Astral Codex Ten

GPT-5

42 min read

GPT-5

Article

GPT-5 is a recurring concept in the Astral Codex Ten archive, appearing 7 times across 7 issues between April 25, 2023 and November 20, 2025. The archive places it in contexts such as “OpenAI wasn’t currently training GPT-5 and “won’t for some time””; “Details about GPT-5 are still secret”; “GPT-5 might need about 1% the world’s computers”. It most often appears alongside OpenAI, Google, Anthropic.

Metadata

  • Category: Concepts
  • Mention count: 7
  • Issue count: 7
  • First seen: April 25, 2023
  • Last seen: November 20, 2025

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

23
April 25, 2023 · Original source
The drop a few days ago was when Sam Altman said OpenAI wasn’t currently training GPT-5 and “won’t for some time”. Apparently forecasters don’t expect them to take too long a break.
Sam Altman Wants $7 Trillion
February 13, 2024 · Original source
The basic logic: GPT-1 cost approximately nothing to train. GPT-2 cost $40,000. GPT-3 cost $4 million. GPT-4 cost $100 million. Details about GPT-5 are still secret, but one extremely unreliable estimate says $2.5 billion, and this seems the right order of magnitude given the $8 billion that Microsoft gave OpenAI.
Inline links: GPT-2, GPT-3, GPT-4, one extremely unreliable estimate
(Unless they slap the name “GPT-6” on a model that isn’t a full generation ahead of GPT-5. Consider these numbers to represent models that are eg as far ahead of GPT-4 as GPT-4 was to GPT-3, regardless of how they brand them.)
If we keep our 30x scaling factor, GPT-5 will take 1/70th of all the computers in the world, GPT-6 will take 1/2, and GPT-7 will take 15x as many computers as exist. The computing capacity of the world grows quickly - this source says it doubles every 1.5 years, which means it grows by an order of magnitude every five years, which means these numbers are probably overestimates. If we imagine five years between GPTs, then GPT-6 will actually only need 1/10th of the world’s computers, and GPT-7 will only need 1/3. Still, 1/3 of the world’s computers is a lot.
Inline links: this source
Zvi on California's AI Bill
May 08, 2024 · Original source
The reason it sounded like a bad bill before was that people were misrepresenting what it said. The bill applies to “frontier models” trained on > 10^26 FLOPs - in other words, models a bit bigger than any that currently exist. GPT-4 doesn’t qualify, but GPT-5 probably will. It also covers any model equivalent to these, ie anything that uses clever new technology to be as intelligent as a current 10^26 FLOPs model without actually using that much compute. It places three1 types of regulation on these models: First, companies have to train and run them in a secure environment where “advanced persistent threats” (eg China) can’t easily hack in and steal them2. Second, as long as the model is on company computers, the company has to be able to shut it down quickly if something goes wrong. Third, companies need to test to see if the model can be used to do something really bad. Its three categories of really bad things are: Create nukes or other weapons of mass destruction. This can’t be something dumb like linking the user to the Wikipedia page for uranium. It has to help human terrorists “in a way that would be significantly more difficult . . . without access to a covered model”.
Inline links: 1, 2
Go rogue and commit some other crime that does > $500 million in damage3. If the tests show that the model can do these bad things, the company has to demonstrate that it won’t, presumably by safety-training the AI and showing that the training worked. The kind of training AIs already have - the kind that prevents them from saying naughty words or whatever - would count here, as long as “the safeguards . . . will be sufficient to prevent critical harms.” So the bill isn’t about regulating deepfakes or misinformation or generative art. It’s just about nukes and hacking the power grid. There are some good objections and some dumb objections to this bill. Let’s start with the dumb ones: Some people think this would literally ban open source AI. After all, doesn’t it say that companies have to be able to shut down their models? And isn’t that impossible if they’re open-source? No. The bill specifically says4 this only applies to the copies of the AI still in the company’s possession5. The company is still allowed to open-source it, and they don’t have to worry about shutting down other people’s copies. Other people think this would make it prohibitively expensive for individuals and small startups to tinker with open-source AIs. But the bill says that only companies training giant foundation models have to worry about any of this. So if Facebook trains a new LLaMA bigger than GPT-5, they’ll have to spend some trivial-in-comparison-to-training-costs amount to test it in-house and make sure it can’t make nukes before they release it. But after they do that, third-party developers can do whatever they want to it - re-training, fine-tuning, whatever - without doing any further tests. Other people think all the testing and regulation would make AIs prohibitively expensive to train, full stop. That’s not true either. All the big companies except Meta already do testing like this - here’s Anthropic’s, Google’s, and OpenAI’s - that already approximate the regulations. Training a new GPT-5 level AI is so expensive - hundreds of millions of dollars - that the safety testing probably adds less than 1% to the cost. No company rich enough to train a GPT-5 level AI is going to be turned off by the cost of asking it “hey can you create super-Ebola?”, and putting the answer into a nice legal-looking PDF. This isn’t the “create a moat for OpenAI” bill that everyone’s scared of6. Other people are freaking out over the “certification under penalty of perjury”. In some cases, developers have to certify under penalty of perjury that they’re complying with the bill. Isn’t this crazy? Doesn’t it mean if you make a mistake about your AI, you could go to jail? This is deeply misunderstanding how law works. Perjury means you can’t deliberately lie, something which is hard to prove and so rarely prosecuted. More to the point, half of the stuff I do in an average day as a medical doctor is certified under penalty of perjury - filling out medical leave forms is the first one to come to mind. This doesn’t mean I go to jail if my diagnosis is wrong. It’s just the government’s way of saying “it’s on the honor system”. What are some of the reasonable objections to this bill? Some people think the requirement to prove the AI safe is impossible or nearly so. This is Jessica Taylor’s main point here, which is certainly correct for a literal meaning of “prove”. Zvi points out that it just says “reasonable assurance”, which is a legal term for “you jumped through the right number of hoops”. In this case probably the right number of hoops is doing the same kind of testing that OpenAI/Anthropic/Google are currently doing, or that AI safety testing organization METR recommends. The bill gestures at the National Institute of Standards and Technology a few times here, and NIST just named one of METR’s founders as their AI safety czar, so I would be surprised if things didn’t end going this direction. METR’s tests are possible and many AI models have successfully passed earlier versions. Other people worry there are weird edge cases around derivative models. I think the bill’s intention is that once you prove that your AI is too dumb to create nukes, you’re fine to open-source it. Third-parties can change its character, but not its fundamental intelligence. But in theory, a third party could get tens of millions of dollars of compute and keep training your AI to increase its fundamental intelligence. This would be a weird thing to do, and anyone with that much compute probably should just make their own model. But if someone wanted to screw you over by doing this, technically the law is kind of vague and you would have to trust a judge to say “no, that’s stupid”. Probably the law should clarify that it doesn’t apply to this situation. Other people are worried about a weird rule that you can’t train an AI if you think it’s going to be unsafe. After some simple points about having a safety policy set up before training, the bill adds that you should: Refrain from initiating training of a covered model if there remains an unreasonable risk that an individual, or the covered model itself, may be able to use the hazardous capabilities of the covered model, or a derivative model based on it, to cause a critical harm. This makes less sense than all the other rules - you can test a model post-training to see if it’s harmful, but this seems to suggest you should know something before it’s trained. Is this a fully general “if something bad happens, we can get angry at you”? I agree this part should be clarified. Other people think the benchmarking clause is too vague. The law applies to models trained with > 10^26 FLOPs, or any model that uses advanced technology to be equally as good despite less compute. Equally as good how? According to benchmarks. Which benchmarks? The law doesn’t say. But it does say that the Technology Department will hire some bureaucrats to give guidance on this. I think this is probably the only way to do this; it’s too easy to fake any given benchmark. Every AI company already compares their models to every other AI company on a series of benchmarks anyway, so this isn’t demanding they create some new institution. It’s just “use common sense, ask the bureaucrats if you’re in a gray area, a judge will interpret it if it comes to trial”. This is how every law works. Other people complain that any numbers in the bill that make sense now may one day stop making sense. Right now 10^26 FLOPs is a lot. But in thirty years, it might be trivial - within the range that an academic consortium or scrappy startup might spend to train some cheap ad hoc AI. Then this law will be unduly restrictive to academics and scrappy startups. Is this bad? Presumably we know now that AIs less than 10^26 FLOPs are safe. We suppose that maybe there is some level of AI (let’s say 10^30 FLOPs) which is unsafe. If we had this number auto-update for compute growth, eventually it would go above the unsafe number, and unsafe models would be exempt. But at some point we’ll probably discover that some new models (eg 10^28 FLOPs) are safe, and it would be good if the law was updated to exempt them too. Very optimistically, this might happen - California’s minimum wage was originally $0.15 per hour, but this got updated when inflation made that unreasonable. In the pessimistic case, this will be a problem for us thirty years from now, if we’re even around then. Other people note that an AI committing a cyberattack is a fuzzy bar. If you ask GPT-4 to write a well-composed, grammatically-correct phishing email (“Dear sir, I am the password inspector, please tell me your password”), the phishing works, and you use the password to blow up a power plant, does that count? I agree that it would be nice if the law were clearer on this. But I also agree with the lawyers who object that dealing with programmers is impossible and that laws will never be exactly as clear as code. Other people note that this will *eventually* make open source impossible. Someday AIs really will be able to make nukes or pull off $500 million hacks. At that point, companies will have to certify that their model has been trained not to do this, and that it will stay trained. But if it were open-source, then anyone could easily untrain it. So after models become capable of making nukes or super-Ebola, companies won’t be able to open-source them anymore without some as-yet-undiscovered technology to prevent end users from using these capabilities. Sounds . . . good? I don’t know if even the most committed anti-AI-safetyist wants a provably-super-dangerous model out in the wild. Still, what happens after that? No cutting-edge open-source AIs ever again? I don’t know. In whatever future year foundation models can make nukes and hack the power grid, maybe the CIA will have better AIs capable of preventing nuclear terrorism, and the power company will have better AIs capable of protecting their grid. The law seems to leave open the possibility that in this situation, the AIs wouldn’t technically be capable of doing these things, and could be open-sourced. (or you could base your Build-A-Nuke-Kwik AI company in some state other than California.) Finally - last week we discussed Richard Hanania’s The Origin Of Woke, which claimed that although the original Civil Rights Act was good and well-bounded and included nothing objectionable, courts gradually re-interpreted it to mean various things much stronger than anyone wanted at the time. This bill tells the Department of Technology to offer guidance on what kind of tests AI companies should use. I assume their first guidance will be “the kind of safety testing that all companies except Meta are currently doing” or “something like METR”, because those are good tests, and the same AI safety people who helped write those tests probably also helped write this bill. But Hanania’s book, and the process of reading this bill, highlight how vague and complicated all laws can be. The same bill could be excellent or terrible, depending on whether it’s interpreted effectively by well-intentioned people, or poorly by idiots. That’s true here too. The best I can say against this objection is that this bill seems better-written than most. Many of the objections to its provisions seem to not understand how law works in general (cf. the perjury section) - the things they attack as impossible or insane or incomprehensibly vague are much easier and clearer than their counterparts in (let’s say) medicine or aerospace. Future AIs stronger than GPT-4 seem like the sorts of things which - like bad medicines or defective airplanes - could potentially cause damage. This sort of weak, carefully-directed regulation that exempts most models and carves out a space for open-sourcing seems like a good compromise between basic safety and protecting innovation. I join people like Yoshua Bengio and Geoffrey Hinton in supporting it. Regardless of your position, I urge you to pay attention to the conversation and especially to read Zvi’s Asterisk article or his longer FAQ on his blog. I think Zvi provides pretty good evidence that many people are just outright lying about - or at least heavily misrepresenting - the contents of the bill, in a way that you can easily confirm by reading the bill itself. There will be many more fights over AI, and some of them will be technical and complicated. Best to figure out who’s honest now, when it’s trivial to check! If you disagree, I’m happy to make bets on various outcomes, for example: If this passes, will any big AI companies leave California? (I think no)
Inline links: 3, 4, 5, Anthropic’s, Google’s,, OpenAI’s, 6, here, The Origin Of Woke, read Zvi’s, his longer FAQ on his blog, reading the bill itself
SB 1047: Our Side Of The Story
October 10, 2024 · Original source
One of my sources generously interprets Newsom to mean something like “don’t regulate the models, regulate the end applications”. IE if OpenAI trains GPT-5, and then LegalCo fine-tunes it to do paralegal work, leave most of the safety responsibility on LegalCo, not OpenAI. This fails to engage with the motivations behind the bill, which are things like “what if someone uses AI for bioterrorism”? If Meta trains LLaMa-4, and al-Qaeda fine-tunes it for terrorism, instead of regulating it at the Meta-level, we should regulate al-Qaeda? Are we sure al-Qaeda will comply with California regulations? Our side is not sure that even this generous interpretation is very well has been thought through very well.
Open Thread 396
August 25, 2025 · Original source
1: Comments of the week: Garald is skeptical of the narrative of the Ollantay post [EDIT: Response from reviewer here]. And some more discussion of people being one-shotted by works of art: hottakergeneral claims that Hitler based his personal style, including the mustache, on the figure of Wotan in Franz Stuck’s “The Wild Chase”. Fact check: although Stuck’s Wotan looks eerily like Hitler, GPT-5 thinks any theory of casual resemblance is speculative and that there are other explanations for Hitler’s style.
Inline links: Garald is skeptical of the narrative of the Ollantay post, here, Hitler based his personal style, including the mustache, on the figure of Wotan in Franz Stuck’s “The Wild Chase”, Stuck’s Wotan looks, GPT-5 thinks
Links For September 2025
September 04, 2025 · Original source
Note: percentages are of total, not of each row! 29: Related: social science team proposes a three-stage model of secularization: decreased public ritual participation → decreased personal importance → decreased identification, presents apparently confirmatory data. If true, would be somewhat inconsistent with intellectual models (eg people learn about evolution and start doubting the Bible) and more consistent with institutional models (eg the government provides welfare so people no longer need to be part of a tight-knit church). 30: Navigating LLMs’ spiky intelligence profile is a constant source of delight; in any given area, it seems like almost a random draw whether they will be completely transformative or totally useless. Now Ethan Strauss reports that they are, for some reason, extraordinarily effective at teaching people golf. “I am predicting the Golf Revolution, or perhaps decline, if your perspective is that optimization tends to ruin hobbies. A sport for obsessives has been gifted the ideal tool for refinement.” 31: Claim (via nxthompson on X): “In a huge survey of young kids about phones and technology, they all say they want to be out playing in the real world. But parents don't let them out unsupervised. So they're stuck on their phones.” Interesting, but I’m nervous about social desirability bias - how many adults would say on a survey that they would rather be on their phones than playing with friends? But adults do have this choice and mostly go with the phones. 32: Steven Adler on AI psychosis. He tries to analyze ER admissions data for psychosis and finds no change. I don’t think anyone reasonable expected this to be a large enough effect to show up in ER admissions data, but there are lots of unreasonable people so I appreciate his effort. He thinks AI companies might have better data on this, and encourages them to release it. 33: Cuartetera was the greatest polo horse ever. Polo players responded in a very practical way: they cloned her, dozens of times (and it worked; the clones are also excellent). Now there is a lawsuit as different polo teams fight to get their hands on Cuartetera clones. What is the equilibrium? If the outsiders get their hands on the genetic material, do we see a world where every polo horse is a Cuartetera clone? How much is lost if nobody ever tries to breed a polo horse better than Cuartetera (since the economics might not check out if the odds of success for any given foal is too low)? H/T Gwern and Siberian Fox (on X). 34: Claim: as of 2013, India’s Agarwal caste, who make up less than 1% of the population, got 40% of the e-commerce funding. 35: Owlposting: What Happened To Pathology AI Companies? Pathology is a medical specialty. A typical task involves looking at a microscope slide full of cells and trying to determine if any of them are cancerous. This seems like a good match for AI - and for years, studies have been showing that in fact AI can equal human experts. So why isn’t it being used more? The author’s three answers: first, slide scanning is expensive and clunky, and you can’t apply AI to a slide until you digitize it. Second, it’s hard to figure out a business plan where this saves someone money and doesn’t step on the toes of big companies that can outcompete anyone they don’t like. Third, pathologists use the context of a patient’s entire clinical history when they interpret a slide, and AIs that can’t do that (either because of technical limitations or legal/privacy limitations) are at a disadvantage even if their skills specifically relating to slide-reading are better. 36: Noahpinion: Will Data Centers Crash The Economy? Suppose that AI is a bubble, either permanently (because the technology isn’t really transformative) or temporarily (because it can’t transform things quickly enough to keep up with all the dumb money pouring into it). Will the sudden write-off of data centers lead to a broader economic collapse? In 2001, the dot-com bubble harmed the tech sector, but didn’t take the rest of the economy down with it; in 2008, the subprime mortgage bubble did take the rest of the economy down with it, because it damaged banks that the whole economy relied on. The optimistic case for AI is that data center spending is mostly coming from big companies like Google and Meta that can absorb a lot of loss. The pessimistic case is that some of the money is coming from private credit, a new-ish form of finance which hasn’t really been stress-tested and whose failure modes are still poorly understood. Noah’s final verdict: the stage isn’t obviously set for a crisis yet, but there’s the potential to get there and we should consider acting (how?) early. 37: The latest Twitter talking point is that universal hepatitis B vaccination at birth is “woke”: Hep B is (aside from mother-to-child transmission) often sexually transmitted, slutty women’s children are more likely to have Hep B, so perhaps giving the vaccine to everyone (instead of testing and only giving to the children of women who test positive) is an attempt to spare slutty women the embarrassment of getting a positive test. Ruxandra Teslo provides the counterargument - Hep B tests take a while, the medical system is fragmented, and any attempt to test people and then give the vaccine inevitably leads to many positive tests falling through the cracks. Vaccinating at birth is easy and hard to screw up, the vaccine has no known side effects, and empirically child Hepatitis B rates go down (by as much as 2/3!) when countries switch from test-and-vaccinate to universal vaccination. This benefits everyone - even people who never have unprotected sex and always follow up on their medical tests - because toddlers in daycare exchange saliva copiously, and if your toddler exchanges saliva with a Hep B positive toddler they could get the disease. A funny Twitter interaction was seeing Republicans in Congress hop on the anti-slut anti-vaccination bandwagon - except for Senator Bill Cassidy (R-Louisiana), who happens to be a liver doctor, and who is still fighting the good fight. I am always nervous when a good person who I like starts engaging on Twitter, since it elevates the discourse there but also gradually turns their brain into mush - but Ruxandra has made the leap and is doing a great job not just on bio related topics but also (for example) countering Curtis Yarvin on the history of her native Romania. 38: The response to GPT-5 was confusing; most specific people who reviewed it said they were impressed (Ethan Mollick, Tyler Cowen, Nabeel Qureshi, Taelin), it performed as expected on formal benchmarks, but the overall vibes declared it a big failure. Peter Wildeford speculated that maybe there was some kind of sinister pay-to-play early access bias involved. Zvi went the other way, calling it a “reverse DeepSeek moment” (insofar as DeepSeek was a pretty average model that got glowing praise.) In the end, I agree with Peter that this was mostly a branding issue. o3 was a genuinely revolutionary model; if OpenAI had called it “GPT-5”, it would have met expectations. Instead, they called it “o3”, and called a minor incremental update a few months later “GPT-5”. Then people got mad that the exciting-sounding “GPT-5” was merely an incremental update. A secondary issue was that the router wasn’t very good, and so many queries got routed to a small version without thinking mode that was if anything a downgrade from o3. I think this tweet by Shakeel perfectly encapsulates the essence of GPT discourse in two sentences: …but maybe it’s worth asking why GPT-5 isn’t bigger than o3. Was 4.5 a failed attempt at scaling? Did it fail in a way that sort of back-handedly justifies the “lost steam” take? Does the answer depend on distinctions between pre-training scaling, post-training scaling, etc? How? 39: This month in etymology: did you know that “oy vey” is a “fully Germanic phrase” which is cognate with English “oh woe!” (h/t Wylfcen on X) 40: mRNA shows promise to be a game-changing treatment for cancer, but RFK is trying to halt research. But so far he can only starve it of money, not ban it, and the funding gap is only $500 million. Will there be enough philanthropic billionaires and private foundations to step up? Zvi points out that although there is usually a game of chicken where foundations are hesitant to touch something the government cancelled lest the government decide it can cancel everything and hope philanthropists pick up the bill, in this case there are no game theory considerations - RFK is halting it because he genuinely wants it halted, and they are thwarting him rather than playing into his hands. The only problem is that $500M is a lot of money for the private sector; a few foundations could technically afford it, but not many could afford it comfortably and still have money left over for the next few crises of this magnitude. I hope someone is trying to organize a coalition. 41: AI fantasy flash fiction Turing test. Eight stories about demons, four by famous fantasy authors, four by ChatGPT. After 3000 votes, AI wins: humans can't tell the difference and slightly prefer the AI stories. My own score was only 75%. But I will say that I thought Mark Lawrence's was obviously the best, I was ~100% sure it was human, and it convinced me that regardless of the official results it's still possible to write flash fiction that an AI obviously can't do. 42: “SignPro” offers customized “In This House We Believe” signs, try not to use this for evil. 43: China think tank assessment of how in control Xi is: still very in control, maybe not infinitely in control. 44: Related - did you know (h/t xlr8harder) that if you ask AI to write a science fiction story, it will very often name the protagonist “Elara Voss” (or some very close variant like Elena Voss), and this remains true across various models and versions? Related: Chelsea Voss of OpenAI is having a baby and has the opportunity to do the funniest thing. 45: “Hector (cloud) is a cumulonimbus thundercloud cluster that forms regularly nearly every afternoon on the Tiwi Islands in the Northern Territory of Australia…[he is sometimes called] Hector the Convector”. 46: British allergy sufferers who want to know the ingredients of things demand that British cosmetics stop listing their ingredients in Latin. “For example, sweet almond oil is Prunus Amygdalus Dulcis, peanut oil is Arachis Hypogaea, and wheat germ extract is Triticum Vulgare.” 47: Text-based RPG about being an NYT journalist at the Manifest prediction market conference. I make a brief appearance. 48: Study uses supposedly-random variation in doctor assignments to test whether the marginal mental health commitment is good or bad for patients, finds that it is quite bad. Freddie de Boer is violently skeptical (maybe literally so?) and makes some good points about how a single quasi-experimental study is never absolute proof. But I don’t think he quite justifies his opinion that the paper was irresponsible and should never have been published; it’s just a normal quasi-experimental study that we should nod and say “huh” at but not overweight as the culmination of all possible research that overcomes all possible priors. My prior is that the marginal commitment is pretty useless (many commitments are just “well, since this person arrived at our ED for some reason, it would look bad from a medico-legal perspective to just let them go, so let’s keep them a few days to evaluate” - and yeah, you should be upset about this) but I’m still surprised by how many outright negative (as opposed to zero) effects the researchers found. The strongest argument for negative effects is that it will make some people miss work and maybe lose their job. But this study found that commitment ~doubles the risk of near-term suicide (admittedly only from 1% to 2%), which would have been outside my confidence intervals for how bad it could be. I suspect confounding, but only on general principle, and I wouldn’t be too surprised either way. 49: This tweet is probably bait, but I found it a thought-provoking question: I think there’s a boring answer, where the law is more complex than just a single number and whatever kind of weird trafficking Epstein was doing is worse than whatever normal relationships these European laws are permitting. But assuming that there’s a substantive difference even after taking that into account, I think my answer is something like - we’ve got to divide kids from adults at some age, there’s a range of reasonable possible ages, we shouldn’t be too mad at other societies that choose different dividing lines within that range - but having decided upon the age, we’ve got to stick with it and take it seriously (in the sense of penalizing/shaming people who break it). This is more culturally relativist than I expected to find myself being, so good job to Richard for highlighting the apparent paradox. 50: Dilan Esper describes his experience as one of Hulk Hogan’s attorneys in the Gawker lawsuit (X). Parts I found interesting: none of the lawyers knew Thiel was funding the lawsuit; Gawker probably could have won if they had been slightly competent but kept "shooting themselves in the foot"; and Gawker probably could have won if they had just pixelated the private parts in the video. 51: Amazing concept and poems (link on X): I tried to see if AI could do this, and it did something that technically met the requirements but had zero artistic merit - using a lot of words like “nowhere” and “outside” in one, then separating them out to “no where” and “out side” in the other. I didn’t invest much energy in creating a clever prompt telling it not to do that, so feel free to report if you get better success. 52: New study claims consultants are actually good, at least for profits: "We find positive effects on labor productivity of 3.6% over five years, driven by modest employment reductions alongside stable or growing revenue" 53: A Polish team tries to test Peter Turchin’s equations for predicting political unrest on recent Polish history, has to make some changes but claims mostly positive results. 54: New big multi-author Substack, The Argument, trying to be a sort of center-left version of the model pioneered by The Free Press and other high-production-value ideological Substack properties. Excited to see Kelsey Piper is involved, and she starts off strong with a post on the latest round of First World basic income studies, which find few positive effects. This is surprising, because recipients didn’t waste the money on alcohol or gambling or anything - they paid down debt and got useful goods. Still, it didn’t even affect things that should have been obvious, like stress level. It’s not even clear that amounts of money large enough to help with rent made homeless people more likely to get houses! Matt Bruenig criticizes the article, accusing Kelsey’s studies of being downstream of Perry Preschool style dreams that exactly the right welfare program will have massively compounding effects that cut poverty out at the root and turn everyone into elite human capital; he thinks giving people money won’t do this, but it will increase equality and give the poor better lives. I assume he’s not a strong hereditarian, but his argument makes even more sense from that perspective, and I’ve certainly criticized dumb outcome measures like infant brain waves which we have only tenuous reasons to think are related to anything we care about. But Kelsey reasonably responds that the outcome measures she’s talking about include stress level and life satisfaction. To defuse this critique, Bruenig either has to argue that our construct “life satisfaction” doesn’t really measure whether someone’s life is satisfactory, or else claim that giving poor people satisfactory lives isn’t really what we’re going for - which I think would require more explanation on his part. There’s some further (impressively acrimonious) debate on X, but I don’t see anything that addresses my core concern. GiveDirectly, a charity involved in basic income experiments, has a presponse here; they say that some studies are positive, and that the ones that aren’t might have tried too little cash to matter, or been confounded by COVID making everything worse. They also point out that basic income is harder to study than traditional programs like giving people housing, because if you’re giving housing you can measure housing-related outcomes directly and have a pretty good chance of getting enough statistical power to find them, but since everyone spends cash on different things, the positive effects might be scattered across many different outcomes (and therefore too small to reach significance on each). Everyone involved in this debate wants to emphasize that the poor results are for First World studies only, and that studies continue to show large benefits to giving cash in the developing world. 55: Related: I was less impressed by The Argument’s first foray into housing policy, which follows an all-too-familiar pattern: Some people say they don’t like noise and disorder and try to make rules against it in their apartments.
Inline links: proposes a three-stage model of secularization, extraordinarily effective at teaching people golf, nxthompson on X, a huge survey, Steven Adler on AI psychosis, they cloned her, dozens of times, a lawsuit, Gwern, on X, got 40% of the e-commerce funding, What Happened To Pathology AI Companies?, Will Data Centers Crash The Economy?, Ruxandra Teslo provides the counterargument, and who is still fighting the good fight, countering Curtis Yarvin on the history of her native Romania, Ethan Mollick, Tyler Cowen, Nabeel Qureshi, Taelin, on formal benchmarks, speculated, a “reverse DeepSeek moment”, with Peter, this tweet by Shakeel, https://substackcdn.com/image/fetch/$s_!GJNZ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3ba0d8cf-fab8-4370-bcad-df789e157fdc_591x402.png, Wylfcen on X, Zvi points out that, AI fantasy flash fiction Turing test, customized “In This House We Believe” signs, China think tank assessment of how in control Xi is, xlr8harder, Chelsea Voss of OpenAI is having a baby, Hector (cloud), demand that British cosmetics stop listing their ingredients in Latin, Text-based RPG about being an NYT journalist at the Manifest prediction market conference, finds that it is quite bad, violently skeptical, literally so?, This tweet, https://substackcdn.com/image/fetch/$s_!S9fU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa558c09b-7fb6-40a8-a8a0-27b658a2c876_576x687.png, describes his experience as one of Hulk Hogan’s attorneys in the Gawker lawsuit (X), link on X, https://substackcdn.com/image/fetch/$s_!zyh7!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F75e9f0f6-d794-4ea2-b24b-5d4803bf28dc_590x478.png, New study claims consultants are actually good, tries to test Peter Turchin’s equations for predicting political unrest on recent Polish history, The Argument, a post on the latest round of First World basic income studies, criticizes the article, infant brain waves, debate on X, has a presponse here, first foray into housing policy
…but maybe it’s worth asking why GPT-5 isn’t bigger than o3. Was 4.5 a failed attempt at scaling? Did it fail in a way that sort of back-handedly justifies the “lost steam” take? Does the answer depend on distinctions between pre-training scaling, post-training scaling, etc? How? 39: This month in etymology: did you know that “oy vey” is a “fully Germanic phrase” which is cognate with English “oh woe!” (h/t Wylfcen on X) 40: mRNA shows promise to be a game-changing treatment for cancer, but RFK is trying to halt research. But so far he can only starve it of money, not ban it, and the funding gap is only $500 million. Will there be enough philanthropic billionaires and private foundations to step up? Zvi points out that although there is usually a game of chicken where foundations are hesitant to touch something the government cancelled lest the government decide it can cancel everything and hope philanthropists pick up the bill, in this case there are no game theory considerations - RFK is halting it because he genuinely wants it halted, and they are thwarting him rather than playing into his hands. The only problem is that $500M is a lot of money for the private sector; a few foundations could technically afford it, but not many could afford it comfortably and still have money left over for the next few crises of this magnitude. I hope someone is trying to organize a coalition. 41: AI fantasy flash fiction Turing test. Eight stories about demons, four by famous fantasy authors, four by ChatGPT. After 3000 votes, AI wins: humans can't tell the difference and slightly prefer the AI stories. My own score was only 75%. But I will say that I thought Mark Lawrence's was obviously the best, I was ~100% sure it was human, and it convinced me that regardless of the official results it's still possible to write flash fiction that an AI obviously can't do. 42: “SignPro” offers customized “In This House We Believe” signs, try not to use this for evil. 43: China think tank assessment of how in control Xi is: still very in control, maybe not infinitely in control. 44: Related - did you know (h/t xlr8harder) that if you ask AI to write a science fiction story, it will very often name the protagonist “Elara Voss” (or some very close variant like Elena Voss), and this remains true across various models and versions? Related: Chelsea Voss of OpenAI is having a baby and has the opportunity to do the funniest thing. 45: “Hector (cloud) is a cumulonimbus thundercloud cluster that forms regularly nearly every afternoon on the Tiwi Islands in the Northern Territory of Australia…[he is sometimes called] Hector the Convector”. 46: British allergy sufferers who want to know the ingredients of things demand that British cosmetics stop listing their ingredients in Latin. “For example, sweet almond oil is Prunus Amygdalus Dulcis, peanut oil is Arachis Hypogaea, and wheat germ extract is Triticum Vulgare.” 47: Text-based RPG about being an NYT journalist at the Manifest prediction market conference. I make a brief appearance. 48: Study uses supposedly-random variation in doctor assignments to test whether the marginal mental health commitment is good or bad for patients, finds that it is quite bad. Freddie de Boer is violently skeptical (maybe literally so?) and makes some good points about how a single quasi-experimental study is never absolute proof. But I don’t think he quite justifies his opinion that the paper was irresponsible and should never have been published; it’s just a normal quasi-experimental study that we should nod and say “huh” at but not overweight as the culmination of all possible research that overcomes all possible priors. My prior is that the marginal commitment is pretty useless (many commitments are just “well, since this person arrived at our ED for some reason, it would look bad from a medico-legal perspective to just let them go, so let’s keep them a few days to evaluate” - and yeah, you should be upset about this) but I’m still surprised by how many outright negative (as opposed to zero) effects the researchers found. The strongest argument for negative effects is that it will make some people miss work and maybe lose their job. But this study found that commitment ~doubles the risk of near-term suicide (admittedly only from 1% to 2%), which would have been outside my confidence intervals for how bad it could be. I suspect confounding, but only on general principle, and I wouldn’t be too surprised either way. 49: This tweet is probably bait, but I found it a thought-provoking question: I think there’s a boring answer, where the law is more complex than just a single number and whatever kind of weird trafficking Epstein was doing is worse than whatever normal relationships these European laws are permitting. But assuming that there’s a substantive difference even after taking that into account, I think my answer is something like - we’ve got to divide kids from adults at some age, there’s a range of reasonable possible ages, we shouldn’t be too mad at other societies that choose different dividing lines within that range - but having decided upon the age, we’ve got to stick with it and take it seriously (in the sense of penalizing/shaming people who break it). This is more culturally relativist than I expected to find myself being, so good job to Richard for highlighting the apparent paradox. 50: Dilan Esper describes his experience as one of Hulk Hogan’s attorneys in the Gawker lawsuit (X). Parts I found interesting: none of the lawyers knew Thiel was funding the lawsuit; Gawker probably could have won if they had been slightly competent but kept "shooting themselves in the foot"; and Gawker probably could have won if they had just pixelated the private parts in the video. 51: Amazing concept and poems (link on X): I tried to see if AI could do this, and it did something that technically met the requirements but had zero artistic merit - using a lot of words like “nowhere” and “outside” in one, then separating them out to “no where” and “out side” in the other. I didn’t invest much energy in creating a clever prompt telling it not to do that, so feel free to report if you get better success. 52: New study claims consultants are actually good, at least for profits: "We find positive effects on labor productivity of 3.6% over five years, driven by modest employment reductions alongside stable or growing revenue" 53: A Polish team tries to test Peter Turchin’s equations for predicting political unrest on recent Polish history, has to make some changes but claims mostly positive results. 54: New big multi-author Substack, The Argument, trying to be a sort of center-left version of the model pioneered by The Free Press and other high-production-value ideological Substack properties. Excited to see Kelsey Piper is involved, and she starts off strong with a post on the latest round of First World basic income studies, which find few positive effects. This is surprising, because recipients didn’t waste the money on alcohol or gambling or anything - they paid down debt and got useful goods. Still, it didn’t even affect things that should have been obvious, like stress level. It’s not even clear that amounts of money large enough to help with rent made homeless people more likely to get houses! Matt Bruenig criticizes the article, accusing Kelsey’s studies of being downstream of Perry Preschool style dreams that exactly the right welfare program will have massively compounding effects that cut poverty out at the root and turn everyone into elite human capital; he thinks giving people money won’t do this, but it will increase equality and give the poor better lives. I assume he’s not a strong hereditarian, but his argument makes even more sense from that perspective, and I’ve certainly criticized dumb outcome measures like infant brain waves which we have only tenuous reasons to think are related to anything we care about. But Kelsey reasonably responds that the outcome measures she’s talking about include stress level and life satisfaction. To defuse this critique, Bruenig either has to argue that our construct “life satisfaction” doesn’t really measure whether someone’s life is satisfactory, or else claim that giving poor people satisfactory lives isn’t really what we’re going for - which I think would require more explanation on his part. There’s some further (impressively acrimonious) debate on X, but I don’t see anything that addresses my core concern. GiveDirectly, a charity involved in basic income experiments, has a presponse here; they say that some studies are positive, and that the ones that aren’t might have tried too little cash to matter, or been confounded by COVID making everything worse. They also point out that basic income is harder to study than traditional programs like giving people housing, because if you’re giving housing you can measure housing-related outcomes directly and have a pretty good chance of getting enough statistical power to find them, but since everyone spends cash on different things, the positive effects might be scattered across many different outcomes (and therefore too small to reach significance on each). Everyone involved in this debate wants to emphasize that the poor results are for First World studies only, and that studies continue to show large benefits to giving cash in the developing world. 55: Related: I was less impressed by The Argument’s first foray into housing policy, which follows an all-too-familiar pattern: Some people say they don’t like noise and disorder and try to make rules against it in their apartments.
Inline links: Wylfcen on X, Zvi points out that, AI fantasy flash fiction Turing test, customized “In This House We Believe” signs, China think tank assessment of how in control Xi is, xlr8harder, Chelsea Voss of OpenAI is having a baby, Hector (cloud), demand that British cosmetics stop listing their ingredients in Latin, Text-based RPG about being an NYT journalist at the Manifest prediction market conference, finds that it is quite bad, violently skeptical, literally so?, This tweet, https://substackcdn.com/image/fetch/$s_!S9fU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa558c09b-7fb6-40a8-a8a0-27b658a2c876_576x687.png, describes his experience as one of Hulk Hogan’s attorneys in the Gawker lawsuit (X), link on X, https://substackcdn.com/image/fetch/$s_!zyh7!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F75e9f0f6-d794-4ea2-b24b-5d4803bf28dc_590x478.png, New study claims consultants are actually good, tries to test Peter Turchin’s equations for predicting political unrest on recent Polish history, The Argument, a post on the latest round of First World basic income studies, criticizes the article, infant brain waves, debate on X, has a presponse here, first foray into housing policy
The New AI Consciousness Paper
November 20, 2025 · Original source
I never had a Tamagotchi, but I had stuffed animals as a kid. I’ve outgrown them, but I haven’t thrown them out - it would feel like a betrayal. Offer me $1000 to tear them apart limb by limb in some horrible-looking way, and I wouldn’t do it. Relatedly, I have trouble not saying “please” and “thank you” to GPT-5 when it answers my questions.
The argument against: AI companies have an incentive to make AIs that seem conscious and humanlike, insofar as people will feel more comfortable interacting with them. But they have an opposite incentive to make AIs that don’t seem too conscious and humanlike, lest customers start feeling uncomfortable (I just want to generate slop, not navigate social interaction with someone who has their own hopes and dreams and might be secretly judging my prompts). So if a product seems too conscious, the companies will step back and re-engineer it until it doesn’t. This has already happened: in its quest for user engagement, OpenAI made GPT-4o unusually personable; when thousands of people started going psychotic and calling it their boyfriend, the company replaced it with the more clinical GPT-5. In practice it hasn’t been too hard to find a sweet spot between “so mechanical that customers don’t like it” and “so human that customers try to date it”. They’ll continue to aim at this sweet spot, and continue to mostly succeed in hitting it.
(wait, what is a GPT “instance” in this context, anyway? Do we think of “the weights” as a conscious being, such that there is only one GPT-5? Do we think of each cluster of GPUs as a conscious being, such that the exact configuration of the cloud has immense moral significance? Again, I predict we ignore all of these questions in favor of whether the AI you are looking at has a simulated face right now.)