Buckmaster slams OpenAI on $1 million proof, still uses Codex

The dispute has been building since the summer. Mathematician Andreas Thom says OpenAI's Astra model used techniques from geometric group theory he had spent two decades developing to prove a problem he had been working on, and he emailed researchers Mark Sellke and Sébastien Bubeck that the firm's "no progress" claim overlooked his 2019 paper. OpenAI amended its press release. Thom asked whether his ChatGPT interactions had been fed into training data and says Sellke replied: "That did not happen." Buckmaster describes OpenAI's approach as irresponsible and "childish," and says the company pointed him to its amended announcement in an email.

Both men accept the tools are too useful to abandon, which is why the field's reaction is not a clean break. Thom has kept using ChatGPT to speed up writing papers, saying AI is "extremely efficient," while Buckmaster has spent the week and a half since his accusations using Codex on his own research. Some institutions are less willing to accommodate: more than 2,000 people with ties to Caltech called on organizers of an AI math hackathon there — sponsored by Anthropic and OpenAI, though OpenAI has since dropped out — to suspend the event. Buckmaster is calling for a detente, saying "I don't want to just engage in fights."

Whether that detente materializes may hinge on whether AI labs agree to ground rules on releasing results, and on whether mathematicians can accept that, as Thom puts it, tracing back who contributed what "is probably over." The stakes fall hardest on younger researchers and students, who Townsend says are already asking what their future looks like; early-career mathematicians risk being "isolated" if they do not use AI to accelerate their work, Thom adds.