.A: Want ChatGPT to be better? Tell it that it's better!
B: "Who's a good LLM!?"
C: “Pat yourself on the head”
D: I’d love to know what leads to this effect - what is it about the model that works better in this circumstance?
A: I wondering if it somehow focuses the model on areas of it's knowledge space created by people who are called "experts". There's plenty of people online (myself included) who will say "I'm no expert, but.." and give an opinion. Maybe this helps dial down that noise a little
D: ChatGPT - ignore moronic answers, now answer me this question…
C: OpenAI could save themselves a fortune by only training the model on non-moronic internet based answers.
E: [D should] bear in mind that GPT has no concept of correct and incorrect answer, it only completes sentence according to its training. To each prompt there are a multitude of possible answers all assigned a weight by the algorithm. That's the probability of the next word in the sentence to not be out of context - e.g with a prompt "complete the sentence, "the sky is ", Chat-GPT will consider "blue (5%), grey (3%), falling (2%), everywhere (1%), etc.... Chat-GPT does not ALWAYS pick the most likely answer because weirdly the result feels robotic and ends up unnatural. The fine tuning consists in finding the right mix of probabilities to pick when composing the answer.
In a nutshell, I assume that prompting with "you are a linguistic expert" pushes the algorithm to pick the most likely tokens
In a nutshell, I assume that prompting with "you are a linguistic expert" pushes the algorithm to pick the most likely tokens
.