News for hackers. Stuff that matters.

A comment on Ask HN: Are we sure LLMs are that useful in a web search application?. Parent comment

Re:Ask HN: Are we sure LLMs are that useful in a web search ... by eptcyka on Thursday February 09, 2023 @10:51AM (#34725755)
It's not designed to generate plausible sounding text, it's designed to produce text that is statistically likely to be sound. It's a markov chain, but there might be multiple roots and some backpropped probabilities for which leafs should be picked. The statistical significance doesn't have to track with logic.
[ Parent | on HN ]
Re:Ask HN: Are we sure LLMs are that useful in a web search ... by stingraycharles on Thursday February 09, 2023 @11:23AM (#34726394)
So what you’re saying is that they not just trained it on text, but also verified the answers and trained it in such a way that it would get a negative feedback if the model gave the wrong answer?
[ Parent | on HN ]
Re:Ask HN: Are we sure LLMs are that useful in a web search ... by pwinnski on Thursday February 09, 2023 @11:27AM (#34726480)
It got negative feedback if the model gave wrong-sounding answers, which is not the same as wrong answers.

I have seen no indication that they had anybody verify the factual content of answers in the training process, and many indications that they did not.

[ Parent | on HN ]