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

NYU professor Tristan Buckmaster and Anthropic mathematician Levent Alpöge revealed three proofs and a preliminary result on the Navier–Stokes existence and smoothness problem, a longstanding Millennium Prize question in mathematical physics. The collaborators say they leaned heavily on AI assistance—primarily OpenAI’s Codex, with additional use of Anthropic’s Claude—during their work. Their approach targeted a less commonly pursued route in the problem’s framework, which the pair say few others were exploring. The announcement immediately became entangled with an internal dispute over whether OpenAI had leveraged knowledge of Buckmaster and Alpöge’s progress to advance its own effort. Buckmaster reports being told that OpenAI had achieved a full proof and that an internal team had used large amounts of compute; he says the timing suggested OpenAI’s prompts and work followed information about his group’s research. Sebastian Bubeck, who leads mathematical research at OpenAI, rejected those characterizations and called them inflammatory, promising a more complete response. The disagreement touches on practical and ethical questions about AI-assisted research. Buckmaster noted that users’ Codex interactions can be used to train OpenAI models unless they opt out, raising the possibility—unproven in this case—that his interactions informed OpenAI’s attempts. He also described a fraught exchange in which, he says, Bubeck proposed removing Alpöge’s credit as part of a compromise and warned that going public could harm Buckmaster’s career; Bubeck has framed his own participation as following academic norms. Beyond the immediate personnel dispute, the episode revives broader concerns about corporate incentives and transparency in AI-driven mathematics. The Navier–Stokes problem carries a $1 million prize and a high profile in the mathematics community; when proprietary systems and large compute budgets enter research workflows, questions arise about credit, reproducibility, and how model training policies might affect independent academic work. Buckmaster has said he has not seen OpenAI’s proof and is not accusing anyone, but he has publicized the timeline and exchanges to counter announcements he believes would otherwise present a misleading account of events.