A parrot that only continues
Last time, AI read a whole library. But at that point it's like a parrot: it can only continue sentences, not answer you.
You ask one thing, it says another
Ask it how to sleep better, and it might continue with a list of other questions, because online, questions often sit next to each other like that.
Learning from sample conversations
Next, people write lots of sample conversations: if asked this, answer like that. AI copies them, and starts answering the actual question.
People give scores
Then AI writes a few different answers, and people score which is better. Helpful, polite, honest answers get rewarded. Made-up or dangerous answers lose points.
Why AI knows how to say no
This is also when AI learns to refuse harmful requests, and learns the polite tone you see every day.
Ask twice, get two answers?
But why does asking the same question twice give two different answers? Part six has the answer.
This article is based on the video From "parrot" to assistant from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.