Ethan (@torchcompiled) argues eliminating LLM hallucinations requires exact memorization, which destroys the utility of generative models

Thinking, reasoning, experience and about everything else is just controlled hallucination constantly and imperfectly regressing towards congruence with observed reality. The only way to escape hallucination is to regurgitate exact memorizations, which defeats the whole purpose. All things are hallucinations, but some hallucinations are useful