Prompt Refusal
Data Skeptic
00:00
How to Get SpongeBob to Give an Apology
It could be that when human raters were looking at chat GPT outputs, they saw chat GPT apologizing for certain things and they said, yes, do more of that. It's also fun to see some of our research was noticed by people who have discussions on Hacker News or Y Combinator that discussion website. But patiently, they would add the test cases and slowly, slowly the wrinkles would get ironed out. Though we can't see it and we don't know what's in it, I think that test corpus was probably growing behind the scenes. And they're just watching what we brag about on Twitter and quietly patching things in the background.
Transcript
Play full episode
