Today's briefing →

OpenAI's math results and the backlash from mathematicians

15 updates Ongoing

Synthesis updated: Oct 31, 2026

In early September OpenAI said an internal model had solved the Navier-Stokes problem, one of the Millennium Prize Problems, and in October it released more than 700 AI-written math manuscripts at once. Mathematicians have objected to claims that OpenAI took credit for others' work, to unverified proofs announced in a rush, and to results released without papers. The Association for Human Mathematics, chaired by Terence Tao, called for a boycott. The Clay Mathematics Institute is reviewing the 166-page candidate proof, and Sam Altman says OpenAI will not stop releasing results.

The Where it stands and What to watch sections just below are our judgment, not a record of facts. The Full timeline at the bottom is the factual record, each entry checkable in that day's reporting.

Where it stands

The credit dispute

It began with an announcement in early September. OpenAI said a model it began training on August 28 solved the Navier-Stokes problem after about 10,000 agents ran for 88 hours. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge, who were working on the same problem, had published a proof of a simplified version a day earlier. Buckmaster says OpenAI rushed after learning of their progress and offered him alone near-unlimited compute and sole authorship, which he called a bribe.

The dispute spread to training data. Buckmaster raised the possibility that the Codex prompts he wrote while working on the problem were used in training, and German mathematician Andreas Thom said his own unpublished work may have been used too. After an investigation, OpenAI amended its announcement to say Buckmaster's prompts could not have influenced the system, but he says he doesn't believe it. Thom also got OpenAI to correct an August announcement that had ignored prior work.

Open letter and advisory group

25 Fields Medal winners signed an open letter criticizing proofs announced in a rush with no time for verification. OpenAI formed AGMAI, a nine-member advisory group of mathematicians, with Princeton's Institute for Advanced Study, and announced at the same time that its internal model had solved more than 100 open problems. The group advises only on how results are released and has no authority to slow research, and only one of its founding members had signed the letter. It said it was independent, but many took it for an OpenAI body.

A flood of manuscripts

Mathematicians objected again when OpenAI planned to post solutions on GitHub without papers. About 40 mathematicians OpenAI convened in August had asked for papers and say they were ignored. OpenAI then released 722 manuscripts from about 4,000 problems and said not all were verified. Only 300 top-line results, about 42 percent, were formalized in Lean, and OpenAI has retracted three papers.

The Association for Human Mathematics called the release a show of power that ignored AGMAI's recommendation and urged a boycott, and more than three dozen mathematicians said it could take years to make sense of the results. The Navier-Stokes proof is also contested. Noam Brown admitted nobody knows what the swarm of agents contributed, and several groups of mathematicians dispute the 166-page candidate proof.

OpenAI presses on

Sam Altman said he sympathizes with mathematicians but will not stop. He said OpenAI holds more discoveries than the hundreds of proofs already released, and that it has spent hundreds of millions of dollars on math, around $10 million of it on Navier-Stokes. It has not released details of a Yang-Mills theory result, so OpenAI is now also faulted for holding results back.

What to watch next

Full timeline (newest first)

Each entry below is a factual record you can verify in that day's reporting.

  1. Pressing on

    Sam Altman said he sympathizes with mathematicians but will not stop releasing results, adding that the Navier-Stokes solution alone cost around $10 million.

    Read it in that day's briefing → The Verge AI

  2. Proof disputed

    Noam Brown admitted nobody knows what the roughly 10,000 agents contributed, and several groups of mathematicians dispute the 166-page candidate proof.

    Read it in that day's briefing → The Decoder

  3. Verification

    Only 300 results from 719 manuscripts were formalized in Lean, and mathematicians said it could take years to make sense of them.

    Read it in that day's briefing → The Verge AI

  4. Boycott

    The Association for Human Mathematics, chaired by Terence Tao, called for a boycott, saying OpenAI ignored AGMAI's recommendation.

    Read it in that day's briefing → The Decoder

  5. Mass release

    OpenAI released 722 manuscripts from about 4,000 problems and said not all of them were verified.

    Read it in that day's briefing → AI타임스

  6. Show 10 earlier updates
    1. Release plan

      Mathematicians who had asked for papers objected again to OpenAI's plan to post hundreds of solutions on GitHub without them.

      Read it in that day's briefing → Wired AI

    2. Confusion

      The AGMAI advisory group said it was independent of OpenAI, but OpenAI's louder announcement left many thinking it was an OpenAI body.

      Read it in that day's briefing → The Verge AI

    3. 100 problems

      OpenAI said an internal model trained for about a month from August 28 solved more than 100 long-standing math problems.

      Read it in that day's briefing → The Decoder

    4. Advisory group

      OpenAI formed a math advisory group with Princeton's Institute for Advanced Study, but only one of its nine founding members had signed the open letter.

      Read it in that day's briefing → TechCrunch AI

    5. Corrections

      OpenAI amended its announcement to say Buckmaster's Codex prompts could not have influenced the system, and corrected an August announcement after Thom pointed out it ignored prior work.

      Read it in that day's briefing → Wired AI

    6. Offer rejected

      OpenAI offered Buckmaster alone near-unlimited compute and sole authorship, and he rejected it as a bribe.

      Read it in that day's briefing → The Verge AI

    7. Open letter

      25 Fields Medal winners criticized rushed, unverified proofs in an open letter, and OpenAI withdrew its sponsorship of a Caltech math event.

      Read it in that day's briefing → TechCrunch AI

    8. Second accusation

      German mathematician Andreas Thom said OpenAI's non-sofic groups result may have drawn on his unpublished work and criticized its evasive denial.

      Read it in that day's briefing → The Verge AI

    9. Prior proof

      Buckmaster and Alpöge had published a proof of a simplified version a day before OpenAI's announcement, and OpenAI denied training on Alpöge's unpublished work.

      Read it in that day's briefing → MIT Technology Review AI

    10. Origin

      OpenAI announced its internal model had solved the Navier-Stokes problem, and Tristan Buckmaster said he and Levent Alpöge had been working toward it first.

      Read it in that day's briefing → Wired AI

This page collects the articles about a single event from our daily briefings and lays them out in order. Each entry links back to that day's briefing, where you can reach the original reporting and the full context.