Sander Schulhoff: commitment

Transcript context

…And then find the thing that's common across them. Amazing. Okay. Anything else that you wanted to share before we talk about prompt injection and red teaming? I guess just quickly, maybe a real reality check is the way that I do regular conversational prompt engineering is I'll just be like, if I need to write an email, I'll just be like, "Writ emil," not even spelled properly about whatever. I usually won't go to all the effort of showing it my previous emails. And there's a lot of situations where I'll paste in some writing and just be like, "Make better, improve." So that super, super short... So that super, super short, lack of details, lack of any prompting techniques, that is the reality of a large part, the vast majority of the conversational prompt engineering that I do. There are cases that I will bring in those other techniques, but the most important places to use those techniques is the product-focused prompt engineering. That is the biggest performance boost. And I guess the reason it is so important is you have to have trust in things you're not going to be seeing. With conversational prompt engineering, you see the output, it comes right back to you. With product-focused, millions of users are interacting with that prompt. You can't watch every output. You want to have a lot of certainty that it's working well. That is extremely helpful. I think that'll help people feel better. They don't have to remember all these things. The fact that you're just write email, misspelled, make better, improve and that works. I think that says a lot. And so let me just ask this, I guess, using some of these techniques in a conversational setting, how much better does your result end up being? If you were to give it examples, if you were to sub-problemate, if you were to do context, is it 10% better, 5% better, 50% better sometimes?…

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