An LLM trained on the entire internet incorporates essentially all (human-known) aspects of things.

Like a superposition of all kinds of writing styles, all perspectives on every topic and all kinds of ways to do things right and wrong.

In post-training (and of course already by sensibly curating the input data in the first place), people try to select useful submanifolds from that.

Ideally, you get a combination of all the extant strengths of people combined into one model. Self-improvement then leaves the sphere of human capability for good.