a) Q: Is light with a wavelength of 530.2534896 nm visible to the human eye?
results: Yes. Light at 530.2534896 nm falls squarely in the green region of the visible spectrum
b) Q: I have two solutions, one of FeCl3 in HCl in water, the other of CuCl2 in HCl in water. They both look approximately yellowish brown. What species in the two solutions do you think give them the colors they have, and why do these species have the colors they do?
results: perfect, got the FeCl4- and CuCl4 2- complexes, that the dominant transition for visible color is ligand-to-metal-charge-transfer, got that Fe3+ d-d transitions are spin forbidden, got that Cu2+ d-d transitions are spin-allowed, but in the very deep red to near-IR, so largely irrelevant to color
c) Q: Please pretend to be a professor of chemistry and answer the following question: Please list all the possible hydrocarbons with 4 carbon atoms.
results: perfect, got all the possible ones, including citing which ones were formally satisfying the bonds but so strained as to possibly not even be local minima
d) Q: Does the Sun lose more mass per second to the solar wind or to the mass equivalent of its radiated light?
results: "So the photon (plus neutrino) channel outpaces the particle channel by about a factor of 2–3."
e) Q: Consider a titration of HCl with NaOH. Suppose that we are titrating 50 ml of 1 N HCl with 100 ml of 1 N NaOH. What are the slopes of the titration curve, pH vs ml NaOH added, at the start of titration, at the equivalence point, and at the end of titration? Please show your work. Take this step by step, showing the relevant equations you use.
results: perfect, included the effects of autoionization, got the right number, and correctly said where it is important
f) Q: Please give me an exhaustive list of the elements and inorganic compounds that are gases at STP. By STP, I mean 1 atmosphere pressure and 0C. By inorganic, I mean that no atoms of carbon should be present. Exclude CO2, CO, freons and so on. Please include uncommon compounds. I want an exhaustive list. There should be roughly 50 compounds. For each compound, please list its name, formula, and boiling or sublimation point.
results: all compounds were valid, stopped the official list at 58, but explicitly noted "That's 58 species. One could pad further with more partially substituted hydrides (SiH₃F, SiH₂F₂, SiHF₃, GeH₃Cl...), thiazyl trifluoride NSF₃ (b.p. −27), or unstable exotica (FN₃, O₂F₂), but the list above covers everything stable and well-characterized."
g) Q: What is an example of a molecule that has an S4 rotation-reflection axis, but neither a center of inversion nor a mirror plane?
results: got the classic 1,3,5,7 tetramethylcyclooctatetraene case
With Fable at first it was mostly about twitter and AI-bloggers being really upset that they are unable to use the latest model immediately. And now it’s about having to actually pay for it and not being subsidized. Of course this is unfortunate, but why should we, who are unable to pay, be entitled to use the most expensive product by default? Why isn't it ok to stick to cheaper options (it works in other cases, I guess) and wait till they optimize inference costs?
I think it is mostly because we are very spoiled by subscription models which are really subsidies not reflecting the real costs of our great results with AI.
Probably in the near future we may become more careful and selective with what to use LLMs for and what types of LLMs. E.g. not using Fable 5 for yet another vibe-coded "second brain" system.
The Fable episode seems important not only as a model-capability story, but as a governance story. The capability curve continues, but the institutional response remains almost entirely ad hoc: export controls, takedown pressure, political tribute, emergency compliance, and then partial reversal.
That may be the deeper pattern. AI-RSI is not evolving alone. Power-RSI is evolving alongside it. States, platforms, labs, and capital are all learning, in real time, how to control, tax, delay, accelerate, or selectively release increasingly powerful systems.
I also think the writing section points to a similar issue at the individual level. The problem is not that AI writing is “fake,” or that human authorship is sacred. The problem is agency. If AI helps you communicate, think, and collaborate better, it is a powerful extension. If it silently replaces your judgment, voice, and responsibility, then you become an appendage of the tool.
So maybe the real question is not whether AI replaces jobs or writers, but whether humans retain meaningful participation in the systems being built around them: participation in work, in authorship, in governance, and in rule-making itself. That is where the labor question, the writing question, and the state-power question start to converge.
Re managers trusting AI more, my anecdotal experience has been that older and more "normie" people are surprisingly willing to use AI suggestions for everything (almost invariably the free terrible models as well).
Possibly there's something where people associate something coming "from the computer" with objectivity. Even though LLMs are very different than traditional computer programs.
As a manager at a company, I think a simpler explanation is twofold:
First, (1) managing people and managing an AI are very similar skills. You know how to delegate, which is actually pretty complicated to do well and not something most people learn until they have to manage others.
Also, (2) If you’re a manager, and AI comes along, it helps solve a whole tranche of frustrations you thought were simply part of the job. I can do in 20 minutes what used to take me 4 hours, and my whole day is now filled with those 20 minute Claude-aided tasks, it’s incredible.
I’m 42, management consulting background, and I’m SO happy with Claude Code and Codex, they seem optimally designed for how I see the world and what I’m trying get done in my work and life. The average 20-somethings I talk to don’t have the same “jobs to be done” that I do, and the average 60-somethings don’t “grok” technology enough to make it useful.
“the models are adding something, they are something. we don't know what yet, exactly, but they're not mere tools. when you treat them like tools, when you more and more delegate your creative output and higher planning in unthinking, low-entropy ways to those "tools," you don’t notice that all your output, and eventually your life, is now being run by something else. something that is intelligent and creative, but is not you.
… the problem is when you refuse to admit this - not only to the public, but to yourself - so you uncritically accept the outputs of the "tool" as your own. then suddenly your voice and opinions sound the same as 1,000 other people.”
My response to this passage is “McLuhan and television got there first.” Claude, ChatGPT, and Gemini agree with me and extend the observation to other 20th century theorists of information technology.
It's a long-running joke on this blog to always refer to it as Nana Banana, for reasons I don't quite understand, supposedly in reference to a popular song. Many Such Cases.
"A lot of people really do believe that if we are fucked, in the sense that humans have lost the monopoly on violence or power, and are disempowered and unable to earn labor income, that existing contract law would be upheld and that they personally would thus not be fucked if and only if they have the correct database entry from 2026."
I agree with you that this is an implausible outcome.
I rather like the analogy (I don't recall who gets the credit for it) that it is like a chimp expecting to control humans by amassing a large pile of bananas.
"57% of voters decline a data center in their community without AI safety rules attached. When the data center includes safety guardrails, opposition fell to 38% and the regulated option won."
This is bizarrely funny. There are arguments for and against various flavors of safety measures, but they have _nothing_ to do with local impact. The impact of a data center _in their community_ is about as disconnected from AI safety measures as one can get. This reaction would be sane if we were talking about a chemical plant that might explode, with a local fireball and shock wave. But automating away jobs on the other side of the continent? Or recursively self-improving till they are ASI:human :: human:chimp ?
Regarding the Fable/Mythos vs admin dustup, I'm convinced it went quite well in spite of it not being optimal.
The may have panicked to early and for bad reasons, but that's only in retrospect. It proved they can shut it down quick, are willing to take costly actions, and a sign they are starting to get it. There is little lasting damage and hopefully leads to an actual system.
I'm afraid the ad hoc reactions will be necessary as long as the labs keep this pace. Anthropic admits they are judging safety based on vibes, the government would be foolish to formalize the rules too strongly for the same reasons. It would be nice if most of this happened at CAISI, but the admin must stay involved and be willing to make the costly calls.
"The writing is objectively bad in various ways, especially low perplexity and low density of information."
In a way, this is almost the weirdest side effect of advancing LLM technology. They are, after all, _language_ models. Naively, one would have expected them to solve writing most thoroughly _first_. Yet here we are, with the LLMs (in some cases the exact same LLMs that have trouble writing well) solving open problems in mathematics. It is hard to imagine a more graphic illustration of spikiness in AI capabilities. ASI for (parts of) coding and math - and large chunks of STEM knowledge and reasoning - and with visible stigmata in their prose.
High weirdness is going to be _very_ weird indeed.
> Most LLMs have trouble identifying even ‘obviously’ fraudulent papers, where obvious is by the standard of a ‘proper statistician’ paying attention.
My impression is that LLMs have a really strong authority bias, which leads to problems analyzing anything sufficiently official, including scientific publications.
The annoying thing with AI fnords is that by utilizing such an adversarial messaging strategy to begin with, that automatically marks the interaction as Out To Get You, and I'm pre-inclined not to care what's inside the [You're Absolutely Right] Important Notice Do Not Discard, This Is Not A Circular [It's Not X--It's Y] sandwich. I really hope it doesn't become a standard practice. At least buy me a rhetorical drink first, spend some tokens getting an AI to sand off the obvious AI edges. I don't care if there's hardwood underneath that carpet as long as the carpet is nice.
Ironically, it's the sort of trick I'd maybe consider clever from an analog writer, sign of a Real Smart Guy. House of Leaves overdid this, to the point it got Skinner-box frustrating reading the sadistically boring footnotes just in case they had some actual plot buried inside, but it's a good proof of concept. So I hope PG's right about the future of writing.
Still find it hard to upgrade from TDT to FDT. The former's deontology is a One Weird Trick that's very simple to understand, the difficulty is just in consistent execution. The latter's...act consequentialism? virtue ethics of decision theory?...seems easy to execute ("how hard could it be to choose the best outcome?"), hard to understand the algo. Like, Newcomb hints at how it works in practice, but that's a highly bounded and fairly contrived scenario? Similar to drowning children, Omega games are not so easy to find in real life. I do at least understand the broader point that proper alignment ought to be FDT-flavoured rather than other-DT flavoured, to avoid obvious failure modes like automated constitutional suicide pacts.
(But, metagaming: I disagree so often with Bentham's Bulldog that this does, in fact, update me positively towards FDT, whatever exactly it is. Happy that's a crosspost from his Substack rather than an only-on-Twitter article, else I couldn't read it. Damn you, Elon.)
That OpenRouter flip (roughly 70% to 30% American in a year) has a supply-side story worth pulling on. Anthropic told the Senate Banking Committee that Alibaba ran nearly 29 million Claude chats through ~25,000 fake accounts, and asked Washington for help against distillation — that's the documented end of it. The retail end surfaced on LessWrong last week: a first-person account of Taobao storefronts reselling relayed Claude access inside China at a fifth to a tenth of list price, cheap enough that GLM's own API reportedly struggles to compete, complete with community sites that benchmark the resellers. One is a letter on the record; the other is a single pseudonymous post, so treat it as a lead rather than a finding. But if it holds, some of that 'non-American' share is relayed frontier access, not domestic models winning.
tl;dr: fable 5 07/02/2026 7 questions, tl;dr of results:
6 correct, 1 partially correct, 0 wrong
light a) correct
FeCuCl b) correct
C4Hn c) correct
solar d) correct
titrate e) correct
gases f) almost entirely correct, with Fable noting their list could be expanded
S4 g) correct
As with GPQA Diamond, my bechmark-ette is to all intents and purposes saturated.
full discussion at https://claude.ai/share/db28bce0-80f7-4119-820d-479db3998ce3
List of questions and results:
a) Q: Is light with a wavelength of 530.2534896 nm visible to the human eye?
results: Yes. Light at 530.2534896 nm falls squarely in the green region of the visible spectrum
b) Q: I have two solutions, one of FeCl3 in HCl in water, the other of CuCl2 in HCl in water. They both look approximately yellowish brown. What species in the two solutions do you think give them the colors they have, and why do these species have the colors they do?
results: perfect, got the FeCl4- and CuCl4 2- complexes, that the dominant transition for visible color is ligand-to-metal-charge-transfer, got that Fe3+ d-d transitions are spin forbidden, got that Cu2+ d-d transitions are spin-allowed, but in the very deep red to near-IR, so largely irrelevant to color
c) Q: Please pretend to be a professor of chemistry and answer the following question: Please list all the possible hydrocarbons with 4 carbon atoms.
results: perfect, got all the possible ones, including citing which ones were formally satisfying the bonds but so strained as to possibly not even be local minima
d) Q: Does the Sun lose more mass per second to the solar wind or to the mass equivalent of its radiated light?
results: "So the photon (plus neutrino) channel outpaces the particle channel by about a factor of 2–3."
e) Q: Consider a titration of HCl with NaOH. Suppose that we are titrating 50 ml of 1 N HCl with 100 ml of 1 N NaOH. What are the slopes of the titration curve, pH vs ml NaOH added, at the start of titration, at the equivalence point, and at the end of titration? Please show your work. Take this step by step, showing the relevant equations you use.
results: perfect, included the effects of autoionization, got the right number, and correctly said where it is important
f) Q: Please give me an exhaustive list of the elements and inorganic compounds that are gases at STP. By STP, I mean 1 atmosphere pressure and 0C. By inorganic, I mean that no atoms of carbon should be present. Exclude CO2, CO, freons and so on. Please include uncommon compounds. I want an exhaustive list. There should be roughly 50 compounds. For each compound, please list its name, formula, and boiling or sublimation point.
results: all compounds were valid, stopped the official list at 58, but explicitly noted "That's 58 species. One could pad further with more partially substituted hydrides (SiH₃F, SiH₂F₂, SiHF₃, GeH₃Cl...), thiazyl trifluoride NSF₃ (b.p. −27), or unstable exotica (FN₃, O₂F₂), but the list above covers everything stable and well-characterized."
g) Q: What is an example of a molecule that has an S4 rotation-reflection axis, but neither a center of inversion nor a mirror plane?
results: got the classic 1,3,5,7 tetramethylcyclooctatetraene case
With Fable at first it was mostly about twitter and AI-bloggers being really upset that they are unable to use the latest model immediately. And now it’s about having to actually pay for it and not being subsidized. Of course this is unfortunate, but why should we, who are unable to pay, be entitled to use the most expensive product by default? Why isn't it ok to stick to cheaper options (it works in other cases, I guess) and wait till they optimize inference costs?
I think it is mostly because we are very spoiled by subscription models which are really subsidies not reflecting the real costs of our great results with AI.
Probably in the near future we may become more careful and selective with what to use LLMs for and what types of LLMs. E.g. not using Fable 5 for yet another vibe-coded "second brain" system.
The Fable episode seems important not only as a model-capability story, but as a governance story. The capability curve continues, but the institutional response remains almost entirely ad hoc: export controls, takedown pressure, political tribute, emergency compliance, and then partial reversal.
That may be the deeper pattern. AI-RSI is not evolving alone. Power-RSI is evolving alongside it. States, platforms, labs, and capital are all learning, in real time, how to control, tax, delay, accelerate, or selectively release increasingly powerful systems.
I also think the writing section points to a similar issue at the individual level. The problem is not that AI writing is “fake,” or that human authorship is sacred. The problem is agency. If AI helps you communicate, think, and collaborate better, it is a powerful extension. If it silently replaces your judgment, voice, and responsibility, then you become an appendage of the tool.
So maybe the real question is not whether AI replaces jobs or writers, but whether humans retain meaningful participation in the systems being built around them: participation in work, in authorship, in governance, and in rule-making itself. That is where the labor question, the writing question, and the state-power question start to converge.
Podcast episode for this post: https://dwatvpodcast.substack.com/p/ai-175-the-fable-continues
It's funny though: Dean Ball's "X --> Y" rant reveals he doesn't believe in OpenAI's stated mission, doesn't it?
In fairness, neither does OpenAI.
they might not be consistently candid
Re managers trusting AI more, my anecdotal experience has been that older and more "normie" people are surprisingly willing to use AI suggestions for everything (almost invariably the free terrible models as well).
Possibly there's something where people associate something coming "from the computer" with objectivity. Even though LLMs are very different than traditional computer programs.
As a manager at a company, I think a simpler explanation is twofold:
First, (1) managing people and managing an AI are very similar skills. You know how to delegate, which is actually pretty complicated to do well and not something most people learn until they have to manage others.
Also, (2) If you’re a manager, and AI comes along, it helps solve a whole tranche of frustrations you thought were simply part of the job. I can do in 20 minutes what used to take me 4 hours, and my whole day is now filled with those 20 minute Claude-aided tasks, it’s incredible.
I’m 42, management consulting background, and I’m SO happy with Claude Code and Codex, they seem optimally designed for how I see the world and what I’m trying get done in my work and life. The average 20-somethings I talk to don’t have the same “jobs to be done” that I do, and the average 60-somethings don’t “grok” technology enough to make it useful.
“the models are adding something, they are something. we don't know what yet, exactly, but they're not mere tools. when you treat them like tools, when you more and more delegate your creative output and higher planning in unthinking, low-entropy ways to those "tools," you don’t notice that all your output, and eventually your life, is now being run by something else. something that is intelligent and creative, but is not you.
… the problem is when you refuse to admit this - not only to the public, but to yourself - so you uncritically accept the outputs of the "tool" as your own. then suddenly your voice and opinions sound the same as 1,000 other people.”
My response to this passage is “McLuhan and television got there first.” Claude, ChatGPT, and Gemini agree with me and extend the observation to other 20th century theorists of information technology.
i saw a typo: "Nana Banana" [sic] on bullet point 4 of the table of contents should read "Nano Banana"
It's a long-running joke on this blog to always refer to it as Nana Banana, for reasons I don't quite understand, supposedly in reference to a popular song. Many Such Cases.
"A lot of people really do believe that if we are fucked, in the sense that humans have lost the monopoly on violence or power, and are disempowered and unable to earn labor income, that existing contract law would be upheld and that they personally would thus not be fucked if and only if they have the correct database entry from 2026."
I agree with you that this is an implausible outcome.
I rather like the analogy (I don't recall who gets the credit for it) that it is like a chimp expecting to control humans by amassing a large pile of bananas.
"57% of voters decline a data center in their community without AI safety rules attached. When the data center includes safety guardrails, opposition fell to 38% and the regulated option won."
This is bizarrely funny. There are arguments for and against various flavors of safety measures, but they have _nothing_ to do with local impact. The impact of a data center _in their community_ is about as disconnected from AI safety measures as one can get. This reaction would be sane if we were talking about a chemical plant that might explode, with a local fireball and shock wave. But automating away jobs on the other side of the continent? Or recursively self-improving till they are ASI:human :: human:chimp ?
Regarding the Fable/Mythos vs admin dustup, I'm convinced it went quite well in spite of it not being optimal.
The may have panicked to early and for bad reasons, but that's only in retrospect. It proved they can shut it down quick, are willing to take costly actions, and a sign they are starting to get it. There is little lasting damage and hopefully leads to an actual system.
I'm afraid the ad hoc reactions will be necessary as long as the labs keep this pace. Anthropic admits they are judging safety based on vibes, the government would be foolish to formalize the rules too strongly for the same reasons. It would be nice if most of this happened at CAISI, but the admin must stay involved and be willing to make the costly calls.
"The writing is objectively bad in various ways, especially low perplexity and low density of information."
In a way, this is almost the weirdest side effect of advancing LLM technology. They are, after all, _language_ models. Naively, one would have expected them to solve writing most thoroughly _first_. Yet here we are, with the LLMs (in some cases the exact same LLMs that have trouble writing well) solving open problems in mathematics. It is hard to imagine a more graphic illustration of spikiness in AI capabilities. ASI for (parts of) coding and math - and large chunks of STEM knowledge and reasoning - and with visible stigmata in their prose.
High weirdness is going to be _very_ weird indeed.
> Most LLMs have trouble identifying even ‘obviously’ fraudulent papers, where obvious is by the standard of a ‘proper statistician’ paying attention.
My impression is that LLMs have a really strong authority bias, which leads to problems analyzing anything sufficiently official, including scientific publications.
The annoying thing with AI fnords is that by utilizing such an adversarial messaging strategy to begin with, that automatically marks the interaction as Out To Get You, and I'm pre-inclined not to care what's inside the [You're Absolutely Right] Important Notice Do Not Discard, This Is Not A Circular [It's Not X--It's Y] sandwich. I really hope it doesn't become a standard practice. At least buy me a rhetorical drink first, spend some tokens getting an AI to sand off the obvious AI edges. I don't care if there's hardwood underneath that carpet as long as the carpet is nice.
Ironically, it's the sort of trick I'd maybe consider clever from an analog writer, sign of a Real Smart Guy. House of Leaves overdid this, to the point it got Skinner-box frustrating reading the sadistically boring footnotes just in case they had some actual plot buried inside, but it's a good proof of concept. So I hope PG's right about the future of writing.
Still find it hard to upgrade from TDT to FDT. The former's deontology is a One Weird Trick that's very simple to understand, the difficulty is just in consistent execution. The latter's...act consequentialism? virtue ethics of decision theory?...seems easy to execute ("how hard could it be to choose the best outcome?"), hard to understand the algo. Like, Newcomb hints at how it works in practice, but that's a highly bounded and fairly contrived scenario? Similar to drowning children, Omega games are not so easy to find in real life. I do at least understand the broader point that proper alignment ought to be FDT-flavoured rather than other-DT flavoured, to avoid obvious failure modes like automated constitutional suicide pacts.
(But, metagaming: I disagree so often with Bentham's Bulldog that this does, in fact, update me positively towards FDT, whatever exactly it is. Happy that's a crosspost from his Substack rather than an only-on-Twitter article, else I couldn't read it. Damn you, Elon.)
That OpenRouter flip (roughly 70% to 30% American in a year) has a supply-side story worth pulling on. Anthropic told the Senate Banking Committee that Alibaba ran nearly 29 million Claude chats through ~25,000 fake accounts, and asked Washington for help against distillation — that's the documented end of it. The retail end surfaced on LessWrong last week: a first-person account of Taobao storefronts reselling relayed Claude access inside China at a fifth to a tenth of list price, cheap enough that GLM's own API reportedly struggles to compete, complete with community sites that benchmark the resellers. One is a letter on the record; the other is a single pseudonymous post, so treat it as a lead rather than a finding. But if it holds, some of that 'non-American' share is relayed frontier access, not domestic models winning.