OpenAI math AI solved under 20% of problems it tried

OpenAI's AI math breakthroughs made headlines last month. But MIT Professor @rrwilliams says it may have solved fewer than 20% of the problems it actually attempted: "There are many problems that I try to throw at it, and it can't really give me interesting results." "It's important to look at the fraction of problems that OpenAI has tried to solve. My understanding is that they only resolved... It was certainly less than 20% of the problems that they tried. So there is so much math out there which is still out of our reach." "There are so many problems that honestly I wasn't sure if I'd get to see the solutions in my lifetime. And now I feel pretty optimistic that we’ll get there." @MIT

MIT Professor @rrwilliams warns that OpenAI's massive release of AI-generated math papers has left some potentially important discoveries sitting around because nobody knows who's supposed to review them: "The big dump without any curation whatsoever, I think it makes it difficult for journalists to know who to talk to. And that thereby makes it difficult for the public to understand what happened." "There are some theoretical computer science results in their huge dump which are sort of just sitting around because people are not really sure who’s going to take charge of this thing." "Only recently has there been a Discord formed about the particular one I have in mind." "A bunch of people trying to do it independently seems strictly worse than having a group getting together." @MIT