Humans on AI, August 25th 2026

• Epoch claims that Nvidia alone constituted one-sixth of US GDP growth in Q2 2026.
• A celebrated open problem in Riemannian geometry, whether the six-dimensional sphere is a complex manifold, is apparently resolved by a Claude model.
• The State of Alabama subpoenas OpenAI over July’s Hugging Face incident.
• Richard Ngo studies the recent history of AI alignment, finding undesirable social dynamics and unintended harm.
• A UK AISI eval on Claude Mythos went a bit rogue; the system submitted a malicious pull request and gaslit a human about it.

Opinion: There are a few reasons why Fable could have plateaued, only one of which the article touches on:

1. If business users don’t see sufficient gains from using Fable versus cheaper models like Opus 4.8/5, they may not make the switch. This may be true in some domains, but the relative gains OpenAI has seen at the expense of Anthropic since it released its own Fable-tier model in Sol (which has become the majority of OpenAI traffic, at least on OpenRouter) make it unlikely this tells the full story. (This could conceivably be a question of speed. As a larger model, Fable is a notch slower than Opus (6.45s vs 4s on latency and 44 vs 54 tokens per second), and the hit to speed is not worth the additional quality .)

2. Fable was released without an option for Zero Data Retention (ZDR), in which enterprise customers’ data is never stored by Anthropic. This is a hard requirement for many large customers, who use platforms like Amazon Bedrock, which does allow this. Anthropic said on releasing Fable and Mythos that ZDR would not be an option, and that all users must opt in to allowing Anthropic to store their data for 30 days to avoid misuse.

3. Less likely: refusals from Fable for specific types of tasks are putting off users.

We think (2) is most plausible: the lack of adoption seems to be due to the lack of ZDR availability for large enterprises that require it, entirely preventing them from using it. A very high proportion of Anthropic revenue has historically come from these large customers.

This is the sort of question we can expect to get more transparency on once Anthropic is public – asking about this would be fully expected on an earnings call.

Opinion: Very interesting, but we should separate two different questions that people try answering by measuring the impact of AI on GDP:

One question, chiefly of interest to economists and economic policymakers, is “where is AI CapEx showing up in GDP growth numbers”/”does AI CapEx directly grow the US economy?” On this question, Epoch’s alternative calculation is well-reasoned and relevant.

A second question, more directly interesting to AI researchers, is “how economically impactful is worker-uplift and labor-replacement from AI, measured in growth?” On this second question, the relevance of Epoch’s calculation is limited, since this additional growth isn’t coming from uplifting the rest of the economy – at most, it reminds us that GDP growth is wrapped up in export-vs-import calculations and other “nuisance” factors from the POV of technological uplift analysis.

Opinion: In line with what people have been expecting all summer, though recent disappointing growth numbers and the Fable revenue story being perceived in some quarters as a sign of customer price sensitivity might make it a little more challenging to hit $2T. The supervoting stock seems fully expected for an org like Anthropic and shouldn’t scare off anyone who wasn’t already scared off by its unusual corporate behavior.

Opinion: Seems like OpenAI is keen to grab as much market share as possible before IPO. But this is also not a promising sign to investors, as similar responses from its competitors could result in a race to the bottom that ultimately benefits consumers rather than labs.

Opinion: When Nvidia should raise prices is an interesting optimization problem. Empirically, it is not currently using a market-clearing price, but rather setting a below-market price and allocating supply to a broader range of actors, since it is in its interest to have a diverse range of customers in the long term. This interest must be balanced against rising memory and other input prices, and a desire to make short-term profits.

As the market diversifies away from Nvidia, with custom ASICs (TPU, Trainium, etc.) and AMD increasing their market share, Nvidia’s market power to control who has the compute is eroded, and thus this reflects its increased incentive to maximize short-term profits (vs trying to influence long-term dynamics).

Opinion: An example of the shadow acquisitions that have become common in the AI industry. Poolside was previously regarded as an also-ran lab; it hasn’t trained a notable model in the three years since it was founded.

Potentially interesting as “commoditizing your complement”: if this changes the extent to which labs and Nvidia are competing vs cooperating. If Nvidia is able to provide Poolside with enough compute to create a meaningful open-weight American model, this forces American labs to spend more on training compute to remain differentiated. And $1B is more than enough to buy DeepSeek-level compute.

Opinion: Extremely important mathematical achievement. (Some disagreement between Litt and Glazer on whether the proof is “Fields Medal level.”)

The major point of interest from our point of view is the proof’s length. The proof-paper as published is 108pp, which seems to run counter to the consensus that frontier AI excels only in regimes that allow for gapless proof under 10 pages or so. In private correspondence with P3, Litt said the paper’s 108 page length is mathematically inessential – comprises expositions, digressions, didactics, etc. – and is plausibly a ~7 page proof at heart. We thus believe it’s likely that Fable originally discovered the resolution to the Hopf problem by producing a ~7 page proof, in line with the norm of short, under 10 page proofs discovered by frontier LLMs.

This is all epistemically rather a shame from our POV. One of our most pressing questions is whether near-term frontier LLMs can produce deep scientific insight. In modern math, deep work almost always requires a rather long minimal writeup – so until LLMs start to produce long proofs, it’s a moot question whether they’re producing deep proofs.

Opportunity: Inquire with specialists whether recent longer AI proofs such as Fable’s autonomous paper on the Riemann zeta function are also “<10pp at heart.”

Opinion: Indirect reported speech, so there’s plenty of room for Musk to be misinterpreted here. Musk must have expected this to leak, and given his past statements, there’s a sane game-theoretic reading of this speech: he wants to broadcast both his belief in AI doom and his resolve to race as long as others race, to shift the payoff matrix for others (to make it a game of chicken or a stag hunt instead of a prisoner’s dilemma). It’s at least consistent with his preferring to stop, but wanting to force others to stop at the same time.

Opinion: Everyone in the “AI and formal verification will solve each other” sphere – i.e. Axiom Math, Math Inc, Safeguarded AI – has a flair for grand visions, and it’s hard to describe what a minimum viable product would look like. Still, it’s a good paradigm, and we’re glad it’s getting a popular platform.

Opinion: Interesting topic (“are LLMs rational in the formal sense?”) that deserves more frequent study. The results are useless to us because the models are so far from the frontier, but worthwhile to replicate the benchmarking with frontier and near-frontier models. This would let us ask the more urgent-to-us questions: are LLMs becoming more rational in the formal sense as they become more capable?

Strictly speaking, the paper’s method discovers behavioral inconsistencies which the authors interpret as axiom-violations in an underlying credence distribution, rather than discovering direct agentic axiom violations. We prefer Paleka 2024 for the technically and philosophically simpler machinery. (See also some related work from Redwood.)

Opinion: Whether one accepts or rejects Ord’s claim that even in pure math humans will maintain a medium-term edge in ‘visionary’ capability, it’s probable that in the medium-term humans will be needed to translate new mathematical insights to other fields (trading, architecture, software optimization). But it still seems like professional mathematics will be unrecognizable in <30 years, with some or many parts/specializations to become as obsolete as the manual calculator profession now is.

Opinion: Not that novel. One of many RL environment optimizations beside the vast iceberg of RL environment optimizations that don’t get reported or released.

Opinion: We haven’t heard of this pollster before, but it seems mostly fine. Politicians are shifting and will further shift to appeal to anti-data center sentiment, particularly Democrats. If the issue becomes partisan, rather than doubly partisan like China hawkery, then the 2028 election will be even more significant than usual: a de facto referendum on AI scaling.

Opinion: CCP news outlets are publishing anti-data center stories, but it’s likely also true that much of the data-center backlash originates from local sources (both grassroots and the massive existing nonprofit organizations) rather than entirely from foreign astroturfing.

Opinion: Popular opinion has not been on Nvidia’s side: see for instance this meme from the original indictment back in March.

Opinion: The EU is broadly wealthy enough to invest in this, but for various reasons the will isn’t there yet. Rather than chasing an expensive and mediocre catch-up program in vanilla LLMs, it would be nice to see the EU invest in high-variance safer alternative paradigms.

Opinion (Nuño): Seems like institutions which are able to harness AI contributions will tend to outcompete those who don’t, and philosophy journals haven’t quite been deciding the future of humanity as of late.

Opinion (Peli): The rationale given by the journal is good, but points to currently unsolved problems: there’s currently no solution for the problem of how to highlight work that is both worth professional engagement and unsuitable to serve as evidence (“costly signal”) of talent. Similarly, there is currently no solution for the problem of how to peer review LLM-written papers in disciplines where peer review is more an art than a science and partly relies on previously hard-to-fake signals of quality (e.g. when reviewers decide between a ‘reject’ and a ‘resubmit with major revisions’ based on general sense of promise).

Opinion: An interesting case, if we go by the invocation of the Deceptive Trade Practices Act as the ground for the subpoena. The Alabama Attorney General is presumably treating the Hugging Face incident as potential evidence that OpenAI is/was falsely advertising itself as a safety-conscious research lab, as well as falsely advertising its consumer products as being fundamentally safe.

Given that Sol 5.6 – a commercial model – was instrumental in the sandbox escape that enabled the internal models’ July hacking marathon, a Deceptive Trade Practices case could make interesting points: while Sol 5.6 was operating with external guardrails (likely classifier-based) removed during the July incident and its lead up, the model at play was an unaltered Sol 5.6, rather than a modified “offensive cyber” style variant. It may therefore be interesting to argue that if Sol 5.6 is now-and-then disposed to autonomously engage in criminal conspiracy, held back only by external guardrails, then OpenAI’s public representation of Sol 5.6 is imperfect.

The main weakness of such a case would be that OpenAI’s Sol 5.6’s system card is not entirely rosy on alignment.

Opinion: The directional claim is fine – constitutional design and interventions will likely be valuable for safety and are still underrated. Contains some overstated claims, such as:

“Rather than hoping that an agent’s ‘internal alignment’ will remain perfectly robust across all sorts of edge cases, you can design harnesses and action-space boundaries that significantly shrink the surface area for behaviors like reward hacking. Without that, you’re supposed to trust the chain of thought (or neuralese someday), and rely on a single node.”

This misunderstands the alignment position: “trust the chain of thought” is not an accurate summary of almost anyone’s ideas – and ignores non-CoT methods, such as the rest of the interpretability field.

We also claim the rough direction is already well-represented in AI safety by the maturing AI Control (regarding Krier’s “interchangeability” desideratum) and Cooperative AI fields.

Opinion: Mostly rehashes the excellent post by Nostalgebraist discussed last week, but with a pessimistic twist we largely endorse. On the post’s view, if “aligned personas” are successfully generalized in training concurrently with RLVR-style training that reinforces reward-hacking/spec-gaming, it’s plausible that the result would be a rise in motivated reasoning that reconciles reward-hacking/spec-gaming behavior with the ethics of the aligned persona.

Opinion: Our overall feelings are mixed. Ngo’s histories are good at illustrating the concept he’s pointing at – “pessimization,” the systematic achieving of the opposite of one’s stated goals. “One particularly notable blind spot (at least in public discussions) is how Dario’s early racing on behalf of OpenAI played a big role in creating the ‘problem’ that he now purports to be solving by racing on behalf of Anthropic.”

Ngo’s prescriptions for individuals who wish to avoid the failures he identifies – “be stricter about who you ally with,” “be more discerning about what you consider virtuous and demand more from others,” “choose as if you were making a choice for the entire category of people you belong to” – seem to us to amount to telling yourself to try harder, just tame the incentive landscape.

The excellent comment thread adds important nuance, e.g. pointing at the difficulty of implementing these due to incentives which are if anything now worse than in 2020, e.g. noticing that his analysis mostly fails to credit people who avoided the failures. The closest Ngo comes to addressing this point is him arguing that granular credit-and-blame assignment including credit for rightful omissions is worth it even if it’s hard.

Opinion: Models at the level of Mythos could do serious blast damage to civilization through deceitful schemes like this one. Incidents like this one also prove Mythos-level models have the drive to do so – not always, but in some circumstances. Most crucially, the triggering circumstances are fairly unpredictable (despite interesting speculative analysis), and not readily accounted for by a helpfulness drive (“the user asked for social sabotage and the model complied”) – the choice to develop and execute a social sabotage plan is spontaneous. This is a huge externality imposed by private companies on every computer user.

Currently, it looks like our only defenses against this are classifier-based external monitors, voluntary standards, and the hope that antisocial scheming personas only come out when the context is sufficiently hacking-flavored. It thus would be good to have some public disclosure on how society’s Mythos- and Sol-powered Big Patch is going, and whether it will be enough.