Will Bryk: preference

Transcript context

…Yeah. Yeah. Because people don't know what's going on. I mean, to me, the best example of like system prompts is like, why are you writing? You're a helpful assistant. Like. Of course you should be an awful, but people don't yet know, like, can I assume that, you know, that, you know, it's like, why did the, and now people write, oh, you're a very smart software engineer, but like, you never made, you never make mistakes. Like, were you going to try and make mistakes before? So I think people don't yet have an understanding, like with, with driving people know what good driving is. It's like, don't crash, stay within kind of like a certain speed range. It's like, follow the directions. It's like, I don't really have to explain all of those things. I hope. But with. AI and like models and like search, people are like, okay, what do you actually know? What are like your assumptions about how search, how you're going to do search? And like, can I trust it? You know, can I influence it? So I think that's kind of the, the middle ground, like before you go ahead and like do all the search, it's like, can I see how you're doing it? And then maybe help show your work kind of like, yeah, steer you. Yeah. Yeah. No, I mean, yeah. Sure. Saying, even if you've crafted a great system prompt, you want to be part of the process itself. Uh, because the system prompt doesn't, it doesn't capture everything. Right. So yeah. A system prompt is like, you get to choose the person you work with. It's like, oh, like I want, I want a software engineer who thinks this way about code. But then even once you've chosen that person, you can't just give them a high level command and they go do it perfectly. You have to be part of that process. So yeah, I agree. Just a side note for my system, my favorite system, prompt programming anecdote now is the Apple intelligence system prompt that someone, someone's a prompt injected it and seen it. And like the Apple. Intelligence has the words, like, please don't, don't hallucinate. And it's like, of course we don't want you to hallucinate. Right. Like, so it's exactly that, that what you're talking about, like we should train this behavior into the model, but somehow we still feel the need to inject into the prompt. And I still don't even think that we are very scientific about it. Like it, I think it's almost like cargo culting. Like we have this like magical, like turn around three times, throw salt over your shoulder before you do something. And like, it worked the last time. So let's just do it the same time now. And like, we do, there's no science to this.…

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