I expect that once we can edit LLMs semantically, we can produce compact core LLMs that may not be good at style imitation, but that can quickly and reliably extract knowledge from text and store and exchange it in a kind of universally interpretable mental representation code.

→ View original post on X — @plinz,