It is well known that LLMs are bad at questions like "Which is bigger, 9.11 or 9.9?", "How many r's in strawberry?", etc. The cause of this is also known to be tokenization.
Is it fruitful to draw an analogy between these questions and optical illusions? In both case, we exploit quirks in how the perception system works, in order to force a wrong answer. Are there more aspects of LLMs that this analogy helps explain?
I suppose it comes down to whether you think that these questions for LLMs and optical illusions for humans are an artifact of training data or architecture.
LLMs operate by sequentially sampling a token out of a distribution conditioned on the previous text it has seen and generated. For your example, this process doesn't trigger a boolean "is 9.9 > 9.11?" operation but rather something like: P("9.9" | "Which is bigger, 9.11 or 9.9?"). Folks with experience in LLMs will know that this isn't entirely correct, but for the sake of scientific communication it is mostly correct.
I'd argue this is an architectural problem. Sure given an infinite amount of training data perhaps this (usually stochastic) sampling operation could mimic the boolean operation perfectly, but the formulation of the model is not conducive to supporting that.
On the other hand, my understanding of optical illusions is that they are based on exploiting our brains making "processing shortcuts" they have learned based on visual data from the real world. I am not a Neuroscientist, but I'd hypothesize that if we had things that looked like optical illusions in the real world, our brains would learn to make them not optical illusions- hence it is a training "problem."
What do Michael Bloomberg, Zohran Mamdani, and Nick Fuentes have in common? All are successful people who worked hard to get where they are. All are subject to a never-ending, high-pitched stream of "it's not faaaaiiiirrrr." This similarity is increasingly important as one side of the political spectrum becomes defined by the glorification of mediocrity.
Let's start with Fuentes. A while back he said, in a video I was not able to find, that he was frustrated with the people in his movement who act like he was given his position on a platter, when really he got it because he put in the effort and was good at drawing attention to himself. Perhaps this accounts for his increasing centrism - a common arc for far-right leaders. He's attacked conservatives for being "openly hostile to all the good things about liberals" and being "low-IQ hillbillies who take pride in being simple and hate the rich,"[1] and is now supporting Gavin Newsom for President.
Onto Mamdani, on track to become the mayor to eight million people at 33-years-old. An intelligent conservative movement would look at his success and ask how they can emulate it. Instead, one influencer on X called him a "failson"[2] perhaps because he spent his early adulthood taking low-income political jobs that put him on the path to be Mayor of NYC, instead of being the dentist who works out of a New Jersey strip mall. Another attacked him for having never having "done a single day of real, hard work" the way guys on a construction site do.[3] If accurate, it seems to have worked out for him. Maybe conservatives should try to emulate his success, put their kids on the path to take political power and undo Mamdani's harm, instead of going on about how much better they are because they hauled cement that summer. Another Tweeter bitched that the attendees at his recent scavenger hunt "have the same vibe as Disney adults."[4] Instead of telling Republicans how they can replicate a successful voter mobilization strategy, she'd rather they smugly feel superior and then lose elections.
Does all of this matter? I argue it does. Our "pro-capitalist" party could be talking about how much richer West Germany was than East Germany, how capitalism can make you rich, too. Instead, it's telling people they should be proud of being poor. The Mike Bloomberg's of the world don't like Mamdani's ideology or policies, but he doesn't give off the stench of resentful loserdom the way so much of the Right does. The result will be a Right that finds itself increasingly struggling to recruit competent people and raise money. On the bright side, this dearth of EHC provides an opportunity for high-IQ, high-agency people who are willing to hold their nose and enter Republican politics.
The right is just following the left's lead into identity politics. They very much used to make ideological arguments in favor of capitalism but the last 20-30 years have demonstrated that tribal politics trump well-reasoned arguments so the right looked around, realized that white people still comprise a majority in this country, and so decided to go with that.
> Onto Mamdani, on track to become the mayor to eight million people at 33-years-old. An intelligent conservative movement would look at his success and ask how they can emulate it.
...He's a populist. The right's already doing that, they don't need to copy him. He still needs to be attacked, because he's opposition.
A common explanation for political polarization in the US is that most congressional races are non-competitive, and therefore politicians only need to worry about losing to someone more radical from their own party, leading to them preemptively becoming more radical themselves.
But this argument seems like a leap in logic. Even if they only need to worry about same-party challenges, why do they only need to worry about more-radical challengers? Why don't they also need to worry about more moderate same-party challenges who would force them in a less radical direction?
Fair point, but then we might expect the average winner to be pretty close to the preferences of the primary electorate, at which point it should be equally easy to challenge them from either direction.
In practice I think that electability concerns and some kind of influence from pragmatic centrists might make the candidates tend to be more centrist than the primary electorate would really prefer, though. And in that case it would play out as described.
We are actively working on improving the conversation experience on poor internet.
Here is how it was made: We had the best psychologists in the country talk us through every conversation they would have for each insomnia case, even down to the exact word choices, analogies, and conversation pace when teaching these well-validated techniques to their patients. We then wrote hundreds pages of instructions for generative AI to follow in its weekly sessions, so that it only responds with human-crafted material, while keeping the conversation collaborative, interactive, and highly personalized. This kind of dynamic, personalized dialogue is what makes in-person sessions effective at improving adherence, which remains the biggest challenge in this gold-standard, first-line treatment.
I truly believe Sheep to be the best digital sleep program based on CBTI, and so do our users.
This is not a proper argument. It is almost certainly intentionally misleading. This does not take long to establish.
Two quotes:
"Instead, Hitler gained power not because a majority of Germans agreed with his aims, but because key leaders, most notably Franz von Papen, thought they could use Hitler to achieve their aims, that they could sand off all of the nasty rhetoric and instead employ Hitler as a cudgel (against the socialists)."
"What I want to note here are two key commonalities: First, fascists were only able to take power because of the gullibility of those who thought they could ‘use’ the fascists against some other enemy (usually communists). Traditional conservative politicians (your Mitch McConnell and Lindsey Graham types) and conservative business leaders (your Elon Musks) fooled themselves into believing that, because the would-be tyrant seemed foolish, buffoonish, and uneducated that such an individual could be controlled to their ends, shaped in more productive, more ‘moderate,’ more ‘business friendly’ directions."
This is misleading because it misframes the decisions of German moderates and conservatives by downplaying Communist violence at the time. The Nazis during the Weimar Republic were certainly violent and they advertised their own street violence but every contemporary account I've ever read has Communists being just as violent, if not more so, in the street, and it takes less than 5 minutes to mass documented violence by communists at this time, such as the Reichstag Bloodbath.
Like, you can just read a list of political violence here:
A lot of political violence by a lot of different actors.
Second, German moderates and conservatives were not choosing between the Nazis and, like, modern Scandinavian socialists. Hitler's '32-'33 rise is contemporary to Stalin's genocide of the Ukrainians in the Holodomor. There's a very small list of people who can claim to be as evil as Hitler but Stalin is right there at this time, at his most powerful and most Stalin. I haven't read too much on Van Papen in particular, maybe he was genuinely naive, but other contemporary accounts are very clear and strident in their denunciations of contemporary Communism and Hitler's anti-Communism shines through "Mein Kampf" almost as strongly as his antisemitism.
Maybe Devereaux has insight here that I lack but I doubt it. From memory, he's primarily a classicist (like, Rome and Athens and stuff) and when he's stepped outside that field into areas I have some above-average familiarity (like Qing history) he has not acquitted himself well. I suspect that's what he's done here.
What the hell are you talking about? You are blaming the Reichstag bloodbath on the communists? It was paramilitary police killing socialist protesters with machine guns and hand grenades...
> Independent and communists, on the other hand, emphasized that the shooting had been done for no reason and without warning. It is unclear whether the warnings existed. Almost all the dead and injured were found south of the Reichstag, on the opposite sidewalk and in the adjacent zoo, according to reports from various sides. There, on Simsonstrasse, the crowd was at least four meters away from the police. So there were no violent attacks during the storming of the building. Most of the victims were hit here. After the shots broke out the crowd fled in panic, the Sipo fired several more minutes with their rifles and machine guns. Nowhere in the sources claims that demonstrators would have been shot back.
I think it’s kind of silly to split hairs arguing whether something meets Mussolini’s definition of fascism.
In “Fascism — A Warning” Madeleine Albright argued that the term has come to have a general sense. Written during Trump’s first term, she wasn’t saying he was a fascist then, just giving a heads up.
“In the book, I try to argue that fascism is not an ideology; it’s a process for taking and holding power. A fascist is somebody who identifies with one group — usually an aggrieved majority — in opposition to a smaller group. It’s about majority rule without any minority rights. Which is why fascists tend to single out the smaller group as being responsible for or the cause of their grievances.
The important thing is that fascists aren’t actually trying to solve problems; they’re invested in exacerbating problems and deepening the divisions that result from them.
They reject the free press and denounce the institutional structures within a society — like Congress or the judiciary.
I’d also add that violence is a crucial element of fascism. Whatever else it is, fascism involves the endorsement and use of violence to achieve political goals and stay in power. It’s a bully with an army, really.”
A bully with an army was the key point.
—————
October 14, 2024 — CNN
“Former President Donald Trump suggested using the military to handle what he called “the enemy from within” on Election Day, saying that he isn’t worried about chaos from his supporters or foreign actors, but instead from “radical left lunatics.”
“I think the bigger problem are the people from within. We have some very bad people. We have some sick people. Radical left lunatics,” Trump said told Fox News’ Maria Bartiromo in an interview on “Sunday Morning Futures.”
“I think it should be very easily handled by, if necessary, by National Guard, or if really necessary, by the military, because they can’t let that happen,” he added.”
IMO part of the reason it is worth splitting hairs on definitions is because vague, general definitions like the one you quoted lead to motte-and-bailey arguments where individuals are using the terms to imply all the super awful things about a label that aren't part of the much more general definition.
Trump doesn't even seem to really match the definition you gave, as I don't see how anyone can accuse him of not wanting to solve problems. We may disagree with his solutions but he definitely seems interested in solving the problems.
This still feels a bit self-serving, though. Fascists are trying to solve problems! Hitler and Mussolini were not, in fact, *just* trying to placate their base and consolidate their power. Now, the fact that Hitler's idea of 'problems' included things like 'Germany doesn't control Poland's farmland' and 'Jews exist' was, of course, a bit of an issue. But the notion that either he or Mussolini were invested in 'exacerbating problems' or 'deepening divisions' after they had seized power is entirely wide of the mark. Hitler, in particular, was not a cynic who pretended to sincere antisemitism and grandiose visions to rise to the top. The world would have been better off if he were! The problem was that he was a genuine fanatic!
Except that behind those problems (Poland and the Jews) was the deeper problem, at least according to them, that Germany and Italy had lost their national identity and sense of purpose in the aftermath of World War I. Between the Depression, the Treaty of Versailles, and the general rising tide of nihilism, people had lost the sense that the future was glorious, and were starting to experiment with things like socialism to at least fill the material void since the spiritual one was seemingly irreparable. The scapegoat was the Jews, but the actual solution offered was perpetual conquest. Aligning your identity wholly with the fatherland, throwing yourself gloriously into battle, and dying if need be (but really killing your enemy and taking his land) was the thing that filled that spiritual void and gave everybody purpose. In other words, I believe you that Hitler was a fanatic, but as we've seen, his fanaticism could only be realized by a never-ending string of invasions, because his vision of what the Third Reich was could only ever be fully realized in a time of war. If Hitler had tried to govern and enact policies without the militarization, without the constant expansion, people might have grown dissatisfied. And of course, that leads you to go to war with the world, which obviously created more problems than the Nazis could solve.
Colonel Bogey was the theme tune to an old British kids TV show called “The Machine Gunners”. Had the probably unintended consequence of a whole extra generation of kids learning about the rumored testicular deficiencies of Nazi leaders
This gave me pause for a moment. I think the best way to rescue the utilitarian view is longtermism. There are a lot of long-term outcomes the utilitarians can effect that are better than the sadists alone would choose to, since they optimize for utilitarian goals while the sadists optimize for hedonistic/sadistic goals.
I think the best utilitarian strategy would be to aim for trans-whatever-species-they-are-ism in the long term, and global control in the medium term.
If the dilemma is rescued by stipulating a stronger form of inability to interact than a vast sea and an inability to meaningfully change the long-term future, then I think the problem just simplifies to “are the utilitarians’ lives worth living in isolation”, since the other species no longer matters. That’s answered by whether or not they’re able to structure their lives such that their equation turns out net positive.
My scenario’s too flimsy to capture what I was after, which is how from a utilitarian perspective, the existence of cruel, immoral actors who are resistant to suffering might be preferable to the existence of kind, moral actors who are prone to suffering.
I'll be in San Francisco in a month or so for a wedding and I'll have a few extra days to explore. Does anyone have any recommendations for restaurants/cafes/neighbourhoods to check out? I've already done most of the tourist activities there, so just looking for any walks/neighbourhoods people enjoy
Go to a show at the Symphony, Opera, or a SF Jazz, go out in Hayes Valley before/after the show. Few different cuisine/price range options for dinner that are all great: RT Rotisserie, Dumpling Home, Doppio Zero.
I've been thinking about emergent misalignment (EM) and more generally emergent behaviors in AI. Some unwanted emergent behaviors may likely be due to training data. For instance, a leaked document from META's chatbot showed that they were giving false medical information and making racialist and/or racist arguments. Of course, there's a huge amount of quack medicine, quack nutrition, sketchy medical studies, and medical conspiracy theories circulating on the Internet, so it's hard for me to see how the training data could *not* get contaminated by quackery and conspiracies. Likewise, the MAGA and the HBD folks have a very vocal presence on social media. Again, for me, it's not surprising that the pseudoscientific pronouncements of the Lynns, Cremieuxs, and Sailers of the world get sucked up with the training data and are then spat out by LLMs that have no capability of evaluating scientific research.
Of course, then there are the weird and unexpected EMs. From the same leaked Meta document, Meta’s AI chatbots, during internal testing, allowed romantic or sensual conversations with children. It's unclear to me whether Meta's alignment policies allowed romantic/sensual discussions with adults, and it just swept children into its feel-good sensual vortex. But it would be deeply weird if it were focusing its romantic/sensual wiles on children and not adults. Does anyone know more details about this?
While EMs can be disturbing or even dangerous, I don't buy into EY's handwringing about AIs being existential threats to humanity. The current crop of LLMs haven't evolved beyond stochastic parrothood (per Emily Bender's 2022 observation of their language behavior).
According to Kaczér, et al, KL-divergence regularization toward a safe reference model offers the best defense against EM...
I don't claim to understand how KL regularization works, but if the empirical results show it does, I'm willing to accept that they know what they're talking about. But, given that most LLMs today are trained on the publicly available internet (which includes a good amount of unsafe, biased, dangerous, and harmful content), I suspect that safety measures will have limited effectiveness because of the impossibility of fully curating the training data. So, I wonder if LLMs won't continue to feed us bullshit even when the alignment problem is solved (if it gets solved).
Is it just me, or are all children's learn-the-alphabet toys bizarrely bad?
I ordered https://www.amazon.com/Melissa-Doug-Alphabet-Sound-Puzzle/dp/B0158IMAWO, a puzzle, where you put a letter into the slot and it speaks the name of the letter. But the slots use a motion detector, so if you put the letter on top of the wrong slot, it says the name of the slot rather than the letter. And the slots don't really look like letters and have pictures of objects (eg an apple for A) instead of the letter. So if the child puts things in the wrong slot, they'll learn the wrong letter associations. Also, it says things like "A is for apple" instead of the letter name "A".
Then I ordered https://www.amazon.com/Electronic-Alphabet-Learning-Interactive-Educational/dp/B0CXDK7NSV. But the pictures are much bigger than the letters (the letters are barely visible) and if you press the picture it says the name of the picture (eg "apple") rather than the name of the letter. Also, there are lots of different modes and if the children press buttons randomly it will quickly start doing random things. Also, if you haven't used it in a few minutes it starts complaining and telling you to come back.
Is there any alphabet toy which actually gets children to focus on learning the names of the letters, and which is robust against children who use it slightly differently than intended?
I can’t answer your question, but thought I’d toss out a coupla ideas. General thrust of what I have to say is that for small children I think clarify and purity in phonics instruction (what the ideal gizmo you’re searching for would have) weighs less than the teaching material’s fun and engagement factor. Teaching approaches that maximize the latter have of a lot of “noise,” but the right kind of noise creates engagement, and kids are very good at extracting information from noisy experiences. The language they hear is a hugely complicated chaotic mess, harboring patterns, all with exceptions — and yet they learn the basics without systematic instruction. So here are a few teaching ideas consistent with that belief.
Let them watch Sesame Street. When I was college-aged I taught nursery school for a couple years, and about 30% of the 3 and 4 year olds could read some, and some could read well. Virtually all had picked up the skill from watching Sesame Street.
Day-to-day life is what’s interesting and delicious to them. Play little games based on day-to-day life that involves their reading things. Scatter them through the day, at times when you’re with them.
-When my daughter was 2 I made a grocery shopping list for her with an image of each item next to the word. Few of the words on it were perfect phonic examples, where the pronunciation could be constructed purely by combining letter sounds, but that didn’t matter much. She learned to read “banana” and “soup” anyhow. She’d check off the items on the list she wanted, then help me find them when we shopped. Later I removed the images.
-Find the item: Gave her a one word written clue, like “bed,” and leave something interesting — a little toy, a snack — in the spot. (Or you could leave a little token, and there could be some treat won for getting 3 tokens.)
-Silly games with alphabet blocks. Yup, dumb old-timey alphabet blocks. Come up with a question, like “where do you sleep?” (Or let her come up with one.) Write a silly answer that rhymes with the right one, e.g. “head,” using the blocks. Asked her to fix it using the blocks. Or she can make a sillier answer. If she makes non-word, like “zed” that’s funny too, if you’re little. You cry, “oh I am so tired! I am going to get in my zed, zed, zed. Or maybe my ked.”
I’m with you here. A lot of children learning toys seem to be low quality and lack thoughtfulness in their product decisions. We have alphabet flash cards where some of the letters have the worst/least useful word associations: “b for brownie”? cmon… what about “ball”. These were the best flash cards I was able to find in Amazon with 30 minutes of searching. Definitely room for improvement, there’s probably a market here
Speak & Spell is the best one I can think of off the top of my head. There's a remade version that costs about $25, or you can find used vintage models from the 70s and 80s on eBay.
It's aimed primarily older kids who can read well enough to play the spelling games, but it also has a mode where pressing a key will say the letter out loud, and a mode where it shows you random common words and reads them aloud to you. My daughter played with hers quite a bit as a toddler, lost interest in it for a while, and then rediscovered it when she was six-ish.
I think the Speak & Spell was a relatively minor contributor to her learning to read. That was mostly from us reading to her and from her going through a period of wanting to watch Alphablocks and various phonics-themed children's music videos as much as we would let her.
I try to spend a little time with our son every day on this toy, or other similar books or toys, pointing to a letter and saying the letter name. It didn't take him long to figure out what the exercise is about though he is still struggling to produce the correct sounds.
I've been thinking a bit about the future of AI as it currently exists, and have come to a rather unwelcome conclusion.
"One machine can do the work of 100 ordinary men. No machine can do the work of one extraordinary man." -- Elbert Hubbard
LLMs can do an adequate, though never superlative, job even in thinking professions, such as programming computers, producing artwork, legal writing, etc. They will get better, but the way they work, token prediction, isn't actual intelligence, and can't make the leap to produce truly new things. We need another advance in the technology to do that.
So only the most creative/genius people will be able to produce output machines could not. The great majority of humanity wouldn't be able to produce what the machines could, or at least not as fast.
Some jobs today still are tough for robots to do, such as picking strawberries. People are still more cost-effective at these, but as technology improves, even without a huge breakthrough, these human jobs will be replaced by machines. Nonetheless, jobs like these fall into the category of things "100 ordinary men" (or women) can do.
When machines are better than the great majority of people, what purpose do people then serve? My conclusion: creating more people, hoping some of them will be the "extraordinary" people that cannot be replaced by machines. And to support these hordes, machines will do work to keep society running, effectively putting everyone on Universal Basic Income (UBI).
If you're on UBI, what incentive do you have to determine whether you're extraordinary? After all, to find out will require a lot of effort over many years, whereas you could just...not do that.
> If you're on UBI, what incentive do you have to determine whether you're extraordinary? After all, to find out will require a lot of effort over many years, whereas you could just...not do that.
I see this as the same question as "why not just wirehead?"
Sure, a few might choose to do that. Generally, though, humans find meaning in more things than lazing about all day partaking in simple physical pleasures.
"My conclusion: creating more people, hoping some of them will be the "extraordinary" people that cannot be replaced by machines."
Ah, but with embryo selection and genetic engineering, we won't even need all those people to randomly roll the genetic dice the old-fashioned way. We just take all the extraordinary people, extract their eggs and sperm, pick the best embryo(s) of the bunch, let those develop all the way through into adult, rinse and repeat.
Eventually we may have enough superior genetic material we don't even need the adult specimens, we can just clone etc.
Aren't you happy to gaze into the future and see this outcome?
> They will get better, but the way they work, token prediction, isn't actual intelligence, and can't make the leap to produce truly new things.
LLMs are made of math. You're made of meat. There are otherwise moderate-but-not-yet-strong similarities between how the two work. In particular, portions of the meat appear to be doing prediction as a key component of how it navigates the world. There is undoubtedly more that the meat is doing, but many of the core atoms _are_ prediction.
As a personal anecdote, I currently am getting superlative performance out of LLMs as far as programming computers and managing the software development process more largely. Getting this performance has only been possible for approximately the last six weeks, is only possible currently with Claude Code, and even there is definitely _not_ the out-of-the-box experience, requiring considerable customization. I estimate that superlative programming performance _will_ be the out-of-the-box experience in no more than six months (no great breakthroughs needed, just blocking-and-tackling productization).
>They will get better, but the way they work, token prediction, isn't actual intelligence, and can't make the leap to produce truly new things. We need another advance in the technology to do that.
I don't think that's true, in that we have no way of knowing right now what is "actual intelligence" and if token prediction is not actual intelligence. People saying those kind of things usually make predictions that gets proven wrong quickly, and then tend to become extraordinary at moving goalposts.
>When machines are better than the great majority of people, what purpose do people then serve?
Even with machines humans like scarcity/non fungible things, so for example concerts will still be a thing I think. Sure a robot may sing (well, they already do through autotune/playback kinda) but you are in the presence of <celebrity> and that's something a robot can't do since the robot isn't <celebrity>. I don't bet much on the "physical presence of another human", humans don't have infinite endurance/patience and "good enough with infinite endurance/patience" beats human most of the time I think, at least for many people it will.
Also, just because someone is better at something than you doesn't mean there is no meaning in doing it. See chess, go, or anyone doing a sport without having a shot at becoming the best. People like to strive towards goal and be better than the them of yesterday, that kind of things.
Another may be consumption. Maybe for some reason humans are more interesting to make products/experiences for and AI assigns themselves value based on that? Sounds super sci-fi-ish and may never happens, but also, don't humans like doing that? Lots of humans dedicate their life to making stuff, like movies, photos, tiktoks, youtube videos, sculptures, pottery, painting, etc, and people really enjoy having more of those in the world, and niches can be so narrow that there may not really be someone better than you at it. Even with AIs, total computing power of the human civilisation is limited and we have code to write, a lot of it actually.
Possibly humans can fulfil the purpose of being someone's child, but that seems to be dropping by a lot lately.
Still, I feel like this is a trend that has started around the industrial revolution and before it seems like the answer to "how can old people adapt to the new world" was "by making children and dying basically", now with less children and technology moving way faster I don't know.
What makes you think extraordinary people will be better than AI? That isn't true in chess, even though Magnus is arguably the greatest chess player who ever lived
I agree with you, but chess is a bad example, in that it's not even close to being an AI-complete problem. (People thought it was back in the seventies, but they turned out to be wrong; Deep Blue was built at a time when AI was otherwise stagnant, using techniques that don't generalize to most other forms of reasoning.)
Happy to help out with astronomy consulting. I have a PhD in planetary geophysics. I'd be willing to try to help with aerospace stuff but much less likely that I'm capable.
Edit: also possible I could be useful for climate/geoengineering stuff? I got a lot of passive exposure to that just being in a geoscience department. But I'm guessing you have plenty of volunteers.
I have long been beset by the sort of insomnia that keeps me from falling back to sleep at 3am (that is, I never have a problem falling asleep at the usual time, but have had periods of several weeks where I struggled to sleep through the night).
Back in March I learned about 'cognitive shuffling' and have since found it about 98% effective - even waking at 5am I can always get back to sleep. Used it this morning to go back to sleep at 4:30a. Not sure how it compares to CBTi, but it doesn't require an app - here's how Gemini describes it:
1. Pick a random word: Choose a neutral word with no repeating letters, such as " Pluto" or "mask".
2. Visualize the first letter: Focus on the first letter of your chosen word.
3. Generate a word list: Think of as many unrelated words as possible that start with that letter. For example, with "Pluto," you might think of "plane, poodle, play".
4. Visualize each word: Briefly imagine each word you think of.
5. Move to the next letter: Once you run out of words for the first letter, move on to the next letter of your original word and repeat the process.
I've never applied 4, I just think of random words. I'm almost always gone by the third letter of the index word. Apparently a lot of people find it highly effective for all types of insomnia, so putting it out here as a public service.
This may seem silly, but since the whole verse goes "Whistle while you work/Hitler is a jerk/Mussolini bit his weenie, now it doesn't work", I believe the weenie that's been bitten and now no longer works is Hitler's. That's my interpretation anyway
As an Italian I can add to this rich literary patrimony by citing that we instead have a rhyme going "Cosa é successo? / Mussolini é caduto nel cesso." (What happened? / Mussolini fell in the toilet).
Hello, I don't know whether this comment will be seen and/or answered, but I thought it was worth a shot. I am a long time lurker, still not comprehending a lot of what is being written in this space but fascinated by it. Im also finishing my bachelor in psychology and very interested in what are the best ways to help people with mental ilnesses. Specifically im interested in the debate of psychiatry vs psicotherapy. And which one can help people, in which way, which helps the most and which helps certain kinds of mental ilnessess and not the others. I have yet to see a thorough breakdown of this questions but this feels like the space where it could happen. The reason im posting is also that, beyond a possible career change I am contemplating, I had a discussion today with a professor which strongly favours the humanistic approach to the psychiatric one and believes most of the psychiatric one to not actually help people. Next week we are resuming the debate, but since she has decades of experience and a lot of material at hand, while I dont. i was wondering if anybody could point me in the right direction to find some good material pro and contra psychiatry and/or psychotherapies.
A good place to start is to ask an LLM which mental disorders and scenarios respond best to talk therapy and which respond best to medicine.
Here are some of my thoughts
Talk therapy can't adjust somebody's innate asthetics through which they experience the world. But chemicals often can, to some extent. Extreme fluctuating moods - excitement and mania - can be tempered by chemicals. To a limited extent people with marginally functional reward or attention systems can have improved experiences and performance with chemicals.
But many issues are also a product of a person-environment interaction. When feasible to change the environment, these can be temporarily resolved. Talk therapy can be useful to diagnose these. Also, people who are distraught, about a specific issue and need counsel, but are otherwise able to feel pleasure and pain, can be have their process and identify their thoughts process in talk therapy, sometimes this can lead to corrective actions.
For many individuals, their disorders are untreatable by either. Often chemicals can be given to assist or reduce distress with minor efficacy.
- I think there's pretty strong evidence that both therapy and psychiatric intervention (i.e., drugs) are effective at treating mental illness. In many cases, either one is good, and both is better.
- I don't know of a good comprehensive review of which things tend to respond better to therapy vs. drugs. I would naively expect such differences to basically amount to "which things have we discovered good drugs for," rather than reflecting some important underlying trend among conditions, but that's just a guess and I could be totally wrong about that.
- When I hear someone who is broadly against psychiatry as an entire field/approach (as opposed to narrower criticisms like "this drug class is overrated" or "this particular disorder is hard to medicate" or "doctors' incentives lead to them over-prescribing certain kinds of medication"), I find it hard to take them seriously. I've never encountered a convincing evidence-based case that medication, broadly construed, doesn't work to treat mental illness. Instead, people advocating against psychiatry as a whole tend to be arguing from philosophical positions like "mental illness isn't actually bad and shouldn't be treated, it's just part of the diversity of human experience" or "drugs are unnatural and it's bad to use them to modify cognition, regardless of whether it works." I think these positions are dumb.
- Here are some relevant links to Scott's writing that, while they aren't directly about your exact question, might be relevant to your interests and/or give you a sense of how the rational-sphere thinks about these issues:
I don't think I understand the specific thing you're trying to figure out. It's hard to assess the value of highly general buzzword things as opposed to specific interventions. If this "debate" is an existing thing, do you have a link to a summary of it or something?
I just finished “Rationality: From AI to Zombies.” I am curious how it has aged, but it is hard for me to judge since the writings of Eliezer 2006-2009 constitute the majority of my exposure to many areas. For instance:
* At time of writing, Eliezer had already disowned his pre-2002 opinions. Given R:A-Z completed 16 years ago, that’s long enough for him to have made multiple major revisions since. Has he, or has the Rationalist community in general, revised takes on major elements of the book?
* How has his explanation of quantum physics and many-worlds aged among rationalists and physicists?
* How has his explanation of consciousness and reductionism aged among rationalists and psychologists/neuroscientists/philosophers?
* This was written during the height of New Atheism and contains several New Atheist talking points. That movement lost popularity – in no small part due to crashing on the rocks of wokeism – though atheism is still slowly trending up. Has the median Rationalist’s take on religion, or the conception of social good, changed?
* In the final posts, Eliezer speculates about Rationalist dojos teaching the Way. ACX has meetups, orgs like 80,000 Hours exist, and Effective Altruism has grown into a movement, but Eliezer was making comparisons to the Catholic Church in terms of community-building, mobilization, and so on. Are there any dojos? Does anyone still dream this big?
The sequences cover a lot, and I’m interested in any and all updates since it was written, be it a point above or anything else that you in particular know and are passionate about.
1) He didn't realize that *practicing* rationality is much harder than knowing the theory, so he didn't put enough emphasis on practice in the book.
2) He chose impressive abstract problems [such as quantum physics] as examples, but now he thinks it would be better to talk about solving problems in *everyday life*.
3) Similarly, too much focus on rational beliefs, not enough focus on rational *action*.
4) A bit disorganized content on the web page, hopefully arranged better in the book.
5) Too much expressing contempt towards stupid ideas. Yes we should reject wrong ideas, but there is a certain tradeoff where too much mockery is bad [Eliezer does not specify how bad, my guess is that it makes it more costly to admit mistakes, and can attract the wrong kind of people]. Luckily, Scott Alexander provided an antidote to this attitude.
Points 1 through 3 would have been incremental steps towards making LessWrong more like a dojo and less like an intellectual community of nerds, but I don't really think they'd have sufficed; if you want to be a dojo I think you have to do a lot of things differently. (And I don't wish he had; it's rare for anything to succeed as much as LW did, and so it's likely that the dojo wouldn't have.)
On point 4, I think Eliezer is just wrong. The densely inter-hyperlinked nature of the original Sequences posts makes them into a trap for the unwary reader who was planning on doing anything else with their day (compare https://xkcd.com/609/), and I think this was instrumental in getting people to actually read them. Rationality: From AI to Zombies is much more of a slog. Furthermore, the blog posts are good blog posts, but the book is not a good book, because each chapter is basically a blog post and that's not a good flow for a book; fixing this would have required substantially rewriting the content to make it more genuinely longform.
Point 5 I read as basically the same point Scott made in https://slatestarcodex.com/2014/04/15/the-cowpox-of-doubt/, which seems slightly different from your interpretation, but I could be wrong. I'm not sure how much difference it made, in any case, especially since Eliezer did not exactly become less habitually contemptuous towards stupid ideas after 2015.
There have been a lot of changes in what's fashionable to talk about, but not that many in what people actually believe. Fewer people argue about interpretations of quantum mechanics, or about consciousness, or about atheism, in our corner of the internet these days, but the basic disagreements are still there, over basically the same questions as in 2009. Philosophy doesn't move that fast (and interpretations of quantum mechanics are philosophy, not physics; no new experimental evidence has come in, except in the sense of experiments that could have falsified quantum theory but didn't).
The one topic where people's substantive views have evolved is community building, because that project was in its infancy in 2009 and we've since spent sixteen years learning what it looks like in practice. Eliezer in particular has lost faith in the project, because he thought that if he taught people rationality they'd come to agree with his specific technical views on AI safety, and that hypothesis has been falsified. (AI safety has of course gained prominence in the discourse, but mostly not in the sense of widespread agreement with those specific views.) As for whether there's anything else to be salvaged, it depends on the specifics of what you're trying to do:
* As noted above, AI safety is now a big field, bigger than the place it started from, with its own exciting new developments and pathologies.
* EA has been pretty successful as a project for building rationalist-flavored world-improvement communities. (People complain about it being diluted by MOPs or captured by society's power centers, but that was always going to happen if the movement got anywhere.)
* The specific project of "learn to avoid cognitive biases"/"rationality as a martial art" has been on a downswing lately, especially with CFAR having curtailed its public-facing activity, but there are still people interested in it. I think this waning of attention was partly because a lot of people weren't all that impressed with CFAR's success at developing a teachable discipline of rationality; they kept talking about publishing their work and then not doing it. (Back when it was more of a topic of discussion, there was periodic discourse from people who'd been under the impression that this was *the entire point* of the rationality community, and felt betrayed that only a subset of rationalists were actually interested enough to devote significant time and energy to it. I feel for them, but don't in fact wish that none of the other stuff had existed, because it's valuable in its own right and probably wouldn't have sprung up counterfactually without the Sequences.)
* There are also lots of little sub-subcultures and social groups where people hang out and have fulfilling social interactions. The one I personally have the most experience with is glowfic; I think there are others I don't know, consisting of people who are into startups or circling or drugs or whatever else. The role of the Sequences was to bring together a lot of people with a lot of different intellectual interests that shared a certain common ethos; once that was finished, having one big community didn't make all that much sense, so instead we became a bunch of sub-communities with some cross-pollination.
There are also some specific claims in the Sequences that have (at least arguably) failed to survive fact checks and/or the replication crisis, but these are mostly not big general theses (though skeptics sometimes argue that these failures should lead us to distrust the entire source, including the big general theses).
"Eliezer in particular has lost faith in the project, because he thought that if he taught people rationality they'd come to agree with his specific technical views on AI safety, and that hypothesis has been falsified."
Somehow that reminds me of Martin Luther: "just print the Bible in the vernacular, let the people read it, and they will interpret it their own way, no priestly mediators necessary!"
"Wait, no, not like *that*!"
(Seemingly Luther did not, in fact, say "Every man has a pope in his belly" about the multitude of small new religious sects and sectaries springing up when everyone could interpret the Bible according to their own conscience and, believe it or not, not coming to the same conclusions as Herr Doktor Luther, but it's too good a quote to give up).
"and interpretations of quantum mechanics are philosophy, not physics; no new experimental evidence has come in, [...]"
A side note, just because:
Take a specific experiment: Alain Aspect's seminal work from 1982 (for which he got the Nobel Prize two years ago). The experiment shows convincingly that if the Copenhagen Interpretation is used, information would have to travel 10 to 50 times faster than light from cause (observation at point P) to consequence (collapse somewhere else).
Does that refute the Copenhagen Interpretation? It does to me, but proponents of the interpretation just shrug and say "yeah, QM is non-local, so what?" or "yeah, true, but you can't use that trick to send information yourself so it doesn't have consequences."
Well, obviously you can't send information faster than light, we know that from other theories, but then why is it necessary for the Copenhagen Interpretation to work?
The MWI doesn't suffer from that flaw. When you measure part of an entangled system you learn (get information) from that system, namely which branch of the entanglement ("world") you (that particular component of the superposition that up to that moment made up "you") live in. No information travels from the measured object to any far away object.
So. It is not that the Interpretation of QM is not physics (it is, at least with my definition of "physics", which has to do with understanding reality). Instead, there are deep psychological hurdles to accept an interpretation that challenges our sense of identity, even if it is the interpretation the most compatible with known physics, the one that doesn't deny objective reality, that is deterministic, and local. (It is also quite possible that historic contingencies are a better explanation for why CPI is preferred; I can talk about that too.)
Newtonian Mechanics has 3 laws:
- Existence of an inertial frame of reference
- F=ma
- action = reaction
Now, imagine that I have a new interpretation of Newtonian Mechanics:
- Existence of an inertial frame of reference
- F=ma
- action = reaction
- all forces above are invisible angels that move instantaneously from place to place and pull and push the amount required by these axioms.
Nobody would take my new interpretation seriously. It has more axioms that the original, and it violates known physics (e.g. angels travel faster than light).
But my answers to criticisms would be "you can't talk to the angels so you can't use them to send information anyway", and "I make the same predictions as you, so both interpretations are equivalent", or "well, it's just philosophy anyway."
This is exactly the situation we have with the Copenhagen Interpretation. MWI has fewer axioms, is compatible with realism, determinism, and locality, but we still insist that angels must be flying faster than light to carry information from place to place.
As I'm sure you're aware, the Bell Inequalities exclude local realism, not simply locality. No physicist thinks that rejecting MWI means abandoning Lorentz invariance.
My point is that everyone, including people favoring every interpretation, correctly predicted the results of Aspect's experiment, because the only thing he tested was whether quantum mechanics would continue to observably behave in the exact same way that it has been understood to behave since 1925. And surprise, it did. (Not to suggest that it was worthless; the theory hadn't been tested in a setting where we'd notice a speed-of-light delay, so it was good to get confirmation that this doesn't change things. But it wasn't a *surprise*.)
The arguments are all about what counts as most parsimonious, what conclusions about the nature of reality are valid to draw from what experimental observations, etc. This work is legitimately intellectually interesting and valuable (well, at least some of it is), and some of the arguments that people make are better than others; I don't think it's necessarily a mistake to have a preferred interpretation. But it's not science, it's philosophy of science.
Aspect empirically proved non-locality of the Copenhagen Interpretation, that's all. At the time his results were framed as a refutation of Einstein's local hidden variable theory, and it absolutely *is* a refutation, but they also emphasized something "wrong" with the Copenhagen Interpretation, at least if you care about locality.
As to what counts as parsimonious, it seems self-evident that the number of axioms is the first step. (But it is not in this case because, I think, people confuse "parsimonious" with "intuitive".)
What would convince a proponent of the Copenhagen Interpretation? (Or of Newtonian Mechanics with invisible angels?) Imagine that an experiment shows that the interpretation breaks another principle, let's say conservation of energy, the proponents would just say "so what, you can't use that yourself to create a perpetual motion machine, so no paradox here." Of course, you can't, because that's just not physical. But then why does your theory requires that energy not be conserved?
So yes, I understood your point well. I agreed with your first post that people have not changed their mind; I only made the side note that they wouldn't have changed their mind even with new experimental data, because the reasons people believe what they believe are not entirely determined by experimental results.
I’m not that familiar with many-worlds. Why doesn’t branching the entire universe require faster-than-light information? Since it’s just an interpretation and not new physics, it seems like it would still be somehow non-local, but handled differently?
> Why doesn’t branching the entire universe require faster-than-light information?
Because it doesn't happen at the same time everywhere. The branching starts locally (when two particles either interact or they don't) and from there it propagates outwards at the speed of light. Except that there are zillion bubbles like this spreading simultaneously.
In the famous cat-in-the-box experiment, as long as the box is hypothetically isolated from the rest of the universe, there are two universes in the box but only one universe outside. When you open the box, the branching continues to propagate outside the box.
As a visual help, instead of imagining two branches as two parallel sheets of paper moving further away from each other, imagine two stickers being peeled off each other. At the place you are pulling them, they are already separated, but sufficiently far away from that place they are still together.
But this is all just an imprecise metaphor. It does not explain e.g. how the universes *interfere* with each other, and that is the key part of the quantum physics. One universe splitting into two is a relatively boring thing, and easier to imagine. Two universes canceling each other, because they have the same contents but an opposite phase, that's the fun part that is difficult to imagine by an analogy, because it has little analogy in the classical world.
I think a common misunderstanding of MWI is that a brand new universe is created every time someone observes something. Not so. The superposition already existed, but the components weren't distinguished.
Maybe this comment I made a few weeks ago will help:
I actually disagree that MWI and CI are the same physics. They are different explanations with a different set of axioms. And yes MWI is local, as I explain in the link above, I think.
(Maybe MWI is wrong; there are *other* reasons to be skeptical; but the usual arguments are unsound.)
How does the CBT-i app stuff differ from whatever program is used by Sleepio? I found that very useful back in 2020, and they never required a prescription, though I don't know what it cost (it was paid for by my work and offered as a health benefit).
I'm not familiar with Sleepio, but I did a course of CBTi almost 20 years ago. It worked, and it's insane that anyone is charging $300/mo for it, because it is an incredibly straightforward protocol that can be summarized in a fairly short pdf. Maybe things have changed significantly in the past 20 years, and I guess I can imagine a fancy version of it that involves an automated sleep tracker, but even that wouldn't get you close to $300/mo. It's basically a combo of sleep hygiene stuff and standard CBT stuff. In a very small nutshell:
1. Track how long you are actually asleep each night.
2. Give yourself that amount of time + 30 minutes in bed every night, with rigid bedtime/rise time schedule. This is brutal but effective.
3. Do some CBT on counterproductive thoughts like "I'm never going to fall asleep" or "I'm going to be a wreck tomorrow"
4. Gradually extend your sleep window as the percentage of time you spend asleep while in bed grows.
There's more to it than just that, but that's the crux of it, if I'm remembering correctly. The only way you could possibly get to $300/mo is if a) it's a prescription program that insurance is paying for and/or b) there is an actually doctor at the other end of it reviewing your progress and providing feedback.
The point that it's simple does not directly have anything to do with the price; a thing is worth what people will pay for it, producers charge what the market will bear. If I needed this thing, I would easily pay $300/month for it if I couldn't have it for cheaper.
The right framing is, since it's so straightforward to produce, why hasn't the price been competed down? And it sounds like the answer to that is largely just that it's niche, without that many providers and without widespread enough demand to entice people who are currently doing something else into the market.
I understand how pricing works. I create software products for a living, have an MBA, etc. This is not worth $300/month. There isn't enough value created, there are credible alternatives, the marginal cost of delivery is too low, there is no defensible proprietary advantage that I am aware of, etc.
I don't know what is going on here, but my guess is that it has something to do with regulatory approval/arbitrage.
What may end up outperforming CBTi at treating insomnia is targeted adjustments to the gut microbiome, at least according to this new Mendelian randomization study:
I'm posting about this study on my blog Thursday evening. One takeaway is that to the extent that it's possible to change the levels of specific gut bacteria without creating unhealthy imbalances, these changes might help some (if not all) chronic insomniacs, roughly analogous to the way that antidepressants can help some (if not all) people with clinical depression.
In a different response to Scott's embryo selection article, I wrote a weirdly-long post on why we use exponential time discounting in general and why he's wrong not to use it when computing the cost/benefits of embryo selection
(This makes the cost/benefit ratio come out bad when used correctly, which I find weird - the benefits seem like they should be no brainers. Maybe the IQ benefits are just severely underpriced in Scott's review, so much so that even after discount they come out ahead? The health benefits might plausibly be as small as described though, especially if they're mostly just maybe-pushing people from just over to just under the clinical diagnosis line).
The recommended CBTi app seems to revolve around an AI sheep, I assume that back in 2021 CBTi did not resolve around LLM chatbots as they barely existed. Is this really the same thing?
It's not that uncommon these days for app developers to inject LLMs into things that were previously done without them, believing (rightly or wrongly) that this improves the user experience.
It's not even that, app developers often shoehorn the current trendy technology into their product to get funding. A while ago everything was on the blockchain, there was a time when VR was the hot new thing, and then there was the NFT fad. When the people writing the checks want your app to include X, it is hard to say no.
There are many CBTi apps available now without a prescription and for a similar $300 price range, although their price is less clear when covered by insurance. For example, Sleep Reset, Stellar Sleep, and Sleepio all offer something similar, and they also claim to be backed by Stanford, Harvard or science. There's also the free CBT-i Coach from the VA.
The issue I've found with these (and likely CBTi and sleep restriction in general) is they don't seem to work as well for early morning insomnia. The apps themselves may also not be as well-tailored to your specific issue as a sleep therapist, even if many of them still use human sleep coaches.
Something I think has slipped under a lot of people's radar is the impending "temporary" emergency scheduling of the newly popular legal opiate 7-OH. It seems this is inevitable and I could see it pushing the opiate crisis into overdrive. This is a compound found in small amounts in Kratom but it's obscenely powerful in pure form and now available at most gas stations where I live. A lot of addicts have switched from counterfeit heroin/oxy/whatever, that is to say fentanyl, to this compound. The advantage being it's rather euphoric, cheaper, easy to come by, strong enough to appease all but the hardest users appetites, not cut with anything else, and despite trying really really hard no one has yet managed to die from an overdose. To top it off it's very addictive both psychologically and physically and can cause physical withdrawals after just a week or two of use, so in addition to these old time addicts there are a bunch of newly minted ones. When they get cut off I fear many will turn to fentanyl and we'll see a bunch of overdoses from people trying to match the high or avoid withdrawals.
Are you comfortable saying which state you live? I don’t see anything like this in mine. In the shops where I occasionally pick up THC gummies I see kratom for sale put not this potent extract.
>With $8 in his pocket, he bought a black packet from behind the counter: a tablet sold by the distributor Pure Leaf Kratom
>Allshouse bought a 30-milligram tablet.
>As police await a toxicology report, a cause of death remains unknown, and the Henrico Police Department would only say that drug paraphernalia were found in the van beside his body.
I think we can say with some certainty this man didn't die from an overdose of 7-oh. 30mg would be a large dose for a person without any tolerance to opioids, but not enough to cause unconsciousness let alone death. There was paraphernalia, perhaps he tried to inject the tablet and died from the shot? But the substance is hardly soluble in water so in that case he most likely died from an air embolism or whatever was used to bind the tablet.
The article contains misinformation that anyone with Google should have caught, for example calling it a hyper concentrated form of Kratom, which is like calling heroin a hyper concentrated form of poppies. I predict a lot of these types of articles as we get closer to a national ban.
By the way, in case it wasn't clear, I strongly recommend people not touch this compound even in moderation, unless they are already opiate addicts and cannot quit or obtain a prescription for maintenance.
Junkies generally don't only do one drug at a time, so yeah, very likely he had a ton of other stuff in his system and this was just the topping on the dessert, as it were.
I suppose it depends on what you mean by junkie. The least functional drug users, like those unable to hold down a job because of their addiction, will usually be spending all their money on a single "drug". In quotes because many street drugs are now kind of a crap shoot, like illicit "Xanax" for example is always going to be some random RC benzo. "Heroin" is usually fentanyl but it might also be a nitazene compound or have Xylazine in it, it might even occasionally be actual heroin.
Most exceptions to this will I think be mixing in just alcohol or weed, of course a lot of drugs don't play well with alcohol, and it should go without saying none of the above applies to the wealthy.
And you're certainly right that a lot of heavy drug users are poly substance abusers. I don't have the stat at hand but I think the majority of overdose deaths result from mixing substances. That might have flipped with fentanyl and nitazenes in the picture, neither of which have any trouble killing a person flat with < 1mg
"S" was actually his ostensible middle initial ("Scott S. Alexander"). In reality, his name is Scott Alexander Siskind, as he publicly revealed in his first post on ACX.
The actual reason it's not a proper anagram is because he had to drop an N.
I still think superintelligent AI is not the danger, it'll be bog-standard stupid AI of the kind we have today (such as the thing that wiped the database) and people turning more decision-making - or even just 'get the AI to do the job' - over to it.
"Oopsies! I wiped the safety protocols for the process. I panicked, me so sowwy, tee-hee!" and there's a smoking hole in the ground where the processing plant used to be. That kind of dumb accident because we anthropomorphise *everything* and if the makers of the AI program it to pretend to be a person (so the rubes can interact with it like it's a human instead of the human it replaced), then we will convince ourselves it's a real human with the same capabilities as a real human, and we'll forget it's a box of gears, and we'll assume "well even the dumbest guy we hire to run this process would not be dumb enough to wipe all the safety protocols" and then we end up with smoking hole in the ground. And then the AI will issue a fake "I panicked" 'explanation' as to why this happened, as if it had a mind or feelings or emotions, and that is supposed to be enough to reassure us that this thing can indeed think instead of simply spewing out the script it was trained to produce.
AI is not the danger, we are, because we are stupid, greedy and vain.
Yes, as our tools get more powerful our mistakes become more impactful. That's ok because its more than compensated for by the higher average output. Should we go back to hunting with spears just because there's a chance that someone will accidentally shoot themselves?
This depends on the risk-benefit profile of the individual tool. E.g., as Kelsey Piper has explained on Twitter, doing gain-of-function research on respiratory viruses has genuine scientific value that can improve our ability to fight disease. Yet we shouldn't do it, because the extent of that scientific value is not high enough to compensate for how bad it would be if something went wrong.
What does this have to do with superintelligence? Like it seems like both situations could be true, where someone dumb uses AI badly to own themselves, but superintelligence is still the same old danger
In past discussion, you gave that link to that website: https://brght.org/iq/country/ which has inflated scores all over the board.
Averaging five largest European countries (Russia, Germany, UK, France, Italy) gets 108.35 score, Ireland is 101.95 in this dataset. So Ireland is lower just like it was in Lynn's...
Yeah, but we're "within normal range" lower, not "intellectually retarded" lower. 100 being the normed value, if we Paddies score (rounded up) 102 IQ, it means we're... average! normal!
Yay, we're normal! 😁
(If you seriously thought I was ever boasting of my gigantic brain, you got off at the wrong exit, brother).
Do you really endorse the "the danger isn't X, it's Y" construction?
For example, would you endorse "the danger isn't smart humans, it's stupid humans"? Stupid humans can definitely cause some problems (like car crashes, or equipment malfunctions). But smart humans can also cause some problems (like inventing the nuclear bomb, or lab leak). These are both kinds of problems we have to deal with! Saying "the problem isn't smart humans, it's stupid humans" is not even wrong.
In the same way, I'm sure there will be stupid AIs who cause the AI equivalent of "human error". These will cause the AI equivalent of car crashes and factory explosions. But damage is kind of limited, in that if too many factories explode, we won't use that AI. The damage from smart AIs, the equivalents of the ones that can invent nukes or do lab leaks, seems unbounded, so I am more concerned about that.
I think it's much more likely that stupid AI will be widespread, and widely used, and we'll manage to screw up a lot of small but important things, very much important when all are summed up, before we get super-duper no really it's a thinking entity AI at all, much less turned loose on the world.
More people have died in car accidents than in wars:
Only in our particular Everett branch/possible world, and only for the last 80 years. If a nuclear war occurs in the future (a risk that only started when nukes were invented, and will remain until we either all die or create a powerful aligned AGI or similar that can stop nukes), then nukes will have been very bad for the world's safety and peace. Likewise if there've been close calls where nuclear war was only averted due to dumb luck or the anthropic principle; there have been two well-known candidate such incidents and a number of lesser-known ones.
(Also, this is beside the point of AI safety, as a nuclear arsenal controlled by an unaligned AI is an obvious danger to humanity in ways that the game theory that currently governs nukes doesn't necessarily mitigate. But I don't know if you even disagree with that.)
For what little it's worth, I'm not sure superintelligent AI is possible. I think stupid use of existing AIs is already happening, and we're barely starting to see the ill effects.
The problem isn't level of intelligence or intent; what matters is the orthogonal problem of power, the ability to affect things we care about. Few would complain if you keep your hyper-intelligent (and/or hyper-malicious) agent in a box, airgapped in a secure cleanroom, akin to a bioweapons lab. In there, the agent can be as intelligent and malicious as it wants to be.
However, if I give an AI control over my nuclear arsenal, either directly or in a sufficiently trusted advisory role, then the worst case of global thermonuclear war is the same whether the AI is malicious or stupid. If "a few factories explode" is your worst case, then it's because you didn't give the agent more power than to blow up a few factories.
Yes, we are already liable to both factories blowing up (Chernobyl, Bhopal etc.) and global thermonuclear war just fine without any AI involvement. However, as the saying goes: computers make very fast, very accurate mistakes. This includes AI and is strictly worse than making slow, inaccurate mistakes.
> Few would complain if you keep your hyper-intelligent (and/or hyper-malicious) agent in a box, airgapped in a secure cleanroom, akin to a bioweapons lab. In there, the agent can be as intelligent and malicious as it wants to be.
Eliezer would like a word with you, I think.
Having an AI system in a box is safe if you do not talk to it. It is also completely useless.
To make use of an ASI, you need to talk to it, and implement some of its suggestions for the problems you are facing. The suggestions are likely going to be plausible and work on the short term, but they might have side effects which maneuver you into a situation where you will eventually release the AI from its box.
As an intuition pump, imagine a bunch of 6yo shipwrecked on some uninhabited island together with a highly intelligent psychopath with multiple professional degrees in a straight-jacket. The kids know that their prisoner is a very persuasive murderer, and have a firm pre-commitment not to let him go free. However, they also depend on his knowledge for survival. My prediction is that this situation will go poorly, and sooner or later, the psychopath will be able to get some of the kids to free him. Perhaps he convinces his guard that he is actually harmless, or engineers a situation where everyone will die unless he is released (and trusts the kids to follow causal decision theory).
In this scenario, the intelligence gap between the kids and the psychopath matters. If the psychopath is just a median 10yo, then it seems much less likely that he can outsmart the 6yo's. Likewise, humans would probably be able to keep a von-Neumann level intelligence boxed in. As far as we know, no highly intelligent humans have taken over the world, after all. But ASI is another ballpark. At best, it is "the smartest human alive ever but also 1000x faster". At worst, it will make us look like chimps in comparison.
"As far as we know, no highly intelligent humans have taken over the world, after all."
But the ones who had limited success towards that goal were not the most intelligent, rather they were extremely persuasive and/or had a particular talent that enabled them to mobilise lots of ordinary schlubs to do their bidding.
That's what I'm talking about; not the von Neumann level AI plotting our downfall, but all the 105-120 IQ people in business, government, and private users who make use of the multiplicity of new AI models on offer to replace humans, take over human work, and be the first port of call for advice, research, and therapy/emotional support.
Any one error on its own won't be immediately fatal, but add them all up? Like the guy who got advice from the AI about cutting out salt from his diet and ending up dosing himself with sodium bromide instead of sodium chloride, and that despite the AI including that this was for cleaning purposes:
This is how we're gonna do it: "well the AI said it so it must be okay and I don't need to check it out any further"/"don't bother checking the AI results, that only wastes time and reduces efficiency, if it says do this then do it, Jenkins!"
The stupid (by comparison with the Platonic Ideal superintelligent AI) AIs are *already* out there, *already* being pushed by businesses, *already* in use.
Recreate "Five Nights At Freddy's"/Ray Bradbury's "The Veldt" in your own kids' bedrooms today!
I wrote the Ollantay review. Wasn't planning on responding in the comments (seemed inappropriate), but as the Comment of the Week was skeptical I figure I can respond here. Perhaps this should have been a footnote. Ah well.
The way that the story of the rebellion is usually told is something like this: Tupac took some incredibly bold moves and gathered a huge army that would have crushed Cuzco had he continued the initiative that he already had (more than half of the curacas in Peru publicly supported him!). But instead he got cold feet or became a coward. Sometimes he's compared to Father Hidalgo, who also gathered a giant army but failed to march on Mexico City, but while Hidalgo seemed to be concerned about the giant social revolution he was accidentally leading, Tupac was perfectly content to watch the social rebellion unfold chaotically and violently. The final months of rebellion do not make sense on their own.
But then I read Ollantay, and then I learned that the early European translators and commentators of the play (Markham and his contemporaries) linked the play to Tupac Amaru. In his introduction to the translation (https://www.gutenberg.org/files/9068/9068-h/9068-h.htm), Markham calls out the Tupac Amaru rebellion directly and does in fact say that Tupac watched the play in 1775; implied is that this is a big reason for his interest in the play! And it seems as if that 1835 periodical that brought Ollantay back in the public consciousness did the same (I haven't been able to find a copy of the article and can't get to Lima for a while). This introduction doesn't specifically say that Ollantay inspired Tupac, but he primes us to read the play with Tupac in mind.
And when we do, we see all those connections that I noted in the review. And then we ask "wait, did you say this play happened *before* the rebellion?" Which all of the early translators do indeed say. And they note the provenance of this claim (priests and monks and family members). And for me, that answers the major questions of the rebellion.
So Garald is sort of correct that we have no direct evidence that Tupac saw Ollantay on stage. But we do have pretty good indirect evidence (Markham said that Tupac saw Ollantay, so somebody told him that, but he doesn't cite his source). This review was an argument that he did see it and that it changed his life. I structured the review as a story with the argument assumed, and not as the argument itself. I think it flows better that way, but it caused some alarm bells to ring and I can't be mad about that.
It may also have been an assumption on his part that he didn't *need* to march on Mexico city, the authorities there would see that their power was over because of the successful (up till then) revolution and besides the king would back him up, so why fight a battle he didn't need to fight?
Of course, it turned out he *did* need to fight it, but we don't know exactly how deep into his fantasy he was; he may well have been convinced that God was on His side, or at least destiny, and that it was inevitable he would win as the sacred rightful descendant of the emperor or something.
Here's a monograph (just 8 pages, in Spanish) about a previous (largest?) indigenous rebellion in colonial Peru, which might provide some context for the later Tupac Amaru rebellion, and also mentions the importance of taking the main centers of power (Cusco and/or Lima) for the rebellion to succeed:
"[the rebells] controlled not only the province but also the roads that transported the products that supplied Lima's markets. As Mela says, most of the mountain products arrived in Lima via the narrow paths of the Lurín and Mala rivers, and by cutting the bridges used by muleteers, the rebels could cut off Lima's contact with the interior."
"...the residents of Lima, one of whom, writing to a relative in Spain, told him that 'if they [the Huarochirí Indians] jump to Tarma, Jauja, and Cusco, the kingdom of Peru will end for the Spanish.'"
"The viceroy noted that Huarochírí was the gateway linking Lima to the interior, adding that 'if this province continues in rebellion, its proximity to this capital would become a refuge for criminals who would disturb its tranquility and cause much damage.'"
(*) Rebelión colonial: Huarochirí, 1750, by Karen Spalding
Yes, Lima. No idea why I said Mexico City. Clearly I am not cut out to be a South American revolutionary, I would be not attacking the wrong city I intended not to attack!
I loved the review and wasn't highlighting Garald's comment as a criticism, just to provide extra perspective. I've edited your response into the relevant section above.
I'm a long time lurker who finally pulled the trigger and started a blog. Here is our first post: https://whattotelltherobot.com/p/elephants-dont-write-sonnets Elephants Don't Write Sonnets, about why LLMs aren't *really* intelligent and what it would take to cross that line, writing from a robotics perspective.
The post doesn't really present any evidence for why we should think that embodiment matters beyond Rodney Brooks saying it does. Its response to all the accomplishments of the past few years seem to be that this isn't "true" intelligence, but that embodied intelligence would be, without predicting what things that current AIs can't do an embodied AI would be able to do.
I'm more likely to do modus tollens here and think of Rodney as one of the many people with bold theories of how you can't do AI unless X which were totally falsified by later events. I think his 2018 predictions, which he commendably updates every year, broadly confirm this story - for example, he said you wouldn't get an AI that "seems as intelligent, as attentive, and as faithful as a dog" until 2048. I think this shows that Rodney wasn't expecting AIs which were competent at basic tasks but lacked some kind of "true intelligence", and that positing the "true intelligence" thing is a retrospective invention intended to explain away why AIs that don't follow his embodiment theory are succeeding beyond what he predicted.
I think if you are going to blog about this, you need to address this perspective rather than taking Rodney's perspective as a given (unless you intend for your blog to be just be a discussion forum for people who are already operating within that perspective).
Thank you for this comment! Engaging with you and other readers of ACX was a major motivation for me pulling the trigger on this, so I am *extremely* excited to get feedback from you! Is it TMI to say I was dancing around our living room when I saw you had replied to this comment?
As it happens, we did address this in section 1.3 of our chapter, which is linked from the blog post and also linked here: https://h2r.cs.brown.edu/wp-content/uploads/tellexwatkins2026.pdf , but we didn't mention this part much if at all in the teaser post. We will plan another post just on section 1.3 based on your feedback here. TLDR: We define behaviors in terms of embodied language (language paired with actions/behaviors in the physical world), that robots/AIs don't yet do, along with learning architectures that point towards doing them.
I've been doing language+robotics for 15 years so I promise it's not a retrospective invention for me, although I won't deny a lot of angst as many of my predictions have been falsified in the past 5 years. :-) Regarding Rod's dog example in particular, my lab is collaborating with Daphna Buschbaum who studies human-dog dyads, so we can try to make our quadruped robots produce similar behaviors, for example to interpret human pointing gestures to find objects. I think Rod would say that behaviorally, the AI needs to be attentive and faithful *in the physical world*, which means processing high-dimensional space/time input and producing high-dimensional motor output. Not just typing to you, but walking up to you, looking at you, nudging your hand for pets, and barking at the big dog coming down the street. I am confident in saying AIs don't do this yet (while being sure an LLM could *talk* about doing it). But Rod has a chapter in the same book as us, so maybe he will say more there!
Can you play Magic: the Gathering over the phone, but without a referee and provably without cheating? Something equivalent to blindfolded chess, with the obvious complication that a lot of information in MtG can transition between various shades of hidden and public, whereas all information in chess is public.
For example, a game could start by each player assigning a generic ID to all the cards in their deck, publish that assignment except it's encrypted, and as long as the library remains shuffled, drawing the top card as normal is mathematically equivalent to drawing a random card that both players agree on by exchanging random numbers. At the end of the game, each player would reveal all their hidden choices and it can be proven that cheating has or hasn't occurred.
Is there such a protocol that covers all possible actions allowed by Magic cards, or is there proof (as in an actual paper) that such a protocol can't exist?
Yes, this is definitely possible. What you are looking for is called secure multi-party computation, and there are protocols like Yao's garbled circuits that can be adapted to not just play MTG, but to perform *any* computation with someone else without revealing your inputs to the computation.
The most straightforward construction would be to run a multi-party computation of an MTG game engine, where the RNG seed is given as the sum of two random values you and your opponent provide. There are still some implementation details that need to work out, but they are not too bad:
1. You need to run multiple rounds of computation, not just a single input/output. You just keep an encrypted state variable that is fed from one computation to the next (again encrypted with a key derived from a combination of secrets you and your opponent provide).
2. You need the game engine to provide private information to just one of the players. To do so you just have it output that information after encrypting it with a key that only one player has.
There are probably more efficient schemes which take advantage of the specific structure of the MTG game. (Also note MTG is Turing Complete)
I'd say no unless you set some limitations. How would you deal with Chaos Orb or Falling Star? You also need to assume trust on possession of cards or allowing proxies I assume, not only because of the initial state, but because Research/Development exists.
Now, if you keep it to tournament legal I'm not really seeing any problem, you'd have to keep track of card order in some scenarios, you'd also need to be able to generate random numbers for coin tosses or "select a random card" effects, but beyond that I'm not really coming up with any issues.
You could track possession of cards by having a trusted third party (maybe WotC or a tournament league) verify that you own the card and issue a cryptographically signed certificate that allows you to use it in the app.
This could still be cheated by selling the physical card after you get the certificate, but that could be mitigated with expiration dates that force you to revalidate periodically. Setting the expiration date to be short enough to be a useful control but long enough that the hassle of revalidating isn't too burdensome on both players and certifiers is a nontrivial problem. You might be able to mitigate by having common cards be taken on trust or non-expiring while only particularly rare cards would need to be revalidated annually or more often. Or you could have the expiration be a soft rather than hard limit, with you and your opponent agreeing to a limit of how old a certificate each of you will accept from the other as part of the terms of the game.
Another alternative would be to decouple app cards from physical cards and use either NFTs or a central registry to track ownership. In the latter case, you'd still use signed certificates so you can verify ownership without pinging the registry every time, but the registry would make validation trivial and make it feasible to set the expiry to a period of days or weeks instead of months or years. I think I remember someone I used to follow on livejournal talking about working on an app that worked on the central registry model, some time around 2005-2008.
Each player shares salted hashes of the cards in their deck with the other player, which the opponent's app shuffles and tracks. When player A draws from their deck, their app asks player B's app which card they draw (by hash) and player A knows which card this is. Player B can confirm that the card actually corresponds to the hash, but doesn't know in advance because of the salting.
I don't think this is a complete solution for MtG, since I seem to recall there being situations where a specific card gets shuffled back into the deck, since player B could then identify that particular card in the deck and track or futz with its position.
I think that if you shuffle a card back into the deck, you just rehash the entire library, i.e. apply new salts to the hashes (or to the revealed card) and provide these as hashes.
For a Scry N, your opponent will just give you N hashes, and you will tell them in which order you place them before and after your library.
I am unsure if there is an effect which will pick a card where neither party knows what it is, but that could be simulated by your opponent hashing the hash they picked for you.
(Of course, your opponent could also provide you with a shuffled library that way -- simply an ordered list of salted hashes of your salted hashes of your cards. Whenever you draw a card, they just reveal how they generated the next hash on the list.)
If this is doable in practice depends a lot on if the telephone refers to a smartphone or just a voice call, because most humans are not great at calculating cryptographic hashes in their head.
When I say "playing over telephone" it is a metaphor for "not being there physically" to check the legality of play [1]. Of course you'd use computers to communicate and to do the mathematical operations.
Yes, in the sense that it would be useless to the would-be cheater. The protocol, if it exists, would ensure that such hidden information is either unknown to players who are not entitled to it, or that it would be useless because the mechanism of choosing a card would not rely on the current order.
Let's say my library consists of Forest, Forest, Grizzly Bear. I assigned IDs 1 through 3 to these cards. Both I and my opponent know that cards 1-3 are in my library. I know that the Grizzly Bear has ID 3, but when it's time for me to draw a card, I don't decide alone which card to draw. My opponent and I would, through the protocol, agree that I'm going to draw card #2, because neither of us knows the order of cards and the order of cards in the library can be abstracted away.
I could of course claim that I drew Grizzly Bears instead of Forest, but I have previously committed to the assignment and the protocol would require that I disclose it eventually, making the cheat provable.
The problem isn't illegally drawing cards. The problem is knowledge of upcoming draws that should be hidden is valuable and influences decisions.
Let's say my opponent taps out to play a very good creature. Do I counterspell, or hold my counterspell for a potential future better spell? If I know my next draw is creature removal, I just draw and use the removal. If I know my next draw is another counter, I probably counter. I think that alone makes such a protocol impossible (without continuous camera + machine vision proctoring to ensure hidden information remains hidden).
Also, many cards make you shuffle, and that's a pain. You have to let the protocol shuffle, and then rearrange your deck according to the new randomized order, which for 20-90 cards (including commander) is annoying. This second concern doesn't make it impossible, but it does make it unpleasant.
You misunderstand my question. I ask about hypothetical "blindfolded magic" which does not require physical cards any more than blindfolded chess requires a physical board and pieces. You can, in practice, use some representation of the gamestate as a mental help, but that is for your own benefit only; what you do with that representation is neither necessary nor sufficient for changing the gamestate, so there are no cards that would need to be physically verified.
The actual gamestate exists only as the result of a process, not because of some physical component. It would be manipulated and documented entirely through a set of non-repudiable, potentially encrypted messages according to an agreed-upon protocol. Deviation from that protocol would be provable to a 3rd party.
Well, no, that's not what I mean. In MTGA/O, a central server keeps track of the gamestate and acts as referee at all times. I'm asking about a peer-to-peer protocol, not about a client-server protocol. In my previous comment I did say "would be provable to a 3rd party.", but not in the sense of a referee keeping track of the rules at all times.
You do have to ban some cards like Chaos Orb that require physical objects. Barring those (which are overrepresented in unsets), I don't see any obstacles to reimplementing a game engine as non-repudiable instead of MTGA's server-based with account logging.
How does this work with scrying? If you scry your library, the next cards become known to you but not your opponent, and you can no longer simply draw the next card by generating a random number. I think you'd have to "draw" scried cards into a separate zone and record their order, so that your opponent can verify you scried only those cards and they were drawn in the order specified.
Or even worse, what about fateseal? How do you select N cards from your opponent's deck and reorder them, without your opponent knowing which cards you selected?
Edit: Also, scrying doesn't simply bury cards, it specifically puts them on the bottom, which will be a problem to keep track of if the game goes all the way through your deck. I think you need a way to cryptographically record the entire order of the deck without cheating.
Yes, these are the complications I'm asking about, whether they can all be overcome.
For the Scry case, I would say that we determine a random card as above, let's say Card #15. I know what card #15 is because of my previous ID-assignment so I announce my choice of putting it back on top or to the bottom.
The next time the game wants to do something with my top card (e.g. when I draw a card, or it gets milled, and so on), we would agree to choose that card #15 instead of a random card because the agreed game state says that #15 is currently on top. If I have put #15 on bottom, then the next time I would draw a card I draw a random card *except* #15, because there are others still "on top" of #15.
I've played entirely digital Magic over the internet from time to time. A friend wrote a little app that tracks your deck state and supports draw, scry, Brainstorm, shuffle etc. That app was trust-based for playing with friends, but most of its functions could be easily adapted to share appropriate hashes after each library interaction. Scry and surveil are completely fine, as you say. But even that app would have trouble with Fateseal and Praetor's Grasp.
And yet having seen some of the proofs that are possible in the space of card algorithms, I feel like this might still be possible. (And if it's not already solved, there's probably a paper at the FUN With Algorithms conference in it for whoever solves it!)
I think that for fateseal & friends, a simple hash-based system is probably not sufficient, and you want a full public key crypto system instead.
In particular, Fateseal could be implemented using a commutative crypto system, e.g. one where you can encrypt a text twice with different keys and it does not matter in which order you do the decryptions. (There might be a more correct name than commutative for this property, though.)
Basically, each card in the library is encrypted twice: once by each player (who can also reorder the cyphertexts, to shuffle the cards).
If you draw a card, then your opponent decrypts the next-index twice-encrypted card. They do not learn anything from that. However, you can now decrypt the result into a card name (plus salt).
If instead you get targeted by Fateseal 1, then you have to decrypt the top card of your library for them. This will allow them to know what card it was (without you being the wiser).
One difficulty is that both parties would have to apply a per-encryption salt if they use public key systems, which will likely clash with the commutativity requirement.
Searching for "commutative encryption" leads to [1], which cites [2]. Seems the trick is to use SRA, which is RSA but with different stuff kept secret, to play "mental poker" -- which is probably mostly equivalent to MtG from the required crypto primitives.
So my idea is not entirely nonsense, just some 45 years late.
Edit: upon reading the SRA paper [2], it appears that the system depends on disclosing the key at the end of the game. This is fine for MtG but absolutely terrible for poker. The fact that you will wonder for the rest of your life if your opponent was bluffing is very much a key part of poker!
Yes, I'm looking for a zero-trust protocol that can deal with the entire tournament-legal card pool except of course cards like Chaos Orb that interact with the physical space.
I'm going to try to turn off my usual misanthropy for a moment, because I think I might have a useful idea, and some of the people here probably know more than me on this subject. (Maybe even Scott himself--he's a psychiatrist?)
On the recent discussion on men in the Bay Area, it occurs to me there's probably a large amount of older therapy literature from the period when most of the people writing about it were men. Granted it's probably got all kinds of obsolete psychoanalytic ideas running through it, but there might be stuff about ego, self-actualization, etc. that might be useful in the current problems of trying to make effective therapy for men?
That's an interesting thought, but I think maybe slightly misguided.
Men actually are well-represented among founders of different therapy schools, both in the past and the present (I think therapy school founder requires a similar neurotype to entrepreneur or something). So I don't think the bottleneck is that there aren't therapy schools that understand the male mind or whatever. I think the bottlenecks are:
- the therapists implementing the schools, including the male-founded schools, are mostly women.
- many therapists aren't going off a school and are just kind of operating on vibes. There is widespread debate about whether this is better or worse, but in any case most of the therapists doing this are female, most of the vibes are female vibes, and insofar as these vibe-based practices then congeal into a coherent philosophy of therapy, it's a female philosophy.
- men are less interested in going to therapy in the first place, and might prefer some kind of figuring-out-your-life practice which is not quite therapy in the traditional sense.
Nice to get an actual informed opinion! Thank you!
But aren't there still older books that might be more useful to men if updated for modern conditions? I've seen tapes of Albert Ellis and he had that sarcastic midcentury New York Jewish edge I think a lot of (non-antisemitic) young guys might like.
A question I posed to my wife and my brother in law: is there anyone else in the history of the world who's name is more known and used than the Earl of Sandwich? Think about it: his name became basically just a part of the English language. His name is used all of the time, even to describe things that he legitimately had nothing to do with because they're from completely separate cultures or predated him. I've seen Indian flatbread sandwiches be called, well, sandwiches, and for Passover every year Jews everywhere read the Haggadah and do the ritual of eating the Hillel sandwich. Sandwich's name even became a verb for putting non-edible things between other things; I can be sandwiched between two people on the bus or a house can be sandwiched in the middle of two buildings. My brother in law proposed that Caesar may be a contender, which is a good proposal with Kaiser, Tsar, C-sections, and salads, though I still think maybe Sandwich has him beat.
Jesus Christ is a widely known figure, and you got Spanish-speaking countries mentioning him each time someone sneezes. Then if you also count any mention on interjections and such, I think it really adds up.
The Virgin Mary might be another contender, given the popularity of the Hail Mary prayer among practicing Catholic. I think you say her name over a hundred times while praying a rosary (twice per prayer for each of 50+ "Hail Mary" beads).
Duke of York? Every time someone says new york, mentions the new york times etc. I think that's likely to be said very often, although I have no idea whether people working at subway say sandwich all the time.
I think this has to be the winner. "America" is a word used by speakers of languages all over the planet including plenty of people who don't know a complete sentence's worth of English words. Plus it appears in multiple places on every world map, classroom globe, etc.
Most people who say "sandwich" have no idea that it's named after anyone, but if that doesn't matter...
More well-known? I would guess that the most well-known names would be some prominent religious figures surpasses it: Adam, Abraham, and Moses are all widely known among adherents of the Abrahamic religions, which are the majority of the world's population.
More frequently used? I don't speak Mandarin, so I don't know if any common words in it are named after people, but if there are, it would probably beat sandwich, given that much of the world either doesn't often eat sandwiches or doesn't speak a language where the word descends from the Earl's title. Caesar's name probably gets more uses from the month of July than from any other source, I would guess.
If there's any actual human whose name was the source behind the various deities that have weekdays named after them (Thor's Day, Saturn's Day, etc), those would probably beat sandwich.
> for Passover every year Jews everywhere read the Haggadah and do the ritual of eating the Hillel sandwich.
Most seders aren't conducted in English, and therefore do not have any direct reference to "sandwich".
I'd guess there are more people who are familiar with sandwiches worldwide than there are Chinese people. In any case in Chinese it is also called a sinicized version of "sandwich": 三明治. Funny how that 三 (sān, meaning 3 when on its own) looks a lot like a sandwich.
4. A shabby or disreputable fellow, particularly one wearing clothes resembling the worn-out cast offs traditionally used to construct a sense-2 or sense-3 guy.
5. The modern meaning, similar to 4 but having lost the negative aspects of its meaning.
The line in the song "I've Got a Little List" from the Gilbert & Sullivan opera "Mikado", where the Lord High Executioner mentions "The lady from the provinces who dresses like a guy", he means sense 2 or 3 with shading of sense 4. A modern American lyricist trying to convey a similar meaning might say something like "The woman from the country who dresses like a scarecrow".
Emperor Augustus was named after the adjective, not the other way round. Also, the adjective is probably less used than all the combined non-July things named after Julius Caesar.
Does anyone know a good writing forum? I've been trying to get back into fiction writing but am currently managing about four sentences every three days.
I remember the children's ditty about Mussolini. It was sung to the tune of "Whistle While You Work" (as sung by the seven dwarfs of Snow White fame), and it was ambiguous as to whose member (his own, or der Fuhrer's) Mussolini had bitten.
This was 20ish years after the end of WW2. I assumed it originated with the uncles and older brothers of my peers.
I recall it as well, though in the version I knew, he "pulled his weenie." I had always assumed it was his own weenie. My understanding is that the ditty was very popular during the war; I am pretty sure I learned it from my mother (along with several unrelated dirty limericks), but I might be mistaken.
For the uninitiated:
"Whistle while you work. Hitler is a jerk. Mussolini pulled his weenie; now it doesn't work!"
I would suggest ignoring both the hype and the anti-hype – journalists seem to like writing headlines like "gpt5 was a massive failure which proves LLMs have plateaued", but I think it's actually mostly in-line with trends (Peter Wildeford said it's roughly what he forecasted; I'm likewise; a poll in the ACX discord of gpt5-predictions before release showed most people getting what they expected or being at most mildly disappointed). For instance, the "time horizon" benchmarks are still scaling strongly. Personally I think Zvi Mowshowitz gives some good overviews of this stuff.
I've been wondering the same thing after noticing that the release of GPT-5 didn't make that big of a splash. Although it's still relatively early, so who knows what might pop up.
I think AI models with the ability to do contiuous learning could be an interesting next step. Quickly integrating new information seems essential for getting good use out of an AI model. That way, a model could improve on the job, so to speak.
I imagine something like telling an LLM the kind of mistake patterns it makes leading to the model avoiding those kinds of mistakes in the future.
But that seems like a big safety issue, because it would probably be very hard to predict what new behaviors such models might develop. What kinds of changes should a model welcome and what kind of changes should it resist? How could it be guaranteed that the model will uphold some core principles if it's always open to change?
Note that the amount of compute that went into training GPT-5 was a much smaller relative increase over GPT-4, compared to GPT-4 over GPT-3 or GPT-3 over GPT-2. So the correspondingly smaller capability gains are expected and don't refute the scaling hypothesis; OpenAI is just engaging in version number inflation. Training a model that's that kind of step up from GPT-4 requires constructing a bunch of new data centers from scratch, and there hasn't been time for that.
No, unfortunately, depending on what you mean. It doesn't seem like labs are banking on just 10xing the parameters on last thing that worked and throwing most of the compute at pre-training. The new hotness is synthetic data, which is expensive to produce but unlike internet data is potentially unbounded. The methods so far used to train the top models are quite crude, at least those methods described in publicly available documents, still they have been rather effective and capabilities continue to increase quickly and steadily despite model size kind of leveling out or sometimes even shrinking.
So what's likely going to be evolving, in the near term, is training architecture. Unlike the transformer where, while there are many ideas floating around it's not super clear what to try next other than fiddling with little details to squeeze out more efficiency, there are some blindingly obvious next steps on the training front. For example, engineers would I imagine very much like to have the ability to train a really big model with something like MCTS, it's been tried but the hardware needed to make that viable apparently isn't there yet and simpler but less compute heavy schemes have won out. Despite that, the trend of getting close to maxing every benchmark more than a couple years continues. And the hardware needed for the next few steps is coming online very soon. Seatbelts buckled, we ain't seen nothing yet
I don't know CBTi, coding, or entrepreneurship, and I'm not sure we're quite at the part of the glorious AI future where the best person to do something that requires three skills is someone who has zero of the skills.
Is the technology at the point where somebody without any expertise in software engineering can make an app and expect it to work? That was not my impression.
I experimented with it a bit over the weekend. My impression was it works until it doesn't, at which point it became a complete nightmare to untangle everything and figure out what was going wrong. And the LLM itself was not much help at that!
Also, though, it didn't work well when I said "make an app to do xyz". It only really worked when I broke down the project myself into a series of logical steps, and then asked the LLM to work on each step at a time.
Web app kind of if you already have the kind of "grit" that you would need to code (try something, fail, try to understand why you fail without getting frustrated too much, try again), but you will get results way faster than learning to code from scratch. Mobile apps it's harder from what I understand, I see less of them.
Admittedly I was trolling a little bit. But only a little! I really am curious what Scott has to say about this.
I think it is possible to get some sort of working app if you spin the slot machine often enough, even if you don't know any code. But also I assume that Scott has access to the latest and greatest (or at least most expensive) models, so his experience might be totally different and much better. If not, I'd like to hear about that too.
The toothbrush moustache sported by Hitler was already popular in Germany and its origin country USA by the time he probably adopted it. There are photographs of Hitler during WW1 that show him wearing a Kaiser moustache instead. Wotan's moustache in the painting also looks much broader than a toothbrush, extending beyond the nose on the sides more like a painter's brush moustache.
To summarize, the "trench moustache" was part of Hitler's and the NSDAP's propaganda that was built on his personal recognizability as a trench soldier from the war, everybody in Germany in the post-WW1 time would recognize it and would understand his message "I am one of you soldiers". For the same reason, he added a trenchcoat to his civil suit and abstained from wearing his lederhosen. Moreover, it was a more modern kind of moustache compared to the establishment style.
Not the only time someone thought that certain pictures must have been the inspiration for later work.
From a draft letter of Tolkien, 1971:
" A few years ago I was visited in Oxford by a man whose name I have forgotten (though I believe he was well-known). He had been much struck by the curious way in which many old pictures seemed to him to have been designed to illustrate The Lord of the Rings long before its time. He brought one or two reproductions. I think he wanted at first simply to discover whether my imagination had fed on pictures, as it clearly had been by certain kinds of literature and languages. When it became obvious that, unless I was a liar, I had never seen the pictures before and was not well acquainted with pictorial Art, he fell silent. I became aware that he was looking fixedly at me. Suddenly he said: 'Of course you don't suppose, do you, that you wrote all that book yourself?'
Pure Gandalf! I was too well acquainted with G. to expose myself rashly, or to ask what he meant. I think I said: 'No, I don't suppose so any longer.' I have never since been able to suppose so. An alarming conclusion for an old philologist to draw concerning his private amusement. But not one that should puff any one up who considers the imperfections of 'chosen instruments', and indeed what sometimes seems their lamentable unfitness for the purpose."
See also here for possible (postcard) inspiration:
"Tolkien had a postcard labelled Der Berggeist ("the mountain spirit"), and on the paper cover in which he kept it, he wrote "the origin of Gandalf" at some point. The postcard reproduces a painting of a bearded figure, sitting on a rock under a pine tree in a mountainous setting. He wears a wide-brimmed round hat and a long cloak and a white fawn is nuzzling his upturned hands. Humphrey Carpenter in his 1977 biography said that Tolkien had bought the postcard during his 1911 holiday in Switzerland. However, Manfred Zimmerman discovered that the painting was by German artist Josef Madlener and dates to the late 1920s. Carpenter concluded that Tolkien was probably mistaken about the origin of the postcard himself. Tolkien must have acquired the card at some time in the early 1930s, at a time when The Hobbit had already begun to take shape."
As soon as I feel like I've heard all of the details about Tolkien, his thoughts, what he was going for, all of his letters, here's one I had not known about.
The letters are so good, the only problem is that they are selected and edited, so sometimes Humphrey Carpenter leaves out interesting bits and obviously there are a ton more letters I, for one, would like a peek at 😀
I thought he cut it to fit under a gas mask and just decided to keep it that way.
WW1 was when beards and mustaches really fell out of style, and it’s largely because they needed to be trimmed or cut completely to fit under the gas masks at the time.
That is the story, yes - other stories say that he adopted the toothbrush to leech off of Charlie Chaplin's popularity. But the reason doesn't really matter; the point is that there is photographic proof that he didn't have the toothbrush before well into adulthood, and that he didn't adopt it "as a youth" as the highlighted comment relays.
That then raises the question, why did Chaplin use the toothbrush moustache? If it was a style at the time, no need to wonder where Hitler picked up the idea of grooming that way.
According to Wikipedia, it originated in America and was introduced to Germany by visiting Americans:
"The toothbrush moustache was introduced to Germany in the late 19th century by visiting Americans. Previously, the most popular style was the imperial moustache, also known as the "Kaiser moustache", which was perfumed and turned up at the ends, as worn by German emperor Wilhelm II. By 1907, enough Germans were wearing the toothbrush moustache to elicit notice by The New York Times under the headline "'TOOTHBRUSH' MUSTACHE; German Women Resent Its Usurpation of the [Kaiser moustache]". The toothbrush was taken up by German automobile racer and folk hero Hans Koeppen in the famous 1908 New York to Paris Race, cementing its popularity among young gentry. Koeppen was described as "Six-feet in height, slim, and athletic, with a toothbrush mustache characteristic of his class, he looks the ideal type of the young Prussian guardsman." By the end of World War I, even some of the German royals were sporting the toothbrush; Crown Prince Wilhelm can be seen with a toothbrush moustache in a 1918 photograph that shows him about to be sent into exile."
So if it was a popular style worn by role models, I am not surprised Hitler adopted it when he's becoming a big cheese in the Party in Munich; this would show he's not some hick from the provinces.
Have you actually seen a Chaplin film? They're brilliant. I'd compare them to a Coen Brothers' film. Sure there's humor but there's also deep artistic wrestling with the existential realities of life. They also have a breathtaking visual style. They weren't just slapstick.
"According to Hitler's bodyguard Rochus Misch, Hitler "loved" Chaplin films, a number of which he watched at his teahouse near the Berghof (built c. 1936)."
I recently got a vasectomy as a push present to my wife after she birthed our third child. Being a good ACXer and man anxious about having my junk operated on, I researched the procedure obessively. The medical information was mostly unsurprising. The demographic information was surprising.
From the abstract: "11.4% of men aged 30–45 years reported having a vasectomy, representing approximately 3.6 million American men. While 14.1% of white men had a vasectomy, only 3.7% of black and 4.5% of Hispanic men reported vasectomy. On multivariate analysis, a significant difference in the odds of vasectomy by race/ethnicity remained, with black (OR 0.20, 0.09–0.45) and Hispanic men (OR 0.41, 0.18–0.95) having a significantly lower rate of vasectomy independent of demographic, partner, and socioeconomic factors. Having ever been married, fathering two or more children, older age, and higher income were all associated with vasectomy."
I had assumed getting a vasectomy was like a standard rite of passage for American men entering middle age. My Dad had one, and many of my friends' dads got snipped. Same for my wife's family. It turns out that vasectomies in the US are not *that* common and those who get them are overwhelmingly middle class white males, and we were raised in a white middle class bubble. Mind blown.
The procedure itself wasn't bad. Local anesthesia and I didn't feel the incision or anything more than pressure/tugging. Prior to the incision, the doc discovered one of my vas was out of place, so he had to manipulate it closer to the skin with his fingers. That hurt quite a bit, even with the anesthesia.
The initial recovery was quick (stopped ice and painkiller about 36 hours in) but I wasn't able to resume running and lifting without discomfort until about 2 months after.
Before my vasectomy almost 30 years ago, the doctor showed me a short educational film, which included an apprehensive Black man asking whether, after the procedure, he "would still be a man".
Consistent with the data, I am a white middle-class male.
Why don't LLM apps implement normal messaging apps functions such as responses?
When an LLM answer touches on multiple subjects, I don't want to have to say "about this subject could you tell me... regarding that one I disagree..."
The LLM should break up its response into multiple messages, and you should be able to answer to each message before launching the inference process
Not sure if this is quite what you had in mind, but there are a few tools out there that allow you to "multiverse" with LLMs -- i.e., branch off a discussion from a certain point. Loom (https://github.com/socketteer/loom) is one, but it's specialized in fiction. Raycast also lets you do what they call "chat branching" (https://www.raycast.com/changelog/1-101-0). Afaik neither of these have the ability to merge the branched chats back into a single thread.
But Claude also has always had it? Meaning - you can take any of your replies and edit it, turning it into a splitting point. Then you can switch between the branches. It's very inconvenient.
If you make just one branching point, it's alright. But when you make multiple branches, multiple splits it becomes hard to keep track of what is where, and if you exit the chat the system doesn't remember where you were and next time you open it you may end up on some other node and have to look for the right place. Basically, it's not a system that was meant for heavy use, but it's convenient enough if you don't push it too far.
The other day I had the sudden realization that in most accents (and, more importantly, in General American, the accent I try to consciously emulate when I speak English) the pronoun "they" is pronounced [ðeɪ].
I have gone all my life pronouncing it [ðεɪ] like some sort of barbarian! I guess it's one of the effects of having started to learn English very young: you learn some things approximately when you are six, and they lock in forever. I still sometimes catch myself pronouncing "is" as [iz] rather than [ɪz]. How utterly embarrassing.
I don't really have a point with this, but it's a bittersweet realization: for however long and deeply you studied a non-native language, it will never be truly yours. You really are stranded where you were born, so to speak.
Ok since the phonetics aren’t obvious here to many I asked ChatGP, which writes.
[ðeɪ] — is the standard pronunciation of they in English. The vowel [eɪ] is that diphthong we use in words like day, say, play.
Meanwhile
But if you’ve been saying [ðεɪ] with the open-mid front vowel [ε] (like the vowel in bed), you’ve basically been saying something closer to “theh” sliding into a weak y. Not far off, but it would sound a bit unusual to most English speakers — like a slight accent twist.
And when I asked was this kinda posh, the model said not anymore but it’s close to the clipped RP back in the day. The Queen in 1950 but not 2000.
Since British English is the gold standard (and frankly no need for the British qualifier) then I think the op should be proud
That's interesting to know, but my concern was that the sound I am making is, as you mentioned, unusual. The part that messes me up is that I'm pretty sure I make that specific vowel combination ONLY when pronouncing "they". I pronounce, for instance "day" as [deɪ] i.e. the standard way. I have this idiosyncratic pronounciation exclusively for the word "they", in all of my English vocabulary.
That's not what a native speaker of any language would do, regardless of their accent: they would just pronounce "day, grey, may, they" as rhyming, regardless of the actual sounds.
But in my idiolect, "they" doesn't rhyme with anything!
> day, grey, may, they" as rhyming, regardless of the actual sounds.
That’s not obvious at all. After all there are differences between a and e in many words. (And probably isn’t how the Queen used to say “They went that day”)
I sympathise; minor inaccuracies in pronunciation can be difficult to notice or correct once they become entrenched. I'd been speaking French for years before I realised there was supposed to be a difference between répondre [ɔ̃] and répandre [ɑ̃].
I've heard this exact same worry before, from a French ex-girlfriend who could hardly speak English when we first met and felt terrible about it (despite my still worse French) and we discussed it quite alot. So I actually feel somewhat equipped to disagree with you!
1) I think it's pretty extreme to realise you're mispronouncing a word and conclude that the language isn't yours: many - perhaps most - native speakers mispronounce many words for their whole lives. For example, young British people lamentably mispronounce (and misspell) words the American way because apparently they learn their English from sodding Netflix and Twitter now instead of BBC Radio Four and Jane Austen like they bloody well ought to - and yet despite their horrible misshapen American-English-with-British-Accents the language belongs to them far more so than it does to me! For another example, there are English words I'd only read, never heard, and only upon hearing them decades later did I realise that I'd been mispronouncing the words my whole life! (There have been plenty, but "cloaca" springs to mind...)
2) The fact that you're sufficiently self-aware to be even _capable_ of noticing such things (and sufficiently well-educated to be capable of using the IPA..) puts you _miles_ ahead of the typical native English speaker. You may not have Tennyson's accent but you're nevertheless "doing English" the way he did! I think self-awareness and education (including self-education) reveal to you the beauty, uniqueness, poetry, and expressive power of a language far, far more than does having it programmed into you when you're a toddler and then literally never voluntarily giving the language another thought for the rest of your life and just using whatever misspelled, unpunctuated, minimum-effort English you absorb by accident from social media.
3) I think "belonging to" a language (or at least, the good part of belonging to a language...) is about how you think about what you say, and choose your words for their poetry or semantic content (as opposed to the quasi-meaningless slop that people only seem to say because they heard it on Netflix... "energy", "based", "gave", "called out", "hits"....) much more so than it is about what pronunciation you use.
4) There are a great many accents and dialects in English, sometimes all-but unintelligible to one another, and it's only really a total accident of geography that the Glaswegian accent "counts" and (say) the Sicilian accent doesn't; there's nothing fundamentally worse about the latter (some might even consider it more beautiful or expressive than the former!) If anything, I think that having both - provided they can at least understand one another, so the accent isn't an insurmountable barrier to communication - makes the language richer and more interesting.
I get what you are saying, and admittedly it's not particularly bad, but it's the difference with my native language that gets me. I mispronounce really common words, while native speakers usually tend to mispronounce less common words; the less common the more mispronounced.
But I was mispronouncing "they"! It's probably in the top 30 most used English words, and I must have heard it spoken literally tens of thousands of times!
I get what you are saying, about foreign vs. native accents, and I agree in principle.
Still, when I can't keep my accent from slipping out, it feels like a failure of self-control.
I guess I would say that (again, you're not the first person I've had this exact same discussion with, sorry if this is somewhat weird coming from a total stranger!) if "not letting your accent from slipping out" is a fulfilling goal that you derive satisfaction and wellbeing from achieving, then sure: keep striving towards it and don't give up until you achieve it.
...but! A) Definitely think very hard about whether this is genuinely a satisfying and fulfilling goal to aim for, because the effort seems to be expensive to you and failures seem to affect you and maybe it's possible that you would derive more fulfillment from learning to embrace the uniquely Emanuellean "they" and focusing your efforts on some other problem, and B) even if, upon consideration, you do consider this a worthwhile and fulfilling goal that you genuinely do want to strive for, that's perfectly fine - but please don't conflate it with "better English" or "English belonging more to one", because that genuinely is more about mental properties than about perfectly mastering very very specific pharyngeal muscle movements...
(As for "they": I know many native English speakers who mispronounce "February", or use "criteria" in the singular, literally every day, and aren't remotely curious or contemplative enough to ever notice they're doing it. I believe that you are a better English speaker than these people, no matter how you pronounce "they"!)
Just today I realized that in a lot of accents, honor/honour is not pronounced the same as "on a", and so they don't find the name Honor Blackman to be all that funny.
Both yes. Are you serious about this by the way? Do you really expect that native speakers ( of many varied accents ourselves) care about minor mispronunciations?
I wasn't serious at all, no. Sorry if I offended you! I just found "...and she's vaguely acceptable" very droll, is all - if it was meant purely factually I misread it, and I'm terribly sorry!
I am Luomei Lyu, the founder and developer of Sheep-Sleep. I refunded Kevin his $15 (not a few hundred bucks as he claimed) the second I woke up and saw the message. Poor Wi-Fi can cause latency in AI responses, and we are actively improving this every day.
The real value and focus of Sheep-Sleep is in the content. Every single AI response — down to the exact wording — comes from thousands of hours of discussion and two years of work with some of the most experienced sleep psychologists in the US. Sheep has already helped many people sleep better: https://www.gnsheep.com/case-studies
As for the internet issues, that’s on me!
—Luomei
P.S., Here is how it was made: We had the best psychologists in the country talk us through every conversation they would have for each insomnia case, even down to the exact word choices, analogies, and conversation pace when teaching these well-validated techniques to their patients. We then wrote hundreds pages of instructions for generative AI to follow in its weekly sessions, so that it only responds with human-crafted material, while keeping the conversation collaborative, interactive, and highly personalized. This kind of dynamic, personalized dialogue is what makes in-person sessions effective at improving adherence, which remains the biggest challenge in this gold-standard, first-line treatment.
Sure. First of all, I ran the iphone app on my desktop and didn't have the best Wi-Fi, so that may have caused some problems below. Anyway, my main issues are:
- It's a SaaS, but there's no login (you literally get a code after paying), and there's no way to cancel your subscription except to email them and wait (I'm still waiting btw), which is just dumb. I thought 298/mo would at least get you an account!
- The conversation with the AI chatbot was unsatisfying for a bunch of reasons, eg, it takes it like 20s of thinking to ask me a preprogrammed question like "Are you having trouble falling asleep or staying asleep?" Couldn't they just have a button that I can press? The first few questions are all like this, with it thinking for 10+ seconds, then asking a very simple question. Also, in its second answer, it repeated the answer twice (idk why), and in its 3rd or 4th answer, it didn't recognize my speech and asked me to speak again. Finally, this might be my aesthetic preference, but talking to an AI-generated cartoon of a sleep just feels...awful. Can't they put a bit more effort and actually have someone draw the character?
- I couldn't find their email/contact on the website, the ads/promotion vids are, well like most fancy new medical products, distasteful and exaggerated––but these are just small complaints.
- Oh and if the founder sees this, I'm still waiting for you to cancel my subscription!
It doesn’t look aesthetically how one would expect an expensive health service to look, so it will be judged on that, and not on the ability to improve health outcomes.
Of course, for all I know the app could be garbage at actually improving sleep too, but I notice that that doesn’t come up in your list of complaints.
It's a heuristic. Having such a badly designed app is correlated with insufficient funding, a lack of professionalism, not really caring much about people using the service, etc.
Okay, but that's not a useful user review, because users can see for themselves that it's unpolished. The point of a user review is so that users can find out whether it actually works, from someone who has direct knowledge of this!
Actually, that's *exactly* what I would want in a user review for something like this. I can't see for myself that it's crude and unpolished and frustrating to talk to without paying £300-odd a month (and a very difficult to cancel £300-odd, too); I'd much rather find these things out before opening my wallet....
If it's so dissatisfying to use that somebody gets frustrated, tries to cancel, and can't even cancel, then as a user I've found out from this review that it doesn't work
Yea, it's not. If someone can pay a few hundred bucks to try this and come back and write a review, that'd be more useful/reliable. (I'm just making a guess about the nature of this thing based on what I saw, if you still want to try the app after reading my rant, by all means go ahead.)
SheepSleep (mentioned in the post above). It looks like they wrote the app in a high school hackathon. I bought a trial and canceled it. Charging 298/mo is diabolical.
Ah yes, the pitfalls of coming straight to the comments.
It shouldn’t be too hard to vibe code a… competitor, especially if someone does the public service of signing up for a subscription and documenting all the app screens and features with screenshots…
Sadly, it's too late to apply for an ACX grant to do a randomized trial on whether the $298/month insomnia app helps people sleep or not.
I'm not sure how you blind the trial. Give half the people a placebo app which costs $298/month but is somehow guaranteed to not contain helpful advice?
I think the problem with blinding here is that $298/month is already most of the treatment effect. If you told people they were in a sleep study and charged them $5, they would probably sleep worse out of spite.
As for AI being a demon: demons are usually unreliable tricksters. LLMs are more like over-enthusiastic interns who take your vague instructions much too literally.
Regarding AI, even the founders know it's diabolical:
Elon Musk: Artificial Intelligence is our biggest existential threat. ... AI is summoning the demon. Holy water will not save you.
DWave Founder Gordie Rose (A Tip of the AI Spear): When you do this, beware. Because you think - just like the guys in the stories - that when you do this: you're going to put that little guy in a pentagram and you're going to wave your holy water at it, and by God it's going to do everything you say and not one thing more. But it never works out that way. ... The word demon doesn't capture the essence of what is happening here: ... Through AI, Lovecraftian Old Ones are being summoned in the background and no one is paying attention and if we’re not careful it’s going to wipe us all out.
The question of how hard it is to control a powerful AI is separate from the question of how soon powerful AI is coming, though of course there's some common interest in figuring out countermeasures.
I dunno, I think pouring a pint of it straight into the innards of the server might do *something* 😁
Ha ha, and you thought I was crazy for having my rosary beads, votive candle of St Martha, and picture of St Therese of Lisieux on my desk around my PC!
They think they can make money and a 20-30% chance of killing everyone isn't enough to keep them from taking a chance on making some more money; entrepreneurs have faced worse odds, I think.
Also perhaps running the AI can improve their chances of being able to control it--to continue the Lovecraftian metaphor, if everyone's summoning Old Ones, maybe if you're the guy who summons Cthulhu you can send him after the rest of the world first.
I love this community but you guys have way too much reverence for successful business people.
Who exactly is "you guys" here? AFAICT people have mostly been unimpressed with, e.g., Musk ignoring his own stated concerns about the control problem.
The cultists imagine that by serving the Old Ones they will be top dogs in the new post-sweeping Earth, that they will have all kinds of powers and pleasures:
“The time would be easy to know, for then mankind would have become as the Great Old Ones; free and wild and beyond good and evil, with laws and morals thrown aside and all men shouting and killing and revelling in joy. Then the liberated Old Ones would teach them new ways to shout and kill and revel and enjoy themselves, and all the earth would flame with a holocaust of ecstasy and freedom.”
The reality, of course, is that all humans are like vermin to the Old Ones and they have no attachment to the cultists or servitors, and they will be devoured and consumed like the rest of us.
Good, because the Great Old Ones will chomp him up along with the rest of us. It doesn't matter what the humans plan or intend, they're just in the way and must be cleared off with the rest of the rubbish.
I sometimes wonder if that's actually Elon's plan, to be eaten first, just like that joke about the Cthulhu cultists: https://www.entrelineas.org/pdf/assets/who-will-be-eaten-first-howard-hallis-2004.pdf. If you stand before the maw when the Hellgate opens, because *you're* the guy who opened it, indeed fought your way to the front of the queue to be the guy who opens it, then...
I kind of feel like this brand of rhetoric isn't helping. It's not that it's wrong; it's that there is a *technical disagreement* about what factors make it hard to steer a powerful AI and how hard they are to overcome. Alignment optimists aren't going to find this kind of comparison compelling, they're just going to conclude that we're mindkilled. Better to focus on the specific technical reasons for pessimism, rather than the vibes.
The same reason they built the LHC, they are creating a portal at CERN which literally sits above where Apollo was thought to reside in the underworld, “to another dimension (CERN own words) which something might come through” (direct quote).
Meanwhile, whatever comes through can use the quantum computers (again messing with the fabric of reality) as a host to interact with the world.
As far as I understand, the LHC is not doing anything that doesn't already happen naturally in the atmosphere. It's just recording the results better.
And quantum computers are quite weird, phyisically speaking, but there's zero reason to believe it's "messing with the fabric of reality". Especially in a universe where things like neutron stars and super novas exist.
AI is not a threat to the universe any more than humans are, but (1) it's still a threat, and (2) at a much shorter timescale. Comparing AI capabilities researchers to LHC and QC engineers is not fair.
Alas, the ones who are building it are telling you directly that they can do everything I mentioned (if you read the article and watch the video the evidence for the LHC and Quantum computers is right there), but you know better than they do… so nevermind, i guess. Move along.
I opened the page. I saw that it said “ AI quantum computers act as hosts for disembodied Nephilim spirits who are stuck in this realm unable to escape:”.
haha, that made me laugh. you should scroll to the section where gordi is giving a presentation and talks about how standing next to his d-wave machines - it feels like they are alive and sentient.
They just can't resist the challenge of building the Torment Nexus from the classic sci-fi novel Don't Create The Torment Nexus. Also, if if it's not them who destroys the world, the Chinese will, and do you really want us to fall behind in a race with the Chinese?
Well, charitably, each of them believe that an ASI created by themselves has a higher (and potentially non-negative) expected value than the ones currently being developed by others and so they are essentially forced to run the race in to save humanity from the more monstrous creation of the next guy over.
There is also a significant amount of legitimate disagreement about how hard it is to make a powerful AI do what you want.
(This doesn't excuse Musk, who I don't think has a principled commitment to an anti-China foreign policy either; I think he's just crazy. Rose isn't working on AGI as far as I know; his former company put out some breathless AGI-sounding marketing material but in reality they were working on narrower robotics stuff.)
That is a good point. AFAIK some like Amodei are not so impressed by the ‘superintelligence explosion’ argument and generally believe we’ll be able to align powerful AIs by researching interpretability and improving model design, prompting, finetuning, etc.
There is also potentially the economist-brained viewpoint that alignment won’t be a problem because such systems would generally act rational economic actors and prefer. Or that they’ll be trained on the human corpus which largely encodes ‘good’ values and so will be good (enough) by default.
I was focused on making a charitable case for Musk, Altman, etc, who do, I think, believe in that case *at best*.
Do you happen to know what Altman has had to say about the control problem? IIRC when OpenAI was founded his stated concern about AI was the possibility that economically useless humans might be liquidated, and his proposed solution was a UBI indexed to GDP. He still gestures at this general "governance" class of concern and says that OpenAI is taking it seriously, but never provides specifics or commits to anything.
It is well known that LLMs are bad at questions like "Which is bigger, 9.11 or 9.9?", "How many r's in strawberry?", etc. The cause of this is also known to be tokenization.
Is it fruitful to draw an analogy between these questions and optical illusions? In both case, we exploit quirks in how the perception system works, in order to force a wrong answer. Are there more aspects of LLMs that this analogy helps explain?
I suppose it comes down to whether you think that these questions for LLMs and optical illusions for humans are an artifact of training data or architecture.
LLMs operate by sequentially sampling a token out of a distribution conditioned on the previous text it has seen and generated. For your example, this process doesn't trigger a boolean "is 9.9 > 9.11?" operation but rather something like: P("9.9" | "Which is bigger, 9.11 or 9.9?"). Folks with experience in LLMs will know that this isn't entirely correct, but for the sake of scientific communication it is mostly correct.
I'd argue this is an architectural problem. Sure given an infinite amount of training data perhaps this (usually stochastic) sampling operation could mimic the boolean operation perfectly, but the formulation of the model is not conducive to supporting that.
On the other hand, my understanding of optical illusions is that they are based on exploiting our brains making "processing shortcuts" they have learned based on visual data from the real world. I am not a Neuroscientist, but I'd hypothesize that if we had things that looked like optical illusions in the real world, our brains would learn to make them not optical illusions- hence it is a training "problem."
What do Michael Bloomberg, Zohran Mamdani, and Nick Fuentes have in common? All are successful people who worked hard to get where they are. All are subject to a never-ending, high-pitched stream of "it's not faaaaiiiirrrr." This similarity is increasingly important as one side of the political spectrum becomes defined by the glorification of mediocrity.
Let's start with Fuentes. A while back he said, in a video I was not able to find, that he was frustrated with the people in his movement who act like he was given his position on a platter, when really he got it because he put in the effort and was good at drawing attention to himself. Perhaps this accounts for his increasing centrism - a common arc for far-right leaders. He's attacked conservatives for being "openly hostile to all the good things about liberals" and being "low-IQ hillbillies who take pride in being simple and hate the rich,"[1] and is now supporting Gavin Newsom for President.
Onto Mamdani, on track to become the mayor to eight million people at 33-years-old. An intelligent conservative movement would look at his success and ask how they can emulate it. Instead, one influencer on X called him a "failson"[2] perhaps because he spent his early adulthood taking low-income political jobs that put him on the path to be Mayor of NYC, instead of being the dentist who works out of a New Jersey strip mall. Another attacked him for having never having "done a single day of real, hard work" the way guys on a construction site do.[3] If accurate, it seems to have worked out for him. Maybe conservatives should try to emulate his success, put their kids on the path to take political power and undo Mamdani's harm, instead of going on about how much better they are because they hauled cement that summer. Another Tweeter bitched that the attendees at his recent scavenger hunt "have the same vibe as Disney adults."[4] Instead of telling Republicans how they can replicate a successful voter mobilization strategy, she'd rather they smugly feel superior and then lose elections.
Does all of this matter? I argue it does. Our "pro-capitalist" party could be talking about how much richer West Germany was than East Germany, how capitalism can make you rich, too. Instead, it's telling people they should be proud of being poor. The Mike Bloomberg's of the world don't like Mamdani's ideology or policies, but he doesn't give off the stench of resentful loserdom the way so much of the Right does. The result will be a Right that finds itself increasingly struggling to recruit competent people and raise money. On the bright side, this dearth of EHC provides an opportunity for high-IQ, high-agency people who are willing to hold their nose and enter Republican politics.
1. https://x.com/FuentesUpdates/status/1908187813117411525
2. https://x.com/feelsdesperate/status/1955655602031734888
3. https://x.com/michelletandler/status/1955648799071683055
4. https://x.com/InezFeltscher/status/1959996136279712162
The right is just following the left's lead into identity politics. They very much used to make ideological arguments in favor of capitalism but the last 20-30 years have demonstrated that tribal politics trump well-reasoned arguments so the right looked around, realized that white people still comprise a majority in this country, and so decided to go with that.
> Onto Mamdani, on track to become the mayor to eight million people at 33-years-old. An intelligent conservative movement would look at his success and ask how they can emulate it.
...He's a populist. The right's already doing that, they don't need to copy him. He still needs to be attacked, because he's opposition.
"What do Michael Bloomberg, Zohran Mamdani, and Nick Fuentes have in common?"
They haven't demonstrated they can bench press 135lbs? Well, I don't know about Fuentes, but I do about Mamdani and (probably) Bloomberg:
https://thehill.com/homenews/campaign/5468735-adams-cuomo-mock-mamdani-bench-press-effort/
https://nypost.com/2025/08/25/us-news/riley-gaines-latest-to-mock-zohran-mamdanis-failed-bench-press/
Reading the Praise/Prompt Machine: An Interface Criticism Approach to ChatGPT
https://dl.acm.org/doi/full/10.1145/3744169.3744194
A common explanation for political polarization in the US is that most congressional races are non-competitive, and therefore politicians only need to worry about losing to someone more radical from their own party, leading to them preemptively becoming more radical themselves.
But this argument seems like a leap in logic. Even if they only need to worry about same-party challenges, why do they only need to worry about more-radical challengers? Why don't they also need to worry about more moderate same-party challenges who would force them in a less radical direction?
Because primary voters are further from the center than the general electorate.
Fair point, but then we might expect the average winner to be pretty close to the preferences of the primary electorate, at which point it should be equally easy to challenge them from either direction.
In practice I think that electability concerns and some kind of influence from pragmatic centrists might make the candidates tend to be more centrist than the primary electorate would really prefer, though. And in that case it would play out as described.
Hey all, I'm Luomei Lyu, the founder/dev of Sheep here!
Sheep has been tested and loved by many sleep physicians and their patients: https://www.gnsheep.com/case-studies
We are actively working on improving the conversation experience on poor internet.
Here is how it was made: We had the best psychologists in the country talk us through every conversation they would have for each insomnia case, even down to the exact word choices, analogies, and conversation pace when teaching these well-validated techniques to their patients. We then wrote hundreds pages of instructions for generative AI to follow in its weekly sessions, so that it only responds with human-crafted material, while keeping the conversation collaborative, interactive, and highly personalized. This kind of dynamic, personalized dialogue is what makes in-person sessions effective at improving adherence, which remains the biggest challenge in this gold-standard, first-line treatment.
I truly believe Sheep to be the best digital sleep program based on CBTI, and so do our users.
> fascism is when you do things liberals don’t like
And things historians don't like, but they have proper arguments to back up their position: https://acoup.blog/2024/10/25/new-acquisitions-1933-and-the-definition-of-fascism/
This is not a proper argument. It is almost certainly intentionally misleading. This does not take long to establish.
Two quotes:
"Instead, Hitler gained power not because a majority of Germans agreed with his aims, but because key leaders, most notably Franz von Papen, thought they could use Hitler to achieve their aims, that they could sand off all of the nasty rhetoric and instead employ Hitler as a cudgel (against the socialists)."
"What I want to note here are two key commonalities: First, fascists were only able to take power because of the gullibility of those who thought they could ‘use’ the fascists against some other enemy (usually communists). Traditional conservative politicians (your Mitch McConnell and Lindsey Graham types) and conservative business leaders (your Elon Musks) fooled themselves into believing that, because the would-be tyrant seemed foolish, buffoonish, and uneducated that such an individual could be controlled to their ends, shaped in more productive, more ‘moderate,’ more ‘business friendly’ directions."
This is misleading because it misframes the decisions of German moderates and conservatives by downplaying Communist violence at the time. The Nazis during the Weimar Republic were certainly violent and they advertised their own street violence but every contemporary account I've ever read has Communists being just as violent, if not more so, in the street, and it takes less than 5 minutes to mass documented violence by communists at this time, such as the Reichstag Bloodbath.
Like, you can just read a list of political violence here:
https://en.wikipedia.org/wiki/Political_violence_in_Germany
A lot of political violence by a lot of different actors.
Second, German moderates and conservatives were not choosing between the Nazis and, like, modern Scandinavian socialists. Hitler's '32-'33 rise is contemporary to Stalin's genocide of the Ukrainians in the Holodomor. There's a very small list of people who can claim to be as evil as Hitler but Stalin is right there at this time, at his most powerful and most Stalin. I haven't read too much on Van Papen in particular, maybe he was genuinely naive, but other contemporary accounts are very clear and strident in their denunciations of contemporary Communism and Hitler's anti-Communism shines through "Mein Kampf" almost as strongly as his antisemitism.
Maybe Devereaux has insight here that I lack but I doubt it. From memory, he's primarily a classicist (like, Rome and Athens and stuff) and when he's stepped outside that field into areas I have some above-average familiarity (like Qing history) he has not acquitted himself well. I suspect that's what he's done here.
What the hell are you talking about? You are blaming the Reichstag bloodbath on the communists? It was paramilitary police killing socialist protesters with machine guns and hand grenades...
> Independent and communists, on the other hand, emphasized that the shooting had been done for no reason and without warning. It is unclear whether the warnings existed. Almost all the dead and injured were found south of the Reichstag, on the opposite sidewalk and in the adjacent zoo, according to reports from various sides. There, on Simsonstrasse, the crowd was at least four meters away from the police. So there were no violent attacks during the storming of the building. Most of the victims were hit here. After the shots broke out the crowd fled in panic, the Sipo fired several more minutes with their rifles and machine guns. Nowhere in the sources claims that demonstrators would have been shot back.
https://en.wikipedia.org/wiki/Reichstag_Bloodbath
Okay, and the modern equivalent of communists is the woke. What's your point?
I think it’s kind of silly to split hairs arguing whether something meets Mussolini’s definition of fascism.
In “Fascism — A Warning” Madeleine Albright argued that the term has come to have a general sense. Written during Trump’s first term, she wasn’t saying he was a fascist then, just giving a heads up.
“In the book, I try to argue that fascism is not an ideology; it’s a process for taking and holding power. A fascist is somebody who identifies with one group — usually an aggrieved majority — in opposition to a smaller group. It’s about majority rule without any minority rights. Which is why fascists tend to single out the smaller group as being responsible for or the cause of their grievances.
The important thing is that fascists aren’t actually trying to solve problems; they’re invested in exacerbating problems and deepening the divisions that result from them.
They reject the free press and denounce the institutional structures within a society — like Congress or the judiciary.
I’d also add that violence is a crucial element of fascism. Whatever else it is, fascism involves the endorsement and use of violence to achieve political goals and stay in power. It’s a bully with an army, really.”
A bully with an army was the key point.
—————
October 14, 2024 — CNN
“Former President Donald Trump suggested using the military to handle what he called “the enemy from within” on Election Day, saying that he isn’t worried about chaos from his supporters or foreign actors, but instead from “radical left lunatics.”
“I think the bigger problem are the people from within. We have some very bad people. We have some sick people. Radical left lunatics,” Trump said told Fox News’ Maria Bartiromo in an interview on “Sunday Morning Futures.”
“I think it should be very easily handled by, if necessary, by National Guard, or if really necessary, by the military, because they can’t let that happen,” he added.”
https://amp.cnn.com/cnn/2024/10/13/politics/trump-military-enemy-from-within-election-day
IMO part of the reason it is worth splitting hairs on definitions is because vague, general definitions like the one you quoted lead to motte-and-bailey arguments where individuals are using the terms to imply all the super awful things about a label that aren't part of the much more general definition.
Trump doesn't even seem to really match the definition you gave, as I don't see how anyone can accuse him of not wanting to solve problems. We may disagree with his solutions but he definitely seems interested in solving the problems.
This still feels a bit self-serving, though. Fascists are trying to solve problems! Hitler and Mussolini were not, in fact, *just* trying to placate their base and consolidate their power. Now, the fact that Hitler's idea of 'problems' included things like 'Germany doesn't control Poland's farmland' and 'Jews exist' was, of course, a bit of an issue. But the notion that either he or Mussolini were invested in 'exacerbating problems' or 'deepening divisions' after they had seized power is entirely wide of the mark. Hitler, in particular, was not a cynic who pretended to sincere antisemitism and grandiose visions to rise to the top. The world would have been better off if he were! The problem was that he was a genuine fanatic!
Except that behind those problems (Poland and the Jews) was the deeper problem, at least according to them, that Germany and Italy had lost their national identity and sense of purpose in the aftermath of World War I. Between the Depression, the Treaty of Versailles, and the general rising tide of nihilism, people had lost the sense that the future was glorious, and were starting to experiment with things like socialism to at least fill the material void since the spiritual one was seemingly irreparable. The scapegoat was the Jews, but the actual solution offered was perpetual conquest. Aligning your identity wholly with the fatherland, throwing yourself gloriously into battle, and dying if need be (but really killing your enemy and taking his land) was the thing that filled that spiritual void and gave everybody purpose. In other words, I believe you that Hitler was a fanatic, but as we've seen, his fanaticism could only be realized by a never-ending string of invasions, because his vision of what the Third Reich was could only ever be fully realized in a time of war. If Hitler had tried to govern and enact policies without the militarization, without the constant expansion, people might have grown dissatisfied. And of course, that leads you to go to war with the world, which obviously created more problems than the Nazis could solve.
Nothing is as sublime as:
“Hitler has only one left ball
Goering has two but they are small
Himmler has something similar
And Goebbels has no balls at all!”
(Sung to the tune of Colonel Bogey’s March, as heard in “The Bridge on the River Kwai”).
Surely "Hitler has only got one ball"? Who even has two /left/ ones?
There's hundreds of versions of that song, the one I grew up with (don't ask) goes "the other is in the Albert hall".
This was meant to be a follow-up to the discussion concerning Mussolini’s dick.
Colonel Bogey was the theme tune to an old British kids TV show called “The Machine Gunners”. Had the probably unintended consequence of a whole extra generation of kids learning about the rumored testicular deficiencies of Nazi leaders
Imagine a world with only two types of beings. They know of each other, but are separated by a vast sea and cannot interact.
Type A: sadistic hedonists who are biologically resistant to suffering but capable of great well being.
Type B: compassionate utilitarians who are biologically resistant to well being but capable of great suffering.
Should Type B consider humanely eliminating themselves?
It seems like maybe yes, but what a world to sign off on.
This gave me pause for a moment. I think the best way to rescue the utilitarian view is longtermism. There are a lot of long-term outcomes the utilitarians can effect that are better than the sadists alone would choose to, since they optimize for utilitarian goals while the sadists optimize for hedonistic/sadistic goals.
I think the best utilitarian strategy would be to aim for trans-whatever-species-they-are-ism in the long term, and global control in the medium term.
If the dilemma is rescued by stipulating a stronger form of inability to interact than a vast sea and an inability to meaningfully change the long-term future, then I think the problem just simplifies to “are the utilitarians’ lives worth living in isolation”, since the other species no longer matters. That’s answered by whether or not they’re able to structure their lives such that their equation turns out net positive.
That’s a great answer.
My scenario’s too flimsy to capture what I was after, which is how from a utilitarian perspective, the existence of cruel, immoral actors who are resistant to suffering might be preferable to the existence of kind, moral actors who are prone to suffering.
I'll be in San Francisco in a month or so for a wedding and I'll have a few extra days to explore. Does anyone have any recommendations for restaurants/cafes/neighbourhoods to check out? I've already done most of the tourist activities there, so just looking for any walks/neighbourhoods people enjoy
Go to a show at the Symphony, Opera, or a SF Jazz, go out in Hayes Valley before/after the show. Few different cuisine/price range options for dinner that are all great: RT Rotisserie, Dumpling Home, Doppio Zero.
I've been thinking about emergent misalignment (EM) and more generally emergent behaviors in AI. Some unwanted emergent behaviors may likely be due to training data. For instance, a leaked document from META's chatbot showed that they were giving false medical information and making racialist and/or racist arguments. Of course, there's a huge amount of quack medicine, quack nutrition, sketchy medical studies, and medical conspiracy theories circulating on the Internet, so it's hard for me to see how the training data could *not* get contaminated by quackery and conspiracies. Likewise, the MAGA and the HBD folks have a very vocal presence on social media. Again, for me, it's not surprising that the pseudoscientific pronouncements of the Lynns, Cremieuxs, and Sailers of the world get sucked up with the training data and are then spat out by LLMs that have no capability of evaluating scientific research.
Of course, then there are the weird and unexpected EMs. From the same leaked Meta document, Meta’s AI chatbots, during internal testing, allowed romantic or sensual conversations with children. It's unclear to me whether Meta's alignment policies allowed romantic/sensual discussions with adults, and it just swept children into its feel-good sensual vortex. But it would be deeply weird if it were focusing its romantic/sensual wiles on children and not adults. Does anyone know more details about this?
https://www.reuters.com/investigates/special-report/meta-ai-chatbot-guidelines/
While EMs can be disturbing or even dangerous, I don't buy into EY's handwringing about AIs being existential threats to humanity. The current crop of LLMs haven't evolved beyond stochastic parrothood (per Emily Bender's 2022 observation of their language behavior).
According to Kaczér, et al, KL-divergence regularization toward a safe reference model offers the best defense against EM...
https://arxiv.org/abs/2508.06249
I don't claim to understand how KL regularization works, but if the empirical results show it does, I'm willing to accept that they know what they're talking about. But, given that most LLMs today are trained on the publicly available internet (which includes a good amount of unsafe, biased, dangerous, and harmful content), I suspect that safety measures will have limited effectiveness because of the impossibility of fully curating the training data. So, I wonder if LLMs won't continue to feed us bullshit even when the alignment problem is solved (if it gets solved).
Is it just me, or are all children's learn-the-alphabet toys bizarrely bad?
I ordered https://www.amazon.com/Melissa-Doug-Alphabet-Sound-Puzzle/dp/B0158IMAWO, a puzzle, where you put a letter into the slot and it speaks the name of the letter. But the slots use a motion detector, so if you put the letter on top of the wrong slot, it says the name of the slot rather than the letter. And the slots don't really look like letters and have pictures of objects (eg an apple for A) instead of the letter. So if the child puts things in the wrong slot, they'll learn the wrong letter associations. Also, it says things like "A is for apple" instead of the letter name "A".
Then I ordered https://www.amazon.com/Electronic-Alphabet-Learning-Interactive-Educational/dp/B0CXDK7NSV. But the pictures are much bigger than the letters (the letters are barely visible) and if you press the picture it says the name of the picture (eg "apple") rather than the name of the letter. Also, there are lots of different modes and if the children press buttons randomly it will quickly start doing random things. Also, if you haven't used it in a few minutes it starts complaining and telling you to come back.
Is there any alphabet toy which actually gets children to focus on learning the names of the letters, and which is robust against children who use it slightly differently than intended?
I can’t answer your question, but thought I’d toss out a coupla ideas. General thrust of what I have to say is that for small children I think clarify and purity in phonics instruction (what the ideal gizmo you’re searching for would have) weighs less than the teaching material’s fun and engagement factor. Teaching approaches that maximize the latter have of a lot of “noise,” but the right kind of noise creates engagement, and kids are very good at extracting information from noisy experiences. The language they hear is a hugely complicated chaotic mess, harboring patterns, all with exceptions — and yet they learn the basics without systematic instruction. So here are a few teaching ideas consistent with that belief.
Let them watch Sesame Street. When I was college-aged I taught nursery school for a couple years, and about 30% of the 3 and 4 year olds could read some, and some could read well. Virtually all had picked up the skill from watching Sesame Street.
Day-to-day life is what’s interesting and delicious to them. Play little games based on day-to-day life that involves their reading things. Scatter them through the day, at times when you’re with them.
-When my daughter was 2 I made a grocery shopping list for her with an image of each item next to the word. Few of the words on it were perfect phonic examples, where the pronunciation could be constructed purely by combining letter sounds, but that didn’t matter much. She learned to read “banana” and “soup” anyhow. She’d check off the items on the list she wanted, then help me find them when we shopped. Later I removed the images.
-Find the item: Gave her a one word written clue, like “bed,” and leave something interesting — a little toy, a snack — in the spot. (Or you could leave a little token, and there could be some treat won for getting 3 tokens.)
-Silly games with alphabet blocks. Yup, dumb old-timey alphabet blocks. Come up with a question, like “where do you sleep?” (Or let her come up with one.) Write a silly answer that rhymes with the right one, e.g. “head,” using the blocks. Asked her to fix it using the blocks. Or she can make a sillier answer. If she makes non-word, like “zed” that’s funny too, if you’re little. You cry, “oh I am so tired! I am going to get in my zed, zed, zed. Or maybe my ked.”
I’m with you here. A lot of children learning toys seem to be low quality and lack thoughtfulness in their product decisions. We have alphabet flash cards where some of the letters have the worst/least useful word associations: “b for brownie”? cmon… what about “ball”. These were the best flash cards I was able to find in Amazon with 30 minutes of searching. Definitely room for improvement, there’s probably a market here
It's better than "b is for bdellium".
Speak & Spell is the best one I can think of off the top of my head. There's a remade version that costs about $25, or you can find used vintage models from the 70s and 80s on eBay.
It's aimed primarily older kids who can read well enough to play the spelling games, but it also has a mode where pressing a key will say the letter out loud, and a mode where it shows you random common words and reads them aloud to you. My daughter played with hers quite a bit as a toddler, lost interest in it for a while, and then rediscovered it when she was six-ish.
I think the Speak & Spell was a relatively minor contributor to her learning to read. That was mostly from us reading to her and from her going through a period of wanting to watch Alphablocks and various phonics-themed children's music videos as much as we would let her.
Albert Brooks once used one of those in a Tonight Show bit
You could be a ventriloquist without learning how to speak without moving your lips.
“You were on vacation in Mexico, did you have a good time?”
“C”
“Do you want to say hi to Johnny?”
“Y?”
https://m.youtube.com/watch?v=p6uFHC9lfzk
We have various toys and books that cover the alphabet, and personally I think it would be best not to lean too much on any one tool so that you don't end up overfitting. For example, we have this thing and our son seems to like it: https://www.walmart.com/ip/Melissa-Doug-ABC-123-Abacus-Classic-Wooden-Educational-Toy-With-36-Letter-and-Number-Tiles/50382744
I try to spend a little time with our son every day on this toy, or other similar books or toys, pointing to a letter and saying the letter name. It didn't take him long to figure out what the exercise is about though he is still struggling to produce the correct sounds.
My wife spent a ton of time looking into these, ordered several, and returned almost all of them. This is the one that she approved of: https://www.amazon.com/LEARNING-BUGS-Interactive-Preschool-Kindergarten/dp/B09HR5FPRW/ref=sr_1_5
Great recommendation, saving for later (my child is a bit too young atm). I like the addition of phonics/letter sounds instead of just letter names.
I've been thinking a bit about the future of AI as it currently exists, and have come to a rather unwelcome conclusion.
"One machine can do the work of 100 ordinary men. No machine can do the work of one extraordinary man." -- Elbert Hubbard
LLMs can do an adequate, though never superlative, job even in thinking professions, such as programming computers, producing artwork, legal writing, etc. They will get better, but the way they work, token prediction, isn't actual intelligence, and can't make the leap to produce truly new things. We need another advance in the technology to do that.
So only the most creative/genius people will be able to produce output machines could not. The great majority of humanity wouldn't be able to produce what the machines could, or at least not as fast.
Some jobs today still are tough for robots to do, such as picking strawberries. People are still more cost-effective at these, but as technology improves, even without a huge breakthrough, these human jobs will be replaced by machines. Nonetheless, jobs like these fall into the category of things "100 ordinary men" (or women) can do.
When machines are better than the great majority of people, what purpose do people then serve? My conclusion: creating more people, hoping some of them will be the "extraordinary" people that cannot be replaced by machines. And to support these hordes, machines will do work to keep society running, effectively putting everyone on Universal Basic Income (UBI).
If you're on UBI, what incentive do you have to determine whether you're extraordinary? After all, to find out will require a lot of effort over many years, whereas you could just...not do that.
> If you're on UBI, what incentive do you have to determine whether you're extraordinary? After all, to find out will require a lot of effort over many years, whereas you could just...not do that.
I see this as the same question as "why not just wirehead?"
Sure, a few might choose to do that. Generally, though, humans find meaning in more things than lazing about all day partaking in simple physical pleasures.
"My conclusion: creating more people, hoping some of them will be the "extraordinary" people that cannot be replaced by machines."
Ah, but with embryo selection and genetic engineering, we won't even need all those people to randomly roll the genetic dice the old-fashioned way. We just take all the extraordinary people, extract their eggs and sperm, pick the best embryo(s) of the bunch, let those develop all the way through into adult, rinse and repeat.
Eventually we may have enough superior genetic material we don't even need the adult specimens, we can just clone etc.
Aren't you happy to gaze into the future and see this outcome?
As it happens, I thought hard enough about this question that I wrote a story to explore one (quite possible) outcome. Check it out if you like: https://sisyphusofmyth.substack.com/p/in-the-garden-of-eden-baby?r=5m1xrv
> They will get better, but the way they work, token prediction, isn't actual intelligence, and can't make the leap to produce truly new things.
LLMs are made of math. You're made of meat. There are otherwise moderate-but-not-yet-strong similarities between how the two work. In particular, portions of the meat appear to be doing prediction as a key component of how it navigates the world. There is undoubtedly more that the meat is doing, but many of the core atoms _are_ prediction.
As a personal anecdote, I currently am getting superlative performance out of LLMs as far as programming computers and managing the software development process more largely. Getting this performance has only been possible for approximately the last six weeks, is only possible currently with Claude Code, and even there is definitely _not_ the out-of-the-box experience, requiring considerable customization. I estimate that superlative programming performance _will_ be the out-of-the-box experience in no more than six months (no great breakthroughs needed, just blocking-and-tackling productization).
>They will get better, but the way they work, token prediction, isn't actual intelligence, and can't make the leap to produce truly new things. We need another advance in the technology to do that.
I don't think that's true, in that we have no way of knowing right now what is "actual intelligence" and if token prediction is not actual intelligence. People saying those kind of things usually make predictions that gets proven wrong quickly, and then tend to become extraordinary at moving goalposts.
>When machines are better than the great majority of people, what purpose do people then serve?
Even with machines humans like scarcity/non fungible things, so for example concerts will still be a thing I think. Sure a robot may sing (well, they already do through autotune/playback kinda) but you are in the presence of <celebrity> and that's something a robot can't do since the robot isn't <celebrity>. I don't bet much on the "physical presence of another human", humans don't have infinite endurance/patience and "good enough with infinite endurance/patience" beats human most of the time I think, at least for many people it will.
Also, just because someone is better at something than you doesn't mean there is no meaning in doing it. See chess, go, or anyone doing a sport without having a shot at becoming the best. People like to strive towards goal and be better than the them of yesterday, that kind of things.
Another may be consumption. Maybe for some reason humans are more interesting to make products/experiences for and AI assigns themselves value based on that? Sounds super sci-fi-ish and may never happens, but also, don't humans like doing that? Lots of humans dedicate their life to making stuff, like movies, photos, tiktoks, youtube videos, sculptures, pottery, painting, etc, and people really enjoy having more of those in the world, and niches can be so narrow that there may not really be someone better than you at it. Even with AIs, total computing power of the human civilisation is limited and we have code to write, a lot of it actually.
Possibly humans can fulfil the purpose of being someone's child, but that seems to be dropping by a lot lately.
Still, I feel like this is a trend that has started around the industrial revolution and before it seems like the answer to "how can old people adapt to the new world" was "by making children and dying basically", now with less children and technology moving way faster I don't know.
What makes you think extraordinary people will be better than AI? That isn't true in chess, even though Magnus is arguably the greatest chess player who ever lived
I agree with you, but chess is a bad example, in that it's not even close to being an AI-complete problem. (People thought it was back in the seventies, but they turned out to be wrong; Deep Blue was built at a time when AI was otherwise stagnant, using techniques that don't generalize to most other forms of reasoning.)
Happy to help out with astronomy consulting. I have a PhD in planetary geophysics. I'd be willing to try to help with aerospace stuff but much less likely that I'm capable.
Edit: also possible I could be useful for climate/geoengineering stuff? I got a lot of passive exposure to that just being in a geoscience department. But I'm guessing you have plenty of volunteers.
Oh cool! I'm an Aquarius, should I give up hope for the rest of the month or give it one more shot?
I have long been beset by the sort of insomnia that keeps me from falling back to sleep at 3am (that is, I never have a problem falling asleep at the usual time, but have had periods of several weeks where I struggled to sleep through the night).
Back in March I learned about 'cognitive shuffling' and have since found it about 98% effective - even waking at 5am I can always get back to sleep. Used it this morning to go back to sleep at 4:30a. Not sure how it compares to CBTi, but it doesn't require an app - here's how Gemini describes it:
1. Pick a random word: Choose a neutral word with no repeating letters, such as " Pluto" or "mask".
2. Visualize the first letter: Focus on the first letter of your chosen word.
3. Generate a word list: Think of as many unrelated words as possible that start with that letter. For example, with "Pluto," you might think of "plane, poodle, play".
4. Visualize each word: Briefly imagine each word you think of.
5. Move to the next letter: Once you run out of words for the first letter, move on to the next letter of your original word and repeat the process.
I've never applied 4, I just think of random words. I'm almost always gone by the third letter of the index word. Apparently a lot of people find it highly effective for all types of insomnia, so putting it out here as a public service.
I memorize the letters from the NY Times' Spelling Bee before I go to bed and work out words as I fall asleep. It works great!
I’ve tried the cognitive shuffling a bit. It does seem to help stop ruminating.
This may seem silly, but since the whole verse goes "Whistle while you work/Hitler is a jerk/Mussolini bit his weenie, now it doesn't work", I believe the weenie that's been bitten and now no longer works is Hitler's. That's my interpretation anyway
This makes sense but I was 8 years old when I heard it and was giggling about hearing the word ‘weenie’ too much to really think it through.
As an Italian I can add to this rich literary patrimony by citing that we instead have a rhyme going "Cosa é successo? / Mussolini é caduto nel cesso." (What happened? / Mussolini fell in the toilet).
From a human anatomy point of view, your interpretation seems a lot more plausible!
And you could write essays relating it to international relations at the time, which is always a bonus.
Hello, I don't know whether this comment will be seen and/or answered, but I thought it was worth a shot. I am a long time lurker, still not comprehending a lot of what is being written in this space but fascinated by it. Im also finishing my bachelor in psychology and very interested in what are the best ways to help people with mental ilnesses. Specifically im interested in the debate of psychiatry vs psicotherapy. And which one can help people, in which way, which helps the most and which helps certain kinds of mental ilnessess and not the others. I have yet to see a thorough breakdown of this questions but this feels like the space where it could happen. The reason im posting is also that, beyond a possible career change I am contemplating, I had a discussion today with a professor which strongly favours the humanistic approach to the psychiatric one and believes most of the psychiatric one to not actually help people. Next week we are resuming the debate, but since she has decades of experience and a lot of material at hand, while I dont. i was wondering if anybody could point me in the right direction to find some good material pro and contra psychiatry and/or psychotherapies.
A good place to start is to ask an LLM which mental disorders and scenarios respond best to talk therapy and which respond best to medicine.
Here are some of my thoughts
Talk therapy can't adjust somebody's innate asthetics through which they experience the world. But chemicals often can, to some extent. Extreme fluctuating moods - excitement and mania - can be tempered by chemicals. To a limited extent people with marginally functional reward or attention systems can have improved experiences and performance with chemicals.
But many issues are also a product of a person-environment interaction. When feasible to change the environment, these can be temporarily resolved. Talk therapy can be useful to diagnose these. Also, people who are distraught, about a specific issue and need counsel, but are otherwise able to feel pleasure and pain, can be have their process and identify their thoughts process in talk therapy, sometimes this can lead to corrective actions.
For many individuals, their disorders are untreatable by either. Often chemicals can be given to assist or reduce distress with minor efficacy.
- I think there's pretty strong evidence that both therapy and psychiatric intervention (i.e., drugs) are effective at treating mental illness. In many cases, either one is good, and both is better.
- I don't know of a good comprehensive review of which things tend to respond better to therapy vs. drugs. I would naively expect such differences to basically amount to "which things have we discovered good drugs for," rather than reflecting some important underlying trend among conditions, but that's just a guess and I could be totally wrong about that.
- When I hear someone who is broadly against psychiatry as an entire field/approach (as opposed to narrower criticisms like "this drug class is overrated" or "this particular disorder is hard to medicate" or "doctors' incentives lead to them over-prescribing certain kinds of medication"), I find it hard to take them seriously. I've never encountered a convincing evidence-based case that medication, broadly construed, doesn't work to treat mental illness. Instead, people advocating against psychiatry as a whole tend to be arguing from philosophical positions like "mental illness isn't actually bad and shouldn't be treated, it's just part of the diversity of human experience" or "drugs are unnatural and it's bad to use them to modify cognition, regardless of whether it works." I think these positions are dumb.
- Here are some relevant links to Scott's writing that, while they aren't directly about your exact question, might be relevant to your interests and/or give you a sense of how the rational-sphere thinks about these issues:
https://slatestarcodex.com/2016/03/31/book-review-my-brother-ron/
https://slatestarcodex.com/2019/11/20/book-review-all-therapy-books/
https://slatestarcodex.com/2016/02/24/two-attitudes-in-psychiatry/
https://slatestarcodex.com/2017/12/28/adderall-risks-much-more-than-you-wanted-to-know/
https://www.astralcodexten.com/p/you-dont-want-a-purely-biological
https://lorienpsych.com/2020/10/30/ontology-of-psychiatric-conditions-taxometrics/
https://lorienpsych.com/2020/11/11/ontology-of-psychiatric-conditions-dynamic-systems/
https://lorienpsych.com/2021/02/10/ontology-of-psychiatric-conditions-tradeoffs-vs-failures/
I don't think I understand the specific thing you're trying to figure out. It's hard to assess the value of highly general buzzword things as opposed to specific interventions. If this "debate" is an existing thing, do you have a link to a summary of it or something?
Given that we are in a rationalist-adjacent place, this may be a good time to remind ourselves of the "virtue of narrowness": https://www.readthesequences.com/The-Virtue-Of-Narrowness
For example, more can be said about a specific mental illness or a specific therapy than about "mental illnesses" or "therapies" in general.
I just finished “Rationality: From AI to Zombies.” I am curious how it has aged, but it is hard for me to judge since the writings of Eliezer 2006-2009 constitute the majority of my exposure to many areas. For instance:
* At time of writing, Eliezer had already disowned his pre-2002 opinions. Given R:A-Z completed 16 years ago, that’s long enough for him to have made multiple major revisions since. Has he, or has the Rationalist community in general, revised takes on major elements of the book?
* How has his explanation of quantum physics and many-worlds aged among rationalists and physicists?
* How has his explanation of consciousness and reductionism aged among rationalists and psychologists/neuroscientists/philosophers?
* This was written during the height of New Atheism and contains several New Atheist talking points. That movement lost popularity – in no small part due to crashing on the rocks of wokeism – though atheism is still slowly trending up. Has the median Rationalist’s take on religion, or the conception of social good, changed?
* In the final posts, Eliezer speculates about Rationalist dojos teaching the Way. ACX has meetups, orgs like 80,000 Hours exist, and Effective Altruism has grown into a movement, but Eliezer was making comparisons to the Catholic Church in terms of community-building, mobilization, and so on. Are there any dojos? Does anyone still dream this big?
The sequences cover a lot, and I’m interested in any and all updates since it was written, be it a point above or anything else that you in particular know and are passionate about.
There is a preface written by Eliezer in 2015, where he mentions the following mistakes he regrets: https://www.readthesequences.com/#preface
1) He didn't realize that *practicing* rationality is much harder than knowing the theory, so he didn't put enough emphasis on practice in the book.
2) He chose impressive abstract problems [such as quantum physics] as examples, but now he thinks it would be better to talk about solving problems in *everyday life*.
3) Similarly, too much focus on rational beliefs, not enough focus on rational *action*.
4) A bit disorganized content on the web page, hopefully arranged better in the book.
5) Too much expressing contempt towards stupid ideas. Yes we should reject wrong ideas, but there is a certain tradeoff where too much mockery is bad [Eliezer does not specify how bad, my guess is that it makes it more costly to admit mistakes, and can attract the wrong kind of people]. Luckily, Scott Alexander provided an antidote to this attitude.
Points 1 through 3 would have been incremental steps towards making LessWrong more like a dojo and less like an intellectual community of nerds, but I don't really think they'd have sufficed; if you want to be a dojo I think you have to do a lot of things differently. (And I don't wish he had; it's rare for anything to succeed as much as LW did, and so it's likely that the dojo wouldn't have.)
On point 4, I think Eliezer is just wrong. The densely inter-hyperlinked nature of the original Sequences posts makes them into a trap for the unwary reader who was planning on doing anything else with their day (compare https://xkcd.com/609/), and I think this was instrumental in getting people to actually read them. Rationality: From AI to Zombies is much more of a slog. Furthermore, the blog posts are good blog posts, but the book is not a good book, because each chapter is basically a blog post and that's not a good flow for a book; fixing this would have required substantially rewriting the content to make it more genuinely longform.
Point 5 I read as basically the same point Scott made in https://slatestarcodex.com/2014/04/15/the-cowpox-of-doubt/, which seems slightly different from your interpretation, but I could be wrong. I'm not sure how much difference it made, in any case, especially since Eliezer did not exactly become less habitually contemptuous towards stupid ideas after 2015.
There have been a lot of changes in what's fashionable to talk about, but not that many in what people actually believe. Fewer people argue about interpretations of quantum mechanics, or about consciousness, or about atheism, in our corner of the internet these days, but the basic disagreements are still there, over basically the same questions as in 2009. Philosophy doesn't move that fast (and interpretations of quantum mechanics are philosophy, not physics; no new experimental evidence has come in, except in the sense of experiments that could have falsified quantum theory but didn't).
The one topic where people's substantive views have evolved is community building, because that project was in its infancy in 2009 and we've since spent sixteen years learning what it looks like in practice. Eliezer in particular has lost faith in the project, because he thought that if he taught people rationality they'd come to agree with his specific technical views on AI safety, and that hypothesis has been falsified. (AI safety has of course gained prominence in the discourse, but mostly not in the sense of widespread agreement with those specific views.) As for whether there's anything else to be salvaged, it depends on the specifics of what you're trying to do:
* As noted above, AI safety is now a big field, bigger than the place it started from, with its own exciting new developments and pathologies.
* EA has been pretty successful as a project for building rationalist-flavored world-improvement communities. (People complain about it being diluted by MOPs or captured by society's power centers, but that was always going to happen if the movement got anywhere.)
* The specific project of "learn to avoid cognitive biases"/"rationality as a martial art" has been on a downswing lately, especially with CFAR having curtailed its public-facing activity, but there are still people interested in it. I think this waning of attention was partly because a lot of people weren't all that impressed with CFAR's success at developing a teachable discipline of rationality; they kept talking about publishing their work and then not doing it. (Back when it was more of a topic of discussion, there was periodic discourse from people who'd been under the impression that this was *the entire point* of the rationality community, and felt betrayed that only a subset of rationalists were actually interested enough to devote significant time and energy to it. I feel for them, but don't in fact wish that none of the other stuff had existed, because it's valuable in its own right and probably wouldn't have sprung up counterfactually without the Sequences.)
* There are also lots of little sub-subcultures and social groups where people hang out and have fulfilling social interactions. The one I personally have the most experience with is glowfic; I think there are others I don't know, consisting of people who are into startups or circling or drugs or whatever else. The role of the Sequences was to bring together a lot of people with a lot of different intellectual interests that shared a certain common ethos; once that was finished, having one big community didn't make all that much sense, so instead we became a bunch of sub-communities with some cross-pollination.
There are also some specific claims in the Sequences that have (at least arguably) failed to survive fact checks and/or the replication crisis, but these are mostly not big general theses (though skeptics sometimes argue that these failures should lead us to distrust the entire source, including the big general theses).
Have I missed anything you were curious about?
"Eliezer in particular has lost faith in the project, because he thought that if he taught people rationality they'd come to agree with his specific technical views on AI safety, and that hypothesis has been falsified."
Somehow that reminds me of Martin Luther: "just print the Bible in the vernacular, let the people read it, and they will interpret it their own way, no priestly mediators necessary!"
"Wait, no, not like *that*!"
(Seemingly Luther did not, in fact, say "Every man has a pope in his belly" about the multitude of small new religious sects and sectaries springing up when everyone could interpret the Bible according to their own conscience and, believe it or not, not coming to the same conclusions as Herr Doktor Luther, but it's too good a quote to give up).
Thanks for the detailed response. If others have takes I want to hear them, but this is exactly the kind of response I hoped for.
Is there anything like a directory or explainer for the subcultures and split off groups?
I don't think so. There was https://slatestarcodex.com/blog_images/ramap.html but it's eleven years out of date by now.
"and interpretations of quantum mechanics are philosophy, not physics; no new experimental evidence has come in, [...]"
A side note, just because:
Take a specific experiment: Alain Aspect's seminal work from 1982 (for which he got the Nobel Prize two years ago). The experiment shows convincingly that if the Copenhagen Interpretation is used, information would have to travel 10 to 50 times faster than light from cause (observation at point P) to consequence (collapse somewhere else).
Does that refute the Copenhagen Interpretation? It does to me, but proponents of the interpretation just shrug and say "yeah, QM is non-local, so what?" or "yeah, true, but you can't use that trick to send information yourself so it doesn't have consequences."
Well, obviously you can't send information faster than light, we know that from other theories, but then why is it necessary for the Copenhagen Interpretation to work?
The MWI doesn't suffer from that flaw. When you measure part of an entangled system you learn (get information) from that system, namely which branch of the entanglement ("world") you (that particular component of the superposition that up to that moment made up "you") live in. No information travels from the measured object to any far away object.
So. It is not that the Interpretation of QM is not physics (it is, at least with my definition of "physics", which has to do with understanding reality). Instead, there are deep psychological hurdles to accept an interpretation that challenges our sense of identity, even if it is the interpretation the most compatible with known physics, the one that doesn't deny objective reality, that is deterministic, and local. (It is also quite possible that historic contingencies are a better explanation for why CPI is preferred; I can talk about that too.)
Newtonian Mechanics has 3 laws:
- Existence of an inertial frame of reference
- F=ma
- action = reaction
Now, imagine that I have a new interpretation of Newtonian Mechanics:
- Existence of an inertial frame of reference
- F=ma
- action = reaction
- all forces above are invisible angels that move instantaneously from place to place and pull and push the amount required by these axioms.
Nobody would take my new interpretation seriously. It has more axioms that the original, and it violates known physics (e.g. angels travel faster than light).
But my answers to criticisms would be "you can't talk to the angels so you can't use them to send information anyway", and "I make the same predictions as you, so both interpretations are equivalent", or "well, it's just philosophy anyway."
This is exactly the situation we have with the Copenhagen Interpretation. MWI has fewer axioms, is compatible with realism, determinism, and locality, but we still insist that angels must be flying faster than light to carry information from place to place.
As I'm sure you're aware, the Bell Inequalities exclude local realism, not simply locality. No physicist thinks that rejecting MWI means abandoning Lorentz invariance.
My point is that everyone, including people favoring every interpretation, correctly predicted the results of Aspect's experiment, because the only thing he tested was whether quantum mechanics would continue to observably behave in the exact same way that it has been understood to behave since 1925. And surprise, it did. (Not to suggest that it was worthless; the theory hadn't been tested in a setting where we'd notice a speed-of-light delay, so it was good to get confirmation that this doesn't change things. But it wasn't a *surprise*.)
The arguments are all about what counts as most parsimonious, what conclusions about the nature of reality are valid to draw from what experimental observations, etc. This work is legitimately intellectually interesting and valuable (well, at least some of it is), and some of the arguments that people make are better than others; I don't think it's necessarily a mistake to have a preferred interpretation. But it's not science, it's philosophy of science.
Aspect empirically proved non-locality of the Copenhagen Interpretation, that's all. At the time his results were framed as a refutation of Einstein's local hidden variable theory, and it absolutely *is* a refutation, but they also emphasized something "wrong" with the Copenhagen Interpretation, at least if you care about locality.
As to what counts as parsimonious, it seems self-evident that the number of axioms is the first step. (But it is not in this case because, I think, people confuse "parsimonious" with "intuitive".)
What would convince a proponent of the Copenhagen Interpretation? (Or of Newtonian Mechanics with invisible angels?) Imagine that an experiment shows that the interpretation breaks another principle, let's say conservation of energy, the proponents would just say "so what, you can't use that yourself to create a perpetual motion machine, so no paradox here." Of course, you can't, because that's just not physical. But then why does your theory requires that energy not be conserved?
So yes, I understood your point well. I agreed with your first post that people have not changed their mind; I only made the side note that they wouldn't have changed their mind even with new experimental data, because the reasons people believe what they believe are not entirely determined by experimental results.
I’m not that familiar with many-worlds. Why doesn’t branching the entire universe require faster-than-light information? Since it’s just an interpretation and not new physics, it seems like it would still be somehow non-local, but handled differently?
> Why doesn’t branching the entire universe require faster-than-light information?
Because it doesn't happen at the same time everywhere. The branching starts locally (when two particles either interact or they don't) and from there it propagates outwards at the speed of light. Except that there are zillion bubbles like this spreading simultaneously.
In the famous cat-in-the-box experiment, as long as the box is hypothetically isolated from the rest of the universe, there are two universes in the box but only one universe outside. When you open the box, the branching continues to propagate outside the box.
As a visual help, instead of imagining two branches as two parallel sheets of paper moving further away from each other, imagine two stickers being peeled off each other. At the place you are pulling them, they are already separated, but sufficiently far away from that place they are still together.
But this is all just an imprecise metaphor. It does not explain e.g. how the universes *interfere* with each other, and that is the key part of the quantum physics. One universe splitting into two is a relatively boring thing, and easier to imagine. Two universes canceling each other, because they have the same contents but an opposite phase, that's the fun part that is difficult to imagine by an analogy, because it has little analogy in the classical world.
I think a common misunderstanding of MWI is that a brand new universe is created every time someone observes something. Not so. The superposition already existed, but the components weren't distinguished.
Maybe this comment I made a few weeks ago will help:
https://www.astralcodexten.com/p/open-thread-393/comment/142906889
I actually disagree that MWI and CI are the same physics. They are different explanations with a different set of axioms. And yes MWI is local, as I explain in the link above, I think.
(Maybe MWI is wrong; there are *other* reasons to be skeptical; but the usual arguments are unsound.)
How does the CBT-i app stuff differ from whatever program is used by Sleepio? I found that very useful back in 2020, and they never required a prescription, though I don't know what it cost (it was paid for by my work and offered as a health benefit).
I'm not familiar with Sleepio, but I did a course of CBTi almost 20 years ago. It worked, and it's insane that anyone is charging $300/mo for it, because it is an incredibly straightforward protocol that can be summarized in a fairly short pdf. Maybe things have changed significantly in the past 20 years, and I guess I can imagine a fancy version of it that involves an automated sleep tracker, but even that wouldn't get you close to $300/mo. It's basically a combo of sleep hygiene stuff and standard CBT stuff. In a very small nutshell:
1. Track how long you are actually asleep each night.
2. Give yourself that amount of time + 30 minutes in bed every night, with rigid bedtime/rise time schedule. This is brutal but effective.
3. Do some CBT on counterproductive thoughts like "I'm never going to fall asleep" or "I'm going to be a wreck tomorrow"
4. Gradually extend your sleep window as the percentage of time you spend asleep while in bed grows.
There's more to it than just that, but that's the crux of it, if I'm remembering correctly. The only way you could possibly get to $300/mo is if a) it's a prescription program that insurance is paying for and/or b) there is an actually doctor at the other end of it reviewing your progress and providing feedback.
The point that it's simple does not directly have anything to do with the price; a thing is worth what people will pay for it, producers charge what the market will bear. If I needed this thing, I would easily pay $300/month for it if I couldn't have it for cheaper.
The right framing is, since it's so straightforward to produce, why hasn't the price been competed down? And it sounds like the answer to that is largely just that it's niche, without that many providers and without widespread enough demand to entice people who are currently doing something else into the market.
I understand how pricing works. I create software products for a living, have an MBA, etc. This is not worth $300/month. There isn't enough value created, there are credible alternatives, the marginal cost of delivery is too low, there is no defensible proprietary advantage that I am aware of, etc.
I don't know what is going on here, but my guess is that it has something to do with regulatory approval/arbitrage.
It sounds extremely similar.
What may end up outperforming CBTi at treating insomnia is targeted adjustments to the gut microbiome, at least according to this new Mendelian randomization study:
https://gpsych.bmj.com/content/gpsych/38/4/e101855.full.pdf
I'm posting about this study on my blog Thursday evening. One takeaway is that to the extent that it's possible to change the levels of specific gut bacteria without creating unhealthy imbalances, these changes might help some (if not all) chronic insomniacs, roughly analogous to the way that antidepressants can help some (if not all) people with clinical depression.
In a different response to Scott's embryo selection article, I wrote a weirdly-long post on why we use exponential time discounting in general and why he's wrong not to use it when computing the cost/benefits of embryo selection
https://open.substack.com/pub/shakeddown/p/exponential-discounts-the-future
(This makes the cost/benefit ratio come out bad when used correctly, which I find weird - the benefits seem like they should be no brainers. Maybe the IQ benefits are just severely underpriced in Scott's review, so much so that even after discount they come out ahead? The health benefits might plausibly be as small as described though, especially if they're mostly just maybe-pushing people from just over to just under the clinical diagnosis line).
The recommended CBTi app seems to revolve around an AI sheep, I assume that back in 2021 CBTi did not resolve around LLM chatbots as they barely existed. Is this really the same thing?
It's not that uncommon these days for app developers to inject LLMs into things that were previously done without them, believing (rightly or wrongly) that this improves the user experience.
It's not even that, app developers often shoehorn the current trendy technology into their product to get funding. A while ago everything was on the blockchain, there was a time when VR was the hot new thing, and then there was the NFT fad. When the people writing the checks want your app to include X, it is hard to say no.
Is this thing venture-funded? I didn't see any particular indication of it.
Also, inserting an unnecessary LLM (or "AI") is often seen as a marketing plus.
Much like doing stuff on the blockchain when a small relational database would work fine was a "thing" ~5 years ago.
There are many CBTi apps available now without a prescription and for a similar $300 price range, although their price is less clear when covered by insurance. For example, Sleep Reset, Stellar Sleep, and Sleepio all offer something similar, and they also claim to be backed by Stanford, Harvard or science. There's also the free CBT-i Coach from the VA.
The issue I've found with these (and likely CBTi and sleep restriction in general) is they don't seem to work as well for early morning insomnia. The apps themselves may also not be as well-tailored to your specific issue as a sleep therapist, even if many of them still use human sleep coaches.
CBTi - there appears to be a free version
https://freecbti.com/
Something I think has slipped under a lot of people's radar is the impending "temporary" emergency scheduling of the newly popular legal opiate 7-OH. It seems this is inevitable and I could see it pushing the opiate crisis into overdrive. This is a compound found in small amounts in Kratom but it's obscenely powerful in pure form and now available at most gas stations where I live. A lot of addicts have switched from counterfeit heroin/oxy/whatever, that is to say fentanyl, to this compound. The advantage being it's rather euphoric, cheaper, easy to come by, strong enough to appease all but the hardest users appetites, not cut with anything else, and despite trying really really hard no one has yet managed to die from an overdose. To top it off it's very addictive both psychologically and physically and can cause physical withdrawals after just a week or two of use, so in addition to these old time addicts there are a bunch of newly minted ones. When they get cut off I fear many will turn to fentanyl and we'll see a bunch of overdoses from people trying to match the high or avoid withdrawals.
Are you comfortable saying which state you live? I don’t see anything like this in mine. In the shops where I occasionally pick up THC gummies I see kratom for sale put not this potent extract.
I'm in Texas. Not an extract, it's a semi synthetic usually sold in pill form.
> and despite trying really really hard no one has yet managed to die from an overdose.
It looks like this might be about 4 hours out of date: https://richmond.com/news/local/article_5be0b9ee-de6b-4b35-b64f-ace154b99ce4.html
>With $8 in his pocket, he bought a black packet from behind the counter: a tablet sold by the distributor Pure Leaf Kratom
>Allshouse bought a 30-milligram tablet.
>As police await a toxicology report, a cause of death remains unknown, and the Henrico Police Department would only say that drug paraphernalia were found in the van beside his body.
I think we can say with some certainty this man didn't die from an overdose of 7-oh. 30mg would be a large dose for a person without any tolerance to opioids, but not enough to cause unconsciousness let alone death. There was paraphernalia, perhaps he tried to inject the tablet and died from the shot? But the substance is hardly soluble in water so in that case he most likely died from an air embolism or whatever was used to bind the tablet.
The article contains misinformation that anyone with Google should have caught, for example calling it a hyper concentrated form of Kratom, which is like calling heroin a hyper concentrated form of poppies. I predict a lot of these types of articles as we get closer to a national ban.
By the way, in case it wasn't clear, I strongly recommend people not touch this compound even in moderation, unless they are already opiate addicts and cannot quit or obtain a prescription for maintenance.
Junkies generally don't only do one drug at a time, so yeah, very likely he had a ton of other stuff in his system and this was just the topping on the dessert, as it were.
I suppose it depends on what you mean by junkie. The least functional drug users, like those unable to hold down a job because of their addiction, will usually be spending all their money on a single "drug". In quotes because many street drugs are now kind of a crap shoot, like illicit "Xanax" for example is always going to be some random RC benzo. "Heroin" is usually fentanyl but it might also be a nitazene compound or have Xylazine in it, it might even occasionally be actual heroin.
Most exceptions to this will I think be mixing in just alcohol or weed, of course a lot of drugs don't play well with alcohol, and it should go without saying none of the above applies to the wealthy.
And you're certainly right that a lot of heavy drug users are poly substance abusers. I don't have the stat at hand but I think the majority of overdose deaths result from mixing substances. That might have flipped with fentanyl and nitazenes in the picture, neither of which have any trouble killing a person flat with < 1mg
What does the name of this substack mean or stand for?
Previous site was Slate Star Codex https://slatestarcodex.com/ and this is an anagram of that
Reminds me of when Metamagical Themas took over for Mathematical Games in Scientific American
Funny, I tried "decyphering" the anagram and, unless I did it wrong, I'm missing a second "s".
Does it technically count as an anagram if you have to "repeat" letters?
"S" was actually his ostensible middle initial ("Scott S. Alexander"). In reality, his name is Scott Alexander Siskind, as he publicly revealed in his first post on ACX.
The actual reason it's not a proper anagram is because he had to drop an N.
... is that the reason the icon for SSC has an N on it?
I believe so, yes.
Nah, Slate Star Codex was a "near-anagram". Scott fixed it for the new blog name.
It is an anagram of Scott Alexander.
Wow! That's really cool!
I still think superintelligent AI is not the danger, it'll be bog-standard stupid AI of the kind we have today (such as the thing that wiped the database) and people turning more decision-making - or even just 'get the AI to do the job' - over to it.
"Oopsies! I wiped the safety protocols for the process. I panicked, me so sowwy, tee-hee!" and there's a smoking hole in the ground where the processing plant used to be. That kind of dumb accident because we anthropomorphise *everything* and if the makers of the AI program it to pretend to be a person (so the rubes can interact with it like it's a human instead of the human it replaced), then we will convince ourselves it's a real human with the same capabilities as a real human, and we'll forget it's a box of gears, and we'll assume "well even the dumbest guy we hire to run this process would not be dumb enough to wipe all the safety protocols" and then we end up with smoking hole in the ground. And then the AI will issue a fake "I panicked" 'explanation' as to why this happened, as if it had a mind or feelings or emotions, and that is supposed to be enough to reassure us that this thing can indeed think instead of simply spewing out the script it was trained to produce.
AI is not the danger, we are, because we are stupid, greedy and vain.
Yes, as our tools get more powerful our mistakes become more impactful. That's ok because its more than compensated for by the higher average output. Should we go back to hunting with spears just because there's a chance that someone will accidentally shoot themselves?
This depends on the risk-benefit profile of the individual tool. E.g., as Kelsey Piper has explained on Twitter, doing gain-of-function research on respiratory viruses has genuine scientific value that can improve our ability to fight disease. Yet we shouldn't do it, because the extent of that scientific value is not high enough to compensate for how bad it would be if something went wrong.
What does this have to do with superintelligence? Like it seems like both situations could be true, where someone dumb uses AI badly to own themselves, but superintelligence is still the same old danger
In past discussion, you gave that link to that website: https://brght.org/iq/country/ which has inflated scores all over the board.
Averaging five largest European countries (Russia, Germany, UK, France, Italy) gets 108.35 score, Ireland is 101.95 in this dataset. So Ireland is lower just like it was in Lynn's...
Yeah, but we're "within normal range" lower, not "intellectually retarded" lower. 100 being the normed value, if we Paddies score (rounded up) 102 IQ, it means we're... average! normal!
Yay, we're normal! 😁
(If you seriously thought I was ever boasting of my gigantic brain, you got off at the wrong exit, brother).
Do you really endorse the "the danger isn't X, it's Y" construction?
For example, would you endorse "the danger isn't smart humans, it's stupid humans"? Stupid humans can definitely cause some problems (like car crashes, or equipment malfunctions). But smart humans can also cause some problems (like inventing the nuclear bomb, or lab leak). These are both kinds of problems we have to deal with! Saying "the problem isn't smart humans, it's stupid humans" is not even wrong.
In the same way, I'm sure there will be stupid AIs who cause the AI equivalent of "human error". These will cause the AI equivalent of car crashes and factory explosions. But damage is kind of limited, in that if too many factories explode, we won't use that AI. The damage from smart AIs, the equivalents of the ones that can invent nukes or do lab leaks, seems unbounded, so I am more concerned about that.
I think it's much more likely that stupid AI will be widespread, and widely used, and we'll manage to screw up a lot of small but important things, very much important when all are summed up, before we get super-duper no really it's a thinking entity AI at all, much less turned loose on the world.
More people have died in car accidents than in wars:
https://scrantonlawfirm.com/death-war-vs-death-motor-vehicle-collisions/
The world has been safer & more peaceful since the invention of the nuclear bomb.
Only in our particular Everett branch/possible world, and only for the last 80 years. If a nuclear war occurs in the future (a risk that only started when nukes were invented, and will remain until we either all die or create a powerful aligned AGI or similar that can stop nukes), then nukes will have been very bad for the world's safety and peace. Likewise if there've been close calls where nuclear war was only averted due to dumb luck or the anthropic principle; there have been two well-known candidate such incidents and a number of lesser-known ones.
(Also, this is beside the point of AI safety, as a nuclear arsenal controlled by an unaligned AI is an obvious danger to humanity in ways that the game theory that currently governs nukes doesn't necessarily mitigate. But I don't know if you even disagree with that.)
For what little it's worth, I'm not sure superintelligent AI is possible. I think stupid use of existing AIs is already happening, and we're barely starting to see the ill effects.
The amusing? possibility is breaking civilization to the point where really capable AIs aren't possible.
The problem isn't level of intelligence or intent; what matters is the orthogonal problem of power, the ability to affect things we care about. Few would complain if you keep your hyper-intelligent (and/or hyper-malicious) agent in a box, airgapped in a secure cleanroom, akin to a bioweapons lab. In there, the agent can be as intelligent and malicious as it wants to be.
However, if I give an AI control over my nuclear arsenal, either directly or in a sufficiently trusted advisory role, then the worst case of global thermonuclear war is the same whether the AI is malicious or stupid. If "a few factories explode" is your worst case, then it's because you didn't give the agent more power than to blow up a few factories.
Yes, we are already liable to both factories blowing up (Chernobyl, Bhopal etc.) and global thermonuclear war just fine without any AI involvement. However, as the saying goes: computers make very fast, very accurate mistakes. This includes AI and is strictly worse than making slow, inaccurate mistakes.
> Few would complain if you keep your hyper-intelligent (and/or hyper-malicious) agent in a box, airgapped in a secure cleanroom, akin to a bioweapons lab. In there, the agent can be as intelligent and malicious as it wants to be.
Eliezer would like a word with you, I think.
Having an AI system in a box is safe if you do not talk to it. It is also completely useless.
To make use of an ASI, you need to talk to it, and implement some of its suggestions for the problems you are facing. The suggestions are likely going to be plausible and work on the short term, but they might have side effects which maneuver you into a situation where you will eventually release the AI from its box.
As an intuition pump, imagine a bunch of 6yo shipwrecked on some uninhabited island together with a highly intelligent psychopath with multiple professional degrees in a straight-jacket. The kids know that their prisoner is a very persuasive murderer, and have a firm pre-commitment not to let him go free. However, they also depend on his knowledge for survival. My prediction is that this situation will go poorly, and sooner or later, the psychopath will be able to get some of the kids to free him. Perhaps he convinces his guard that he is actually harmless, or engineers a situation where everyone will die unless he is released (and trusts the kids to follow causal decision theory).
In this scenario, the intelligence gap between the kids and the psychopath matters. If the psychopath is just a median 10yo, then it seems much less likely that he can outsmart the 6yo's. Likewise, humans would probably be able to keep a von-Neumann level intelligence boxed in. As far as we know, no highly intelligent humans have taken over the world, after all. But ASI is another ballpark. At best, it is "the smartest human alive ever but also 1000x faster". At worst, it will make us look like chimps in comparison.
"As far as we know, no highly intelligent humans have taken over the world, after all."
But the ones who had limited success towards that goal were not the most intelligent, rather they were extremely persuasive and/or had a particular talent that enabled them to mobilise lots of ordinary schlubs to do their bidding.
That's what I'm talking about; not the von Neumann level AI plotting our downfall, but all the 105-120 IQ people in business, government, and private users who make use of the multiplicity of new AI models on offer to replace humans, take over human work, and be the first port of call for advice, research, and therapy/emotional support.
Any one error on its own won't be immediately fatal, but add them all up? Like the guy who got advice from the AI about cutting out salt from his diet and ending up dosing himself with sodium bromide instead of sodium chloride, and that despite the AI including that this was for cleaning purposes:
https://www.acpjournals.org/doi/full/10.7326/aimcc.2024.1260
This is how we're gonna do it: "well the AI said it so it must be okay and I don't need to check it out any further"/"don't bother checking the AI results, that only wastes time and reduces efficiency, if it says do this then do it, Jenkins!"
The stupid (by comparison with the Platonic Ideal superintelligent AI) AIs are *already* out there, *already* being pushed by businesses, *already* in use.
Recreate "Five Nights At Freddy's"/Ray Bradbury's "The Veldt" in your own kids' bedrooms today!
https://techcrunch.com/2025/08/16/ai-powered-stuffed-animals-are-coming-for-your-kids/
https://www.rte.ie/news/business/2025/0612/1518159-barbie-maker-mattel-teams-up-with-openai/
LLM psychosis for four year olds, it's the must-have toy every parent is getting this Christmas!
A bromide is also a cliche, and people are already suffering from cliche overload.
I wrote the Ollantay review. Wasn't planning on responding in the comments (seemed inappropriate), but as the Comment of the Week was skeptical I figure I can respond here. Perhaps this should have been a footnote. Ah well.
The way that the story of the rebellion is usually told is something like this: Tupac took some incredibly bold moves and gathered a huge army that would have crushed Cuzco had he continued the initiative that he already had (more than half of the curacas in Peru publicly supported him!). But instead he got cold feet or became a coward. Sometimes he's compared to Father Hidalgo, who also gathered a giant army but failed to march on Mexico City, but while Hidalgo seemed to be concerned about the giant social revolution he was accidentally leading, Tupac was perfectly content to watch the social rebellion unfold chaotically and violently. The final months of rebellion do not make sense on their own.
But then I read Ollantay, and then I learned that the early European translators and commentators of the play (Markham and his contemporaries) linked the play to Tupac Amaru. In his introduction to the translation (https://www.gutenberg.org/files/9068/9068-h/9068-h.htm), Markham calls out the Tupac Amaru rebellion directly and does in fact say that Tupac watched the play in 1775; implied is that this is a big reason for his interest in the play! And it seems as if that 1835 periodical that brought Ollantay back in the public consciousness did the same (I haven't been able to find a copy of the article and can't get to Lima for a while). This introduction doesn't specifically say that Ollantay inspired Tupac, but he primes us to read the play with Tupac in mind.
And when we do, we see all those connections that I noted in the review. And then we ask "wait, did you say this play happened *before* the rebellion?" Which all of the early translators do indeed say. And they note the provenance of this claim (priests and monks and family members). And for me, that answers the major questions of the rebellion.
So Garald is sort of correct that we have no direct evidence that Tupac saw Ollantay on stage. But we do have pretty good indirect evidence (Markham said that Tupac saw Ollantay, so somebody told him that, but he doesn't cite his source). This review was an argument that he did see it and that it changed his life. I structured the review as a story with the argument assumed, and not as the argument itself. I think it flows better that way, but it caused some alarm bells to ring and I can't be mad about that.
It may also have been an assumption on his part that he didn't *need* to march on Mexico city, the authorities there would see that their power was over because of the successful (up till then) revolution and besides the king would back him up, so why fight a battle he didn't need to fight?
Of course, it turned out he *did* need to fight it, but we don't know exactly how deep into his fantasy he was; he may well have been convinced that God was on His side, or at least destiny, and that it was inevitable he would win as the sacred rightful descendant of the emperor or something.
> that he didn't *need* to march on Mexico city,
I take it you meant Lima.
Here's a monograph (just 8 pages, in Spanish) about a previous (largest?) indigenous rebellion in colonial Peru, which might provide some context for the later Tupac Amaru rebellion, and also mentions the importance of taking the main centers of power (Cusco and/or Lima) for the rebellion to succeed:
https://beta.acuedi.org/storage/books/pdf/4951.pdf (*)
A few relevant quotes (translated from Spanish):
"[the rebells] controlled not only the province but also the roads that transported the products that supplied Lima's markets. As Mela says, most of the mountain products arrived in Lima via the narrow paths of the Lurín and Mala rivers, and by cutting the bridges used by muleteers, the rebels could cut off Lima's contact with the interior."
"...the residents of Lima, one of whom, writing to a relative in Spain, told him that 'if they [the Huarochirí Indians] jump to Tarma, Jauja, and Cusco, the kingdom of Peru will end for the Spanish.'"
"The viceroy noted that Huarochírí was the gateway linking Lima to the interior, adding that 'if this province continues in rebellion, its proximity to this capital would become a refuge for criminals who would disturb its tranquility and cause much damage.'"
(*) Rebelión colonial: Huarochirí, 1750, by Karen Spalding
Yes, Lima. No idea why I said Mexico City. Clearly I am not cut out to be a South American revolutionary, I would be not attacking the wrong city I intended not to attack!
I loved the review and wasn't highlighting Garald's comment as a criticism, just to provide extra perspective. I've edited your response into the relevant section above.
I'm a long time lurker who finally pulled the trigger and started a blog. Here is our first post: https://whattotelltherobot.com/p/elephants-dont-write-sonnets Elephants Don't Write Sonnets, about why LLMs aren't *really* intelligent and what it would take to cross that line, writing from a robotics perspective.
The post doesn't really present any evidence for why we should think that embodiment matters beyond Rodney Brooks saying it does. Its response to all the accomplishments of the past few years seem to be that this isn't "true" intelligence, but that embodied intelligence would be, without predicting what things that current AIs can't do an embodied AI would be able to do.
I'm more likely to do modus tollens here and think of Rodney as one of the many people with bold theories of how you can't do AI unless X which were totally falsified by later events. I think his 2018 predictions, which he commendably updates every year, broadly confirm this story - for example, he said you wouldn't get an AI that "seems as intelligent, as attentive, and as faithful as a dog" until 2048. I think this shows that Rodney wasn't expecting AIs which were competent at basic tasks but lacked some kind of "true intelligence", and that positing the "true intelligence" thing is a retrospective invention intended to explain away why AIs that don't follow his embodiment theory are succeeding beyond what he predicted.
I think if you are going to blog about this, you need to address this perspective rather than taking Rodney's perspective as a given (unless you intend for your blog to be just be a discussion forum for people who are already operating within that perspective).
Thank you for this comment! Engaging with you and other readers of ACX was a major motivation for me pulling the trigger on this, so I am *extremely* excited to get feedback from you! Is it TMI to say I was dancing around our living room when I saw you had replied to this comment?
As it happens, we did address this in section 1.3 of our chapter, which is linked from the blog post and also linked here: https://h2r.cs.brown.edu/wp-content/uploads/tellexwatkins2026.pdf , but we didn't mention this part much if at all in the teaser post. We will plan another post just on section 1.3 based on your feedback here. TLDR: We define behaviors in terms of embodied language (language paired with actions/behaviors in the physical world), that robots/AIs don't yet do, along with learning architectures that point towards doing them.
I've been doing language+robotics for 15 years so I promise it's not a retrospective invention for me, although I won't deny a lot of angst as many of my predictions have been falsified in the past 5 years. :-) Regarding Rod's dog example in particular, my lab is collaborating with Daphna Buschbaum who studies human-dog dyads, so we can try to make our quadruped robots produce similar behaviors, for example to interpret human pointing gestures to find objects. I think Rod would say that behaviorally, the AI needs to be attentive and faithful *in the physical world*, which means processing high-dimensional space/time input and producing high-dimensional motor output. Not just typing to you, but walking up to you, looking at you, nudging your hand for pets, and barking at the big dog coming down the street. I am confident in saying AIs don't do this yet (while being sure an LLM could *talk* about doing it). But Rod has a chapter in the same book as us, so maybe he will say more there!
Can you play Magic: the Gathering over the phone, but without a referee and provably without cheating? Something equivalent to blindfolded chess, with the obvious complication that a lot of information in MtG can transition between various shades of hidden and public, whereas all information in chess is public.
For example, a game could start by each player assigning a generic ID to all the cards in their deck, publish that assignment except it's encrypted, and as long as the library remains shuffled, drawing the top card as normal is mathematically equivalent to drawing a random card that both players agree on by exchanging random numbers. At the end of the game, each player would reveal all their hidden choices and it can be proven that cheating has or hasn't occurred.
Is there such a protocol that covers all possible actions allowed by Magic cards, or is there proof (as in an actual paper) that such a protocol can't exist?
Yes, this is definitely possible. What you are looking for is called secure multi-party computation, and there are protocols like Yao's garbled circuits that can be adapted to not just play MTG, but to perform *any* computation with someone else without revealing your inputs to the computation.
The most straightforward construction would be to run a multi-party computation of an MTG game engine, where the RNG seed is given as the sum of two random values you and your opponent provide. There are still some implementation details that need to work out, but they are not too bad:
1. You need to run multiple rounds of computation, not just a single input/output. You just keep an encrypted state variable that is fed from one computation to the next (again encrypted with a key derived from a combination of secrets you and your opponent provide).
2. You need the game engine to provide private information to just one of the players. To do so you just have it output that information after encrypting it with a key that only one player has.
There are probably more efficient schemes which take advantage of the specific structure of the MTG game. (Also note MTG is Turing Complete)
I'd say no unless you set some limitations. How would you deal with Chaos Orb or Falling Star? You also need to assume trust on possession of cards or allowing proxies I assume, not only because of the initial state, but because Research/Development exists.
Now, if you keep it to tournament legal I'm not really seeing any problem, you'd have to keep track of card order in some scenarios, you'd also need to be able to generate random numbers for coin tosses or "select a random card" effects, but beyond that I'm not really coming up with any issues.
Chaos Orb, Falling Star would not be implemented, true. Legal ownership is out of scope, I only care about the game mechanics.
You could track possession of cards by having a trusted third party (maybe WotC or a tournament league) verify that you own the card and issue a cryptographically signed certificate that allows you to use it in the app.
This could still be cheated by selling the physical card after you get the certificate, but that could be mitigated with expiration dates that force you to revalidate periodically. Setting the expiration date to be short enough to be a useful control but long enough that the hassle of revalidating isn't too burdensome on both players and certifiers is a nontrivial problem. You might be able to mitigate by having common cards be taken on trust or non-expiring while only particularly rare cards would need to be revalidated annually or more often. Or you could have the expiration be a soft rather than hard limit, with you and your opponent agreeing to a limit of how old a certificate each of you will accept from the other as part of the terms of the game.
Another alternative would be to decouple app cards from physical cards and use either NFTs or a central registry to track ownership. In the latter case, you'd still use signed certificates so you can verify ownership without pinging the registry every time, but the registry would make validation trivial and make it feasible to set the expiry to a period of days or weeks instead of months or years. I think I remember someone I used to follow on livejournal talking about working on an app that worked on the central registry model, some time around 2005-2008.
Does this at all stop people from looking at their library to see what order the cards are in?
Each player shares salted hashes of the cards in their deck with the other player, which the opponent's app shuffles and tracks. When player A draws from their deck, their app asks player B's app which card they draw (by hash) and player A knows which card this is. Player B can confirm that the card actually corresponds to the hash, but doesn't know in advance because of the salting.
I don't think this is a complete solution for MtG, since I seem to recall there being situations where a specific card gets shuffled back into the deck, since player B could then identify that particular card in the deck and track or futz with its position.
I think that if you shuffle a card back into the deck, you just rehash the entire library, i.e. apply new salts to the hashes (or to the revealed card) and provide these as hashes.
For a Scry N, your opponent will just give you N hashes, and you will tell them in which order you place them before and after your library.
I am unsure if there is an effect which will pick a card where neither party knows what it is, but that could be simulated by your opponent hashing the hash they picked for you.
(Of course, your opponent could also provide you with a shuffled library that way -- simply an ordered list of salted hashes of your salted hashes of your cards. Whenever you draw a card, they just reveal how they generated the next hash on the list.)
If this is doable in practice depends a lot on if the telephone refers to a smartphone or just a voice call, because most humans are not great at calculating cryptographic hashes in their head.
When I say "playing over telephone" it is a metaphor for "not being there physically" to check the legality of play [1]. Of course you'd use computers to communicate and to do the mathematical operations.
[1] https://dl.acm.org/doi/10.1145/1008908.1008911
I was assuming smart phone. For a voice call, I don't have any immediate thoughts on how to do it more securely than the honor system.
Yes, in the sense that it would be useless to the would-be cheater. The protocol, if it exists, would ensure that such hidden information is either unknown to players who are not entitled to it, or that it would be useless because the mechanism of choosing a card would not rely on the current order.
Let's say my library consists of Forest, Forest, Grizzly Bear. I assigned IDs 1 through 3 to these cards. Both I and my opponent know that cards 1-3 are in my library. I know that the Grizzly Bear has ID 3, but when it's time for me to draw a card, I don't decide alone which card to draw. My opponent and I would, through the protocol, agree that I'm going to draw card #2, because neither of us knows the order of cards and the order of cards in the library can be abstracted away.
I could of course claim that I drew Grizzly Bears instead of Forest, but I have previously committed to the assignment and the protocol would require that I disclose it eventually, making the cheat provable.
The problem isn't illegally drawing cards. The problem is knowledge of upcoming draws that should be hidden is valuable and influences decisions.
Let's say my opponent taps out to play a very good creature. Do I counterspell, or hold my counterspell for a potential future better spell? If I know my next draw is creature removal, I just draw and use the removal. If I know my next draw is another counter, I probably counter. I think that alone makes such a protocol impossible (without continuous camera + machine vision proctoring to ensure hidden information remains hidden).
Also, many cards make you shuffle, and that's a pain. You have to let the protocol shuffle, and then rearrange your deck according to the new randomized order, which for 20-90 cards (including commander) is annoying. This second concern doesn't make it impossible, but it does make it unpleasant.
You misunderstand my question. I ask about hypothetical "blindfolded magic" which does not require physical cards any more than blindfolded chess requires a physical board and pieces. You can, in practice, use some representation of the gamestate as a mental help, but that is for your own benefit only; what you do with that representation is neither necessary nor sufficient for changing the gamestate, so there are no cards that would need to be physically verified.
The actual gamestate exists only as the result of a process, not because of some physical component. It would be manipulated and documented entirely through a set of non-repudiable, potentially encrypted messages according to an agreed-upon protocol. Deviation from that protocol would be provable to a 3rd party.
https://en.wikipedia.org/wiki/Non-repudiation
Oh, well then yes, obviously. See MTGO or MTGA for implementations.
Well, no, that's not what I mean. In MTGA/O, a central server keeps track of the gamestate and acts as referee at all times. I'm asking about a peer-to-peer protocol, not about a client-server protocol. In my previous comment I did say "would be provable to a 3rd party.", but not in the sense of a referee keeping track of the rules at all times.
You do have to ban some cards like Chaos Orb that require physical objects. Barring those (which are overrepresented in unsets), I don't see any obstacles to reimplementing a game engine as non-repudiable instead of MTGA's server-based with account logging.
How does this work with scrying? If you scry your library, the next cards become known to you but not your opponent, and you can no longer simply draw the next card by generating a random number. I think you'd have to "draw" scried cards into a separate zone and record their order, so that your opponent can verify you scried only those cards and they were drawn in the order specified.
Or even worse, what about fateseal? How do you select N cards from your opponent's deck and reorder them, without your opponent knowing which cards you selected?
Edit: Also, scrying doesn't simply bury cards, it specifically puts them on the bottom, which will be a problem to keep track of if the game goes all the way through your deck. I think you need a way to cryptographically record the entire order of the deck without cheating.
Yes, these are the complications I'm asking about, whether they can all be overcome.
For the Scry case, I would say that we determine a random card as above, let's say Card #15. I know what card #15 is because of my previous ID-assignment so I announce my choice of putting it back on top or to the bottom.
The next time the game wants to do something with my top card (e.g. when I draw a card, or it gets milled, and so on), we would agree to choose that card #15 instead of a random card because the agreed game state says that #15 is currently on top. If I have put #15 on bottom, then the next time I would draw a card I draw a random card *except* #15, because there are others still "on top" of #15.
I've played entirely digital Magic over the internet from time to time. A friend wrote a little app that tracks your deck state and supports draw, scry, Brainstorm, shuffle etc. That app was trust-based for playing with friends, but most of its functions could be easily adapted to share appropriate hashes after each library interaction. Scry and surveil are completely fine, as you say. But even that app would have trouble with Fateseal and Praetor's Grasp.
And yet having seen some of the proofs that are possible in the space of card algorithms, I feel like this might still be possible. (And if it's not already solved, there's probably a paper at the FUN With Algorithms conference in it for whoever solves it!)
I think that for fateseal & friends, a simple hash-based system is probably not sufficient, and you want a full public key crypto system instead.
In particular, Fateseal could be implemented using a commutative crypto system, e.g. one where you can encrypt a text twice with different keys and it does not matter in which order you do the decryptions. (There might be a more correct name than commutative for this property, though.)
Basically, each card in the library is encrypted twice: once by each player (who can also reorder the cyphertexts, to shuffle the cards).
If you draw a card, then your opponent decrypts the next-index twice-encrypted card. They do not learn anything from that. However, you can now decrypt the result into a card name (plus salt).
If instead you get targeted by Fateseal 1, then you have to decrypt the top card of your library for them. This will allow them to know what card it was (without you being the wiser).
One difficulty is that both parties would have to apply a per-encryption salt if they use public key systems, which will likely clash with the commutativity requirement.
Searching for "commutative encryption" leads to [1], which cites [2]. Seems the trick is to use SRA, which is RSA but with different stuff kept secret, to play "mental poker" -- which is probably mostly equivalent to MtG from the required crypto primitives.
So my idea is not entirely nonsense, just some 45 years late.
Edit: upon reading the SRA paper [2], it appears that the system depends on disclosing the key at the end of the game. This is fine for MtG but absolutely terrible for poker. The fact that you will wonder for the rest of your life if your opponent was bluffing is very much a key part of poker!
[1] https://xianmu.github.io/posts/2018-09-19-commutative-encryption.html#ref-shamir1981mental
[2] https://dspace.mit.edu/bitstream/handle/1721.1/148953/MIT-LCS-TM-125.pdf?sequence=1
Yes, I'm looking for a zero-trust protocol that can deal with the entire tournament-legal card pool except of course cards like Chaos Orb that interact with the physical space.
I'm going to try to turn off my usual misanthropy for a moment, because I think I might have a useful idea, and some of the people here probably know more than me on this subject. (Maybe even Scott himself--he's a psychiatrist?)
On the recent discussion on men in the Bay Area, it occurs to me there's probably a large amount of older therapy literature from the period when most of the people writing about it were men. Granted it's probably got all kinds of obsolete psychoanalytic ideas running through it, but there might be stuff about ego, self-actualization, etc. that might be useful in the current problems of trying to make effective therapy for men?
That's an interesting thought, but I think maybe slightly misguided.
Men actually are well-represented among founders of different therapy schools, both in the past and the present (I think therapy school founder requires a similar neurotype to entrepreneur or something). So I don't think the bottleneck is that there aren't therapy schools that understand the male mind or whatever. I think the bottlenecks are:
- the therapists implementing the schools, including the male-founded schools, are mostly women.
- many therapists aren't going off a school and are just kind of operating on vibes. There is widespread debate about whether this is better or worse, but in any case most of the therapists doing this are female, most of the vibes are female vibes, and insofar as these vibe-based practices then congeal into a coherent philosophy of therapy, it's a female philosophy.
- men are less interested in going to therapy in the first place, and might prefer some kind of figuring-out-your-life practice which is not quite therapy in the traditional sense.
Nice to get an actual informed opinion! Thank you!
But aren't there still older books that might be more useful to men if updated for modern conditions? I've seen tapes of Albert Ellis and he had that sarcastic midcentury New York Jewish edge I think a lot of (non-antisemitic) young guys might like.
A question I posed to my wife and my brother in law: is there anyone else in the history of the world who's name is more known and used than the Earl of Sandwich? Think about it: his name became basically just a part of the English language. His name is used all of the time, even to describe things that he legitimately had nothing to do with because they're from completely separate cultures or predated him. I've seen Indian flatbread sandwiches be called, well, sandwiches, and for Passover every year Jews everywhere read the Haggadah and do the ritual of eating the Hillel sandwich. Sandwich's name even became a verb for putting non-edible things between other things; I can be sandwiched between two people on the bus or a house can be sandwiched in the middle of two buildings. My brother in law proposed that Caesar may be a contender, which is a good proposal with Kaiser, Tsar, C-sections, and salads, though I still think maybe Sandwich has him beat.
Jesus Christ is a widely known figure, and you got Spanish-speaking countries mentioning him each time someone sneezes. Then if you also count any mention on interjections and such, I think it really adds up.
The Virgin Mary might be another contender, given the popularity of the Hail Mary prayer among practicing Catholic. I think you say her name over a hundred times while praying a rosary (twice per prayer for each of 50+ "Hail Mary" beads).
That doesn't count because you're referring to her by name, it's not a word with a separate definition etymologically derived from her name. Contrast, e.g., https://en.wiktionary.org/wiki/Jesus_Christ#Interjection
The original question was about whose name is most known and used, although we very quickly drifted away from that
Duke of York? Every time someone says new york, mentions the new york times etc. I think that's likely to be said very often, although I have no idea whether people working at subway say sandwich all the time.
It was his name in the sense that it was what people called him. (And the foodstuff is named after him specifically, not after the title.)
What about Amerigo Vespucci?
Sandwich was technically his title not his name, which was John Montagu.
Amerigo is a good one. He has the benefit of being used cross-languages.
Sandwich is also used cross-languages. The wikidata item, https://wikidata.org/wiki/Q28803 , shows that the name of the thing in most languages is either "sandwich" or an adaptation of it to the language's writing system. This includes chinese 三明治 : see https://en.wiktionary.org/wiki/%E4%B8%89%E6%98%8E%E6%B2%BB .
I think this has to be the winner. "America" is a word used by speakers of languages all over the planet including plenty of people who don't know a complete sentence's worth of English words. Plus it appears in multiple places on every world map, classroom globe, etc.
Quite a few words are named after someone or other. Especially Greek and Latin words.
Some that come to mind:
Tantalize - King Tantalus
Narcissism - Narcissus
Draconian - Draco (Athens)
Mentor - Mentor (Odyssey)
Boycott - Charles Boycott (Some Irish guy from what I remember)
Also a ton in science terms like Pasteurize, Diesel, Ampere/Volt and quite a lot of other stuff I’m sure.
Definitely not more common than "sandwich", but also: Mausolus -> mausoleum
I think only real historical figures are supposed to count.
Most people who say "sandwich" have no idea that it's named after anyone, but if that doesn't matter...
More well-known? I would guess that the most well-known names would be some prominent religious figures surpasses it: Adam, Abraham, and Moses are all widely known among adherents of the Abrahamic religions, which are the majority of the world's population.
More frequently used? I don't speak Mandarin, so I don't know if any common words in it are named after people, but if there are, it would probably beat sandwich, given that much of the world either doesn't often eat sandwiches or doesn't speak a language where the word descends from the Earl's title. Caesar's name probably gets more uses from the month of July than from any other source, I would guess.
If there's any actual human whose name was the source behind the various deities that have weekdays named after them (Thor's Day, Saturn's Day, etc), those would probably beat sandwich.
> for Passover every year Jews everywhere read the Haggadah and do the ritual of eating the Hillel sandwich.
Most seders aren't conducted in English, and therefore do not have any direct reference to "sandwich".
I'd guess there are more people who are familiar with sandwiches worldwide than there are Chinese people. In any case in Chinese it is also called a sinicized version of "sandwich": 三明治. Funny how that 三 (sān, meaning 3 when on its own) looks a lot like a sandwich.
We considered Mandarin, but there are more global speakers of English then Mandarin, it's more widespread as a second language.
My guess is that the most commonly used English word that's a specific person's name is "guy".
Also, re: your brother-in-law's suggestion, don't forget to include July.
Is the word guy named after Guy Fawkes or someone? I always assumed it was like a Tom, Dick, and Harry thing.
Yes. The chain of meanings was something like:
1. Male given name, cognate of Gaius or Guido
2. Effigy of Guido "Guy" Fawkes
3. Dummy or effigy in general
4. A shabby or disreputable fellow, particularly one wearing clothes resembling the worn-out cast offs traditionally used to construct a sense-2 or sense-3 guy.
5. The modern meaning, similar to 4 but having lost the negative aspects of its meaning.
The line in the song "I've Got a Little List" from the Gilbert & Sullivan opera "Mikado", where the Lord High Executioner mentions "The lady from the provinces who dresses like a guy", he means sense 2 or 3 with shading of sense 4. A modern American lyricist trying to convey a similar meaning might say something like "The woman from the country who dresses like a scarecrow".
Yes, it's named after Guy Fawkes: https://www.oed.com/dictionary/guy_n2?tab=meaning_and_use
And August? (which can also be an adjective)
Emperor Augustus was named after the adjective, not the other way round. Also, the adjective is probably less used than all the combined non-July things named after Julius Caesar.
Does anyone know a good writing forum? I've been trying to get back into fiction writing but am currently managing about four sentences every three days.
A bunch of my friends in the community use https://4thewords.com.
Worth a shot, thanks.
I remember the children's ditty about Mussolini. It was sung to the tune of "Whistle While You Work" (as sung by the seven dwarfs of Snow White fame), and it was ambiguous as to whose member (his own, or der Fuhrer's) Mussolini had bitten.
This was 20ish years after the end of WW2. I assumed it originated with the uncles and older brothers of my peers.
I recall it as well, though in the version I knew, he "pulled his weenie." I had always assumed it was his own weenie. My understanding is that the ditty was very popular during the war; I am pretty sure I learned it from my mother (along with several unrelated dirty limericks), but I might be mistaken.
For the uninitiated:
"Whistle while you work. Hitler is a jerk. Mussolini pulled his weenie; now it doesn't work!"
Have LLM’s hit a scaling wall?
If so, what next? A new architecture goes further? A financial crash?
If not, then what?
I would suggest ignoring both the hype and the anti-hype – journalists seem to like writing headlines like "gpt5 was a massive failure which proves LLMs have plateaued", but I think it's actually mostly in-line with trends (Peter Wildeford said it's roughly what he forecasted; I'm likewise; a poll in the ACX discord of gpt5-predictions before release showed most people getting what they expected or being at most mildly disappointed). For instance, the "time horizon" benchmarks are still scaling strongly. Personally I think Zvi Mowshowitz gives some good overviews of this stuff.
I've been wondering the same thing after noticing that the release of GPT-5 didn't make that big of a splash. Although it's still relatively early, so who knows what might pop up.
I think AI models with the ability to do contiuous learning could be an interesting next step. Quickly integrating new information seems essential for getting good use out of an AI model. That way, a model could improve on the job, so to speak.
I imagine something like telling an LLM the kind of mistake patterns it makes leading to the model avoiding those kinds of mistakes in the future.
But that seems like a big safety issue, because it would probably be very hard to predict what new behaviors such models might develop. What kinds of changes should a model welcome and what kind of changes should it resist? How could it be guaranteed that the model will uphold some core principles if it's always open to change?
Note that the amount of compute that went into training GPT-5 was a much smaller relative increase over GPT-4, compared to GPT-4 over GPT-3 or GPT-3 over GPT-2. So the correspondingly smaller capability gains are expected and don't refute the scaling hypothesis; OpenAI is just engaging in version number inflation. Training a model that's that kind of step up from GPT-4 requires constructing a bunch of new data centers from scratch, and there hasn't been time for that.
Then it seems that OpenAI is trying to keep the hype going while it manages to get those new data centers up and running.
I wonder what reasons can convince investors to be patient enough to not withdraw their money while waiting for the new infrastructure.
No, unfortunately, depending on what you mean. It doesn't seem like labs are banking on just 10xing the parameters on last thing that worked and throwing most of the compute at pre-training. The new hotness is synthetic data, which is expensive to produce but unlike internet data is potentially unbounded. The methods so far used to train the top models are quite crude, at least those methods described in publicly available documents, still they have been rather effective and capabilities continue to increase quickly and steadily despite model size kind of leveling out or sometimes even shrinking.
So what's likely going to be evolving, in the near term, is training architecture. Unlike the transformer where, while there are many ideas floating around it's not super clear what to try next other than fiddling with little details to squeeze out more efficiency, there are some blindingly obvious next steps on the training front. For example, engineers would I imagine very much like to have the ability to train a really big model with something like MCTS, it's been tried but the hardware needed to make that viable apparently isn't there yet and simpler but less compute heavy schemes have won out. Despite that, the trend of getting close to maxing every benchmark more than a couple years continues. And the hardware needed for the next few steps is coming online very soon. Seatbelts buckled, we ain't seen nothing yet
Why don’t you just vibecode the CBTi app yourself, Scott?
I don't know CBTi, coding, or entrepreneurship, and I'm not sure we're quite at the part of the glorious AI future where the best person to do something that requires three skills is someone who has zero of the skills.
Fair enough!
Is the technology at the point where somebody without any expertise in software engineering can make an app and expect it to work? That was not my impression.
I experimented with it a bit over the weekend. My impression was it works until it doesn't, at which point it became a complete nightmare to untangle everything and figure out what was going wrong. And the LLM itself was not much help at that!
Also, though, it didn't work well when I said "make an app to do xyz". It only really worked when I broke down the project myself into a series of logical steps, and then asked the LLM to work on each step at a time.
Web app kind of if you already have the kind of "grit" that you would need to code (try something, fail, try to understand why you fail without getting frustrated too much, try again), but you will get results way faster than learning to code from scratch. Mobile apps it's harder from what I understand, I see less of them.
Admittedly I was trolling a little bit. But only a little! I really am curious what Scott has to say about this.
I think it is possible to get some sort of working app if you spin the slot machine often enough, even if you don't know any code. But also I assume that Scott has access to the latest and greatest (or at least most expensive) models, so his experience might be totally different and much better. If not, I'd like to hear about that too.
The toothbrush moustache sported by Hitler was already popular in Germany and its origin country USA by the time he probably adopted it. There are photographs of Hitler during WW1 that show him wearing a Kaiser moustache instead. Wotan's moustache in the painting also looks much broader than a toothbrush, extending beyond the nose on the sides more like a painter's brush moustache.
https://en.wikipedia.org/wiki/Toothbrush_moustache
https://www.google.com/search?q=Painter's+brush+moustache
The German news magazine Stern's website recently had an article about Hitler's moustache.
https://www.stern.de/panorama/wissen/adolf-hitler--wie-er-zu-seinem--hitler-bart--kam---und-was-er-bedeutet-30826594.html
To summarize, the "trench moustache" was part of Hitler's and the NSDAP's propaganda that was built on his personal recognizability as a trench soldier from the war, everybody in Germany in the post-WW1 time would recognize it and would understand his message "I am one of you soldiers". For the same reason, he added a trenchcoat to his civil suit and abstained from wearing his lederhosen. Moreover, it was a more modern kind of moustache compared to the establishment style.
Not the only time someone thought that certain pictures must have been the inspiration for later work.
From a draft letter of Tolkien, 1971:
" A few years ago I was visited in Oxford by a man whose name I have forgotten (though I believe he was well-known). He had been much struck by the curious way in which many old pictures seemed to him to have been designed to illustrate The Lord of the Rings long before its time. He brought one or two reproductions. I think he wanted at first simply to discover whether my imagination had fed on pictures, as it clearly had been by certain kinds of literature and languages. When it became obvious that, unless I was a liar, I had never seen the pictures before and was not well acquainted with pictorial Art, he fell silent. I became aware that he was looking fixedly at me. Suddenly he said: 'Of course you don't suppose, do you, that you wrote all that book yourself?'
Pure Gandalf! I was too well acquainted with G. to expose myself rashly, or to ask what he meant. I think I said: 'No, I don't suppose so any longer.' I have never since been able to suppose so. An alarming conclusion for an old philologist to draw concerning his private amusement. But not one that should puff any one up who considers the imperfections of 'chosen instruments', and indeed what sometimes seems their lamentable unfitness for the purpose."
See also here for possible (postcard) inspiration:
https://tolkiengateway.net/wiki/Gandalf#Inspiration
"Tolkien had a postcard labelled Der Berggeist ("the mountain spirit"), and on the paper cover in which he kept it, he wrote "the origin of Gandalf" at some point. The postcard reproduces a painting of a bearded figure, sitting on a rock under a pine tree in a mountainous setting. He wears a wide-brimmed round hat and a long cloak and a white fawn is nuzzling his upturned hands. Humphrey Carpenter in his 1977 biography said that Tolkien had bought the postcard during his 1911 holiday in Switzerland. However, Manfred Zimmerman discovered that the painting was by German artist Josef Madlener and dates to the late 1920s. Carpenter concluded that Tolkien was probably mistaken about the origin of the postcard himself. Tolkien must have acquired the card at some time in the early 1930s, at a time when The Hobbit had already begun to take shape."
As soon as I feel like I've heard all of the details about Tolkien, his thoughts, what he was going for, all of his letters, here's one I had not known about.
The letters are so good, the only problem is that they are selected and edited, so sometimes Humphrey Carpenter leaves out interesting bits and obviously there are a ton more letters I, for one, would like a peek at 😀
I thought he cut it to fit under a gas mask and just decided to keep it that way.
WW1 was when beards and mustaches really fell out of style, and it’s largely because they needed to be trimmed or cut completely to fit under the gas masks at the time.
That is the story, yes - other stories say that he adopted the toothbrush to leech off of Charlie Chaplin's popularity. But the reason doesn't really matter; the point is that there is photographic proof that he didn't have the toothbrush before well into adulthood, and that he didn't adopt it "as a youth" as the highlighted comment relays.
https://www.astralcodexten.com/p/your-review-ollantay/comment/148004547
That then raises the question, why did Chaplin use the toothbrush moustache? If it was a style at the time, no need to wonder where Hitler picked up the idea of grooming that way.
According to Wikipedia, it originated in America and was introduced to Germany by visiting Americans:
https://en.wikipedia.org/wiki/Toothbrush_moustache
"The toothbrush moustache was introduced to Germany in the late 19th century by visiting Americans. Previously, the most popular style was the imperial moustache, also known as the "Kaiser moustache", which was perfumed and turned up at the ends, as worn by German emperor Wilhelm II. By 1907, enough Germans were wearing the toothbrush moustache to elicit notice by The New York Times under the headline "'TOOTHBRUSH' MUSTACHE; German Women Resent Its Usurpation of the [Kaiser moustache]". The toothbrush was taken up by German automobile racer and folk hero Hans Koeppen in the famous 1908 New York to Paris Race, cementing its popularity among young gentry. Koeppen was described as "Six-feet in height, slim, and athletic, with a toothbrush mustache characteristic of his class, he looks the ideal type of the young Prussian guardsman." By the end of World War I, even some of the German royals were sporting the toothbrush; Crown Prince Wilhelm can be seen with a toothbrush moustache in a 1918 photograph that shows him about to be sent into exile."
So if it was a popular style worn by role models, I am not surprised Hitler adopted it when he's becoming a big cheese in the Party in Munich; this would show he's not some hick from the provinces.
Why would Hitler have wanted to channel a comedy actor anyway?
Have you actually seen a Chaplin film? They're brilliant. I'd compare them to a Coen Brothers' film. Sure there's humor but there's also deep artistic wrestling with the existential realities of life. They also have a breathtaking visual style. They weren't just slapstick.
Allegedly he loved Chaplin's movies:
https://en.wikipedia.org/wiki/Toothbrush_moustache
"According to Hitler's bodyguard Rochus Misch, Hitler "loved" Chaplin films, a number of which he watched at his teahouse near the Berghof (built c. 1936)."
I recently got a vasectomy as a push present to my wife after she birthed our third child. Being a good ACXer and man anxious about having my junk operated on, I researched the procedure obessively. The medical information was mostly unsurprising. The demographic information was surprising.
https://pmc.ncbi.nlm.nih.gov/articles/PMC2784091/
From the abstract: "11.4% of men aged 30–45 years reported having a vasectomy, representing approximately 3.6 million American men. While 14.1% of white men had a vasectomy, only 3.7% of black and 4.5% of Hispanic men reported vasectomy. On multivariate analysis, a significant difference in the odds of vasectomy by race/ethnicity remained, with black (OR 0.20, 0.09–0.45) and Hispanic men (OR 0.41, 0.18–0.95) having a significantly lower rate of vasectomy independent of demographic, partner, and socioeconomic factors. Having ever been married, fathering two or more children, older age, and higher income were all associated with vasectomy."
I had assumed getting a vasectomy was like a standard rite of passage for American men entering middle age. My Dad had one, and many of my friends' dads got snipped. Same for my wife's family. It turns out that vasectomies in the US are not *that* common and those who get them are overwhelmingly middle class white males, and we were raised in a white middle class bubble. Mind blown.
This sorta makes sense to me. A vasectomy is kinda like a prenup: they're only useful if you have something to lose.
How was the procedure itself? I am thinking about doing the same after our next child, but am not looking forward to it.
The procedure itself wasn't bad. Local anesthesia and I didn't feel the incision or anything more than pressure/tugging. Prior to the incision, the doc discovered one of my vas was out of place, so he had to manipulate it closer to the skin with his fingers. That hurt quite a bit, even with the anesthesia.
The initial recovery was quick (stopped ice and painkiller about 36 hours in) but I wasn't able to resume running and lifting without discomfort until about 2 months after.
Oh wow, that's quite a long break to take from sports. Thanks for the reply!
Why do a vasectomy when you could just use an IUD?
It was a present to my wife. Requiring her to undergo a (by most reports) painful procedure after 3 C-sections seemed unsporting.
I'm staying on my IUD forever because it stops me from having a period. It's a gift that keeps on giving!
Before my vasectomy almost 30 years ago, the doctor showed me a short educational film, which included an apprehensive Black man asking whether, after the procedure, he "would still be a man".
Consistent with the data, I am a white middle-class male.
Honestly? Feminists want me to do it, and I always used condoms when I was still dating anyway.
Why don't LLM apps implement normal messaging apps functions such as responses?
When an LLM answer touches on multiple subjects, I don't want to have to say "about this subject could you tell me... regarding that one I disagree..."
The LLM should break up its response into multiple messages, and you should be able to answer to each message before launching the inference process
Not sure if this is quite what you had in mind, but there are a few tools out there that allow you to "multiverse" with LLMs -- i.e., branch off a discussion from a certain point. Loom (https://github.com/socketteer/loom) is one, but it's specialized in fiction. Raycast also lets you do what they call "chat branching" (https://www.raycast.com/changelog/1-101-0). Afaik neither of these have the ability to merge the branched chats back into a single thread.
It's not quite what I meant
I was thinking of a single chat with a shared context, it's just that the answers could be split into chunks so that you can answer to specific parts
Hasn't ChatGPT had this feature for years?
Oh, have they? I guess I haven't noticed it, maybe because I'm mostly a Claude guy.
But Claude also has always had it? Meaning - you can take any of your replies and edit it, turning it into a splitting point. Then you can switch between the branches. It's very inconvenient.
Huh, I thought I remembered not being able to do this in Claude but it does seem to be available.
What's inconvenient about it? I find it quite useful. Or are you just referring to the inability to merge branches?
If you make just one branching point, it's alright. But when you make multiple branches, multiple splits it becomes hard to keep track of what is where, and if you exit the chat the system doesn't remember where you were and next time you open it you may end up on some other node and have to look for the right place. Basically, it's not a system that was meant for heavy use, but it's convenient enough if you don't push it too far.
The other day I had the sudden realization that in most accents (and, more importantly, in General American, the accent I try to consciously emulate when I speak English) the pronoun "they" is pronounced [ðeɪ].
I have gone all my life pronouncing it [ðεɪ] like some sort of barbarian! I guess it's one of the effects of having started to learn English very young: you learn some things approximately when you are six, and they lock in forever. I still sometimes catch myself pronouncing "is" as [iz] rather than [ɪz]. How utterly embarrassing.
I don't really have a point with this, but it's a bittersweet realization: for however long and deeply you studied a non-native language, it will never be truly yours. You really are stranded where you were born, so to speak.
Ok since the phonetics aren’t obvious here to many I asked ChatGP, which writes.
[ðeɪ] — is the standard pronunciation of they in English. The vowel [eɪ] is that diphthong we use in words like day, say, play.
Meanwhile
But if you’ve been saying [ðεɪ] with the open-mid front vowel [ε] (like the vowel in bed), you’ve basically been saying something closer to “theh” sliding into a weak y. Not far off, but it would sound a bit unusual to most English speakers — like a slight accent twist.
And when I asked was this kinda posh, the model said not anymore but it’s close to the clipped RP back in the day. The Queen in 1950 but not 2000.
Since British English is the gold standard (and frankly no need for the British qualifier) then I think the op should be proud
That's interesting to know, but my concern was that the sound I am making is, as you mentioned, unusual. The part that messes me up is that I'm pretty sure I make that specific vowel combination ONLY when pronouncing "they". I pronounce, for instance "day" as [deɪ] i.e. the standard way. I have this idiosyncratic pronounciation exclusively for the word "they", in all of my English vocabulary.
That's not what a native speaker of any language would do, regardless of their accent: they would just pronounce "day, grey, may, they" as rhyming, regardless of the actual sounds.
But in my idiolect, "they" doesn't rhyme with anything!
This is also in reply to @Wasserschweinchen
EDIT: spelling
> day, grey, may, they" as rhyming, regardless of the actual sounds.
That’s not obvious at all. After all there are differences between a and e in many words. (And probably isn’t how the Queen used to say “They went that day”)
I sympathise; minor inaccuracies in pronunciation can be difficult to notice or correct once they become entrenched. I'd been speaking French for years before I realised there was supposed to be a difference between répondre [ɔ̃] and répandre [ɑ̃].
Could be worse. I took a long time to realise there was supposed to be a difference between "Je suis en route" et "Je suis en rut".....
Hahahahaha
According to A Course in Phonetics by Ladefoged & Johnson, [εɪ] is a normal realization of /eɪ/ in GenAm.
I've heard this exact same worry before, from a French ex-girlfriend who could hardly speak English when we first met and felt terrible about it (despite my still worse French) and we discussed it quite alot. So I actually feel somewhat equipped to disagree with you!
1) I think it's pretty extreme to realise you're mispronouncing a word and conclude that the language isn't yours: many - perhaps most - native speakers mispronounce many words for their whole lives. For example, young British people lamentably mispronounce (and misspell) words the American way because apparently they learn their English from sodding Netflix and Twitter now instead of BBC Radio Four and Jane Austen like they bloody well ought to - and yet despite their horrible misshapen American-English-with-British-Accents the language belongs to them far more so than it does to me! For another example, there are English words I'd only read, never heard, and only upon hearing them decades later did I realise that I'd been mispronouncing the words my whole life! (There have been plenty, but "cloaca" springs to mind...)
2) The fact that you're sufficiently self-aware to be even _capable_ of noticing such things (and sufficiently well-educated to be capable of using the IPA..) puts you _miles_ ahead of the typical native English speaker. You may not have Tennyson's accent but you're nevertheless "doing English" the way he did! I think self-awareness and education (including self-education) reveal to you the beauty, uniqueness, poetry, and expressive power of a language far, far more than does having it programmed into you when you're a toddler and then literally never voluntarily giving the language another thought for the rest of your life and just using whatever misspelled, unpunctuated, minimum-effort English you absorb by accident from social media.
3) I think "belonging to" a language (or at least, the good part of belonging to a language...) is about how you think about what you say, and choose your words for their poetry or semantic content (as opposed to the quasi-meaningless slop that people only seem to say because they heard it on Netflix... "energy", "based", "gave", "called out", "hits"....) much more so than it is about what pronunciation you use.
4) There are a great many accents and dialects in English, sometimes all-but unintelligible to one another, and it's only really a total accident of geography that the Glaswegian accent "counts" and (say) the Sicilian accent doesn't; there's nothing fundamentally worse about the latter (some might even consider it more beautiful or expressive than the former!) If anything, I think that having both - provided they can at least understand one another, so the accent isn't an insurmountable barrier to communication - makes the language richer and more interesting.
I get what you are saying, and admittedly it's not particularly bad, but it's the difference with my native language that gets me. I mispronounce really common words, while native speakers usually tend to mispronounce less common words; the less common the more mispronounced.
But I was mispronouncing "they"! It's probably in the top 30 most used English words, and I must have heard it spoken literally tens of thousands of times!
I get what you are saying, about foreign vs. native accents, and I agree in principle.
Still, when I can't keep my accent from slipping out, it feels like a failure of self-control.
I guess I would say that (again, you're not the first person I've had this exact same discussion with, sorry if this is somewhat weird coming from a total stranger!) if "not letting your accent from slipping out" is a fulfilling goal that you derive satisfaction and wellbeing from achieving, then sure: keep striving towards it and don't give up until you achieve it.
...but! A) Definitely think very hard about whether this is genuinely a satisfying and fulfilling goal to aim for, because the effort seems to be expensive to you and failures seem to affect you and maybe it's possible that you would derive more fulfillment from learning to embrace the uniquely Emanuellean "they" and focusing your efforts on some other problem, and B) even if, upon consideration, you do consider this a worthwhile and fulfilling goal that you genuinely do want to strive for, that's perfectly fine - but please don't conflate it with "better English" or "English belonging more to one", because that genuinely is more about mental properties than about perfectly mastering very very specific pharyngeal muscle movements...
(As for "they": I know many native English speakers who mispronounce "February", or use "criteria" in the singular, literally every day, and aren't remotely curious or contemplative enough to ever notice they're doing it. I believe that you are a better English speaker than these people, no matter how you pronounce "they"!)
There was an enjoyable discussion here (ACX) a few months ago, about mispronouncing words one had read but not heard.
This phenomenon was given the name "Calliope Syndrome", with Calliope rhyming with "hope" rather than "ropy".
It's associated with children who are avid readers.
Just today I realized that in a lot of accents, honor/honour is not pronounced the same as "on a", and so they don't find the name Honor Blackman to be all that funny.
I’m not sure what the phonetics are here but my wife uses Dey and dis for This, and she’s vaguely acceptable.
>and she's vaguely acceptable
Pronunciation-wise, or overall?
Both yes. Are you serious about this by the way? Do you really expect that native speakers ( of many varied accents ourselves) care about minor mispronunciations?
I wasn't serious at all, no. Sorry if I offended you! I just found "...and she's vaguely acceptable" very droll, is all - if it was meant purely factually I misread it, and I'm terribly sorry!
Oh I was being droll.
The OP is talking about the vowel. They're pronouncing the consonant as a voiced th sound like in every standard variety of English.
See Michael Caine in De Prestige
That would be more like [dεɪ], with th-stopping, wouldn't it?
Anyway, thinking about it that way makes it sound more endearing
Were you meaning more like a z sound?
The OP is only talking about the vowel, not the consonant. They're pronouncing the consonant as a standard voiced th sound.
True, I guess I meant th=d generally
I rarely comment, but the sheep program is hard to use and not worth your money.
Hi all,
I am Luomei Lyu, the founder and developer of Sheep-Sleep. I refunded Kevin his $15 (not a few hundred bucks as he claimed) the second I woke up and saw the message. Poor Wi-Fi can cause latency in AI responses, and we are actively improving this every day.
The real value and focus of Sheep-Sleep is in the content. Every single AI response — down to the exact wording — comes from thousands of hours of discussion and two years of work with some of the most experienced sleep psychologists in the US. Sheep has already helped many people sleep better: https://www.gnsheep.com/case-studies
As for the internet issues, that’s on me!
—Luomei
P.S., Here is how it was made: We had the best psychologists in the country talk us through every conversation they would have for each insomnia case, even down to the exact word choices, analogies, and conversation pace when teaching these well-validated techniques to their patients. We then wrote hundreds pages of instructions for generative AI to follow in its weekly sessions, so that it only responds with human-crafted material, while keeping the conversation collaborative, interactive, and highly personalized. This kind of dynamic, personalized dialogue is what makes in-person sessions effective at improving adherence, which remains the biggest challenge in this gold-standard, first-line treatment.
Tell me more?
Sure. First of all, I ran the iphone app on my desktop and didn't have the best Wi-Fi, so that may have caused some problems below. Anyway, my main issues are:
- It's a SaaS, but there's no login (you literally get a code after paying), and there's no way to cancel your subscription except to email them and wait (I'm still waiting btw), which is just dumb. I thought 298/mo would at least get you an account!
- The conversation with the AI chatbot was unsatisfying for a bunch of reasons, eg, it takes it like 20s of thinking to ask me a preprogrammed question like "Are you having trouble falling asleep or staying asleep?" Couldn't they just have a button that I can press? The first few questions are all like this, with it thinking for 10+ seconds, then asking a very simple question. Also, in its second answer, it repeated the answer twice (idk why), and in its 3rd or 4th answer, it didn't recognize my speech and asked me to speak again. Finally, this might be my aesthetic preference, but talking to an AI-generated cartoon of a sleep just feels...awful. Can't they put a bit more effort and actually have someone draw the character?
- I couldn't find their email/contact on the website, the ads/promotion vids are, well like most fancy new medical products, distasteful and exaggerated––but these are just small complaints.
- Oh and if the founder sees this, I'm still waiting for you to cancel my subscription!
Perhaps the slow responses were strategically used to make you sleepy?
Ah, the old MetaMed problem.
https://thezvi.wordpress.com/2015/06/30/the-thing-and-the-symbolic-representation-of-the-thing/
It doesn’t look aesthetically how one would expect an expensive health service to look, so it will be judged on that, and not on the ability to improve health outcomes.
Of course, for all I know the app could be garbage at actually improving sleep too, but I notice that that doesn’t come up in your list of complaints.
It's a heuristic. Having such a badly designed app is correlated with insufficient funding, a lack of professionalism, not really caring much about people using the service, etc.
Okay, but that's not a useful user review, because users can see for themselves that it's unpolished. The point of a user review is so that users can find out whether it actually works, from someone who has direct knowledge of this!
That's nonsensical, what he's talking about directly ties into the usability of the application.
Hey, thank you for this!
Improving anyone's sleep def takes more than 10 minutes.
Actually, that's *exactly* what I would want in a user review for something like this. I can't see for myself that it's crude and unpolished and frustrating to talk to without paying £300-odd a month (and a very difficult to cancel £300-odd, too); I'd much rather find these things out before opening my wallet....
If it's so dissatisfying to use that somebody gets frustrated, tries to cancel, and can't even cancel, then as a user I've found out from this review that it doesn't work
Yea, it's not. If someone can pay a few hundred bucks to try this and come back and write a review, that'd be more useful/reliable. (I'm just making a guess about the nature of this thing based on what I saw, if you still want to try the app after reading my rant, by all means go ahead.)
What is the sheep program?
SheepSleep (mentioned in the post above). It looks like they wrote the app in a high school hackathon. I bought a trial and canceled it. Charging 298/mo is diabolical.
Ah yes, the pitfalls of coming straight to the comments.
It shouldn’t be too hard to vibe code a… competitor, especially if someone does the public service of signing up for a subscription and documenting all the app screens and features with screenshots…
Idk, but I think maybe ACX should do some basic background check before featuring a service that's this scammy...
What about an open source CBT app?
Well, somebody has to build it, and it doesn't seem like anyone's interested.
1. Build open source project.
2. ????
3. Receive no money and lots of abuse.
FOSS
Sadly, it's too late to apply for an ACX grant to do a randomized trial on whether the $298/month insomnia app helps people sleep or not.
I'm not sure how you blind the trial. Give half the people a placebo app which costs $298/month but is somehow guaranteed to not contain helpful advice?
I’m pretty sure this is the placebo app.
My favorite part of their website is the video titled:
"how to get off 14 years of sleep meds in 5 days"
Like, great, now you're paying too much for an app and you're also going through some horrific drug withdrawal.
On what basis have you made this judgement?
I think the problem with blinding here is that $298/month is already most of the treatment effect. If you told people they were in a sleep study and charged them $5, they would probably sleep worse out of spite.
As for AI being a demon: demons are usually unreliable tricksters. LLMs are more like over-enthusiastic interns who take your vague instructions much too literally.
The more competent demons pretend to be over enthusiastic interns.
For $300 a month, they could send someone round to your place to knock you over the head with a mallet!
The second paragraph seems like it's replying to the wrong comment?
Yeah messed up was half asleep
Regarding AI, even the founders know it's diabolical:
Elon Musk: Artificial Intelligence is our biggest existential threat. ... AI is summoning the demon. Holy water will not save you.
DWave Founder Gordie Rose (A Tip of the AI Spear): When you do this, beware. Because you think - just like the guys in the stories - that when you do this: you're going to put that little guy in a pentagram and you're going to wave your holy water at it, and by God it's going to do everything you say and not one thing more. But it never works out that way. ... The word demon doesn't capture the essence of what is happening here: ... Through AI, Lovecraftian Old Ones are being summoned in the background and no one is paying attention and if we’re not careful it’s going to wipe us all out.
Musk and Rose saying this: https://old.bitchute.com/video/CHblsEoL6xxE [4:29mins]
There’s this and then there’s the reports of ChatGPT and other top level models slowing down in their progression. What’s a boy to believe.
The question of how hard it is to control a powerful AI is separate from the question of how soon powerful AI is coming, though of course there's some common interest in figuring out countermeasures.
"Holy water will not save you."
I dunno, I think pouring a pint of it straight into the innards of the server might do *something* 😁
Ha ha, and you thought I was crazy for having my rosary beads, votive candle of St Martha, and picture of St Therese of Lisieux on my desk around my PC!
Despite actually knowing Lisieux fairly well, I couldn't help misreading that as "St Thérèse of Linux"..
Hey, we may yet get a saint for that!
I think rye whiskey would be better than holy water, more conductive you know. Plus it has a moderate stat boost for courage.
Very spiritual, or at least spirits? 😁
Why the fuck are they building it then
Money. Power. The usual.
They think they can make money and a 20-30% chance of killing everyone isn't enough to keep them from taking a chance on making some more money; entrepreneurs have faced worse odds, I think.
Also perhaps running the AI can improve their chances of being able to control it--to continue the Lovecraftian metaphor, if everyone's summoning Old Ones, maybe if you're the guy who summons Cthulhu you can send him after the rest of the world first.
I love this community but you guys have way too much reverence for successful business people.
Who exactly is "you guys" here? AFAICT people have mostly been unimpressed with, e.g., Musk ignoring his own stated concerns about the control problem.
That's fair. Maybe I spent too much time on Sneer Club.
To be clear I'm a lot closer to ACX than the sneerers, but I do think people here have way too much faith in capitalism.
The cultists imagine that by serving the Old Ones they will be top dogs in the new post-sweeping Earth, that they will have all kinds of powers and pleasures:
“The time would be easy to know, for then mankind would have become as the Great Old Ones; free and wild and beyond good and evil, with laws and morals thrown aside and all men shouting and killing and revelling in joy. Then the liberated Old Ones would teach them new ways to shout and kill and revel and enjoy themselves, and all the earth would flame with a holocaust of ecstasy and freedom.”
The reality, of course, is that all humans are like vermin to the Old Ones and they have no attachment to the cultists or servitors, and they will be devoured and consumed like the rest of us.
I'm familiar with the Cthulhu Mythos lore, but I doubt Musk really is planning to be anyone's servant.
Good, because the Great Old Ones will chomp him up along with the rest of us. It doesn't matter what the humans plan or intend, they're just in the way and must be cleared off with the rest of the rubbish.
I sometimes wonder if that's actually Elon's plan, to be eaten first, just like that joke about the Cthulhu cultists: https://www.entrelineas.org/pdf/assets/who-will-be-eaten-first-howard-hallis-2004.pdf. If you stand before the maw when the Hellgate opens, because *you're* the guy who opened it, indeed fought your way to the front of the queue to be the guy who opens it, then...
I kind of feel like this brand of rhetoric isn't helping. It's not that it's wrong; it's that there is a *technical disagreement* about what factors make it hard to steer a powerful AI and how hard they are to overcome. Alignment optimists aren't going to find this kind of comparison compelling, they're just going to conclude that we're mindkilled. Better to focus on the specific technical reasons for pessimism, rather than the vibes.
The same reason they built the LHC, they are creating a portal at CERN which literally sits above where Apollo was thought to reside in the underworld, “to another dimension (CERN own words) which something might come through” (direct quote).
Meanwhile, whatever comes through can use the quantum computers (again messing with the fabric of reality) as a host to interact with the world.
You can read more about this here:
https://archive.is/k9xon
and watching this highly informative video:
AMONG THE MOST FASCINATING PRESENTATIONS ON BOOK OF ENOCH, FALLEN ANGELS, NEPHILIM, GIANTS, SPIRITS
https://old.bitchute.com/video/CVLBF3QP6PlE
You are not dealing with sane people. They will happily kills us all to gain favor with the unseen realm
That video link leads to a 404 page.
Apologies, thank you for the heads up, fixed now
As far as I understand, the LHC is not doing anything that doesn't already happen naturally in the atmosphere. It's just recording the results better.
And quantum computers are quite weird, phyisically speaking, but there's zero reason to believe it's "messing with the fabric of reality". Especially in a universe where things like neutron stars and super novas exist.
AI is not a threat to the universe any more than humans are, but (1) it's still a threat, and (2) at a much shorter timescale. Comparing AI capabilities researchers to LHC and QC engineers is not fair.
Alas, the ones who are building it are telling you directly that they can do everything I mentioned (if you read the article and watch the video the evidence for the LHC and Quantum computers is right there), but you know better than they do… so nevermind, i guess. Move along.
I opened the page. I saw that it said “ AI quantum computers act as hosts for disembodied Nephilim spirits who are stuck in this realm unable to escape:”.
I closed the page.
haha, that made me laugh. you should scroll to the section where gordi is giving a presentation and talks about how standing next to his d-wave machines - it feels like they are alive and sentient.
From the horses mouth. Cya
They just can't resist the challenge of building the Torment Nexus from the classic sci-fi novel Don't Create The Torment Nexus. Also, if if it's not them who destroys the world, the Chinese will, and do you really want us to fall behind in a race with the Chinese?
Mr President, we cannot allow a Torment Nexus gap!
Well, charitably, each of them believe that an ASI created by themselves has a higher (and potentially non-negative) expected value than the ones currently being developed by others and so they are essentially forced to run the race in to save humanity from the more monstrous creation of the next guy over.
Moloch strikes again.
There is also a significant amount of legitimate disagreement about how hard it is to make a powerful AI do what you want.
(This doesn't excuse Musk, who I don't think has a principled commitment to an anti-China foreign policy either; I think he's just crazy. Rose isn't working on AGI as far as I know; his former company put out some breathless AGI-sounding marketing material but in reality they were working on narrower robotics stuff.)
That is a good point. AFAIK some like Amodei are not so impressed by the ‘superintelligence explosion’ argument and generally believe we’ll be able to align powerful AIs by researching interpretability and improving model design, prompting, finetuning, etc.
There is also potentially the economist-brained viewpoint that alignment won’t be a problem because such systems would generally act rational economic actors and prefer. Or that they’ll be trained on the human corpus which largely encodes ‘good’ values and so will be good (enough) by default.
I was focused on making a charitable case for Musk, Altman, etc, who do, I think, believe in that case *at best*.
"such systems would generally act rational economic actors"
So what happens when spherical cow world meets messy human reality?
Do you happen to know what Altman has had to say about the control problem? IIRC when OpenAI was founded his stated concern about AI was the possibility that economically useless humans might be liquidated, and his proposed solution was a UBI indexed to GDP. He still gestures at this general "governance" class of concern and says that OpenAI is taking it seriously, but never provides specifics or commits to anything.
Are they AI responses, or did you spend two years crafting them?
If the latter, why does it repeat itself and take 20s to think of its next answer?
How did you go from "I should charge to make people respect this" to "I should charge $596 for treatment"?