Erdős problems form a substantial fraction of all mathematical problems that have been explicitly stated but not solved; are sufficiently famous that people care about them; and are sufficiently uninteresting that people have not spent that much effort trying to solve them.

Solving problems people have already stated is a niche activity in mathematical research. More often, people study something they find interesting, try to frame it in a way that can be solved with the tools they have, and then try to come up with a solution. And in the ideal case, both the framing and the solution will be interesting on their own.

Erdos problems are easier to state, thus they make a great benchmark for the first year of AI mathematics.

Afaik this is because there is a community and database around them.

They're just famous because Erdos was a great mathematician, kinda like the Hilbert problems a century earlier.

It's not just Erdos problems - https://news.ycombinator.com/item?id=48213189

I was promised a cure for cancer, but all I got was this disproof of an Erdos problem.

It's a large set of problems that are both interesting and difficult, but not seen as foundational enough or important enough that they have already had sustained attention on them by mathematicians for decades or centuries, and so they might actually be solvable by an LLM.

The models can't actually so good work on practical problems so openai tasks them on stuff nobody cares about