There’s also a fascinating proposal I read somewhere where you create a training set with a knowledge cutoff of 1900 or 1930 and see if the resulting AI could predict the future or independently discover later scientific breakthroughs.