OpenAI’s 372 Math Results Force Departments to Rethink Hiring

On October 6, OpenAI released 372 families of AI math results, then pulled three papers the next day.

The remaining 719 manuscripts sit on GitHub. Math departments are now arguing about who still gets trained to ask the next question.

An 88-Hour Swarm and a Forced Blowup

On September 8, OpenAI said an internal model, still being trained and stronger than GPT-6 Astra, had produced a finite-time singularity with smooth forcing for the three-dimensional Navier-Stokes equations. The write-up came with a Lean proof. The company said it does not intend to claim the Clay Mathematics Institute’s $1 million Millennium Prize.

Claude-Louis Navier and George Gabriel Stokes wrote the equations in the nineteenth century to treat a fluid as a continuum. In 1934, Jean Leray showed that solutions exist in a generalized sense. Whether they stay smooth has stood as an open question for roughly 90 years. In 2000 the Clay Institute named it one of seven Millennium Prize Problems.

OpenAI’s construction is a vortex that spirals inward and stretches like spaghetti. Speed grows without bound while energy stays finite, and a smooth external force remains in the picture. That matches statements C and D in Clay’s official formulation, the breakdown route with forcing. Statements A and B, the unforced smoothness question most people mean by Navier-Stokes, are still open.

THE SEPTEMBER RUN

• The swarm: The Navier-Stokes group used on the order of 10,000 concurrent agents, after about 100 agents spent about 50 hours on a related Euler blowup.
• The clock: Training of the internal model began on August 28. Agents launched on September 1 and reached a resolution on Saturday, September 5, about 88 hours later, with 17 more hours of Lean work via GPT-6 Astra.
• The traffic: The Navier-Stokes agents sent 2.7 million messages and about 130 billion output tokens. Across every problem in the run, the totals were 4.9 million messages and about 300 billion tokens.
• The bill: OpenAI put a customer rerun of the Navier-Stokes effort at about $15 million.

The company said it started the sweep on September 1 after rumors that two Millennium problems had fallen. It later named those rumors as work by Tristan Buckmaster, a mathematician at New York University, and Levent Alpöge, who works at Anthropic. OpenAI cedes priority on forced Euler to that pair, claims unforced Euler and forced Navier-Stokes for its own system, and says no user inputs after July 3, including Buckmaster’s Codex prompts, could have shaped the model.

Javier Gómez-Serrano, a mathematician at Brown University, told a hastily called Harvard audience on September 11 that the write-up was incomprehensible even as he called a theorem of this class a genuine accomplishment. Dan Roberts, a research lead at OpenAI, described the human role as a bumblebee, carrying useful fragments from one agent group to another.

Why the Million-Dollar Prize Is Still Untouched

Clay announced the seven problems in Paris on May 24, 2000, with $1 million on each. Grigori Perelman’s proof of the Poincaré conjecture is the only prize it has paid. Navier-Stokes is listed as Active, not solved. OpenAI posted a manuscript and a Lean file, not a journal paper, and said it would not seek the money.

WHAT CLAY REQUIRES BEFORE IT EVEN LOOKS

• A qualifying outlet: The institute does not take direct submissions. A proposed solution must appear in a refereed mathematics journal of worldwide repute, or in another outlet the board has approved.
• Two years: The scientific board will not consider a solution until at least two years after publication, under rules adopted on September 26, 2018.
• General acceptance: The proof must also win general acceptance in the global mathematics community, as Clay judges it.

Those three tests run in order. Even a clean Lean file does not start the clock. Until a qualifying journal publishes the argument, there is no two-year wait to finish, and there is no prize committee to name. The unforced problem remains untouched, so the version of Navier-Stokes that lives in popular memory is still open in Clay’s own A/B sense.

Priority fights do not help. Buckmaster has alleged pressure and leakage around the weekend of the swarm. OpenAI denies both. The institute’s process was built for a single human proof that the field could sit with. It was not built for a 88-hour agent run and a GitHub folder.

372 Families Landed in a Single GitHub Drop

The Navier-Stokes post was the loud one. The October 6 drop was the one that hit hiring. OpenAI published a broad set of new results from the same internal frontier model, after talking with the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. The average accepted result, the company said, used three hours of ChatGPT Pro thinking. The model was posed about 4,000 problems. Ten abridged reasoning summaries came with the set, including claimed work on the irrationality exponent of π, the Mahler conjectures, and the three-dimensional relativistic Vlasov-Maxwell system.

THE GITHUB CATALOGUE

The public repo now lists 719 manuscripts in 372 families, grouped by discipline, under an Apache 2.0 license. OpenAI warns that some unformalized results could have issues. A sign error on October 7 broke one argument and two papers that depended on it, which is why three files came down and fourteen others were repaired. Exceptions to the fixed three-hour procedure include a zero-free region for the Riemann zeta function and a Hodge claim for CM abelian varieties.

On October 15, OpenAI said it would give $5 million to a Simons Foundation program so mathematicians can examine and build on AI-written results, and would give Simons-supported researchers access to models and compute. That money buys workshops. It does not turn 719 PDFs into theses, referee reports, or a hiring standard.

Cornell’s Town Hall Was About the Next Hire

Steven Strogatz, an applied mathematician at Cornell University, has been walking around his house in Ithaca since May, when AI systems began clearing problems his students had hoped to live with. After the October 6 dump, the department called faculty, staff, and students into a town hall. The agenda was not a celebration. It was how to judge a job candidate, and what to expect of a Ph.D. student, once a model can finish a dissertation problem and draft the thesis.

Alex Townsend, a younger colleague in numerical linear algebra, had already gathered about 200 open problems in his area, then moved toward releasing about 600 long-standing problems across applied math. He wants friends and students to work them, with AI if needed, before a lab stamps them. Some colleagues found that plan exhilarating. Others called him a collaborator in wrecking the training path, because famous unsolved problems were how young people learned to be mathematicians.

Justin Moore, who works in logic and set theory, had spent 19 years on a question about non-amenable groups and Thompson’s F. In 2010 he bought a bottle of XO cognac in Luminy, outside Marseille, to open when it fell. The machine got there first. He had not read the proof. He said he was genuinely busy, and that it does not help that AI does not write mathematics particularly well. He still plans to open the bottle once peers he trusts agree the argument is correct and they collectively understand it.

I have a feeling that the cognac tasting will feel more like a wake than a celebration.

Justin Moore, Cornell mathematician, in an email

Aria Beaupré, a graduate student at that meeting, asked what the point of the subject now is. Curtis McMullen, a Fields Medalist at Harvard, had already put the pure-math side in an email to Strogatz: mathematics is not radiology, and human understanding is what it is about. Applied people can use a tool they do not fully explain. Pure math treats the explanation as the work. When the explanation arrives as an unread file, the apprenticeship has nothing to hand to the next person.

Mathematicians Pledge to Stop Feeding the Labs

The Association for Human Mathematics answered the dump on October 7. “Mathematicians did not ask for this work to be done,” it said, and “releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power.” It urged colleagues to stop working with OpenAI. The association now lists 1,008 members. Of those, 654 sit in an AI-free caucus that will not use models to find proofs, hunt examples, or pose research questions. Others will refuse lab jobs and AI-written papers but still touch chatbots at the margin.

On September 11, 25 Fields Medalists issued a letter on a “severe misalignment” between AI labs and mathematics. Solutions, they wrote, are announced in a rush, with no time for a proper writeup, isolation of new methods, or citation of earlier work, which raises attribution questions and threatens the human transmission chain. Fields Medalist Jacob Tsimerman took the other door and joined OpenAI this year, arguing that models will soon do the job better and faster.

WHERE EXPERTS DISAGREE

• The boycott camp: The Association for Human Mathematics treats the GitHub drop as a show of force and wants mathematicians off OpenAI’s payroll and, for the stricter caucus, off the models themselves.
• The adaptation camp: Townsend and Tsimerman treat resistance as futile and want the old problems cleared so people can spend their time on questions, applications, and interpretation.
• The Fields letter: Twenty-five medalists, writing before the October dump, warned that racing famous problems as benchmarks burns the students and ideas the profession actually runs on.

Outsiders hear a boycott and ask why anyone would slow discovery so humans can keep the puzzles. That is the wrong objection. The sharper one is that a lab can mint theorems faster than a department can turn them into people. Hide work in progress, and you starve collaboration. Publish 722 files in a night, and you starve the slow reading that used to be a career.

The Lean Check Leaves the Argument Unread

Lean is the proof assistant almost every lab now points to when a human says “I cannot read this.” A kernel checks each step. That is a real gain, and it is also how a 88-hour swarm can claim a Millennium-class result before a single referee has slept. Thomas Hales, in a guest post on Terence Tao’s blog, walked through soundness bugs found in Lean in July and August, a stretch he calls the Summer of Soundness Bugs, including an illicit Collatz disproof. The bugs were patched. Mathlib was re-checked. Hales still wants a human audit that the formal statement is the statement anyone meant.

Tao told a Caltech audience on October 9, in a talk he billed as Math 2.0, that further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics. He has floated a journal that would require a Lean check on every proof, under the motto of understanding over priority. Anthropic, on September 4, said Claude had formalized Fermat’s Last Theorem in 11 days and about 13 million lines of Lean. Line count is not comprehension. About 42 percent of OpenAI’s top-line dump results have Lean proofs at all. The rest are PDFs that still need people.

Machine-checked correctness and human understanding have come apart. Generating an answer is now cheaper than standing behind it. A department that once trained students by handing them a beacon problem now has to train them to interrogate a file, to decide which of 372 families matter, and to keep asking questions a swarm has not been pointed at. Moore still has not opened the bottle. Clay’s two-year clock has not started either, because the Navier-Stokes write-up is not in a qualifying journal.