Tristan Buckmaster had the pieces of a legendary problem in hand when OpenAI’s agents announced a solution, and he says his own published work supplied them. A $1M bounty sat on the result.
The New York University professor told Wired the arrangement leaves him little room, describing the frontier labs as a monopoly with no real alternative for mathematicians.
Since going public ten days ago, he has used Codex to tidy his papers and to retrace how OpenAI’s agents might have travelled from his earlier results to a finished proof.
Tens of thousands of agents went to work on Navier-Stokes existence and smoothness only after OpenAI learned how close the equation was to being settled, Buckmaster says. He had pursued that same problem with Anthropic researcher Levent Alpoge, drawing on both Codex and Claude. The company opened an investigation and amended its September 8 announcement to say his prompts could not have influenced the system in any way.
Andreas Thom describes a parallel case. The German mathematician has spent two decades on geometric group theory techniques that only a handful of people grasp. OpenAI said in August that Astra had used them to settle a problem he was working on, leaving him to wonder how the company learned of them. He wrote to researchers about a 2019 paper the announcement had overlooked. The press release was amended, and he was assured his ChatGPT sessions never entered training.
What he says he cannot recover is provenance. Models, in his view, have ended the idea that contributions can be traced to the people who made them.
Cornell’s Alex Townsend reports feeling excited and nervous at once, and wondering what his purpose becomes. Colleagues have begun trading notes on which subscriptions unlock the strongest models.
