OpenAI fought dirty on career-making math problem, says NYU mathematician

TechCrunch reports that NYU mathematics professor Tristan Buckmaster announced preliminary findings and proofs toward a major unsolved problem with Anthropic mathematician Levent Alpöge, using Codex and Claude. Shortly after, OpenAI published a full proof of the Navier–Stokes existence and smoothness Millennium Prize problem, attributing it to an unreleased next-generation model after a week-long effort consuming 300 billion output tokens—$22.5 million of compute at current Astra rates.

Buckmaster says information about their progress reached OpenAI while they were finalizing results; when contacted, OpenAI said it already had a full proof, then grew evasive on start date and human input. He says it emerged an entire team had worked the problem with massive compute, and that the first prompt was sent in the past few days after news of their work reached OpenAI. OpenAI’s own post says the latest effort began September 1, inspired by rumors two Millennium problems had been solved, and confirms talks with Buckmaster and Alpöge.

Buckmaster argues almost nobody else was attacking via the smooth-force route Luis and Diego opened—the path he and Alpöge quietly chose—and that one does not arrive there in a few days from the problem statement alone. He also alleges OpenAI’s Sébastien Bubeck asked him to remove Alpöge’s credit as a compromise and, when Buckmaster pushed to go public, said “Why would you ruin your career?” then “If you don’t want me to be nice, then I don’t have to be nice.”

Signal Desk’s read: treat this as a priority fight over scientific credit in the age of industrial compute—not as settled proof that OpenAI stole a manuscript, and not as a tidy lab-triumph story either.

The load-bearing claim is timing plus method rarity. If OpenAI’s push truly started after rumors of Buckmaster/Alpöge progress, and if it followed the same uncommon Fefferman-options path few others were on, then “independent rediscovery in a few days” is the charitable story OpenAI needs evidence for. OpenAI’s post denies seeing their work before public release and says no specific user data was accessed, while conceding it cannot rule out that de-identified product-usage data helped models; it also argues the proofs differ and even the Euler-case results differ (forced vs unforced). That is a partial rebuttal, not a full clearing of the incentive problem Buckmaster raises: labs with Astra-scale budgets can finish formalizations faster once a human research direction is known.

Alpöge’s Anthropic affiliation—while he was not doing this for Anthropic—appears to have poisoned the diplomacy. Asking to strip a collaborator’s credit as “compromise,” then escalating to career-ruin language if true as Buckmaster recounts, is institutional ugliness that will stick regardless of how Clay eventually scores the proofs. The Codex training-data worry is the second fuse: extensive Codex use plus OpenAI’s right to train on interactions (with opt-out) makes regurgitation plausible to outsiders even when OpenAI calls it unlikely.

What is over-stated is any implication the Millennium Prize is already pocketed; verification culture still has to grind. What is under-stated is how this episode trains mathematicians to treat frontier labs as rival claimants with asymmetric compute, not neutral tools. Buckmaster’s instinct—to put maximum information in public—is the only workable disinfectant.

Navier–Stokes existence and smoothness is one of seven Clay Millennium Prize problems, each carrying a $1 million bounty. The equations are central to fluid mechanics yet poorly understood theoretically. OpenAI’s release followed Buckmaster’s statement the same day; TechCrunch updated after OpenAI published its Navier–Stokes write-up.