— We are entering a new era of discovery now, Sam Altman says on X, followed by OpenAI president Greg Brockman’s comment that we are moving — towards acceleration of scientific discovery and improving quality of life for everyone.
The problems span deep and wide across the field, with solutions on anything from number theory to computer science and mathematical physics to complex geometry. There are so many solutions that OpenAI says their listings themselves are not rankings, in a document that spans 41 pages with as many as ten problems per page.
In OpenAI’s announcement, they say they have been consulting with Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study «to develop best practices» and to «inform how we release these results,» after getting backlash from the mathematical community on earlier problem releases. The group says:
— This release is the beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge.
«Many» of the proofs are formalized in Lean, which means they are computer-verifiable, with more to come «as we obtain them,» OpenAI says. Along with the proofs themselves, there are ten summaries of model reasoning and estimates on how much compute was spent to solve each problem. OpenAI says the average time spent was three hours of ChatGPT Pro-level reasoning.
— We want to directly empower scientists with state-of-the-art capabilities and are working to responsibly release the model that produced these results, OpenAI writes.
Read more: OpenAI’s presentation, GitHub repository, PDF overview, and AGMAI’s statement. NYT and Interesting Engineering. Discussion on Hacker News and r/Singularity.