It's interesting. He says ( 32:10 ) "A problem, you can solve or not solve, and that's a scoring function. And AI is pretty good at optimizing scoring functions. But a conjecture could be useful, or not useful, ... you can generate random conjectures, but to generate a really fruitful conjecture, ... we don't have a good scoring function for that." See Terence Tao on SAIR and Jazz Maia on Manifestation . Subscribe to SAIR .