Research note

Research Notes: The OpenAI Math Drop (October 2026)

What happened

On October 6, 2026, OpenAI published a large batch of AI-generated solutions to open mathematics problems on GitHub, accompanied by a page titled "Sharing AI progress in mathematics." The release covered algebra, group theory, topology, mathematical logic, and theoretical computer science.

The scale figures conflict by source. Joshua Gans counts 722 papers and called it possibly "the biggest day of scientific advancement in history." The Guardian reported "over 370 results." WIRED reported more than 100 long-standing open problems among hundreds of results. These counts are not reconciled in the available coverage; treat all of them as claims, not as a settled inventory.

The drop did not arrive cold. In September 2026, OpenAI had claimed a solution to the Navier–Stokes Millennium Prize problem, which had already triggered dispute over credit, proprietary models, and process.

The AGMAI objection

In August 2026, OpenAI convened mathematicians, and an Advisory Group on Mathematics and AI (AGMAI) was formed at the Institute for Advanced Study. Per a contributor to Proofs and Prompts quoting OpenAI's own release page: AGMAI asks AI companies to stop evaluating their proprietary models on open research problems; OpenAI states it believes this evaluation is important. The October 6 drop proceeded regardless. Critics (among them Henry Wilton and Johannes Schmitt) describe prior community input as partially ignored: no human authors, no journal process, no disclosed problem-selection criteria or failure rates.

Verification status

  • Some results reportedly carry partial Lean certificates; many do not.
  • Wilton, after a cursory glance: the write-ups "vary wildly in quality"; checking all of them will take months.
  • Tristan Humbert, on the Katok's entropy conjecture paper: "mostly unreadable slop."
  • At least one contributor demonstrated that an OpenAI example (a 116-generator Artin group) was improvable to 12 generators in ten minutes with another tool — offered as evidence that "many papers were never reviewed by a human" and that the drop prioritized quantity.
  • One reaction of note on process: AGMAI's existence raised expectations of carefully explained companion papers; what arrived was "a massive GitHub drop." ("What did we even need AGMAI for?")

The reactions

Collected at Proofs and Prompts, "100+ reactions to 100+ solutions" (October 8, 2026; updated October 9). Representative, verified quotes:

  • Matt Zaremsky (Boone–Higman conjecture, four-year research program): "it feels like we've been working on an archeological dig for 4 years, gradually discovering more and more of a really cool-looking dinosaur skeleton, and then a trillion-dollar company showed up and just blasted the whole thing with TNT, handed us the whole skeleton, and walked away." Also: "it's believable we could have gotten there in another year or two."
  • Tristan Humbert (Katok's entropy conjecture, his PhD subject): sadness at his favorite problem being "killed"; then panic — he is applying for postdocs and spent the morning rewriting his research plan and emailing advisors; then anger at the paper's quality. His conclusion: OpenAI's standards "harming mathematical research maybe in an irreversible way," with the community "accepting to work for free" to validate the proofs.
  • Tasmin Chu: "I realize now that what I wanted more than to know that it is true was the time and space to think about it for the next few years. I feel a profound sense of grief for the working conditions I believed I would have."
  • Henry Wilton: chief feelings "disorientation and bewilderment"; the greatest worry is the pipeline — "the necessary incentives may no longer exist for bright young people to invest that effort."
  • One contributor reports friends and colleagues "experiencing severe anxiety, depression and loss of purpose," adding: "What they need most at this moment is not more powerful AI, but the time to process the transformation of their discipline."
  • The optimistic register exists too: Gans frames a division of labor (AI generates, humans select and sign off); Tobias Osborne and Martin Bridson argue the field is entering a bigger era, not ending.

Two first-person video reactions, both published October 9, 2026: Sophie Maclean, "I am sad about AI in maths..." (process grief, first person), and the developer channel Awesome, "The AI skeptics are changing sides..." (the parallel conversion wave in software).

What it does not support

  • A settled count. 722, 370+, and "hundreds, including 100+ long-standing" are unreconciled claims.
  • Correctness. "Solved" means "a proof was published on GitHub." Community verification is months away at best; at least some write-ups are, by expert judgment, unreadable.
  • "Math is dead." WIRED reports some OpenAI employees believe it; OpenAI denies it; much of the community argues transformation, not death. All three positions are advocacy.
  • AGMAI endorsement. The documented record runs the other way: the advisory group's central request was declined.

How this corpus uses it

Linked at first mention from the post Who Is AI Coming for First? as the event of record. The post narrates; this note holds the facts.

Sources

↓ Markdown