Transcript 0:00 I had a pleasure talking to Peter Norvig. Thousands of students know him as the author of this monumental book, and I love it too. 0:09 The funny fact is Peter Norvig once thought ordinary businesses would never be able to use the internet at scale. Setting up a web server was too difficult. There simply were not enough people who knew how to do it. 0:22 He was right about the difficulty, but he was completely wrong about how long it would last. 0:28 At AI Conference, I wanted to talk to Peter because he has watched AI change for more than forty years and helped generations of people understand it through Artificial Intelligence: A Modern Approach, which he wrote with Stuart Russell. 0:42 Today, he's watching students who have never programmed build their first apps. So what should we still be learning? What does all this new capability make possible? 0:52 And where might the people who understand the technology best be getting the future wrong? I don't have a very good track record for prediction. 1:01 So I remember when, uh, Al Gore invented the internet, when, uh, there was an ARPANET, and I was on that as a student at Berkeley. 1:09 And then, uh, Gore and others had this idea that there was gonna be internet that was open to everybody, and anybody could sign up for it and, you know, businesses would sign up for it. 1:21 And I said, "That'll never work because there's only, like, a couple hundred people in the world who really know how to set up a web server and, and do it right. How are ten million companies gonna all do that? 1:34 They won't be able to do it." 1:35 But I didn't realize is when there's a demand, people will make the software tools that make that process easy, so you don't have to be this Unix wizard in order to, to set up a network, that you just buy this package from Netscape or Microsoft and answer the questions and it works. 1:52 And so I think we'll see more of that, of, you know, maybe I'm the w- the worst person to be asking or, you know, us insiders who are saying, you know, "I know all the details of how you build these language models." 2:06 No, you shouldn't be asking somebody like that. You should be asking an end user, "What tools do you need to make your life work?" 2:13 I think we'll see that, and we'll see this more powerful tools to do things that we used to think were hard, and then they'll become easy. 2:21 Let's go back to the start of how was our book different than the previous books, right? 2:27 'Cause I, Stuart Russell, my coauthor and I, we had been teaching out of the other books, and we thought, you know, the other books that are available were good, but it felt like the field has changed, that there was- Since '95. 2:37 Yeah. And maybe, you know, over the last decade up to '95. So I think maybe around '85, you know, there was a big push away from logic and towards probability. 2:48 So, uh, Judea Pearl was, was probably the leader of that, and I think everybody else saw, yes, this is the way to go, that we were too constrained by Boolean logic, and AI should really be the study of uncertainty, and probability was the right way to do that. 3:03 So that was one change. And then, uh, machine learning rather than hand coding was another change. The move towards big data was another change. 3:13 And so we kinda felt like the older books were, were missing those changes to the field, and we've... and we named them The Modern Approach. 3:21 And I guess all of those are still true, but now to a much bigger extreme, that, you know, sort of everything is now just scaling and having more data and more compute. 3:32 So sometimes people ask me, uh, "Well, do, do students still need to read the, the first half of your book?" So the second half is on, uh, machine learning, and we cover to some extent, uh, transformers and so on. 3:45 But the first half is knowledge representation and logic. And, and I guess my answer is the, the students probably don't have to read all of that, but the LLM should definitely read all of it. 3:54 And so, like, if you look at AlphaGo, so the reason AlphaGo was able to beat the world champion at Go sort of decades before people thought that was gonna be possible is because it used neural nets in a clever way to encode what people knew about a board position. 4:11 And that would've been enough to demolish amateurs like me, but in order to beat the world champion, you have to take that and then also put in the Monte Carlo tree search, which is this traditional old school algorithm. 4:25 And I think that's still true, that going forward we'll have this mix of saying you need the neural nets to understand the complexity and uncertainty of the world and so on, but you also need clever algorithms. 4:36 I think you defined artificia- artificial intelligence as, um, the way to understand rational agents. Mm-hmm. Right? Yeah. So what are rational agents now? How do you see that? 4:48 Over the last year, we've seen th- this re-emphasis on the idea of agents, right? So it... So we started it back in, uh, 1995. You were probably one of the first to- Yeah... describe agency. 4:57 And in fact, uh, you know, one of the things that our, uh, publisher did for us is we had the book, and they were trying to say, "Well, what's unique about this?" 5:06 And they put a little banner on the side of the book, right? So we had agreed on the title, but then they also put a little banner that said, "The Intelligent Agent Book." 5:14 And at the time I thought, you know, this is kinda like cheap marketing Kind of thing. You know, I expect, uh, a sticker to be on my, uh, new and im-improved soap, but not on a serious textbook. 5:27 But in retrospect, that was probably a good move to say that clearly, that that's the way we were going to describe what AI is. How do you see agents now? 5:35 Because I used your book a lot when I was writing this series about agentic workflows, and you have a category there. Yeah. I guess one of the things we really missed is how do you build a agent? 5:46 And then, uh, we did have, uh, a chapter on, uh, game theory and decision theory that talked about, uh, communities of agents and how they interact, and so on. But it was less of a focus than we're seeing today, right? 6:01 So right now we're, we're seeing the main way we want to interact is to have a swarm of agents and, and figure out how they work with each other, and just having one is, is not enough. 6:11 Intelligence is both deterministic and probabilistic. Do you see the emergence, re-emergence of old concepts that more deterministic? Yeah. So there definitely is some of that. It's a... 6:21 So there are some advocates as saying, "Well, we should go back to symbolic AI, and then maybe combine them in some way of having, uh, neurosymbolic AI." 6:30 And I agree with that, but I guess I think the current system of language models are probably approximately on the right path. So I don't think that we need to tear it all down and start over again. 6:44 Rather, I think the core should be something like what we have now, a neural net with attention, and then on top of that, we should learn how to do logic. And I think that's the way humans evolved, right? 6:56 We aren't born, uh, knowing how to do modus ponens. We're born knowing how to interact and see the world and move and so on. 7:05 And then we learn how to do logical reasoning on top of that, and I think our AI systems should be built the same way. So you said that you think we will continue to develop the neural networks with attention mechanism. 7:17 What do you think about world models in physically- Yeah. 7:21 You know, so I've been in AI for 40 something years, and, and every decade it's like the robots are coming next decade, and they haven't got here yet, but maybe, you know? Some decade that prediction's going to be right. 7:31 I guess I feel like language models are world models, but they could be better m- world models. Why is it a language model a world model? 7:40 Because it knows that, uh, the cat chase the rat is more likely than the rat chase the cat, and so on, right? So it knows a lot about the world. A language model obviously can't do real time 3D scene generation. 7:55 And so, so that's something that's useful. But I guess I see a smaller distinction between language model and a, and a world model. And I look at, like, uh, Yann LeCun, you know, I agree with everything he says, right? 8:08 He says, "Models should be able to do A, B, C, and D." I say, "Yeah, all those things are important. You zeroed in on things that we don't do so well today that we should do well in, in the future." 8:18 And then his conclusion is we have to tear it all down and start over, and mine is, no, I think if we do small tweaks to what we have now, we'll end up in the same place. 8:28 Peter and Yann LeCun agree on several things AI needs to get better at. The disagreement is about how to get there. 8:35 Peter thinks we can build on the systems we already have, but how do we find out what those systems are still missing? And so I remember at Google in the pre-AI days we said, "Well, we want lots of kinds of feedback." 8:49 We wanna measure the clicks, and that gives us feedback that's not very specific, but we have billions of data points, and so we can sort of see and we can do experiments and change things and see if the clicks change, but that'll only tell you so much. 9:02 We didn't call it reinforcement learning with human feedback, but we had human raters. And we said, "Which is better? This set of search results or this set of search results?" 9:11 And we felt like we couldn't have done it without that, right? It would've been incomplete to just have the click data. So we had that as well. 9:18 And then we had a third level where we would say, "Let's take an individual person and, you know, put them in the room with a one-way mirror and record their every interaction and ask them, 'What are you thinking about now? 9:32 Try to search for such and such. What search terms would you come up with? What do you think about this result page? What is this telling you?'" And do this very in, in depth. 9:41 So, you know, we had one level where there's billions of pieces of, of not very useful data, an intermediate level where there's tens of thousands of more detailed data, and then another level where there's dozens of pieces of very detailed data. 9:56 And it felt like it would've been incomplete without any one of those levels. And then there is the question of what people need to learn, what our students are supposed to learn in this new era. 10:08 If a student can get AI to complete the programming assignment, what does that assignment tell us anymore at all? That's a long evaluation. 10:16 Peter's answer gets into which students computer science should serve, and maybe computer science should be for far many more people. And also he asked what we should ask the students to do. 10:27 And going to education since you wrote this, uh, the universities, you mentioned at the keynote that you listed a few things that needs to be changed, uh, because of AI, and education is one of them. Yeah. 10:37 How do you see it, it's going to change? Yeah. You know, that was in the context of, uh, what's software engineering. 10:43 And so one of the things that's changed is if you look at the assignments that were given out, like, three years ago, AI could do every one of them perfectly. 10:51 And so it doesn't really make sense to be giving those assignments. So we have to think of, well, what are the skills that you want now? Another thing is I've noticed that, uh, there seems to be a decrease in CS majors. 11:03 There was a big jump, part 'cause people said, "Oh, well, I can get a great job if I add this." Now we see enrollments are down a bit. 11:09 What I've told my colleagues at, at Stanford and Berkeley is saying, "That's fine," and said, "Don't concentrate on the number of majors you serve." Instead, why don't you try to optimize for the number of minors? 11:21 And I think what's more powerful is, is for somebody to say, "Well, I wanna be a physics major or a biology major or a sociology major. I wanna know the AI and computing tools that I can use in my profession." 11:35 I think, uh, they should kind of be focusing more that way, and that's basically what mathematics has done, right? 11:41 So very few people become professional mathematicians, but lots of people use calculus or statistics or whatever in their work, and so I think CS education should be more like that. 11:53 Then another thing people are worried of, uh, oh, in many fields we relied on essay writing to prove that you know the material, and that's kind of out the window. 12:03 I mean, to some extent you could always hire somebody to write your essay for you, but now the barriers are much lower. 12:08 One of the answers that people have come up with for that is to say, "Well, rather than doing answers, maybe you should assign the students to come up with problems." 12:18 And if they can come up with a good problem, whether they solve it using AI or solving it on their own, that exercise of coming up with a good problem, that was really the creative part, and that's what they should be judged on. 12:28 And then other people are just saying, "Well, I'm gonna have, uh, oral exams," or, "I'm gonna have discussion rather than writing." And I think writing is a really useful tool. 12:36 You know, I spend a lot of time doing it, but being able to have a discussion... 12:40 So one of the weaknesses of writing is you don't have to answer challenges and say, "Well, here's my points A, B, C, and D" and you can judge that of are those points all coherent. 12:51 But in an essay, nobody can say, "Well, what about X, Y, and Z?" But in a oral discussion you can, and so maybe more of education be focused that way. So I think there's a lot of opportunities. 13:03 Uh, but for sure we can't just stand still and do more of what we've been doing in the past. What excites you the most of the current AI era? 13:10 I mean, so it's gotta be the coding, 'cause, 'cause that's what I do day to day. And, and just seeing it, it's really only been one year that it went from 13:19 kind of an interesting, uh, possibility to a day-to-day tool for, you know, a large portion of the programmers in the world. Do you still teach to the students? Yeah, I, uh, so I still have a position at Stanford. 13:32 I'm not teaching this quarter, but I will be teaching in the winter, and I'll be teaching the, uh, human-centered AI class. Yeah, I'm just wondering how you approach this class. 13:41 So we try to make the class very interactive, right? So we shouldn't be just lectures, should be students involved. 13:48 So we want them to ask questions and be involved in, in discussions, and then we also break up the class and have them doing stuff. 13:54 Last time we taught it for the first time, uh, we said, "Okay, starting now, everyone's going to make an app." 14:01 And our class, about half of them were CS majors or other engineering majors who had lots of programming experience, but half weren't, and they were lawyers or sociologists or other fields that didn't have programming experience. 14:14 But in that one class, everybody successfully wrote an app, and some of them said, "Wow, you know, I had no idea you could do that." And so I think this next time we teach it, we'll have more of that. 14:24 So we'll say, we'll take some of the class time to say, "Okay, now today everybody's gonna build something to do this, and next week they'll build something else to do s- do another one." 14:34 And I like the idea of people doing it in teams and people doing it in an environment where others are commenting on what they're doing rather than having them go home and, and doing it on their own. How do you see AI? 14:47 Was it a positive force or there is a tension? Yeah, there's certainly a tension. I'm concerned about that. 14:54 I think that we're seeing, generally even without AI, we're seeing this push of enabling of stronger technologies to be used by smaller groups of people. 15:07 You know, it used to be we felt like we were in kind of this stable standoff. 15:12 So we had a couple superpowers, and they had nuclear missiles, and they had aircraft carriers, but they knew they couldn't really go to war against each other 'cause the outcome would be bad, and so it was stable. 15:22 But now we're seeing, well, you don't need an aircraft carrier. You, you invest a couple tens of thousand dollars in some drones, and now you can exert your power in a way that wasn't possible before. 15:34 And we're seeing more and more of that, of where smaller groups can be more powerful through technology and, and AI is adding to that, and I think that does destabilize the world in some ways, that there's more threats to worry about. 15:48 That's a very interesting comparison of two superpowers against each other and now, and DARPA was basically helping with developing AI, so it was very of military- Yeah... knowledge of. 15:59 I want to, um, sit with this thought a little bit. You see, uh, people like, uh, Jeff Hinton, one, one of the 16:06 leaders of the field, understands it too very well and, and is very concerned at this point, and I think he has a history of, of being aware of this, right? 16:15 So he left Carnegie Mellon to go to Toronto in part because he didn't like seeing all this DARPA funding of Carnegie Mellon. But thanks to it, we have internet and also- Yeah... other technologies. Yeah, yeah. 16:28 But, uh, so it's hard balance. Right. Do you have conversations with the leaders of these companies? Yeah, I've had some. Yeah. How do you see them building this, um... 16:39 Basically, I think my, my question is, do you trust them leading this? I guess I'm worried, right? 16:44 I think trust can only last so long, and maybe there's this corrupting power that, uh, even if people start out with the best of intentions and so on, sometimes they can be led down a bad path. 16:59 And so, so I do worry about this concentration of, of wealth and power. 17:03 I'm somewhat optimistic that, you know, if you had looked four years ago, you would've s- or at least I would've predicted, oh, all the attention is gonna be on the biggest frontier labs, and it's gonna be this oligarchy, and just a small number of companies are gonna be in charge of everything. 17:21 Now you're seeing DeepSeek and Kimi and all these others saying, "Well, we can have much smaller models, and they're pretty good." 17:28 So I think I'm optimistic that that means it won't be, uh, a winner take most, that it'll be m- the power will be much more widely distributed, and I, I think that's a safer place to be in. 17:39 I mean, I hope that the idea is that we're building this technology to make people's lives better, 17:46 and we should think about that and think about how we can help humanity and not just how g- how your company can get rich. Thank you, Peter. It was a huge pleasure and honor to talk to you. 17:58 And thank you the AI Conference team for organizing this interview. Dear viewers, please let me know your thoughts in the comment section. I love talking to you. 18:05 And if you enjoyed this conversation, subscribe and hit like, even hype. See you soon. [upbeat music]