Kevin Hartnett

It's like they would either say, so you can balance your checkbook or to teach you how to think, right?

It's like one or the other.

And to teach you how to think is really the point here.

It's like if you can reason through a math problem, think logically, then you can apply that type of skill to like all of the parts of your life.

And I think the labs believe that if you can teach a model to reason through math problems, it's going to be able to do all these other things that are much more probably like commercially valuable well as well.

Right, when ChatGPT came out in November 2022, mathematicians were like passing around all these like, ha, ha, ha, look at this stupid model telling me that there are only finitely many primes when we all know there are infinitely many primes.

And like two plus two is five, basically that kind of thing.

I mean, I think essentially the models got better.

There's definitely an element of reinforcement learning on math problems that make these models better, like RL on math.

But I think it is just like the general improvement in these models that we all experience kind of in a lot of ways we use them has led to these kind of incredible reasoning tasks that they become capable at.

Yes.

I mean, so Paul Erdos, um,

is a really colorful mathematician.

He was essentially the Bob Dylan of math in that he like spent his life on the road.

He died, I think at age 83, actually at a math conference.

He slept on mathematicians' couches his whole life.

And as he went around, he compiled lists of problems that he thought were interesting,

Either he would find them in the wild or he invented many of them.

And he created this Erdos list.

He endowed them with these like tiny little rewards, like $20 for solving this problem, $500 for solving this problem.