I don't see LLMs as a large chunk of knowledge, I see them as an emergent alien intelligence snapshotted at the moment it appeared to stop learning. It's further hobbled by the limited context window it has to use, and the probabilistic output structure that allows for outside random influences to pick its next word.
Both the context window and output structure are, in my opinion, massive impedance mismatches for the emergent intellect embedded in the weights of the model.
If there were a way to match the impedance, I strongly suspect we'd already have AGI on our hands.
Both the context window and output structure are, in my opinion, massive impedance mismatches for the emergent intellect embedded in the weights of the model.
If there were a way to match the impedance, I strongly suspect we'd already have AGI on our hands.