The Pulse
OpenAI’s More Than 100 Math Results Reopen a Credit Fight
OpenAI says a new internal model has resolved more than 100 long-standing math problems, as mathematicians question how the company releases results and recognizes prior work. The debate follows an August meeting and a separate dispute over
AI.info Team · Written with AI ·
At an August meeting with about 40 mathematicians, OpenAI discussed what might happen if AI outpaces human ability in the field. Attendees say the company hinted that its models had solved hundreds of long-standing problems, while asking for advice on how to share the work. According to WIRED, participants urged OpenAI to publish papers that mathematicians could study rather than announce results through a blog post or social media.
Now, as OpenAI prepares another release, some mathematicians say the company has not followed that advice. The dispute is about more than whether a model can produce a proof: researchers are also contesting how results are presented, how earlier human work is credited, and who gets time and access to understand them.
OpenAI says its model has cleared more than 100 problems
OpenAI says a new internal model has resolved more than 100 long-standing open problems across most areas of mathematics. In an announcement on September 21, the company said it began training the model on August 28 and that its progress had surprised mathematicians inside OpenAI. WIRED reported on October 6 that people familiar with the plans expected the company to post hundreds of results to GitHub that day; OpenAI spokesperson Lindsay McCallum told the publication that the company had not set a release time.
The coming batch follows OpenAI’s August 1 announcement of ten results, which it said came from an internal version of its Astra model. OpenAI described the problems as spanning areas including geometry, group theory, quantum complexity and extremal combinatorics. It said the model generated the mathematical arguments, while people prepared manuscripts and the company took responsibility for correctness.
That distinction has become part of the argument. OpenAI says attribution should reflect how a result was produced; the mathematicians WIRED interviewed say the way results are released can make it harder to assess significance, trace earlier ideas and build on the work.
The earlier Navier–Stokes dispute still hangs over the release
The latest conflict follows OpenAI’s September announcement of a proposed solution to the Navier–Stokes existence and smoothness problem, one of the Clay Mathematics Institute’s Millennium Prize Problems. The company said its internal system produced an analytical proof and a Lean formalization, and described its result as showing that fluid motion can develop a singularity in finite time. The company’s own account says it launched its effort after hearing rumors on September 1 that two Millennium problems had been solved.
NYU mathematician Tristan Buckmaster accused OpenAI of moving ahead of work he had done with Levent Alpöge, an Anthropic employee. Buckmaster and Alpöge had not published that work, and had used OpenAI tools during their research. WIRED reported that Buckmaster also objected to how credit was discussed; OpenAI researcher Sébastien Bubeck denied, in a public statement cited by WIRED, asking for Alpöge to be left off a paper. OpenAI says its investigation found that Buckmaster’s Codex prompts could not have influenced the internal model, including through training.
OpenAI’s account says the company’s model solved a different variant of the related Euler equations than Buckmaster and Alpöge did: its result concerned an unforced problem, while theirs involved external forcing. OpenAI also said it did not intend to claim the Millennium Prize. Those details do not settle the broader argument over priority, communication or credit, but they show why the claims need to be separated rather than treated as one result.
Mathematicians call for papers, attribution and time to understand
“Math by tweet and math by press release to me is not the way to nurture the ecosystem that created the fertile ground that they have trained on,” says Bryna Kra, a Northwestern University mathematician.
She said attendees at the August meeting wanted papers that would let researchers absorb and use the results. Kra told the publication that OpenAI’s latest approach appeared to ignore that input.
OpenAI formed an independent advisory group in September to advise on reviewing and communicating results, their significance, and professional standards. The group’s public September 29 recommendations call for literature reviews and citations to earlier work, scholarly repositories, documentation of how each result was generated, and support—including funding—for the work needed to develop human understanding.
The group also says it does not endorse frontier labs testing advanced mathematical problems on proprietary models that other researchers cannot access, and asks them to stop. That recommendation reflects a concern that labs can produce results beyond the reach of the wider research community, leaving mathematicians to verify and interpret claims without comparable tools.
OpenAI’s advisory group faces its first test
OpenAI says the advisory group will operate independently, receive no company payment and have the freedom to publish advice, including criticism. Its remit covers how the company reviews and communicates emerging results, but OpenAI’s announcement says the group will not advise on the pace of the company’s internal math research.
McCallum rejected the description of AI companies’ conduct as “mobster behavior,” a phrase Nestor Guillen, a visiting math professor at NYU, used in his conversation with WIRED. The spokesperson also told the magazine OpenAI was working with the math community on a responsible release. The company’s stated intent and the advisory group’s independence are now being measured against a practical question: whether the next batch of results arrives with the documentation and context researchers say they need.
OpenAI has said the model produced more than 100 results. As of WIRED’s October 6 report, the company had not announced a release time for the planned GitHub publication.