OpenAI has just released a vast number of new results in preprint form, see here. These are the products of an internal frontier model. About the AGMAI, they say they “have drawn on their advice and public recommendations to inform how we release these results.”, although one of the recommendations at the top was “we ask them to stop testing advanced mathematical problems on proprietary models.”
Like everyone else, I’ll spend some of this evening sorting through the long list to see what is of interest.
Update: Some quick thoughts about the slopocalypse. Note that this is happening in a seriously polluted and dysfunctional information environment. Maybe it’s my incompetence at dealing with social media, but the only way I’ve found of finding out relevant news about the slopocpalypse is by supporting Elon Musk’s Fascist enterprise and scrolling through an endless sewer of crap on X to find the few instances where well-informed mathematicians are quickly tweeting out relevant information. Lots of problems with getting your information this way, including walking around all the time depressed and disgusted at the state of the world. That said:
- A striking aspect of the paper dump is how sloppy it is at every level. The dump is a disorganized mess: after scrolling through long lists of papers and trying to read a couple, I realized that the huge list I’d gone through was only a fraction of the 722 papers, and that trying to find things this way would take forever. It’s hard not to think the people responsible for this just don’t care. If Astra is so great, couldn’t someone have taken 30 seconds to type into it a prompt “Please organize this in a coherent and useful way”?
- When the Navier-Stokes trainwreck happened, there was a concerted effort from within the math community to get to the public a counter-narrative to that of OpenAI. I’ve seen nothing like that this time. The AGMAI group supposedly dealing with OpenAI last week put out an excellent list of recommendations. OpenAI completely ignored them, and AGMAI then issued a statement I don’t understand at all. They are the people in the best position to evaluate whether their recommendations were followed, but instead of doing this, they write “it is ultimately up to the mathematical community to assess the extent to which our recommendations were followed successfully.”
- One thing you can learn on X today: some of the proofs are just wrong.
- Another one: some of the slop is supposedly verified by Lean. Can we trust this? Maybe not for Navier-Stokes, what about the hundreds of others?.
- Is Astra getting these proofs from its access to mathematician’s private work? See the first paragraph on page 3 here.
- According to OpenAI, these proofs were all produced with simple prompts to next version of Astra, applied to 4000 open problems in mathematics. What was the point of doing this? If the goal of OpenAI researchers is to improve their product, it would make sense to try a much more modest number of problems, then use the results to figure out how to do better (e.g. provide writeups that are readable…).
It seems to that the answer to this is that they were trying to exploit a weakness in the mathematics research incentive structure, where people earn credits by solving these sorts of problems, even though this is a proxy for what is really important (new understanding). What OpenAI chose to do was to carpet-bomb mathematics, grabbing as many as possible of the chips on the table. - I have a lot of trouble knowing what to think of the “AI alignment problem”, that of ensuring alignment between AI behavior and wholesome human goals. It does seem clear that the math community (and the rest of humanity) has a serious and immediate “OpenAI alignment problem”, with a huge misalignment between that corporation’s behavior and what is best for mathematics in particular and humanity in general.
- Another thing you learn on Elon Musk’s sewer is that a factor generating enthusiasm for our tech overlords and their new world is the exploitation of resentment of “elites”. In particular, criticism of the slopocalypse is widely denounced as nothing but the whining of pointy-headed over-privileged intellectuals who finally are getting theirs.
Update: There’s a very interesting collection of reactions to the slopocalypse here.


Dear Peter, as in any technological revolution there are positive effects and very dangerous developments. Personally I do not see all this “apocalyptic” future on mathematics itself (the subject will probably start to resemble more an experimental natural science, with many “experimentally AI-achieved results” that have to wait for “human understanding”, when that will be possible). The adoption of AI is often just giving more visibility to issues that were already deeply problematic in the professional community (hyper-production of papers, incorrect usage of of biblio-metric indexes for evaluation, excessive importance of problem-driven and incremental research, scarce availability and contraction of academic positions).
I am much more concerned about the effects on the open sharing of scientific results; the further privatization of research and the increased discrimination between communities that will have easy access to the new technology and those that will not.
On partially related (positive) news, there is a new preprint archive (https://hexagonmath.org/) that might be of relevance as a standard repository for results not suitable for arXiv (either because of AI-usage or because of arXiv endorsment/censorship policies); quoting from the web-site of the repository:
“Hexagon was created in response to the staggering changes that artificial intelligence has induced in the mathematics research landscape. It accepts research works of all kinds, provided that the contributor(s) believe in good faith that the work is correct and of interest to the mathematical community, and have made a good faith effort to ensure that it accurately cites prior work that should be cited. This includes works that the contributor may not wish to, or may not be able to, submit to other repositories such as the arXiv. Works that are also suitable for the arXiv are equally welcome, and a work may be posted to both.”
Best Regards. Paolo
Calling this “slop” seems to me another weird political opinion. An interesting question is why AI replaced human mathematicians before replacing physicists or truck drivers. Plausibly because math is a compact powerful language and mathematicians insistence on formalisation and axioms made their field perfect for sharp AI training and execution. Then maybe human physicists will survive by outsourcing math aspects to AI?
They proved Haldane’s conjecture in the “hardest” case for the spin-1 Heisenberg antiferromagnet in 1d with periodic bc’s, that the spectral gap is not zero. They used the Aizenman-Nachtergaele graphical representation which relates to a percolation model.
There is also a proof of the mass gap for the O(4) classical Heisenberg model on a two dimensional lattice.
@Alessandro: The writeups are pure disgusting slop. I looked at about twenty of them, and all of them are like that. It’s not hard to see this. Also, human mathematicians haven’t been replaced at all.
None of this slop. Why the snarky title?
Alessandro Strumia: it’s because math is relatively easily determined to be correct or not, which makes objective training easier. And it’s close to language. The same reasons (although in reverse order of significance) explain why AI “solved” programming first.
The adoption of AI is often just giving more visibility to issues that were already deeply problematic in the professional community (hyper-production of papers, incorrect usage of of biblio-metric indexes for evaluation, excessive importance of problem-driven and incremental research, scarce availability and contraction of academic positions).
Maybe, but it also seems to be exacerbating rather than ameliorating every one of those issues.
A prominent mathematician told me recently that the definition of a mathematician is “someone who teaches calculus to the unwilling.” So indeed by that definition, there is no imminent danger of them being replaced. On a related note, however, the AI companies are eager to hire boatloads of recent graduates with technical training, with math PhDs at the top of the pecking order, but at the same time that the demand for these people is growing, the opportunities to grow them are being cut: I heard that Berkeley math admitted only 5 PhD students this year due to the dire funding situation. Harvard is also in bad shape, having cut 50% of PhD admissions in all of FAS last year (they are partially rebounding this year). Personally I’m on strike from mentoring PhD students until Anthropic or OpenAI fund them, which would cost them peanuts.
Dear Clone of Weyl, you raise an interesting question: why AI writeups seem weird and inelegant, while having better content than humans? It could be just bad instructions. Or it could be something interesting: AI superior thinking power + “context” avoids human bias for elegance.
Broadly, mathematics can be thought of as problem-solving plus theory-building. AI has shown itself potent at problem-solving (e.g. at coming up with the Navier-Stokes counterexample). Question: Has it shown any aptitude at theory-building–at doing something like inventing Galois theory to settle the solvability of the quintic? Is its architecture suitable for creating new abstract ontologies? Has it done something like this already? My (fairly ignorant) impression is: no.
I have tried looking more closely at the O(4) Heisenberg model paper. The proofs are not at all clear, which makes me suspicious. I am not saying it is nonsense at this stage, but…
What’s confusing is that it replaces the lattice with a “Brownian regulator” and says these have some correspondence/equivalence. The paper contains stuff on rigorous inequalities (like GKS II), which I happen to know don’t work for this model (but maybe I am misunderstanding how these are applied). Either brilliant or just plain weird.
If someone else has tried to read this, maybe they have a different perspective.
Oh, looking more carefully, GKS II is applied only to Ising spins, which is probably okay.
A Theoretical Physics Faculty,
I hear that Harvard math did manage to admit the usual number of students, due to help from private funding. Very unclear what the future is for math phd students (will AI take over Calculus teaching?), or how to train them.
I feel like I’m hearing a lot of annoyance/disgust about OpenAI,Anthropic, Google, Meta, etc, often coming together with strategizing about how to get ahold of some of their trillions.
Jim Holt,
To me, that’s the big question right now.
On of my basic problems here is that I don’t understand how the current math AI agents are doing what they are doing. Looking at the results, they seem to be very good at finding intricate solutions to hard problems using known methods, combined with pretty terrible at explaining what they are doing. I’ll try to find time to look at more of these documents, the couple I did look at seemed to be pretty unreadable. Unfortunately human-written math papers sometimes are this way too.
What I haven’t seen is an AI agent solving a problem in an insightful way, or at least explaining an insight. In the past, when someone solved an old problem, sometimes it was by grinding away with known techniques, but often it was with a new insight. A different way of thinking about the problem that made the path to a solution clear. If among the 722 papers, there are any that basically say “in the past people tried to use X, Y, Z to solve this, but these can’t work because of this specific aspect of the problem. Recognizing this, method Q can be applied instead and it works. Here are the details….”
If any of the 722 papers are recognizably like this, I’d be curious to hear about it.
Because of the human brain’s problems dealing with too much complexity, great mathematical ideas have been ones that simplify things (often by “theory-building”). This allows non-silicon based life forms to grasp what is going on and work with it. Are current AI agents able to do this? Are there ones in the labs being developed? I’m very curious to see how this will play out.
Using AI to prove theorems is not going to stop. I understand why many (most?) mathematicians wouldn’t want to use AI but you just need a small subset who will use these tools to make it impractical for others to not use them.
Are there attempts by the mathematical community to post-train (with reinforcement learning + appropriate reward functions) open models (Qwen, deepseek, GLM) on math problems/theorems to improve readability or to align them with our notions of mathematical elegance? This would require significant GPU resources but not even close to the scale of pre-training. Simons Foundation, the Allen Institute, Red Hat etc. might be able to help.
This is obviously not AI slop in the original meaning of the term (e.g., the scripts and poster adds for that Willy Wonka fiasco in Glasgow some years ago). First, they are actually true (highly unlikely they are technically wrong given the Leans). Second, they have actual insight, since they solved hard problems that the best human mathematicians have been trying for decades and failing.
All that said, they do seem to retain some typical characteristics of slop. They are usully derivative of previous work rather than proposing a completely new way of looking at things, as human so-called “genius” often do. This is, I think, intrinsic to all modern AI by design, since they can only generalize “in-distribution” rather than doing big OOD leaps and still be correct as humans seem to do sometimes. No one in modern AI research has a clear idea of how to make systems capable of systematically doing this by design.
Also, the “explanations” they write of their findings are highly terse, technical, and assume a background from the readers as if they were an AI themselves. They often start mentioning highly technical points from the development of the proof before introducing any of it. I also found this in my own experience with using agents. This can be reverted, but requires additional prompting work to make them give a more human friendly explanation of things.
I suspect the Haldane conjecture proof may be clearer than others. It is only about 7 pages. I have not understood it completely yet, but I am hopeful. One sloppy thing about that paper: there are lots of citations. But there is no indication for any citation where in the paper where the cited ideas were used. For example, they use twisted boundary conditions, but they do not cite Sutherland or anybody else for originating this idea.
You ask “What was the point of doing this?” and I can only assume your question is rhetorical. I’m far from the first person to make this observation but: I consider it axiomatic that none of these companies care about math (or theoretical physics); math is just the battleground on which they are fighting their war at the moment. Let’s hope something of what we love is left standing when this is all over.
Peter, that paper on autoformalization is exactly to the point. The corpus was interpreted by the generations of humans who produced it, and LLM completions are interpreted when we take them up, but between the corpus and our taking up the completions there’s simply a statistical calculation. It’s frankly amazing to see people attribute to the machine that carries out the calculation characteristics that only make sense on the assumption that the machine is interpreting the strings of tokens. As concerns this, Scholze’s comment at proofsandprompts.com is the most sensible (in brief: if the aim of a proof is to increase human understanding of mathematics – he’s paraphrasing William Thurston’s famous line here – then the appearance of LLMs needn’t *essentially* change the practice).
Re: nihilists on x, and frankly some commenters here, it’s not surprising to see them embrace these machines as humanity’s comeuppance. They’re driven by resentment. Silly notions about a machine that implements a statistical calculation give oxygen to their grim hopes.
Before anyone complains that I’m putting my head in the sand: I’m not. These machines are, in fact, amazing, but not for the reasons that many people are saying. The machines also present serious challenges — see e.g. Scholze’s worry re: cryptography. But it’s time we start seeing with clear eyes what’s actually amazing about them, putting to one side sensationalizations (many originating fromm the companies themselves, for obvious reasons).
I don’t think it’s reasonable to suggest that the primary goal of OpenAI was to earn credits. If that was the goal, why not release solutions as soon as they are found, instead of all at once, to avoid getting scooped? A subtler point is that many results which are proven in the literature conditional on the Riemann Hypothesis in fact follow from the Quasi-Riemann Hypothesis they proved, and why not establish all of these results instead of just one (Artin’s primitive root conjecture)? Why not write at least one of the results up nicely and submit it to the Annals?
I believe “improve the product” is a factor. If they want to evaluate and compare different model’s problem-solving capability for any reason (RL, measuring improvement, evaluating different design choices) it makes sense to try a large number of problems instead of a smaller number to get higher-quality data.
However, OpenAI employees have stated that the biggest source of problems is that employees are allowed to use the models for what they want, and many of them want to solve these problems. A lot of OpenAI employees are mathematicians and physicists – the exact number depends on if you count those who have degrees in those fields or just those that currently have positions at academic institutions or just those who have put papers on the arXiv in the last year (including their own work and not simply AI work), but there are significant numbers of all of these. They of course care about the answers to certain problems, and asked the internal model to solve them.
Hi Will,
One problem here is that things like a “goal of OpenAI” are ill-defined since “OpenAI” refers to a bunch of people with different goals. Among the mathematicians there, probably different ones have different goals. Then the physicists have another set of goals, the leaders of these groups something different, and so on up to Sam Altman.
Your question “Why not write at least one of the results up nicely and submit it to the Annals?” is a good one. I’m guessing the answer is that “they” feel that they already showed they could do this kind of thing with their solution of Millennium problem.
Another version of the question though is why “they” didn’t at least try to follow the recommendations of the AGMAI group, e.g. unless one of the staff understood a result, collaborate on understanding it with experts and writing it up in understandable form? The professional math component of “they” surely understood the concerns raised by AGMAI about a dump, so why did it happen?
I’m guessing a lot of this is that “they” (here the professional math component) a while back made the decision to gather as huge a database of open math problems as they could, and when a new model was ready for use, set it to work on all of them (with a relatively small budget/problem). Who knows why they did this, but I’m guessing the frame of mind “what’s valuable in math is solving unsolved problems, so let’s solve as many as possible” had something to do with it.
After the Navier-Stokes blowback, “they” realized that things were not so simple, there were problems with the idea of publicly dumping these results. In the debate though over what to do, maybe the argument “these are of great value, we can’t deny them to the public” won the day.
The SMF has launched a new seminar, occurring twice a month, called “Séminaire Math & IA à midi” (or “séminaire MIAM”, and if you know French you may find this quite funny). The first talk was given by Hugo Duminil-Copin and can be watched on https://www.youtube.com/watch?v=iRjasLnigsU. (I learnt about this thanks to an article in the newspaper “le Monde” today: https://www.lemonde.fr/sciences/article/2026/10/10/nouveau-coup-de-massue-d-openai-sur-les-mathematiciens-ils-ont-annihile-mon-domaine-de-recherche_6790951_1650684.html.)
The first time I read the list of preprints, I missed another paper, arguing for a gap in the O(N) Heisenberg model, for N>3. Unlike the O(4) paper (which in Section 6 uses the Bethe Ansatz), this paper uses more standard statistical-mechanical methods.
A student and I are working on the O(3) problem from a different angle: we have to hurry up and finish!
I meant N>2 in the Open AI paper.
Here’s another thought that might be germane. When a theorem has many conceptually different proofs (e.g. the multitudinous analytic/algebraic/topological proofs of the fundamental theorem of algebra), would all those proofs turn out to be the “same” proof (modulo permutations of the logical steps) if fully formalized in the same foundational language (e.g. set theory)? In my experience, mathematicians have conflicting intuitions on this question.
No. For example, if two proofs are essentially the same, they should work in the same context. Analysts prove that modular curves have models over number fields by looking at Fourier expansions at the cusps; algebraists prove it by looking at the functor they represent. The second proof works for some Shimura curves that are compact, hence have no cusps. These proofs are really different.
FYI, Peter, pangram says that 100% of Jeremy Jr. Thomas’s comment is AI generated. I’m surprised you didn’t catch it since it has a pronounced “AI voice” style to it. It can also be abbreivated to 2 sentences, but instead wastes 4 paragraphs of our time… I honestly don’t know what’s the point of these comments or who makes them.
G.P.,
Yes, likely AI, deleted. Appropriate for the “slopocalypse” posting. One thing AI is really good at is this sort of thing, so we’re well on our way to a world where AI generated slop and scams will overwhelm human generated material just about everywhere.
OpenAI donated $15M to the Simons Fund, to help mathematics institutes working on AI-coordinated research: https://www.simonsfoundation.org/funding/targeted-grants-to-mathai-institutes-2027/
That will surely have some influence.
OpenAI, Anthropic and Simons Foundation each contributed $5 million to the grant.
The point about organisation is fair. If 722 papers are released at once with no index by field, status or verification level, the burden of sorting them falls on exactly the people whose time is scarcest. Even a simple table listing the problem, the claimed result, and whether a human expert has checked it would change how the release is received. It also makes it harder for the community to produce the kind of coordinated assessment you mention from the Navier-Stokes episode, because nobody can tell quickly which items deserve attention first.