Researchers Say OpenAI’s Bot Shows Signs Of Mathematical Genius

Thirty leading mathematicians from around the world secretly met at UC Berkeley to test OpenAI’s “o4-mini” powerful artificial intelligence. The bot received its most challenging math problems during a weekend competition which left the participants amazed by its responses.

“I have colleagues who literally said these models are approaching mathematical genius,” said Ken Ono, a mathematician at the University of Virginia and a judge at the event, as reported by Scientific American (SCI AM).

The developers trained O4-mini as a compact yet powerful version of ChatGPT to handle complex problem-solving tasks, as reported by SCI AM.

The researchers were amazed when O4-mini solved 20% of 300 unpublished math questions during the FrontierMath test, which Epoch AI developed as a nonprofit organization. Traditional models solved fewer than 2%, noted SCI AM.

“I came up with a problem which experts in my field would recognize as an open question in number theory,” Ono said, as reported by SCI AM. The bot spent two minutes reviewing the literature, tried a simpler version first, and then solved it, adding, “No citation necessary because the mystery number was computed by me!”

“It was starting to get really cheeky […] That’s frightening” Ono added, as reported by SCI AM.

The group discovered ten problems that the AI system could not solve but many participants were amazed by the rapid advancement of the technology. “This is what a very, very good graduate student would be doing—in fact, more,” said Yang Hui He of the London Institute, as reported by SCI AM.

The mathematicians explored potential scenarios where humans would direct AI systems instead of performing mathematical solutions independently. As Ono warned, “It’s a grave mistake to say that generalized artificial intelligence will never come […] These models are already outperforming most of our best graduate students,” as reported by SCI AM