That's interesting. I suppose it could even be possible to test his theories. Just applied the exact same training methodology to smaller models or slightly easier problems and study what happens.