Sam Schillace: prediction

Transcript context

…story? Hey! Yeah, that's kind of funny. I'm actually kind of a modest person and so I'm sure I feel about having my name attached to them. Although I do agree with all, I believe all of them because I wrote all of them. This is like a designer, John Might, who works with me, decided to stick my name on them and put them out there. Seriously, but like, well, but like, so this was just I, I'm not, I don't build models. Like I'm not an AI engineer in the sense of, of like AI researcher that's like doing inference. Like I'm somebody who's like consuming the models. Exactly. So it's kind of funny when you're talking about AI engineering, like it's a good way of putting it. Cause that's how like I think about myself. I'm like, I'm an app builder. I just want to build with this tool. Yep. And so we spent all of the fall and into the winter in that first year, like Just trying to build stuff and learn how this tool worked. And I guess those are a little bit in the spirit of like Robert Bentley's programming pearls or something. I was just like, let's kind of distill some of these ideas down of like. How does this thing work? I saw something I still see today with people doing like inference is still kind of expensive. GPUs are still kind of scarce. And so people try to get everything done in like one shot. And so there's all this like prompt tuning to get things working. And one of the first laws was like, break it into pieces. Like if it's hard for you, it's going to be hard for the model. But if it's you know, there's this kind of weird thing where like, it's. It's absolutely not a human being, but starting to think about, like, how would I solve the problem is often a good way to figure out how to architect the program so that the model can solve the problem. So, like, that was one of the first laws. That came from me just trying to, like, replicate a test of a, like, a more complicated, There's like a reasoning process that you have to go through that, that Google was, was the react, the react thing, and I was trying to get GPT 4 to do it on its own. And, and so I'd ask it the question that was in this paper, and the answer to the question is like the year 2000. It's like, what year did this particular author who wrote this book live in this country? And you've kind of got to carefully reason through it. And like, I could not get GPT 4 to Just to answer the question with the year 2000. And if you're thinking about this as like the kernel is like a pipelined orchestrator, right? It's like very Unix y, where like you have a, some kind of command and you pipe stuff to the next parameters and output to the next thing. So I'm thinking about this as like one module in like a pipeline, and I just want it to give me the answer. some kind of command and you pipe stuff to the next parameters and output to the next thing. So I'm thinking about this as like one module in like a pipeline, and I just want it to give me the answer. I don't want anything else. And I could not prompt engineer my way out of that. I just like, it was giving me a paragraph or reasoning. And so I sort of like anthropomorphized a little bit and I was like, well, the only way you can think about stuff is it can think out loud because there's nothing else that the model does. It's just doing token generation. And so it's not going to be able to do this reasoning if it can't think out loud. And that's why it's always producing this. But if you take that paragraph of output, which did get to the right answer and you pipe it into a second prompt. That just says read this conversation and just extract the answer and report it back. That's an easier task. That would be an easier task for you to do or me to do. It's easier reasoning. And so it's an easier thing for the model to do and it's much more accurate. And that's like 100 percent accurate. It always does that. So like that was one of those, those insights on the that led to the, the choice loss. I think one of the other ones that's kind of interesting that I think people still don't fully appreciate is that GPT 4 is the rough equivalent of like a human being sitting down for centuries or millennia and reading all the books that they can find. It's this vast mind, right, and the embedding space, the latent space, is 100, 000 K, 100, 000 dimensional space, right? Like it's this huge, high dimensional space, and we don't have good, um, Intuition about high dimensional spaces, like the topology works in really weird ways, connectivity works in weird ways. So a lot of what we're doing is like aiming the attention of a model into some part of this very weirdly connected space. That's kind of what prompt engineering is. But that kind of, like, what we observed to begin with that led to one of those laws was You know, ask smart to get smart. And I think we've all, we all understand this now, right? Like this is the whole field of prompt engineering. But like, if you ask like a simple, a simplistic question of the model, you'll get kind of a simplistic answer. Cause you're pointing it at a simplistic part of that high dimensional space. And if you ask it a more intelligent question, you get more intelligent stuff back out. And so I think that's part of like how you think about programming as well. It's like, how are you directing the attention of the model? And I think we still don't have a good intuitive feel for that. To me, the most interesting thing is how do you tie the ask smart, get smart with the syntax and semantics piece. I gave a talk at GDC last week about the rise of full stack employees and how these models are like semantic representation of tasks that people do. But at the same time, we have code. Also become semantic representation of code. You know, I give you the example of like Python that sort it's like really a semantic function. It's not code, but it's actually code underneath. How do you think about tying the two together where you have code? To then extract the smart parts so that you don't have to like ask smart every time and like kind of wrap them in like higher level functions.…

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