722 papers in one night: is OpenAI harming mathematics?

On the morning of 6 October 2026, a repository appeared on GitHub. It belonged to OpenAI, and it held 722 full mathematical papers, organized into 372 families. Each one claimed to resolve, or at least to advance substantially, an open problem in mathematics or theoretical computer science. The whole collection had been written by a single unreleased internal model in the days before. Someone who wanted to put the same work on arXiv would need more than three hundred accounts, because arXiv allows one person two uploads a month. So OpenAI published it on GitHub instead.

The scale is what makes people stop. The release spans seventeen directions, from number theory and algebraic geometry to theoretical computer science and mathematical physics. OpenAI says the results came from testing more than four thousand research problems, and that each result cost on average about three hours of compute, roughly what a ChatGPT Pro user spends in one long session. A handful of the 372 families target problems that have sat open for decades, and in some cases for a century.

Take the Kakeya problem. In its three-dimensional maximal version and its four-dimensional Hausdorff dimension version, Wang Hong and Joshua Zahl spent years building the proof that reached the Fields Medal conversation. OpenAI claims its model produced related results on the same line of questions in a few hours. Nobody is saying the two are directly comparable; a proof written quickly still has to be read, checked, and believed. But the difference in time is the thing the field cannot stop talking about.

The release also carries work touching the quasi-Riemann hypothesis, the Unique Games conjecture, special cases of the Hodge conjecture, and one form of Hilbert’s tenth problem. None of these is a small claim. A paper that settles any one of them would normally make a career. Here they arrive in bulk, in a folder on the internet, with a protocol for revisions and citations but no journal and no referees.

Scott Aaronson called it one of the biggest days in mathematical history. That may be hyperbole, but it is not the empty sort. The practical problem is verification. A human mathematician reads a paper, checks the logic, tries to break it. With 722 papers from a machine, that workflow does not scale. The community’s first reaction was not excitement; it was worry about who will read all of this. One well-known response compared it to a code review problem: code you did not write, arriving faster than anyone can review it, some of it probably right, some of it probably subtly wrong, and no easy way to tell which is which.

That is the honest core of it. Mathematics was never really about who writes the first line of a proof. It is about who can stand behind it, who can check it, and who can build the next thing on top of it without it collapsing. A machine that produces candidate proofs in bulk changes the bottleneck. The bottleneck stops being production and becomes curation and verification. That is a genuine shift, and it is worth being clear-eyed about it.

A gloomy column about the end of the profession would be easy to write, and so would a triumphant one about machines outproducing Fields medallists. Both are too simple. What the release actually does is force a distinction that mathematicians have quietly relied on for a long time: the difference between finding an argument and understanding one. The first part, the search over possible moves, is what a model trained on every proof ever published might do at speed. The second part, knowing which steps matter and why, and knowing when a gap is real rather than cosmetic, is still a human activity. A collection like this does not replace that activity; it floods it with material.

There are also real worries worth naming. One is provenance: 722 papers with no named authors, no institution, no record of how each one was produced is a difficult thing to cite, to teach from, or to reward. Another is reliability. OpenAI itself has been careful, on other occasions, to point out that its outputs can be confidently wrong. A claimed proof of a hard conjecture is exactly the kind of thing a language model can hallucinate with total conviction. The repository is a starting point, not an answer key.

And yet the interesting thing is how quickly the research community moved from “is this real?” to “what is inside it?”. Within weeks, papers were appearing that treated the release as raw material. One preprint on arXiv, titled What the Machine Invented, proposed to mine the 722 manuscripts for new lemmas, constructions, inequalities and proof mechanisms, not to adjudicate the claims but to see whether the machine had found moves that humans had not. That is a remarkable reaction, and, to my mind, a hopeful one. It treats the output for what it is: a very large pile of candidate mathematics, some of which will turn out to be valuable.

The event also raises a question about how mathematics is shared at all. Universities have journals, arXiv, and careers built on authorship. A company that can generate a thousand papers a week does not fit that infrastructure. OpenAI published on GitHub and said it was consulting an advisory group about best practice for revisions and citations. The community is right to push for standards. If machine-generated mathematics is going to be part of the discipline, it needs agreed rules for what counts as a contribution, how errors get reported and fixed, and who is accountable when a proof turns out to be wrong.

I grew up in a place where the practical value of mathematics was always obvious: it was the route out, the subject that rewarded effort no matter your background. Nothing about this release changes that. If anything, it makes the skill of judging an argument more valuable, not less. The person who can look at a claimed proof and tell whether it is worth believing will be the one who is scarce.

So I do not think this is the end of mathematics as a human activity. I think it is the start of a strange new division of labour, where machines do the searching and humans do the deciding. The day the 722 papers landed was genuinely historic, in the narrow sense that a lot of things changed at once. What it did not change is the oldest rule of the subject: a result is only as good as the person who can defend it. That part still falls to us.