陶哲轩系学者谈 AI 解出 Navier–Stokes 与奖项争议

Tony Feng @tonylfeng原推发布于 2026.09.12

在 X 上打开原推

Second, I'm disturbed that an event which could have been celebrated as a triumph of human understanding, is instead headlined by ugly disputes. The obsession with priority and prizes will become (much more) unhealthy in the presence of AI. My opinion is that all mathematical prizes should be abolished soon, with perhaps a gentle phaseout in the next year or two allowing only consideration of work that appeared pre June 2026. Traditional norms will have to be revisited, and egos put aside; I think the concept of individual priority and achievement will lose meaning in mathematics research, and the focus should be entirely on collective scientific understanding. This will be upsetting to many people, but I do not see how the community can proceed harmoniously otherwise. In this spirit, I would like to give a generous account of events. Looking at the progress of Alpoge–Buckmaster, I am confident that they could eventually have gotten to Navier–Stokes on their own before long. This would also have been controversial, as accounts suggest that their progress (or at least, the proofs) came from internal Claude models, leveraged with industrial scale compute from Anthropic. Whatever the noble intentions at the origin of the project, it’s hard to imagine that alternate ending would not also have been paraded around as a corporate achievement. I am also confident that OpenAI would have solved Navier–Stokes even if Buckmaster had never typed any information into ChatGPT. As I mentioned, other people were trying to push Sol and Astra along the same program. In the end, even the solutions to the Euler equation are appreciably different. But regardless, can OpenAI just come out and clearly answer the question: if users turn off “Improve the model for everyone”, do you in fact abstain from training on it altogether? If not, what exactly does that button do anyway?