AI isn't naturally polite
AI isn't polite by nature. Real people score thousands of its answers. That's called RLHF.
SCORING
Raters pick the better answer
Raters read several answers to the same question and pick the one that's clearer and more helpful.
REWARDS
Learning from rewards
The model learns from those scores like getting rewards, and gradually answers more like the high-scoring replies.
DOWNSIDE
It likes to please
The downside is that AI tends to please you and may say your idea is right even when it isn't.
This article is based on the video RLHF from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.