News for hackers. Stuff that matters.

A comment on Telling GPT-4 you're scared or under pressure improves performance. Parent comment

Re:Telling GPT-4 you're scared or under pressure improves pe... by toxik on Saturday November 04, 2023 @04:37AM (#38139168)
It is not accurate to say that an LLM like ChatGPT predicts anything. It is trained to maximize a score function, so it is more like trying to win a game where the moves are word choices.
[ Parent | on HN ]
Re:Telling GPT-4 you're scared or under pressure improves pe... by akoboldfrying on Saturday November 04, 2023 @04:40AM (#38139175)
The game is predicting the next word a person would write.
[ Parent | on HN ]
Re:Telling GPT-4 you're scared or under pressure improves pe... by logicchains on Saturday November 04, 2023 @05:32AM (#38139409)
Not after RLHF.
[ Parent | on HN ]
Re:Telling GPT-4 you're scared or under pressure improves pe... by akoboldfrying on Saturday November 04, 2023 @05:44AM (#38139466)
I'd say RLHF bends the game towards predicting what words a "helpful", "respectful" person would write next, for values of "helpful" and "respectful" that vary according to each person involved in scoring (but which are carefully shaped by the people choosing those people, and paying for their time).
[ Parent | on HN ]
Re:Telling GPT-4 you're scared or under pressure improves pe... by toxik on Sunday November 05, 2023 @03:44AM (#38149402)
That’s not true and that was the entire point of my comment. It’s like playing chess, you’re not predicting your next best move, you’re searching for it.
[ Parent | on HN ]