Comment Re: Still hasn't proven to be useful for mainstre (Score 1) 40
I think the principle you are missing is that LLMs can ONLY hallucinate. That's all their probability based "reasoning" allows. Their training data is not complete, and even if it were, there'd still be biases built into it. The reason we judge some result from them accurate is because we have ulterior sources of information against which we check. We judge something hallucination because it conflicts with our seemingly independent information. But all LLMs can do is follow markov chains, they have no concept of an accurate response or an inaccurate response.
Also, the world is open ended, an LLM's world stops at their training data.