Someone on HN asked why Erdos problems are so heavily represented in AI math, and got a confident response that it’s because they’re a significant proportion of open problems. This is a god tier “what are you talking about?” response. A significant proportion of graduate math textbooks will contain a few dozen open problems at the time of writing, and the vast majority of conjectures/open problems don’t make it into a textbook. What the Erdos problems have is convenience and a name. Unlike the conjectures that are scattered throughout papers and textbooks, they have a website, and they’re associated with a very famous mathematician. So they’re a very natural target if you want to test the theorem finding capabilities of your AI model.