Casey Noon

Okay, let's bring in Kevin Hartnett.

Kevin Hartno, welcome to Hard Fork.

Thanks, Kevin.

Well, Kevin, we've brought you here today to talk to you about AI and math.

You just wrote a book called The Proof in the Code, which is, I'll just say it, the most interesting book I've ever read about math.

So when we last checked on the field of mathematics and AI, it was last summer, and three of the big AI labs, Google DeepMind, OpenAI, and Harmonic, had all reported that their math models had achieved a gold medal score at the International Math Olympiad.

That was something that people had been saying for years would be impossible for, or would take many, many decades for computers to be able to do, but their AI models did it last summer.

What has been happening in the field of AI and math since then?

Why were the labs so focused on the IMO and on math in general?

Was it because that was just like a very hard challenge that they liked, or was it because they thought that being able to do math at a high level would enable their models to do other important things?

How did mathematicians view them?

They said, something isn't adding up here.

Thank you, Casey.

I've been following the story of AI and math in part through people like Terence Tao, who is widely considered the greatest living mathematician.

He has been sort of experimenting and writing and making videos about his experiments with these AI programs for use in sort of frontier math research for a number of years now.

And when he started sort of working with them, he was like, oh, these aren't that helpful, or maybe they're like a sort of mediocre grad student who you'd have assisting you.

And more recently, he seems to be actually saying like this is revolutionary for the kind of frontier math research that I and other professional mathematicians do.

He recently made a video with OpenAI talking about how he can now just try a bunch of sort of crazy ideas and experiments because the sort of cognitive friction of using these models means that you can just sort of have an idea and give it to the model and say, go test this a bunch of different ways and figure out if there's anything there.

Is that a widely held view among mathematicians, or is he just sort of on the extreme AI-pilled end of the field?

I do.