Interesting that for the hardest problems, AI does better with a massive training set than deducing from assumptions, while humans seem to think they do better reasoning from first principles than using a bigger training set