Evolution figured out how to get learning to produce hierarchical models for perception and planning.
It had to find a solution to the collapse problem. What you describe is "classical" reinforcement learning. We know that's way too inefficient.

→ View original post on X — @ylecun,