DotScott - Joining the dots on your AI journey
AI services
Our collection of services designed for small and medium businesses to embrace AI skills.
AI Questionnaire
.training

Learn the essentials at your own pace, in plain English.

.consultancy

Find where AI can save you time and money in your business.

.workshops

Hands-on sessions for you and your team.

your AI system

A practical, documented set of tools and processes you can keep using.

Website Design & SEO services
A complete range of website design and SEO services, available individually or combined into projects.
Website Design

OpenAI Just Dropped 719 Maths Papers on GitHub. The Mathematicians Are the Canary.

Skip to main content
< All Topics
Posted

OpenAI has dumped around 719 maths manuscripts onto GitHub, claiming answers to problems people have spent years, sometimes decades, trying to solve. Some of the mathematicians closest to those problems are, understandably, devastated.

If your first reaction is “brilliant, more maths, that’s progress”, imagine opening your laptop and finding the work you hoped would define your career already sitting on GitHub. You might be delighted by the discovery. You might also be wondering what you’re doing next Monday morning.

I think the mathematicians are the canary for the rest of us.

Writers and artists have already felt it and programmers are living through it right now. Next comes scientific research, then the professions: legal, medical, financial, accountancy. If you sell your expertise in any way, shape or form, watch how this plays out.

What OpenAI actually released

A month ago OpenAI announced an independent advisory group of mathematicians to help it “responsibly share advances in AI and mathematics”, including how to assess and communicate new results. Sounds great. The group published its first recommendations on 29 September. A week later, on 6 October, OpenAI put 722 manuscripts, organised into 372 “families”, on GitHub, with a Lean library so some of the proofs can be machine-checked.

So: we want to learn how to communicate better with you. Oh, and by the way, we solved all this maths you’ve been struggling with. Suckers!

Back in August I covered Astra’s ten maths advances, and the argument then was whether ten results held up. This is 719. They won’t all be unique, they won’t all be correct and they won’t all be useful. But the machine is now producing proposed answers faster than the people in the field can absorb them, and every one of those papers is work the profession has to go through.

Here is how the last three months have unfolded:

Date What happened
1 August 2026 OpenAI credits ten maths and computer science results to an internal model called Astra, each checked in the Lean proof assistant.
8 September OpenAI claims a solution to the Navier-Stokes problem, one of the seven Millennium Prize problems, produced by a swarm of around 10,000 AI agents at a cost of millions of dollars in computing.
11 September More than two dozen Fields Medallists sign a declaration titled “A Severe Misalignment of AI in Mathematics”, objecting to AI labs treating unsolved problems as benchmarks.
21 September OpenAI announces the Advisory Group on Mathematics and Artificial Intelligence, nine mathematicians hosted at the Institute for Advanced Study in Princeton. The same post says the model has resolved “more than 100” further open problems.
29 September The advisory group publishes its recommendations: release the model, the exact prompt and the compute time behind every result.
6 October 722 manuscripts in 372 families go onto GitHub. OpenAI publishes average compute time and ten abbreviated prompts, not the model and not the prompts for the other 362 families.
7 October Three manuscripts are withdrawn after a sign error is found in one argument that two others depended on. Fourteen more are revised. The count settles at 719.

The claims cover algebra, number theory, theoretical computer science, logic, probability and topology. Among them: a partial result towards the Riemann hypothesis, a proof of the Hodge conjecture for one class of objects, a proof of the unique games conjecture, a faster matrix multiplication algorithm, and a negative answer to Hilbert’s tenth problem over the rationals. If those names mean nothing to you, the short version is that several of them are problems that have had careers built around them.

What “solved” actually means here

This is where it gets interesting, and where the parallel with every other profession starts.

Only about 42% of the headline results have been formalised in Lean. Lean is software that checks whether a formal proof really does follow from its formal statement. It is very good at that. What it cannot do is confirm that the formal statement matches what the paper claims in English. OpenAI’s own README says some of the unformalised results “could have issues”, and its researchers admitted they hadn’t read all the papers before release.

Three UK mathematicians, Alexander Bastounis at King’s College London and Fabian Circelli and Anders Hansen at Cambridge, went through the Navier-Stokes proof and its Lean code and found two places where they diverge. In one, the code proves a weaker estimate than the paper states. They aren’t saying the proof is wrong. They are saying a machine check is not a substitute for a human reading it, which is a point worth holding on to.

The model is not available. OpenAI says the model tried roughly 4,000 problems and that the average result took about three hours of computing. Nobody outside OpenAI can re-run any of it. Andrew Sutherland at MIT put it plainly: claims of one-shot solutions should be treated as unverified until the model is released and results can be replicated. “We should ask for receipts.”

The papers are hard to read. Dana Moshkovitz, one of the experts on the unique games conjecture, said the preprint on her own problem was so badly written it was hard to get through without AI help. Peter Sarnak at Princeton says he won’t call any of it a “theorem” until humans have verified and understood it.

So the honest position is: some of these results are probably right and important, some are probably right and trivial, some are probably wrong, and the mathematical community has been handed the job of sorting out which is which. That job will take months, possibly years, and nobody asked for it.

How the mathematicians are taking it

Not well, on the whole.

Hugo Duminil-Copin, who won the Fields Medal in 2022, told a French Mathematical Society call on 8 October that the release included claimed solutions to 16 questions he had spent years on, and that nearly every open problem he had flagged in his lectures, papers and grant applications had been resolved in one go. He said he cried so hard one night that it woke his daughter, and that OpenAI had “annihilated” his field of research. Six weeks earlier he had written an essay warning that AI proofs could “nuke the mathematical landscape”. He was right, and it didn’t help.

The Association for Human Mathematics issued a statement saying it rejects OpenAI’s claim that the release advances the subject, and that putting out over 700 files at once “is not a demonstration of scholarship, but a demonstration of power”. It called for a boycott of OpenAI. Terence Tao, who had earlier described the coming flood of results as “toxic waste”, reshared it.

Even OpenAI’s own advisory group, which includes Timothy Gowers and Martin Hairer, said of the practice of testing proprietary models on advanced open problems: “we do not endorse this practice.” OpenAI’s response is that it takes the guidelines seriously, isn’t bound by them, and will not slow down because these problems are an essential test of whether its AI is getting smarter.

Gowers, who declined to sign the Fields Medallists’ letter, has a more measured worry, and I think it is the one that matters most. A flood of new results isn’t a disaster in itself if it expands what we know. The danger is to the pipeline: if the big open problems get picked off by a machine, why would a bright 22-year-old spend five years on a PhD trying to crack one?

Why the mathematicians are the canary

Mathematics was supposed to be one of the last things to go. It’s abstract, it requires long chains of careful reasoning, and it has a brutal standard of truth: a proof is either right or it isn’t. If anything was going to stay human for a while, this was the candidate.

Instead it has become the clearest example yet of the pattern that has already hit everyone else:

  • Writers and artists found their output could be imitated at volume, cheaply, by people with no training in the craft.
  • Programmers are watching tools write most of the code, and finding their job has quietly become reviewing it.
  • Mathematicians have now had their problems answered faster than they can check the answers.

What the three have in common is not that the machine is better than the best people. It mostly isn’t, yet. What they have in common is that the machine produces proposed work faster than the profession can evaluate it, and the burden of evaluation lands on the people whose jobs are being disrupted. OpenAI doesn’t need to prove all 719 papers are right. It only needs enough of them to be right that ignoring the rest becomes impossible.

Scientific research is next, and the mechanics are identical: generate thousands of hypotheses, experiments and papers, let the humans sort them. Then the professions. A legal argument, a medical differential, a tax position and a set of accounts are all things that can be proposed in volume and then checked. The checking is the hard bit, and it is now the job.

What this means if you sell your expertise

I’m not going to pretend I know how this ends for mathematics. But the pattern is now clear enough to act on, whatever profession you’re in.

  1. The value is moving from producing the answer to standing behind it. Nobody is going to pay a solicitor, an accountant or a consultant for a first draft any more. They’ll pay for the judgement about which draft is right, and for someone to be accountable when it isn’t. If your pricing is still based on the time it takes to produce things, that is the first thing to look at.
  2. Verification is a skill, and it’s a different one. The mathematicians who are coping best are the ones treating the dump as material to be audited, not as a verdict. Reading AI output critically, quickly and with a nose for where it’s likely to be wrong is now a core professional skill in every field, and it is not the same skill as doing the work yourself.
  3. Volume is the weapon. Note that OpenAI did not release one perfect paper. It released 719 imperfect ones. Expect the same in your market: not one competitor that does your job better, but a flood of cheap, roughly adequate output that changes what clients think the work is worth. Your answer to that cannot be “ours is better”. It has to be visible, specific and provable.
  4. Insist on receipts. Sutherland’s line applies to every AI tool you’re sold. Which model, what prompt, what did it cost to run, and can I reproduce it? If a supplier can’t answer, you are taking on the verification cost and they are keeping the margin.
  5. Protect your pipeline. Gowers’s worry about PhD students applies to trainees, juniors and apprentices in every firm. If AI does the grunt work that used to train people, you have to find another way to train them, or in ten years you’ll have nobody who can check the machine.

The mathematicians didn’t get a say in when this happened to them. You still do, a bit. Use it.

Need a hand?

If you’d like to think through what this means for your business, or how to put AI to work without handing the checking to your clients, get in touch.

Sources

Posted
Table of Contents