Saying AI is "matrices" is like saying human cognition is "neurons." Maybe true at some level, but it's an implementation detail. The important part of a language model is the function that maps tokens to contextual embeddings.