François CholletVerified account

used to think that ML people didn't know enough math, that this was what was limiting the field. Now I think the field needs more humanities

Why we would need more math is clear: the field is largely empirical and mostly free of theoretical foundations (true before DL, worse now)

Opening up a math textbook (e.g. information geometry) at any random page can easily lead to a good and novel ML paper. Math under-exploited

But being better at building solutions is less valuable if you don't pick the right problems to solve and don't apply your solutions wisely