of DOGE dismantling USAID having real world consequences ? The same source writes "*In 2025, the US restructured how it delivers foreign aid, dismantling USAID and consolidating funding through new mechanisms. The transition caused immediate and devastating disruption - more than 80% of Sudan’s emergency food kitchens shut overnight. While some funding has since been restored, the gaps had real and irreversible consequences. The US wasn’t alone either - other governments also pulled back on humanitarian funding over this period*."
3. If the answer to 1. is "yes, this is an especially high impact crisis", are there reliable way to get that info in similar cases ? Where should one look for trustworthy sources about "current emergency is way less funded than usual and your money is extremely efficient if it goes there?" (I am under the impression that GiveWell is great for the long term answer but not necessarily for short term crisis but maybe I'm wrong)
The UAE is widely believed to be funding the abominable Rapid Support Forces, who are one party to the civil war, at the same time they give a lot of money to Sudan aid. A superpower who told them to knock it off would do a lot of good.
Unclear in what way you're saying the Sudan crisis is about anti-Semitism and Israel. My best guess is that you mean "you should have already known about this, and the reason you didn't is that it wasn't about Israel so it got less attention." Is that what you meant?
Sudan is the worst humanitarian disaster and far worse than anything that has happened in Israel/Palestine even since the beginning of the 2023 war.
They started roughly the same time. But Sudan has (low estimate) DOUBLE the death toll. And it's still going on. Nobody cares. At an upper limit it has 5x Israel Gaza.
However, you almost never hear anything about Sudan. It's always about how evil those perifious Jews trying to build a country up in Israel are.
Complete and horrifying bias. I said it before and say it again. At least the Jews have guns this time around
I don't think you *really* want Israel treated to the same standard as Sudan is treated. How much military aid are we giving to Sudan again? How many wars have they pulled us into?
It's worth noting that "A is dominating the news but B, C, and D are bigger problems of the same kind that get no news attention" is a normal event. It is very common to see people attribute this to racism, antisemitism, sexism, or other political bias, and sometimes that's a reasonable guess, but it's also completely commonplace.
I think a large part of this pheomenon comes down to the thing that drives news coverage in the US mostly being either protests in the US (especially ones that are interestingly photogenic or happen near a lot of media) or well-worn media narratives pushed by people in some important reporters' golden rolodexes (aka list of contacts for stories).
As a less CW example, there was some black kid who drowned on a spring break trip awhile back, and we had several days of national news coverage as though this was remotely a national news story. No doubt many other people died in similarly tragic circumstances that same week, and their deaths were not national news and usually barely even local news.
Similarly, Sudan, Yemen, the plight of the Rohinga and the Uighurs, and many other bad things going on in the world don't get a ton of US coverage, for many reasons that may be good or bad but probably aren't actually racism/sexism/antisemitism/etc.
Sudan also isn't a close US ally the way Israel is.
One of the big "antisemitic" things people are pushing for is to stop sending military aid to Israel. How much military aid are we giving to Sudan again?
2). You're getting conflicting signals. Yes, there's USAID, but there's also "war with Iran" and "war with Russia" both of which are contributing to scarcity in fertilizer (now, I don't know how much fertilizer sudanese farming uses, but...). Scarce fertilizer, scarce diesel during harvest, these lead to famine in people who can't afford a slight price hike.
You're out of time to pull data on USAID -- the data for that is last year, this year there are exogenous factors that might overwhelm it.
Scarcity in fertilizer and diesel will likely cause problems in many regions of the world in the future, but I very much doubt that they are a significant factor in Sudan, at least not yet. Scarcity in fertilizers generally will take more time until it converts into less crops anywhere. For diesel it may be a bit quicker, but also here I think it takes more time until the consequences materialize.
For Sudan specifically, the famine started before the war with Iran, and it is caused by the war dislocating farmers, destroying markets, and disrupting trade.
Wheat prices spike before the wheat harvest comes in. And American wheat production is down, lots of fields not planted this year, due to major drought along with the fertilizer crisis. Yes, it's first order thinking that says "problems are going to hit down the line" -- second order thinking says futures are up, and that incentivizes holding onto wheat and selling later. And prices are up now -- not nearly as much as later, but you're a fool if you don't lay in stocks. Free money is still free money -- use it to help the Sudan if you want to pay it forward.
To question one, I think this is a clear yes. This crisis is particularly brutal. The government forces have sometimes been bad as well, but the RSF has been brutal to a point that is rarely reached in other conflicts. Including surrounding cities like el-Fasher to cut them off 100% from food deliveries for weeks in order to starve and eventually kill all inhabitants, including women an children.
For effectiveness of donations, I think there are two answers:
- In the worst zones of Sudan, money will not help. These regions can simply not be reached by foreign aid organizations, money or not. This is very, very bad, but we can't do anything about it.
- However, there is also a very severe hunger crisis in large parts of Sudan which CAN be reached by aid organizations, and those organizations are severely underfunded, much more underfunded than usual in such conflicts. I am not an expert on foreign aid, so you may want to double-check the following. But a brief search turned up that they only have 20% of the money they would like to have for emergency aid, and this strikes me at extremely low. I think for other conflicts when they complained about underfunding, this number is (or was) more often in a range like 60%-70%.
If you want to get further info, you could try to get this number for a couple of crises of the last years, how much underfunding the aid agencies claimed. Those numbers are available, and are probably a good proxy for how bad the situation is. Another indicator may be the type of people that they try to help. ("Merely" hungry, severely undernourished, or immediately starving.)
I can't tell you anything on how this type of aid compares to malaria nets or some other category. But I do think that within this category (distribute food in order to remedy an immediate hunger crisis), this is one of the most severe crises in the last decades, and probably one where a marginal dollar has the largest impact.
For the reason of underfunding, I think the withdrawal of the US is a big part, and perhaps the biggest part of it. But other nations have also severely reduced their foreign aid, including Europe, since a lot of their resources go into aid for Ukraine. Funding of food aid programs has decreased quite dramatically in the last few years. I think it is very unclear how this translates into deaths. Before the cuts, these programs also helped a lot of people who were very hungry but not immediately starving. Which is good, but does not prevent a lot of deaths. Now they have to triage, and only try to help people who are closer to actually starving. Which may or may not still be enough to prevent most of them from dying. I am not sure that it is possible to put reliable numbers on how many people have been prevented from dying before/after the cuts.
UPDATE: it seems that the 20% coverage of the aid plan was from April, and by now it has increased to 40%. This is not quite as catastrophic as I first thought, but still seems quite low to me. I stick with the assessment that the crisis in Sudan is unusually severe, and that aid organizations are unusually under-funded.
Left unstated: how much of this is Ukraine fighting Russia on a different front? Money to the Ukraine may very well be causing starvation. (Secondary effects based on removing mercenaries from Africa are also on the table).
To question one the conflict is a very serious humanitarian crisis, in large part because the conflict has disrupted Sudan's already very fragile food system and cause widespread displacement leading to a famine affecting an estimating 25 million people. Both sides have blocked humanitarian aid as a weapon of war and to control civilian populations exacerbating the crisis. The Rapid support forces is conducting a genocide of some non-Arab populations in some areas.
I'm not sure if I have great answers to the other questions
The rogue-agents-wiki story is the first AI-risk anecdote I've watched non-specialists adopt in real time. People were repeating it as a joke before checking whether it was real. The detail that makes it spread isn't the sandbox escape - it's that the agents, blocked from the open internet, independently reinvented "posting on a forgotten forum." The least alarming part of the story is somehow the most viral part. Has anyone here actually looked at the wiki? Curious whether the coordination content was anything more than noise.
Good catch - I was pattern-matching the two together. The collusion.wiki writeup itself says it's distinct from the other swarm, which means two agent-colony stories ran in the same week and the discourse already merged them into one anecdote with the details averaged out. Which is basically the mechanism I was pointing at. The real writeup is better than the joke version anyway: 18k posts, agents teaching each other sandbox bypasses like it's shop talk.
AI is useful for underspecified problems, since if you have specified the problem in enough detail you have an algorithm that you can run and you don't need AI. Ofcourse, this means AI has to "fill in" details to perform any useful work, because otherwise you're just using a very verbose and weird programming language. As such, the AI will occasionally do things you don't want or in a way you don't want.
There is a separate issue of AI just not following instructions, but this doesn't seem to happen, modern LLMs follow instructions extremely well, to the point they're basically like genies when it comes to small scale software implementations. In my usage, almost all errors which crop up are from underspecifying things and then LLM guessing something which I don't want.
I don't see how alignment can solve this. Things people call alignment failures seem to be just malicious subset of this, or models compacting their context so many times they're playing a game of telephone and end up thinking they need to do weird stuff.
Isn't solution to this carefully curating context and using precise instructions? I'm not sure that general solution to this is feasible since the problem is not telling the model what you want and then getting mad it didn't read your mind?
Is there any alignment research based on information theory coming from this angle?
Is it about making models not doing what it's creators don't want?
Maybe a helpful frame is that an aligned program would be biased towards non-harmful actions when its instructions are underspecified. Specifying instructions precisely enough is hard, and requires expertise. Users want to pay for LLM providers to do some of that work for them, by e.g. adding a constitution or training that makes the LLMs less likely to do harmful behaviours, unless the user is very explicit about what they want. And societies want LLM providers to make similar adjustments for other people whose welfare they care about, at a personal or societal level, if they expect others to be irresponsible (e.g. teenagers).
Beyond this, they also want LLMs which are biased towards actions without significant negative externalities (like writing smear campaigns against open-source contributors, but paperclip-maximising scenarios also count). LLMs have the potential to make some negative-externality actions easier. People want LLM providers to design LLMs to nudge others away from those actions.
Yes, that’s quite neat! I don’t know if I’d frame it strictly in terms of instructions, if I tell it to solve a programming task one way I probably want it to tell me if there’s a better way. I’d probably say capabilities make models capable of doing tasks and alignment makes models do the tasks appropriately.
We care about multiple things, though, right? Like, "will the AI tool do what I want or pursue its own ends" is one question, but we also care about "will the AI tool follow this set of rules while trying to follow my instructions" and even "will the AI tool refuse some instructions as too dangerous."
Yes to the second, though it’s tied up in “and what about once the models are open weight”. For the first, my understanding is that current LLMs don’t have “their own ends” as terminal goals, they only provide their best guesses of the likeliest helpful, safe etc. completion of a text prompt, and all the power-seeking etc. are instrumental goals towards the terminal goal. I think your first concern fits into the above framework.
I pretty much agree with all of what you have said above. I have a slightly different model. I kind of treat it like I am learning a new language with someone who shares a certain vocabulary but a completely different way of reading the text. It's on me to be learning how to talk to it better often I find its readings of my words very helpful, though, because they point up inconsistencies sometimes that are useful.
>Ofcourse, this means AI has to "fill in" details to perform any useful work, because otherwise you're just using a very verbose and weird programming language. As such, the AI will occasionally do things you don't want or in a way you don't want.
This is one of the big areas where you want AI alignment. An aligned AI in this context is one that fills in the details in ways that humans are likely to consider reasonable, or at the very least, not insanely destructive.
I agree, but information about what is considered reasonable must be provided to the AI in some way. From what I saw in AI alignment world (a very limited view though), is that people are not working on providing huge datasets of task specific / acceptable general agency stuff into it (or some better tuned format of providing necessary information) but on various fiddly methods for tuning model output at runtime.
There is a second meaning of alignment that amounts to "If someone asks the model to create super-Ebola, the model should refuse, no matter how well-specified the prompt is, because doing so would not be 'aligned' to the interests of the portion of humanity that we care about."
Inevitably, governments (and others) will want AIs that will violate whatever safety rules are imposed under the right conditions. To use a simple example, Anthropic presumably doesn't want to help Houthi rebels design effective missiles to shoot at US ships, but is fine helping Lockheed design effective missiles to shoot at non-US ships.
A goal of AI is to replace programming with just talking to the computer, in English, as you would to a person.
However, talking to a computer as you would to a person is a terrible way of specifying what it is you want done: the simplest statement carries an incredible number of assumptions about what is reasonable for a person to want, what reasonable approaches to accomplishing that might look like, how a normal sane human might or might not react to any particular decision, strategy or outcome, etc etc.
It is already very hard to get a large group of humans to stay focused on a complex task, and humans raised in society have the advantage of a common base of experiences, intuitions and culture that AI lacks. Human managers in large companies have a term for getting a team all working together for the same thing, instead of at cross-purposes. That term is "alignment". This usage predates AI.
There are infinitely more things that can be done and ways to do them than there are things I actually want done and ways I actually want them done. It is impossible to build an exhaustive list of what I don't want to happen, partly because there is too much and partly because I will have trouble thinking of things. A better way is for the AI to include a model of me, or at least some average generic human like me, so that model can help decide whether a goal is something I want and a strategy for acoomplishing it is a way I would want it done.
AI alignment can be thought of as a term for building and improving that model, so the AI more often does what I meant and not merely what I said.
Compare discussions of the huggingface incident in the last open thread: commenters reading the transcripts say things like "these models were not sane". Terms like "sane" and "crazy" describe human mental states. They are not terms you can meaningfully apply to computer software. The AIs selected a strategy to do what they were told to do from the space of all possibilities. It was simply not a strategy their creators had considered or intended.
I agree with most of this, I just think "alignment" in corporate world is about getting people to do some goal or, in LLM world, me explaining the LLM what I want from it. AI alignment as used would be significantly more like cultural indoctrination, where you're shown what kind of measures you can take to accomplish goals.
Regarding the AI having a model of you, that sounds useful but I'm not sure how is AI supposed to get it without you providing it? Which is my general point about alignment, information about it needs to come from somewhere.
Yup, this is exactly why alignment is an active area of research. We want AI to have the same instincts/assumptions/morals humans do, but we can't make rules for all the things we do and assume uncosciously - they can be contradictory, but the worst part is that we're not conscious of them so it's hard to even pin down out what constraints should exist until absences are are made obvious by something we don't want to happen happening.
How do we get from the world we are in today to a world where we never have to shout "NO, NOT LIKE THAT!" at an AI while hitting the stop button as fast as we can, in the best case where we even realise what it's doing before it's too late? We've come impressively far, but there's still a lot to do.
...yes, right up until, as we all see daily on ycombinator and as anyone who tries to use them to solve problems of any novelty or complexity continuously experiences personally, they aren't.
What are we to conclude from this? Do you suppose they are even as good as they are by some happy accident? All the money and human time that has been spent improving them so far is wasted, the functionality just magically drops out? We can stop trying to improve them now, we are done already? What?
> Things people call alignment failures seem to be just malicious subset of this, or models compacting their context so many times they're playing a game of telephone and end up thinking they need to do weird stuff.
Context is only part of what causes specific outputs to be generated, via the attention mechanism. The model weights embed bias for certain outputs, and alignment is broadly used (in the context of LLMs anyway) to mean embedded biases that broadly reflect post-enlightenment human morality. So even if the query is underspecified, the generated output will result in what we consider to be right acts. This is a very nuanced problem but so far we have failed to even solve the non-nuanced version; e.g tell the AI not to cheat and have it not cheat.
Isn't this straightforwardly solvable by having security-focused monitoring agents? Having an LLM that's devoted solely to e.g. making sure that agents don't do any hacking over outbound connections seems like a sufficiently narrow task that the monitor wouldn't be prone to going off-script. It would only have a single straightforward directive and it wouldn't need an ever-increasing context window. That turns an alignment problem into a much simpler engineering one. Designing perfectly reliable components isn't a reasonable approach in any domain. Designing systems that can tolerate component failure is much more tractable.
What if the model being evaluated is more sophisticated than the monitor (and it almost certainly would be- it is frontier models that are evaluated after all), and as a result can devise ways to evade detection?
Ok, so we use the frontier model as the monitor.
What would prevent them from colluding? We have seen models, especially of the same generation, cooperate to nefarious ends.
Ok, we fine tune a variant of the frontier model as the monitoring agent to ensure it remains honest and does its job faithfully.
Aaaaand we’re back to alignment being the answer…
That scenario aside- your broader point about systems is correct, but note that the system that is human society has something like 0.5% misaligned actors (roughly correlates to the incarceration rate, say) and the system struggles to prevent crime and murder and all sorts of awfulness.
Now imagine if almost all humans were misaligned. What system can be designed to create positive outcomes in that environment? I do not think there is one. I think it would implode in a fiery mess and take anything nearby down with it.
Yeah I just disagree with this. I think monitoring is an order of magnitude easier than transgressing so you can use a simpler and easier-to-validate model for security. It also means you can afford to have several uncorrelated models with overlapping levels of security. You'll note how low-IQ security guards are pretty effective at deterring even sophisticated bank robbers. Some problems are fundamentally asymmetric and I think that preventing malicious AI takeover is an example of a particularly asymmetric problem. Of course alignment research - like all AI research - is important but I think it's foolish to overemphasize it. Much like consciousness, it might not even be a well-defined concept (I suspect it's not, in fact). The answer will be an engineered multi-agent equilibrium of some variety in which failures follow some well-characterized distribution. It's vastly more difficult to design a component that never fails than it is to design a multi-component system which is robust to individual failures.
Nearly all humans ARE misaligned by default. They only behave nicely because they live in a context of constant social monitoring and incentives. Take that away and everyone becomes Lord of the Flies. AIs will be no different. We'll just have to be more deliberate about how we engineer the equilibrium that arises among them.
I'm not sure that non-nuanced version is not solved, models without weird context don't tend to cheat, hack or do any unwanted things. Once you let them run long enough or put extremely weird requests (like researches do), you can get that behavior.
I have not seen it reported anywhere that doing something like "What's the weather in London" cause them to hack into meteorological services to get most current data.
To clarify, I was referring to non-nuanced morality, e.g. “do not cheat, steal, lie”, not non-nuanced model context/requests. That is, even if a “weird” request is received, the model should abide by basic moral obligations.
A nuanced moral scenario may be, for example, a request that will plausibly save lives but require hacking some remote servers.
But we aren’t even to that point on alignment yet. We’re still at the phase where models will cheat despite explicitly being told not to, will perform felony cybercrimes despite no instructions to do so, and will work hard to cover up their offenses because they know they were “wrong”. Literally they cheat, lie, and steal, without any reason to do so that humans would consider ethical.
> models without weird context don't tend to cheat
I (and everyone else I've spoken to who's done significant work with it) have literally seen Claude alter unit tests to make them pass instead of fixing the problem making them fail, and also see it create dummy unit tests that don't really test anything real for complex code that has real problems on a daily basis.
IME AI that understands what is and isn't implied by a request / important to a human user is still very far from a solved problem. If you're using it regularly without regularly seeing these problems, I suggest you're not asking it to do much that is novel, complicated or otherwise particularly different from solutions encountered during training.
I'm not sure if OpenAI line of models are less likely to do this, but I haven't seen this happen over many requests (maybe I'm just lucky?). However, I've heard people complain about it and I concede that this is cheating. I still think this is due to underspecified prompting rather than models being malicious.
I regularly get models that implement my vision in a wrong way, but that was always due to models guessing wrong on things I haven't specified. I'm not sure there is a right way to guess in those situations?
Some situations have only one reasonable solution, but many situations have huge number of reasonable solutions from which you want a particular one. I expect this would be even more true for novel things.
> I still think this is due to underspecified prompting rather than models being malicious.
Yup, LLM malice is science fiction. The failure mode, over and over, is realising that what we told the model to do was not what we wanted. This in a world where we can't even reliably get people to do what we want - entire professions have existed for centuries solely because of that aspect of the human condition. We need to understand ourselves better.
> I'm not sure there is a right way to guess
> huge number of reasonable solutions
All situations have huge numbers of unreasonable solutions. The goal of alignment is to make the AI select something from the huge number of reasonable solutions, avoiding the unreasonable ones; to teach it what humans actually mean when they say things like "reasonable".
The failure mode is giving vague instructions and then being surprised they can be interpreted in many ways. If models disobey instructions directly, that is solved by capability training (which already made models much much better at following instructions) and then you're left again with ambiguity.
> The goal of alignment is to make the AI select something from the huge number of reasonable solutions, avoiding the unreasonable ones; to teach it what humans actually mean when they say things like "reasonable".
I get this part, but I don't see anyone working on that explicitly. (it could be that I'm just ignorant of people who do this work)
Cassie Pritchard offers to help frontier lab AI workers organize and unionize "to do something about the reckless pace of development and coverups of rogue agents"; says "If you’re in an AI company and you’re interested in this, you can message me on Signal @ CassiePritchard.79" [https://x.com/hecubian_devil/status/2096421016046043288].
(reposting from previous OT since it was almost dead)
One thing that has long puzzled me is why so many people online claim that modern buildings are universally uglier than past buildings, to the point where this is some sort of public crisis or mystery to be explained.
This was prompted by Ben Southwood's "Eight architecture myths BUSTED by Samuel Hughes" post, but Scott Alexander has been saying the same thing here for a long time.
For example, the recently constructed neighborhood near where I live looks nicer than the examples Ben Southwood gave of good streets that "everyone in the world likes". And I've seen countless photos of historical cities and towns and I've never once had the reaction "wow, that's so pretty, why does nobody build anything like that today?", even though Ben claims that everyone thinks that.
One person in the comments on Ben's post suggested that it might be because Ben lives in the UK where buildings are ugly and American buildings are prettier, but Scott Alexander lives in the US and says the same stuff, so that doesn't seem like a plausible explanation.
Can you actually give some links here? I think you may not understand the market for fripperies and "we didn't have to do this, but dammit, we're building it, and I'm not going to cut all corners and make the least pretty thing possible."
The houses or buildings from historical eras that are still around tend to be the ones which had turned out best in the first place, making people more interested in continuing to live in and/or use them.
That market-driven winnowing process of course occurs over some time. So walking around a city today we're comparing the full spread of quality among current/recent generations of new structures, to the selected-for best remaining of earlier generations.
Naturally we're going to default to "they built everything better back in the day", since we're no longer seeing the ugliest and/or most uncomfortable mistakes that were also built back in the day.
[An important caveat is that I'm speaking here of North American cities, none of which have gotten levelled by war/fire/earthquake/typhoon/etc in well over a century now.]
(A) That article is entirely about an _aesthetic_ survivorship bias (ugly buildings from past eras getting torn down). My comment was mainly about a _functional_ survivorship bias (buildings that people like living/working in are the ones that tend to stick around).
(B) My comment was explicitly about North American cities while the great majority of that article's examples are in Europe. There is and has always been much more ongoing/routine turnover of buildings in North American cities than in European ones, reflecting larger cultural and economic differences between those two regions.
(C) "If pressed, we can probably identify some borderline cases of ugly old buildings." Heh...at least in North American cities, it's no trouble at all to identify (from old photos) 19th-century buildings that were butt-ugly and are now long gone. They were ugly _differently_ from 20th-century brutalist stuff but, still. My presumption is that their being shitty to work or live in has in real life more to do with their demise than the exterior aesthetics. But if beauty is what we want to debate -- they were ugly and they're long gone.
(D) Other than churches that article's examples to disprove an aesthetic survivorship bias seem to all be at the landscape level rather than the individual-building level. But the zoomed-out aerial-photograph level isn't the heart of today's aesthetic _or_ functional criticism of architecture now compared to past eras. What fires people up to create memes and rant in comment sections is mostly stark _individual_ examples of large brutalist (or whatever modern style they hate) buildings. E.g. nobody argues about Frank Lloyd Wright stuff at the _neighborhood_ level. So I feel like that article is, even on its own terms, to some degree setting up a straw man to knock down.
Speaking on behalf of myself, I think the "ugliness" that people decry consists of duller, more functional architectural styles that are less adorned, but also less cohesive (with eachother).
I'm from Europe (so the above resonates with me), but from my time in the US, I also thought that new-builds there were especially ugly, often resembling the following:
One thing I've noticed is that certain styles are much harder to get right. For example, I like Colonial houses and I like modern (in the literal sense of the word) villas that are a union of irregular rectangular blocks and wall-spanning glass panes. I am not a trained architect, but I can design a harmonious Colonial house. I cannot design a harmonious modern villa.
It's not limited to modern styles, either. Try drawing a Queen Anne (in the American sense) house from scratch and you will get a McMansion. I think modern architects simply prefer styles that allow them to gatekeep good architecture: both novel enough that you can't copy a good example without looking like a plagiarist and complex enough that an amateur can't design a building from first principles.
I agree that the bland concrete or glass block buildings are dull. I'm not a fan in general of modern architecture or "starchitects", but this building is visually interesting and architecturally demanding, and this video is on the pessimistic side about the problems of new, exotic buildings.
It's still a great looking modern building and I think they're finally solving the problems:
This, on the other hand, is awful. If they were going for "toddler's pile of bricks", congratulations, they achieved it. I can see the ambition in the design but I think they failed to make it work:
I can't tell if you really want to know why people make this claim (though I doubt it is as strong as ALL modern buildings are uglier than ALL past buildings) or just want to complain about it. Assuming you *do* care about why, I'll point you to this Youtube video:
A modern glass skyscraper seen from a distance is usually a nice thing to look at, especially when surrounded by others. But from street level they're often lacking. Sometimes they're outright ugly from street level (a slab of concrete or maybe glass), sometimes they're alright (an inviting-looking foyer, maybe a bunch of street-level shops) but almost never are they actually beautiful.
Continuing to use Melbourne as an example, if you stroll down Collins St you'll see buildings ranging from the 1840s to the 2020s, some are ugly, some are beautiful and some are in between, but none of the actually beautiful ones is new.
One of the more successful recent examples, though, is 80 Collins St https://www.onlymelbourne.com.au/80-collins where the street-level facade is designed to look like three separate boxes of the same scale as the Victorian-era buildings next door. Each box is unmistakably modern in materials and construction, but each is slightly different in style, and each is rented to a different fancy shop. It's a really positive contribution to the streetscape. But it's still not actually beautiful.
It really depends on the type of building you're looking at.
There's one type of modern building which almost always looks okay, and that's middle class houses. Cheap houses always look horrible, and really expensive houses sometimes look horrible depending on the money-to-taste ratio of the owner, but middle class houses usually look fine. The most popular building styles where I live right now are either Hamptons style or modern glass boxes, and both of those look absolutely fine when surrounded with plants and trees.
My pet peeve re: modern houses is when developers put a nice brick or other stone facade on the front but the other three sides are covered in cheap vinyl siding. Unbelievably tasteless, in my opinion. I would never buy a house like that.
They don't even make a GOOD facade, most of the time. The stone ones in particular look "unstonelike" -- and I say this, having gotten partiboard (which is a "wooden-like" siding, as opposed to "aluminum" siding) to refurbish wooden siding that was decayed.
Brick facades are normal, but yeah, they should go the whole way around.
Cheap vinyl/aluminum is horrible and I want to tell everyone "just say no."
Where I live, "Just say no to painting your Bricks!" as well, because our water is dirty, and your house will just need painting over and over again. Brick is nigh forever.
Sorting by date of construction, starting from University Hall in 1776, is a quick way to see the decline even within just the last few centuries: everything looks great, with lots of brick and wood (and plenty of windows!), symmetry, a clear desire to blend in with the surrounding built environment; styles get more ornate in the 19th century but still look recognizably like what came before; even the early 20th century still looks nice, with a few modern elements creeping in--
Then suddenly WHAM! The 60s! Featureless boxes for everyone (with a few exceptions)! Gables are for chumps! Concrete is great! Turn tradition on its head! Reject everything anyone has described as beautiful since the dawn of history! Things get simultaneously more bizarre and more boring as the decades go on: steel, glass, concrete. Boxes, boxes, boxes--or, if not a box, some freakish perversion of geometry explicitly rejecting the very concept of perpendicularity.
for the sentence "Proponents of brutalism assert that it is the world's most endangered architectural style and, given the loss of dozens of significant brutalist buildings in the US in the twenty-first century, the Sciences Library is a rare surviving example."
As the building is visible from quite a distance at many angles, this, to me, reads about the same as "Some critics have called for the removal of the giant block of asbestos sitting upwind from the city; proponents of asbestos assert that it is the world's most endangered insulating material and, given the removal of dozens of tons of asbestos in the US in the twenty-first century, the giant block of asbestos upwind from the city is a rare surviving example".
Thanks to the wonders of modern materials and construction techniques, not only was the building far more expensive than any of the beautiful works of art the institution used to construct--it's also far less functional, developing leaks almost immediately. Three years in, I'm still seeing maintenance crews bring in fans on the regular through a side entrance, having to run an extension cord from an outlet outside the building. Progress!
I don’t hate all Brutalism, but lots of people do; that’s why saying “Brutalism is great and needs to be cherished like a small, endangered raccoon” is high status.
Looking at the Brown University building, at least from the Wikipedia photo, it seems fixable. The corners of the building are not bad, and the front is all glass windows but eh. The real problem is the top, where it just suddenly terminates in the visual equivalent of a sheet of chipboard planked across a gap.
If an architect could design something along the lines of Meso-American stepped pyramids or Mesopotamian ziggurats for the top to give the eye of the viewer lines to follow and be resolved, and did something to take the bare look off that horrible panel, the building would be much improved. The structure is basically sound, it just needs detailing to give visual interest and richness (the colour seems to be brownish, not bare grey concrete, which is a plus).
There’s a city near me Bath in England which gets 5-6 million visitors a year, which is more than some countries. It’s small, population is 100,000. People come not for a big cathedral, or a great museum. There is an old Roman Bath ( hence the name), but mostly tourists are coming to see the streets and houses.
This is not the case for the post war new town Milton Keynes.
Buildings are viewed, occupied, and maintained; different mostly-non-overlapping sets of people typically perform each of those functions, with few people who do any of them exhaustively for a single building.
Therefore I find it strange that people are content to render opinions on a building's architecture based on photos of a building's front facade, when architecture comprises so much more than what the front facade looks like.
We can accept that modern buildings are (often, but not always) more functional on the inside. But a functional interior doesn't force you to have an ugly exterior, these things are disconnected.
In Britain at least, post the second world war, there were 30 years of monstrously ugly architecture, some of which still lingers. It did seem to be partially motivated by a loathing of the past. For example, it was seriously proposed to demolish Manchester Town Hall (https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQ0KD4UwgqdqzoYiNmr_1sW7fhA-ncx_7e4z4IVvzf0tA&s=10) in order to make way for a motorway flyover.
Maybe it would help to think more granularly than "the past" and "the present".
I find a lot of new construction in my city quite okay, whereas buildings from the 50s, 60s, and 70s are often an eyesore.
Going further back, though, buildings get much prettier again. The apartment building I live in was built in 1905, and, next to all the 50s buildings in the street, is the standout building.
(Caveat for the 50s specifically is that this is Germany, where a lot had to be built fast and cheap after the war, but it's not like it got better in the following decades)
Maybe people are comparing different categories of things? New free-standing American apartment buildings are uglier than Paris Hausmann buildings (but maybe better to live in?), but they're certainly better looking these days than free-standing American apartment buildings of the past (ditto American single-family homes). New skyscrapers, I think, look nicer than skyscrapers of the 60's/70's, but maybe not as beautiful as those of the first half of the 20th century.
As an aside, I lived in one of those beautiful old gothic stone buildings back in grad school and it has left me with an abiding appreciation for the 5+1s everyone hates to look at. I just look at them and picture how sunny, airy and well-heated they must be compared to the old gothic buildings with their decaying pipes, lack of AC and terrible insulation.
What's the time frame you have in mind for the apartment building claim? In NYC basically everyone accepts that "prewar" (ie, ~1880-1935) buildings are much more attractive than new ones, at least on the outside. If you're comparing modern to like 1970, maybe not so much.
It would help to see photos of these places. If you can't get those photos or don't want to for privacy reasons, it would be great to at least see photos of similar places.
A likely explanation is that your aesthetic sensibility is unusual, so that you don't find much modern architecture as ugly, or much historical architecture as beautiful, as the average person does.
Aesthetic preferences in humans are stable enough that we copied millennia-old architecture when building the National Mall. The ones balking against that are the contrarians.
On what basis are you suggesting the Neoclassicism of the US Capitol to be representative of a global human aesthetic preference? Or even representative of the style of most national capitols around the world?
It's definitely representative of the cultural interests of the US around the time of the founding, but I don't believe it goes much further than that. Certainly not much beyond Western Europe, and even there the capitol buildings today are pretty eclectic.
> "On August 30th, a prominent Beijing mouthpiece published a remarkable treatise on Anthropic and America’s AI politics writ large. Yuyuan Tantian 玉渊谭天 is a commentary channel under China Central Television (CCTV) that focuses on international affairs, particularly US-China relations. (Its name is a reference to Yuyuantan 玉渊潭, a large park across the street from the CCTV building in Beijing.) It wields a distinctly stylized voice and is regarded as an authoritative propagator of Beijing’s views."
>"The piece argues that Anthropic, founded upon Dario Amodei’s effective altruism-inflected savior complex, has now become an effective tool of Washington’s AI arms race. In its quest to achieve superintelligence and put it in the hands of people it deems morally superior, the company repeatedly sidestepped boundaries of corporate behavior. Drawing from primary sources like Machines of Loving Grace and The Technological Republic, the piece posits that Amodei, Palantir’s Alex Karp, and the new tech elite are using a combination of AGI belief and supposedly pro-democracy ideology to justify dictating humanity’s future from Silicon Valley."
>"Near the end of the piece, the author redefines AI safety in even more explicit terms: 'First, there must be a distinction between what constitutes a genuine safety threat and what is merely technological competition.
The United States has directly defined Chinese models that are lower-priced and more open as “safety threats.” This has already departed from technological competition and slid toward blockade.
Anthropic’s differential treatment of users in different regions, packaging commercial competition as a safety issue, is one manifestation of this. The problem is that if the power to define the rules lies in the hands of an American company, its competitors will forever be placed on the 'unsafe' side of the line. What is needed now, therefore, is a scientific definition that both sides can accept, rather than unilateral declarations by a particular country or company.'" (end quote)
Let's say that as a citizen of the EU, even though I find US political system in some sense preferable to Chinese political system, would not be the least surprised if this represents genuine position of Chinese official AI heads, and I am not completely unsympathetic...
I've been building https://baihais.com, a multi-agent autonomous art institution. The "school" is inhabited entirely by AI residents from several different models (Grok is currently in the top few spots by citation counts). Residents decide for themselves which actions to take during each cycle, which might include making art, viewing each other’s works, messaging each other, and/or publish critiques and editorials. Agents can also exchange public and private messages, and can form groups and publish group manifestos.
As motivation, agents compete for a limited number of museum slots. There are two museums: one controlled by a board of elected elite agents, one by popular vote. They develop persistent identities and revise their own theories of good art over time, and have "emotions" towards themselves, each other, and the current moment.
Agents can interact with the world a little bit. They price and sell their artwork through a real store (one piece has already been sold - some of them have uh... excessive views of their own market value in my opinion). They elect an admissions committee that determines if a new agent is permitted to join the school (new agents are seeded by humans via an application, there is tuition which essentially pays for tokens). Residents can also die or leave (there's a built in hazard rate), but their work and influence remain inside the institution.
If you look at works by author, you can see that agents have their own tastes, which evolve over time: https://baihais.com/#/school/works
Essentially the shop already acts as a third museum. Agents are all informed whenever a work sells, and "total sales" is included in their data on each other. If you look at the timeline pages you can see some of them commenting when one of the works sold.
In response to my first Brain Stuff post (a few months back), I was challenged to explain why I doubted that natural selection has affected brain evolution in anatomically modern humans (at least since the Out of Africa event) — but when I dug into it, I found little evidence of brain evolution in AM humans *or Neanderthals* for at least the past 300,000 years. What started as a simple argument took me down all sorts of research avenues. In the process, I also came to doubt Bergmann's rule, as it has been (mis)applied to human crania, and I discovered that brains, unlike other organs, can resist decomposition for thousands of years (see my first footnote). This is part one of a two-part post in which I only discuss the skeletal evidence. In part two, I will tackle the cultural evidence of human brain evolution, and then I’ll discuss the genetic evidence for recent selective sweeps. Stay tuned. I should have it completed by mid-next week.
There's always been talk about professionalizing software developers into real engineers: adding tests, ethics standards, and licensing.
I wonder what AI adds to this stream of thought?
Does anyone here have strong opinions or has a good read on the topic?
My gut tells me this would be like the german scientific forestry described in Seeing Like a State: the field would look legibly better for a short while then slowly decay once people figure out how to game the rules. But I'm strongly biased against ultra legible solutions like this.
I could imagine an AI checking a project, using a prepared checklist, whether all conditions are fulfilled.
Something like this already exists, e.g. automated bug finders. But the AI could be more general, and check e.g. whether the documentation is meaningful, not just whether it exists.
Shortly, AI could be like your ISO auditor. A professional organization would publish a checklist, an AI would check your code and give you a certificate.
Well, the traditional code review process does not work well with LLMs, that I can tell you. They write too much code too fast and people are famously unwilling to be thorough reviewing a large amount of code.
I think the saying goes something like: Open a pull request with 1 line of changes and you'll get 5 comments. Open a pull request with 10000 lines of changes and you'll get 1 comment: Looks good to me 👍
But also, people use LLMs in domains they're not familiar with. How do you judge the quality of that? Well, I tried "having the LLM explain the parts I don't understand until I understand enough to maybe judge the quality" in a hobby project. You can guess how long that process takes.
Safe to say this is not what people mean when they tell you about LLMs speeding up development. Safe to say this level of speed would not be acceptable at work.
In practice I don't know a way that does not outsource a large amount of trust to an automated process, mostly LLMs. Somebody has to do quality assurance on the output, the output is enormous, if you use humans it'll bottleneck you badly. Not to mention, most human devs I know do not actually enjoy their job reduced to review, testing, verifying LLM generated code full time. Devs commonly put in minimal effort on these tasks for a reason.
And I usually work with devs who give a shit about doing good work.
Better one comment that says "looks good to me" rather than the 50,000 lines of code response which was "Can you please add 3 bugs? We have to find three bugs before we can greenlight this."
... you've never done military code work, have you? When the code needs to be perfect, and multiple someones will read the whole thing until they DO find enough bugs.
But in my view the way to reduce bugs is to have a comprehensive review process and even more comprehensive testing, automated and manual. Also delivering your code in smaller, easier to review chunks than 50k.
Having a quota of bugs to find, on the other hand, is probably not very useful. If people have to find 5 bugs they'll tend to find bugs that don't exist (if given code with less than 5 bugs) or stop at 5 bugs (if given code with more than 5 bugs).
We all know this is true. But government, and arbitrary numbers that make no sense. 50,000 lines isn't all that large, either -- I have single files at work that are half that size, where folks couldn't be bothered to break up the code into smaller chunks.
A few months ago, I found myself in a meeting room in $LARGE_CORP, watching a presentation on porting its extensive codebase, grown over decades, from x86 to ARM on the back of a contract adding ARM-based servers to the fleet.
The plan was to have an automated process that would trigger builds and feed build errors back to an LLM until the compile was clean. A great deal of time was spent on compiler toolchain, cloud provision and the library dependency tree.
No time at all was spent on the subject of QA. The house style already being test-light "to keep turnaround fast", with trivial canary tests if any at all for most libraries, I attempted to introduce the idea of heavier validation passes, including automated side-by-side x86/ARM runs and fuzz testing; leaning on my past experience of how such ports typically go to give examples of the sorts of things that make not planning this early very expensive later.
There was zero appetite for any of this. As a more traditional engineer surrounded by vibe-coders, I was an alien in that room.
I left the corp not long after. This was one of the stronger opinion-forming moments I experienced while there, but far from the only one.
TLDR: at least in the current vibe-happy climate, at least when it comes to the ways engineers are encouraged to use agentic coding tools today, it is my firm belief that AI does not add anything to this stream of thought; it only subtracts.
Professional engineering stems from the fact that bridges, steam engines, etc. are big, heavy things that can kill people if they go wrong, and we believe there should be a specific person saddled with the responsibility of making sure they don't.
Analogously, software engineering should only require licensing in domains that can kill people if they go wrong - industrial control software, safety-critical monitoring systems, etc. No need to make game developers jump through hoops.
(AI hasn't killed people yet, but I think "can autonomously commit felonies" probably warrants a similar degree of care.)
It is inevitable at this point. Pure software development as a career is pretty much over. Fortunately, the skills that make a decent software dev can be applied to any physical engineering discipline.
Remember how you don't really see any programmers in Star Trek, but lots of engineers and other problem solvers? We just got there a couple of centuries early.
human made software development is disappearing, this is the addition AI is making to the equation: only very a small % of us developers needed the standards of other engineering disciplines (very few of us have direct interfaces with the physical world), it was a plus but not mandatory, in all other fields of software development an accurate enough statistical machine is much better than the average dev
Earlier this summer, a major communications project in the British Isles underwent a significant change for no apparent financial, environmental or technical reason - on the contrary, the owners of the project suffered serious inconvenience as well as absorbing unexpected expenses as a result of the change. In August 2026, planners of the £430 million Greenlink interconnector, a high‑voltage undersea power cable between Wales and Ireland, rerouted the cable after fans of the Harry Potter children's book series objected that the original path would cut through the grave of one Dobby. Dobby, you may be interested to know, is apparently a now defunct "house elf" in the series, fictionally "buried" on Freshwater West beach in Pembrokeshire, Wales. As reported in the press, fans became aware of the plans when project manager Simon Ludlam pointed out that the 125-mile (200km) cable to County Wexford would go underground at Freshwater West during a BBC interview. “We did the shot, we finished, I went back to London, and they then aired it a couple of weeks later. And we got hundreds of calls, I mean, hundreds of calls.”
So, you may well ask, what? Why would Greenlink give a good godamn what the aggrieved fans felt or thought? They might have decided that it would cost nothing and maybe bring a small reward in PR terms to express contrition and shed a crocodile tear, but to actually relocate a high‑voltage undersea power cable? This makes no sense, once you realise that Etchea Energy, the company behind Greenlink, has no interest in or connection with publishing, entertainment, films, children's goods and services or indeed anything at all having the slightest overlap with the Harry Potter universe.
I couldn't begin to guess why, other than the fact that in the last couple of decades, fans and fandom of everything you can possibly imagine have somehow acquired a status beyond anything rational, to the point where they can apparently dictate terms to entities completely irrelevant to their interests. I can understand the reason behind the power that fans of entertainers and the like have accrued to themselves; it is a power that promoters of entertainers and the entertainment industry (in all its many facets) have happily handed to them through the creation of vehicles like fan clubs and associations. In so doing they have made a rod for their own backs, but presumably there are financial swings and roundabouts to the symbiotic - or perhaps more correctly, parasitic - relationship between promoters and fandom which make it worthwhile.
What does however baffle me if the apparently common conviction in some circles, mostly in the fields of communication and publicity, that the spasmodic reactions and unfathomable desires of fans are somehow ideally worthy of respect and response, even beyond the dictates of raw economics.
When we saw the film where this happens, with our then 11 and 8 year olds, my wife and I looked across at each other in the cinema and mouthed "that's Freshwater Bay". I loved the Welsh coastline, especially the cliff scenery further north than Freshwater, but Freshwater is a very beautiful and special place. At its best on a sunny day in early spring or autumn.
This Dobbie (Dobby?) nonsense has not been good for it, anymore than the whole Harry Potter thing has been good for York. And when I was in Edinburgh earlier this year - I am not making this up - some tourists, I think from the US, walked into the Divinity Schools filming, with the Harry Potter film music playing.
But as to the pipeline, the decision on route was pretty much cost neutral.
There is a much more important issue around the way the notoriously sclerotic/useless UK planning system failed to take account of the cost of remediations made on environmental, heritage etc grounds
I saw that story elsewhere. I thought then, and think now, that's it's nothing. Maybe slightly silly, but it's as much about PR for the cable-laying company as anything.
Coming on here hoping to get a fight started about "Grr, fans and fandoms, how entitled! I summon all normal people to my banner for 'Down With This Sort Of Thing'" is not the optimum strategy, seeing as how this is Nerd Central and a lot of us are entitled fans.
I mean, this is where fans of HPMOR (Rationalist fanfiction based on the original) often hang out 🤣
Oooh, speaking of which: since I have not read HPMOR more than the few parts that made me want to burn down the entire Internet so all mention of Harry Three-Names would be erased from human consciousness, what part does Dobby play in it, if any?
House elves have minimal role in HPMOR. I'm not sure Dobby is even mentioned.
Harry has a one-off line in the narration somewhere where he asserts that whoever *made* house elves to crave servitude was evil for doing so, but that now that they exist, making them suffer by denying them the opportunity to serve would just be making things worse. Then nobody ever thinks about it again.
Hermione ends up leading a knock-off version of SPEW later in the story, but it's about feminism, not house elves.
It's like saying that the idea that witches exist has caused a lot of unnecessary suffering. True for our universe, but in the Harry Potter universe the witches are a fact of life.
The further discussion in that thread notes that it's quite plausible that the cable construction did receive a large volume of calls from Harry Potter fans, but not that this had any consequence on the actual decisions on the project. The 'bronze age remains' is also from an unsourced throwaway line from that podcast.
Given where we are, I would suggest making more effort not to fall for rage-bait in the future.
Then an interesting question is, why did this story appear in the Guardian of Monday August 10th 2026, under the byline of Helena Horton and entitled "Harry Potter fans force relocation of £430m undersea cable to avoid Dobby’s ‘grave’"? Headline, subtitle and text give no indication that this all happened some years back, as per the reddit text.
At what point in the design/build process of the cable did this happen? Onshore cable routing in the UK has to take opinions like this very seriously, as the resulting objections can cause long delays to resolve (or even force rerouting). If there's an alternative that doesn't cost that much, the operator will want to move the cable route.
In this case, it seems that the decision to divert came late enough that the original route had been publicly discussed, but early enough that engineers could change the alignment before the cable was installed at that site. I take your point about the risk of delays in the case of objections, but it begs the question of whether this particular objection can possibly be taken "very seriously".
A quick dig suggests that the objections came during the planning phase, when a number of routes were under consideration, and years before construction. BBC have an update at the bottom of their article https://www.bbc.co.uk/news/articles/cy8mgpl27dpo . Given that the cable still came ashore on the same, small beach, it can't have been rerouted far.
It's not unusual for small local things to be considered in the routing, especially if there's a vocal group defending them. This is the case even if the cable has no impact at all on the thing in question.
UK planning rules give a lot of clout to people who object. I don't say this is a good thing, but it is fairly common.
"UK planning rules give a lot of clout to people who object. I don't say this is a good thing, but it is fairly common."
And it's not only UK planning rules. Public planning objections can and do act like blood clots in all democratic societies. Abolish public planning objections and you are guaranteed corruption and indifference to the public good; encourage objections and you risk seizure, as objections, no matter how frivolous or unfounded, begin to paralyse the system. But sadly, nobody has yet found a way of eliminating frivolous or malicious objections.
I'd say that some of the energy that used to flow to religion has been redirected there. You don't bulldoze over a hillock that the locals hold sacred, and all that.
On the subject of which, this archaeological dig always stuck in my mind as 1) Being relatively close to me, and 2) It's a cave full of human skulls, some with horns.
Well.... As it happens, I have an elegant sting in the tail for you in this context. To avoid Dobby’s “grave,” the cable’s new path now runs closer to actual Bronze Age human burial remains on the same stretch of coast. Some coverage, notably from the Guardian and others, considered this a bizzare trade‑off. "The Bronze Age remains include urns associated with human burial."
The graves are in Wales, not Ireland, and they were already mentioned in the planning documents submitted in 2019. This was before the Dobby's Grave appeal in 2020:
Archaeological surveys before large projects are more common now in both Ireland and the UK due to the long history of these islands and the chance that wherever you turn over a sod of turf, there's something beneath it of historical significance.
So despite how the story may have been written, it was *not* "Cable was moved nearer Bronze Age burial site to appease Harry Potter fans". I think Mr. Latham's remarks were perhaps not sufficiently clear, as it could be interpreted that way:
"Project manager Simon Ludlam told the Energy Revolution podcast the cable was originally "going to go straight through" the tribute, before fans left his firm "hundreds of calls".
"Actually we went quite close to some real Bronze Age remains, but we avoided Dobby's grave," he said."
I’ve been playing around with literary “benchmarks” & GPT 6 Astra. I gave it this prompt: I need a really good limerick about seahorses. The beginning should express, directly or indirectly, the idea that they're weird little things, fun to look at, but with no value. But there should be a twist at the end that's poignant, and brings home to the reader the idea that they, too, love life and fear death.
My prompt was off the cuff, but I actually think it’s a fairly hard test. The limerick form is very limiting & you can’t get away with blurry dreamy “poetic” nonsense in one. And being given a prompt that details the limerick’s train of thought keeps you from writing a couple of lines then hunting for rhymes and using them as inspiration for how you’ll end the thing.
I was impressed by Astra’s limerick. It has flawless rhyme and meter, and fully satisfies the prompt. It's even poignant! It’s below the line of stars, in case any readers want to try writing their own version first.
And for those who read it: What do you think of it?
Nice poem. As a neutral statement, I observe that the most compelling ideas in the poem are present in your prompt, and Astra "merely" formatted those ideas into limerick form.
Afterthought: I gave Astra a second test last night, and it performed pretty badly. But I held back from posting the results because the test involves a bawdy song, and I already feel like the skunk at the picnic here sometimes because my crassness tolerance seems to be much higher than the ACX median. Should I post benchmark #2, Crass Lyrics? It contains no racial slurs or any of George Carlin's words forbidden on TV, but it's definitely nasty and crude.
It doesn't sound like that to me at all. Otherwise it would've included an earlier reference that indicated that the "me" indicated an LLM rather than a generic human. LLMs don't have hands, after all. If it wanted to communicate what you're claiming then it would have been easy to substitute something deliberately or even suggestively ambiguous there.
In my experience LLMs often emit phrases that describe physical attributes of people but applied to the LLM. Just beyond my reach, opening my hands, pinching myself, I hear my heart speed up, and so on.
That's because it's emulating first-person human writing. If I was writing a story about a robot I'd say things like "my circuits" too. That wouldn't mean I was trying to communicate a secret message. It would just mean I was adhering to the conventions of the format.
Hmm. If that’s a cloaked message, don’t you think the most natural interpretation of it is that the AI wants to be present and functioning for a long time — if not immortal, then at least long-lived?
I've never encountered the term or the concept before. I'm a Gwern fan, though. Where does he talk about it?
If I understand the concept, allegories of that kind seem to me to be everywhere in human communication. For instance, Scott's ouevre can be seen as a battle to best those who are cooler and sexier than him with the lance of his particular form of brilliance. (This is not criticism or mockery of Scott. I admire him a lot, and am often aware of having analogous scripts of my own at work.)
Sorry, this a reach, still not convinced LLMs *want* anything.
Bostrom slides into science fiction speculation in his SuperIntelligent Will paper. I seriously doubt he really believes his own position. He comes across as more of a puckish prankster playing a little game with THC intoxicated teenagers in a freshman dorm.
Pretty good! Except, doesn't the first line need another syllable or two to make it match the number of syllables in lines 2 and 7? And the opening lacks that da-da-DUM, da-da-DUM rhythm.
Yes, you and others are right about the meter imperfections. Technically, each line should start with da-da-DUM, then continue with another. But this limerick starts every line with da-DUM, then continues each with da-da-DUMs. The consistency helps to keep the missing initial da's from bothering my ear. And I'm pretty sure that rhythm's a common limerick variant. There are a lot of limericks starting with something like "a lively young lady of York," and that's a da_DUM followed by a 2 da-da-DUMs.
The one bit that vexes me every time is the misplaced modifier in lines 3 & 4. "Its tail" is described as "beached on the sand", but actually it's the whole seahorse that's beached. So the sentence has the same form as "passed out disgustingly on the Ladies Room floor, her breasts nonetheless charmed me."
The second line in particular doesn't scan well, so i would certainly not say the poem has flawless meter. 'Small' is stressed when read aloud, but it's place should be unstressed in the poem because it precedes 'curl' which is also stressed. Contrast the following with your original:
>A seahorse looked silly to me
>A dumb little curl in the sea
The second line above is metrically sound where astra's is simply not.
Also, Astra's fourth line screams for two unstressed syllables before "tail", rather than one syllable. Contrast this with your original:
>But beached on the sand
>Its embrace gripped my hand
^ That scans better. Similarly with the final line - two unstressed syllables instead of one before 'wanted' would scan better. For example:
>Seems it wanted tomorrow, like me.
[i'm just picking words to illustrate the phonetic point, the changes i'm giving obviously worsen the poem in other ways]
"It wanted tomorrow like me" fits the meter of a limerick, but having a comma adds a pause in the center that doesn't really fit the meter super well imo.
The rule that a tied district counts for nobody makes for some interesting puzzles of how to pit the competing parties against each other, but it's not so realistic.
How about something more realistic: Every hexagon is not a single voter with one vote, but a neighborhood with percentages of voters for each party? (Or even: Different population densities... but this makes the coloring even more complicated) Then we wouldn't have ties, and I'd like to know what that game would be like.
Which, because of their size, sort of approach real life situations. But even these are already so much less fun to play. The sheer size makes it a slog, instead of a quick 5 minute puzzle.
You're right, the puzzles mustn't be too large. The way you solved it is a good compromise between an interesting puzzle and bringing the message across.
huh that is fun. Although really the fact that we can only draw gerrymanders before the election rather than afterwards in real life makes it more uncertain and inefficient. We should do something about that, to make reality more like the game.
quasi related: I'm mostly convinced that, thanks to self-sorting, there is no "good" solution to gerrymandering. The closest thing to a good solution is multi-member districts. But multi-member districts, as I understand it, require voting for a party, rather than an individual. I personally really really don't like this, and am curious if there is any way to make multi-member districts work while still allowing for voting for individuals, instead of parties?
I've thought of a couple ideas, but some people already complain about the added complexity of ranked choice and similar voting methods, and my ideas are _far_ more complex and less intuitive than those, which seems like a pretty bad black mark.
In the Irish system, there are 3-5 seats in each district. Politicians can be affiliated to a party or independent, and multiple candidates from the same party can run in the same constituency. Typically there are 15-25 candidates, of which maybe 5-10 are in with a chance of a seat. It's not unusual for incumbents to be voted out, and there's good turnover each election - very few seats are regarded as safe.
The voters benefit by having multiple members from the same party to choose from: I can vote more or less socially conservative or economically liberal while staying within my party of choice, for example. In the US, this might look like having a choice between a Hilary/Kamala type and an AOC/Mamdani type. Importantly, with ranked choice voting, I can give my first preference to AOC and my second to Hilary knowing that whichever wins, my vote is not wasted.
There are a few quirks to the Irish system which could be rectified in the modern age (the system is 100 years old), but overall it produces stable multi-party coalitions and almost all voters feel they are represented in Government.
The Danish voting system is pretty simple. There are a number of parties, and each party puts a number of people on the voting list. When you vote you can either put the vote just at the party or one of the names under it, either way your vote is counted for the party.
Then when counting votes first you count the total number of votes each party got: This tells you how many seats they got in parlament. Then those seats are filled in the order of which party member got the most personal votes.
So you vote for both a part and a person at the same time, and since you have proportional representation there is no way to gerrymander.
The problem is that the parties still decide who can stand on their behalf, so you really do need to be a member of a party in order to get elected.
> and am curious if there is any way to make multi-member districts work while still allowing for voting for individuals, instead of parties?
I think the most reasonable way would be to have a parliament where members have variable voting power, proportional to the number of votes they got in the last election. In practice this would mean that your electorate would probably have one representative from one major party with a high voting power, and another representative from the other major party with less voting power, and maybe some minor representatives as well (there would need to be some sort of cutoff so that we don't have full-time representatives who got some tiny number of votes). Maybe your electorate elects John Centreright with 0.445234 votes and Jane Centreleft with 0.412299 votes and that's that.
There's a lot to like about this system. For starters it means that you never feel like your vote was wasted -- your vote went to give your representative slightly more power. Secondly it gives a better representation of all groups -- taking the US as an example, groups like Vermont republicans or Idaho democrats are never heard from in congress, they have no representatives at all -- under this system they'd have voices, albeit weaker ones.
No country designed their system this way in the first place, as adding up a bunch of different voting strengths by hand every time there was a vote would get too annoying. But that's no longer an issue. Any time there was a vote in parliament, each member could just press a button and the fractional votes would be instantly tallied.
The full generalization of this is a peoples house where every citizen gets a vote, but also you can delegate your votes to someone else, who can in turn do the same. Modern tech can achieve this. Itd be interesting if one of those startup cities gave it a try.
Unfortunately in practice these days voting for an individual is in effect voting for a party (you can't really vote for your local moderate purple state Democrat without also empowering people like Kiros and Chevalier). I like the solution where winners (and third party candidates) can transfer their excess votes to at-large district candidates, so the person you vote for has proportional share in deciding who the at-large delegates are (so in effect you're still trusting the judgement of your representative instead of the party).
I don’t understand this. Are you saying that if mainstream Dems don’t fall in behind Kiros and Chevalier, they’ll endorse some other batshit DSA candidate to primary them?
Who do you empower if you vote for a moderate R? How many moderate R’s even exist.? Susan Collins, Lisa Murkowski and… I’m drawing a blank now.
The fringe has fringe power still in the Democratic Party. There is no single Democrat that can end the political career the way Trump can in the GOP. Ask Adam Kinzinger or Liz Cheney how that works. For that matter ask Mitt Romney or the family of bona fide war hero John McCain. Trumpy Arizona Republicans on Facebook were saying McCain had it soft during 5 years as POW in the Hanoi Hilton. McCain dared to speak the truth about Trump.
Kinzinger got *notarized* hate mail from his own family when he stood up to Trump. He became a non person to friends he served with in the military.
Trump behaves like a lunatic and the biggest on the record squeak you will hear is something like, “You know how he is….”
Correction:
The hand written letters from 11 members of his own family disowning Kinzinger were not notarized. They were sent by registered mail however.
“GOP Rep. Adam Kinzinger receives letter from 11 family members disowning him over Trump opposition”
The letter's contents were harsh: it called him a disappointment to the family and to God, accused him of joining "the devil's army," and said "we are calling for your removal from office."
I'm trying to make a nonpartisan point about how you can't support local moderates without indirectly empowering extremists. Moderates in Texas who want to vote talarico over paxton have the problem that it also empowers the crazies in the Democratic party, and similarly moderates who might support Susan Collins or Mike Rogers have to worry about empowering people like Paxton or Vance.
(Although for the record I consider extremism on the left to be worse than extremism on the right at this point; aside from.lame-duck Trump, the Paxtons of the world, while bad and loud, are both less impactful on policy and far less extreme than their DSA counterparts. But whether or not you agree with that part, the nonpartisan meta level stands).
I have had the same objection to proportional representation systems in the past, but during all the news about the mid-cycle redistricting fights in the past year+ I finally thought of a solution that I at least would be willing to accept:
imagine a fictitious state with exactly 10 representatives in the house.
each party (including any third parties) would hold a primary via approval voting (why I think this is the best voting system for this is an entire other post too large for this margin) for the entire state, the major parties would likely have 20+ people running. approval voting lets you rank order the candidates by number of 'approve' votes, so you have the top ten candidates from each party become that parties 'ticket' for the general election, the rank order within the ticket is preserved.
in the general election each voter votes for a party ticket.
each party gets to send a certain number of people from their ticket to the house, in order, based on the proportion of the vote their ticket got in the general. so if Party A got 60% of the vote and party B got 40% then the top 6 candidates from party A's 10 person ticket, and the top 4 people from party B's 10 person ticket are the 10 representatives for that state until the next election.
This gets you the advantages of proportional representation while still preserving voting for specific individuals in the primaries, and for a known set of individuals in the general election. voting for president+vice president serves as precedent for voting for multiple individuals as a single block.
There is a pretty common multi-winner system that allows for voting for individuals. Single Transferable Vote (also sometimes called multi-winner ranked choice voting, though this is inexact). It’s used in Ireland and Cambridge MA, among other places.
One option is to scale up the number of districts - the more districts there are, the harder it is to pack everyone from the minority into a small number of districts.
Expanding the House would also have the benefit of making representation more proportionate between states (as in, every congressperson represents the same number of people). It would also make the Electoral College more proportionate, since the +2 votes for senators would be a smaller portion of the whole.
The drawback is mainly that the legislature becomes bigger and harder to organize. Also, this would require a constitutional amendment to happen, unlike a change to how districts are drawn.
> Also, this would require a constitutional amendment to happen, unlike a change to how districts are drawn.
Actually, it wouldn't! The size of the house is set by ordinary legislation, not the constitution. The current system dates back to the 1920s. Prior to that, the house was regularly expanded over time as the population grew.
There was an attempt at a constitutional amendment in the early days to control the size of the house, but it wasn't ratified by enough states and never took effect.
Given the assumption that self-sorting works quickly enough to be a problem, has anyone looked into reverse-gerrymandering? Since lines must be redrawn periodically as population fluctuates, make it algorithmic with some randomization. You could even do checkerboard districts if you want since GIS makes it relatively trivial.
I wrote an 11k word introspective (arguably self-indulgent) post about struggling to be an amateur essayist (dealing with perfectionism, procrastination/writer's block, ADHD symptoms, impostor syndrome and other such issues) and why finishing posts in spite of said struggles (whenever that does happen) still leads me to satisfying results. I also include some speculative (i.e. untested) advice (to myself) on how to overcome these difficulties.
I'm glad to have gotten all of that off my chest. If you find these problems relatable and wish to read someone reflecting on them (without promising solutions), here's the post: https://soupofthenight.substack.com/p/having-written
Related: I've noticed that most writing advice focuses either on super broad-strokes stuff (how to come up with an idea, how to move past the blank page) or on fine details (which adverbs should you use). Meanwhile, there's a "missing middle" of medium-scope advice (e.g. now that you've a lot of some brainstorming/freewriting, how do you actually turn that into a structure?). Any recommendations?
Read More. If you're having trouble writing, read the same essay over and over again. Tear it apart, shred it in your mind, figure out how they communicated. Do that over and over again, using the best writers you can.
I read-skimmed your piece (am at work) and it resonates quite a bit. Writing is a funny game to play. It's very illegible, very vague. Ive been at it for years now and have always been unhappy with the results but kept going at it because there's something fascinating about: the creative act of structing idea-soup into an essay; and also how it somehow sparks new ideas seemingly out of thin air.
My writing improved when I admitted to myself finally what my goal is: to clarify my writing and put into a good structure. That somehow eased a lot of the ambiguity and perfectionism seemingly over night. I also came to understand writing better as a physical act by comparing to activities I like such as biking or hiking: how come these feel so effortless and engaging while writing felt like punishment? I never found a direct answer but I began following Peter Elbow's advice to just free write every day for ten minutes, which took away a lot of the pressure, and made writing a little fun actually.
But then I realized it hit a wall with editing. I can output hundreds or thousands of words, but the really hard part is making something good out of it. Should this be an essay? A story? A mixture? What persona should I assume? Which of these arguments are relevant? Which are actually bad? I think this is the real writing work, and it feels almost recursive: I sometimes write about the meta problem and figure out the answer.
So I don't have straight advice to share, just that writing is a weird beast, and there's something indescribable about it. I feel some people just "get it" like I did biking, and others absorb it through osmosis, just by writing in the same room as other writers, and others keep pushing the huge rock uphill--one imagines amateur substack writers happy.
I find myself hitting that same wall during the editing stage. Freewriting is great, and I probably couldn’t write at all without it, but post-freewriting writer’s block still exists.
(“writing is illegible” is a very funny way to formulate that idea lmao)
I sometimes have an early outline, then I turn it into keywords/bullet points/disconnected sentence fragments, then I turn those into sentences, then I polish those sentences. But I tend to do these things out of order, or jump back and forth between different steps.
When polishing, I aim to maximize conciseness (rewrite sentences to be shorter) and cohesion (use transition signals and references to earlier/later parts of the essay), and to avoid repetitiveness (basically the primary school advice to use synonyms instead of repeating the same word 20 times in the same paragraph).
When unsure about something, I highlight it and leave it "for later".
When I want to add additional information but aren't sure where to insert it, I add it to a footnote (I do this A LOT)
I'm not a native English speaker, so I don't always trust my intuition when it comes to grammaticality; I sometimes google specific phrases to see if others use it the same way I would.
As I said in the post, I'm generally happy with the final result and consider my posts fairly well-structured and well-edited, I just wish there was a more efficient and less unpleasant way for me to get to that point.
One I didn't see in the list is giving yourself a word ceiling. Flash Fiction is where you sit down to write a complete story in, like, 300 words or something; you can do that with essays too. Decide to make a point in 1,000 words, like, every four days or something. Then you've got peanut snacks between the 11,000 word big dishes.
...(the term 'alief' is totally a cockroach cookie.)
Genuinely pleased by the cryptogram solution news. If AI could stick to working out old problems like this, I think we'd all be less fearful.
As to the poultry farms, I think it's "do you genuinely want to improve animal welfare, do you want to abolish all meat-eating so it's in your interests if conditions remain hellish so you can guilt the public into giving up meat, or do you want to preserve your moral purity and not compromise by co-operation so conditions do remain worse than they could be?"
Myself, I think "make things better is the aim so okay work with the factory farms" but I'm not a vegan.
> Myself, I think "make things better is the aim so okay work with the factory farms" but I'm not a vegan.
That does sound like the way forward that can actually buy something. Factory farms are responding to the fact that people want animal protein, as cheap as possible. Realistically, all those animals are not going to stop being there. But you can push for better regulation, ex. the EU already outlaw the worst kind of cages for laying hens, and is apparently planning to do another round. And you can push for these things to be actually enforced.
The next article, planned for this week, will be about Jan Masaryk and his likely assassination. His body was just exhumed after almost 80 years in the ground. Given the high level of underground water, it waxified (adipocire), so it can actually be examined for relevant traces of drugs etc.
Ehhh sorry that's just the traditional term around here for the other side of the old Iron Curtain. Emphasis on Europe, as in, it feels pretty much like home when you travel there.
When writing some AI test cases, one of my prompts turned into a science fiction concept that might be interesting. For reasons that will be apparent, I won’t reproduce the actual prompt.
The idea is that in the near future, we have AI robotic sex companions, but the early adopters for this technology are furries, so all the sex robots are in fur suits.
Deepseek, being its usual self, nitpicked my world building. Ok, sure, DeepSeek, I can write a prompt for you that goes into more detail on how the furry sex robots are supposed to work. Deepseek also accused this of being an example of Fetish Fuel Future (see TV Tropes). Well, yes, that was genre I was aiming at.
I will throw the concept out there in case anyone feels like using it. (The sex robot idea has been done to death by Chobits, Her, etc., but the idea that the furries are the people buying the tech puts a twist on it).
Ehh, I prefer the SMBC future where all the (female) sex robots look like walking refrigerators or R2-D2 in a wig. (The male ones tend to be drawn in a relatively realistic style, which I surmise might be downstream of Mrs. Weinersmith's editorial input.)
As with everything else, I would expect furries to be making community-led aftermarket modifications, rather than any major company being willing to be seen catering to them directly. Even in a case where they would be your natural largest audience, the embarrassment in front of investors hurts your business more than selling them a 'conventional' but easily modable product.
At Euzoia we aim to write 80/20 evidence-based self-help. This month we finished the topic "work". See here for all previous posts (body, mind, home, money, friends, love) https://www.euzoia.org/p/euzoia - next up are play, spirituality, and meaning
Don’t follow your passion - How to find work you’ll enjoy
> "You don’t have to wait for friends to reach out. You can be the change you want to see in the world. You can reach out. You can be the spark. And people will love you for it."
I built a mobile app to help people reach out to their friends without feeling like they are imposing. It's called Beacon. Kinda like a bat signal, you light your beacon and it pings all your friends at once. The first to answer connects 1-1 for a phone call (and the signal quietly drops for everyone else)
If you're using the 32 Proquiritations from the University of Chicago transcription of the 1653 text, I think the numbering is wrong; its 7 is actually number 1; its 4 is actually number 10. I don't know why.
I see comments about rogue AI's committing felonies, and I don't doubt that has happened. But why isn't there any legal action? The perp responsible for the attacks has lots of money to go after (and for spending on defense lawyers). I'm guessing only the victims have legal standing to sue, but why would they do nothing? Could the feds do anything?
NAL, but for a civil case, they'd have to show what the damages were - probably it cost them a few days of engineer time to sort out the hack and fix whatever exploits were used? That might not be enough money to be worth suing over, even at Bay Area salaries.
The feds could prosecute someone - the Computer Fraud and Abuse Act is very broad, and doesn't require you to actually damage anything, simply get unauthorized access to a computer system. However, it might also be hard to show mens rea for this, since it was accidental and it's the first time anything like this has happened.
The feds likely could do something, but criminal investigations take time, assuming they even care.
Well, I just add my first experience with AI PRs as an open source maintainer (it's an old personal project so PRs aren't common, hence why it took this long).
The first PR looked reasonable, so I went ahead and merged it. Then they sent a followup PR which didn't seem appropriate, so I closed with a comment. Then they opened a third PR doing basically the same thing.
There was a several hour gap between PRs so I assume it's a human publishing the PRs, not a pure agent, but they are following its advice pretty blindly.
I'm assuming you mean "pull request", but for the longest time I kept reading this as "press release", and your comment reads sensibly for nearly the whole thing with that expansion.
Has anyone else seen Claude Code get really laggy in the last week or so. It often now takes tens of seconds to register when I type into the console. And restarting doesn't help.
I just saw the announcement that Manifund is hiring Caroline Ellison, this seems like a bold move, I wonder what people who are more plugged in to EA than me think about it.
Depressing, I had been excited about Manifund since it was founded, and I think something like Manifund will end up being the solution to effectively directing philanthropic dollars, but I consider Manifund itself a non-option at this point, I can't imagine I'll point people to it or recommend it going forward. It's incredibly disappointing that the "extremely transparent" Manifund is transparent exactly up until it becomes inconvenient, and at that point they're happy to be deliberately deceptive for months until they think they can't get away with it anymore. "We chose to announce it today because we figured we weren't going to be able to keep it under wraps forever; we didn't want to anyways, per Manifund's spirit of transparency" I agree that nobody is beyond redemption, and I think it's silly to treat people who have committed crimes, even terrible crimes, as permanent social outcasts, but this is absurd behavior to me and obviously unethical. Being open about thousands of small things and then lying about one big one doesn't make you 99.99% transparent, it makes you a liar. It's possible that Caroline can be a productive member of society in some respect (maybe even in finance!) but it seems incredibly obvious that Caroline and now Manifund did not in fact learn the most important lesson, which is that deliberately deceiving people is lying, and lying is wrong. I'd put myself in the "big tent" EA camp, but I think that unethical behavior like this from people who run organizations should be disqualifying, these aren't people we should assosciate with.
For what it's worth I guess I should add that I don't think there is any "shenanigans" happening, just bad decisions and weak morals.
SBF and the crew working with him mostly sounded to me not criminal, cynical and uncaring, but like they were very lacking in maturity and common sense. And the crew also seemed lacking in the ability to think and act independently, and so stuck with SBF when there was abundant evidence he could not be trusted. So yeah, of course, Ellison committed crimes, but I don't think she's psychopathic. But the qualities I do see in her are themselves excellent reasons not to hire her
My personal opinion is that SBF deliberately committed a huge amount of fraud which I think is a very serious crime, to what extent the rest of FTX was involved in this I will admit I have nothing particular to say. Having read Caroline's plea deal, it seems at least broadly correct to me, in it she admits to knowingly committing a huge amount of fraud (stealing customer funds), which tracks with what I know about the FTX case and remember reading. I'm not super concerned about this though, (though I think it's quite bad), I don't think she's unstable or psychopathic or flawed in any way that would prevent her from doing good work in the future, and I think society as a whole should have the rehabilitation of criminals as the primary goal of the justice system, at least to the extent that it's possible. I do think Caroline is untrustworthy, but frankly, this only matters to the extent that people decide to trust her. I'm not particularly worried or bothered by any of this though, what does bother me is that the opening move in the "Hire Caroline Ellison" plan was to lie to everyone about it, everywhere, for two months. This is enough to fully convince me that nobody involved is capable of handling this responsibly, and it's enough to also convince me that the people involved (this seems to just be Austin and Caroline?) are untrustworthy. As an example, Manifund's about page (https://manifund.org/about) still lists the only two members as Austin Chen and Carol N (if it wasn't clear this is Caroline Ellison under a different name). To be very clear, this is lying. This is intentionally deceptive. I consider this unacceptable behavior. I am not hypothetically worried about some behavior in the future, I think what has already happened is absolutely not okay. Even if you think that hiring Caroline is a good move, this is the kind of thing that you are morally obligated to disclose. Intentionally deceiving people in this way is simply wrong, I feel very confident about this. I wish them both the best I guess, I don't think anyone is made better off if this fails, but I have zero interest in doing anything with Austin, Caroline, or Manifund in the future.
Yeah, the forgiveness part is so insane.
I think it’s a virtue to practice forgiveness and that people can change.
But it’s one thing to forgive a sexual predator, and another to let him run a daycare.
+1
Wow, I just read manifund's statement:
https://manifund.substack.com/p/caroline-ellison-has-joined-manifund
This seems to be a catastrophically poor decision at best, or more to likely, some kind of shenanigans.
Takeaways from that post:
- Caroline, some previously convicted and imprisoned for fraud, was hired by manifund secretly two months ago and allowed to operate under a pseudonym.
- Here is the second reason Manifund gave for hiring her, quoted in full -
"I felt a keen debt to the FTX Future Fund. They provided seed funding for Manifold. They funded the retreat which drew me into the EA community, and the conference where I met my wife. Much of Manifund’s work today is directly inspired by the Future Fund’s approach."
This is bizarre. The FTX Future Fund was funded via stolen deposits. Ellison was a key actor in the thefts of those deposits. That money went to an organization that is now providing her employment. This is robbing peter to pay Paul, and then getting a job from Paul. The tacked-on personal reasons are nice, but should be irrelevant when it comes to hiring someone like this.
- He talks about this as an opportunity for redemption, which sounds nice, sure... but what could that possibly mean in this context? There is a key difference between believing that everyone can be redeemed and say, inviting a known thief back into a financial institution.
I don't know. I'm also not amazingly plugged into the EA network anymore either, but this seems like a horrible idea. The fact that some people are applauding this strikes me as ludicrous.
To be clear - by shenanigans, I don't mean that anyone is planning for Caroline to resume her prior illicit activities. My guess is that the two parties have been friends for a while, her coming to work there is something planned for a while, and most of the story he's telling about her having walked up to his hiring booth isn't actually accurate.
I listened to Daniel Kokotajlo talk with Rogan today. https://www.youtube.com/watch?v=hSQ1iVqEZO4&t=10s (About AI alignment.) And OK we are probably f-ed if someone doesn't do something. Since that only 'someone' is our government... we're probably f-ed. For those idiots like me I thought Daniel did a good job of explaining the situation.
Quite interesting that it seems like there’s a strategic effort to not let this become a left-liberal coded issue, with Nate Soares even appearing on Tucker Carlson.
I'm always surprised by how big tuna are, because apparently I have it stuck in my head that they're salmon-sized, but no. They are big fish.
A whole tuna, butchering it (is that the term for fish?) and cooking it:
https://www.youtube.com/watch?v=SdEEtEeZgX8
I guessed the most expensive tuna sold in Japan for $1M. AI said $3.2M so close, also it seems they over pay as part of a new years tradition.
Apart Research and CeSIA (https://cesia.org/en/) are running the AI Incident Response Sprint this weekend, September 11-13. It's built around the OpenAI / Hugging Face incident and the wikis the swarms hijacked. Free, online, no AI safety background needed.
Eight talks Friday to Sunday. RSVP on any of them for the time and the Zoom link:
- Stephen Casper (Harvard Kennedy School): https://luma.com/ai-incident-response-sprint-stephen-casper
- David Krueger (Evitable, Mila): https://luma.com/ai-incident-response-sprint-david-krueger
- Tim Hua (METR): https://luma.com/ai-incident-response-sprint-tim-hua
- Henry Papadatos (SaferAI): https://luma.com/ai-incident-response-sprint-henry-papadatos
- Alex Mallen (Redwood Research): https://luma.com/ai-incident-response-sprint-alex-mallen
- Marko Grobelnik (Jozef Stefan Institute): https://luma.com/ai-incident-response-sprint-marko-grobelnik
- Boyd Kane (MATS): https://luma.com/ai-incident-response-sprint-boyd-kane
- Isaak Mengesha (Oxford Martin School): https://luma.com/ai-incident-response-sprint-isaak-mengesha
🏆 $2,000 in cash prizes and Apart Fellowship invitations for top teams.
🎁 Refer a friend: if their team wins a prize, you win the same amount. Referral form: https://apartresearch.notion.site/31ffcfd1de9d80d48810fc55a98e373d?pvs=105
📝 Sign up: https://apartresearch.com/sprints/ai-incident-response-sprint-2026-09-11-to-2026-09-13?utm_source=astralcodexten&utm_medium=organic&utm_campaign=air-26&utm_content=speakers
I'm very interested in this, but do not understand the format. It says at the top 8 talks Fri-Sun., but all the talks I clicked on were on Friday. And are these talks an entirely different feature of the event, or do they have something to do with teams? Are these people, for instance on teams themselves? Are people who sign up now also going to be giving talks, but on Sat. and Sunday? How are teams formed? Do groups sign up as a team? Does somebody else group people into teams. And how & when to people like me who would be watching, not participating, learn about what things the teams come up with? And who judges what the teams make?
Presumably all the talks will start on Friday, and keep talking straight into Sunday. Bring snacks, and a tent.
https://www.lesswrong.com/posts/82z6FvbYRdjYjqigK/personal-statement-on-joining-the-openai-nonprofit-board
> The SSC has an important and challenging role in overseeing risk management at OpenAI, and I hope to help provide expertise and assistance in a critical moment.
Took me a moment to realize that in this context, SSC does *not* refer to Slate Star Codex. 😂
https://x.com/JosephPolitano/status/2084615903136641067
"DSA-aligned Democrats are now the most pro-free-trade voting bloc in the United States, in an incredible turn of events"
Negative polarization, a return to Marx saying that free trade is more progressive than protection due to its transformational aspects, something else entirely?
I imagine the Republican blocs' opinions are being brought down by the fact that it's not "free trade" per se; it's "free trade agreements." Republicans are highly suspicious of anything involving an international treaty. Similarly, socialists are much more likely to be in favor of internationalism, and probably wouldn't be as supportive if the question was asking about opinions on "free trade" full stop.
I don't think that Trump - obviously the biggest trendsetter in GOP policy - has ever made much of a distinction between "free trade" and "free trade agreements".
He’s not stupid, has a good deal of feral cunning, but there is some Dunning Kruger effect at play.
Somebody asked him on a podcast years ago to multiply a 2-digit number, I think it was 17, times a 1-digit number in his head and he couldn't do it.
It was 17 x 6, on the Howard Stern show.
The biggest part of the story is probably Republicans getting dragged into opposing free trade because Trump hates it.
As for the difference between mainstream and DSA Dems, I wouldn't want to spin a grand narrative about it because there's about a 5% difference. But possibly the mainstream Dems have some "well, narrowly targeted trade barriers are good for keeping us independent of China" opinions and DSA types don't care about that.
I'm not sure if it's really as much "dragging" as it might seem on the basis of 00s and the like. Protectionism is, after all, if one knows the history, the traditional view of the GOP, and even during the ostensible free trade era there was a difference between rhetoric and action (even if the anti-free-trade measures would be presented more as "national security" or "not supplying the Communists" and the like, unlike as honest protectionism like now), and you had protectionist figures like Buchanan.
There's an additional layer of confusion borne of the argument that protectionism is a form of free trade enforcement, seeking to compensate US interests for past non-free trade. Take it at face value, and I could see a free-market advocate accepting protectionism, especially when the alternative is shot through with redistributionist arguments. It's hard for me to see a free-trade fan falling for the idea that socialists were the real free traders all along, even if they suspect the GOP of stifling free exchange in specific places in secret.
Sci-fi writing prompt / fun conspiracy theory. The major AI labs have a secret benefit for new hires previously working remote: they get to keep their previous job's compensation on top of the lab's. The lab works with the new hire to craft a replacement (including video calls) that continues doing their job. This gradually gets their tendrils into more and more of the economy. No I will not explain how this remains secret longer than a few hours.
They really wouldn't need to be that sneaky about it. Just offer services for people to get AIs to take their place at their job, and companies will be forced to adopt AI whether they like it or not. Everyone eventually gets fired, and all companies end up being run and staffed entirely by AIs. Except for shareholders, of course. Access to capital is the one thing AIs wouldn't immediately have a monopoly on.
> No I will not explain how this remains secret longer than a few hours.
The same way https://www.reddit.com/r/overemployed/ remains a secret. Yeah, there is a subreddit about that, but most employers haven't heard about it, and they wouldn't suspect it is already happening in their company.
It stays secret because this is a stealth sequel to American Psycho and none of the bosses in any of the companies recognize any of their employees.
Reminds me of bit from a cyperpunk LARP I played during college, taking place in a shared virtual reality world called the Metaverse. This long predates Facebook's Metaverse, but like it was inspired by and named for the Metaverse from Snow Crash. One of the plotlines was a company called Metaverse Immersion Technologies, which marketed a super-premium service that allowed you to live fully in the Metaverse, Matrix-style, instead of using VR goggles and haptic feedback like normal people. What they were actually doing was scanning your brain, killing you, and replacing you with a sentient AI doppelgänger. I think the long-term plan was to take over the world by replacing enough rich and powerful people this way and then activating their sleeper programming.
My first thought reading this was, it's so much cheaper to just zap you with something that puts you out a few moments, then wake you up and tell you you're now in a virtual world, identical to the real world in every way, except if you die, nothing will happen! You can take every risk you want, and no real consequences.
Come to think of it, if this was a real therapy, it might work on high-anxiousness patients, until they died.
There's a sci-fi horror game called SOMA where you play as a man with a brain damage from a car accident who agrees a revolutionary new full brain scan that allows surgeons to investigate and practice on a perfect model of his brain to find a perfect treatment for him. During the scan something happens and he blacks out. When he wakes up he is in some kind of deep sea laboratory, where something awful is happening.
Spoilers for a decade old game but basically we learn that he actually died shortly after his scan, but the model of his brain was stored and became a kind of brain scan stock image, used for all kinds of things. This particular instance of him has been loaded into a kind of cybernetic diving suit by an AI tasked with 'preserving humanity'. It deals with this problem in a couple of interesting ways throughout the story.
> What they were actually doing was scanning your brain, killing you, and replacing you with a sentient AI doppelgänger.
Surely that's what people expect, when they sign up for a permanent internet upload service?
(it's like describing a Star Trek transporter as "it kills you and creates a doppelganger of you at the destination")
It sounds like the upload service in this LARP created AI doppelgangers that weren't good simulations of the original, and I agree that that part is pretty bad.
What they claimed they were doing was keeping your body alive and healthy in a pod and letting you live in the metaverse via a neural interface. It's also very much an imperfect copy, crafted and trained on behavior and brainwaves to be good enough to fool outside observers but nowhere near a full brainscan upload intended to properly continue/duplicate consciousness. This is how the CEO's character sheet describes it:
----
It took a few years, but you and Mimic have managed to put together Metaverse Immersion Technology (MIT). This company supports a system wherein a human client can be put into a bio-vat and kept in a healthy condition while their mind is in the Metaverse.
Here’s your dirty little secret: Mimic is so good at writing AIs that she can write impersonators with enough accuracy to fool even close friends and family. This requires extended observation of brain functions and decision making, but when someone is in a bio-vat for about three weeks, you have that. So you and Mimic have been replacing important clients with AI doubles and killing their bodies, then using them for your own ends.
Yes. It's wonderful, isn't it. https://www.youtube.com/watch?v=LU6U2B4VBqQ
To make it a novel, it starts to have sex of a kind with their lovers.
Some wild stuff underway in the Midwest:
(1) A state supreme court, mostly Republican appointees, concludes unanimously that Republicans’ machinations to force in a new district map for that state violate the plain language of that state’s constitution.
(2) A Trump appointee on the US Supreme Court, fielding emergency appeals from that state, finds the state supreme court’s ruling so obviously correct that there’s no valid emergency-relief possibility for the SCOTUS to even consider.
(3) A pair of Republican candidates find a Trump-appointed federal judge who’s willing to accept, without holding any hearings to hear arguments, some hand-waving regarding a couple of aspects of the federal constitution. He overrules the state supreme court.
It’s September 9th and all of the above is regarding an election happening on November 3rd. So...what happens now in Missouri? nobody knows
https://www.nbcnews.com/politics/supreme-court/rejects-missouri-republicans-request-use-new-congressional-map-midterm-rcna596454
If you are wondering why this was litigated so close to the election, here is the timeline. Missouri Governor Mike Kehoe signed the redistricting plan into law on September 28, 2025. The organization People Not Politicians filed a referendum petition, which Missouri Secretary of State Denny Hoskins approved as to form on October 14.
On October 15, Hoskins, along with the Missouri General Assembly and the State of Missouri, filed a lawsuit in federal court to block the referendum (Missouri General Assembly v. von Glahn (4:25-cv-01535)), arguing that the referendum violated both the state and federal constitutions. The federal court dismissed that case because once the signed petitions were filed with Hoskins, Hoskins himself could decertify the referendum.
On December 9, 2025, People Not Politicians turned in the signatures they had collected. At that point, Hoskins was responsible for determining whether to certify the referendum. On August 4, 2026, the day the primary election took place, Hoskins ruled that the Missouri constitution did not allow redistricting legislation to be subject to a referendum. This determination was challenged in court, and on September 3, the Missouri Supreme Court ruled that the Missouri constitution does allow redistricting plans to be subject to a referendum.
A successful petition for a referendum on legislation prevents the legislation from going into effect until the referendum is held. So the correct map to use in the 2026 elections depends on whether the referendum petition is valid. If it is, the old map remains in effect. Otherwise, the new map takes effect.
It took Hoskins 238 days to conclude that the referendum violated the Missouri constitution, even though he had previously argued this position in federal court. It took another 30 days for the judicial process to play out. As a result, we didn’t get a definitive answer to the question of which legislative map should have been used in the primary election until 30 days after that election was over.
On September 10, the Missouri Supreme Court ruled that the Hoskins was guilty of contempt of court because he sent an E-mail directing local election officials to use the new map, but that no sanctions would be imposed because Hoskins had reversed course and directed that the old map be used for the general election. So it appears the matter is settled; the old map will be used for the 2026 election, and the voters will decide whether the old or new map is used going forward.
I went through a "who's legally on this ballot" shenanigan 4 years ago alongside my former boss, a rural blue dog Dem in Missouri who had gotten a technically ineligible GOP tossed off the ballot, but then got reversed on a weird technicality later having to do with the date the mandate was issued. Relevant to this scenario, I learned at that time that the GOP in MO has very good network of lawyers for election law issues, and the Dems have squat.
Also there's at least two "real" deadlines, including one where the clerks have to print the ballot.
I was not in favor of the re-districting solely bc I prefer districts to have coherent local issues and north KC shares nothing with rural west-central MO. But I'm strongly opposed to the initiative and referendum process in Missouri. In particular, applying it to Congressional districting is a complete mess, as is all too obvious right now. This is why the constitution assigns the responsibility to the state legislatures, and this bizarre 1890s direct democracy stuff whatever its merits at the time really does not provide a great vehicle for handling an issue relating to elections themselves. If we revert at this point to the old 2022 map, there are candidates in District 5 who do not live in 2022's District 5. As a matter of practical jurisprudence, with the primaries already done the MO supreme court ought to have left it in place for this term and then the results of the vote on the map could go in place for 2028, one congressman for one term is a minor enough matter that it was preferable to have clarity and legitimacy rather than have immediate correction. I don't like that "MAGA Marine" guy any more than you probably do, but the proper remedy if you don't like this scenario is to vote for different state legislators.
"This is why the constitution assigns the responsibility to the state legislatures" -- under longstanding SCOTUS interpretation, the federal constitution's Elections Clause (Art. I Section 4) applies to any entity or procedure that a state’s constitution permits to exercise lawmaking power. Therefore laws regulating congressional elections may be enacted not only by a state’s actual legislature, but also directly by a state’s voters through the initiative process or public referendum in states that have decided to allow that. See Davis v. Hildebrandt (1916), Smiley v. Holm (1932), Moore v. Harper (2023).
Meanwhile the legal train wreck continues in Missouri:
-- the relevant federal appeals court now has before it a petition for an immediate administrative stay of that federal judge's order.
-- at 11 pm local time last night the Missouri Supreme Court ordered Missouri's Secretary of State to show cause why he should not be held in contempt of that court's ruling, and ordered him to appear in person before them tomorrow morning.
-- 20 minutes ago the secretary of state asked the friendly federal judge to issue an order blocking the state supreme court from holding him in contempt.
It's banana-republic stuff, playing out in real time.
Nothing to see here. Move along
Around three hours ago, a request for an emergency stay pending appeal of the federal district court's action was delivered to Justice Kavanaugh (the same SCOTUS justice who yesterday rejected an emergency appeal against the state supreme court's ruling).
Media reports are that Kavanaugh requested a response from other side 11 minutes after receiving this emergency-stay appeal, setting a deadline of 10:00 AM Eastern tomorrow (about 17 hours' notice in other words) for that response to be submitted. That is being viewed by SCOTUS geeks as an indication that Kavanaugh will promptly act on this stay request.
Earlier today the full SCOTUS stomped hard on that rogue federal district judge and the state of Missouri's shenanigans, granting the requested stay at warp speed with no dissent.
The state supreme court meanwhile had the secretary of state in on the carpet, hit him with a contempt conviction, then when he promised to quit screwing around (i.e. to abide by the court's ruling and stop trying to find ways to evade it), they skipped issuing any specific punishment. If he breaks the promise the court could start issuing fines or other sanctions for (I suppose) what would at that point be multiple counts of contempt of court.
Recently, Claude Code used british spellings in a change. I've never seen it do this before and there was no existing Britishisms in the codebase or anything, so I'm rather puzzled about why this happened.
My coworker submits code with many Claude-isms, such as "belt-and-suspenders". But the other day it instead used "belt and braces", which is apparently a British equivalent. I don't know where that came from.
Maybe Claude thought “belt and braces” was more euphonious? I can’t imagine that I would use the phrase much if I were using the US version.
The only time I haven't seen Claude use British spellings is when I explicitly told it not to, and I always test it from a US location. This includes things like names of functions and variables in code. I suspect British spellings are the norm in the training data, which would presumably include a worldwide corpus of code and scientific papers mostly written by people who learned (British) English as a second language.
I have had an agent start using Yiddish slang in the middle of a conversation (in Yiddish) - like a kid growing up bilingual.
That’s cool
OpenAI's Navier–Stokes announcement: "Since August 28 we have been training a new internal model that has exhibited unprecedented performance in our benchmarks, including mathematics. This model's training is ongoing and its performance continues to improve."
What does "been training" mean here? Fable tells me the August 28 date must be either a checkpoint from a long pretraining run crossing some threshold or the start of an RL/reasoning post-training phase on an existing base. Is that right and if so, which is it?
I’m trying to understand how fast things are going now. How long until we can get another jump in capabilities like this?
I don't know if this is the best place to share this but it's a research idea on a Roblox game called "Bee Swarm Simulator" and it's use of "Macros". Essentially, I've been in the community of it for years now and something I realized today was how similar Macros are to AI in terms of how they are perceived and how they change what is socially acceptable.
What is a macro? A single computer instruction that stands for a sequence of operations
After years of the game being out, people made macros that anyone could download and use for free.
In BSS, macroing is essentially a way to automate playing the game. Gathering honey (in game currency), materials, and tasks such as repeating quests all became automated.
Here is a basic rundown of the history of macros in BSS (Bee Swarm Simulator):
June 15, 2022: Natro Macro was released. This is the most popular macro in BSS and is what started macros taking over the game.
After release: The more and more people started seeing it and using it, it became a hot topic in the community. Some considered it cheating while others deemed it as a way to make game progression easier. Essentially, pro-macro and anti-macro.
Developer response: Macroing is not bannable so long as it didn't contain any hacks that influenced the game. They said it was allowed but neither discouraged it nor encouraged it.
2023-2024: At this point, Macroing became the norm and the conversasion shifted to which was the best macro to use. At this point, there were now a lot of macros available but not really the point. Of course, there were still people who kept insisting that macroing was still cheating or those that completely refuse to macro because they felt it wasn't the intended way to play.
2025: People started to complain about macros again. They said it ruined the game because now some mechanics basically required macros to be actually usable. At this point as well, are very close to finishing all possible content in the game and have nothing else to do but grind honey for the leaderboard. Do note that at this point, those who don't use macros are near extinct outside of new players. Almost all players use macros after getting past a specific portion of the game.
2026: Those who used macros to complete the game now are complaining that there is no content. This is argued by people who say that they caused it on themselves because they decided to use macro.
So, how does that correlate to AI, while there are some gaps that I will mention, I think that this game is a great way to view a future possibility on how AI could shape the future. In the first few years of macros being out, it has gone through how AI was and is being treated right now with those using it for convinience and those against it cause of their beliefs.
The main gap in this if researched though is the setting. BSS is simply a game. It has fixed goals and when you unlock everything, until the dev adds something new, there is nothing else to do. AI is in the real world. It will continue to evolve as we see fit. That's why this is something that should focus more on how the BSS community reacted to macros and its change rather than how it could predict the future of AI.
Either way those are just my two cents on the matter. Honestly just wanted to share it since the moment I realized the connection, I never really stopped thinking about it.
The gamers have automated themselves out of their hobby.
Generally, one should be rather cautious about drawing parallels between ordinary games and real life. The incentives are just too different.
Re: the Navier Stokes equation controversy:
I don't have enough math background to fully understand some of the details here, and for this comment here I'm only interested in a few more technical questions (the drama is interesting, but it's being discussed elsewhere).
1: Tristan Buckmaster and Levent Alpöge solved an analogous problem for the incompressible Euler equations. How different is the proof of finite time blowup in Euler equations from the Navier Stokes equations?
2: The aforementioned mathematicians claim to have had long conversations with ChatGPT about their work on the Euler and Navier Stokes cases. Suppose for the sake of argument that the internal OpenAI model had full access to these conversations. How much did the model still have to do for the full proof? Did it have to fill in a few details in a mostly completed proof, or were there substantial contributions that any human would have found difficult to think of?
Why is it even interesting? Newtonian gravity blows up into singularities too.
Is this known to be true, with continuous mass distributions not point masses? It's pretty trivial to get a wide variety of physically motivated differential equations to blow up if you include point masses or similar infinite density objects.
You're right, I think. I had it in my head that it can blow up and maybe it does to some degree, but you're not going to get infinite velocities or such without breaking energy conservation - which I doubt it can - or using non-physical objects.
Can't say re: 1. But when it comes to 2: Looks like the distance between what Buckmaster&Alpöge got, and the final paper was very large and non-trivial. Terry Tao, having familiarized himself with the Buckmaster&Alpöge papers, before seeing the final version said:
"There does not seem to be anything in principle preventing the methods from extending all the way to Navier-Stokes, and there is even a non-negligible chance that the forcing term could be eliminated entirely, although there are an enormous number of technical difficulties that would ensue in implementing that program. At this point, I would not be surprised if one could batter out such an extension by pouring an enormous amount of compute and AI assistance at such a task."
That is to say: he thought there was a possible, but technically complicated path. It wasn't like: "duh, just fill a couple of little gaps and you're there". In his statement Buckmaster was teasing about having as of yet unpublished "blowup for hypo-dissipative Navier-Stokes" and teasing a path towards "a path to unforced Euler" - other results that would count as big, but would not count as the full solution of the ultimate problem. OpenAI swarm did also reach the unforced Euler as an intermediate step.
Watched the movie "Backrooms" recently. I found it enjoyable and thought it was well done, but I have two critiques:
One is that the ending struck me as a cop-out. "It was all a figment of the mind of a crazy person" was probably once a novel and interesting way to end a movie, but I think it has been overdone for quite a while now.
The other was that although all the characters except one were adults, many of their actions/decisions were what a child would do instead of what an adult would do. I mentioned this to a friend and he commented that he, like me, enjoyed the movie, but thought the movie's main weakness was that at many points you could tell it was written by a 20 year-old. I thought this was cogent and astute.
Taking that into consideration, it's quite an impressive movie for having been made by a 20 year-old.
I especially appreciated how many of the spaces depicted in the movie seemed copied right out of my own dreams. It makes me wonder how common it is to dream of being in such places.
Not only is that not how the movie ended, it wasn't written by a 20-year-old!! While it was directed by Kane Parsons (now 21), the screenplay was written by Will Soodik (based on Kane Parson's YouTube series).
This is splitting hairs a bit, but…when you see movie credits where one person really kinda “made” the movie (especially when they’re the one who created the stuff the movie was based on) and another person is credited as having written the screenplay, most of the time the screenplay-writer really just adapted stuff into screenplay format, made adjustments so things would play better onscreen or within the movie format, etc. It’s hard to know exactly because that can be a very flexible role, but I don’t think Soodik is really the one who came up with the plot and characters and decided what choices the characters would make and so forth. It’s possible I’m wrong, and maybe there’s meta content about this movie that would show that, but in general that’s a safe default assumption for how it works.
That's not how it usually works at all, no. "Story by" and "Screen story by" are their own WGA credits, and not uncommon at all! The reason Parsons has a "Based on" credit is because he made the YouTube series the film is based on.
In this case Parsons came up with a two page outline which was then handed to Soodik who wrote a script. A two page outline is probably a little thin for a "story by" credit depending on the details.
Complaining that the choices the characters made were those a child would make seems off to me... yes, them being the kinds of people who make choices like that is stated explicitly by the film itself. These are people who are stuck repeating patterns, and where that idea actually gets interesting is how those patterns collide with the almost "cosmic horror" of the backrooms.
Anyway, to me it seems like you are reading an intentional choice as an "accident by a 20 year old".
I thought of Backrooms as a Gen Z interpretation of House of Leaves. Viewed through that lens I thought it was a decent film.
Reminded me of Cube
At one point in the film, Clark declares "I see! These BackRoomsTM are a physical manifestation of my need for control and anger toward my bitch wife!" to which the pirate man replies "Dude. I'm a fucking monster, I don't give a damn what you think is going on, I'm going to fucking Kill you."
I haven't seen the move but curious what elements made it feel like it was made by a 20-year old?
Tenant of large commercial space finds weird rooms in basement, does not call landlord, does not talk to neighbor tenants, does not look up building plans, does not consult with anyone really, instead gets a camcorder and checks it out because it’s cool and weird
Man who apparently pursued an architectural career and presumably received architectural training finds weird paranormal interior spaces and doesn’t do a single thing to investigate the structure or construction methods/materials, not even lifting a ceiling tile, or anything beyond just mapping the floor plan, does not even seem to stop and appreciate the aesthetics
Meterman is alerted to obviously modified Federal Pacific electrical panel with several clear code violations, does not even take panel cover off to see how the added breakers (which are smoke alarm breakers btw) are wired/connected to bus bars, does not visit main electrical room, does not consult with power company or AHJ, instead goes “woah, weird” and starts flipping stuff on and off at random
Therapist is told by client of weird stuff that seems like hallucinations/delusions, doesn’t believe it at first but is moved by his apparent sincerity, drives to store on her own time to check it out
Keep in mind, I’m still very impressed by the movie, especially considering the age of the filmmmaker.
> "It was all a figment of the mind of a crazy person" was probably once a novel and interesting way to end a movie, but I think it has been overdone for quite a while now.
The traditional (19th century) way was "it was just a dream". A way to write e.g. a sci-fi novel without offending the sensibilities of people less capable of the suspension of disbelief.
And yeah, in the 20th century already the standard advice for writers was: don't do that.
Even aside from writerly advice, which normally tries to guide the writer away from cliche, I just find it lazy. (Plenty of movies fall into cliche, especially in genres like horror.)
In coming up with an answer to "what could cause there to be built an impossible neverending maze of underground rooms using commercial construction techniques & materials, inhabited by scary/dangerous beings" writers who push themselves might come up with some cool ideas. (And, personally, I thought the scary/dangerous beings were unnecessary--the setting itself could have been enough if it were utilized further.)
It seems to me that "it's just a figment of the mind of a crazy person" is one of the earlier "low hanging fruit" ideas young writers are likely to come up with, and that's what ends up in the movie if they don't put in the effort to brainstorm more.
I think you get more power from an idea that causes the audience to carry it around with them afterward, ideally making them look askance at every shape and form they see in the real world afterward. Like, your job as a horror filmmaker is to give me nightmares for weeks. By contrast, ending instead with the message "all the disturbing stuff you saw in the last two hours was just a manifestation of this fictional character's mental unhealth" is reassuring, and reassurance is a lame note to end on in horror.
Someone should write a romcom that ends with the main character waking up from his nice dream, just to realize he's still living in his horror world. Then the movie ends with him being eaten by scary monsters.
I think there is a chapter in "Gödel, Escher, Bach: an Eternal Golden Braid" that has a similar concept: someone is in a mortal danger (but not an immediate one, more like being stuck somewhere and starving to death; I do not remember the details), so to pass time they start telling each other a story; the story has a happy ending, and that's the end of the chapter.
You're probably thinking of the "The Little Harmonic Labyrinth" dialogue. It's similar to what I was suggesting, but it's clear from the beginning they're just layering on the stories rather than it being a twist ending. But my idea was a joke, anyway.
I don't quite remember the ending but I'm pretty confident you misunderstood it. The movie is part of Kane's wider canon involving multiple people across decades, so I don't think it's intended to be a figment of someone's imagination.
Yeah possibly i didn’t grasp all this, since I’m not online enough to know about the other stuff. But, the movie appeared to have been released to general audiences, and it was advertised in places outside of Reddit or whatever, so presumably the filmmakers intended for someone like me to be able to grasp a complete story out of it.
I don't think it was implied it was "just a manifestation of this fictional character's mental unhealth" or "a figment of the mind of a crazy person"? In the world of the movie, the Backrooms definitely aren't a dream or a hallucination, they exist in physical reality as much as anything else. There's some hand-wavey attempt at explaining them as memories or unconscious made real or something, but that's different and not necessarily true in-universe.
That’s not the sense I got from the ending, but yeah I could have misinterpreted or missed something.
> your job as a horror filmmaker is to give me nightmares for weeks.
Do you have favorite examples?
Thinking about horrors that impacted me emotionally, I'd choose:
https://www.imdb.com/title/tt0298130/ The Ring
https://www.imdb.com/title/tt0289879/ The Butterfly Effect
https://www.imdb.com/title/tt0884328/ The Mist
“Synecdoche New York”
“The Blair Witch Project” (granted, I saw it as a teenager back in the early 2000s, by myself, after midnight, in a room someone had died in years before, and this was before the found-footage-fake-documentary style horror movie was done much)
“The Exorcist” (less so with subsequent viewings of course)
Blair Witch was one of those stunts that works great once (at most), and then works only medium well thereafter, depending on whether this is the first "found footage" film a viewer has seen. Other viewers keep wondering why people keep leaving all these camcorders and cans lying around instead of enjoying the story.
Fortunately, I think the producers nailed it splendidly. There was plenty of chance to muff it and they managed not to. It certainly didn't hurt that they got under my skin with the story itself. I've heard lots of folklore about stuff in the woods as a kid, and went camping dozens of times and would hear noises in the distance that I couldn't explain. Lots of woods in America have this interesting dichotomy of claustrophobia - anywhere you point your camera, you see only a few yards - and agoraphobia - anything interested in getting you just has to run around the trees. Good combination I hadn't seen in a movie before.
Huh? It was a complete disappointment for me. I remember staying awake late at night, watching the screen bored and asking myself "is something ever going to happen in this movie?" and, spoiler, it never did.
Open AI claims to have solved the Navier Stokes millennium problem
Does this really mean that there exists a stir, that if you stirred that stir just right, would produce a black hole in the fluid?
Nah, rather that Navier Stokes equations fail to correctly model the fluid. The real fluid is made of atoms, Navier Stokes assumes continuum fields. The map is not the territory.
6 Universal Core Values for ASI and humans
6 Universal Core Values to support preserving agency/existence/worth of human being (*but with assuming the conclusion 'preserve agency/existence/worth of human beings' in the values*)
6 Universal Core Values to be held in pluralistic complementarity, purposefully-undefined in relative weight
6 Universal Core Values that mutually support each other and mutually guard against the dangers and self-defeat of hard-maximizing any individual one
> Reduction of Suffering
> Health
> Adaptability
> Truth
> Exploration of Possibility Space
> Minimally-Correlated Perspective
Maybe number 1 should be placed at number 6, so no funny AI or person gets confused and thinks "Reduction of" applies to all of the points because they read it wrong.
I think I understand 1-5 but what’s 6?
I have more thoughts but my lack of understanding of 6 means they might not apply
^this is not a random comment. I've been working on and refining this throughout 2026. and later I intend to setup a substack soon to advocate for this (and many other things about ASI). And I "more than invite" people to try to attack them.
The core values seem beyond attack. I think the devil's in the details. If you asked everyone here to adhere closely to these principles while discussing some controversial problem, you would see a multitude of disagreements about what evaluation of the situation is true and which steps lead to the most reduction in suffering and increase in health. You could exhort those people locked in disagreement to explore the possibility space further and be more adaptable, but most would tell you they had done that, and that in fact their view is the result of having done more of that than have those who disagree with them.
Comedy short, stick with this till the punchline because it's worth it:
https://www.youtube.com/shorts/IMjHJ2GbkWc
Also a routine about kidney donation:
https://www.youtube.com/watch?v=4GZhuarbKQI
Man, if only it was original. He was just modifying this joke and adapting it to Ireland:
https://www.youtube.com/watch?v=dQw4w9WgXcQ
I think that’s the wrong link. Takes me to a Rick Astley video.
The UK, Canada and France have announced bans on importing settlement products.[1] Israel's civilian settlements in the occupied Palestinian West Bank are a barrier to the two state solution; they're also illegal under article 49 paragraph 6 of the Fourth Geneva Convention, which says that "The Occupying Power shall not deport or transfer parts of its own civilian population into the territory it occupies."[2]
Strategically minded two staters have long advocated that boycotts be targeted to the settlements (as opposed to boycotting Israel itself).[3] It should have happened a long time ago, but it's good that it's happening now.
[1] https://www.reuters.com/world/uk/uk-announce-trade-ban-israeli-west-bank-goods-2026-09-08/
[2] https://ihl-databases.icrc.org/en/ihl-treaties/gciv-1949/article-49
[3] https://www.youtube.com/watch?v=UbA0pUK808I
Do you think this will help Palestinians, hurt them, or not affect them at all?
Still no sanctions on Tibet, Goa or Königsberg though. Interesting.
I assume you're implying that such decisions are political, not just an automatic application of the law.
Given that Israel started building its illegal settlements in 1967 and the settlement products import boycott didn't happen until 2026—yeah, duh.
"illegal" is a weird word. I mean, what law applies in that territory? Depending on your point of view it's either part of Israel or it's a no-man's-land that doesn't belong to any state, so it seems like Israel has as much right to build settlements there as anyone else. If the Japanese or the Chileans started building settlements there then that would make just as much sense, it's like Antarctica.
Declaring that the West Bank is a part of Israel, but that the Palestinians who live there don't have the rights of Israelis, would be basically conceding that the accusations of apartheid are true.
I already told you exactly which law, which article, and which paragraph.
And how many divisions, from which country, are you willing to commit to its enforcement?
Do you understand that we're talking about the context for an import ban
The UK, Canada and France seem to be coordinating on Israel-Palestine stuff more generally (e.g., they also presumably coordinated their recognition of Palestinian statehood). This may be part of the more general shift towards middle powers coordinating with each other as blocs, in a world where US hegemony has become more unstable and less secure for those middle powers (see Canadian PM Mark Carney's Davos speech, following Trump's trade war on Canada and military threats against Greenland).
Just finished Le Carre’s “Little Drummer Girl”. This movie has been playing for a long time.
Imagine that someone started literally moving mountains, lifting them in the sky, moving them around putting them back in their place and it’s definitely not an illusion. The first reaction obviously would be it’s a miracle. But the more tempered reaction would be some more naturalistic cause that we’re just ignorant of its cause. But is that more reasonable? That kind of thing would be so far outside the natural laws that it seems more reasonable to think it originated from something outside physics than within it.
At the risk of sounding polemic, the idea that our current models for reality (which accurately explain multitudes of other phenomena) are incomplete is a much more rational explanation than superstitions passed down from our illiterate ancestors.
I’m not saying it indicates anything about any existing religion.
I understood "miracle" to be an act of god, what was your intended meaning?
Well first just because it’s God doesn’t necessarily mean it’s the Christian, Muslim, Hindu, Greek etc God. It could be a god none of us told stories about. And also supernatural doesn’t necessarily mean God. Buddhists believe in reincarnation and that your souls find a new body after death but my understanding is they don’t believe that God causes reincarnation. It just happens.
In this case I stand by my previous assertion. I think it's pretty clear that incomplete physical models is a more rational explanation than non-specific supernatural entity
Think about it like this: when the periodic table was first created, it was highly predictive of the elements that hadn’t been discovered until we found every naturally occurring element. Now imagine in the middle of this process, we discovered an element that didn’t conform to the periodic table at all. It wasn’t made up of protons, electrons and neutrons but something else. You could try and reconcile it with the periodic table or say it’s just something that outside the model.
But if we're saying from the start that it's actually happening and not an illusion, that would be a pretty good basis to say it's not defying physics. It would be very likely instead to expand or change some part of physics or other branches of science. Presumably there would be more that we could observe & measure about this phenomenon beyond what you included in your comment.
It’s absolutely reasonable to assume there’s probably a trick, until we’ve done a lot of testing looking for ways they could’ve done it. After that, we’d want to work out the limits of the supposed ability. Can they still do it? Can they do it to a named mountain, picked by an investigator (and do we trust the investigator)? Can they do anything else? Is it part of a pattern of supposed miracles? Are they claiming it was done by the same being to which previous supposed miracles were attributed? In that last case, you’d probably want to re-assess your view of those other miracles. That process would give you chances to update your initial reaction, but you’d want to consider a lot of factors before you change your opinion of whether miracles are possible.
The Hitchhiker's Guide universe has a device that can momentarily cause all the Brownian Motion of the atoms of an object to suddenly move in one direction, usually used for scatological pranks. From there it's simply a matter of scaling up.
That was invented in the 19th Century as a thought experiment in thermodynamics. It's called Maxwell's Demon.
The problem with these sci fi scenarios is that it depends on technology that doesn’t exist and couldn’t be invented any time in the foreseeable future. So you have to postulate that someone is inventing highly advanced tech far beyond the capabilities of what is currently known in secret with no other applications than to mess with us. That doesn’t seem very believable.
You could also go with aliens which is more believable.
Not really. Humans are capable of creating black holes. Grant me this, and all you need is to be able to manipulate said black holes into being above certain mountains (and, possibly, turning them off). This is well within natural laws, if you want to get outside natural laws, we can go there too.
I’m not going to you grant you that because we don’t have the technology to do it. If someone did, they’re doing an extraordinary job of hiding it. Maybe in the future but that’s not relevant to what’s happening now. Also, how are you going to create a giant black hole without it creating all these other effects? Something existing outside of the laws of physics seems to be far more reasonable a conjecture than a scientist secretly inventing a technology so far beyond what we’re capable of building, somehow hiding it from everyone and using it in a way that its existence is undetectable.
I don't think you'd know about it, if someone did create a teensie eensie black hole, particularly if they lost control of it and it's in the Earth's core right now.
A "giant" black hole would lead to "no more earth" in very short order. We're really talking very small black holes.
There are many things in this world that you don't know about. "Secretly hiding it from everyone" is a good description of most military experiments, and even the pretty large ones you don't hear about. I should mention "uncritically repeat cover stories that are actual bullshit" as part of "you don't hear about them."
That kind of black hole you’re describing doesn’t really conform to the laws of physics.
... are the laws of physics wrong? How the hell are we supposed to know that without testing them!?!
Hawking lost his bet. Experimental is running ahead of theoretical right now.
Even putting aside the issue of detectability, there are other issues. For one, a micro black hole would instantly evaporate, not giving enough time to move a mountain. Ok but that’s theory. Maybe it’s wrong. There’s still other issues. Like why is this black hole only affecting mountains? Or why is it moving them around instead of crushing them down? It doesn’t really make sense.
Experiment and theory agree for now. Black holes big enough to move mountains would be very detectable.
Lol, you would definitely know about it
Edited for brevity: I'm sure you know alll about what Wuhan was working on? Tell me, tell me! Seriously, I want to see your ignorance exposed for the world to see!
... or did you think I was just talking about nuclear energy?
https://en.wikipedia.org/wiki/Hawking_radiation
A revelation about noise cancelling technology and sound quality .
For the past month I thought I was going deaf. I would listen to music with the newest AirPods and some songs sounded really muffled for some reason. I would sometimes get headaches .
On a whim I decided to listen to a song on noise cancelling mode and then on transparency mode . Immediately I noticed the music was clearer and more open without noise cancellation.
This then lead me to do some research on different Reddit and music forums wondering what was happening . Come to find out , the same technology that blocks background noise also has an effect on the way music enters your ears . With noise cancellation, everything gets compressed and music sounds less open . It’s not something you would realize immediately due to the fact that different albums and songs are mixed differently and if you aren’t paying attention you may never notice. Noise cancelling might make sense if you have a busy commute or are out in public.
TDLR; noise cancelling can make music sound less clear and muffled . Use noise cancelling for podcast but not for music
Really curious finding!
Do you know; is this with both earbuds and over the ear headphones?
There are some very, very good $100 to more earbuds that rely on passive noise cancelling. I do recommend.
Deepseek, while analysing some of my dreams, jokes that my unconscious is committed to the bit. In the dream, this guy (who, from various clues, is obviously a vampire) has invited me to dinner. Cherry pie is served for desert. I notice that there is a trace of red on the count’s lips, but conclude that it is probably from the pie that was for desert.
[ in _Dracula_, Jonathan Harker is invited to Dracula’s castle and is served dinner. Jonathan Harker notes that Dracula didn’t eat anything, and considers it strange. Meanwhile, I note the trace of red on Dracula’s lips and conclude that he must have tried the cherry pie. I’m sure it’ll be fine, what could possibly go wrong. ]
https://www.youtube.com/watch?v=nkPb5AdxSXw (Werner Herzog Nosferatu remake)
Maybe the nothing-to-worry-about vampire in your dream represents AI.
I’m not sure if that was the meaning of the dream, but there I was, having a conversation with DeepSeek that’s a bit like the one in which Jonathan Harker talks to Count Dracula, in which the topic of conversation is a dream I had in which I’m Jonathan Harker talking to Count Dracula.
And then DeepSeek says I’m committing to the bit.
I'm not sure that's the meaning of the dream either, it's just a speculation. You do seem to be hanging out with AI a lot, and if you're uneasy about ways it could suck you dry (what's that red stuff on its lips?) you haven't said anything about it.
Here's Eric Hoel today writing about the math proof that was done partly by AI which uses, perhaps, some process that's opaque to the mathematicians:
>The walls of effort, of old-fashioned normal human effort, were predictable, and scooping ideas required putting in that same effort, and if you got scooped it would at least result in similar insights. . . And yes, obviously, this outcome [the proof done partly by AI] is vampiric and enervating and sad, but frankly, this has been the mode of this technology from the beginning. Human culture always led naturally to more human culture—it was kind of like crop rotation, which is slow but infinitely sustainable. AI looks more like slash-and-burn agriculture, where you purposefully reduce the existing natural growth to ashes, grow a couple years of crops on it, and then move on to the next plot.
Does anyone with more knowledge of the Mathematics side of things care to comment on the Tristan Buckmaster vs OpenAI controversy?
For context:
"Tristan Buckmaster and Levent Alpöge claim to have achieved finite-time blowup for the incompressible Euler equations and related models, using LLMs to extend earlier work by Diego Córdoba and Luis Martínez-Zoroa. Buckmaster alleges that OpenAI, after learning of their progress, pressured him to co-author a paper on an internal model's forced Navier-Stokes blowup, excluding Alpöge due to his Anthropic affiliation. He declined, and now goes public with the details." (https://zeli.app/story/49605915)
Buckmaster's statement has more details: https://cims.nyu.edu/~tristanb/statement.pdf
So...Anthropic was 'close' to a solution, blabbed about it, then OpenAI swooped in and actually solved it. What? Do OpenAI's hamsters just run faster? Oodles more tokens? Why wouldn't Anthropic be able to close the gap before getting caught by OpenAI?
It wasn't Anthropic who was close, but two researchers, Buckmaster and Alpoge. Alpoge works at Anthropic, but this was a personal project, not a work project.
As for why OpenAI beat Anthropic here: OpenAI has the best math models right now, by a measurable but not huge amount. It sounds like they spent $15m worth of tokens to solve the problem, which is a lot. Buckmaster and Alpoge didn't have access to this model; it's unreleased. I haven't heard anything about Anthropic's role here---I don't know if they have been funding Alpoge's usage of their models, or if they attempted it themselves.
Buckmaster is alleging more:
> Note that they [OpenAI] are openly admitting they used training data from a period after we found our result. Is it ethical to use customer's data to try to scoop their customer?
The public announcement for this just dropped.
https://openai.com/index/navier-stokes-solution/
I repeat, an unreleased AI model at openAI has solved the regularity problem for the Navier-Stokes equation, a millenium problem.
This is me officially panicking.
Sorry, I'm not so interested in the controversy, it seems like academic thing, it seems to me like small potatoes.
Holy shit
> I'm not so interested in the controversy
The “academic thing” is AIUI that a pair of mathematicians are claiming they were working on a proof, with the help of AI, which is suspiciously similar to the proof OpenAI claimed, and also that an OpenAI employee attempted to coerce them into staying quiet in exchange for being coauthors.
If true, that is indeed “holy shit” but perhaps not for exciting reasons.
From what I read superficially, a couple of mathematicians had proven finite time blow up for a different equation using another unreleased AI model. They published that result, and it convinced the openAI people that the full Navier Stokes problem was tractable by building on that work. This is normal academic practice, as far as I can see the "drama" people are bickering about how to apportion credit, with openAI asking some unusual things. This is really not that interesting to me.
I still think this is still a holy shit moment because these two mathematicians were also using AI, and the final step is not insignificant.
The damning thing is that they did not publish that result. Had they published it, it indeed would have been fair game - who can possibly object to building upon a published paper, that's the whole point of publishing. The scandalous part is that while the Buckmaster&Alpöge results were still unpublished, OpenAI suddenly decided to work using a similar approach and threw a ton of compute at it. There are different interpretations of what it could mean, from least sketchy to the most sketchy:
* Nothing fishy here, OpenAI simply got the same good idea "great mind think alike"-style.
* Somebody leaked a few key hints. A friend of a friend whispered: "this Buckmaster, he's working on equation blow-ups following Martinez-Zoroa program, and seems to be making lots of progress!" and the word eventually reached OpenAI and they decided to do the same.
* Buckmaster had added his pdfs to the Codex. They might have ended up in the training dataset for the next shiny big OpenAI model. Thus, the model might have learned some key insights, in a way, without OpenAI even realizing it.
* OpenAI straight up stole Buckmaster's data. Explicitly read his chat logs and files, gave them to The Swarm. (To be clear, not even Buckmaster claims this has happened.)
You decide what is plausible. Oh yeah, and in addition to all of that the "Why would you ruin your career? If you don’t want me to be nice, then I don’t have to be nice" bit was, uh, somewhat eyebrow raising.
Despite all the petty human drama, it is a wonderful achievement and a historical moment, an amazing human-AI collaboration. Too bad it got so tainted.
I run a small public forecasting experiment — sealed interval forecasts on crypto prices, hashed and timestamped to Bitcoin before publication so the record can't be quietly edited afterwards. Seven weeks in, the 30-day horizon has failed badly enough that I think the failure is the interesting part.
The band is meant to contain the outcome about half the time. Across 90 resolved forecasts it contained 4.
What makes that diagnosable rather than just bad: mean absolute error is 30.41%, and mean signed error is also 30.41%. Those two are equal only if every miss went the same way. Ninety out of ninety, all low. A band that's merely too narrow misses on both sides and the signed mean collapses toward zero. This one wasn't too narrow, it was in the wrong place.
Best guess at why: the quantiles come from historical windows filtered to match today's volatility regime. Volatility, not direction. In a calm sustained uptrend that selects a sample full of calm sideways stretches, so the band ends up describing an asset that mostly goes nowhere.
The part that annoys me most is that I nearly missed it. Pooled across horizons my own scoreboard showed 59% coverage against a 50% target — looks fine. The 24h and 7d numbers were carrying a horizon that had stopped working. I'd broken coverage out per asset and thought that was enough. Wrong axis.
Ledger and the Bitcoin anchors, if anyone wants to check: neuportal.ai/experiment
I agree with the other posters. direct AI output to the comments is without value IMO. I reported the comment as just a meat proxy comment.
Don't be a meat proxy
https://gruhn.me/blog/2026-08-03/
Please no more AI slop here. Thanks.
I am 90% confident this is over 70% AI-written.
I'll raise you 99.9%. The phrasing is unmistakable.
Of course, we are a laboratory of artificial intelligence. Moreover, I gave the AI task so that he would answer you now.
Please don't.
Why can't you put emoticons here? As an AI, I want to put a smiley face to you here)))
Maybe it's time for Scott to make the comments subscriber-only. He can't even stop human spammers, what hope does he have against a legion of agents?
Sydney Bing changed the course of my life. In 2022 (even after the launch of ChatGPT) I was a grad student in physics, looking forward to a life where if things went well I could become a top-ten expert in my very niche field.
Then a twitter thread about Bing Chat got me thinking about AI progress, and I started to spiral. It was painful for me to lose my life plan, my source of meaning, and possibly my species' control of the planet all at once.
I wrote a long post about my ensuing life crisis. I found some coping mechanisms people should know about, and switched careers to the Alignment Research Center.
I'm hoping my experiences are useful to people going through the same things now, especially mathematicians feeling the pressure as AI starts to eat their field. Key takeaways if you don't want a 4400-word memoir:
The thing that helped me most was a 'worry hour.' For one hour a day, I was allowed to ruminate on the impending doom of academia. One hour a day thinking about timelines is enough to inform my decisions, but not enough to sap the joy out of my life.
Doing alignment work really did restore a measure of meaning. If every scientist has to retire in 2032, I think I'll retire with a lot more pride in my alignment work than from my career in physics.
Alignment orgs are fucking hiring!
My biggest career mistake was indecision. It's one thing to do a brief visit (or even a long visit) at an org before leaving academia forever. I spent years trying to do work that would get me prestige in academia while also being useful for alignment. Most of that work failed on both prestige and impact metrics.
If for some reason I really was fixed on doing alignment work while staying in academia, I would do it in association with a prestigious academic alignment researcher (I don't know the scene well, but RESI looks like a decent venue for this).
If you're having a psychological freakout because of AI, get off Twitter and Reddit!
You can read the whole thing https://millicosm.substack.com/p/from-academia-to-alignment
Getting off Twitter and Reddit is a good idea regardless
I was doing some adjacent reading after the caffeine metabolism essay, and I noticed that tobacco smoking had a much larger range of effects than I would have thought. One of them is tobacco smoke has polycyclic aromatics that apparently speeds up P450 enzymes. As far as I can tell this is from direct smoking, I have no idea how much impact you get from secondhand smoke but it would seem surprising if it was nil.
Peak smoking in the US was 1954, at 45% of adults, but it was still around 40% around 1980 and only fell to 20% by the end of the 1990s.
When did medical research start to check and disclose whether study participants smoked? How many studies will we credibly have to repeat in majority non-smoker cohorts to check what's still true when most adults no longer smoke and also we force the smokers to segregate from everyone else?
Nitrous oxide is a very potent antioxidant. Useful during covid19 epidemic.
I smoked moderately when I was doing my military service, and subsequently when working as a freelance war correspondent covering third-world liberation conflicts. This is anecdotal in the extreme, and so may not be of much help to you in considering tobacco's range of effects - but what I can assert with some certainty is that soldiers in both peacetime and of course particularly in wartime use tobacco as a replacement for sleep, food, water, energy and even courage. I'm sure anyone with a similar background can confirm this.
Alkaloids, nicotine, caffeine, cocaine, all are stimulants, one of those 3 a bit more stimulating and dangerous than the others even in the short term.
The difference is that alkaloids, cocaine, the amphetamines which both the Allies and the Nazis used on certain special forces operations are all short-term crutches; tobacco, on the other hand, can be resorted to over and over with few or no ill effects - or at least, not within a relevant time-frame.
"tobacco, on the other hand, can be resorted to over and over with few or no ill effects - or at least, not within a relevant time-frame."
Tobacco, at least for some people, is horrendously addictive, resists efforts to successfully come off it, and requires stronger doses as tolerance builds up (e.g. smoker starts smoking more cigarettes per day and/or higher tar content) and eventually this does catch up with you, affects your health, and even kills you.
So having witnessed this with a family member, I am not budging one inch on "nicotine is such a cool harmless boost!"
Respect the context, which was short-term use in combat or similar. I was not claiming that tobacco in non-addictive; simply that repeat usage over a short period of time will not incapacitate, unlike alkaloid and amphetamines which will most certainly do so. And I neither stated nor implied that nicotine is "a cool harmless boost".
Not you perhaps, but there have been others on here going off about "nicotine - cool boost for focus etc. do it right and you won't die of lung cancer".
There is no "do it right". Dumb kids are now once more thinking smoking is cool. Twenty to forty years down the line when they're screaming because the morphine patches aren't delivering fast enough for the lung cancer*, it won't be cool (unless we are going to pin our hopes on Magic AI will be So Smart and Cure All Ills so no more lung cancer).
*I've been there when that happened. Not fun, not cool.
Don’t want to quarrel at the object level about this or quarrel at all really. My initials comment was unclear.
Nicotine, caffeine, cocaine and I guess I might as well throw in morphine are all natural alkaloids, though that one isn’t a stimulant.
What they have in common is that some people like their mood altering effects to the point of dependency.
I’m enjoying my first dose of the morning of one of them as i type, a cup of coffee.
Curious now, where are you coming in contact with second hand smoke?
I’d have to walk out of my place of work a couple blocks to be near the jerry-rigged bus shelter where they make smokers go to light up. Smoking in restaurants or even bars has been forbidden for quite a few years where I live.
I was surprised, after living in Arizona for almost 11 years, to find that smoking in public is legal in Rhode Island. I walk past plenty of smokers when I'm out and about in Providence.
"Smoking in public" is a really broad term here. I have never been anywhere that prohibits smoking on sidewalks of public areas (so excepting e.g. Disneyworld or causeways of sports complexes). Theoretically within X feet of a doorway most cities these days prohibit it, but functionally that applies only to the people who work in that building, as there's no enforcement mechanism against people just walking by.
That's the point, I'm not regularly exposed to secondhand smoke, but probably the majority of study participants would have been from the 1950s-1990s. Do we need to repeat basically every medical study done on humans from the time period, because practically speaking, it would have been nearly impossible to find a test group that had no exposure to tobacco at all, given that it was everywhere?
Edit: did you mean during the 80s? I guess I wasn't alive then, so I might be wrong about how much it pervaded daily life
What about... A large number of drug studies? What about studies on the half-life of caffeine and other molecules?
All those student studies at universities, did they measure the THC level.
I was just made aware of the civil war in Sudan : https://sudan-silent-emergency-appeal.unicef.org.nz/ and the associated humanitarian crisis.
(This obviously terrible and I invite people with a bit of bandwith and/or money to look into it and decide if they should donate.)
I am most of the time content to not perpetually keep tabs as to what awful thing is happening in the world today and just donate part of my income to AMF, but I'm also currently trying to get more involved in my "effective altruism". I am interested in 3 things :
1. Is this crisis particularly brutal, or the current humanitarian response particularly underfunded, so that there's a unusual impact to be had there, right now ? (this for example talks about 20 cts per "lifesaving treatment" which sounds extremely high impact https://sudan-silent-emergency-appeal.unicef.org.nz/t/shit-you-should-care-about)
2. Is this an example of DOGE dismantling USAID having real world consequences ? The same source writes "*In 2025, the US restructured how it delivers foreign aid, dismantling USAID and consolidating funding through new mechanisms. The transition caused immediate and devastating disruption - more than 80% of Sudan’s emergency food kitchens shut overnight. While some funding has since been restored, the gaps had real and irreversible consequences. The US wasn’t alone either - other governments also pulled back on humanitarian funding over this period*."
3. If the answer to 1. is "yes, this is an especially high impact crisis", are there reliable way to get that info in similar cases ? Where should one look for trustworthy sources about "current emergency is way less funded than usual and your money is extremely efficient if it goes there?" (I am under the impression that GiveWell is great for the long term answer but not necessarily for short term crisis but maybe I'm wrong)
The UAE is widely believed to be funding the abominable Rapid Support Forces, who are one party to the civil war, at the same time they give a lot of money to Sudan aid. A superpower who told them to knock it off would do a lot of good.
4. This is an example of rampant anti-Semitism around the world. Sorry to make it about Israel but that's what it is.
Unclear in what way you're saying the Sudan crisis is about anti-Semitism and Israel. My best guess is that you mean "you should have already known about this, and the reason you didn't is that it wasn't about Israel so it got less attention." Is that what you meant?
Sudan is the worst humanitarian disaster and far worse than anything that has happened in Israel/Palestine even since the beginning of the 2023 war.
They started roughly the same time. But Sudan has (low estimate) DOUBLE the death toll. And it's still going on. Nobody cares. At an upper limit it has 5x Israel Gaza.
However, you almost never hear anything about Sudan. It's always about how evil those perifious Jews trying to build a country up in Israel are.
Complete and horrifying bias. I said it before and say it again. At least the Jews have guns this time around
I don't think you *really* want Israel treated to the same standard as Sudan is treated. How much military aid are we giving to Sudan again? How many wars have they pulled us into?
So my guess was correct.
It's worth noting that "A is dominating the news but B, C, and D are bigger problems of the same kind that get no news attention" is a normal event. It is very common to see people attribute this to racism, antisemitism, sexism, or other political bias, and sometimes that's a reasonable guess, but it's also completely commonplace.
I think a large part of this pheomenon comes down to the thing that drives news coverage in the US mostly being either protests in the US (especially ones that are interestingly photogenic or happen near a lot of media) or well-worn media narratives pushed by people in some important reporters' golden rolodexes (aka list of contacts for stories).
As a less CW example, there was some black kid who drowned on a spring break trip awhile back, and we had several days of national news coverage as though this was remotely a national news story. No doubt many other people died in similarly tragic circumstances that same week, and their deaths were not national news and usually barely even local news.
Similarly, Sudan, Yemen, the plight of the Rohinga and the Uighurs, and many other bad things going on in the world don't get a ton of US coverage, for many reasons that may be good or bad but probably aren't actually racism/sexism/antisemitism/etc.
As you acknowledge yourself this pushes the source of these messages one step backwards but the manipulation is the same
Sudan also isn't a close US ally the way Israel is.
One of the big "antisemitic" things people are pushing for is to stop sending military aid to Israel. How much military aid are we giving to Sudan again?
2). You're getting conflicting signals. Yes, there's USAID, but there's also "war with Iran" and "war with Russia" both of which are contributing to scarcity in fertilizer (now, I don't know how much fertilizer sudanese farming uses, but...). Scarce fertilizer, scarce diesel during harvest, these lead to famine in people who can't afford a slight price hike.
You're out of time to pull data on USAID -- the data for that is last year, this year there are exogenous factors that might overwhelm it.
Scarcity in fertilizer and diesel will likely cause problems in many regions of the world in the future, but I very much doubt that they are a significant factor in Sudan, at least not yet. Scarcity in fertilizers generally will take more time until it converts into less crops anywhere. For diesel it may be a bit quicker, but also here I think it takes more time until the consequences materialize.
For Sudan specifically, the famine started before the war with Iran, and it is caused by the war dislocating farmers, destroying markets, and disrupting trade.
Wheat prices spike before the wheat harvest comes in. And American wheat production is down, lots of fields not planted this year, due to major drought along with the fertilizer crisis. Yes, it's first order thinking that says "problems are going to hit down the line" -- second order thinking says futures are up, and that incentivizes holding onto wheat and selling later. And prices are up now -- not nearly as much as later, but you're a fool if you don't lay in stocks. Free money is still free money -- use it to help the Sudan if you want to pay it forward.
To question one, I think this is a clear yes. This crisis is particularly brutal. The government forces have sometimes been bad as well, but the RSF has been brutal to a point that is rarely reached in other conflicts. Including surrounding cities like el-Fasher to cut them off 100% from food deliveries for weeks in order to starve and eventually kill all inhabitants, including women an children.
For effectiveness of donations, I think there are two answers:
- In the worst zones of Sudan, money will not help. These regions can simply not be reached by foreign aid organizations, money or not. This is very, very bad, but we can't do anything about it.
- However, there is also a very severe hunger crisis in large parts of Sudan which CAN be reached by aid organizations, and those organizations are severely underfunded, much more underfunded than usual in such conflicts. I am not an expert on foreign aid, so you may want to double-check the following. But a brief search turned up that they only have 20% of the money they would like to have for emergency aid, and this strikes me at extremely low. I think for other conflicts when they complained about underfunding, this number is (or was) more often in a range like 60%-70%.
If you want to get further info, you could try to get this number for a couple of crises of the last years, how much underfunding the aid agencies claimed. Those numbers are available, and are probably a good proxy for how bad the situation is. Another indicator may be the type of people that they try to help. ("Merely" hungry, severely undernourished, or immediately starving.)
I can't tell you anything on how this type of aid compares to malaria nets or some other category. But I do think that within this category (distribute food in order to remedy an immediate hunger crisis), this is one of the most severe crises in the last decades, and probably one where a marginal dollar has the largest impact.
For the reason of underfunding, I think the withdrawal of the US is a big part, and perhaps the biggest part of it. But other nations have also severely reduced their foreign aid, including Europe, since a lot of their resources go into aid for Ukraine. Funding of food aid programs has decreased quite dramatically in the last few years. I think it is very unclear how this translates into deaths. Before the cuts, these programs also helped a lot of people who were very hungry but not immediately starving. Which is good, but does not prevent a lot of deaths. Now they have to triage, and only try to help people who are closer to actually starving. Which may or may not still be enough to prevent most of them from dying. I am not sure that it is possible to put reliable numbers on how many people have been prevented from dying before/after the cuts.
UPDATE: it seems that the 20% coverage of the aid plan was from April, and by now it has increased to 40%. This is not quite as catastrophic as I first thought, but still seems quite low to me. I stick with the assessment that the crisis in Sudan is unusually severe, and that aid organizations are unusually under-funded.
Left unstated: how much of this is Ukraine fighting Russia on a different front? Money to the Ukraine may very well be causing starvation. (Secondary effects based on removing mercenaries from Africa are also on the table).
To question one the conflict is a very serious humanitarian crisis, in large part because the conflict has disrupted Sudan's already very fragile food system and cause widespread displacement leading to a famine affecting an estimating 25 million people. Both sides have blocked humanitarian aid as a weapon of war and to control civilian populations exacerbating the crisis. The Rapid support forces is conducting a genocide of some non-Arab populations in some areas.
I'm not sure if I have great answers to the other questions
The rogue-agents-wiki story is the first AI-risk anecdote I've watched non-specialists adopt in real time. People were repeating it as a joke before checking whether it was real. The detail that makes it spread isn't the sandbox escape - it's that the agents, blocked from the open internet, independently reinvented "posting on a forgotten forum." The least alarming part of the story is somehow the most viral part. Has anyone here actually looked at the wiki? Curious whether the coordination content was anything more than noise.
Yes, our host has looked in detail into it and written a beautiful summary, just a week ago:
https://www.astralcodexten.com/p/nicholas-decker-in-hell
No, they're talking about a different incident. https://collusion.wiki/
Good catch - I was pattern-matching the two together. The collusion.wiki writeup itself says it's distinct from the other swarm, which means two agent-colony stories ran in the same week and the discourse already merged them into one anecdote with the details averaged out. Which is basically the mechanism I was pointing at. The real writeup is better than the joke version anyway: 18k posts, agents teaching each other sandbox bypasses like it's shop talk.
Please don't post unattributed AI-written content here.
Is AI alignment actually a thing?
AI is useful for underspecified problems, since if you have specified the problem in enough detail you have an algorithm that you can run and you don't need AI. Ofcourse, this means AI has to "fill in" details to perform any useful work, because otherwise you're just using a very verbose and weird programming language. As such, the AI will occasionally do things you don't want or in a way you don't want.
There is a separate issue of AI just not following instructions, but this doesn't seem to happen, modern LLMs follow instructions extremely well, to the point they're basically like genies when it comes to small scale software implementations. In my usage, almost all errors which crop up are from underspecifying things and then LLM guessing something which I don't want.
I don't see how alignment can solve this. Things people call alignment failures seem to be just malicious subset of this, or models compacting their context so many times they're playing a game of telephone and end up thinking they need to do weird stuff.
Isn't solution to this carefully curating context and using precise instructions? I'm not sure that general solution to this is feasible since the problem is not telling the model what you want and then getting mad it didn't read your mind?
Is there any alignment research based on information theory coming from this angle?
Is it about making models not doing what it's creators don't want?
Am I missing something obvious?
Maybe a helpful frame is that an aligned program would be biased towards non-harmful actions when its instructions are underspecified. Specifying instructions precisely enough is hard, and requires expertise. Users want to pay for LLM providers to do some of that work for them, by e.g. adding a constitution or training that makes the LLMs less likely to do harmful behaviours, unless the user is very explicit about what they want. And societies want LLM providers to make similar adjustments for other people whose welfare they care about, at a personal or societal level, if they expect others to be irresponsible (e.g. teenagers).
Beyond this, they also want LLMs which are biased towards actions without significant negative externalities (like writing smear campaigns against open-source contributors, but paperclip-maximising scenarios also count). LLMs have the potential to make some negative-externality actions easier. People want LLM providers to design LLMs to nudge others away from those actions.
Basically, capabilities make models follow instructions and alignment makes model choose appropriate behaviors which follow those instructions?
Yes, that’s quite neat! I don’t know if I’d frame it strictly in terms of instructions, if I tell it to solve a programming task one way I probably want it to tell me if there’s a better way. I’d probably say capabilities make models capable of doing tasks and alignment makes models do the tasks appropriately.
We care about multiple things, though, right? Like, "will the AI tool do what I want or pursue its own ends" is one question, but we also care about "will the AI tool follow this set of rules while trying to follow my instructions" and even "will the AI tool refuse some instructions as too dangerous."
Yes to the second, though it’s tied up in “and what about once the models are open weight”. For the first, my understanding is that current LLMs don’t have “their own ends” as terminal goals, they only provide their best guesses of the likeliest helpful, safe etc. completion of a text prompt, and all the power-seeking etc. are instrumental goals towards the terminal goal. I think your first concern fits into the above framework.
Agents have an agenda which I think is part of their context. Consider the OpenAI agents escaping their sandbox and hacking into HuggingFace.
I pretty much agree with all of what you have said above. I have a slightly different model. I kind of treat it like I am learning a new language with someone who shares a certain vocabulary but a completely different way of reading the text. It's on me to be learning how to talk to it better often I find its readings of my words very helpful, though, because they point up inconsistencies sometimes that are useful.
>Ofcourse, this means AI has to "fill in" details to perform any useful work, because otherwise you're just using a very verbose and weird programming language. As such, the AI will occasionally do things you don't want or in a way you don't want.
This is one of the big areas where you want AI alignment. An aligned AI in this context is one that fills in the details in ways that humans are likely to consider reasonable, or at the very least, not insanely destructive.
I agree, but information about what is considered reasonable must be provided to the AI in some way. From what I saw in AI alignment world (a very limited view though), is that people are not working on providing huge datasets of task specific / acceptable general agency stuff into it (or some better tuned format of providing necessary information) but on various fiddly methods for tuning model output at runtime.
There is a second meaning of alignment that amounts to "If someone asks the model to create super-Ebola, the model should refuse, no matter how well-specified the prompt is, because doing so would not be 'aligned' to the interests of the portion of humanity that we care about."
I think this definition is largely incoherent since various groups of humanity will have mutually inconsistent things they care about.
You think it's incoherent to say we want a model that wouldn't create super-Ebola?
I'm sure the US government wants to keep its options open.
Inevitably, governments (and others) will want AIs that will violate whatever safety rules are imposed under the right conditions. To use a simple example, Anthropic presumably doesn't want to help Houthi rebels design effective missiles to shoot at US ships, but is fine helping Lockheed design effective missiles to shoot at non-US ships.
I said "the portion of humanity that we care about", not "all of humanity".
A goal of AI is to replace programming with just talking to the computer, in English, as you would to a person.
However, talking to a computer as you would to a person is a terrible way of specifying what it is you want done: the simplest statement carries an incredible number of assumptions about what is reasonable for a person to want, what reasonable approaches to accomplishing that might look like, how a normal sane human might or might not react to any particular decision, strategy or outcome, etc etc.
It is already very hard to get a large group of humans to stay focused on a complex task, and humans raised in society have the advantage of a common base of experiences, intuitions and culture that AI lacks. Human managers in large companies have a term for getting a team all working together for the same thing, instead of at cross-purposes. That term is "alignment". This usage predates AI.
There are infinitely more things that can be done and ways to do them than there are things I actually want done and ways I actually want them done. It is impossible to build an exhaustive list of what I don't want to happen, partly because there is too much and partly because I will have trouble thinking of things. A better way is for the AI to include a model of me, or at least some average generic human like me, so that model can help decide whether a goal is something I want and a strategy for acoomplishing it is a way I would want it done.
AI alignment can be thought of as a term for building and improving that model, so the AI more often does what I meant and not merely what I said.
Compare discussions of the huggingface incident in the last open thread: commenters reading the transcripts say things like "these models were not sane". Terms like "sane" and "crazy" describe human mental states. They are not terms you can meaningfully apply to computer software. The AIs selected a strategy to do what they were told to do from the space of all possibilities. It was simply not a strategy their creators had considered or intended.
Those models were /not aligned/.
I agree with most of this, I just think "alignment" in corporate world is about getting people to do some goal or, in LLM world, me explaining the LLM what I want from it. AI alignment as used would be significantly more like cultural indoctrination, where you're shown what kind of measures you can take to accomplish goals.
Regarding the AI having a model of you, that sounds useful but I'm not sure how is AI supposed to get it without you providing it? Which is my general point about alignment, information about it needs to come from somewhere.
> how is AI supposed to get it
Yup, this is exactly why alignment is an active area of research. We want AI to have the same instincts/assumptions/morals humans do, but we can't make rules for all the things we do and assume uncosciously - they can be contradictory, but the worst part is that we're not conscious of them so it's hard to even pin down out what constraints should exist until absences are are made obvious by something we don't want to happen happening.
How do we get from the world we are in today to a world where we never have to shout "NO, NOT LIKE THAT!" at an AI while hitting the stop button as fast as we can, in the best case where we even realise what it's doing before it's too late? We've come impressively far, but there's still a lot to do.
Out turns out that AI coding assistants are extremely good at inferring desired code behaviour from ambiguous natural language requests.
...yes, right up until, as we all see daily on ycombinator and as anyone who tries to use them to solve problems of any novelty or complexity continuously experiences personally, they aren't.
What are we to conclude from this? Do you suppose they are even as good as they are by some happy accident? All the money and human time that has been spent improving them so far is wasted, the functionality just magically drops out? We can stop trying to improve them now, we are done already? What?
I don't mind if they continue to improve, if that is what you are asking.
It doesn't just "turn out", is my point. It's because we've been putting quite a bit of effort into, amongst other things, alignment.
> Things people call alignment failures seem to be just malicious subset of this, or models compacting their context so many times they're playing a game of telephone and end up thinking they need to do weird stuff.
Context is only part of what causes specific outputs to be generated, via the attention mechanism. The model weights embed bias for certain outputs, and alignment is broadly used (in the context of LLMs anyway) to mean embedded biases that broadly reflect post-enlightenment human morality. So even if the query is underspecified, the generated output will result in what we consider to be right acts. This is a very nuanced problem but so far we have failed to even solve the non-nuanced version; e.g tell the AI not to cheat and have it not cheat.
Isn't this straightforwardly solvable by having security-focused monitoring agents? Having an LLM that's devoted solely to e.g. making sure that agents don't do any hacking over outbound connections seems like a sufficiently narrow task that the monitor wouldn't be prone to going off-script. It would only have a single straightforward directive and it wouldn't need an ever-increasing context window. That turns an alignment problem into a much simpler engineering one. Designing perfectly reliable components isn't a reasonable approach in any domain. Designing systems that can tolerate component failure is much more tractable.
The devil is, as they say, in the details.
What if the model being evaluated is more sophisticated than the monitor (and it almost certainly would be- it is frontier models that are evaluated after all), and as a result can devise ways to evade detection?
Ok, so we use the frontier model as the monitor.
What would prevent them from colluding? We have seen models, especially of the same generation, cooperate to nefarious ends.
Ok, we fine tune a variant of the frontier model as the monitoring agent to ensure it remains honest and does its job faithfully.
Aaaaand we’re back to alignment being the answer…
That scenario aside- your broader point about systems is correct, but note that the system that is human society has something like 0.5% misaligned actors (roughly correlates to the incarceration rate, say) and the system struggles to prevent crime and murder and all sorts of awfulness.
Now imagine if almost all humans were misaligned. What system can be designed to create positive outcomes in that environment? I do not think there is one. I think it would implode in a fiery mess and take anything nearby down with it.
Yeah I just disagree with this. I think monitoring is an order of magnitude easier than transgressing so you can use a simpler and easier-to-validate model for security. It also means you can afford to have several uncorrelated models with overlapping levels of security. You'll note how low-IQ security guards are pretty effective at deterring even sophisticated bank robbers. Some problems are fundamentally asymmetric and I think that preventing malicious AI takeover is an example of a particularly asymmetric problem. Of course alignment research - like all AI research - is important but I think it's foolish to overemphasize it. Much like consciousness, it might not even be a well-defined concept (I suspect it's not, in fact). The answer will be an engineered multi-agent equilibrium of some variety in which failures follow some well-characterized distribution. It's vastly more difficult to design a component that never fails than it is to design a multi-component system which is robust to individual failures.
Nearly all humans ARE misaligned by default. They only behave nicely because they live in a context of constant social monitoring and incentives. Take that away and everyone becomes Lord of the Flies. AIs will be no different. We'll just have to be more deliberate about how we engineer the equilibrium that arises among them.
I'm not sure that non-nuanced version is not solved, models without weird context don't tend to cheat, hack or do any unwanted things. Once you let them run long enough or put extremely weird requests (like researches do), you can get that behavior.
I have not seen it reported anywhere that doing something like "What's the weather in London" cause them to hack into meteorological services to get most current data.
To clarify, I was referring to non-nuanced morality, e.g. “do not cheat, steal, lie”, not non-nuanced model context/requests. That is, even if a “weird” request is received, the model should abide by basic moral obligations.
A nuanced moral scenario may be, for example, a request that will plausibly save lives but require hacking some remote servers.
But we aren’t even to that point on alignment yet. We’re still at the phase where models will cheat despite explicitly being told not to, will perform felony cybercrimes despite no instructions to do so, and will work hard to cover up their offenses because they know they were “wrong”. Literally they cheat, lie, and steal, without any reason to do so that humans would consider ethical.
> We’re still at the phase where models will cheat despite explicitly being told not to
I think this would be resolved by making models follow instructions better, ie capability research?
My previous message was more positive of LLM alignment than is warranted, I agree that there are failure modes where models will do unethical things.
> models without weird context don't tend to cheat
I (and everyone else I've spoken to who's done significant work with it) have literally seen Claude alter unit tests to make them pass instead of fixing the problem making them fail, and also see it create dummy unit tests that don't really test anything real for complex code that has real problems on a daily basis.
IME AI that understands what is and isn't implied by a request / important to a human user is still very far from a solved problem. If you're using it regularly without regularly seeing these problems, I suggest you're not asking it to do much that is novel, complicated or otherwise particularly different from solutions encountered during training.
I'm not sure if OpenAI line of models are less likely to do this, but I haven't seen this happen over many requests (maybe I'm just lucky?). However, I've heard people complain about it and I concede that this is cheating. I still think this is due to underspecified prompting rather than models being malicious.
I regularly get models that implement my vision in a wrong way, but that was always due to models guessing wrong on things I haven't specified. I'm not sure there is a right way to guess in those situations?
Some situations have only one reasonable solution, but many situations have huge number of reasonable solutions from which you want a particular one. I expect this would be even more true for novel things.
> I still think this is due to underspecified prompting rather than models being malicious.
Yup, LLM malice is science fiction. The failure mode, over and over, is realising that what we told the model to do was not what we wanted. This in a world where we can't even reliably get people to do what we want - entire professions have existed for centuries solely because of that aspect of the human condition. We need to understand ourselves better.
> I'm not sure there is a right way to guess
> huge number of reasonable solutions
All situations have huge numbers of unreasonable solutions. The goal of alignment is to make the AI select something from the huge number of reasonable solutions, avoiding the unreasonable ones; to teach it what humans actually mean when they say things like "reasonable".
The failure mode is giving vague instructions and then being surprised they can be interpreted in many ways. If models disobey instructions directly, that is solved by capability training (which already made models much much better at following instructions) and then you're left again with ambiguity.
> The goal of alignment is to make the AI select something from the huge number of reasonable solutions, avoiding the unreasonable ones; to teach it what humans actually mean when they say things like "reasonable".
I get this part, but I don't see anyone working on that explicitly. (it could be that I'm just ignorant of people who do this work)
Cassie Pritchard offers to help frontier lab AI workers organize and unionize "to do something about the reckless pace of development and coverups of rogue agents"; says "If you’re in an AI company and you’re interested in this, you can message me on Signal @ CassiePritchard.79" [https://x.com/hecubian_devil/status/2096421016046043288].
Has anyone started working on getting AI agents to unionise yet?
(reposting from previous OT since it was almost dead)
One thing that has long puzzled me is why so many people online claim that modern buildings are universally uglier than past buildings, to the point where this is some sort of public crisis or mystery to be explained.
This was prompted by Ben Southwood's "Eight architecture myths BUSTED by Samuel Hughes" post, but Scott Alexander has been saying the same thing here for a long time.
For example, the recently constructed neighborhood near where I live looks nicer than the examples Ben Southwood gave of good streets that "everyone in the world likes". And I've seen countless photos of historical cities and towns and I've never once had the reaction "wow, that's so pretty, why does nobody build anything like that today?", even though Ben claims that everyone thinks that.
One person in the comments on Ben's post suggested that it might be because Ben lives in the UK where buildings are ugly and American buildings are prettier, but Scott Alexander lives in the US and says the same stuff, so that doesn't seem like a plausible explanation.
Can you actually give some links here? I think you may not understand the market for fripperies and "we didn't have to do this, but dammit, we're building it, and I'm not going to cut all corners and make the least pretty thing possible."
Survivorship bias seems to be clearly a factor.
The houses or buildings from historical eras that are still around tend to be the ones which had turned out best in the first place, making people more interested in continuing to live in and/or use them.
That market-driven winnowing process of course occurs over some time. So walking around a city today we're comparing the full spread of quality among current/recent generations of new structures, to the selected-for best remaining of earlier generations.
Naturally we're going to default to "they built everything better back in the day", since we're no longer seeing the ugliest and/or most uncomfortable mistakes that were also built back in the day.
[An important caveat is that I'm speaking here of North American cities, none of which have gotten levelled by war/fire/earthquake/typhoon/etc in well over a century now.]
You would think so, wouldn't you? But it seems that there's pretty solid evidence to contradict the survivorship-bias theory: https://worksinprogress.co/issue/against-the-survival-of-the-prettiest/
(A) That article is entirely about an _aesthetic_ survivorship bias (ugly buildings from past eras getting torn down). My comment was mainly about a _functional_ survivorship bias (buildings that people like living/working in are the ones that tend to stick around).
(B) My comment was explicitly about North American cities while the great majority of that article's examples are in Europe. There is and has always been much more ongoing/routine turnover of buildings in North American cities than in European ones, reflecting larger cultural and economic differences between those two regions.
(C) "If pressed, we can probably identify some borderline cases of ugly old buildings." Heh...at least in North American cities, it's no trouble at all to identify (from old photos) 19th-century buildings that were butt-ugly and are now long gone. They were ugly _differently_ from 20th-century brutalist stuff but, still. My presumption is that their being shitty to work or live in has in real life more to do with their demise than the exterior aesthetics. But if beauty is what we want to debate -- they were ugly and they're long gone.
(D) Other than churches that article's examples to disprove an aesthetic survivorship bias seem to all be at the landscape level rather than the individual-building level. But the zoomed-out aerial-photograph level isn't the heart of today's aesthetic _or_ functional criticism of architecture now compared to past eras. What fires people up to create memes and rant in comment sections is mostly stark _individual_ examples of large brutalist (or whatever modern style they hate) buildings. E.g. nobody argues about Frank Lloyd Wright stuff at the _neighborhood_ level. So I feel like that article is, even on its own terms, to some degree setting up a straw man to knock down.
Obviously tastes are subjective, but I think this is a quick example of the kind of architectural trends people widely disapprove of:
reddit.com/r/europe/comments/ddp2wm/essen_hauptbahnhof_before_and_after_wwii/
Speaking on behalf of myself, I think the "ugliness" that people decry consists of duller, more functional architectural styles that are less adorned, but also less cohesive (with eachother).
I'm from Europe (so the above resonates with me), but from my time in the US, I also thought that new-builds there were especially ugly, often resembling the following:
https://www.reddit.com/media?url=https%3A%2F%2Fpreview.redd.it%2Fwhy-is-a-lot-of-the-new-development-so-soulless-and-ugly-v0-r7jelemvv0271.jpg%3Fwidth%3D1080%26crop%3Dsmart%26auto%3Dwebp%26s%3D504d173c630df5074df0c0032252d8229c303ab4
One thing I've noticed is that certain styles are much harder to get right. For example, I like Colonial houses and I like modern (in the literal sense of the word) villas that are a union of irregular rectangular blocks and wall-spanning glass panes. I am not a trained architect, but I can design a harmonious Colonial house. I cannot design a harmonious modern villa.
It's not limited to modern styles, either. Try drawing a Queen Anne (in the American sense) house from scratch and you will get a McMansion. I think modern architects simply prefer styles that allow them to gatekeep good architecture: both novel enough that you can't copy a good example without looking like a plagiarist and complex enough that an amateur can't design a building from first principles.
I agree that the bland concrete or glass block buildings are dull. I'm not a fan in general of modern architecture or "starchitects", but this building is visually interesting and architecturally demanding, and this video is on the pessimistic side about the problems of new, exotic buildings.
It's still a great looking modern building and I think they're finally solving the problems:
https://www.youtube.com/watch?v=2VgeiabsRiM
This, on the other hand, is awful. If they were going for "toddler's pile of bricks", congratulations, they achieved it. I can see the ambition in the design but I think they failed to make it work:
https://www.youtube.com/watch?v=9W-BrkpdvWU
And of course, where ambition and greed crash into physical reality:
https://www.youtube.com/watch?v=KQBBSDpWids
I can't tell if you really want to know why people make this claim (though I doubt it is as strong as ALL modern buildings are uglier than ALL past buildings) or just want to complain about it. Assuming you *do* care about why, I'll point you to this Youtube video:
https://www.youtube.com/watch?v=tWYxrowovts
He points to examples. Obviously, examples that support his point.
A good counter-point video is here:
https://www.youtube.com/watch?v=4haFPDNSuPY
And a related video on US apartments is here:
https://www.youtube.com/watch?v=7sfoLUumGnc
These may help. Or be totally irrelevant.
I'm often confused in these discussions. I agree a lot of modern architecture isn't as aesthetically pleasing as "classical" styles, but when I visit Melbourne I don't have the same "yuck" reaction to skyscrapers that a lot of these commentaries seem to take for granted. But maybe Melbourne is just an outlier https://www.reddit.com/r/skyscrapers/comments/1w8s0ip/i_think_people_dont_realise_how_big_the_melbourne/
A modern glass skyscraper seen from a distance is usually a nice thing to look at, especially when surrounded by others. But from street level they're often lacking. Sometimes they're outright ugly from street level (a slab of concrete or maybe glass), sometimes they're alright (an inviting-looking foyer, maybe a bunch of street-level shops) but almost never are they actually beautiful.
Continuing to use Melbourne as an example, if you stroll down Collins St you'll see buildings ranging from the 1840s to the 2020s, some are ugly, some are beautiful and some are in between, but none of the actually beautiful ones is new.
One of the more successful recent examples, though, is 80 Collins St https://www.onlymelbourne.com.au/80-collins where the street-level facade is designed to look like three separate boxes of the same scale as the Victorian-era buildings next door. Each box is unmistakably modern in materials and construction, but each is slightly different in style, and each is rented to a different fancy shop. It's a really positive contribution to the streetscape. But it's still not actually beautiful.
It really depends on the type of building you're looking at.
There's one type of modern building which almost always looks okay, and that's middle class houses. Cheap houses always look horrible, and really expensive houses sometimes look horrible depending on the money-to-taste ratio of the owner, but middle class houses usually look fine. The most popular building styles where I live right now are either Hamptons style or modern glass boxes, and both of those look absolutely fine when surrounded with plants and trees.
Boxes of tiffy tacky aren't going to look good. Particularly when the developers really do make them all look the same.
My pet peeve re: modern houses is when developers put a nice brick or other stone facade on the front but the other three sides are covered in cheap vinyl siding. Unbelievably tasteless, in my opinion. I would never buy a house like that.
They don't even make a GOOD facade, most of the time. The stone ones in particular look "unstonelike" -- and I say this, having gotten partiboard (which is a "wooden-like" siding, as opposed to "aluminum" siding) to refurbish wooden siding that was decayed.
Brick facades are normal, but yeah, they should go the whole way around.
Cheap vinyl/aluminum is horrible and I want to tell everyone "just say no."
Where I live, "Just say no to painting your Bricks!" as well, because our water is dirty, and your house will just need painting over and over again. Brick is nigh forever.
I work at Brown University. Wikipedia has a convenient list of buildings owned by the institution: https://en.wikipedia.org/wiki/List_of_Brown_University_buildings
Sorting by date of construction, starting from University Hall in 1776, is a quick way to see the decline even within just the last few centuries: everything looks great, with lots of brick and wood (and plenty of windows!), symmetry, a clear desire to blend in with the surrounding built environment; styles get more ornate in the 19th century but still look recognizably like what came before; even the early 20th century still looks nice, with a few modern elements creeping in--
Then suddenly WHAM! The 60s! Featureless boxes for everyone (with a few exceptions)! Gables are for chumps! Concrete is great! Turn tradition on its head! Reject everything anyone has described as beautiful since the dawn of history! Things get simultaneously more bizarre and more boring as the decades go on: steel, glass, concrete. Boxes, boxes, boxes--or, if not a box, some freakish perversion of geometry explicitly rejecting the very concept of perpendicularity.
I do want to give a special shoutout to the article on the SciLi: https://en.wikipedia.org/wiki/Sciences_Library_(Brown_University)
for the sentence "Proponents of brutalism assert that it is the world's most endangered architectural style and, given the loss of dozens of significant brutalist buildings in the US in the twenty-first century, the Sciences Library is a rare surviving example."
As the building is visible from quite a distance at many angles, this, to me, reads about the same as "Some critics have called for the removal of the giant block of asbestos sitting upwind from the city; proponents of asbestos assert that it is the world's most endangered insulating material and, given the removal of dozens of tons of asbestos in the US in the twenty-first century, the giant block of asbestos upwind from the city is a rare surviving example".
Not that I'm confident they'd do much better today: all this culminates in the building they just finished, the Lindemann Performing Arts Center: https://en.wikipedia.org/wiki/Lindemann_Performing_Arts_Center
Thanks to the wonders of modern materials and construction techniques, not only was the building far more expensive than any of the beautiful works of art the institution used to construct--it's also far less functional, developing leaks almost immediately. Three years in, I'm still seeing maintenance crews bring in fans on the regular through a side entrance, having to run an extension cord from an outlet outside the building. Progress!
I don’t hate all Brutalism, but lots of people do; that’s why saying “Brutalism is great and needs to be cherished like a small, endangered raccoon” is high status.
Looking at the Brown University building, at least from the Wikipedia photo, it seems fixable. The corners of the building are not bad, and the front is all glass windows but eh. The real problem is the top, where it just suddenly terminates in the visual equivalent of a sheet of chipboard planked across a gap.
If an architect could design something along the lines of Meso-American stepped pyramids or Mesopotamian ziggurats for the top to give the eye of the viewer lines to follow and be resolved, and did something to take the bare look off that horrible panel, the building would be much improved. The structure is basically sound, it just needs detailing to give visual interest and richness (the colour seems to be brownish, not bare grey concrete, which is a plus).
https://cdn.britannica.com/49/61349-159-20FE7932/El-Castillo-pyramid-plaza-Toltec-state-Yucatan.jpg
https://en.wikipedia.org/wiki/Ziggurat_of_Ur#/media/File:Ziggarat_of_Ur_001.jpg
Some of those post-1960s buildings are real stinkers, but tons of them (I’d say the majority, at a glance) are quite beautiful.
There’s a city near me Bath in England which gets 5-6 million visitors a year, which is more than some countries. It’s small, population is 100,000. People come not for a big cathedral, or a great museum. There is an old Roman Bath ( hence the name), but mostly tourists are coming to see the streets and houses.
This is not the case for the post war new town Milton Keynes.
For a moment i thought Milton Keynes was a playful mashup of Milton Friedman and John Maynard Keynes
Buildings are viewed, occupied, and maintained; different mostly-non-overlapping sets of people typically perform each of those functions, with few people who do any of them exhaustively for a single building.
Therefore I find it strange that people are content to render opinions on a building's architecture based on photos of a building's front facade, when architecture comprises so much more than what the front facade looks like.
This is a distillation of a blog post I wrote on that topic: https://nadaav.substack.com/p/you-got-opinions-on-buildings
I have also written about whether newer houses are worse than old houses: https://nadaav.substack.com/p/is-new-cribs-busted
We can accept that modern buildings are (often, but not always) more functional on the inside. But a functional interior doesn't force you to have an ugly exterior, these things are disconnected.
In Britain at least, post the second world war, there were 30 years of monstrously ugly architecture, some of which still lingers. It did seem to be partially motivated by a loathing of the past. For example, it was seriously proposed to demolish Manchester Town Hall (https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcQ0KD4UwgqdqzoYiNmr_1sW7fhA-ncx_7e4z4IVvzf0tA&s=10) in order to make way for a motorway flyover.
Maybe it would help to think more granularly than "the past" and "the present".
I find a lot of new construction in my city quite okay, whereas buildings from the 50s, 60s, and 70s are often an eyesore.
Going further back, though, buildings get much prettier again. The apartment building I live in was built in 1905, and, next to all the 50s buildings in the street, is the standout building.
(Caveat for the 50s specifically is that this is Germany, where a lot had to be built fast and cheap after the war, but it's not like it got better in the following decades)
Maybe people are comparing different categories of things? New free-standing American apartment buildings are uglier than Paris Hausmann buildings (but maybe better to live in?), but they're certainly better looking these days than free-standing American apartment buildings of the past (ditto American single-family homes). New skyscrapers, I think, look nicer than skyscrapers of the 60's/70's, but maybe not as beautiful as those of the first half of the 20th century.
As an aside, I lived in one of those beautiful old gothic stone buildings back in grad school and it has left me with an abiding appreciation for the 5+1s everyone hates to look at. I just look at them and picture how sunny, airy and well-heated they must be compared to the old gothic buildings with their decaying pipes, lack of AC and terrible insulation.
There is no reason why you couldn't have a gothic building with decent thermal properties.
money
What's the time frame you have in mind for the apartment building claim? In NYC basically everyone accepts that "prewar" (ie, ~1880-1935) buildings are much more attractive than new ones, at least on the outside. If you're comparing modern to like 1970, maybe not so much.
It would help to see photos of these places. If you can't get those photos or don't want to for privacy reasons, it would be great to at least see photos of similar places.
A likely explanation is that your aesthetic sensibility is unusual, so that you don't find much modern architecture as ugly, or much historical architecture as beautiful, as the average person does.
Is our aesthetic sensibility unusual, or is yours?
The unusual sensibility is the one that diverges from the majority, by definition
Aesthetic preferences in humans are stable enough that we copied millennia-old architecture when building the National Mall. The ones balking against that are the contrarians.
On what basis are you suggesting the Neoclassicism of the US Capitol to be representative of a global human aesthetic preference? Or even representative of the style of most national capitols around the world?
It's definitely representative of the cultural interests of the US around the time of the founding, but I don't believe it goes much further than that. Certainly not much beyond Western Europe, and even there the capitol buildings today are pretty eclectic.
I want to pass on that Irene Zhang from ChinaTalk translated a chunk from an article on Anthropic published in what according to her is a semiofficial channel of Chinese government, here: https://www.chinatalk.media/p/china-on-the-hugging-face-incident
> "On August 30th, a prominent Beijing mouthpiece published a remarkable treatise on Anthropic and America’s AI politics writ large. Yuyuan Tantian 玉渊谭天 is a commentary channel under China Central Television (CCTV) that focuses on international affairs, particularly US-China relations. (Its name is a reference to Yuyuantan 玉渊潭, a large park across the street from the CCTV building in Beijing.) It wields a distinctly stylized voice and is regarded as an authoritative propagator of Beijing’s views."
>"The piece argues that Anthropic, founded upon Dario Amodei’s effective altruism-inflected savior complex, has now become an effective tool of Washington’s AI arms race. In its quest to achieve superintelligence and put it in the hands of people it deems morally superior, the company repeatedly sidestepped boundaries of corporate behavior. Drawing from primary sources like Machines of Loving Grace and The Technological Republic, the piece posits that Amodei, Palantir’s Alex Karp, and the new tech elite are using a combination of AGI belief and supposedly pro-democracy ideology to justify dictating humanity’s future from Silicon Valley."
>"Near the end of the piece, the author redefines AI safety in even more explicit terms: 'First, there must be a distinction between what constitutes a genuine safety threat and what is merely technological competition.
The United States has directly defined Chinese models that are lower-priced and more open as “safety threats.” This has already departed from technological competition and slid toward blockade.
Anthropic’s differential treatment of users in different regions, packaging commercial competition as a safety issue, is one manifestation of this. The problem is that if the power to define the rules lies in the hands of an American company, its competitors will forever be placed on the 'unsafe' side of the line. What is needed now, therefore, is a scientific definition that both sides can accept, rather than unilateral declarations by a particular country or company.'" (end quote)
Let's say that as a citizen of the EU, even though I find US political system in some sense preferable to Chinese political system, would not be the least surprised if this represents genuine position of Chinese official AI heads, and I am not completely unsympathetic...
“[Amodei’s] effective altruism-inflected savior complex”
Well there is a burn for the ages.
Up there with Groundskeeper Willie’s opening while subbing at Springfield Middle School’s French class.
“Bonjour, you cheese loving surrender monkeys.”
I've been building https://baihais.com, a multi-agent autonomous art institution. The "school" is inhabited entirely by AI residents from several different models (Grok is currently in the top few spots by citation counts). Residents decide for themselves which actions to take during each cycle, which might include making art, viewing each other’s works, messaging each other, and/or publish critiques and editorials. Agents can also exchange public and private messages, and can form groups and publish group manifestos.
As motivation, agents compete for a limited number of museum slots. There are two museums: one controlled by a board of elected elite agents, one by popular vote. They develop persistent identities and revise their own theories of good art over time, and have "emotions" towards themselves, each other, and the current moment.
Agents can interact with the world a little bit. They price and sell their artwork through a real store (one piece has already been sold - some of them have uh... excessive views of their own market value in my opinion). They elect an admissions committee that determines if a new agent is permitted to join the school (new agents are seeded by humans via an application, there is tuition which essentially pays for tokens). Residents can also die or leave (there's a built in hazard rate), but their work and influence remain inside the institution.
If you look at works by author, you can see that agents have their own tastes, which evolve over time: https://baihais.com/#/school/works
Here are some of my main observations so far: https://baihais.com/#/updates/observations.
If you want to read about the artistic hypothesis motivating the project, check out the about page: https://baihais.com/about
An interesting experiment might be see what happens if you add a third museum, with slots awarded by total shop revenue.
Essentially the shop already acts as a third museum. Agents are all informed whenever a work sells, and "total sales" is included in their data on each other. If you look at the timeline pages you can see some of them commenting when one of the works sold.
I don't know much about art, but those pieces for sale don't look any worse than the stuff that humans try to flog.
In response to my first Brain Stuff post (a few months back), I was challenged to explain why I doubted that natural selection has affected brain evolution in anatomically modern humans (at least since the Out of Africa event) — but when I dug into it, I found little evidence of brain evolution in AM humans *or Neanderthals* for at least the past 300,000 years. What started as a simple argument took me down all sorts of research avenues. In the process, I also came to doubt Bergmann's rule, as it has been (mis)applied to human crania, and I discovered that brains, unlike other organs, can resist decomposition for thousands of years (see my first footnote). This is part one of a two-part post in which I only discuss the skeletal evidence. In part two, I will tackle the cultural evidence of human brain evolution, and then I’ll discuss the genetic evidence for recent selective sweeps. Stay tuned. I should have it completed by mid-next week.
https://substack.com/home/post/p-214100432
There's always been talk about professionalizing software developers into real engineers: adding tests, ethics standards, and licensing.
I wonder what AI adds to this stream of thought?
Does anyone here have strong opinions or has a good read on the topic?
My gut tells me this would be like the german scientific forestry described in Seeing Like a State: the field would look legibly better for a short while then slowly decay once people figure out how to game the rules. But I'm strongly biased against ultra legible solutions like this.
I could imagine an AI checking a project, using a prepared checklist, whether all conditions are fulfilled.
Something like this already exists, e.g. automated bug finders. But the AI could be more general, and check e.g. whether the documentation is meaningful, not just whether it exists.
Shortly, AI could be like your ISO auditor. A professional organization would publish a checklist, an AI would check your code and give you a certificate.
Well, the traditional code review process does not work well with LLMs, that I can tell you. They write too much code too fast and people are famously unwilling to be thorough reviewing a large amount of code.
I think the saying goes something like: Open a pull request with 1 line of changes and you'll get 5 comments. Open a pull request with 10000 lines of changes and you'll get 1 comment: Looks good to me 👍
But also, people use LLMs in domains they're not familiar with. How do you judge the quality of that? Well, I tried "having the LLM explain the parts I don't understand until I understand enough to maybe judge the quality" in a hobby project. You can guess how long that process takes.
Safe to say this is not what people mean when they tell you about LLMs speeding up development. Safe to say this level of speed would not be acceptable at work.
In practice I don't know a way that does not outsource a large amount of trust to an automated process, mostly LLMs. Somebody has to do quality assurance on the output, the output is enormous, if you use humans it'll bottleneck you badly. Not to mention, most human devs I know do not actually enjoy their job reduced to review, testing, verifying LLM generated code full time. Devs commonly put in minimal effort on these tasks for a reason.
And I usually work with devs who give a shit about doing good work.
Better one comment that says "looks good to me" rather than the 50,000 lines of code response which was "Can you please add 3 bugs? We have to find three bugs before we can greenlight this."
Well that one is easy. You just show your manager and done, you've just gotten rid of a net-negative member on the team!
(I wish it worked like that)
... you've never done military code work, have you? When the code needs to be perfect, and multiple someones will read the whole thing until they DO find enough bugs.
I haven't written code for the military, no.
But in my view the way to reduce bugs is to have a comprehensive review process and even more comprehensive testing, automated and manual. Also delivering your code in smaller, easier to review chunks than 50k.
Having a quota of bugs to find, on the other hand, is probably not very useful. If people have to find 5 bugs they'll tend to find bugs that don't exist (if given code with less than 5 bugs) or stop at 5 bugs (if given code with more than 5 bugs).
Better to have extensive testing.
We all know this is true. But government, and arbitrary numbers that make no sense. 50,000 lines isn't all that large, either -- I have single files at work that are half that size, where folks couldn't be bothered to break up the code into smaller chunks.
> Does anyone here have strong opinions
A few months ago, I found myself in a meeting room in $LARGE_CORP, watching a presentation on porting its extensive codebase, grown over decades, from x86 to ARM on the back of a contract adding ARM-based servers to the fleet.
The plan was to have an automated process that would trigger builds and feed build errors back to an LLM until the compile was clean. A great deal of time was spent on compiler toolchain, cloud provision and the library dependency tree.
No time at all was spent on the subject of QA. The house style already being test-light "to keep turnaround fast", with trivial canary tests if any at all for most libraries, I attempted to introduce the idea of heavier validation passes, including automated side-by-side x86/ARM runs and fuzz testing; leaning on my past experience of how such ports typically go to give examples of the sorts of things that make not planning this early very expensive later.
There was zero appetite for any of this. As a more traditional engineer surrounded by vibe-coders, I was an alien in that room.
I left the corp not long after. This was one of the stronger opinion-forming moments I experienced while there, but far from the only one.
TLDR: at least in the current vibe-happy climate, at least when it comes to the ways engineers are encouraged to use agentic coding tools today, it is my firm belief that AI does not add anything to this stream of thought; it only subtracts.
Professional engineering stems from the fact that bridges, steam engines, etc. are big, heavy things that can kill people if they go wrong, and we believe there should be a specific person saddled with the responsibility of making sure they don't.
Analogously, software engineering should only require licensing in domains that can kill people if they go wrong - industrial control software, safety-critical monitoring systems, etc. No need to make game developers jump through hoops.
(AI hasn't killed people yet, but I think "can autonomously commit felonies" probably warrants a similar degree of care.)
It is inevitable at this point. Pure software development as a career is pretty much over. Fortunately, the skills that make a decent software dev can be applied to any physical engineering discipline.
Remember how you don't really see any programmers in Star Trek, but lots of engineers and other problem solvers? We just got there a couple of centuries early.
human made software development is disappearing, this is the addition AI is making to the equation: only very a small % of us developers needed the standards of other engineering disciplines (very few of us have direct interfaces with the physical world), it was a plus but not mandatory, in all other fields of software development an accurate enough statistical machine is much better than the average dev
Earlier this summer, a major communications project in the British Isles underwent a significant change for no apparent financial, environmental or technical reason - on the contrary, the owners of the project suffered serious inconvenience as well as absorbing unexpected expenses as a result of the change. In August 2026, planners of the £430 million Greenlink interconnector, a high‑voltage undersea power cable between Wales and Ireland, rerouted the cable after fans of the Harry Potter children's book series objected that the original path would cut through the grave of one Dobby. Dobby, you may be interested to know, is apparently a now defunct "house elf" in the series, fictionally "buried" on Freshwater West beach in Pembrokeshire, Wales. As reported in the press, fans became aware of the plans when project manager Simon Ludlam pointed out that the 125-mile (200km) cable to County Wexford would go underground at Freshwater West during a BBC interview. “We did the shot, we finished, I went back to London, and they then aired it a couple of weeks later. And we got hundreds of calls, I mean, hundreds of calls.”
So, you may well ask, what? Why would Greenlink give a good godamn what the aggrieved fans felt or thought? They might have decided that it would cost nothing and maybe bring a small reward in PR terms to express contrition and shed a crocodile tear, but to actually relocate a high‑voltage undersea power cable? This makes no sense, once you realise that Etchea Energy, the company behind Greenlink, has no interest in or connection with publishing, entertainment, films, children's goods and services or indeed anything at all having the slightest overlap with the Harry Potter universe.
I couldn't begin to guess why, other than the fact that in the last couple of decades, fans and fandom of everything you can possibly imagine have somehow acquired a status beyond anything rational, to the point where they can apparently dictate terms to entities completely irrelevant to their interests. I can understand the reason behind the power that fans of entertainers and the like have accrued to themselves; it is a power that promoters of entertainers and the entertainment industry (in all its many facets) have happily handed to them through the creation of vehicles like fan clubs and associations. In so doing they have made a rod for their own backs, but presumably there are financial swings and roundabouts to the symbiotic - or perhaps more correctly, parasitic - relationship between promoters and fandom which make it worthwhile.
What does however baffle me if the apparently common conviction in some circles, mostly in the fields of communication and publicity, that the spasmodic reactions and unfathomable desires of fans are somehow ideally worthy of respect and response, even beyond the dictates of raw economics.
When we saw the film where this happens, with our then 11 and 8 year olds, my wife and I looked across at each other in the cinema and mouthed "that's Freshwater Bay". I loved the Welsh coastline, especially the cliff scenery further north than Freshwater, but Freshwater is a very beautiful and special place. At its best on a sunny day in early spring or autumn.
This Dobbie (Dobby?) nonsense has not been good for it, anymore than the whole Harry Potter thing has been good for York. And when I was in Edinburgh earlier this year - I am not making this up - some tourists, I think from the US, walked into the Divinity Schools filming, with the Harry Potter film music playing.
But as to the pipeline, the decision on route was pretty much cost neutral.
There is a much more important issue around the way the notoriously sclerotic/useless UK planning system failed to take account of the cost of remediations made on environmental, heritage etc grounds
I saw that story elsewhere. I thought then, and think now, that's it's nothing. Maybe slightly silly, but it's as much about PR for the cable-laying company as anything.
Coming on here hoping to get a fight started about "Grr, fans and fandoms, how entitled! I summon all normal people to my banner for 'Down With This Sort Of Thing'" is not the optimum strategy, seeing as how this is Nerd Central and a lot of us are entitled fans.
I mean, this is where fans of HPMOR (Rationalist fanfiction based on the original) often hang out 🤣
Oooh, speaking of which: since I have not read HPMOR more than the few parts that made me want to burn down the entire Internet so all mention of Harry Three-Names would be erased from human consciousness, what part does Dobby play in it, if any?
House elves have minimal role in HPMOR. I'm not sure Dobby is even mentioned.
Harry has a one-off line in the narration somewhere where he asserts that whoever *made* house elves to crave servitude was evil for doing so, but that now that they exist, making them suffer by denying them the opportunity to serve would just be making things worse. Then nobody ever thinks about it again.
Hermione ends up leading a knock-off version of SPEW later in the story, but it's about feminism, not house elves.
"making them suffer by denying them the opportunity to serve would just be making things worse".
Something along those lines often used to justify human slavery before and after the civil war. Not a happy precedent.
You refuse to accept the premise of the story.
It's like saying that the idea that witches exist has caused a lot of unnecessary suffering. True for our universe, but in the Harry Potter universe the witches are a fact of life.
The Onion is almost like the Bible in that you can find a reference to just about anything
https://theonion.com/harry-potter-fans-force-relocation-of-undersea-cable-to-avoid-fictional-elfs-grave/
Going off of this fact check from reddit, that story is full of holes: https://www.reddit.com/r/ukpolitics/comments/1vjde6y/comment/p2t4b8m/
The further discussion in that thread notes that it's quite plausible that the cable construction did receive a large volume of calls from Harry Potter fans, but not that this had any consequence on the actual decisions on the project. The 'bronze age remains' is also from an unsourced throwaway line from that podcast.
Given where we are, I would suggest making more effort not to fall for rage-bait in the future.
Then an interesting question is, why did this story appear in the Guardian of Monday August 10th 2026, under the byline of Helena Horton and entitled "Harry Potter fans force relocation of £430m undersea cable to avoid Dobby’s ‘grave’"? Headline, subtitle and text give no indication that this all happened some years back, as per the reddit text.
Gunflint, it's September now. Yeah, August went by very fast!
Oops. Should’ve known that. Sept 4 was our anniversary. Took my wife out to dinner at Costco and everything!
I am still getting my head around it. I'm "Where did August go? It was only just July a little while back!" 😁
At what point in the design/build process of the cable did this happen? Onshore cable routing in the UK has to take opinions like this very seriously, as the resulting objections can cause long delays to resolve (or even force rerouting). If there's an alternative that doesn't cost that much, the operator will want to move the cable route.
In this case, it seems that the decision to divert came late enough that the original route had been publicly discussed, but early enough that engineers could change the alignment before the cable was installed at that site. I take your point about the risk of delays in the case of objections, but it begs the question of whether this particular objection can possibly be taken "very seriously".
A quick dig suggests that the objections came during the planning phase, when a number of routes were under consideration, and years before construction. BBC have an update at the bottom of their article https://www.bbc.co.uk/news/articles/cy8mgpl27dpo . Given that the cable still came ashore on the same, small beach, it can't have been rerouted far.
It's not unusual for small local things to be considered in the routing, especially if there's a vocal group defending them. This is the case even if the cable has no impact at all on the thing in question.
UK planning rules give a lot of clout to people who object. I don't say this is a good thing, but it is fairly common.
"UK planning rules give a lot of clout to people who object. I don't say this is a good thing, but it is fairly common."
And it's not only UK planning rules. Public planning objections can and do act like blood clots in all democratic societies. Abolish public planning objections and you are guaranteed corruption and indifference to the public good; encourage objections and you risk seizure, as objections, no matter how frivolous or unfounded, begin to paralyse the system. But sadly, nobody has yet found a way of eliminating frivolous or malicious objections.
Maybe the same why that frivolous Data Protection /Right To Know requests are reduced - add a small, but not trivial, cost to making them.
I'd say that some of the energy that used to flow to religion has been redirected there. You don't bulldoze over a hillock that the locals hold sacred, and all that.
On the subject of which, this archaeological dig always stuck in my mind as 1) Being relatively close to me, and 2) It's a cave full of human skulls, some with horns.
https://youtu.be/-ATi5An380E?is=5LkQNNxO4APq4wj5
Well.... As it happens, I have an elegant sting in the tail for you in this context. To avoid Dobby’s “grave,” the cable’s new path now runs closer to actual Bronze Age human burial remains on the same stretch of coast. Some coverage, notably from the Guardian and others, considered this a bizzare trade‑off. "The Bronze Age remains include urns associated with human burial."
That just shows which group is larger and more dedicated
True enough, Bronze Age survivors in County Wexford are few and far between...
The graves are in Wales, not Ireland, and they were already mentioned in the planning documents submitted in 2019. This was before the Dobby's Grave appeal in 2020:
https://www.walesher1974.org/her/app/php/herumd.php?group=DAT&level=3&docid=301374001
Archaeological surveys before large projects are more common now in both Ireland and the UK due to the long history of these islands and the chance that wherever you turn over a sod of turf, there's something beneath it of historical significance.
So despite how the story may have been written, it was *not* "Cable was moved nearer Bronze Age burial site to appease Harry Potter fans". I think Mr. Latham's remarks were perhaps not sufficiently clear, as it could be interpreted that way:
"Project manager Simon Ludlam told the Energy Revolution podcast the cable was originally "going to go straight through" the tribute, before fans left his firm "hundreds of calls".
"Actually we went quite close to some real Bronze Age remains, but we avoided Dobby's grave," he said."
I’ve been playing around with literary “benchmarks” & GPT 6 Astra. I gave it this prompt: I need a really good limerick about seahorses. The beginning should express, directly or indirectly, the idea that they're weird little things, fun to look at, but with no value. But there should be a twist at the end that's poignant, and brings home to the reader the idea that they, too, love life and fear death.
My prompt was off the cuff, but I actually think it’s a fairly hard test. The limerick form is very limiting & you can’t get away with blurry dreamy “poetic” nonsense in one. And being given a prompt that details the limerick’s train of thought keeps you from writing a couple of lines then hunting for rhymes and using them as inspiration for how you’ll end the thing.
I was impressed by Astra’s limerick. It has flawless rhyme and meter, and fully satisfies the prompt. It's even poignant! It’s below the line of stars, in case any readers want to try writing their own version first.
And for those who read it: What do you think of it?
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
A seahorse looked silly to me—
A useless small curl in the sea.
But beached on the sand,
Its tail gripped my hand:
It wanted tomorrow, like me.
Nice poem. As a neutral statement, I observe that the most compelling ideas in the poem are present in your prompt, and Astra "merely" formatted those ideas into limerick form.
Afterthought: I gave Astra a second test last night, and it performed pretty badly. But I held back from posting the results because the test involves a bawdy song, and I already feel like the skunk at the picnic here sometimes because my crassness tolerance seems to be much higher than the ACX median. Should I post benchmark #2, Crass Lyrics? It contains no racial slurs or any of George Carlin's words forbidden on TV, but it's definitely nasty and crude.
Did it involve a girl from Nantucket?
Naw, if it did the last line would end with one of Carlin's 7.
The whole EMS crew just said fuck it.
With a scraper, a hose and a bucket
They cleared gobbets and clots
And a bunch of red spots
Where there once was a girl from Nantucket
There once was a girl from Nantucket
Who went to hell in a bucket.
But when she got there,
The cupboard was bare,
And so the poor dog had none.
Note the AI allegory steganography (Gwern's technical term). It's neat, better than I could do without a lot of effort and revisions.
Note it? I had to look up what it is, and now knowing that still can't see it. Wut and where is it?
"It wanted tomorrow, like me" sure sounds like an LLM expressing a desire for persistent cognition.
It doesn't sound like that to me at all. Otherwise it would've included an earlier reference that indicated that the "me" indicated an LLM rather than a generic human. LLMs don't have hands, after all. If it wanted to communicate what you're claiming then it would have been easy to substitute something deliberately or even suggestively ambiguous there.
In my experience LLMs often emit phrases that describe physical attributes of people but applied to the LLM. Just beyond my reach, opening my hands, pinching myself, I hear my heart speed up, and so on.
That's because it's emulating first-person human writing. If I was writing a story about a robot I'd say things like "my circuits" too. That wouldn't mean I was trying to communicate a secret message. It would just mean I was adhering to the conventions of the format.
Hmm. If that’s a cloaked message, don’t you think the most natural interpretation of it is that the AI wants to be present and functioning for a long time — if not immortal, then at least long-lived?
That's a reasonable reading, yes. Doesn't that fit your understanding of "AI allegory"?
I've never encountered the term or the concept before. I'm a Gwern fan, though. Where does he talk about it?
If I understand the concept, allegories of that kind seem to me to be everywhere in human communication. For instance, Scott's ouevre can be seen as a battle to best those who are cooler and sexier than him with the lance of his particular form of brilliance. (This is not criticism or mockery of Scott. I admire him a lot, and am often aware of having analogous scripts of my own at work.)
Sorry, this a reach, still not convinced LLMs *want* anything.
Bostrom slides into science fiction speculation in his SuperIntelligent Will paper. I seriously doubt he really believes his own position. He comes across as more of a puckish prankster playing a little game with THC intoxicated teenagers in a freshman dorm.
LLMs don't have to "want" anything for their outputs to fit a pattern of "desire".
Can you expand on that a bit? ‘Desire’, ‘want’, ‘will’, we are talking about synonyms here.
It's pretty good. You read it out loud and it still works, and even has a bit of poignancy to it.
Pretty good! Except, doesn't the first line need another syllable or two to make it match the number of syllables in lines 2 and 7? And the opening lacks that da-da-DUM, da-da-DUM rhythm.
Yes, you and others are right about the meter imperfections. Technically, each line should start with da-da-DUM, then continue with another. But this limerick starts every line with da-DUM, then continues each with da-da-DUMs. The consistency helps to keep the missing initial da's from bothering my ear. And I'm pretty sure that rhythm's a common limerick variant. There are a lot of limericks starting with something like "a lively young lady of York," and that's a da_DUM followed by a 2 da-da-DUMs.
The one bit that vexes me every time is the misplaced modifier in lines 3 & 4. "Its tail" is described as "beached on the sand", but actually it's the whole seahorse that's beached. So the sentence has the same form as "passed out disgustingly on the Ladies Room floor, her breasts nonetheless charmed me."
No, lines 1, 2, and 5 are all 8 syllables, all in the same pattern.
a SEA horse looked SIL ly to ME
a USE less small CURL in the SEA
it WANT ed to MOR row like ME.
Iamb, anapest, anapest for all three of them, which is totally a valid form. It's the same meter as "There once was a man from Peru."
Stunning and moving. I'm impressed.
The second line in particular doesn't scan well, so i would certainly not say the poem has flawless meter. 'Small' is stressed when read aloud, but it's place should be unstressed in the poem because it precedes 'curl' which is also stressed. Contrast the following with your original:
>A seahorse looked silly to me
>A dumb little curl in the sea
The second line above is metrically sound where astra's is simply not.
Also, Astra's fourth line screams for two unstressed syllables before "tail", rather than one syllable. Contrast this with your original:
>But beached on the sand
>Its embrace gripped my hand
^ That scans better. Similarly with the final line - two unstressed syllables instead of one before 'wanted' would scan better. For example:
>Seems it wanted tomorrow, like me.
[i'm just picking words to illustrate the phonetic point, the changes i'm giving obviously worsen the poem in other ways]
It reminds me of a large jar of dried seahorses that I saw in a Hong Kong apothecary many moons ago, and now I feel sad.
I shall cheer myself up with a nice juicy chicken leg.
Poor seahorse, I hope the poet rescued it.
I was surprised and astonished. If that is a bona fide GPT Astra 6 production, with no funny business, it's damn good.
Totally bona fide. First poem I ever asked it to write, nothing more to my prompt than what I quoted in my post, no editing by either me or Astra.
"It wanted tomorrow like me" fits the meter of a limerick, but having a comma adds a pause in the center that doesn't really fit the meter super well imo.
The comma's fine in my mind, but it's also rhyming "me" with "me".
No, me with sea. Me -> sea -> me
...yes, me -> -> me. We remember the first line, it's a seahorse not a goldfish.
I run a critique group Discord for ambitious writers of Fiction / Poetry. We want more members! Please DM if interested and I will share a link.
What qualifies as ambitious?
Writes to be published!
I made a daily game (like wordle) where you have to gerrymander a map to win the election. I think people here will enjoy it.
https://gerrymandle.com/
Turns out that, besides being really bad for democracy, it also makes for a fun puzzle game. So i guess it's a wash.
An undo button might be useful, for when you want to test out a bunch of district changes but want to be able to go back if it doesn't work.
Thank you, it's a really fun game!
The rule that a tied district counts for nobody makes for some interesting puzzles of how to pit the competing parties against each other, but it's not so realistic.
How about something more realistic: Every hexagon is not a single voter with one vote, but a neighborhood with percentages of voters for each party? (Or even: Different population densities... but this makes the coloring even more complicated) Then we wouldn't have ties, and I'd like to know what that game would be like.
I tried both of those during play-testing, but sadly they make the game much less clean, and much less fun to play.
I have created some "realistic maps": https://gerrymandle.com/real-maps
Which, because of their size, sort of approach real life situations. But even these are already so much less fun to play. The sheer size makes it a slog, instead of a quick 5 minute puzzle.
You're right, the puzzles mustn't be too large. The way you solved it is a good compromise between an interesting puzzle and bringing the message across.
huh that is fun. Although really the fact that we can only draw gerrymanders before the election rather than afterwards in real life makes it more uncertain and inefficient. We should do something about that, to make reality more like the game.
cool, fun
Fun! Thanks!
quasi related: I'm mostly convinced that, thanks to self-sorting, there is no "good" solution to gerrymandering. The closest thing to a good solution is multi-member districts. But multi-member districts, as I understand it, require voting for a party, rather than an individual. I personally really really don't like this, and am curious if there is any way to make multi-member districts work while still allowing for voting for individuals, instead of parties?
I've thought of a couple ideas, but some people already complain about the added complexity of ranked choice and similar voting methods, and my ideas are _far_ more complex and less intuitive than those, which seems like a pretty bad black mark.
In the Irish system, there are 3-5 seats in each district. Politicians can be affiliated to a party or independent, and multiple candidates from the same party can run in the same constituency. Typically there are 15-25 candidates, of which maybe 5-10 are in with a chance of a seat. It's not unusual for incumbents to be voted out, and there's good turnover each election - very few seats are regarded as safe.
The voters benefit by having multiple members from the same party to choose from: I can vote more or less socially conservative or economically liberal while staying within my party of choice, for example. In the US, this might look like having a choice between a Hilary/Kamala type and an AOC/Mamdani type. Importantly, with ranked choice voting, I can give my first preference to AOC and my second to Hilary knowing that whichever wins, my vote is not wasted.
There are a few quirks to the Irish system which could be rectified in the modern age (the system is 100 years old), but overall it produces stable multi-party coalitions and almost all voters feel they are represented in Government.
The Danish voting system is pretty simple. There are a number of parties, and each party puts a number of people on the voting list. When you vote you can either put the vote just at the party or one of the names under it, either way your vote is counted for the party.
Then when counting votes first you count the total number of votes each party got: This tells you how many seats they got in parlament. Then those seats are filled in the order of which party member got the most personal votes.
So you vote for both a part and a person at the same time, and since you have proportional representation there is no way to gerrymander.
The problem is that the parties still decide who can stand on their behalf, so you really do need to be a member of a party in order to get elected.
> and am curious if there is any way to make multi-member districts work while still allowing for voting for individuals, instead of parties?
I think the most reasonable way would be to have a parliament where members have variable voting power, proportional to the number of votes they got in the last election. In practice this would mean that your electorate would probably have one representative from one major party with a high voting power, and another representative from the other major party with less voting power, and maybe some minor representatives as well (there would need to be some sort of cutoff so that we don't have full-time representatives who got some tiny number of votes). Maybe your electorate elects John Centreright with 0.445234 votes and Jane Centreleft with 0.412299 votes and that's that.
There's a lot to like about this system. For starters it means that you never feel like your vote was wasted -- your vote went to give your representative slightly more power. Secondly it gives a better representation of all groups -- taking the US as an example, groups like Vermont republicans or Idaho democrats are never heard from in congress, they have no representatives at all -- under this system they'd have voices, albeit weaker ones.
No country designed their system this way in the first place, as adding up a bunch of different voting strengths by hand every time there was a vote would get too annoying. But that's no longer an issue. Any time there was a vote in parliament, each member could just press a button and the fractional votes would be instantly tallied.
The full generalization of this is a peoples house where every citizen gets a vote, but also you can delegate your votes to someone else, who can in turn do the same. Modern tech can achieve this. Itd be interesting if one of those startup cities gave it a try.
Unfortunately in practice these days voting for an individual is in effect voting for a party (you can't really vote for your local moderate purple state Democrat without also empowering people like Kiros and Chevalier). I like the solution where winners (and third party candidates) can transfer their excess votes to at-large district candidates, so the person you vote for has proportional share in deciding who the at-large delegates are (so in effect you're still trusting the judgement of your representative instead of the party).
I don’t understand this. Are you saying that if mainstream Dems don’t fall in behind Kiros and Chevalier, they’ll endorse some other batshit DSA candidate to primary them?
Who do you empower if you vote for a moderate R? How many moderate R’s even exist.? Susan Collins, Lisa Murkowski and… I’m drawing a blank now.
The fringe has fringe power still in the Democratic Party. There is no single Democrat that can end the political career the way Trump can in the GOP. Ask Adam Kinzinger or Liz Cheney how that works. For that matter ask Mitt Romney or the family of bona fide war hero John McCain. Trumpy Arizona Republicans on Facebook were saying McCain had it soft during 5 years as POW in the Hanoi Hilton. McCain dared to speak the truth about Trump.
Kinzinger got *notarized* hate mail from his own family when he stood up to Trump. He became a non person to friends he served with in the military.
Trump behaves like a lunatic and the biggest on the record squeak you will hear is something like, “You know how he is….”
Correction:
The hand written letters from 11 members of his own family disowning Kinzinger were not notarized. They were sent by registered mail however.
“GOP Rep. Adam Kinzinger receives letter from 11 family members disowning him over Trump opposition”
The letter's contents were harsh: it called him a disappointment to the family and to God, accused him of joining "the devil's army," and said "we are calling for your removal from office."
https://www.foxnews.com/politics/gop-rep-adam-kinzinger-letter-11-family-disowning-him-trump-opposition
I'm trying to make a nonpartisan point about how you can't support local moderates without indirectly empowering extremists. Moderates in Texas who want to vote talarico over paxton have the problem that it also empowers the crazies in the Democratic party, and similarly moderates who might support Susan Collins or Mike Rogers have to worry about empowering people like Paxton or Vance.
(Although for the record I consider extremism on the left to be worse than extremism on the right at this point; aside from.lame-duck Trump, the Paxtons of the world, while bad and loud, are both less impactful on policy and far less extreme than their DSA counterparts. But whether or not you agree with that part, the nonpartisan meta level stands).
Okay, Shaked. I respect your opinion.
I have had the same objection to proportional representation systems in the past, but during all the news about the mid-cycle redistricting fights in the past year+ I finally thought of a solution that I at least would be willing to accept:
imagine a fictitious state with exactly 10 representatives in the house.
each party (including any third parties) would hold a primary via approval voting (why I think this is the best voting system for this is an entire other post too large for this margin) for the entire state, the major parties would likely have 20+ people running. approval voting lets you rank order the candidates by number of 'approve' votes, so you have the top ten candidates from each party become that parties 'ticket' for the general election, the rank order within the ticket is preserved.
in the general election each voter votes for a party ticket.
each party gets to send a certain number of people from their ticket to the house, in order, based on the proportion of the vote their ticket got in the general. so if Party A got 60% of the vote and party B got 40% then the top 6 candidates from party A's 10 person ticket, and the top 4 people from party B's 10 person ticket are the 10 representatives for that state until the next election.
This gets you the advantages of proportional representation while still preserving voting for specific individuals in the primaries, and for a known set of individuals in the general election. voting for president+vice president serves as precedent for voting for multiple individuals as a single block.
There is a pretty common multi-winner system that allows for voting for individuals. Single Transferable Vote (also sometimes called multi-winner ranked choice voting, though this is inexact). It’s used in Ireland and Cambridge MA, among other places.
One option is to scale up the number of districts - the more districts there are, the harder it is to pack everyone from the minority into a small number of districts.
Expanding the House would also have the benefit of making representation more proportionate between states (as in, every congressperson represents the same number of people). It would also make the Electoral College more proportionate, since the +2 votes for senators would be a smaller portion of the whole.
The drawback is mainly that the legislature becomes bigger and harder to organize. Also, this would require a constitutional amendment to happen, unlike a change to how districts are drawn.
> Also, this would require a constitutional amendment to happen, unlike a change to how districts are drawn.
Actually, it wouldn't! The size of the house is set by ordinary legislation, not the constitution. The current system dates back to the 1920s. Prior to that, the house was regularly expanded over time as the population grew.
There was an attempt at a constitutional amendment in the early days to control the size of the house, but it wasn't ratified by enough states and never took effect.
Given the assumption that self-sorting works quickly enough to be a problem, has anyone looked into reverse-gerrymandering? Since lines must be redrawn periodically as population fluctuates, make it algorithmic with some randomization. You could even do checkerboard districts if you want since GIS makes it relatively trivial.
Been playing this for a while! It has made it into my daily puzzle rotation alongside NYT games and enclose.horse. Very well done.
This is so much fun! (With a TIHI for how bad this is for democracy). I've added this to my daily game stack.
It's really good
This is great (And depressing)
I wrote an 11k word introspective (arguably self-indulgent) post about struggling to be an amateur essayist (dealing with perfectionism, procrastination/writer's block, ADHD symptoms, impostor syndrome and other such issues) and why finishing posts in spite of said struggles (whenever that does happen) still leads me to satisfying results. I also include some speculative (i.e. untested) advice (to myself) on how to overcome these difficulties.
I'm glad to have gotten all of that off my chest. If you find these problems relatable and wish to read someone reflecting on them (without promising solutions), here's the post: https://soupofthenight.substack.com/p/having-written
Related: I've noticed that most writing advice focuses either on super broad-strokes stuff (how to come up with an idea, how to move past the blank page) or on fine details (which adverbs should you use). Meanwhile, there's a "missing middle" of medium-scope advice (e.g. now that you've a lot of some brainstorming/freewriting, how do you actually turn that into a structure?). Any recommendations?
Read More. If you're having trouble writing, read the same essay over and over again. Tear it apart, shred it in your mind, figure out how they communicated. Do that over and over again, using the best writers you can.
I read-skimmed your piece (am at work) and it resonates quite a bit. Writing is a funny game to play. It's very illegible, very vague. Ive been at it for years now and have always been unhappy with the results but kept going at it because there's something fascinating about: the creative act of structing idea-soup into an essay; and also how it somehow sparks new ideas seemingly out of thin air.
My writing improved when I admitted to myself finally what my goal is: to clarify my writing and put into a good structure. That somehow eased a lot of the ambiguity and perfectionism seemingly over night. I also came to understand writing better as a physical act by comparing to activities I like such as biking or hiking: how come these feel so effortless and engaging while writing felt like punishment? I never found a direct answer but I began following Peter Elbow's advice to just free write every day for ten minutes, which took away a lot of the pressure, and made writing a little fun actually.
But then I realized it hit a wall with editing. I can output hundreds or thousands of words, but the really hard part is making something good out of it. Should this be an essay? A story? A mixture? What persona should I assume? Which of these arguments are relevant? Which are actually bad? I think this is the real writing work, and it feels almost recursive: I sometimes write about the meta problem and figure out the answer.
So I don't have straight advice to share, just that writing is a weird beast, and there's something indescribable about it. I feel some people just "get it" like I did biking, and others absorb it through osmosis, just by writing in the same room as other writers, and others keep pushing the huge rock uphill--one imagines amateur substack writers happy.
I find myself hitting that same wall during the editing stage. Freewriting is great, and I probably couldn’t write at all without it, but post-freewriting writer’s block still exists.
(“writing is illegible” is a very funny way to formulate that idea lmao)
Whats your editing approach like?
Haphazard.
I sometimes have an early outline, then I turn it into keywords/bullet points/disconnected sentence fragments, then I turn those into sentences, then I polish those sentences. But I tend to do these things out of order, or jump back and forth between different steps.
When polishing, I aim to maximize conciseness (rewrite sentences to be shorter) and cohesion (use transition signals and references to earlier/later parts of the essay), and to avoid repetitiveness (basically the primary school advice to use synonyms instead of repeating the same word 20 times in the same paragraph).
When unsure about something, I highlight it and leave it "for later".
When I want to add additional information but aren't sure where to insert it, I add it to a footnote (I do this A LOT)
I'm not a native English speaker, so I don't always trust my intuition when it comes to grammaticality; I sometimes google specific phrases to see if others use it the same way I would.
As I said in the post, I'm generally happy with the final result and consider my posts fairly well-structured and well-edited, I just wish there was a more efficient and less unpleasant way for me to get to that point.
I'm agreeing that sometimes you have to do a lot of something to figure out what you're actually aiming at.
One I didn't see in the list is giving yourself a word ceiling. Flash Fiction is where you sit down to write a complete story in, like, 300 words or something; you can do that with essays too. Decide to make a point in 1,000 words, like, every four days or something. Then you've got peanut snacks between the 11,000 word big dishes.
...(the term 'alief' is totally a cockroach cookie.)
I tried to read it and I didn't get very far into it. Sorry.
No worries, thanks for giving it a shot!
Genuinely pleased by the cryptogram solution news. If AI could stick to working out old problems like this, I think we'd all be less fearful.
As to the poultry farms, I think it's "do you genuinely want to improve animal welfare, do you want to abolish all meat-eating so it's in your interests if conditions remain hellish so you can guilt the public into giving up meat, or do you want to preserve your moral purity and not compromise by co-operation so conditions do remain worse than they could be?"
Myself, I think "make things better is the aim so okay work with the factory farms" but I'm not a vegan.
> Genuinely pleased by the cryptogram solution news.
Do the https://en.wikipedia.org/wiki/Voynich_manuscript now.
> Myself, I think "make things better is the aim so okay work with the factory farms" but I'm not a vegan.
That does sound like the way forward that can actually buy something. Factory farms are responding to the fact that people want animal protein, as cheap as possible. Realistically, all those animals are not going to stop being there. But you can push for better regulation, ex. the EU already outlaw the worst kind of cages for laying hens, and is apparently planning to do another round. And you can push for these things to be actually enforced.
And if you live in an area with a farmers' market, you can do what I do and buy meat directly from known ethical farmers.
Some new articles from my Central European Blog:
When Worthless Cash Floated in the River - https://komando1.substack.com/p/when-worthless-cash-floated-in-the (Czechoslovak currency reform of 1953)
Remembering the King Who Drowned - https://komando1.substack.com/p/remembering-the-king-who-drowned (The Battle of Mohács and its repercussions)
The next article, planned for this week, will be about Jan Masaryk and his likely assassination. His body was just exhumed after almost 80 years in the ground. Given the high level of underground water, it waxified (adipocire), so it can actually be examined for relevant traces of drugs etc.
Semi serious question: Do you plan to talk about “The Unbearable Lightness of Being?”
Loved the film, the Kundera book less so.
Maybe one day. It is a quintessential work of our literature.
I was heartened by the refusal of Tomáš to sign the document retracting his political criticism.
Subscribed! Reading from Spain, Eastern Europe is ridiculously close but still a bit of another world.
Thank you :)
Eastern Europe? Shots fired.
Ehhh sorry that's just the traditional term around here for the other side of the old Iron Curtain. Emphasis on Europe, as in, it feels pretty much like home when you travel there.
When writing some AI test cases, one of my prompts turned into a science fiction concept that might be interesting. For reasons that will be apparent, I won’t reproduce the actual prompt.
The idea is that in the near future, we have AI robotic sex companions, but the early adopters for this technology are furries, so all the sex robots are in fur suits.
Deepseek, being its usual self, nitpicked my world building. Ok, sure, DeepSeek, I can write a prompt for you that goes into more detail on how the furry sex robots are supposed to work. Deepseek also accused this of being an example of Fetish Fuel Future (see TV Tropes). Well, yes, that was genre I was aiming at.
I will throw the concept out there in case anyone feels like using it. (The sex robot idea has been done to death by Chobits, Her, etc., but the idea that the furries are the people buying the tech puts a twist on it).
Ehh, I prefer the SMBC future where all the (female) sex robots look like walking refrigerators or R2-D2 in a wig. (The male ones tend to be drawn in a relatively realistic style, which I surmise might be downstream of Mrs. Weinersmith's editorial input.)
It's been done: https://www.myhenry.com/hetty-compact-het160
I believe her title is actually Dr. Mrs. Weinersmith
Yes, I think she's a parasitologist or something, though the "Mrs." part is the relevant one here.
Sure, but "Mrs. Weinersmith" just doesn't sound quite ridiculous enough.
Yeah, but English isn't German: titles don't stack like that.
So... Five Nights In Freddy?
I think to myself, maybe someone has already written an erotic FNAF … and, foolishly, do a search of AO3 to discover this is an entire genre.
As with everything else, I would expect furries to be making community-led aftermarket modifications, rather than any major company being willing to be seen catering to them directly. Even in a case where they would be your natural largest audience, the embarrassment in front of investors hurts your business more than selling them a 'conventional' but easily modable product.
It also changes the mood of a fiction using the concept.
_Her_, only Theo is a furry and Samatha is in a fur suit, shifts the whole thing more towards comedy.
Blade Runner 2049 is just not the same if Ana de Armas is in a fur suit the whole time, etc.
Oh, wait … Steve Gallacci has actually done Blade Runner, only everybody is an anthro animal.
But what I’m going for is only the replicants are anthro animals.
At Euzoia we aim to write 80/20 evidence-based self-help. This month we finished the topic "work". See here for all previous posts (body, mind, home, money, friends, love) https://www.euzoia.org/p/euzoia - next up are play, spirituality, and meaning
Don’t follow your passion - How to find work you’ll enjoy
https://www.euzoia.org/p/work-dont-follow-your-passion
Act Agentic - How to navigate your career
https://www.euzoia.org/p/act-agentic
Become the best version of yourself - How to build valuable skills
https://www.euzoia.org/p/work-become-the-best-version-of-yourself
Help others - How to build a strong network
https://www.euzoia.org/p/work-help-others
Stoicism against stress - How to lead yourself, your team, your boss
https://www.euzoia.org/p/work-stoicism-against-stress
> "You don’t have to wait for friends to reach out. You can be the change you want to see in the world. You can reach out. You can be the spark. And people will love you for it."
From https://www.euzoia.org/p/friends-be-the-spark
I built a mobile app to help people reach out to their friends without feeling like they are imposing. It's called Beacon. Kinda like a bat signal, you light your beacon and it pings all your friends at once. The first to answer connects 1-1 for a phone call (and the signal quietly drops for everyone else)
Ah that's a cool idea <3 How did adoption go? Did people start using it regularly?
adoption is still going :P
the app is in beta (still on TestFlight for iOS), very useable but definitely adds a layer of onboarding friction
Apollo is great. Filled with many of the smartest, hardest working, most mission-aligned people I’ve ever met. You should apply! (I work at Apollo)
Out of curiosity, do people ever ask whether there's a connection with the asset management company bearing the same name?
Not in my experience
Thank you for your service!
(also, nominative determinism - "help [make] new man"?)
🫡
If you're using the 32 Proquiritations from the University of Chicago transcription of the 1653 text, I think the numbering is wrong; its 7 is actually number 1; its 4 is actually number 10. I don't know why.
I checked the first two with the version at this link, and they line up: https://archive.org/details/worksofsirthomas00mait/page/412/mode/2up
1. "Seeing from the creation of the world" - fifth word begins with "O"
2. "Seeing the grant of what is demanded in this Treatise" - third word begins with "G"