Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by lossolo on Monday October 09, 2023 @08:28PM
(#37827165)
> This is super interesting because you’ll end up with different people with different opinions on “what is the best answer” if they all have different experiences.
In context of grandparent comment it's code, so while we can have different solutions to the same problem, we can verify if the solution is correct because it's not an opinion on some subject.
Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by aarong11 on Monday October 09, 2023 @09:17PM
(#37827517)
There's no such thing as "one correct answer" to most tasks in any language. In mathematics, maybe. If you're asking it to recall a specific quote verbatim, yeah. But other than that there are many (infinite?) different ways you can paraphrase differently to get the same point across (at least in English).
Most of us when using LLMs feel it should only produce the one correct answer; sometimes that’s the right way to think about it, but other times we might want an opinionated answer, when there is no true answer.
So I think if we get better learning and self correcting machines, we also get more opinionated and sometimes incorrect machines. Kind of like people. And sometimes they’re stubborn or confidently incorrect, but that I think actually points to more intelligence (in people at least), since it points to people deciding based on their experiences.