So, what if it would try to automatically challenge statements it has made, before giving a final answer. Do you think, it would be possible to improve LLM's accuracy with this kind of auto validation feedback loop?
Like we need to think before speaking, to consider what we are about to say.
The guys at OpenAI are aware of that. For sure.
The scientific community is aware of that. For sure.
So. Why is this done, like it is done?
>ChatGPT can often correct it's answer, if a mistake is pointed out
*["His" because it was always a boy. Girls aren't so overtly competitive]
By coincidence, here's an example from earlier today when I wanted to know the name of a move I'd played in a few chess games:
----
ME: In chess is there a term for an opening where black sacrifices his black bishop on F2 or white sacrifices his white bishop on F7. Thus forcing the opponent's king to have to move and therefore preventing him from castling later?
ChatGPT: Yes, there is a term for this kind of opening. It is called the "Scholar's Mate" or "Fool's Mate." The moves that lead to this opening are...
[I hit "Stop Generating" here]
ME: That is not the answer to what I asked. Please read the question again
ChatGPT: I apologize for misunderstanding your question. The opening you are referring to is called the "Bishop's Gambit" or "King's Gambit Accepted." In this opening, White sacrifices their f-pawn to gain control of the center of the board and open up lines for their pieces, including the bishop on c4.
If Black accepts the gambit by capturing the pawn, White can then play Bxf7+, sacrificing the bishop to force the king to move and prevent castling. This move is known as the "King's Bishop Gambit." Similarly, Black can play the "Queen's Bishop Gambit" by sacrificing their bishop on f2 to achieve the same goal.
----
A trivial example. but it demonstrates the point. It's like "teacher" started to ask 'What is the move called in Chess...?' and "Little Johnny" immediately leapt to his feet and blurted out 'Scholar's Mate!', without even listening to the question.
I suppose in a way it makes interacting with ChatGPT seem a bit more human. Although I doubt it's meant to be like that....
...or is it? Here's a conspiracy theory for you, that's just occurred to me: What if ChatGPT is programmed this way deliberately so that, in occasionally having to correct it, we feel more like we're collaborating" on a semi-equal footing and less like "the machine" is our infallible invincible superior?