Those are not serious mathematical problems. Those are toy math problems, crafted backwards from known facts, designed to be solved in under 1hr, that are hard for most humans because they lack the memorization and recall and search speed that the computer has.
Sure. But even many high-caliber research mathematicians can’t do Putnam problems in a heartbeat. If we get to the point where an LLM can solve any homework problem that appears in a textbook, including graduate textbooks, that would already be something like a “lemma prover” if not a full-blown “theorem prover”.
Anyway, I think five years ago I was skeptical that ML would even get to the point of being able to solve competition problems, and I was proven wrong, so my priors have been updated.
