An LLM is trained on a wide distribution of data. "Just asking" it to not make mistakes makes it more likely to sample the part of the distribution that contains no mistakes.
Doesn’t that predispose that it knows which parts of the distribution do and don’t have mistakes, and therefore that it knowingly makes mistakes unless you ask it not to? That doesn’t seem right to me and I’d be really surprised if this actually makes it stop hallucinating - seems more like something you’d put in the prompt without knowing why because it “seems to” produce better output (i.e. cargo cult prompt engineering).
Correctness is something it learns. I've read a few papers about hallucinations, and the jury is still out on whether a model knows when it's hallucinating, if we assume hallucinations are orthogonal to correctness
Now this distinction isn't very useful in the grand scheme of things because in the end the output is wrong anyway, but it doesn't make asking to work along the axis of correctness cargo cult
> Doesn’t that predispose that it knows which parts of the distribution do and don’t have mistakes, and therefore that it knowingly makes mistakes unless you ask it not to?
Of course it does, to the extent it "knows" anything. It replies in a way that's average for the distribution. If you tell it this is an important task and not to make mistakes, it will give you a response that's more like that of someone who's been told this is an important task and not to make mistakes.