This paper is arguing that it’s impossible for an LLM to know the answer to every question, therefore it’s impossible to eliminate hallucination. It’s easy to

This paper is arguing that it’s impossible for an LLM to know the answer to every question, therefore it’s impossible to eliminate hallucination. It’s easy to imagine an LLM that responds “I don’t know” to all questions. An LLM like that isn’t very useful, but it also doesn’t hallucinate. Eliminating hallucination by teaching it to recognize what it doesn’t know is probably a more sane approach than teaching an LLM to know literally everything in the universe.