The code that specifies the architecture is tiny.
That's precisely my point.
There is a tiny amount of prior structure and a lot of learning.

→ View original post on X — @ylecun