> Unless mathematics is fundamentally more complicated than chess or go, I don't see why that could not happen.
My usual comparison is Sokoban: there are still lots of levels that humans can beat that all Sokoban AIs cannot, including the AIs that came out of Deepmind and Google. The problem is that the training set is so much smaller, and the success heuristics so much more exacting, that we can't get the same benefits from scale as we do with chess. Math is even harder than that.
(I wonder if there's something innate about "planning problems" that makes them hard for AI to do.)
My usual comparison is Sokoban: there are still lots of levels that humans can beat that all Sokoban AIs cannot, including the AIs that came out of Deepmind and Google. The problem is that the training set is so much smaller, and the success heuristics so much more exacting, that we can't get the same benefits from scale as we do with chess. Math is even harder than that.
(I wonder if there's something innate about "planning problems" that makes them hard for AI to do.)