It does that because all of its training data in the last, post-training stage are of the form [question -> authoritative sounding solution], where the solutions are written by humans. The LLMs just imitate the form/style of that training data.

→ View original post on X — @karpathy,