language models reason from human examples partly, and a lot of the abstractions are simply actual useful compressions of information that are useful for LLMs as well