OpenAI Releases Answers to 377 Math Problems, Shaking Up the Math World

On Tuesday, the artificial intelligence company OpenAI released solutions to 377 math problems that no human had been able to solve. The problems covered many areas of math, including algebra, number theory, and computer science. This big release came just weeks after OpenAI said it had cracked one of the hardest math problems in the world — a problem so tough that solving it came with a $1 million prize. Many mathematicians were surprised, and some were upset, by how the company shared the news.

OpenAI's AI model solved these problems on its own, but not everyone is happy about it. Mathematician Tristan Buckmaster from New York University pointed out that it is still unclear whether the AI used ideas from human mathematicians without giving them credit. He said the company did not do enough careful checking before releasing so many results at once. He also worried that the AI may have taken human work and simply finished it without saying so.

A group called an advisory board was set up to help AI companies share math results in a responsible way. The board is made up of mathematicians and is hosted by the Institute for Advanced Study in Princeton, New Jersey. It is not connected to any AI company, so it can give independent advice. After OpenAI made its big release, the board said this was just the beginning of the process of humans understanding the new math results.

The advisory board made several suggestions to OpenAI. One suggestion was that AI companies should share all the steps their AI used to reach an answer. Another suggestion was that companies should stop using unsolved math problems just to test how powerful their AI models are. OpenAI said it listened to some of the advice, but it did not follow all of the board's recommendations.

Dan Roberts, the research lead at OpenAI, explained why the company kept testing its AI on hard math problems. He said testing was important to build better tools, and that the math solutions were a side effect of that work. The AI model used to solve the problems has not been released to the public. On average, each solution took about three hours of computing time.

To help make sure the answers were correct, many of the solutions were checked using a computer language called Lean. Lean is a special program that checks whether the logic of a math proof is sound. For 10 of the 377 solutions, OpenAI also shared summaries explaining how the AI reached its answer. This was meant to help human mathematicians understand and use the results.

Some mathematicians worry that AI is moving too fast for humans to keep up. Even if AI can solve problems, humans still need time to understand the solutions and learn from them. Kai Shaikh, a math student at the University of Toronto, said this race to solve problems could hurt the deeper understanding that makes math meaningful. Melanie Wood, a Harvard mathematician on the advisory board, said the goal is to make sure AI results can be understood by mathematicians and can truly advance the field.