Yu Deng, Fields Medalist and professor at the University of Chicago, systematically discussed how to do mathematics in the AI era in his keynote speech at the "Zhishifenzi" "Saixiansheng" forum on October 26.
He first pointed out two advantages of AI over humans: knowledge so broad it has no boundaries, and trial-and-error costs so low they can be ignored. But he stressed that whether in mathematics or anything else, humans must be the subjects, and AI is only a tool we use.
Then Yu Deng turned his fire on AI companies. He said that if AI is truly treated as a researcher, academic rules should be followed—when a major conclusion is proven, a reasonable request is to publish the chain of thought. Yet in reality, most companies refuse to disclose it on the grounds of "preventing distillation," and among the proofs that have been published, only a very few actually reveal the chain of thought.
"You can't openly distill all of humanity's online literature and knowledge when only you are doing the distilling, yet when others merely want to know the chain of thought, you say this won't do and that won't do." This barb of irony exposed the double standard of the big companies—the "magistrate allowed to set fires while commoners are forbidden to light lamps." Yu Deng therefore proposed relying on open-source mathematical models so that the mathematical community can keep its own models in its own hands, while constraining data use, chain-of-thought disclosure, and traceability according to academic norms.
25 Fields Medalists co-sign a protest
This was not a lone voice. On September 11 local time, Terence Tao, Yu Deng, Peter Scholze, and 25 Fields Medalists in total co-signed the statement "Serious Misalignment of Artificial Intelligence in Mathematics," warning of a severe disconnect between the goals of AI companies and the mathematical community. The statement noted that AI solutions are often released hastily, with no time for formal write-ups, isolation of new methods, or citation of others' work, giving rise to serious risks of misattribution and plagiarism.