Symbolic imageMathematicians question vetting after OpenAI publishes hundreds of claimed results
OpenAI released a large collection of AI-generated maths results on Tuesday. Mathematicians are impressed but doubt how the results were checked.
OpenAI put the results on GitHub with a blog post on Tuesday, 6 October. FAZ counts more than 700 contributions with solutions, proofs or proof fragments, while The Guardian reports over 370 results, in fields from number theory and algebraic geometry to mathematical logic. According to the Wall Street Journal, cited by FAZ, they include progress on three of the five remaining Millennium Problems, and Alex Kontorovich of Rutgers University said a human achieving this would be sure of a Fields Medal. OpenAI says it put about 4,000 problems to the model and has outlined its method and the computing power required.
Doubts over vetting and access
OpenAI says it consulted experts from the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. The institute itself said it does not endorse testing the hardest problems on proprietary models, warning of a two-tier system and of AI arguments no human can verify. OpenAI promised to give mathematicians a voice but did not say it would stop. NYU mathematician Tristan Buckmaster told the New York Times that prompts may have fed the model information that helped it reach results.