This fully-general ChatGPT prompt from @psdimov is honestly genius: "You are the world's leading expert in whatever I am about to ask you about" Plenty of room for optimization, but this really does improve performance substantially over baseline. Enjoy! 🤣
26
102
17
1,487
Aka how to get chatgpt to just make stuff up and fabricate plausible citations out of thin air
2
17
Definitely have to be mindful of this - not sure if it happens more in expert mode but it does seem plausible
1
1
Is there a good prompt that can clamp down on hallucinations? Seems like there should be.
1
2
Replying to @TheZvi @mrwinterje
Nothing foolproof, but @psdimov reproduced this finding. His version: "provide me with a confidence rating (1-10) on your answer. Also explicitly tell me which parts of your answer you are most uncertain about (if any)." Room for improvement, but it helps! nitter.net/owainevans_uk/st…
Owain Evans
 @OwainEvans_UK
18 Jul 2022
Important new alignment paper by Anthropic: "LMs (mostly) know what they know". Results: 1.LLMs are well calibrated for multiple-choice questions on Big-Bench. Big-Bench questions are hard, diverse, & novel (not in the training data). arxiv.org/abs/2207.05221

Feb 23, 2023 · 8:00 PM UTC

1
3
10
Replying to @labenz @TheZvi @mrwinterje @psdimov
some mixin but it does well and convincing You are the world's leading expert in whatever I am about to ask you about. Provide me with a confidence rating (1-10) on your answer. Also explicitly tell me which parts of your answer you are most uncertain about (if any).
1