OpenAI's Astra proof draws fire from mathematician Andreas Thom

OpenAI's Astra model announced a solution to a question Mikhail Gromov posed in 1999, constructing what the company called the first known non-sofic group. The key step ran through a 2019 paper by Andreas Thom and Gábor Kun. Thom, a mathematician in Dresden, had spent months before the announcement working through the same mathematics with ChatGPT, Wired reports.

Thom emailed OpenAI's Mark Sellke and Sébastien Bubeck with two questions: whether his conversations had entered the model's training data, and whether they had been accessible to the system while it worked the proof. Sellke's full reply, by Thom's account, was one sentence: "Regarding your conversations with ChatGPT: that did not happen." Thom published the email exchange and argues the answer only addresses the second question.

This is the second priority dispute at OpenAI in recent weeks. In September NYU's Tristan Buckmaster and Anthropic's Levent Alpöge posted their own progress on Navier-Stokes-adjacent problems; one day later OpenAI claimed a full Navier-Stokes proof. Bubeck told Wired "We did not use their prompt or proofs to prompt our models." OpenAI also quietly updated the Astra writeup to credit Thom and Kun after mathematicians noted the omission.

The Wired piece made the rounds among three of the AI researchers on our radar.