This is true, but also misleading. We are learning that the models achieve compression by distilling higher level concepts and deriving generalized human like a

This is true, but also misleading. We are learning that the models achieve compression by distilling higher level concepts and deriving generalized human like abilities, for example the recent introspection paper from Anthropic.