It's not an assumption. Reasoning, as I define it, is simply not doable by a system that produces a finite number of tokens, each of which is produced by a neural net with a fixed number of layers.

→ View original post on X — @ylecun,