The neural network that encodes the LLM has no canonical form, its structure depends on the architecture, the training data, and the order in which it is processed. It is still very difficult to extract a semantically interpretable representation from an LLM.

→ View original post on X — @plinz,