I think of understanding as a mapping between observables and causal models. Weights can be a way to represent understanding. LLMs don't always understand, but Gary is himself simulating understanding via prompt completion, which is why he always gets to the same conclusion.

→ View original post on X — @plinz,