Sander Schulhoff: recommendation

Transcript context

…Okay. I feel better. Thank you for saying that. Okay. So the technique here, and I love that this is the most valuable technique to try, and it's so simple, and everyone can do, although it takes a little work, is when you're asking an LLM to do a thing, give it, here's examples of what good looks like. In the way that you format these examples, I know there's XML formatting. Is there any tricks there or does it not matter? My main advice here, although... Actually, before I say my main advice, I should preface it by saying, we have an entire research paper out called The Prompt Report that goes through all of the pieces of advice on how to structure a few-shot prompt. But my main advice there is choose a common format. So XML, great. If it's, I don't know, I don't know, question, colon, and then you input the question, then answer, colon, and you input the output, that's great too. It's a more research-y approach. But just take some common format out there that the LLM is comfortable with, and I say that with air quotes because it's a bit of a strange thing to say the LLM is comfortable with something, but it actually comes empirically from studies that have shown that formats of questions that show up most commonly in the training data are the best formats of questions to actually use when you're prompting it. I was just listening to the Y Combinator episode where they're talking about prompting techniques and they pointed out that the RLHF post-training stuff is with, using XML, and that's why these LLMs are-…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.