Peter Bloem on Nostr: I would say on a higher level ML models are good for modeling "an intuition" about ...

I would say on a higher level ML models are good for modeling "an intuition" about interesting directions. This is helpful in automated proof search (or game playing) as a fast heuristic to guide the search.

For human/AI interaction, the question then becomes whether "an intuition" is "a human-like intuition". There, I think the models have a tendency to diverge from human instincts, which makes it difficult to collaborate. They "experience" the world very differently to us and what is simple to them can be very complex to us.