When we autoregressively sample from a language model, we are, by definition, seeing its hallucination. Users may think otherwise only when the LLM’s “dream” just so happens be an output that is acceptable to the user.

→ View original post on X — @hardmaru,