Post

Well, no. Math is not just a collection of problems for which we want true/false answers. It’s a community, where experts help newcomers develop skills and understanding so that they can be the next generation of experts. And the goal is to advance human understanding.

In the course of their training, PhD students and early-career mathematicians do prove things! But they also *learn* things, and provide new perspectives, and new problems to think about. This new drop from OpenAI does none of that.

The exposition in the papers is not good, according to a person I know who has tried to understand an earlier AI-generated paper. And a preprint I saw today points out that the computer-verified proofs provided with some of the papers are not verifying the English language proofs in the papers.

Human mathematicians are left to sort out this mess. So it’s like someone dropping a burning bag of shit on your doorstep, telling you there’s a silver dollar inside, and wondering why you’re not happy.

Plus, there’s the general issue that there are actual people working on some of these problems. Their work is on hold now while we wait for all this to get sorted out. AND the AI papers I’ve seen are really bad at providing attribution and giving citations to earlier work in a meaningful way.

One can also ask: how much money did OpenAI spend on this? What would have happened if that same amount of money had funded human researchers?

OpenAI doesn’t care about any of this. They want to be able to brag that their models have resolved open problems so that their IPO gets them more money.

There's also an issue of cognitive distortion with these systems. In OpenAI's earlier drop of problems, which were checked by their human mathematicians, they just apparently forgot that the Navier–Stokes equations are a physical model and published a solution that is physically impossible.

I think I missed that aspect of the Navier–Stokes example!