The people mocking "no proof" didn't realize that it was their insistence to have a particular type of theory that was holding them back. They would only work on "trivial" models for which the loss was convex, so they could use fancy semi-definite programming methods and prove

→ View original post on X — @ylecun,