It seems the state of the art with generating GPT-3 speech is to generate multiple responses until you have a good one and cherry-pick it. I’m not sure whether including a disclaimer explaining that process will still be helpful. Yes there’s a sizable number who don’t know about that process or who don’t automatically assume it’s being used, but I’m not sure how big that number is anymore. I don’t think Isusr should explain GPT-3 or link to an OpenAI blog every time he uses it as that’s clearly a waste of time even though there’s still a large number of people who don’t know. So where do we draw the line? For me, every time I see someone say they’ve generated text with GPT-3 I automatically assume it’s a cherry-picked response unless they say something to the contrary. I know from experience that’s the only way to get consistently good responses out of GPT-3 is to cherry pick. I estimate that a lot of people on LW are in the same boat.
It seems the state of the art with generating GPT-3 speech is to generate multiple responses until you have a good one and cherry-pick it. I’m not sure whether including a disclaimer explaining that process will still be helpful. Yes there’s a sizable number who don’t know about that process or who don’t automatically assume it’s being used, but I’m not sure how big that number is anymore. I don’t think Isusr should explain GPT-3 or link to an OpenAI blog every time he uses it as that’s clearly a waste of time even though there’s still a large number of people who don’t know. So where do we draw the line? For me, every time I see someone say they’ve generated text with GPT-3 I automatically assume it’s a cherry-picked response unless they say something to the contrary. I know from experience that’s the only way to get consistently good responses out of GPT-3 is to cherry pick. I estimate that a lot of people on LW are in the same boat.