Post #4327849

@thestrangelet @cmccullough [1/2] One common use case is as a search engine, and I would argue that what counts is not whether those have errors, but how one can reduce errors. In the case of LLM's these seem to mostly be hallucinations. I recommend the following paper: https://arxiv.org/abs/2512.01797 They found that certain neurons they call H-neurons light up whenever models hallucinate, indicating that they have an internal representation of when they hallucinate.