The people who could instruct such AI and evaluate its outputs would be domain experts. Without them how would you even start to approach those problems?
Have you looked at the papers and C.V.s of the people on the front page of x.ai? Many have already worked on machine guided proofs and I see people with pure math and physics backgrounds.
Formal verification research is conducted since 30+ years and LLMs aren't really good at math and logic, so based on what grounds can we expect a breakthrough in machine guided proofs anytime soon?
Various recent math olympiad results from deep mind and openai.
I think we'll see something analogous to the classification of the finite simple groups collaboration, but where some core researchers will do several proofs in the area to show it the kinds of things needed and it can help fill in others. Things where there is a large volume of similar work to be done.
