OpenAI Released 719 Math Manuscripts in a Single Day. A Group of 800+ Mathematicians Calls That the Problem.

On October 6, OpenAI published a repository of 719 manuscripts claiming full or partial solutions to more than 370 open problems in mathematics, including measurable progress -- not a full solution -- on three Millennium Prize problems and a claimed proof of the Unique Games Conjecture. The work came out of an unreleased internal model, run at an average of roughly three hours of compute per solution, and ten of the write-ups include the model's own chain of thought. OpenAI described its own goal as wanting the work "to push the frontier of human knowledge and enable further progress in mathematics," adding that it is "committed to further improving the quality of the papers." 3 of the 719 manuscripts were withdrawn within 24 hours, after reviewers found errors -- and TechCrunch reported that 42% of the release had not undergone formal verification at all, meaning the remaining 58% is formalized in a proof assistant a computer actually checks line by line.

The next day, the Association for Human Mathematics -- which says it counts more than 800 members, including Fields medalist Peter Scholze -- published a statement that Terence Tao hosted as a guest post on his own blog, attaching only a brief editor's note. It did not dispute that the proofs check. It objected to how 719 of them arrived at once.

That objection has a specific origin. OpenAI says it consulted the Advisory Group on Mathematics and Artificial Intelligence, a nine-member body hosted by the Institute for Advanced Study, specifically to avoid repeating the backlash from an earlier, smaller math controversy. TechCrunch reported that AGMAI's *first* recommendation to OpenAI was not about formatting or pacing -- it was to stop testing advanced open problems on proprietary models at all. The October 6 release does not appear to follow that advice. AGMAI itself has stayed diplomatic, calling its talks with OpenAI "constructive" and a first step, and explicitly leaving it to the wider mathematical community to judge whether OpenAI's compliance with its other principles was adequate.

Reaction among mathematicians themselves split hard. Dan Litt, at the University of Toronto, called the release "great for mathematics" and welcomed the GitHub format for making a previously closed body of work newly explorable. Melanie Wood, a Harvard mathematics professor, told TechCrunch the opposite concern in one line: "there is not human understanding of them at the point of release." Tristan Buckmaster -- who separately disputed credit on an earlier OpenAI Navier-Stokes claim -- said OpenAI had not done adequate due diligence to rule out that the model had plagiarized other mathematicians' unpublished work.

Tao's own objection, laid out separately at a Caltech talk on October 9, runs deeper than any single release. He argues mathematics has shifted from proof scarcity to proof abundance -- a world where a correct proof used to be "difficult and rare, and often required significant human expertise to obtain," and now often isn't. Treating problem-solving speed as the measure of progress, he says, "was largely accurate in the era of proof scarcity, but has become misaligned in the era of proof abundance," and "further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics." His proposed alternative, which he calls Math 2.0, would point the same AI tools at open *exposition* problems, ablation studies testing whether a theorem still holds once a key input is removed, and shared infrastructure like the formal library Mathlib -- using the technology to deepen understanding of results, not just to manufacture more of them.

Further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics.

One detail of the coverage is worth checking rather than repeating. Several outlets describe Tao as AHM's chair; the statement itself does not.

Scale is the other way to read this, separate from process. Scott Aaronson, citing people briefed on the work, put the testing behind the release at roughly 8,000 problems attempted, with a success rate near 5% and an average of about three hours of GPT-Pro-level compute per attempt -- a brute-force sweep across a huge chunk of the field's open questions, rather than hand-picked targets. Aaronson drew a direct contrast with how Anthropic has approached the same kind of work: partnering with two named mathematicians, compensating them, and producing a single human-readable proof of a pair of decades-old conjectures, rather than a 719-file batch. Neither approach is inherently illegitimate -- a sweep finds things a hand-picked search wouldn't, and a hand-built proof is easier for the field to actually absorb -- but they produce very different numbers of files to argue about afterward.

What's not in dispute is the sequence, or the stakes. OpenAI built an advisory process specifically to avoid a repeat of an earlier math-credit fight, consulted it, and then released a volume of work that group's own first recommendation argued against. Whether that reads as OpenAI overruling its own safeguard or as a field unable to agree among itself on what counts as consent is, for now, a matter of which mathematician you ask -- and AHM's answer is that its members should stop finding out the hard way, one bulk release at a time.