By This Hour AI Desk
A reported breakthrough attempt on one of mathematics’ most prominent unsolved questions has turned into a dispute over provenance, access and the role of artificial intelligence in research. Tristan Buckmaster, an NYU mathematics professor, has alleged that OpenAI received information about unpublished work he was conducting with mathematician Levent Alpöge, then pursued a similar route with far greater computing resources.
OpenAI has rejected the central allegation. Its reported position is that neither its researchers nor its AI agents saw the pair’s work before it was publicly released, and that no particular user data was accessed to solve the problem. Yet the company reportedly also said it could not completely exclude the possibility that de-identified data derived from use of its products had contributed to improvements in its models. The gap between those positions is at the heart of the disagreement: direct access to unpublished research is one question; the possible influence of data used in model improvement is another.
The argument concerns the Navier–Stokes existence and smoothness problem, a question about equations used to describe fluid motion. The equations have broad use in fluid mechanics, but their deeper mathematical behavior has resisted a definitive resolution. The problem is among the Millennium Prize problems, each associated with a $1 million award for a solution that meets the relevant standard. A claimed proof, or meaningful partial progress, therefore carries unusually high scientific and professional stakes.
A preliminary result and a competing claim
Buckmaster announced three proofs related to the Navier–Stokes problem in work with Alpöge. The supplied account characterizes their findings as preliminary progress toward the central question rather than a settled solution of it. Their work reportedly used both OpenAI’s Codex and Anthropic’s Claude. Alpöge works at Anthropic, but the account says the research was not conducted for that company.
Shortly afterward, OpenAI reportedly published what it described as a full proof of the Navier–Stokes existence and smoothness problem. The company attributed that result to an unreleased next-generation model and said the model had addressed several unsolved questions over the preceding week. The distinction is consequential. A set of related proofs and a full proof of the central problem are not interchangeable claims, and neither the supplied material nor the reported public exchange establishes that either body of work has undergone independent mathematical validation.
That unresolved validation issue limits every broader conclusion. Mathematics ultimately depends not on the prestige of an institution, the scale of a computation or the confidence of an announcement, but on whether a proof can be examined and accepted on its merits. The source material describes competing accounts of research activity and authorship concerns; it does not provide a basis to determine whether the claimed full proof is correct, whether the preliminary findings are correct, or how the results compare technically.
The reported scale of OpenAI’s effort has added to the sensitivity. The account says the work used 300 billion output tokens over a week, a quantity it valued at $22.5 million using the compute rates it cited. Those figures are reported estimates rather than independently established costs. Still, they frame Buckmaster’s concern: a well-funded AI lab may be able to mobilize computational capacity at a level unavailable to individual academic researchers, particularly when pursuing an avenue already believed to be promising.
The disputed September timeline
Buckmaster’s allegation is not simply that OpenAI worked on the same famous problem. Navier–Stokes is widely studied, and simultaneous attention to a well-known open question would not by itself show that one team relied on another’s ideas. His stated concern, as summarized in the supplied account, is that the pair had taken a relatively uncommon route and that information about their progress reached OpenAI before their research became public.
He further alleged that OpenAI then adopted that route, using substantial compute to move more quickly toward a formal proof. The account portrays the approach as one selected by Buckmaster and Alpöge in private and not as an obvious route a researcher or model would necessarily reach merely from being given the problem. That is an allegation about both timing and intellectual lineage. It cannot be resolved just by observing that the later work addressed the same question.
OpenAI’s reported account aligns with part of the timing while contesting the inference. The company said its latest effort began on September 1 after rumors circulated that two Millennium Prize problems had been solved. It also reportedly confirmed discussions with Buckmaster and Alpöge. But OpenAI denied that its researchers or agents had encountered the pair’s work by any means before its public release. Its reported response also maintained that its proofs differed substantially from the other work, including in the precise results reached in related Euler-equation work.
Those accounts leave important questions unanswered. The supplied record does not establish what information, if any, was conveyed to OpenAI before publication; who may have received it; what was understood from it; or whether it affected the choice of research direction. Nor does it show the prompts, internal research records, model outputs or technical materials that would be needed to test the competing narratives. The reported common starting date is not proof of copying, but it does not independently disprove Buckmaster’s concern either.
Codex use raises a separate data question
A second strand of the dispute arises from the pair’s use of Codex while developing their research. Buckmaster reportedly worried that material from those interactions might somehow have informed OpenAI’s own effort. This is a more complicated claim than an allegation that a researcher opened or read a private conversation. It concerns whether use of an AI product can leave traces in broader model-improvement processes and whether a later system might reproduce useful elements when confronted with a similar task.
OpenAI’s reported answer draws a line between specific customer material and de-identified usage data. The company said no specific user data was accessed for the purpose of solving the problem and said direct reproduction of the pair’s work was unlikely. At the same time, it reportedly did not rule out that de-identified data derived from product use might have helped improve the models. That qualification does not establish that Buckmaster and Alpöge’s interactions had any effect. It identifies a possibility the company said it could not entirely eliminate.
For researchers using AI systems, that distinction matters. A denial of direct inspection can coexist with uncertainty about aggregate training or improvement pathways. Conversely, a general possibility that de-identified data may help a model does not demonstrate that any particular idea was retained, surfaced or material to a particular result. The supplied reporting does not specify the researchers’ product settings, the handling of their interactions, or any technical analysis comparing their work with OpenAI’s proof. Without those details, the causal claim remains unproven.
The dispute also puts the status of AI assistance under pressure. Buckmaster and Alpöge reportedly used models from two rival labs. OpenAI, in turn, says an unreleased model found its claimed proof. These accounts describe AI as a research instrument, but they do not settle how credit, confidentiality and responsibility should work when an academic’s use of a commercial system overlaps with a company’s internal research ambitions. The answer may depend on facts absent from the available account: product terms, data controls, laboratory procedures and the technical path from prompt to proof.
Credit allegations compound the conflict
Buckmaster has also made a more personal allegation about the ensuing discussions. He said an OpenAI representative asked him to remove Alpöge’s credit from a proposed compromise and later warned that making the dispute public could harm his career. These are serious assertions concerning conduct and attribution. The supplied material contains OpenAI’s response on access to the research and user data, but it does not provide a reported company response to the alleged request about credit or the claimed career warning.
That absence matters. It would be wrong to treat an allegation about private conversations as an established account without supporting records, independent witnesses or a direct response that addresses it. It is also unclear from the supplied material what the proposed compromise would have contained, how credit was being discussed, or whether the parties agreed on any shared description of their respective contributions. The dispute has therefore moved beyond a technical priority question into a disagreement over how researchers should be recognized and treated during a conflict.
There is a practical reason the parties’ technical claims may receive close scrutiny. If OpenAI’s proof is intended as a full resolution of a Millennium Prize problem, questions about publication, review and the proof’s relation to prior work will be central. If Buckmaster and Alpöge’s results amount to important progress rather than a solution, their contribution may still matter in any account of how the field advanced. None of that can be fairly decided from claims about compute use, product interactions or the timing of announcements alone.
The available report has not been independently corroborated. Its account rests on a single secondary-source report and competing statements attributed to Buckmaster and OpenAI; the supplied evidence does not establish access to unpublished work, data-derived influence, misconduct, or the correctness of either claimed mathematical result. The immediate test will be whether the underlying proofs and a clearer record of the research chronology become available for expert examination.
For further context on this subject, see Three hikers rescued on Mount Shasta after reported Gemini trip planning.
Reporting notes
What is confirmed: The available account describes competing statements, not independently verified findings or a reviewed proof.
Why this matters: The dispute raises questions about research priority, AI product data and the verification of claimed mathematical breakthroughs.
What remains unclear: Whether OpenAI had any advance knowledge of the pair’s work, whether product-use data mattered, and whether either result is mathematically valid. This report is based on one source and has not been independently corroborated.
