WEBVTT 00:00.000 --> 00:03.720 The following is a conversation with Francois Chollet. 00:03.720 --> 00:05.760 He's the creator of Keras, 00:05.760 --> 00:08.080 which is an open source deep learning library 00:08.080 --> 00:11.480 that is designed to enable fast, user friendly experimentation 00:11.480 --> 00:13.600 with deep neural networks. 00:13.600 --> 00:16.680 It serves as an interface to several deep learning libraries, 00:16.680 --> 00:19.040 most popular of which is TensorFlow, 00:19.040 --> 00:22.600 and it was integrated into the TensorFlow main code base 00:22.600 --> 00:24.080 a while ago. 00:24.080 --> 00:27.000 Meaning, if you want to create, train, 00:27.000 --> 00:28.640 and use neural networks, 00:28.640 --> 00:31.040 probably the easiest and most popular option 00:31.040 --> 00:33.840 is to use Keras inside TensorFlow. 00:34.840 --> 00:37.240 Aside from creating an exceptionally useful 00:37.240 --> 00:38.680 and popular library, 00:38.680 --> 00:41.920 Francois is also a world class AI researcher 00:41.920 --> 00:43.680 and software engineer at Google. 00:44.560 --> 00:46.960 And he's definitely an outspoken, 00:46.960 --> 00:50.560 if not controversial personality in the AI world, 00:50.560 --> 00:52.920 especially in the realm of ideas 00:52.920 --> 00:55.920 around the future of artificial intelligence. 00:55.920 --> 00:58.600 This is the Artificial Intelligence Podcast. 00:58.600 --> 01:01.000 If you enjoy it, subscribe on YouTube, 01:01.000 --> 01:02.760 give it five stars on iTunes, 01:02.760 --> 01:04.160 support it on Patreon, 01:04.160 --> 01:06.120 or simply connect with me on Twitter 01:06.120 --> 01:09.960 at Lex Friedman, spelled F R I D M A N. 01:09.960 --> 01:13.840 And now, here's my conversation with Francois Chollet. 01:14.880 --> 01:17.320 You're known for not sugarcoating your opinions 01:17.320 --> 01:19.160 and speaking your mind about ideas in AI, 01:19.160 --> 01:21.160 especially on Twitter. 01:21.160 --> 01:22.760 It's one of my favorite Twitter accounts. 01:22.760 --> 01:26.320 So what's one of the more controversial ideas 01:26.320 --> 01:29.360 you've expressed online and gotten some heat for? 01:30.440 --> 01:31.360 How do you pick? 01:33.080 --> 01:33.920 How do I pick? 01:33.920 --> 01:36.880 Yeah, no, I think if you go through the trouble 01:36.880 --> 01:39.640 of maintaining a Twitter account, 01:39.640 --> 01:41.840 you might as well speak your mind, you know? 01:41.840 --> 01:44.600 Otherwise, what's even the point of having a Twitter account? 01:44.600 --> 01:45.480 It's like having a nice car 01:45.480 --> 01:47.560 and just leaving it in the garage. 01:48.600 --> 01:50.840 Yeah, so what's one thing for which I got 01:50.840 --> 01:53.600 a lot of pushback? 01:53.600 --> 01:56.640 Perhaps, you know, that time I wrote something 01:56.640 --> 02:00.920 about the idea of intelligence explosion, 02:00.920 --> 02:04.520 and I was questioning the idea 02:04.520 --> 02:06.840 and the reasoning behind this idea. 02:06.840 --> 02:09.640 And I got a lot of pushback on that. 02:09.640 --> 02:11.840 I got a lot of flak for it. 02:11.840 --> 02:13.600 So yeah, so intelligence explosion, 02:13.600 --> 02:14.960 I'm sure you're familiar with the idea, 02:14.960 --> 02:18.800 but it's the idea that if you were to build 02:18.800 --> 02:22.920 general AI problem solving algorithms, 02:22.920 --> 02:26.000 well, the problem of building such an AI, 02:27.480 --> 02:30.520 that itself is a problem that could be solved by your AI, 02:30.520 --> 02:31.880 and maybe it could be solved better 02:31.880 --> 02:33.760 than what humans can do. 02:33.760 --> 02:36.840 So your AI could start tweaking its own algorithm, 02:36.840 --> 02:39.520 could start making a better version of itself, 02:39.520 --> 02:43.240 and so on iteratively in a recursive fashion. 02:43.240 --> 02:47.320 And so you would end up with an AI 02:47.320 --> 02:50.080 with exponentially increasing intelligence. 02:50.080 --> 02:50.920 That's right. 02:50.920 --> 02:55.880 And I was basically questioning this idea, 02:55.880 --> 02:59.040 first of all, because the notion of intelligence explosion 02:59.040 --> 03:02.200 uses an implicit definition of intelligence 03:02.200 --> 03:05.360 that doesn't sound quite right to me. 03:05.360 --> 03:10.360 It considers intelligence as a property of a brain 03:11.200 --> 03:13.680 that you can consider in isolation, 03:13.680 --> 03:16.640 like the height of a building, for instance. 03:16.640 --> 03:19.040 But that's not really what intelligence is. 03:19.040 --> 03:22.200 Intelligence emerges from the interaction 03:22.200 --> 03:25.240 between a brain, a body, 03:25.240 --> 03:28.320 like embodied intelligence, and an environment. 03:28.320 --> 03:30.720 And if you're missing one of these pieces, 03:30.720 --> 03:33.800 then you cannot really define intelligence anymore. 03:33.800 --> 03:36.800 So just tweaking a brain to make it smaller and smaller 03:36.800 --> 03:39.120 doesn't actually make any sense to me. 03:39.120 --> 03:39.960 So first of all, 03:39.960 --> 03:43.000 you're crushing the dreams of many people, right? 03:43.000 --> 03:46.000 So there's a, let's look at like Sam Harris. 03:46.000 --> 03:48.680 Actually, a lot of physicists, Max Tegmark, 03:48.680 --> 03:52.120 people who think the universe 03:52.120 --> 03:54.640 is an information processing system, 03:54.640 --> 03:57.680 our brain is kind of an information processing system. 03:57.680 --> 03:59.400 So what's the theoretical limit? 03:59.400 --> 04:03.160 Like, it doesn't make sense that there should be some, 04:04.800 --> 04:07.520 it seems naive to think that our own brain 04:07.520 --> 04:10.000 is somehow the limit of the capabilities 04:10.000 --> 04:11.600 of this information system. 04:11.600 --> 04:13.600 I'm playing devil's advocate here. 04:13.600 --> 04:15.600 This information processing system. 04:15.600 --> 04:17.760 And then if you just scale it, 04:17.760 --> 04:19.360 if you're able to build something 04:19.360 --> 04:20.920 that's on par with the brain, 04:20.920 --> 04:24.040 you just, the process that builds it just continues 04:24.040 --> 04:26.400 and it'll improve exponentially. 04:26.400 --> 04:30.160 So that's the logic that's used actually 04:30.160 --> 04:32.560 by almost everybody 04:32.560 --> 04:36.920 that is worried about super human intelligence. 04:36.920 --> 04:39.120 So you're trying to make, 04:39.120 --> 04:40.960 so most people who are skeptical of that 04:40.960 --> 04:43.000 are kind of like, this doesn't, 04:43.000 --> 04:46.520 their thought process, this doesn't feel right. 04:46.520 --> 04:47.680 Like that's for me as well. 04:47.680 --> 04:49.760 So I'm more like, it doesn't, 04:51.440 --> 04:52.800 the whole thing is shrouded in mystery 04:52.800 --> 04:55.840 where you can't really say anything concrete, 04:55.840 --> 04:57.880 but you could say this doesn't feel right. 04:57.880 --> 05:00.640 This doesn't feel like that's how the brain works. 05:00.640 --> 05:02.400 And you're trying to with your blog posts 05:02.400 --> 05:05.680 and now making it a little more explicit. 05:05.680 --> 05:10.680 So one idea is that the brain isn't exist alone. 05:10.680 --> 05:13.200 It exists within the environment. 05:13.200 --> 05:15.680 So you can't exponentially, 05:15.680 --> 05:18.000 you would have to somehow exponentially improve 05:18.000 --> 05:20.920 the environment and the brain together almost. 05:20.920 --> 05:25.920 Yeah, in order to create something that's much smarter 05:25.960 --> 05:27.840 in some kind of, 05:27.840 --> 05:29.960 of course we don't have a definition of intelligence. 05:29.960 --> 05:31.280 That's correct, that's correct. 05:31.280 --> 05:34.280 I don't think, you should look at very smart people today, 05:34.280 --> 05:37.280 even humans, not even talking about AIs. 05:37.280 --> 05:38.640 I don't think their brain 05:38.640 --> 05:41.960 and the performance of their brain is the bottleneck 05:41.960 --> 05:45.200 to their expressed intelligence, to their achievements. 05:46.600 --> 05:49.960 You cannot just tweak one part of this system, 05:49.960 --> 05:52.840 like of this brain, body, environment system 05:52.840 --> 05:55.960 and expect that capabilities like what emerges 05:55.960 --> 06:00.280 out of this system to just explode exponentially. 06:00.280 --> 06:04.200 Because anytime you improve one part of a system 06:04.200 --> 06:06.760 with many interdependencies like this, 06:06.760 --> 06:09.520 there's a new bottleneck that arises, right? 06:09.520 --> 06:12.280 And I don't think even today for very smart people, 06:12.280 --> 06:15.000 their brain is not the bottleneck 06:15.000 --> 06:17.560 to the sort of problems they can solve, right? 06:17.560 --> 06:19.800 In fact, many very smart people today, 06:20.760 --> 06:22.520 you know, they are not actually solving 06:22.520 --> 06:24.800 any big scientific problems, they're not Einstein. 06:24.800 --> 06:28.280 They're like Einstein, but you know, the patent clerk days. 06:29.800 --> 06:31.920 Like Einstein became Einstein 06:31.920 --> 06:36.080 because this was a meeting of a genius 06:36.080 --> 06:39.480 with a big problem at the right time, right? 06:39.480 --> 06:42.480 But maybe this meeting could have never happened 06:42.480 --> 06:44.960 and then Einstein would have just been a patent clerk, right? 06:44.960 --> 06:48.400 And in fact, many people today are probably like 06:49.760 --> 06:52.240 genius level smart, but you wouldn't know 06:52.240 --> 06:54.800 because they're not really expressing any of that. 06:54.800 --> 06:55.640 Wow, that's brilliant. 06:55.640 --> 06:58.520 So we can think of the world, Earth, 06:58.520 --> 07:02.720 but also the universe as just as a space of problems. 07:02.720 --> 07:05.160 So all these problems and tasks are roaming it 07:05.160 --> 07:06.880 of various difficulty. 07:06.880 --> 07:10.120 And there's agents, creatures like ourselves 07:10.120 --> 07:13.360 and animals and so on that are also roaming it. 07:13.360 --> 07:16.480 And then you get coupled with a problem 07:16.480 --> 07:17.640 and then you solve it. 07:17.640 --> 07:19.880 But without that coupling, 07:19.880 --> 07:22.560 you can't demonstrate your quote unquote intelligence. 07:22.560 --> 07:24.480 Exactly, intelligence is the meeting 07:24.480 --> 07:27.480 of great problem solving capabilities 07:27.480 --> 07:28.760 with a great problem. 07:28.760 --> 07:30.560 And if you don't have the problem, 07:30.560 --> 07:32.280 you don't really express any intelligence. 07:32.280 --> 07:34.760 All you're left with is potential intelligence, 07:34.760 --> 07:36.240 like the performance of your brain 07:36.240 --> 07:38.680 or how high your IQ is, 07:38.680 --> 07:42.080 which in itself is just a number, right? 07:42.080 --> 07:46.520 So you mentioned problem solving capacity. 07:46.520 --> 07:47.360 Yeah. 07:47.360 --> 07:51.800 What do you think of as problem solving capacity? 07:51.800 --> 07:55.160 Can you try to define intelligence? 07:56.640 --> 08:00.000 Like what does it mean to be more or less intelligent? 08:00.000 --> 08:03.000 Is it completely coupled to a particular problem 08:03.000 --> 08:05.720 or is there something a little bit more universal? 08:05.720 --> 08:07.440 Yeah, I do believe all intelligence 08:07.440 --> 08:09.080 is specialized intelligence. 08:09.080 --> 08:12.200 Even human intelligence has some degree of generality. 08:12.200 --> 08:15.320 Well, all intelligent systems have some degree of generality 08:15.320 --> 08:19.400 but they're always specialized in one category of problems. 08:19.400 --> 08:21.880 So the human intelligence is specialized 08:21.880 --> 08:23.560 in the human experience. 08:23.560 --> 08:25.560 And that shows at various levels, 08:25.560 --> 08:30.200 that shows in some prior knowledge that's innate 08:30.200 --> 08:32.040 that we have at birth. 08:32.040 --> 08:35.360 Knowledge about things like agents, 08:35.360 --> 08:38.080 goal driven behavior, visual priors 08:38.080 --> 08:43.080 about what makes an object, priors about time and so on. 08:43.520 --> 08:45.360 That shows also in the way we learn. 08:45.360 --> 08:47.160 For instance, it's very, very easy for us 08:47.160 --> 08:48.600 to pick up language. 08:49.560 --> 08:52.080 It's very, very easy for us to learn certain things 08:52.080 --> 08:54.920 because we are basically hard coded to learn them. 08:54.920 --> 08:58.280 And we are specialized in solving certain kinds of problem 08:58.280 --> 08:59.720 and we are quite useless 08:59.720 --> 09:01.440 when it comes to other kinds of problems. 09:01.440 --> 09:06.160 For instance, we are not really designed 09:06.160 --> 09:08.800 to handle very long term problems. 09:08.800 --> 09:12.880 We have no capability of seeing the very long term. 09:12.880 --> 09:16.880 We don't have very much working memory. 09:18.000 --> 09:20.080 So how do you think about long term? 09:20.080 --> 09:21.360 Do you think long term planning, 09:21.360 --> 09:24.880 are we talking about scale of years, millennia? 09:24.880 --> 09:26.400 What do you mean by long term? 09:26.400 --> 09:28.120 We're not very good. 09:28.120 --> 09:29.760 Well, human intelligence is specialized 09:29.760 --> 09:30.720 in the human experience. 09:30.720 --> 09:32.800 And human experience is very short. 09:32.800 --> 09:34.240 One lifetime is short. 09:34.240 --> 09:35.880 Even within one lifetime, 09:35.880 --> 09:40.000 we have a very hard time envisioning things 09:40.000 --> 09:41.360 on a scale of years. 09:41.360 --> 09:43.240 It's very difficult to project yourself 09:43.240 --> 09:46.960 at a scale of five years, at a scale of 10 years and so on. 09:46.960 --> 09:50.000 We can solve only fairly narrowly scoped problems. 09:50.000 --> 09:52.320 So when it comes to solving bigger problems, 09:52.320 --> 09:53.760 larger scale problems, 09:53.760 --> 09:56.360 we are not actually doing it on an individual level. 09:56.360 --> 09:59.280 So it's not actually our brain doing it. 09:59.280 --> 10:03.040 We have this thing called civilization, right? 10:03.040 --> 10:06.600 Which is itself a sort of problem solving system, 10:06.600 --> 10:10.000 a sort of artificially intelligent system, right? 10:10.000 --> 10:12.120 And it's not running on one brain, 10:12.120 --> 10:14.080 it's running on a network of brains. 10:14.080 --> 10:15.640 In fact, it's running on much more 10:15.640 --> 10:16.760 than a network of brains. 10:16.760 --> 10:20.080 It's running on a lot of infrastructure, 10:20.080 --> 10:23.040 like books and computers and the internet 10:23.040 --> 10:25.800 and human institutions and so on. 10:25.800 --> 10:30.240 And that is capable of handling problems 10:30.240 --> 10:33.760 on a much greater scale than any individual human. 10:33.760 --> 10:37.600 If you look at computer science, for instance, 10:37.600 --> 10:39.840 that's an institution that solves problems 10:39.840 --> 10:42.560 and it is superhuman, right? 10:42.560 --> 10:44.200 It operates on a greater scale. 10:44.200 --> 10:46.880 It can solve much bigger problems 10:46.880 --> 10:49.080 than an individual human could. 10:49.080 --> 10:52.160 And science itself, science as a system, as an institution, 10:52.160 --> 10:57.120 is a kind of artificially intelligent problem solving 10:57.120 --> 10:59.360 algorithm that is superhuman. 10:59.360 --> 11:02.800 Yeah, it's, at least computer science 11:02.800 --> 11:07.720 is like a theorem prover at a scale of thousands, 11:07.720 --> 11:10.400 maybe hundreds of thousands of human beings. 11:10.400 --> 11:14.680 At that scale, what do you think is an intelligent agent? 11:14.680 --> 11:18.280 So there's us humans at the individual level, 11:18.280 --> 11:22.400 there is millions, maybe billions of bacteria in our skin. 11:23.880 --> 11:26.400 There is, that's at the smaller scale. 11:26.400 --> 11:29.160 You can even go to the particle level 11:29.160 --> 11:31.000 as systems that behave, 11:31.840 --> 11:34.360 you can say intelligently in some ways. 11:35.440 --> 11:37.840 And then you can look at the earth as a single organism, 11:37.840 --> 11:39.200 you can look at our galaxy 11:39.200 --> 11:42.160 and even the universe as a single organism. 11:42.160 --> 11:44.680 Do you think, how do you think about scale 11:44.680 --> 11:46.280 in defining intelligent systems? 11:46.280 --> 11:50.440 And we're here at Google, there is millions of devices 11:50.440 --> 11:53.360 doing computation just in a distributed way. 11:53.360 --> 11:55.880 How do you think about intelligence versus scale? 11:55.880 --> 11:59.400 You can always characterize anything as a system. 12:00.640 --> 12:03.600 I think people who talk about things 12:03.600 --> 12:05.320 like intelligence explosion, 12:05.320 --> 12:08.760 tend to focus on one agent is basically one brain, 12:08.760 --> 12:10.960 like one brain considered in isolation, 12:10.960 --> 12:13.200 like a brain, a jaw that's controlling a body 12:13.200 --> 12:16.280 in a very like top to bottom kind of fashion. 12:16.280 --> 12:19.480 And that body is pursuing goals into an environment. 12:19.480 --> 12:20.720 So it's a very hierarchical view. 12:20.720 --> 12:22.880 You have the brain at the top of the pyramid, 12:22.880 --> 12:25.960 then you have the body just plainly receiving orders. 12:25.960 --> 12:27.640 And then the body is manipulating objects 12:27.640 --> 12:28.920 in the environment and so on. 12:28.920 --> 12:32.920 So everything is subordinate to this one thing, 12:32.920 --> 12:34.720 this epicenter, which is the brain. 12:34.720 --> 12:37.120 But in real life, intelligent agents 12:37.120 --> 12:39.240 don't really work like this, right? 12:39.240 --> 12:40.920 There is no strong delimitation 12:40.920 --> 12:43.400 between the brain and the body to start with. 12:43.400 --> 12:45.000 You have to look not just at the brain, 12:45.000 --> 12:46.560 but at the nervous system. 12:46.560 --> 12:48.840 But then the nervous system and the body 12:48.840 --> 12:50.760 are naturally two separate entities. 12:50.760 --> 12:53.960 So you have to look at an entire animal as one agent. 12:53.960 --> 12:57.000 But then you start realizing as you observe an animal 12:57.000 --> 13:00.200 over any length of time, 13:00.200 --> 13:03.160 that a lot of the intelligence of an animal 13:03.160 --> 13:04.600 is actually externalized. 13:04.600 --> 13:06.240 That's especially true for humans. 13:06.240 --> 13:08.880 A lot of our intelligence is externalized. 13:08.880 --> 13:10.360 When you write down some notes, 13:10.360 --> 13:11.960 that is externalized intelligence. 13:11.960 --> 13:14.000 When you write a computer program, 13:14.000 --> 13:16.000 you are externalizing cognition. 13:16.000 --> 13:19.720 So it's externalizing books, it's externalized in computers, 13:19.720 --> 13:21.520 the internet, in other humans. 13:23.080 --> 13:25.400 It's externalizing language and so on. 13:25.400 --> 13:30.400 So there is no hard delimitation 13:30.480 --> 13:32.640 of what makes an intelligent agent. 13:32.640 --> 13:33.880 It's all about context. 13:34.960 --> 13:38.800 Okay, but AlphaGo is better at Go 13:38.800 --> 13:40.200 than the best human player. 13:42.520 --> 13:45.000 There's levels of skill here. 13:45.000 --> 13:48.600 So do you think there's such a ability, 13:48.600 --> 13:52.800 such a concept as intelligence explosion 13:52.800 --> 13:54.760 in a specific task? 13:54.760 --> 13:57.360 And then, well, yeah. 13:57.360 --> 14:00.120 Do you think it's possible to have a category of tasks 14:00.120 --> 14:02.080 on which you do have something 14:02.080 --> 14:05.040 like an exponential growth of ability 14:05.040 --> 14:07.440 to solve that particular problem? 14:07.440 --> 14:10.320 I think if you consider a specific vertical, 14:10.320 --> 14:13.720 it's probably possible to some extent. 14:15.320 --> 14:18.320 I also don't think we have to speculate about it 14:18.320 --> 14:22.280 because we have real world examples 14:22.280 --> 14:26.920 of recursively self improving intelligent systems, right? 14:26.920 --> 14:30.920 So for instance, science is a problem solving system, 14:30.920 --> 14:32.600 a knowledge generation system, 14:32.600 --> 14:36.240 like a system that experiences the world in some sense 14:36.240 --> 14:40.160 and then gradually understands it and can act on it. 14:40.160 --> 14:42.120 And that system is superhuman 14:42.120 --> 14:45.600 and it is clearly recursively self improving 14:45.600 --> 14:47.560 because science feeds into technology. 14:47.560 --> 14:50.200 Technology can be used to build better tools, 14:50.200 --> 14:52.880 better computers, better instrumentation and so on, 14:52.880 --> 14:56.720 which in turn can make science faster, right? 14:56.720 --> 15:00.560 So science is probably the closest thing we have today 15:00.560 --> 15:04.760 to a recursively self improving superhuman AI. 15:04.760 --> 15:08.520 And you can just observe is science, 15:08.520 --> 15:10.320 is scientific progress to the exploding, 15:10.320 --> 15:12.800 which itself is an interesting question. 15:12.800 --> 15:15.560 You can use that as a basis to try to understand 15:15.560 --> 15:17.920 what will happen with a superhuman AI 15:17.920 --> 15:21.000 that has a science like behavior. 15:21.000 --> 15:23.320 Let me linger on it a little bit more. 15:23.320 --> 15:27.600 What is your intuition why an intelligence explosion 15:27.600 --> 15:28.560 is not possible? 15:28.560 --> 15:30.920 Like taking the scientific, 15:30.920 --> 15:33.240 all the semi scientific revolutions, 15:33.240 --> 15:38.080 why can't we slightly accelerate that process? 15:38.080 --> 15:41.200 So you can absolutely accelerate 15:41.200 --> 15:43.120 any problem solving process. 15:43.120 --> 15:46.720 So a recursively self improvement 15:46.720 --> 15:48.640 is absolutely a real thing. 15:48.640 --> 15:51.880 But what happens with a recursively self improving system 15:51.880 --> 15:53.680 is typically not explosion 15:53.680 --> 15:56.520 because no system exists in isolation. 15:56.520 --> 15:58.640 And so tweaking one part of the system 15:58.640 --> 16:00.880 means that suddenly another part of the system 16:00.880 --> 16:02.200 becomes a bottleneck. 16:02.200 --> 16:03.800 And if you look at science, for instance, 16:03.800 --> 16:06.800 which is clearly a recursively self improving, 16:06.800 --> 16:09.040 clearly a problem solving system, 16:09.040 --> 16:12.000 scientific progress is not actually exploding. 16:12.000 --> 16:13.520 If you look at science, 16:13.520 --> 16:16.480 what you see is the picture of a system 16:16.480 --> 16:19.240 that is consuming an exponentially increasing 16:19.240 --> 16:20.520 amount of resources, 16:20.520 --> 16:23.960 but it's having a linear output 16:23.960 --> 16:26.000 in terms of scientific progress. 16:26.000 --> 16:28.960 And maybe that will seem like a very strong claim. 16:28.960 --> 16:31.160 Many people are actually saying that, 16:31.160 --> 16:34.560 scientific progress is exponential, 16:34.560 --> 16:36.120 but when they're claiming this, 16:36.120 --> 16:38.400 they're actually looking at indicators 16:38.400 --> 16:43.080 of resource consumption by science. 16:43.080 --> 16:46.680 For instance, the number of papers being published, 16:47.560 --> 16:49.960 the number of patents being filed and so on, 16:49.960 --> 16:53.600 which are just completely correlated 16:53.600 --> 16:58.480 with how many people are working on science today. 16:58.480 --> 17:00.640 So it's actually an indicator of resource consumption, 17:00.640 --> 17:03.200 but what you should look at is the output, 17:03.200 --> 17:06.680 is progress in terms of the knowledge 17:06.680 --> 17:08.040 that science generates, 17:08.040 --> 17:10.640 in terms of the scope and significance 17:10.640 --> 17:12.520 of the problems that we solve. 17:12.520 --> 17:16.720 And some people have actually been trying to measure that. 17:16.720 --> 17:20.160 Like Michael Nielsen, for instance, 17:20.160 --> 17:21.920 he had a very nice paper, 17:21.920 --> 17:23.720 I think that was last year about it. 17:25.200 --> 17:28.360 So his approach to measure scientific progress 17:28.360 --> 17:33.360 was to look at the timeline of scientific discoveries 17:33.720 --> 17:37.160 over the past, you know, 100, 150 years. 17:37.160 --> 17:41.360 And for each major discovery, 17:41.360 --> 17:44.360 ask a panel of experts to rate 17:44.360 --> 17:46.760 the significance of the discovery. 17:46.760 --> 17:49.600 And if the output of science as an institution 17:49.600 --> 17:50.440 were exponential, 17:50.440 --> 17:55.440 you would expect the temporal density of significance 17:56.600 --> 17:58.160 to go up exponentially. 17:58.160 --> 18:00.960 Maybe because there's a faster rate of discoveries, 18:00.960 --> 18:02.960 maybe because the discoveries are, you know, 18:02.960 --> 18:04.920 increasingly more important. 18:04.920 --> 18:06.800 And what actually happens 18:06.800 --> 18:10.040 if you plot this temporal density of significance 18:10.040 --> 18:11.320 measured in this way, 18:11.320 --> 18:14.520 is that you see very much a flat graph. 18:14.520 --> 18:16.600 You see a flat graph across all disciplines, 18:16.600 --> 18:19.720 across physics, biology, medicine, and so on. 18:19.720 --> 18:22.480 And it actually makes a lot of sense 18:22.480 --> 18:23.320 if you think about it, 18:23.320 --> 18:26.000 because think about the progress of physics 18:26.000 --> 18:28.000 110 years ago, right? 18:28.000 --> 18:30.080 It was a time of crazy change. 18:30.080 --> 18:31.960 Think about the progress of technology, 18:31.960 --> 18:34.360 you know, 170 years ago, 18:34.360 --> 18:35.400 when we started having, you know, 18:35.400 --> 18:37.560 replacing horses with cars, 18:37.560 --> 18:40.000 when we started having electricity and so on. 18:40.000 --> 18:41.520 It was a time of incredible change. 18:41.520 --> 18:44.600 And today is also a time of very, very fast change, 18:44.600 --> 18:48.040 but it would be an unfair characterization 18:48.040 --> 18:50.560 to say that today technology and science 18:50.560 --> 18:52.920 are moving way faster than they did 50 years ago 18:52.920 --> 18:54.360 or 100 years ago. 18:54.360 --> 18:59.360 And if you do try to rigorously plot 18:59.520 --> 19:04.520 the temporal density of the significance, 19:04.880 --> 19:07.320 yeah, of significance, sorry, 19:07.320 --> 19:09.720 you do see very flat curves. 19:09.720 --> 19:12.040 And you can check out the paper 19:12.040 --> 19:16.000 that Michael Nielsen had about this idea. 19:16.000 --> 19:20.000 And so the way I interpret it is, 19:20.000 --> 19:24.160 as you make progress in a given field, 19:24.160 --> 19:26.120 or in a given subfield of science, 19:26.120 --> 19:28.680 it becomes exponentially more difficult 19:28.680 --> 19:30.440 to make further progress. 19:30.440 --> 19:35.000 Like the very first person to work on information theory. 19:35.000 --> 19:36.440 If you enter a new field, 19:36.440 --> 19:37.920 and it's still the very early years, 19:37.920 --> 19:41.160 there's a lot of low hanging fruit you can pick. 19:41.160 --> 19:42.000 That's right, yeah. 19:42.000 --> 19:43.960 But the next generation of researchers 19:43.960 --> 19:48.160 is gonna have to dig much harder, actually, 19:48.160 --> 19:50.640 to make smaller discoveries, 19:50.640 --> 19:52.640 probably larger number of smaller discoveries, 19:52.640 --> 19:54.640 and to achieve the same amount of impact, 19:54.640 --> 19:57.480 you're gonna need a much greater head count. 19:57.480 --> 20:00.040 And that's exactly the picture you're seeing with science, 20:00.040 --> 20:03.760 that the number of scientists and engineers 20:03.760 --> 20:06.520 is in fact increasing exponentially. 20:06.520 --> 20:08.400 The amount of computational resources 20:08.400 --> 20:10.040 that are available to science 20:10.040 --> 20:11.880 is increasing exponentially and so on. 20:11.880 --> 20:15.560 So the resource consumption of science is exponential, 20:15.560 --> 20:18.200 but the output in terms of progress, 20:18.200 --> 20:21.000 in terms of significance, is linear. 20:21.000 --> 20:23.120 And the reason why is because, 20:23.120 --> 20:26.000 and even though science is regressively self improving, 20:26.000 --> 20:28.440 meaning that scientific progress 20:28.440 --> 20:30.240 turns into technological progress, 20:30.240 --> 20:32.960 which in turn helps science. 20:32.960 --> 20:35.280 If you look at computers, for instance, 20:35.280 --> 20:38.480 our products of science and computers 20:38.480 --> 20:41.560 are tremendously useful in speeding up science. 20:41.560 --> 20:43.840 The internet, same thing, the internet is a technology 20:43.840 --> 20:47.480 that's made possible by very recent scientific advances. 20:47.480 --> 20:52.400 And itself, because it enables scientists to network, 20:52.400 --> 20:55.520 to communicate, to exchange papers and ideas much faster, 20:55.520 --> 20:57.440 it is a way to speed up scientific progress. 20:57.440 --> 20:58.440 So even though you're looking 20:58.440 --> 21:01.440 at a regressively self improving system, 21:01.440 --> 21:04.080 it is consuming exponentially more resources 21:04.080 --> 21:09.080 to produce the same amount of problem solving, very much. 21:09.200 --> 21:11.080 So that's a fascinating way to paint it, 21:11.080 --> 21:14.960 and certainly that holds for the deep learning community. 21:14.960 --> 21:18.120 If you look at the temporal, what did you call it, 21:18.120 --> 21:21.240 the temporal density of significant ideas, 21:21.240 --> 21:23.920 if you look at in deep learning, 21:24.840 --> 21:26.960 I think, I'd have to think about that, 21:26.960 --> 21:29.040 but if you really look at significant ideas 21:29.040 --> 21:32.400 in deep learning, they might even be decreasing. 21:32.400 --> 21:37.400 So I do believe the per paper significance is decreasing, 21:39.600 --> 21:41.240 but the amount of papers 21:41.240 --> 21:43.440 is still today exponentially increasing. 21:43.440 --> 21:45.880 So I think if you look at an aggregate, 21:45.880 --> 21:48.840 my guess is that you would see a linear progress. 21:48.840 --> 21:53.840 If you were to sum the significance of all papers, 21:56.120 --> 21:58.640 you would see roughly in your progress. 21:58.640 --> 22:03.640 And in my opinion, it is not a coincidence 22:03.880 --> 22:05.800 that you're seeing linear progress in science 22:05.800 --> 22:07.720 despite exponential resource consumption. 22:07.720 --> 22:10.280 I think the resource consumption 22:10.280 --> 22:15.280 is dynamically adjusting itself to maintain linear progress 22:15.880 --> 22:18.560 because we as a community expect linear progress, 22:18.560 --> 22:21.240 meaning that if we start investing less 22:21.240 --> 22:23.600 and seeing less progress, it means that suddenly 22:23.600 --> 22:26.800 there are some lower hanging fruits that become available 22:26.800 --> 22:31.280 and someone's gonna step up and pick them, right? 22:31.280 --> 22:36.280 So it's very much like a market for discoveries and ideas. 22:36.920 --> 22:38.720 But there's another fundamental part 22:38.720 --> 22:41.800 which you're highlighting, which as a hypothesis 22:41.800 --> 22:45.160 as science or like the space of ideas, 22:45.160 --> 22:48.160 any one path you travel down, 22:48.160 --> 22:51.080 it gets exponentially more difficult 22:51.080 --> 22:54.720 to get a new way to develop new ideas. 22:54.720 --> 22:57.640 And your sense is that's gonna hold 22:57.640 --> 23:01.520 across our mysterious universe. 23:01.520 --> 23:03.360 Yes, well, exponential progress 23:03.360 --> 23:05.480 triggers exponential friction. 23:05.480 --> 23:07.440 So that if you tweak one part of the system, 23:07.440 --> 23:10.680 suddenly some other part becomes a bottleneck, right? 23:10.680 --> 23:14.880 For instance, let's say you develop some device 23:14.880 --> 23:17.160 that measures its own acceleration 23:17.160 --> 23:18.720 and then it has some engine 23:18.720 --> 23:20.800 and it outputs even more acceleration 23:20.800 --> 23:22.360 in proportion of its own acceleration 23:22.360 --> 23:23.320 and you drop it somewhere, 23:23.320 --> 23:25.240 it's not gonna reach infinite speed 23:25.240 --> 23:27.880 because it exists in a certain context. 23:29.080 --> 23:31.000 So the air around it is gonna generate friction 23:31.000 --> 23:34.320 and it's gonna block it at some top speed. 23:34.320 --> 23:37.480 And even if you were to consider the broader context 23:37.480 --> 23:39.840 and lift the bottleneck there, 23:39.840 --> 23:42.240 like the bottleneck of friction, 23:43.120 --> 23:45.120 then some other part of the system 23:45.120 --> 23:48.120 would start stepping in and creating exponential friction, 23:48.120 --> 23:49.920 maybe the speed of flight or whatever. 23:49.920 --> 23:51.920 And this definitely holds true 23:51.920 --> 23:54.960 when you look at the problem solving algorithm 23:54.960 --> 23:58.160 that is being run by science as an institution, 23:58.160 --> 23:59.720 science as a system. 23:59.720 --> 24:01.720 As you make more and more progress, 24:01.720 --> 24:05.800 despite having this recursive self improvement component, 24:06.760 --> 24:09.840 you are encountering exponential friction. 24:09.840 --> 24:13.480 The more researchers you have working on different ideas, 24:13.480 --> 24:14.880 the more overhead you have 24:14.880 --> 24:18.040 in terms of communication across researchers. 24:18.040 --> 24:22.920 If you look at, you were mentioning quantum mechanics, right? 24:22.920 --> 24:26.880 Well, if you want to start making significant discoveries 24:26.880 --> 24:29.680 today, significant progress in quantum mechanics, 24:29.680 --> 24:33.000 there is an amount of knowledge you have to ingest, 24:33.000 --> 24:34.080 which is huge. 24:34.080 --> 24:36.520 So there's a very large overhead 24:36.520 --> 24:39.240 to even start to contribute. 24:39.240 --> 24:40.680 There's a large amount of overhead 24:40.680 --> 24:44.040 to synchronize across researchers and so on. 24:44.040 --> 24:47.440 And of course, the significant practical experiments 24:48.600 --> 24:52.160 are going to require exponentially expensive equipment 24:52.160 --> 24:56.480 because the easier ones have already been run, right? 24:56.480 --> 25:00.480 So in your senses, there's no way escaping, 25:00.480 --> 25:04.480 there's no way of escaping this kind of friction 25:04.480 --> 25:08.600 with artificial intelligence systems. 25:08.600 --> 25:11.520 Yeah, no, I think science is a very good way 25:11.520 --> 25:14.280 to model what would happen with a superhuman 25:14.280 --> 25:16.440 recursive research improving AI. 25:16.440 --> 25:18.240 That's your sense, I mean, the... 25:18.240 --> 25:19.680 That's my intuition. 25:19.680 --> 25:23.400 It's not like a mathematical proof of anything. 25:23.400 --> 25:24.400 That's not my point. 25:24.400 --> 25:26.600 Like, I'm not trying to prove anything. 25:26.600 --> 25:27.920 I'm just trying to make an argument 25:27.920 --> 25:31.160 to question the narrative of intelligence explosion, 25:31.160 --> 25:32.880 which is quite a dominant narrative. 25:32.880 --> 25:35.840 And you do get a lot of pushback if you go against it. 25:35.840 --> 25:39.320 Because, so for many people, right, 25:39.320 --> 25:42.200 AI is not just a subfield of computer science. 25:42.200 --> 25:44.120 It's more like a belief system. 25:44.120 --> 25:48.640 Like this belief that the world is headed towards an event, 25:48.640 --> 25:55.040 the singularity, past which, you know, AI will become... 25:55.040 --> 25:57.080 will go exponential very much, 25:57.080 --> 25:58.600 and the world will be transformed, 25:58.600 --> 26:00.840 and humans will become obsolete. 26:00.840 --> 26:03.880 And if you go against this narrative, 26:03.880 --> 26:06.920 because it is not really a scientific argument, 26:06.920 --> 26:08.880 but more of a belief system, 26:08.880 --> 26:11.240 it is part of the identity of many people. 26:11.240 --> 26:12.600 If you go against this narrative, 26:12.600 --> 26:14.400 it's like you're attacking the identity 26:14.400 --> 26:15.560 of people who believe in it. 26:15.560 --> 26:17.640 It's almost like saying God doesn't exist, 26:17.640 --> 26:19.000 or something. 26:19.000 --> 26:21.880 So you do get a lot of pushback 26:21.880 --> 26:24.040 if you try to question these ideas. 26:24.040 --> 26:26.520 First of all, I believe most people, 26:26.520 --> 26:29.240 they might not be as eloquent or explicit as you're being, 26:29.240 --> 26:30.920 but most people in computer science 26:30.920 --> 26:33.000 are most people who actually have built 26:33.000 --> 26:36.360 anything that you could call AI, quote, unquote, 26:36.360 --> 26:38.080 would agree with you. 26:38.080 --> 26:40.560 They might not be describing in the same kind of way. 26:40.560 --> 26:43.960 It's more, so the pushback you're getting 26:43.960 --> 26:48.080 is from people who get attached to the narrative 26:48.080 --> 26:51.000 from, not from a place of science, 26:51.000 --> 26:53.400 but from a place of imagination. 26:53.400 --> 26:54.760 That's correct, that's correct. 26:54.760 --> 26:56.920 So why do you think that's so appealing? 26:56.920 --> 27:01.920 Because the usual dreams that people have 27:02.120 --> 27:03.960 when you create a superintelligence system 27:03.960 --> 27:05.120 past the singularity, 27:05.120 --> 27:08.600 that what people imagine is somehow always destructive. 27:09.440 --> 27:12.240 Do you have, if you were put on your psychology hat, 27:12.240 --> 27:17.240 what's, why is it so appealing to imagine 27:17.400 --> 27:20.760 the ways that all of human civilization will be destroyed? 27:20.760 --> 27:22.080 I think it's a good story. 27:22.080 --> 27:23.120 You know, it's a good story. 27:23.120 --> 27:28.120 And very interestingly, it mirrors a religious stories, 27:28.160 --> 27:30.560 right, religious mythology. 27:30.560 --> 27:34.360 If you look at the mythology of most civilizations, 27:34.360 --> 27:38.280 it's about the world being headed towards some final events 27:38.280 --> 27:40.480 in which the world will be destroyed 27:40.480 --> 27:42.800 and some new world order will arise 27:42.800 --> 27:44.920 that will be mostly spiritual, 27:44.920 --> 27:49.400 like the apocalypse followed by a paradise probably, right? 27:49.400 --> 27:52.600 It's a very appealing story on a fundamental level. 27:52.600 --> 27:54.560 And we all need stories. 27:54.560 --> 27:58.160 We all need stories to structure the way we see the world, 27:58.160 --> 27:59.960 especially at timescales 27:59.960 --> 28:04.520 that are beyond our ability to make predictions, right? 28:04.520 --> 28:08.840 So on a more serious non exponential explosion, 28:08.840 --> 28:13.840 question, do you think there will be a time 28:15.000 --> 28:19.800 when we'll create something like human level intelligence 28:19.800 --> 28:23.800 or intelligent systems that will make you sit back 28:23.800 --> 28:28.520 and be just surprised at damn how smart this thing is? 28:28.520 --> 28:30.160 That doesn't require exponential growth 28:30.160 --> 28:32.120 or an exponential improvement, 28:32.120 --> 28:35.600 but what's your sense of the timeline and so on 28:35.600 --> 28:40.600 that you'll be really surprised at certain capabilities? 28:41.080 --> 28:42.560 And we'll talk about limitations and deep learning. 28:42.560 --> 28:44.480 So do you think in your lifetime, 28:44.480 --> 28:46.600 you'll be really damn surprised? 28:46.600 --> 28:51.440 Around 2013, 2014, I was many times surprised 28:51.440 --> 28:53.960 by the capabilities of deep learning actually. 28:53.960 --> 28:55.920 That was before we had assessed exactly 28:55.920 --> 28:57.880 what deep learning could do and could not do. 28:57.880 --> 29:00.600 And it felt like a time of immense potential. 29:00.600 --> 29:03.080 And then we started narrowing it down, 29:03.080 --> 29:04.360 but I was very surprised. 29:04.360 --> 29:07.120 I would say it has already happened. 29:07.120 --> 29:10.800 Was there a moment, there must've been a day in there 29:10.800 --> 29:14.360 where your surprise was almost bordering 29:14.360 --> 29:19.360 on the belief of the narrative that we just discussed. 29:19.440 --> 29:20.800 Was there a moment, 29:20.800 --> 29:22.400 because you've written quite eloquently 29:22.400 --> 29:23.960 about the limits of deep learning, 29:23.960 --> 29:25.760 was there a moment that you thought 29:25.760 --> 29:27.720 that maybe deep learning is limitless? 29:30.000 --> 29:32.400 No, I don't think I've ever believed this. 29:32.400 --> 29:35.560 What was really shocking is that it worked. 29:35.560 --> 29:37.640 It worked at all, yeah. 29:37.640 --> 29:40.520 But there's a big jump between being able 29:40.520 --> 29:43.400 to do really good computer vision 29:43.400 --> 29:44.920 and human level intelligence. 29:44.920 --> 29:49.520 So I don't think at any point I wasn't under the impression 29:49.520 --> 29:51.280 that the results we got in computer vision 29:51.280 --> 29:54.080 meant that we were very close to human level intelligence. 29:54.080 --> 29:56.040 I don't think we're very close to human level intelligence. 29:56.040 --> 29:58.520 I do believe that there's no reason 29:58.520 --> 30:01.760 why we won't achieve it at some point. 30:01.760 --> 30:06.400 I also believe that it's the problem 30:06.400 --> 30:08.560 with talking about human level intelligence 30:08.560 --> 30:11.240 that implicitly you're considering 30:11.240 --> 30:14.360 like an axis of intelligence with different levels, 30:14.360 --> 30:16.720 but that's not really how intelligence works. 30:16.720 --> 30:19.480 Intelligence is very multi dimensional. 30:19.480 --> 30:22.480 And so there's the question of capabilities, 30:22.480 --> 30:25.560 but there's also the question of being human like, 30:25.560 --> 30:27.040 and it's two very different things. 30:27.040 --> 30:28.280 Like you can build potentially 30:28.280 --> 30:30.640 very advanced intelligent agents 30:30.640 --> 30:32.640 that are not human like at all. 30:32.640 --> 30:35.240 And you can also build very human like agents. 30:35.240 --> 30:37.840 And these are two very different things, right? 30:37.840 --> 30:38.760 Right. 30:38.760 --> 30:42.240 Let's go from the philosophical to the practical. 30:42.240 --> 30:44.240 Can you give me a history of Keras 30:44.240 --> 30:46.440 and all the major deep learning frameworks 30:46.440 --> 30:48.480 that you kind of remember in relation to Keras 30:48.480 --> 30:52.040 and in general, TensorFlow, Theano, the old days. 30:52.040 --> 30:55.400 Can you give a brief overview Wikipedia style history 30:55.400 --> 30:59.120 and your role in it before we return to AGI discussions? 30:59.120 --> 31:00.720 Yeah, that's a broad topic. 31:00.720 --> 31:04.040 So I started working on Keras. 31:04.920 --> 31:06.240 It was the name Keras at the time. 31:06.240 --> 31:08.320 I actually picked the name like 31:08.320 --> 31:10.200 just the day I was going to release it. 31:10.200 --> 31:14.800 So I started working on it in February, 2015. 31:14.800 --> 31:17.240 And so at the time there weren't too many people 31:17.240 --> 31:20.320 working on deep learning, maybe like fewer than 10,000. 31:20.320 --> 31:22.840 The software tooling was not really developed. 31:25.320 --> 31:28.800 So the main deep learning library was Cafe, 31:28.800 --> 31:30.840 which was mostly C++. 31:30.840 --> 31:32.760 Why do you say Cafe was the main one? 31:32.760 --> 31:36.000 Cafe was vastly more popular than Theano 31:36.000 --> 31:38.920 in late 2014, early 2015. 31:38.920 --> 31:42.400 Cafe was the one library that everyone was using 31:42.400 --> 31:43.400 for computer vision. 31:43.400 --> 31:46.120 And computer vision was the most popular problem 31:46.120 --> 31:46.960 in deep learning at the time. 31:46.960 --> 31:47.800 Absolutely. 31:47.800 --> 31:50.440 Like ConvNets was like the subfield of deep learning 31:50.440 --> 31:53.160 that everyone was working on. 31:53.160 --> 31:57.680 So myself, so in late 2014, 31:57.680 --> 32:00.600 I was actually interested in RNNs, 32:00.600 --> 32:01.760 in recurrent neural networks, 32:01.760 --> 32:05.800 which was a very niche topic at the time, right? 32:05.800 --> 32:08.640 It really took off around 2016. 32:08.640 --> 32:11.080 And so I was looking for good tools. 32:11.080 --> 32:14.800 I had used Torch 7, I had used Theano, 32:14.800 --> 32:17.640 used Theano a lot in Kaggle competitions. 32:19.320 --> 32:20.840 I had used Cafe. 32:20.840 --> 32:25.840 And there was no like good solution for RNNs at the time. 32:25.840 --> 32:28.640 Like there was no reusable open source implementation 32:28.640 --> 32:30.000 of an LSTM, for instance. 32:30.000 --> 32:32.920 So I decided to build my own. 32:32.920 --> 32:35.440 And at first, the pitch for that was, 32:35.440 --> 32:39.960 it was gonna be mostly around LSTM recurrent neural networks. 32:39.960 --> 32:41.360 It was gonna be in Python. 32:42.280 --> 32:44.280 An important decision at the time 32:44.280 --> 32:45.440 that was kind of not obvious 32:45.440 --> 32:50.360 is that the models would be defined via Python code, 32:50.360 --> 32:54.400 which was kind of like going against the mainstream 32:54.400 --> 32:58.000 at the time because Cafe, Pylon 2, and so on, 32:58.000 --> 33:00.600 like all the big libraries were actually going 33:00.600 --> 33:03.520 with the approach of setting configuration files 33:03.520 --> 33:05.560 in YAML to define models. 33:05.560 --> 33:08.840 So some libraries were using code to define models, 33:08.840 --> 33:12.280 like Torch 7, obviously, but that was not Python. 33:12.280 --> 33:16.680 Lasagne was like a Theano based very early library 33:16.680 --> 33:18.640 that was, I think, developed, I don't remember exactly, 33:18.640 --> 33:20.240 probably late 2014. 33:20.240 --> 33:21.200 It's Python as well. 33:21.200 --> 33:22.040 It's Python as well. 33:22.040 --> 33:24.320 It was like on top of Theano. 33:24.320 --> 33:28.320 And so I started working on something 33:29.480 --> 33:32.520 and the value proposition at the time was that 33:32.520 --> 33:36.240 not only what I think was the first 33:36.240 --> 33:38.800 reusable open source implementation of LSTM, 33:40.400 --> 33:44.440 you could combine RNNs and covenants 33:44.440 --> 33:45.440 with the same library, 33:45.440 --> 33:46.920 which is not really possible before, 33:46.920 --> 33:49.080 like Cafe was only doing covenants. 33:50.440 --> 33:52.560 And it was kind of easy to use 33:52.560 --> 33:54.440 because, so before I was using Theano, 33:54.440 --> 33:55.680 I was actually using scikitlin 33:55.680 --> 33:58.320 and I loved scikitlin for its usability. 33:58.320 --> 34:01.560 So I drew a lot of inspiration from scikitlin 34:01.560 --> 34:02.400 when I made Keras. 34:02.400 --> 34:05.600 It's almost like scikitlin for neural networks. 34:05.600 --> 34:06.680 The fit function. 34:06.680 --> 34:07.960 Exactly, the fit function, 34:07.960 --> 34:10.800 like reducing a complex string loop 34:10.800 --> 34:12.880 to a single function call, right? 34:12.880 --> 34:14.880 And of course, some people will say, 34:14.880 --> 34:16.320 this is hiding a lot of details, 34:16.320 --> 34:18.680 but that's exactly the point, right? 34:18.680 --> 34:20.280 The magic is the point. 34:20.280 --> 34:22.680 So it's magical, but in a good way. 34:22.680 --> 34:24.960 It's magical in the sense that it's delightful. 34:24.960 --> 34:26.160 Yeah, yeah. 34:26.160 --> 34:27.640 I'm actually quite surprised. 34:27.640 --> 34:29.600 I didn't know that it was born out of desire 34:29.600 --> 34:32.480 to implement RNNs and LSTMs. 34:32.480 --> 34:33.320 It was. 34:33.320 --> 34:34.160 That's fascinating. 34:34.160 --> 34:36.040 So you were actually one of the first people 34:36.040 --> 34:37.960 to really try to attempt 34:37.960 --> 34:41.000 to get the major architectures together. 34:41.000 --> 34:42.760 And it's also interesting. 34:42.760 --> 34:45.160 You made me realize that that was a design decision at all 34:45.160 --> 34:47.360 is defining the model and code. 34:47.360 --> 34:49.920 Just, I'm putting myself in your shoes, 34:49.920 --> 34:53.200 whether the YAML, especially if cafe was the most popular. 34:53.200 --> 34:54.720 It was the most popular by far. 34:54.720 --> 34:58.480 If I was, if I were, yeah, I don't, 34:58.480 --> 34:59.560 I didn't like the YAML thing, 34:59.560 --> 35:02.840 but it makes more sense that you will put 35:02.840 --> 35:05.720 in a configuration file, the definition of a model. 35:05.720 --> 35:07.200 That's an interesting gutsy move 35:07.200 --> 35:10.040 to stick with defining it in code. 35:10.040 --> 35:11.600 Just if you look back. 35:11.600 --> 35:13.480 Other libraries were doing it as well, 35:13.480 --> 35:16.320 but it was definitely the more niche option. 35:16.320 --> 35:17.160 Yeah. 35:17.160 --> 35:18.360 Okay, Keras and then. 35:18.360 --> 35:21.520 So I released Keras in March, 2015, 35:21.520 --> 35:24.160 and it got users pretty much from the start. 35:24.160 --> 35:25.800 So the deep learning community was very, very small 35:25.800 --> 35:27.240 at the time. 35:27.240 --> 35:30.600 Lots of people were starting to be interested in LSTM. 35:30.600 --> 35:32.440 So it was gonna release it at the right time 35:32.440 --> 35:35.560 because it was offering an easy to use LSTM implementation. 35:35.560 --> 35:37.680 Exactly at the time where lots of people started 35:37.680 --> 35:42.280 to be intrigued by the capabilities of RNN, RNNs for NLP. 35:42.280 --> 35:43.920 So it grew from there. 35:43.920 --> 35:48.920 Then I joined Google about six months later, 35:51.480 --> 35:54.920 and that was actually completely unrelated to Keras. 35:54.920 --> 35:57.080 So I actually joined a research team 35:57.080 --> 35:59.520 working on image classification, 35:59.520 --> 36:00.680 mostly like computer vision. 36:00.680 --> 36:02.320 So I was doing computer vision research 36:02.320 --> 36:03.640 at Google initially. 36:03.640 --> 36:05.520 And immediately when I joined Google, 36:05.520 --> 36:10.520 I was exposed to the early internal version of TensorFlow. 36:10.520 --> 36:13.920 And the way it appeared to me at the time, 36:13.920 --> 36:15.720 and it was definitely the way it was at the time 36:15.720 --> 36:20.760 is that this was an improved version of Theano. 36:20.760 --> 36:24.720 So I immediately knew I had to port Keras 36:24.720 --> 36:26.800 to this new TensorFlow thing. 36:26.800 --> 36:29.800 And I was actually very busy as a noobler, 36:29.800 --> 36:30.720 as a new Googler. 36:31.600 --> 36:34.520 So I had not time to work on that. 36:34.520 --> 36:38.680 But then in November, I think it was November, 2015, 36:38.680 --> 36:41.240 TensorFlow got released. 36:41.240 --> 36:44.560 And it was kind of like my wake up call 36:44.560 --> 36:47.320 that, hey, I had to actually go and make it happen. 36:47.320 --> 36:52.200 So in December, I ported Keras to run on top of TensorFlow, 36:52.200 --> 36:53.320 but it was not exactly a port. 36:53.320 --> 36:55.280 It was more like a refactoring 36:55.280 --> 36:57.920 where I was abstracting away 36:57.920 --> 37:00.480 all the backend functionality into one module 37:00.480 --> 37:02.320 so that the same code base 37:02.320 --> 37:05.080 could run on top of multiple backends. 37:05.080 --> 37:07.440 So on top of TensorFlow or Theano. 37:07.440 --> 37:09.760 And for the next year, 37:09.760 --> 37:14.760 Theano stayed as the default option. 37:15.400 --> 37:20.400 It was easier to use, somewhat less buggy. 37:20.640 --> 37:23.360 It was much faster, especially when it came to audience. 37:23.360 --> 37:26.360 But eventually, TensorFlow overtook it. 37:27.480 --> 37:30.200 And TensorFlow, the early TensorFlow, 37:30.200 --> 37:33.960 has similar architectural decisions as Theano, right? 37:33.960 --> 37:37.440 So it was a natural transition. 37:37.440 --> 37:38.320 Yeah, absolutely. 37:38.320 --> 37:42.960 So what, I mean, that still Keras is a side, 37:42.960 --> 37:45.280 almost fun project, right? 37:45.280 --> 37:49.040 Yeah, so it was not my job assignment. 37:49.040 --> 37:50.360 It was not. 37:50.360 --> 37:52.240 I was doing it on the side. 37:52.240 --> 37:55.840 And even though it grew to have a lot of users 37:55.840 --> 37:59.600 for a deep learning library at the time, like Stroud 2016, 37:59.600 --> 38:02.480 but I wasn't doing it as my main job. 38:02.480 --> 38:04.760 So things started changing in, 38:04.760 --> 38:09.760 I think it must have been maybe October, 2016. 38:10.200 --> 38:11.320 So one year later. 38:12.360 --> 38:15.240 So Rajat, who was the lead on TensorFlow, 38:15.240 --> 38:19.240 basically showed up one day in our building 38:19.240 --> 38:20.080 where I was doing like, 38:20.080 --> 38:21.640 so I was doing research and things like, 38:21.640 --> 38:24.640 so I did a lot of computer vision research, 38:24.640 --> 38:27.560 also collaborations with Christian Zighetti 38:27.560 --> 38:29.640 and deep learning for theorem proving. 38:29.640 --> 38:32.920 It was a really interesting research topic. 38:34.520 --> 38:37.640 And so Rajat was saying, 38:37.640 --> 38:41.040 hey, we saw Keras, we like it. 38:41.040 --> 38:42.440 We saw that you're at Google. 38:42.440 --> 38:45.280 Why don't you come over for like a quarter 38:45.280 --> 38:47.280 and work with us? 38:47.280 --> 38:49.240 And I was like, yeah, that sounds like a great opportunity. 38:49.240 --> 38:50.400 Let's do it. 38:50.400 --> 38:55.400 And so I started working on integrating the Keras API 38:55.720 --> 38:57.320 into TensorFlow more tightly. 38:57.320 --> 39:02.320 So what followed up is a sort of like temporary 39:02.640 --> 39:05.480 TensorFlow only version of Keras 39:05.480 --> 39:09.320 that was in TensorFlow.com Trib for a while. 39:09.320 --> 39:12.200 And finally moved to TensorFlow Core. 39:12.200 --> 39:15.360 And I've never actually gotten back 39:15.360 --> 39:17.600 to my old team doing research. 39:17.600 --> 39:22.320 Well, it's kind of funny that somebody like you 39:22.320 --> 39:27.320 who dreams of, or at least sees the power of AI systems 39:28.960 --> 39:31.680 that reason and theorem proving we'll talk about 39:31.680 --> 39:36.520 has also created a system that makes the most basic 39:36.520 --> 39:40.400 kind of Lego building that is deep learning 39:40.400 --> 39:42.640 super accessible, super easy. 39:42.640 --> 39:43.800 So beautifully so. 39:43.800 --> 39:47.720 It's a funny irony that you're both, 39:47.720 --> 39:49.120 you're responsible for both things, 39:49.120 --> 39:54.000 but so TensorFlow 2.0 is kind of, there's a sprint. 39:54.000 --> 39:55.080 I don't know how long it'll take, 39:55.080 --> 39:56.960 but there's a sprint towards the finish. 39:56.960 --> 40:01.040 What do you look, what are you working on these days? 40:01.040 --> 40:02.160 What are you excited about? 40:02.160 --> 40:04.280 What are you excited about in 2.0? 40:04.280 --> 40:05.760 I mean, eager execution. 40:05.760 --> 40:08.440 There's so many things that just make it a lot easier 40:08.440 --> 40:09.760 to work. 40:09.760 --> 40:13.640 What are you excited about and what's also really hard? 40:13.640 --> 40:15.800 What are the problems you have to kind of solve? 40:15.800 --> 40:19.080 So I've spent the past year and a half working on 40:19.080 --> 40:22.920 TensorFlow 2.0 and it's been a long journey. 40:22.920 --> 40:25.080 I'm actually extremely excited about it. 40:25.080 --> 40:26.440 I think it's a great product. 40:26.440 --> 40:29.360 It's a delightful product compared to TensorFlow 1.0. 40:29.360 --> 40:31.440 We've made huge progress. 40:32.640 --> 40:37.400 So on the Keras side, what I'm really excited about is that, 40:37.400 --> 40:42.400 so previously Keras has been this very easy to use 40:42.400 --> 40:45.840 high level interface to do deep learning. 40:45.840 --> 40:47.280 But if you wanted to, 40:50.520 --> 40:53.040 if you wanted a lot of flexibility, 40:53.040 --> 40:57.520 the Keras framework was probably not the optimal way 40:57.520 --> 40:59.760 to do things compared to just writing everything 40:59.760 --> 41:00.600 from scratch. 41:01.800 --> 41:04.680 So in some way, the framework was getting in the way. 41:04.680 --> 41:07.960 And in TensorFlow 2.0, you don't have this at all, actually. 41:07.960 --> 41:11.040 You have the usability of the high level interface, 41:11.040 --> 41:14.480 but you have the flexibility of this lower level interface. 41:14.480 --> 41:16.800 And you have this spectrum of workflows 41:16.800 --> 41:21.560 where you can get more or less usability 41:21.560 --> 41:26.560 and flexibility trade offs depending on your needs, right? 41:26.640 --> 41:29.680 You can write everything from scratch 41:29.680 --> 41:32.320 and you get a lot of help doing so 41:32.320 --> 41:36.400 by subclassing models and writing some train loops 41:36.400 --> 41:38.200 using ego execution. 41:38.200 --> 41:40.160 It's very flexible, it's very easy to debug, 41:40.160 --> 41:41.400 it's very powerful. 41:42.280 --> 41:45.000 But all of this integrates seamlessly 41:45.000 --> 41:49.440 with higher level features up to the classic Keras workflows, 41:49.440 --> 41:51.560 which are very scikit learn like 41:51.560 --> 41:56.040 and are ideal for a data scientist, 41:56.040 --> 41:58.240 machine learning engineer type of profile. 41:58.240 --> 42:00.840 So now you can have the same framework 42:00.840 --> 42:02.880 offering the same set of APIs 42:02.880 --> 42:05.000 that enable a spectrum of workflows 42:05.000 --> 42:08.560 that are more or less low level, more or less high level 42:08.560 --> 42:13.520 that are suitable for profiles ranging from researchers 42:13.520 --> 42:15.560 to data scientists and everything in between. 42:15.560 --> 42:16.960 Yeah, so that's super exciting. 42:16.960 --> 42:18.400 I mean, it's not just that, 42:18.400 --> 42:21.680 it's connected to all kinds of tooling. 42:21.680 --> 42:24.520 You can go on mobile, you can go with TensorFlow Lite, 42:24.520 --> 42:27.240 you can go in the cloud or serving and so on. 42:27.240 --> 42:28.960 It all is connected together. 42:28.960 --> 42:31.880 Now some of the best software written ever 42:31.880 --> 42:36.880 is often done by one person, sometimes two. 42:36.880 --> 42:40.800 So with a Google, you're now seeing sort of Keras 42:40.800 --> 42:42.840 having to be integrated in TensorFlow, 42:42.840 --> 42:46.800 I'm sure has a ton of engineers working on. 42:46.800 --> 42:51.040 And there's, I'm sure a lot of tricky design decisions 42:51.040 --> 42:52.200 to be made. 42:52.200 --> 42:54.440 How does that process usually happen 42:54.440 --> 42:56.800 from at least your perspective? 42:56.800 --> 42:59.800 What are the debates like? 43:00.720 --> 43:04.200 Is there a lot of thinking, 43:04.200 --> 43:06.880 considering different options and so on? 43:06.880 --> 43:08.160 Yes. 43:08.160 --> 43:12.640 So a lot of the time I spend at Google 43:12.640 --> 43:17.280 is actually discussing design discussions, right? 43:17.280 --> 43:20.480 Writing design docs, participating in design review meetings 43:20.480 --> 43:22.080 and so on. 43:22.080 --> 43:25.240 This is as important as actually writing a code. 43:25.240 --> 43:26.080 Right. 43:26.080 --> 43:28.120 So there's a lot of thought, there's a lot of thought 43:28.120 --> 43:32.280 and a lot of care that is taken 43:32.280 --> 43:34.160 in coming up with these decisions 43:34.160 --> 43:37.160 and taking into account all of our users 43:37.160 --> 43:40.680 because TensorFlow has this extremely diverse user base, 43:40.680 --> 43:41.520 right? 43:41.520 --> 43:43.120 It's not like just one user segment 43:43.120 --> 43:45.480 where everyone has the same needs. 43:45.480 --> 43:47.640 We have small scale production users, 43:47.640 --> 43:49.520 large scale production users. 43:49.520 --> 43:52.800 We have startups, we have researchers, 43:53.720 --> 43:55.080 you know, it's all over the place. 43:55.080 --> 43:57.560 And we have to cater to all of their needs. 43:57.560 --> 44:00.040 If I just look at the standard debates 44:00.040 --> 44:04.000 of C++ or Python, there's some heated debates. 44:04.000 --> 44:06.000 Do you have those at Google? 44:06.000 --> 44:08.080 I mean, they're not heated in terms of emotionally, 44:08.080 --> 44:10.800 but there's probably multiple ways to do it, right? 44:10.800 --> 44:14.040 So how do you arrive through those design meetings 44:14.040 --> 44:15.440 at the best way to do it? 44:15.440 --> 44:19.280 Especially in deep learning where the field is evolving 44:19.280 --> 44:20.880 as you're doing it. 44:21.880 --> 44:23.600 Is there some magic to it? 44:23.600 --> 44:26.240 Is there some magic to the process? 44:26.240 --> 44:28.280 I don't know if there's magic to the process, 44:28.280 --> 44:30.640 but there definitely is a process. 44:30.640 --> 44:33.760 So making design decisions 44:33.760 --> 44:36.080 is about satisfying a set of constraints, 44:36.080 --> 44:39.920 but also trying to do so in the simplest way possible, 44:39.920 --> 44:42.240 because this is what can be maintained, 44:42.240 --> 44:44.920 this is what can be expanded in the future. 44:44.920 --> 44:49.120 So you don't want to naively satisfy the constraints 44:49.120 --> 44:51.880 by just, you know, for each capability you need available, 44:51.880 --> 44:53.960 you're gonna come up with one argument in your API 44:53.960 --> 44:54.800 and so on. 44:54.800 --> 44:59.800 You want to design APIs that are modular and hierarchical 45:00.640 --> 45:04.080 so that they have an API surface 45:04.080 --> 45:07.040 that is as small as possible, right? 45:07.040 --> 45:11.640 And you want this modular hierarchical architecture 45:11.640 --> 45:14.560 to reflect the way that domain experts 45:14.560 --> 45:16.400 think about the problem. 45:16.400 --> 45:17.880 Because as a domain expert, 45:17.880 --> 45:19.840 when you are reading about a new API, 45:19.840 --> 45:24.760 you're reading a tutorial or some docs pages, 45:24.760 --> 45:28.200 you already have a way that you're thinking about the problem. 45:28.200 --> 45:32.320 You already have like certain concepts in mind 45:32.320 --> 45:35.680 and you're thinking about how they relate together. 45:35.680 --> 45:37.200 And when you're reading docs, 45:37.200 --> 45:40.280 you're trying to build as quickly as possible 45:40.280 --> 45:45.280 a mapping between the concepts featured in your API 45:45.280 --> 45:46.800 and the concepts in your mind. 45:46.800 --> 45:48.880 So you're trying to map your mental model 45:48.880 --> 45:53.600 as a domain expert to the way things work in the API. 45:53.600 --> 45:57.040 So you need an API and an underlying implementation 45:57.040 --> 46:00.120 that are reflecting the way people think about these things. 46:00.120 --> 46:02.880 So in minimizing the time it takes to do the mapping. 46:02.880 --> 46:04.680 Yes, minimizing the time, 46:04.680 --> 46:06.560 the cognitive load there is 46:06.560 --> 46:10.920 in ingesting this new knowledge about your API. 46:10.920 --> 46:13.160 An API should not be self referential 46:13.160 --> 46:15.520 or referring to implementation details. 46:15.520 --> 46:19.160 It should only be referring to domain specific concepts 46:19.160 --> 46:21.360 that people already understand. 46:23.240 --> 46:24.480 Brilliant. 46:24.480 --> 46:27.560 So what's the future of Keras and TensorFlow look like? 46:27.560 --> 46:29.640 What does TensorFlow 3.0 look like? 46:30.600 --> 46:33.720 So that's kind of too far in the future for me to answer, 46:33.720 --> 46:37.800 especially since I'm not even the one making these decisions. 46:37.800 --> 46:39.080 Okay. 46:39.080 --> 46:41.240 But so from my perspective, 46:41.240 --> 46:43.200 which is just one perspective 46:43.200 --> 46:46.040 among many different perspectives on the TensorFlow team, 46:47.200 --> 46:52.200 I'm really excited by developing even higher level APIs, 46:52.360 --> 46:53.560 higher level than Keras. 46:53.560 --> 46:56.480 I'm really excited by hyperparameter tuning, 46:56.480 --> 46:59.240 by automated machine learning, AutoML. 47:01.120 --> 47:03.200 I think the future is not just, you know, 47:03.200 --> 47:07.600 defining a model like you were assembling Lego blocks 47:07.600 --> 47:09.200 and then collect fit on it. 47:09.200 --> 47:13.680 It's more like an automagical model 47:13.680 --> 47:16.080 that would just look at your data 47:16.080 --> 47:19.040 and optimize the objective you're after, right? 47:19.040 --> 47:23.040 So that's what I'm looking into. 47:23.040 --> 47:26.480 Yeah, so you put the baby into a room with the problem 47:26.480 --> 47:28.760 and come back a few hours later 47:28.760 --> 47:30.960 with a fully solved problem. 47:30.960 --> 47:33.560 Exactly, it's not like a box of Legos. 47:33.560 --> 47:35.920 It's more like the combination of a kid 47:35.920 --> 47:38.800 that's really good at Legos and a box of Legos. 47:38.800 --> 47:41.520 It's just building the thing on its own. 47:41.520 --> 47:42.680 Very nice. 47:42.680 --> 47:44.160 So that's an exciting future. 47:44.160 --> 47:46.080 I think there's a huge amount of applications 47:46.080 --> 47:48.560 and revolutions to be had 47:49.920 --> 47:52.640 under the constraints of the discussion we previously had. 47:52.640 --> 47:57.480 But what do you think of the current limits of deep learning? 47:57.480 --> 48:02.480 If we look specifically at these function approximators 48:03.840 --> 48:06.160 that tries to generalize from data. 48:06.160 --> 48:10.160 You've talked about local versus extreme generalization. 48:11.120 --> 48:13.280 You mentioned that neural networks don't generalize well 48:13.280 --> 48:14.560 and humans do. 48:14.560 --> 48:15.760 So there's this gap. 48:17.640 --> 48:20.840 And you've also mentioned that extreme generalization 48:20.840 --> 48:23.960 requires something like reasoning to fill those gaps. 48:23.960 --> 48:27.560 So how can we start trying to build systems like that? 48:27.560 --> 48:30.600 Right, yeah, so this is by design, right? 48:30.600 --> 48:37.080 Deep learning models are like huge parametric models, 48:37.080 --> 48:39.280 differentiable, so continuous, 48:39.280 --> 48:42.680 that go from an input space to an output space. 48:42.680 --> 48:44.120 And they're trained with gradient descent. 48:44.120 --> 48:47.160 So they're trained pretty much point by point. 48:47.160 --> 48:50.520 They are learning a continuous geometric morphing 48:50.520 --> 48:55.320 from an input vector space to an output vector space. 48:55.320 --> 48:58.960 And because this is done point by point, 48:58.960 --> 49:02.200 a deep neural network can only make sense 49:02.200 --> 49:05.880 of points in experience space that are very close 49:05.880 --> 49:08.520 to things that it has already seen in string data. 49:08.520 --> 49:12.520 At best, it can do interpolation across points. 49:13.840 --> 49:17.360 But that means in order to train your network, 49:17.360 --> 49:21.680 you need a dense sampling of the input cross output space, 49:22.880 --> 49:25.240 almost a point by point sampling, 49:25.240 --> 49:27.160 which can be very expensive if you're dealing 49:27.160 --> 49:29.320 with complex real world problems, 49:29.320 --> 49:33.240 like autonomous driving, for instance, or robotics. 49:33.240 --> 49:36.000 It's doable if you're looking at the subset 49:36.000 --> 49:37.120 of the visual space. 49:37.120 --> 49:38.800 But even then, it's still fairly expensive. 49:38.800 --> 49:40.920 You still need millions of examples. 49:40.920 --> 49:44.240 And it's only going to be able to make sense of things 49:44.240 --> 49:46.880 that are very close to what it has seen before. 49:46.880 --> 49:49.160 And in contrast to that, well, of course, 49:49.160 --> 49:50.160 you have human intelligence. 49:50.160 --> 49:53.240 But even if you're not looking at human intelligence, 49:53.240 --> 49:56.800 you can look at very simple rules, algorithms. 49:56.800 --> 49:58.080 If you have a symbolic rule, 49:58.080 --> 50:03.080 it can actually apply to a very, very large set of inputs 50:03.120 --> 50:04.880 because it is abstract. 50:04.880 --> 50:09.560 It is not obtained by doing a point by point mapping. 50:10.720 --> 50:14.000 For instance, if you try to learn a sorting algorithm 50:14.000 --> 50:15.520 using a deep neural network, 50:15.520 --> 50:18.520 well, you're very much limited to learning point by point 50:20.080 --> 50:24.360 what the sorted representation of this specific list is like. 50:24.360 --> 50:29.360 But instead, you could have a very, very simple 50:29.400 --> 50:31.920 sorting algorithm written in a few lines. 50:31.920 --> 50:34.520 Maybe it's just two nested loops. 50:35.560 --> 50:40.560 And it can process any list at all because it is abstract, 50:41.040 --> 50:42.240 because it is a set of rules. 50:42.240 --> 50:45.160 So deep learning is really like point by point 50:45.160 --> 50:48.640 geometric morphings, train with good and decent. 50:48.640 --> 50:53.640 And meanwhile, abstract rules can generalize much better. 50:53.640 --> 50:56.160 And I think the future is we need to combine the two. 50:56.160 --> 50:59.160 So how do we, do you think, combine the two? 50:59.160 --> 51:03.040 How do we combine good point by point functions 51:03.040 --> 51:08.040 with programs, which is what the symbolic AI type systems? 51:08.920 --> 51:11.600 At which levels the combination happen? 51:11.600 --> 51:14.680 I mean, obviously we're jumping into the realm 51:14.680 --> 51:16.880 of where there's no good answers. 51:16.880 --> 51:20.280 It's just kind of ideas and intuitions and so on. 51:20.280 --> 51:23.080 Well, if you look at the really successful AI systems 51:23.080 --> 51:26.320 today, I think they are already hybrid systems 51:26.320 --> 51:29.520 that are combining symbolic AI with deep learning. 51:29.520 --> 51:32.520 For instance, successful robotics systems 51:32.520 --> 51:36.400 are already mostly model based, rule based, 51:37.400 --> 51:39.400 things like planning algorithms and so on. 51:39.400 --> 51:42.200 At the same time, they're using deep learning 51:42.200 --> 51:43.840 as perception modules. 51:43.840 --> 51:46.000 Sometimes they're using deep learning as a way 51:46.000 --> 51:50.920 to inject fuzzy intuition into a rule based process. 51:50.920 --> 51:54.560 If you look at the system like in a self driving car, 51:54.560 --> 51:57.240 it's not just one big end to end neural network. 51:57.240 --> 51:59.000 You know, that wouldn't work at all. 51:59.000 --> 52:00.760 Precisely because in order to train that, 52:00.760 --> 52:05.160 you would need a dense sampling of experience base 52:05.160 --> 52:06.200 when it comes to driving, 52:06.200 --> 52:08.880 which is completely unrealistic, obviously. 52:08.880 --> 52:12.440 Instead, the self driving car is mostly 52:13.920 --> 52:18.360 symbolic, you know, it's software, it's programmed by hand. 52:18.360 --> 52:21.640 So it's mostly based on explicit models. 52:21.640 --> 52:25.840 In this case, mostly 3D models of the environment 52:25.840 --> 52:29.520 around the car, but it's interfacing with the real world 52:29.520 --> 52:31.440 using deep learning modules, right? 52:31.440 --> 52:33.440 So the deep learning there serves as a way 52:33.440 --> 52:36.080 to convert the raw sensory information 52:36.080 --> 52:38.320 to something usable by symbolic systems. 52:39.760 --> 52:42.400 Okay, well, let's linger on that a little more. 52:42.400 --> 52:45.440 So dense sampling from input to output. 52:45.440 --> 52:48.240 You said it's obviously very difficult. 52:48.240 --> 52:50.120 Is it possible? 52:50.120 --> 52:51.800 In the case of self driving, you mean? 52:51.800 --> 52:53.040 Let's say self driving, right? 52:53.040 --> 52:55.760 Self driving for many people, 52:57.560 --> 52:59.520 let's not even talk about self driving, 52:59.520 --> 53:03.880 let's talk about steering, so staying inside the lane. 53:05.040 --> 53:07.080 Lane following, yeah, it's definitely a problem 53:07.080 --> 53:08.880 you can solve with an end to end deep learning model, 53:08.880 --> 53:10.600 but that's like one small subset. 53:10.600 --> 53:11.440 Hold on a second. 53:11.440 --> 53:12.760 Yeah, I don't know why you're jumping 53:12.760 --> 53:14.480 from the extreme so easily, 53:14.480 --> 53:16.280 because I disagree with you on that. 53:16.280 --> 53:21.000 I think, well, it's not obvious to me 53:21.000 --> 53:23.400 that you can solve lane following. 53:23.400 --> 53:25.840 No, it's not obvious, I think it's doable. 53:25.840 --> 53:30.840 I think in general, there is no hard limitations 53:31.200 --> 53:33.680 to what you can learn with a deep neural network, 53:33.680 --> 53:38.680 as long as the search space is rich enough, 53:40.320 --> 53:42.240 is flexible enough, and as long as you have 53:42.240 --> 53:45.360 this dense sampling of the input cross output space. 53:45.360 --> 53:47.720 The problem is that this dense sampling 53:47.720 --> 53:51.120 could mean anything from 10,000 examples 53:51.120 --> 53:52.840 to like trillions and trillions. 53:52.840 --> 53:54.360 So that's my question. 53:54.360 --> 53:56.200 So what's your intuition? 53:56.200 --> 53:58.720 And if you could just give it a chance 53:58.720 --> 54:01.880 and think what kind of problems can be solved 54:01.880 --> 54:04.240 by getting a huge amounts of data 54:04.240 --> 54:08.000 and thereby creating a dense mapping. 54:08.000 --> 54:12.480 So let's think about natural language dialogue, 54:12.480 --> 54:14.000 the Turing test. 54:14.000 --> 54:17.000 Do you think the Turing test can be solved 54:17.000 --> 54:21.120 with a neural network alone? 54:21.120 --> 54:24.440 Well, the Turing test is all about tricking people 54:24.440 --> 54:26.880 into believing they're talking to a human. 54:26.880 --> 54:29.040 And I don't think that's actually very difficult 54:29.040 --> 54:34.040 because it's more about exploiting human perception 54:35.600 --> 54:37.520 and not so much about intelligence. 54:37.520 --> 54:39.680 There's a big difference between mimicking 54:39.680 --> 54:42.080 intelligent behavior and actual intelligent behavior. 54:42.080 --> 54:45.360 So, okay, let's look at maybe the Alexa prize and so on. 54:45.360 --> 54:47.480 The different formulations of the natural language 54:47.480 --> 54:50.520 conversation that are less about mimicking 54:50.520 --> 54:52.800 and more about maintaining a fun conversation 54:52.800 --> 54:54.720 that lasts for 20 minutes. 54:54.720 --> 54:56.200 That's a little less about mimicking 54:56.200 --> 54:59.080 and that's more about, I mean, it's still mimicking, 54:59.080 --> 55:01.440 but it's more about being able to carry forward 55:01.440 --> 55:03.640 a conversation with all the tangents that happen 55:03.640 --> 55:05.080 in dialogue and so on. 55:05.080 --> 55:08.320 Do you think that problem is learnable 55:08.320 --> 55:13.320 with a neural network that does the point to point mapping? 55:14.520 --> 55:16.280 So I think it would be very, very challenging 55:16.280 --> 55:17.800 to do this with deep learning. 55:17.800 --> 55:21.480 I don't think it's out of the question either. 55:21.480 --> 55:23.240 I wouldn't rule it out. 55:23.240 --> 55:25.400 The space of problems that can be solved 55:25.400 --> 55:26.920 with a large neural network. 55:26.920 --> 55:30.080 What's your sense about the space of those problems? 55:30.080 --> 55:32.560 So useful problems for us. 55:32.560 --> 55:34.800 In theory, it's infinite, right? 55:34.800 --> 55:36.200 You can solve any problem. 55:36.200 --> 55:39.800 In practice, well, deep learning is a great fit 55:39.800 --> 55:41.800 for perception problems. 55:41.800 --> 55:46.800 In general, any problem which is naturally amenable 55:47.640 --> 55:52.200 to explicit handcrafted rules or rules that you can generate 55:52.200 --> 55:54.960 by exhaustive search over some program space. 55:56.080 --> 55:59.320 So perception, artificial intuition, 55:59.320 --> 56:03.240 as long as you have a sufficient training dataset. 56:03.240 --> 56:05.360 And that's the question, I mean, perception, 56:05.360 --> 56:08.400 there's interpretation and understanding of the scene, 56:08.400 --> 56:10.280 which seems to be outside the reach 56:10.280 --> 56:12.960 of current perception systems. 56:12.960 --> 56:15.920 So do you think larger networks will be able 56:15.920 --> 56:18.280 to start to understand the physics 56:18.280 --> 56:21.080 and the physics of the scene, 56:21.080 --> 56:23.400 the three dimensional structure and relationships 56:23.400 --> 56:25.560 of objects in the scene and so on? 56:25.560 --> 56:28.320 Or really that's where symbolic AI has to step in? 56:28.320 --> 56:34.480 Well, it's always possible to solve these problems 56:34.480 --> 56:36.800 with deep learning. 56:36.800 --> 56:38.560 It's just extremely inefficient. 56:38.560 --> 56:42.000 A model would be an explicit rule based abstract model 56:42.000 --> 56:45.240 would be a far better, more compressed 56:45.240 --> 56:46.840 representation of physics. 56:46.840 --> 56:49.080 Then learning just this mapping between 56:49.080 --> 56:50.960 in this situation, this thing happens. 56:50.960 --> 56:52.720 If you change the situation slightly, 56:52.720 --> 56:54.760 then this other thing happens and so on. 56:54.760 --> 56:57.440 Do you think it's possible to automatically generate 56:57.440 --> 57:02.200 the programs that would require that kind of reasoning? 57:02.200 --> 57:05.360 Or does it have to, so the way the expert systems fail, 57:05.360 --> 57:07.120 there's so many facts about the world 57:07.120 --> 57:08.960 had to be hand coded in. 57:08.960 --> 57:14.600 Do you think it's possible to learn those logical statements 57:14.600 --> 57:18.200 that are true about the world and their relationships? 57:18.200 --> 57:20.360 Do you think, I mean, that's kind of what theorem proving 57:20.360 --> 57:22.680 at a basic level is trying to do, right? 57:22.680 --> 57:26.160 Yeah, except it's much harder to formulate statements 57:26.160 --> 57:28.480 about the world compared to formulating 57:28.480 --> 57:30.320 mathematical statements. 57:30.320 --> 57:34.200 Statements about the world tend to be subjective. 57:34.200 --> 57:39.600 So can you learn rule based models? 57:39.600 --> 57:40.920 Yes, definitely. 57:40.920 --> 57:43.640 That's the field of program synthesis. 57:43.640 --> 57:48.040 However, today we just don't really know how to do it. 57:48.040 --> 57:52.400 So it's very much a grass search or tree search problem. 57:52.400 --> 57:56.800 And so we are limited to the sort of tree session grass 57:56.800 --> 57:58.560 search algorithms that we have today. 57:58.560 --> 58:02.760 Personally, I think genetic algorithms are very promising. 58:02.760 --> 58:04.360 So almost like genetic programming. 58:04.360 --> 58:05.560 Genetic programming, exactly. 58:05.560 --> 58:08.840 Can you discuss the field of program synthesis? 58:08.840 --> 58:14.560 Like how many people are working and thinking about it? 58:14.560 --> 58:17.960 Where we are in the history of program synthesis 58:17.960 --> 58:20.720 and what are your hopes for it? 58:20.720 --> 58:24.600 Well, if it were deep learning, this is like the 90s. 58:24.600 --> 58:29.120 So meaning that we already have existing solutions. 58:29.120 --> 58:34.280 We are starting to have some basic understanding 58:34.280 --> 58:35.480 of what this is about. 58:35.480 --> 58:38.000 But it's still a field that is in its infancy. 58:38.000 --> 58:40.440 There are very few people working on it. 58:40.440 --> 58:44.480 There are very few real world applications. 58:44.480 --> 58:47.640 So the one real world application I'm aware of 58:47.640 --> 58:51.680 is Flash Fill in Excel. 58:51.680 --> 58:55.080 It's a way to automatically learn very simple programs 58:55.080 --> 58:58.200 to format cells in an Excel spreadsheet 58:58.200 --> 59:00.240 from a few examples. 59:00.240 --> 59:02.800 For instance, learning a way to format a date, things like that. 59:02.800 --> 59:03.680 Oh, that's fascinating. 59:03.680 --> 59:04.560 Yeah. 59:04.560 --> 59:06.280 You know, OK, that's a fascinating topic. 59:06.280 --> 59:10.480 I always wonder when I provide a few samples to Excel, 59:10.480 --> 59:12.600 what it's able to figure out. 59:12.600 --> 59:15.960 Like just giving it a few dates, what 59:15.960 --> 59:18.480 are you able to figure out from the pattern I just gave you? 59:18.480 --> 59:19.760 That's a fascinating question. 59:19.760 --> 59:23.320 And it's fascinating whether that's learnable patterns. 59:23.320 --> 59:25.520 And you're saying they're working on that. 59:25.520 --> 59:28.200 How big is the toolbox currently? 59:28.200 --> 59:29.520 Are we completely in the dark? 59:29.520 --> 59:30.440 So if you said the 90s. 59:30.440 --> 59:31.720 In terms of program synthesis? 59:31.720 --> 59:32.360 No. 59:32.360 --> 59:37.720 So I would say, so maybe 90s is even too optimistic. 59:37.720 --> 59:41.080 Because by the 90s, we already understood back prop. 59:41.080 --> 59:43.960 We already understood the engine of deep learning, 59:43.960 --> 59:47.280 even though we couldn't really see its potential quite. 59:47.280 --> 59:48.520 Today, I don't think we have found 59:48.520 --> 59:50.400 the engine of program synthesis. 59:50.400 --> 59:52.880 So we're in the winter before back prop. 59:52.880 --> 59:54.160 Yeah. 59:54.160 --> 59:55.720 In a way, yes. 59:55.720 --> 1:00:00.120 So I do believe program synthesis and general discrete search 1:00:00.120 --> 1:00:02.760 over rule based models is going to be 1:00:02.760 --> 1:00:06.640 a cornerstone of AI research in the next century. 1:00:06.640 --> 1:00:10.200 And that doesn't mean we are going to drop deep learning. 1:00:10.200 --> 1:00:11.880 Deep learning is immensely useful. 1:00:11.880 --> 1:00:17.200 Like, being able to learn is a very flexible, adaptable, 1:00:17.200 --> 1:00:18.120 parametric model. 1:00:18.120 --> 1:00:20.720 So it's got to understand that's actually immensely useful. 1:00:20.720 --> 1:00:23.040 All it's doing is pattern cognition. 1:00:23.040 --> 1:00:25.640 But being good at pattern cognition, given lots of data, 1:00:25.640 --> 1:00:27.920 is just extremely powerful. 1:00:27.920 --> 1:00:30.320 So we are still going to be working on deep learning. 1:00:30.320 --> 1:00:31.840 We are going to be working on program synthesis. 1:00:31.840 --> 1:00:34.680 We are going to be combining the two in increasingly automated 1:00:34.680 --> 1:00:36.400 ways. 1:00:36.400 --> 1:00:38.520 So let's talk a little bit about data. 1:00:38.520 --> 1:00:44.600 You've tweeted, about 10,000 deep learning papers 1:00:44.600 --> 1:00:47.080 have been written about hard coding priors 1:00:47.080 --> 1:00:49.600 about a specific task in a neural network architecture 1:00:49.600 --> 1:00:52.440 works better than a lack of a prior. 1:00:52.440 --> 1:00:55.120 Basically, summarizing all these efforts, 1:00:55.120 --> 1:00:56.920 they put a name to an architecture. 1:00:56.920 --> 1:00:59.280 But really, what they're doing is hard coding some priors 1:00:59.280 --> 1:01:01.560 that improve the performance of the system. 1:01:01.560 --> 1:01:06.880 But which gets straight to the point is probably true. 1:01:06.880 --> 1:01:09.800 So you say that you can always buy performance by, 1:01:09.800 --> 1:01:12.920 in quotes, performance by either training on more data, 1:01:12.920 --> 1:01:15.480 better data, or by injecting task information 1:01:15.480 --> 1:01:18.400 to the architecture of the preprocessing. 1:01:18.400 --> 1:01:21.280 However, this isn't informative about the generalization power 1:01:21.280 --> 1:01:23.080 the techniques use, the fundamental ability 1:01:23.080 --> 1:01:24.200 to generalize. 1:01:24.200 --> 1:01:26.800 Do you think we can go far by coming up 1:01:26.800 --> 1:01:29.920 with better methods for this kind of cheating, 1:01:29.920 --> 1:01:33.520 for better methods of large scale annotation of data? 1:01:33.520 --> 1:01:34.960 So building better priors. 1:01:34.960 --> 1:01:37.280 If you automate it, it's not cheating anymore. 1:01:37.280 --> 1:01:38.360 Right. 1:01:38.360 --> 1:01:41.600 I'm joking about the cheating, but large scale. 1:01:41.600 --> 1:01:46.560 So basically, I'm asking about something 1:01:46.560 --> 1:01:48.280 that hasn't, from my perspective, 1:01:48.280 --> 1:01:53.360 been researched too much is exponential improvement 1:01:53.360 --> 1:01:55.960 in annotation of data. 1:01:55.960 --> 1:01:58.120 Do you often think about? 1:01:58.120 --> 1:02:00.840 I think it's actually been researched quite a bit. 1:02:00.840 --> 1:02:02.720 You just don't see publications about it. 1:02:02.720 --> 1:02:05.840 Because people who publish papers 1:02:05.840 --> 1:02:07.920 are going to publish about known benchmarks. 1:02:07.920 --> 1:02:09.800 Sometimes they're going to read a new benchmark. 1:02:09.800 --> 1:02:12.200 People who actually have real world large scale 1:02:12.200 --> 1:02:13.880 depending on problems, they're going 1:02:13.880 --> 1:02:16.960 to spend a lot of resources into data annotation 1:02:16.960 --> 1:02:18.400 and good data annotation pipelines, 1:02:18.400 --> 1:02:19.640 but you don't see any papers about it. 1:02:19.640 --> 1:02:20.400 That's interesting. 1:02:20.400 --> 1:02:22.720 So do you think, certainly resources, 1:02:22.720 --> 1:02:24.840 but do you think there's innovation happening? 1:02:24.840 --> 1:02:25.880 Oh, yeah. 1:02:25.880 --> 1:02:28.880 To clarify the point in the tweet. 1:02:28.880 --> 1:02:31.160 So machine learning in general is 1:02:31.160 --> 1:02:33.840 the science of generalization. 1:02:33.840 --> 1:02:37.800 You want to generate knowledge that 1:02:37.800 --> 1:02:40.440 can be reused across different data sets, 1:02:40.440 --> 1:02:42.000 across different tasks. 1:02:42.000 --> 1:02:45.280 And if instead you're looking at one data set 1:02:45.280 --> 1:02:50.000 and then you are hard coding knowledge about this task 1:02:50.000 --> 1:02:54.040 into your architecture, this is no more useful 1:02:54.040 --> 1:02:56.760 than training a network and then saying, oh, I 1:02:56.760 --> 1:03:01.920 found these weight values perform well. 1:03:01.920 --> 1:03:05.680 So David Ha, I don't know if you know David, 1:03:05.680 --> 1:03:08.760 he had a paper the other day about weight 1:03:08.760 --> 1:03:10.400 agnostic neural networks. 1:03:10.400 --> 1:03:12.120 And this is a very interesting paper 1:03:12.120 --> 1:03:14.400 because it really illustrates the fact 1:03:14.400 --> 1:03:17.400 that an architecture, even without weights, 1:03:17.400 --> 1:03:21.360 an architecture is knowledge about a task. 1:03:21.360 --> 1:03:23.640 It encodes knowledge. 1:03:23.640 --> 1:03:25.840 And when it comes to architectures 1:03:25.840 --> 1:03:30.440 that are uncrafted by researchers, in some cases, 1:03:30.440 --> 1:03:34.160 it is very, very clear that all they are doing 1:03:34.160 --> 1:03:38.880 is artificially reencoding the template that 1:03:38.880 --> 1:03:44.400 corresponds to the proper way to solve the task encoding 1:03:44.400 --> 1:03:45.200 a given data set. 1:03:45.200 --> 1:03:48.120 For instance, I know if you looked 1:03:48.120 --> 1:03:52.280 at the baby data set, which is about natural language 1:03:52.280 --> 1:03:55.520 question answering, it is generated by an algorithm. 1:03:55.520 --> 1:03:57.680 So this is a question answer pairs 1:03:57.680 --> 1:03:59.280 that are generated by an algorithm. 1:03:59.280 --> 1:04:01.520 The algorithm is solving a certain template. 1:04:01.520 --> 1:04:04.400 Turns out, if you craft a network that 1:04:04.400 --> 1:04:06.360 literally encodes this template, you 1:04:06.360 --> 1:04:09.640 can solve this data set with nearly 100% accuracy. 1:04:09.640 --> 1:04:11.160 But that doesn't actually tell you 1:04:11.160 --> 1:04:14.640 anything about how to solve question answering 1:04:14.640 --> 1:04:17.680 in general, which is the point. 1:04:17.680 --> 1:04:19.400 The question is just to linger on it, 1:04:19.400 --> 1:04:21.560 whether it's from the data side or from the size 1:04:21.560 --> 1:04:23.280 of the network. 1:04:23.280 --> 1:04:25.920 I don't know if you've read the blog post by Rich Sutton, 1:04:25.920 --> 1:04:28.400 The Bitter Lesson, where he says, 1:04:28.400 --> 1:04:31.480 the biggest lesson that we can read from 70 years of AI 1:04:31.480 --> 1:04:34.720 research is that general methods that leverage computation 1:04:34.720 --> 1:04:37.160 are ultimately the most effective. 1:04:37.160 --> 1:04:39.720 So as opposed to figuring out methods 1:04:39.720 --> 1:04:41.840 that can generalize effectively, do you 1:04:41.840 --> 1:04:47.720 think we can get pretty far by just having something 1:04:47.720 --> 1:04:51.520 that leverages computation and the improvement of computation? 1:04:51.520 --> 1:04:54.960 Yeah, so I think Rich is making a very good point, which 1:04:54.960 --> 1:04:57.560 is that a lot of these papers, which are actually 1:04:57.560 --> 1:05:02.800 all about manually hardcoding prior knowledge about a task 1:05:02.800 --> 1:05:04.720 into some system, it doesn't have 1:05:04.720 --> 1:05:08.600 to be deep learning architecture, but into some system. 1:05:08.600 --> 1:05:11.920 These papers are not actually making any impact. 1:05:11.920 --> 1:05:14.800 Instead, what's making really long term impact 1:05:14.800 --> 1:05:18.520 is very simple, very general systems 1:05:18.520 --> 1:05:21.280 that are really agnostic to all these tricks. 1:05:21.280 --> 1:05:23.320 Because these tricks do not generalize. 1:05:23.320 --> 1:05:27.480 And of course, the one general and simple thing 1:05:27.480 --> 1:05:33.160 that you should focus on is that which leverages computation. 1:05:33.160 --> 1:05:36.200 Because computation, the availability 1:05:36.200 --> 1:05:39.400 of large scale computation has been increasing exponentially 1:05:39.400 --> 1:05:40.560 following Moore's law. 1:05:40.560 --> 1:05:44.080 So if your algorithm is all about exploiting this, 1:05:44.080 --> 1:05:47.440 then your algorithm is suddenly exponentially improving. 1:05:47.440 --> 1:05:52.400 So I think Rich is definitely right. 1:05:52.400 --> 1:05:57.120 However, he's right about the past 70 years. 1:05:57.120 --> 1:05:59.440 He's like assessing the past 70 years. 1:05:59.440 --> 1:06:02.360 I am not sure that this assessment will still 1:06:02.360 --> 1:06:04.880 hold true for the next 70 years. 1:06:04.880 --> 1:06:07.160 It might to some extent. 1:06:07.160 --> 1:06:08.560 I suspect it will not. 1:06:08.560 --> 1:06:11.560 Because the truth of his assessment 1:06:11.560 --> 1:06:16.800 is a function of the context in which this research took place. 1:06:16.800 --> 1:06:18.600 And the context is changing. 1:06:18.600 --> 1:06:21.440 Moore's law might not be applicable anymore, 1:06:21.440 --> 1:06:23.760 for instance, in the future. 1:06:23.760 --> 1:06:31.200 And I do believe that when you tweak one aspect of a system, 1:06:31.200 --> 1:06:32.920 when you exploit one aspect of a system, 1:06:32.920 --> 1:06:36.480 some other aspect starts becoming the bottleneck. 1:06:36.480 --> 1:06:38.800 Let's say you have unlimited computation. 1:06:38.800 --> 1:06:41.440 Well, then data is the bottleneck. 1:06:41.440 --> 1:06:43.560 And I think we are already starting 1:06:43.560 --> 1:06:45.720 to be in a regime where our systems are 1:06:45.720 --> 1:06:48.120 so large in scale and so data ingrained 1:06:48.120 --> 1:06:50.360 that data today and the quality of data 1:06:50.360 --> 1:06:53.040 and the scale of data is the bottleneck. 1:06:53.040 --> 1:06:58.160 And in this environment, the bitter lesson from Rich 1:06:58.160 --> 1:07:00.800 is not going to be true anymore. 1:07:00.800 --> 1:07:03.960 So I think we are going to move from a focus 1:07:03.960 --> 1:07:09.840 on a computation scale to focus on data efficiency. 1:07:09.840 --> 1:07:10.720 Data efficiency. 1:07:10.720 --> 1:07:13.120 So that's getting to the question of symbolic AI. 1:07:13.120 --> 1:07:16.280 But to linger on the deep learning approaches, 1:07:16.280 --> 1:07:19.240 do you have hope for either unsupervised learning 1:07:19.240 --> 1:07:23.280 or reinforcement learning, which are 1:07:23.280 --> 1:07:28.120 ways of being more data efficient in terms 1:07:28.120 --> 1:07:31.560 of the amount of data they need that required human annotation? 1:07:31.560 --> 1:07:34.280 So unsupervised learning and reinforcement learning 1:07:34.280 --> 1:07:36.640 are frameworks for learning, but they are not 1:07:36.640 --> 1:07:39.000 like any specific technique. 1:07:39.000 --> 1:07:41.200 So usually when people say reinforcement learning, 1:07:41.200 --> 1:07:43.320 what they really mean is deep reinforcement learning, 1:07:43.320 --> 1:07:47.440 which is like one approach which is actually very questionable. 1:07:47.440 --> 1:07:50.920 The question I was asking was unsupervised learning 1:07:50.920 --> 1:07:54.680 with deep neural networks and deep reinforcement learning. 1:07:54.680 --> 1:07:56.840 Well, these are not really data efficient 1:07:56.840 --> 1:08:00.520 because you're still leveraging these huge parametric models 1:08:00.520 --> 1:08:03.720 point by point with gradient descent. 1:08:03.720 --> 1:08:08.000 It is more efficient in terms of the number of annotations, 1:08:08.000 --> 1:08:09.520 the density of annotations you need. 1:08:09.520 --> 1:08:13.840 So the idea being to learn the latent space around which 1:08:13.840 --> 1:08:17.960 the data is organized and then map the sparse annotations 1:08:17.960 --> 1:08:18.760 into it. 1:08:18.760 --> 1:08:23.560 And sure, I mean, that's clearly a very good idea. 1:08:23.560 --> 1:08:26.080 It's not really a topic I would be working on, 1:08:26.080 --> 1:08:28.040 but it's clearly a good idea. 1:08:28.040 --> 1:08:31.760 So it would get us to solve some problems that? 1:08:31.760 --> 1:08:34.880 It will get us to incremental improvements 1:08:34.880 --> 1:08:38.240 in labeled data efficiency. 1:08:38.240 --> 1:08:43.520 Do you have concerns about short term or long term threats 1:08:43.520 --> 1:08:47.800 from AI, from artificial intelligence? 1:08:47.800 --> 1:08:50.640 Yes, definitely to some extent. 1:08:50.640 --> 1:08:52.800 And what's the shape of those concerns? 1:08:52.800 --> 1:08:56.880 This is actually something I've briefly written about. 1:08:56.880 --> 1:09:02.680 But the capabilities of deep learning technology 1:09:02.680 --> 1:09:05.200 can be used in many ways that are 1:09:05.200 --> 1:09:09.760 concerning from mass surveillance with things 1:09:09.760 --> 1:09:11.880 like facial recognition. 1:09:11.880 --> 1:09:15.440 In general, tracking lots of data about everyone 1:09:15.440 --> 1:09:18.920 and then being able to making sense of this data 1:09:18.920 --> 1:09:22.240 to do identification, to do prediction. 1:09:22.240 --> 1:09:23.160 That's concerning. 1:09:23.160 --> 1:09:26.560 That's something that's being very aggressively pursued 1:09:26.560 --> 1:09:31.440 by totalitarian states like China. 1:09:31.440 --> 1:09:34.000 One thing I am very much concerned about 1:09:34.000 --> 1:09:40.640 is that our lives are increasingly online, 1:09:40.640 --> 1:09:43.280 are increasingly digital, made of information, 1:09:43.280 --> 1:09:48.080 made of information consumption and information production, 1:09:48.080 --> 1:09:51.800 our digital footprint, I would say. 1:09:51.800 --> 1:09:56.280 And if you absorb all of this data 1:09:56.280 --> 1:10:01.440 and you are in control of where you consume information, 1:10:01.440 --> 1:10:06.960 social networks and so on, recommendation engines, 1:10:06.960 --> 1:10:10.200 then you can build a sort of reinforcement 1:10:10.200 --> 1:10:13.760 loop for human behavior. 1:10:13.760 --> 1:10:18.360 You can observe the state of your mind at time t. 1:10:18.360 --> 1:10:21.080 You can predict how you would react 1:10:21.080 --> 1:10:23.800 to different pieces of content, how 1:10:23.800 --> 1:10:27.000 to get you to move your mind in a certain direction. 1:10:27.000 --> 1:10:33.160 And then you can feed you the specific piece of content 1:10:33.160 --> 1:10:35.680 that would move you in a specific direction. 1:10:35.680 --> 1:10:41.800 And you can do this at scale in terms 1:10:41.800 --> 1:10:44.960 of doing it continuously in real time. 1:10:44.960 --> 1:10:46.440 You can also do it at scale in terms 1:10:46.440 --> 1:10:50.480 of scaling this to many, many people, to entire populations. 1:10:50.480 --> 1:10:53.840 So potentially, artificial intelligence, 1:10:53.840 --> 1:10:57.440 even in its current state, if you combine it 1:10:57.440 --> 1:11:01.760 with the internet, with the fact that all of our lives 1:11:01.760 --> 1:11:05.120 are moving to digital devices and digital information 1:11:05.120 --> 1:11:08.720 consumption and creation, what you get 1:11:08.720 --> 1:11:14.480 is the possibility to achieve mass manipulation of behavior 1:11:14.480 --> 1:11:16.840 and mass psychological control. 1:11:16.840 --> 1:11:18.520 And this is a very real possibility. 1:11:18.520 --> 1:11:22.080 Yeah, so you're talking about any kind of recommender system. 1:11:22.080 --> 1:11:26.160 Let's look at the YouTube algorithm, Facebook, 1:11:26.160 --> 1:11:29.720 anything that recommends content you should watch next. 1:11:29.720 --> 1:11:32.960 And it's fascinating to think that there's 1:11:32.960 --> 1:11:41.120 some aspects of human behavior that you can say a problem of, 1:11:41.120 --> 1:11:45.400 is this person hold Republican beliefs or Democratic beliefs? 1:11:45.400 --> 1:11:50.240 And this is a trivial, that's an objective function. 1:11:50.240 --> 1:11:52.600 And you can optimize, and you can measure, 1:11:52.600 --> 1:11:54.360 and you can turn everybody into a Republican 1:11:54.360 --> 1:11:56.080 or everybody into a Democrat. 1:11:56.080 --> 1:11:57.840 I do believe it's true. 1:11:57.840 --> 1:12:03.680 So the human mind is very, if you look at the human mind 1:12:03.680 --> 1:12:05.320 as a kind of computer program, it 1:12:05.320 --> 1:12:07.560 has a very large exploit surface. 1:12:07.560 --> 1:12:09.360 It has many, many vulnerabilities. 1:12:09.360 --> 1:12:10.840 Exploit surfaces, yeah. 1:12:10.840 --> 1:12:13.520 Ways you can control it. 1:12:13.520 --> 1:12:16.680 For instance, when it comes to your political beliefs, 1:12:16.680 --> 1:12:19.400 this is very much tied to your identity. 1:12:19.400 --> 1:12:23.040 So for instance, if I'm in control of your news feed 1:12:23.040 --> 1:12:26.000 on your favorite social media platforms, 1:12:26.000 --> 1:12:29.360 this is actually where you're getting your news from. 1:12:29.360 --> 1:12:32.960 And of course, I can choose to only show you 1:12:32.960 --> 1:12:37.120 news that will make you see the world in a specific way. 1:12:37.120 --> 1:12:41.920 But I can also create incentives for you 1:12:41.920 --> 1:12:44.720 to post about some political beliefs. 1:12:44.720 --> 1:12:47.960 And then when I get you to express a statement, 1:12:47.960 --> 1:12:51.840 if it's a statement that me as the controller, 1:12:51.840 --> 1:12:53.800 I want to reinforce. 1:12:53.800 --> 1:12:55.560 I can just show it to people who will agree, 1:12:55.560 --> 1:12:56.880 and they will like it. 1:12:56.880 --> 1:12:59.280 And that will reinforce the statement in your mind. 1:12:59.280 --> 1:13:02.760 If this is a statement I want you to, 1:13:02.760 --> 1:13:05.320 this is a belief I want you to abandon, 1:13:05.320 --> 1:13:09.600 I can, on the other hand, show it to opponents. 1:13:09.600 --> 1:13:10.640 We'll attack you. 1:13:10.640 --> 1:13:12.840 And because they attack you, at the very least, 1:13:12.840 --> 1:13:16.840 next time you will think twice about posting it. 1:13:16.840 --> 1:13:20.280 But maybe you will even start believing this 1:13:20.280 --> 1:13:22.840 because you got pushback. 1:13:22.840 --> 1:13:28.440 So there are many ways in which social media platforms 1:13:28.440 --> 1:13:30.520 can potentially control your opinions. 1:13:30.520 --> 1:13:35.040 And today, so all of these things 1:13:35.040 --> 1:13:38.240 are already being controlled by AI algorithms. 1:13:38.240 --> 1:13:41.880 These algorithms do not have any explicit political goal 1:13:41.880 --> 1:13:42.880 today. 1:13:42.880 --> 1:13:48.680 Well, potentially they could, like if some totalitarian 1:13:48.680 --> 1:13:52.720 government takes over social media platforms 1:13:52.720 --> 1:13:55.360 and decides that now we are going to use this not just 1:13:55.360 --> 1:13:58.040 for mass surveillance, but also for mass opinion control 1:13:58.040 --> 1:13:59.360 and behavior control. 1:13:59.360 --> 1:14:01.840 Very bad things could happen. 1:14:01.840 --> 1:14:06.480 But what's really fascinating and actually quite concerning 1:14:06.480 --> 1:14:11.280 is that even without an explicit intent to manipulate, 1:14:11.280 --> 1:14:14.760 you're already seeing very dangerous dynamics 1:14:14.760 --> 1:14:18.160 in terms of how these content recommendation 1:14:18.160 --> 1:14:19.800 algorithms behave. 1:14:19.800 --> 1:14:24.920 Because right now, the goal, the objective function 1:14:24.920 --> 1:14:28.640 of these algorithms is to maximize engagement, 1:14:28.640 --> 1:14:32.520 which seems fairly innocuous at first. 1:14:32.520 --> 1:14:36.480 However, it is not because content 1:14:36.480 --> 1:14:42.000 that will maximally engage people, get people to react 1:14:42.000 --> 1:14:44.720 in an emotional way, get people to click on something. 1:14:44.720 --> 1:14:52.200 It is very often content that is not 1:14:52.200 --> 1:14:54.400 healthy to the public discourse. 1:14:54.400 --> 1:14:58.200 For instance, fake news are far more 1:14:58.200 --> 1:15:01.320 likely to get you to click on them than real news 1:15:01.320 --> 1:15:06.960 simply because they are not constrained to reality. 1:15:06.960 --> 1:15:11.360 So they can be as outrageous, as surprising, 1:15:11.360 --> 1:15:15.880 as good stories as you want because they're artificial. 1:15:15.880 --> 1:15:18.880 To me, that's an exciting world because so much good 1:15:18.880 --> 1:15:19.560 can come. 1:15:19.560 --> 1:15:24.520 So there's an opportunity to educate people. 1:15:24.520 --> 1:15:31.200 You can balance people's worldview with other ideas. 1:15:31.200 --> 1:15:33.800 So there's so many objective functions. 1:15:33.800 --> 1:15:35.840 The space of objective functions that 1:15:35.840 --> 1:15:40.720 create better civilizations is large, arguably infinite. 1:15:40.720 --> 1:15:43.720 But there's also a large space that 1:15:43.720 --> 1:15:51.480 creates division and destruction, civil war, 1:15:51.480 --> 1:15:53.160 a lot of bad stuff. 1:15:53.160 --> 1:15:56.920 And the worry is, naturally, probably that space 1:15:56.920 --> 1:15:59.160 is bigger, first of all. 1:15:59.160 --> 1:16:04.920 And if we don't explicitly think about what kind of effects 1:16:04.920 --> 1:16:08.320 are going to be observed from different objective functions, 1:16:08.320 --> 1:16:10.160 then we're going to get into trouble. 1:16:10.160 --> 1:16:14.480 But the question is, how do we get into rooms 1:16:14.480 --> 1:16:18.560 and have discussions, so inside Google, inside Facebook, 1:16:18.560 --> 1:16:21.840 inside Twitter, and think about, OK, 1:16:21.840 --> 1:16:24.840 how can we drive up engagement and, at the same time, 1:16:24.840 --> 1:16:28.200 create a good society? 1:16:28.200 --> 1:16:29.560 Is it even possible to have that kind 1:16:29.560 --> 1:16:31.720 of philosophical discussion? 1:16:31.720 --> 1:16:33.080 I think you can definitely try. 1:16:33.080 --> 1:16:37.280 So from my perspective, I would feel rather uncomfortable 1:16:37.280 --> 1:16:41.560 with companies that are uncomfortable with these new 1:16:41.560 --> 1:16:47.120 student algorithms, with them making explicit decisions 1:16:47.120 --> 1:16:50.440 to manipulate people's opinions or behaviors, 1:16:50.440 --> 1:16:53.480 even if the intent is good, because that's 1:16:53.480 --> 1:16:55.200 a very totalitarian mindset. 1:16:55.200 --> 1:16:57.440 So instead, what I would like to see 1:16:57.440 --> 1:16:58.880 is probably never going to happen, 1:16:58.880 --> 1:17:00.360 because it's not super realistic, 1:17:00.360 --> 1:17:02.520 but that's actually something I really care about. 1:17:02.520 --> 1:17:06.280 I would like all these algorithms 1:17:06.280 --> 1:17:10.560 to present configuration settings to their users, 1:17:10.560 --> 1:17:14.600 so that the users can actually make the decision about how 1:17:14.600 --> 1:17:19.000 they want to be impacted by these information 1:17:19.000 --> 1:17:21.960 recommendation, content recommendation algorithms. 1:17:21.960 --> 1:17:24.240 For instance, as a user of something 1:17:24.240 --> 1:17:26.520 like YouTube or Twitter, maybe I want 1:17:26.520 --> 1:17:30.280 to maximize learning about a specific topic. 1:17:30.280 --> 1:17:36.800 So I want the algorithm to feed my curiosity, 1:17:36.800 --> 1:17:38.760 which is in itself a very interesting problem. 1:17:38.760 --> 1:17:41.200 So instead of maximizing my engagement, 1:17:41.200 --> 1:17:44.600 it will maximize how fast and how much I'm learning. 1:17:44.600 --> 1:17:47.360 And it will also take into account the accuracy, 1:17:47.360 --> 1:17:50.680 hopefully, of the information I'm learning. 1:17:50.680 --> 1:17:55.680 So yeah, the user should be able to determine exactly 1:17:55.680 --> 1:17:58.560 how these algorithms are affecting their lives. 1:17:58.560 --> 1:18:03.520 I don't want actually any entity making decisions 1:18:03.520 --> 1:18:09.480 about in which direction they're going to try to manipulate me. 1:18:09.480 --> 1:18:11.680 I want technology. 1:18:11.680 --> 1:18:14.280 So AI, these algorithms are increasingly 1:18:14.280 --> 1:18:18.560 going to be our interface to a world that is increasingly 1:18:18.560 --> 1:18:19.960 made of information. 1:18:19.960 --> 1:18:25.840 And I want everyone to be in control of this interface, 1:18:25.840 --> 1:18:29.160 to interface with the world on their own terms. 1:18:29.160 --> 1:18:32.840 So if someone wants these algorithms 1:18:32.840 --> 1:18:37.640 to serve their own personal growth goals, 1:18:37.640 --> 1:18:40.640 they should be able to configure these algorithms 1:18:40.640 --> 1:18:41.800 in such a way. 1:18:41.800 --> 1:18:46.680 Yeah, but so I know it's painful to have explicit decisions. 1:18:46.680 --> 1:18:51.080 But there is underlying explicit decisions, 1:18:51.080 --> 1:18:53.360 which is some of the most beautiful fundamental 1:18:53.360 --> 1:18:57.400 philosophy that we have before us, 1:18:57.400 --> 1:19:01.120 which is personal growth. 1:19:01.120 --> 1:19:05.680 If I want to watch videos from which I can learn, 1:19:05.680 --> 1:19:08.080 what does that mean? 1:19:08.080 --> 1:19:11.800 So if I have a checkbox that wants to emphasize learning, 1:19:11.800 --> 1:19:15.480 there's still an algorithm with explicit decisions in it 1:19:15.480 --> 1:19:17.800 that would promote learning. 1:19:17.800 --> 1:19:19.200 What does that mean for me? 1:19:19.200 --> 1:19:22.800 For example, I've watched a documentary on flat Earth 1:19:22.800 --> 1:19:23.640 theory, I guess. 1:19:27.280 --> 1:19:28.240 I learned a lot. 1:19:28.240 --> 1:19:29.800 I'm really glad I watched it. 1:19:29.800 --> 1:19:32.560 It was a friend recommended it to me. 1:19:32.560 --> 1:19:35.800 Because I don't have such an allergic reaction to crazy 1:19:35.800 --> 1:19:37.640 people, as my fellow colleagues do. 1:19:37.640 --> 1:19:40.360 But it was very eye opening. 1:19:40.360 --> 1:19:42.120 And for others, it might not be. 1:19:42.120 --> 1:19:45.560 From others, they might just get turned off from that, same 1:19:45.560 --> 1:19:47.160 with Republican and Democrat. 1:19:47.160 --> 1:19:50.200 And it's a non trivial problem. 1:19:50.200 --> 1:19:52.880 And first of all, if it's done well, 1:19:52.880 --> 1:19:56.560 I don't think it's something that wouldn't happen, 1:19:56.560 --> 1:19:59.280 that YouTube wouldn't be promoting, 1:19:59.280 --> 1:20:00.200 or Twitter wouldn't be. 1:20:00.200 --> 1:20:02.280 It's just a really difficult problem, 1:20:02.280 --> 1:20:05.520 how to give people control. 1:20:05.520 --> 1:20:08.960 Well, it's mostly an interface design problem. 1:20:08.960 --> 1:20:11.080 The way I see it, you want to create technology 1:20:11.080 --> 1:20:16.400 that's like a mentor, or a coach, or an assistant, 1:20:16.400 --> 1:20:20.520 so that it's not your boss. 1:20:20.520 --> 1:20:22.560 You are in control of it. 1:20:22.560 --> 1:20:25.760 You are telling it what to do for you. 1:20:25.760 --> 1:20:27.840 And if you feel like it's manipulating you, 1:20:27.840 --> 1:20:31.760 it's not actually doing what you want. 1:20:31.760 --> 1:20:34.920 You should be able to switch to a different algorithm. 1:20:34.920 --> 1:20:36.440 So that's fine tune control. 1:20:36.440 --> 1:20:38.840 You kind of learn that you're trusting 1:20:38.840 --> 1:20:40.080 the human collaboration. 1:20:40.080 --> 1:20:41.920 I mean, that's how I see autonomous vehicles too, 1:20:41.920 --> 1:20:44.480 is giving as much information as possible, 1:20:44.480 --> 1:20:47.240 and you learn that dance yourself. 1:20:47.240 --> 1:20:50.280 Yeah, Adobe, I don't know if you use Adobe product 1:20:50.280 --> 1:20:52.280 for like Photoshop. 1:20:52.280 --> 1:20:55.040 They're trying to see if they can inject YouTube 1:20:55.040 --> 1:20:57.120 into their interface, but basically allow you 1:20:57.120 --> 1:20:59.840 to show you all these videos, 1:20:59.840 --> 1:21:03.320 that everybody's confused about what to do with features. 1:21:03.320 --> 1:21:07.120 So basically teach people by linking to, 1:21:07.120 --> 1:21:10.280 in that way, it's an assistant that uses videos 1:21:10.280 --> 1:21:13.440 as a basic element of information. 1:21:13.440 --> 1:21:18.240 Okay, so what practically should people do 1:21:18.240 --> 1:21:24.000 to try to fight against abuses of these algorithms, 1:21:24.000 --> 1:21:27.400 or algorithms that manipulate us? 1:21:27.400 --> 1:21:29.280 Honestly, it's a very, very difficult problem, 1:21:29.280 --> 1:21:32.800 because to start with, there is very little public awareness 1:21:32.800 --> 1:21:35.040 of these issues. 1:21:35.040 --> 1:21:38.520 Very few people would think there's anything wrong 1:21:38.520 --> 1:21:39.720 with the unused algorithm, 1:21:39.720 --> 1:21:42.040 even though there is actually something wrong already, 1:21:42.040 --> 1:21:44.480 which is that it's trying to maximize engagement 1:21:44.480 --> 1:21:49.880 most of the time, which has very negative side effects. 1:21:49.880 --> 1:21:56.160 So ideally, so the very first thing is to stop 1:21:56.160 --> 1:21:59.560 trying to purely maximize engagement, 1:21:59.560 --> 1:22:06.560 try to propagate content based on popularity, right? 1:22:06.560 --> 1:22:11.040 Instead, take into account the goals 1:22:11.040 --> 1:22:13.560 and the profiles of each user. 1:22:13.560 --> 1:22:16.920 So you will be, one example is, for instance, 1:22:16.920 --> 1:22:20.800 when I look at topic recommendations on Twitter, 1:22:20.800 --> 1:22:24.480 it's like, you know, they have this news tab 1:22:24.480 --> 1:22:25.480 with switch recommendations. 1:22:25.480 --> 1:22:28.480 It's always the worst coverage, 1:22:28.480 --> 1:22:30.360 because it's content that appeals 1:22:30.360 --> 1:22:34.080 to the smallest common denominator 1:22:34.080 --> 1:22:37.080 to all Twitter users, because they're trying to optimize. 1:22:37.080 --> 1:22:39.040 They're purely trying to optimize popularity. 1:22:39.040 --> 1:22:41.320 They're purely trying to optimize engagement. 1:22:41.320 --> 1:22:42.960 But that's not what I want. 1:22:42.960 --> 1:22:46.080 So they should put me in control of some setting 1:22:46.080 --> 1:22:50.360 so that I define what's the objective function 1:22:50.360 --> 1:22:52.200 that Twitter is going to be following 1:22:52.200 --> 1:22:54.120 to show me this content. 1:22:54.120 --> 1:22:57.360 And honestly, so this is all about interface design. 1:22:57.360 --> 1:22:59.440 And we are not, it's not realistic 1:22:59.440 --> 1:23:01.760 to give users control of a bunch of knobs 1:23:01.760 --> 1:23:03.400 that define algorithm. 1:23:03.400 --> 1:23:06.760 Instead, we should purely put them in charge 1:23:06.760 --> 1:23:09.400 of defining the objective function. 1:23:09.400 --> 1:23:13.240 Like, let the user tell us what they want to achieve, 1:23:13.240 --> 1:23:15.280 how they want this algorithm to impact their lives. 1:23:15.280 --> 1:23:16.680 So do you think it is that, 1:23:16.680 --> 1:23:19.360 or do they provide individual article by article 1:23:19.360 --> 1:23:21.600 reward structure where you give a signal, 1:23:21.600 --> 1:23:24.720 I'm glad I saw this, or I'm glad I didn't? 1:23:24.720 --> 1:23:28.480 So like a Spotify type feedback mechanism, 1:23:28.480 --> 1:23:30.680 it works to some extent. 1:23:30.680 --> 1:23:32.000 I'm kind of skeptical about it 1:23:32.000 --> 1:23:34.880 because the only way the algorithm, 1:23:34.880 --> 1:23:39.120 the algorithm will attempt to relate your choices 1:23:39.120 --> 1:23:41.040 with the choices of everyone else, 1:23:41.040 --> 1:23:45.000 which might, you know, if you have an average profile 1:23:45.000 --> 1:23:47.880 that works fine, I'm sure Spotify accommodations work fine 1:23:47.880 --> 1:23:49.560 if you just like mainstream stuff. 1:23:49.560 --> 1:23:53.960 If you don't, it can be, it's not optimal at all actually. 1:23:53.960 --> 1:23:56.040 It'll be in an efficient search 1:23:56.040 --> 1:24:00.800 for the part of the Spotify world that represents you. 1:24:00.800 --> 1:24:02.960 So it's a tough problem, 1:24:02.960 --> 1:24:07.960 but do note that even a feedback system 1:24:07.960 --> 1:24:10.880 like what Spotify has does not give me control 1:24:10.880 --> 1:24:15.000 over what the algorithm is trying to optimize for. 1:24:16.320 --> 1:24:19.360 Well, public awareness, which is what we're doing now, 1:24:19.360 --> 1:24:21.360 is a good place to start. 1:24:21.360 --> 1:24:25.960 Do you have concerns about longterm existential threats 1:24:25.960 --> 1:24:27.360 of artificial intelligence? 1:24:28.280 --> 1:24:31.040 Well, as I was saying, 1:24:31.040 --> 1:24:33.360 our world is increasingly made of information. 1:24:33.360 --> 1:24:36.240 AI algorithms are increasingly going to be our interface 1:24:36.240 --> 1:24:37.880 to this world of information, 1:24:37.880 --> 1:24:41.480 and somebody will be in control of these algorithms. 1:24:41.480 --> 1:24:45.920 And that puts us in any kind of a bad situation, right? 1:24:45.920 --> 1:24:46.880 It has risks. 1:24:46.880 --> 1:24:50.840 It has risks coming from potentially large companies 1:24:50.840 --> 1:24:53.760 wanting to optimize their own goals, 1:24:53.760 --> 1:24:55.960 maybe profit, maybe something else. 1:24:55.960 --> 1:25:00.720 Also from governments who might want to use these algorithms 1:25:00.720 --> 1:25:03.520 as a means of control of the population. 1:25:03.520 --> 1:25:05.000 Do you think there's existential threat 1:25:05.000 --> 1:25:06.320 that could arise from that? 1:25:06.320 --> 1:25:09.120 So existential threat. 1:25:09.120 --> 1:25:13.240 So maybe you're referring to the singularity narrative 1:25:13.240 --> 1:25:15.560 where robots just take over. 1:25:15.560 --> 1:25:18.320 Well, I don't, I'm not terminating robots, 1:25:18.320 --> 1:25:21.000 and I don't believe it has to be a singularity. 1:25:21.000 --> 1:25:24.800 We're just talking to, just like you said, 1:25:24.800 --> 1:25:27.920 the algorithm controlling masses of populations. 1:25:28.920 --> 1:25:31.120 The existential threat being, 1:25:32.640 --> 1:25:36.760 hurt ourselves much like a nuclear war would hurt ourselves. 1:25:36.760 --> 1:25:37.600 That kind of thing. 1:25:37.600 --> 1:25:39.480 I don't think that requires a singularity. 1:25:39.480 --> 1:25:42.560 That requires a loss of control over AI algorithm. 1:25:42.560 --> 1:25:43.560 Yes. 1:25:43.560 --> 1:25:47.000 So I do agree there are concerning trends. 1:25:47.000 --> 1:25:52.000 Honestly, I wouldn't want to make any longterm predictions. 1:25:52.960 --> 1:25:56.000 I don't think today we really have the capability 1:25:56.000 --> 1:25:58.560 to see what the dangers of AI 1:25:58.560 --> 1:26:01.360 are going to be in 50 years, in 100 years. 1:26:01.360 --> 1:26:04.800 I do see that we are already faced 1:26:04.800 --> 1:26:08.840 with concrete and present dangers 1:26:08.840 --> 1:26:11.560 surrounding the negative side effects 1:26:11.560 --> 1:26:14.960 of content recombination systems, of newsfeed algorithms 1:26:14.960 --> 1:26:17.640 concerning algorithmic bias as well. 1:26:18.640 --> 1:26:21.200 So we are delegating more and more 1:26:22.240 --> 1:26:25.080 decision processes to algorithms. 1:26:25.080 --> 1:26:26.760 Some of these algorithms are uncrafted, 1:26:26.760 --> 1:26:29.360 some are learned from data, 1:26:29.360 --> 1:26:31.920 but we are delegating control. 1:26:32.920 --> 1:26:36.280 Sometimes it's a good thing, sometimes not so much. 1:26:36.280 --> 1:26:39.480 And there is in general very little supervision 1:26:39.480 --> 1:26:41.000 of this process, right? 1:26:41.000 --> 1:26:45.400 So we are still in this period of very fast change, 1:26:45.400 --> 1:26:50.400 even chaos, where society is restructuring itself, 1:26:50.920 --> 1:26:53.160 turning into an information society, 1:26:53.160 --> 1:26:54.520 which itself is turning into 1:26:54.520 --> 1:26:58.360 an increasingly automated information passing society. 1:26:58.360 --> 1:27:02.520 And well, yeah, I think the best we can do today 1:27:02.520 --> 1:27:06.040 is try to raise awareness around some of these issues. 1:27:06.040 --> 1:27:07.680 And I think we're actually making good progress. 1:27:07.680 --> 1:27:11.720 If you look at algorithmic bias, for instance, 1:27:12.760 --> 1:27:14.760 three years ago, even two years ago, 1:27:14.760 --> 1:27:17.040 very, very few people were talking about it. 1:27:17.040 --> 1:27:20.320 And now all the big companies are talking about it. 1:27:20.320 --> 1:27:22.360 They are often not in a very serious way, 1:27:22.360 --> 1:27:24.560 but at least it is part of the public discourse. 1:27:24.560 --> 1:27:27.080 You see people in Congress talking about it. 1:27:27.080 --> 1:27:31.960 And it all started from raising awareness. 1:27:31.960 --> 1:27:32.800 Right. 1:27:32.800 --> 1:27:36.080 So in terms of alignment problem, 1:27:36.080 --> 1:27:39.400 trying to teach as we allow algorithms, 1:27:39.400 --> 1:27:41.520 just even recommender systems on Twitter, 1:27:43.640 --> 1:27:47.080 encoding human values and morals, 1:27:48.280 --> 1:27:50.200 decisions that touch on ethics, 1:27:50.200 --> 1:27:52.600 how hard do you think that problem is? 1:27:52.600 --> 1:27:57.240 How do we have lost functions in neural networks 1:27:57.240 --> 1:27:58.640 that have some component, 1:27:58.640 --> 1:28:01.080 some fuzzy components of human morals? 1:28:01.080 --> 1:28:06.080 Well, I think this is really all about objective function engineering, 1:28:06.080 --> 1:28:10.520 which is probably going to be increasingly a topic of concern in the future. 1:28:10.520 --> 1:28:14.640 Like for now, we're just using very naive loss functions 1:28:14.640 --> 1:28:17.760 because the hard part is not actually what you're trying to minimize. 1:28:17.760 --> 1:28:19.040 It's everything else. 1:28:19.040 --> 1:28:22.840 But as the everything else is going to be increasingly automated, 1:28:22.840 --> 1:28:27.040 we're going to be focusing our human attention 1:28:27.040 --> 1:28:30.240 on increasingly high level components, 1:28:30.240 --> 1:28:32.680 like what's actually driving the whole learning system, 1:28:32.680 --> 1:28:33.960 like the objective function. 1:28:33.960 --> 1:28:36.920 So loss function engineering is going to be, 1:28:36.920 --> 1:28:40.640 loss function engineer is probably going to be a job title in the future. 1:28:40.640 --> 1:28:44.520 And then the tooling you're creating with Keras essentially 1:28:44.520 --> 1:28:47.040 takes care of all the details underneath. 1:28:47.040 --> 1:28:52.720 And basically the human expert is needed for exactly that. 1:28:52.720 --> 1:28:53.920 That's the idea. 1:28:53.920 --> 1:28:57.640 Keras is the interface between the data you're collecting 1:28:57.640 --> 1:28:59.080 and the business goals. 1:28:59.080 --> 1:29:03.480 And your job as an engineer is going to be to express your business goals 1:29:03.480 --> 1:29:06.720 and your understanding of your business or your product, 1:29:06.720 --> 1:29:11.840 your system as a kind of loss function or a kind of set of constraints. 1:29:11.840 --> 1:29:19.480 Does the possibility of creating an AGI system excite you or scare you or bore you? 1:29:19.480 --> 1:29:22.080 So intelligence can never really be general. 1:29:22.080 --> 1:29:26.400 You know, at best it can have some degree of generality like human intelligence. 1:29:26.400 --> 1:29:30.640 It also always has some specialization in the same way that human intelligence 1:29:30.640 --> 1:29:33.440 is specialized in a certain category of problems, 1:29:33.440 --> 1:29:35.440 is specialized in the human experience. 1:29:35.440 --> 1:29:37.280 And when people talk about AGI, 1:29:37.280 --> 1:29:42.520 I'm never quite sure if they're talking about very, very smart AI, 1:29:42.520 --> 1:29:45.080 so smart that it's even smarter than humans, 1:29:45.080 --> 1:29:48.000 or they're talking about human like intelligence, 1:29:48.000 --> 1:29:49.680 because these are different things. 1:29:49.680 --> 1:29:54.760 Let's say, presumably I'm oppressing you today with my humanness. 1:29:54.760 --> 1:29:59.240 So imagine that I was in fact a robot. 1:29:59.240 --> 1:30:01.920 So what does that mean? 1:30:01.920 --> 1:30:04.920 That I'm impressing you with natural language processing. 1:30:04.920 --> 1:30:07.840 Maybe if you weren't able to see me, maybe this is a phone call. 1:30:07.840 --> 1:30:10.000 So that kind of system. 1:30:10.000 --> 1:30:11.120 Companion. 1:30:11.120 --> 1:30:15.040 So that's very much about building human like AI. 1:30:15.040 --> 1:30:18.200 And you're asking me, you know, is this an exciting perspective? 1:30:18.200 --> 1:30:19.440 Yes. 1:30:19.440 --> 1:30:21.760 I think so, yes. 1:30:21.760 --> 1:30:28.000 Not so much because of what artificial human like intelligence could do, 1:30:28.000 --> 1:30:30.880 but, you know, from an intellectual perspective, 1:30:30.880 --> 1:30:34.120 I think if you could build truly human like intelligence, 1:30:34.120 --> 1:30:37.240 that means you could actually understand human intelligence, 1:30:37.240 --> 1:30:39.880 which is fascinating, right? 1:30:39.880 --> 1:30:42.680 Human like intelligence is going to require emotions. 1:30:42.680 --> 1:30:44.400 It's going to require consciousness, 1:30:44.400 --> 1:30:49.720 which is not things that would normally be required by an intelligent system. 1:30:49.720 --> 1:30:53.160 If you look at, you know, we were mentioning earlier like science 1:30:53.160 --> 1:30:59.600 as a superhuman problem solving agent or system, 1:30:59.600 --> 1:31:02.120 it does not have consciousness, it doesn't have emotions. 1:31:02.120 --> 1:31:04.320 In general, so emotions, 1:31:04.320 --> 1:31:07.640 I see consciousness as being on the same spectrum as emotions. 1:31:07.640 --> 1:31:12.280 It is a component of the subjective experience 1:31:12.280 --> 1:31:18.800 that is meant very much to guide behavior generation, right? 1:31:18.800 --> 1:31:20.800 It's meant to guide your behavior. 1:31:20.800 --> 1:31:24.520 In general, human intelligence and animal intelligence 1:31:24.520 --> 1:31:29.280 has evolved for the purpose of behavior generation, right? 1:31:29.280 --> 1:31:30.680 Including in a social context. 1:31:30.680 --> 1:31:32.480 So that's why we actually need emotions. 1:31:32.480 --> 1:31:34.920 That's why we need consciousness. 1:31:34.920 --> 1:31:38.360 An artificial intelligence system developed in a different context 1:31:38.360 --> 1:31:42.800 may well never need them, may well never be conscious like science. 1:31:42.800 --> 1:31:47.960 Well, on that point, I would argue it's possible to imagine 1:31:47.960 --> 1:31:51.480 that there's echoes of consciousness in science 1:31:51.480 --> 1:31:55.480 when viewed as an organism, that science is consciousness. 1:31:55.480 --> 1:31:59.160 So, I mean, how would you go about testing this hypothesis? 1:31:59.160 --> 1:32:07.000 How do you probe the subjective experience of an abstract system like science? 1:32:07.000 --> 1:32:10.400 Well, the point of probing any subjective experience is impossible 1:32:10.400 --> 1:32:13.200 because I'm not science, I'm Lex. 1:32:13.200 --> 1:32:20.520 So I can't probe another entity, it's no more than bacteria on my skin. 1:32:20.520 --> 1:32:24.160 You're Lex, I can ask you questions about your subjective experience 1:32:24.160 --> 1:32:28.440 and you can answer me, and that's how I know you're conscious. 1:32:28.440 --> 1:32:31.840 Yes, but that's because we speak the same language. 1:32:31.840 --> 1:32:35.520 You perhaps, we have to speak the language of science in order to ask it. 1:32:35.520 --> 1:32:40.320 Honestly, I don't think consciousness, just like emotions of pain and pleasure, 1:32:40.320 --> 1:32:44.160 is not something that inevitably arises 1:32:44.160 --> 1:32:47.920 from any sort of sufficiently intelligent information processing. 1:32:47.920 --> 1:32:53.920 It is a feature of the mind, and if you've not implemented it explicitly, it is not there. 1:32:53.920 --> 1:32:58.960 So you think it's an emergent feature of a particular architecture. 1:32:58.960 --> 1:33:00.320 So do you think... 1:33:00.320 --> 1:33:02.000 It's a feature in the same sense. 1:33:02.000 --> 1:33:08.240 So, again, the subjective experience is all about guiding behavior. 1:33:08.240 --> 1:33:15.120 If the problems you're trying to solve don't really involve an embodied agent, 1:33:15.120 --> 1:33:19.520 maybe in a social context, generating behavior and pursuing goals like this. 1:33:19.520 --> 1:33:22.160 And if you look at science, that's not really what's happening. 1:33:22.160 --> 1:33:27.920 Even though it is, it is a form of artificial AI, artificial intelligence, 1:33:27.920 --> 1:33:31.920 in the sense that it is solving problems, it is accumulating knowledge, 1:33:31.920 --> 1:33:35.040 accumulating solutions and so on. 1:33:35.040 --> 1:33:39.440 So if you're not explicitly implementing a subjective experience, 1:33:39.440 --> 1:33:44.000 implementing certain emotions and implementing consciousness, 1:33:44.000 --> 1:33:47.360 it's not going to just spontaneously emerge. 1:33:47.360 --> 1:33:48.080 Yeah. 1:33:48.080 --> 1:33:53.200 But so for a system like, human like intelligence system that has consciousness, 1:33:53.200 --> 1:33:55.840 do you think it needs to have a body? 1:33:55.840 --> 1:33:56.720 Yes, definitely. 1:33:56.720 --> 1:33:59.600 I mean, it doesn't have to be a physical body, right? 1:33:59.600 --> 1:34:03.440 And there's not that much difference between a realistic simulation in the real world. 1:34:03.440 --> 1:34:06.400 So there has to be something you have to preserve kind of thing. 1:34:06.400 --> 1:34:11.840 Yes, but human like intelligence can only arise in a human like context. 1:34:11.840 --> 1:34:16.800 Intelligence needs other humans in order for you to demonstrate 1:34:16.800 --> 1:34:19.040 that you have human like intelligence, essentially. 1:34:19.040 --> 1:34:19.540 Yes. 1:34:20.320 --> 1:34:28.080 So what kind of tests and demonstration would be sufficient for you 1:34:28.080 --> 1:34:30.960 to demonstrate human like intelligence? 1:34:30.960 --> 1:34:31.360 Yeah. 1:34:31.360 --> 1:34:35.600 Just out of curiosity, you've talked about in terms of theorem proving 1:34:35.600 --> 1:34:38.000 and program synthesis, I think you've written about 1:34:38.000 --> 1:34:40.480 that there's no good benchmarks for this. 1:34:40.480 --> 1:34:40.720 Yeah. 1:34:40.720 --> 1:34:42.000 That's one of the problems. 1:34:42.000 --> 1:34:46.320 So let's talk program synthesis. 1:34:46.320 --> 1:34:47.760 So what do you imagine is a good... 1:34:48.800 --> 1:34:51.360 I think it's related questions for human like intelligence 1:34:51.360 --> 1:34:52.560 and for program synthesis. 1:34:53.360 --> 1:34:56.080 What's a good benchmark for either or both? 1:34:56.080 --> 1:34:56.480 Right. 1:34:56.480 --> 1:34:59.200 So I mean, you're actually asking two questions, 1:34:59.200 --> 1:35:02.480 which is one is about quantifying intelligence 1:35:02.480 --> 1:35:06.880 and comparing the intelligence of an artificial system 1:35:06.880 --> 1:35:08.480 to the intelligence for human. 1:35:08.480 --> 1:35:13.440 And the other is about the degree to which this intelligence is human like. 1:35:13.440 --> 1:35:15.120 It's actually two different questions. 1:35:16.560 --> 1:35:18.960 So you mentioned earlier the Turing test. 1:35:19.680 --> 1:35:23.200 Well, I actually don't like the Turing test because it's very lazy. 1:35:23.200 --> 1:35:28.720 It's all about completely bypassing the problem of defining and measuring intelligence 1:35:28.720 --> 1:35:34.160 and instead delegating to a human judge or a panel of human judges. 1:35:34.160 --> 1:35:37.120 So it's a total copout, right? 1:35:38.160 --> 1:35:43.200 If you want to measure how human like an agent is, 1:35:43.760 --> 1:35:46.640 I think you have to make it interact with other humans. 1:35:47.600 --> 1:35:53.760 Maybe it's not necessarily a good idea to have these other humans be the judges. 1:35:53.760 --> 1:35:59.280 Maybe you should just observe behavior and compare it to what a human would actually have done. 1:36:00.560 --> 1:36:05.120 When it comes to measuring how smart, how clever an agent is 1:36:05.120 --> 1:36:11.120 and comparing that to the degree of human intelligence. 1:36:11.120 --> 1:36:12.960 So we're already talking about two things, right? 1:36:13.520 --> 1:36:20.320 The degree, kind of like the magnitude of an intelligence and its direction, right? 1:36:20.320 --> 1:36:23.280 Like the norm of a vector and its direction. 1:36:23.280 --> 1:36:32.000 And the direction is like human likeness and the magnitude, the norm is intelligence. 1:36:32.720 --> 1:36:34.080 You could call it intelligence, right? 1:36:34.080 --> 1:36:41.040 So the direction, your sense, the space of directions that are human like is very narrow. 1:36:41.040 --> 1:36:41.200 Yeah. 1:36:42.240 --> 1:36:48.880 So the way you would measure the magnitude of intelligence in a system 1:36:48.880 --> 1:36:54.640 in a way that also enables you to compare it to that of a human. 1:36:54.640 --> 1:36:59.200 Well, if you look at different benchmarks for intelligence today, 1:36:59.200 --> 1:37:04.160 they're all too focused on skill at a given task. 1:37:04.160 --> 1:37:08.720 Like skill at playing chess, skill at playing Go, skill at playing Dota. 1:37:10.720 --> 1:37:15.600 And I think that's not the right way to go about it because you can always 1:37:15.600 --> 1:37:18.240 beat a human at one specific task. 1:37:19.200 --> 1:37:23.920 The reason why our skill at playing Go or juggling or anything is impressive 1:37:23.920 --> 1:37:28.400 is because we are expressing this skill within a certain set of constraints. 1:37:28.400 --> 1:37:32.320 If you remove the constraints, the constraints that we have one lifetime, 1:37:32.320 --> 1:37:36.080 that we have this body and so on, if you remove the context, 1:37:36.080 --> 1:37:40.480 if you have unlimited string data, if you can have access to, you know, 1:37:40.480 --> 1:37:44.640 for instance, if you look at juggling, if you have no restriction on the hardware, 1:37:44.640 --> 1:37:48.400 then achieving arbitrary levels of skill is not very interesting 1:37:48.400 --> 1:37:52.400 and says nothing about the amount of intelligence you've achieved. 1:37:52.400 --> 1:37:57.440 So if you want to measure intelligence, you need to rigorously define what 1:37:57.440 --> 1:38:02.960 intelligence is, which in itself, you know, it's a very challenging problem. 1:38:02.960 --> 1:38:04.320 And do you think that's possible? 1:38:04.320 --> 1:38:06.000 To define intelligence? Yes, absolutely. 1:38:06.000 --> 1:38:09.760 I mean, you can provide, many people have provided, you know, some definition. 1:38:10.560 --> 1:38:12.000 I have my own definition. 1:38:12.000 --> 1:38:13.440 Where does your definition begin? 1:38:13.440 --> 1:38:16.240 Where does your definition begin if it doesn't end? 1:38:16.240 --> 1:38:21.680 Well, I think intelligence is essentially the efficiency 1:38:22.320 --> 1:38:29.760 with which you turn experience into generalizable programs. 1:38:29.760 --> 1:38:32.800 So what that means is it's the efficiency with which 1:38:32.800 --> 1:38:36.720 you turn a sampling of experience space into 1:38:36.720 --> 1:38:46.000 the ability to process a larger chunk of experience space. 1:38:46.000 --> 1:38:52.560 So measuring skill can be one proxy across many different tasks, 1:38:52.560 --> 1:38:54.480 can be one proxy for measuring intelligence. 1:38:54.480 --> 1:38:58.720 But if you want to only measure skill, you should control for two things. 1:38:58.720 --> 1:39:04.960 You should control for the amount of experience that your system has 1:39:04.960 --> 1:39:08.080 and the priors that your system has. 1:39:08.080 --> 1:39:13.120 But if you look at two agents and you give them the same priors 1:39:13.120 --> 1:39:16.160 and you give them the same amount of experience, 1:39:16.160 --> 1:39:21.360 there is one of the agents that is going to learn programs, 1:39:21.360 --> 1:39:25.440 representations, something, a model that will perform well 1:39:25.440 --> 1:39:28.720 on the larger chunk of experience space than the other. 1:39:28.720 --> 1:39:30.960 And that is the smaller agent. 1:39:30.960 --> 1:39:36.960 Yeah. So if you fix the experience, which generate better programs, 1:39:37.680 --> 1:39:39.520 better meaning more generalizable. 1:39:39.520 --> 1:39:40.560 That's really interesting. 1:39:40.560 --> 1:39:42.400 That's a very nice, clean definition of... 1:39:42.400 --> 1:39:47.280 Oh, by the way, in this definition, it is already very obvious 1:39:47.280 --> 1:39:49.440 that intelligence has to be specialized 1:39:49.440 --> 1:39:51.680 because you're talking about experience space 1:39:51.680 --> 1:39:54.080 and you're talking about segments of experience space. 1:39:54.080 --> 1:39:57.200 You're talking about priors and you're talking about experience. 1:39:57.200 --> 1:40:02.480 All of these things define the context in which intelligence emerges. 1:40:04.480 --> 1:40:08.640 And you can never look at the totality of experience space, right? 1:40:09.760 --> 1:40:12.160 So intelligence has to be specialized. 1:40:12.160 --> 1:40:14.960 But it can be sufficiently large, the experience space, 1:40:14.960 --> 1:40:16.080 even though it's specialized. 1:40:16.080 --> 1:40:19.120 There's a certain point when the experience space is large enough 1:40:19.120 --> 1:40:21.440 to where it might as well be general. 1:40:22.000 --> 1:40:23.920 It feels general. It looks general. 1:40:23.920 --> 1:40:25.680 Sure. I mean, it's very relative. 1:40:25.680 --> 1:40:29.360 Like, for instance, many people would say human intelligence is general. 1:40:29.360 --> 1:40:31.200 In fact, it is quite specialized. 1:40:32.800 --> 1:40:37.120 We can definitely build systems that start from the same innate priors 1:40:37.120 --> 1:40:39.120 as what humans have at birth. 1:40:39.120 --> 1:40:42.320 Because we already understand fairly well 1:40:42.320 --> 1:40:44.480 what sort of priors we have as humans. 1:40:44.480 --> 1:40:46.080 Like many people have worked on this problem. 1:40:46.800 --> 1:40:51.040 Most notably, Elisabeth Spelke from Harvard. 1:40:51.040 --> 1:40:52.240 I don't know if you know her. 1:40:52.240 --> 1:40:56.000 She's worked a lot on what she calls core knowledge. 1:40:56.000 --> 1:41:00.640 And it is very much about trying to determine and describe 1:41:00.640 --> 1:41:02.320 what priors we are born with. 1:41:02.320 --> 1:41:04.720 Like language skills and so on, all that kind of stuff. 1:41:04.720 --> 1:41:05.220 Exactly. 1:41:06.880 --> 1:41:11.440 So we have some pretty good understanding of what priors we are born with. 1:41:11.440 --> 1:41:12.560 So we could... 1:41:13.760 --> 1:41:17.760 So I've actually been working on a benchmark for the past couple years, 1:41:17.760 --> 1:41:18.640 you know, on and off. 1:41:18.640 --> 1:41:20.480 I hope to be able to release it at some point. 1:41:20.480 --> 1:41:21.760 That's exciting. 1:41:21.760 --> 1:41:25.680 The idea is to measure the intelligence of systems 1:41:26.800 --> 1:41:28.640 by countering for priors, 1:41:28.640 --> 1:41:30.480 countering for amount of experience, 1:41:30.480 --> 1:41:34.800 and by assuming the same priors as what humans are born with. 1:41:34.800 --> 1:41:39.520 So that you can actually compare these scores to human intelligence. 1:41:39.520 --> 1:41:43.280 You can actually have humans pass the same test in a way that's fair. 1:41:43.280 --> 1:41:52.320 Yeah. And so importantly, such a benchmark should be such that any amount 1:41:52.960 --> 1:41:55.920 of practicing does not increase your score. 1:41:56.480 --> 1:42:00.560 So try to picture a game where no matter how much you play this game, 1:42:01.600 --> 1:42:05.040 that does not change your skill at the game. 1:42:05.040 --> 1:42:05.920 Can you picture that? 1:42:05.920 --> 1:42:11.040 As a person who deeply appreciates practice, I cannot actually. 1:42:11.040 --> 1:42:16.560 There's actually a very simple trick. 1:42:16.560 --> 1:42:19.440 So in order to come up with a task, 1:42:19.440 --> 1:42:21.760 so the only thing you can measure is skill at the task. 1:42:21.760 --> 1:42:22.320 Yes. 1:42:22.320 --> 1:42:24.800 All tasks are going to involve priors. 1:42:24.800 --> 1:42:25.600 Yes. 1:42:25.600 --> 1:42:29.920 The trick is to know what they are and to describe that. 1:42:29.920 --> 1:42:33.760 And then you make sure that this is the same set of priors as what humans start with. 1:42:33.760 --> 1:42:38.560 So you create a task that assumes these priors, that exactly documents these priors, 1:42:38.560 --> 1:42:42.240 so that the priors are made explicit and there are no other priors involved. 1:42:42.240 --> 1:42:48.960 And then you generate a certain number of samples in experience space for this task, right? 1:42:49.840 --> 1:42:56.320 And this, for one task, assuming that the task is new for the agent passing it, 1:42:56.320 --> 1:43:04.320 that's one test of this definition of intelligence that we set up. 1:43:04.320 --> 1:43:06.880 And now you can scale that to many different tasks, 1:43:06.880 --> 1:43:10.480 that each task should be new to the agent passing it, right? 1:43:11.360 --> 1:43:14.480 And also it should be human interpretable and understandable 1:43:14.480 --> 1:43:16.880 so that you can actually have a human pass the same test. 1:43:16.880 --> 1:43:19.760 And then you can compare the score of your machine and the score of your human. 1:43:19.760 --> 1:43:20.720 Which could be a lot of stuff. 1:43:20.720 --> 1:43:23.040 You could even start a task like MNIST. 1:43:23.040 --> 1:43:28.800 Just as long as you start with the same set of priors. 1:43:28.800 --> 1:43:34.080 So the problem with MNIST, humans are already trying to recognize digits, right? 1:43:35.600 --> 1:43:40.960 But let's say we're considering objects that are not digits, 1:43:42.400 --> 1:43:43.920 some completely arbitrary patterns. 1:43:44.480 --> 1:43:48.880 Well, humans already come with visual priors about how to process that. 1:43:48.880 --> 1:43:54.080 So in order to make the game fair, you would have to isolate these priors 1:43:54.080 --> 1:43:57.280 and describe them and then express them as computational rules. 1:43:57.280 --> 1:44:01.680 Having worked a lot with vision science people, that's exceptionally difficult. 1:44:01.680 --> 1:44:03.120 A lot of progress has been made. 1:44:03.120 --> 1:44:08.080 There's been a lot of good tests and basically reducing all of human vision into some good priors. 1:44:08.640 --> 1:44:10.960 We're still probably far away from that perfectly, 1:44:10.960 --> 1:44:14.640 but as a start for a benchmark, that's an exciting possibility. 1:44:14.640 --> 1:44:24.240 Yeah, so Elisabeth Spelke actually lists objectness as one of the core knowledge priors. 1:44:24.800 --> 1:44:25.920 Objectness, cool. 1:44:25.920 --> 1:44:26.880 Objectness, yeah. 1:44:27.440 --> 1:44:31.520 So we have priors about objectness, like about the visual space, about time, 1:44:31.520 --> 1:44:34.240 about agents, about goal oriented behavior. 1:44:35.280 --> 1:44:39.280 We have many different priors, but what's interesting is that, 1:44:39.280 --> 1:44:43.920 sure, we have this pretty diverse and rich set of priors, 1:44:43.920 --> 1:44:46.880 but it's also not that diverse, right? 1:44:46.880 --> 1:44:50.800 We are not born into this world with a ton of knowledge about the world, 1:44:50.800 --> 1:44:57.840 with only a small set of core knowledge. 1:44:58.640 --> 1:45:05.040 Yeah, sorry, do you have a sense of how it feels to us humans that that set is not that large? 1:45:05.040 --> 1:45:09.600 But just even the nature of time that we kind of integrate pretty effectively 1:45:09.600 --> 1:45:11.600 through all of our perception, all of our reasoning, 1:45:12.640 --> 1:45:17.680 maybe how, you know, do you have a sense of how easy it is to encode those priors? 1:45:17.680 --> 1:45:24.560 Maybe it requires building a universe and then the human brain in order to encode those priors. 1:45:25.440 --> 1:45:28.640 Or do you have a hope that it can be listed like an axiomatic? 1:45:28.640 --> 1:45:29.280 I don't think so. 1:45:29.280 --> 1:45:33.040 So you have to keep in mind that any knowledge about the world that we are 1:45:33.040 --> 1:45:41.120 born with is something that has to have been encoded into our DNA by evolution at some point. 1:45:41.120 --> 1:45:41.440 Right. 1:45:41.440 --> 1:45:45.440 And DNA is a very, very low bandwidth medium. 1:45:46.000 --> 1:45:51.200 Like it's extremely long and expensive to encode anything into DNA because first of all, 1:45:52.560 --> 1:45:57.440 you need some sort of evolutionary pressure to guide this writing process. 1:45:57.440 --> 1:46:03.440 And then, you know, the higher level of information you're trying to write, the longer it's going to take. 1:46:04.480 --> 1:46:13.520 And the thing in the environment that you're trying to encode knowledge about has to be stable 1:46:13.520 --> 1:46:15.280 over this duration. 1:46:15.280 --> 1:46:20.960 So you can only encode into DNA things that constitute an evolutionary advantage. 1:46:20.960 --> 1:46:25.280 So this is actually a very small subset of all possible knowledge about the world. 1:46:25.280 --> 1:46:32.080 You can only encode things that are stable, that are true, over very, very long periods of time, 1:46:32.080 --> 1:46:33.680 typically millions of years. 1:46:33.680 --> 1:46:38.720 For instance, we might have some visual prior about the shape of snakes, right? 1:46:38.720 --> 1:46:43.920 But what makes a face, what's the difference between a face and an art face? 1:46:44.560 --> 1:46:48.080 But consider this interesting question. 1:46:48.080 --> 1:46:56.640 Do we have any innate sense of the visual difference between a male face and a female face? 1:46:56.640 --> 1:46:57.600 What do you think? 1:46:58.640 --> 1:46:59.840 For a human, I mean. 1:46:59.840 --> 1:47:04.000 I would have to look back into evolutionary history when the genders emerged. 1:47:04.000 --> 1:47:06.240 But yeah, most... 1:47:06.240 --> 1:47:09.840 I mean, the faces of humans are quite different from the faces of great apes. 1:47:10.640 --> 1:47:11.600 Great apes, right? 1:47:12.880 --> 1:47:13.600 Yeah. 1:47:13.600 --> 1:47:14.800 That's interesting. 1:47:14.800 --> 1:47:22.800 Yeah, you couldn't tell the face of a female chimpanzee from the face of a male chimpanzee, 1:47:22.800 --> 1:47:23.440 probably. 1:47:23.440 --> 1:47:26.160 Yeah, and I don't think most humans have all that ability. 1:47:26.160 --> 1:47:33.280 So we do have innate knowledge of what makes a face, but it's actually impossible for us to 1:47:33.280 --> 1:47:40.320 have any DNA encoded knowledge of the difference between a female human face and a male human face 1:47:40.320 --> 1:47:50.560 because that knowledge, that information came up into the world actually very recently. 1:47:50.560 --> 1:47:56.400 If you look at the slowness of the process of encoding knowledge into DNA. 1:47:56.400 --> 1:47:57.360 Yeah, so that's interesting. 1:47:57.360 --> 1:48:02.080 That's a really powerful argument that DNA is a low bandwidth and it takes a long time to encode. 1:48:02.800 --> 1:48:05.200 That naturally creates a very efficient encoding. 1:48:05.200 --> 1:48:12.800 But one important consequence of this is that, so yes, we are born into this world with a bunch of 1:48:12.800 --> 1:48:17.600 knowledge, sometimes high level knowledge about the world, like the shape, the rough shape of a 1:48:17.600 --> 1:48:19.520 snake, of the rough shape of a face. 1:48:20.480 --> 1:48:26.960 But importantly, because this knowledge takes so long to write, almost all of this innate 1:48:26.960 --> 1:48:32.080 knowledge is shared with our cousins, with great apes, right? 1:48:32.080 --> 1:48:35.600 So it is not actually this innate knowledge that makes us special. 1:48:36.320 --> 1:48:42.000 But to throw it right back at you from the earlier on in our discussion, it's that encoding 1:48:42.960 --> 1:48:48.320 might also include the entirety of the environment of Earth. 1:48:49.360 --> 1:48:49.920 To some extent. 1:48:49.920 --> 1:48:56.480 So it can include things that are important to survival and production, so for which there is 1:48:56.480 --> 1:49:02.880 some evolutionary pressure, and things that are stable, constant over very, very, very long time 1:49:02.880 --> 1:49:03.380 periods. 1:49:04.160 --> 1:49:06.320 And honestly, it's not that much information. 1:49:06.320 --> 1:49:14.400 There's also, besides the bandwidths constraint and the constraints of the writing process, 1:49:14.400 --> 1:49:21.440 there's also memory constraints, like DNA, the part of DNA that deals with the human brain, 1:49:21.440 --> 1:49:22.640 it's actually fairly small. 1:49:22.640 --> 1:49:25.520 It's like, you know, on the order of megabytes, right? 1:49:25.520 --> 1:49:29.600 There's not that much high level knowledge about the world you can encode. 1:49:31.600 --> 1:49:38.880 That's quite brilliant and hopeful for a benchmark that you're referring to of encoding 1:49:38.880 --> 1:49:39.360 priors. 1:49:39.360 --> 1:49:43.120 I actually look forward to, I'm skeptical whether you can do it in the next couple of 1:49:43.120 --> 1:49:44.320 years, but hopefully. 1:49:45.040 --> 1:49:45.760 I've been working. 1:49:45.760 --> 1:49:49.920 So honestly, it's a very simple benchmark, and it's not like a big breakthrough or anything. 1:49:49.920 --> 1:49:53.200 It's more like a fun side project, right? 1:49:53.200 --> 1:49:55.680 But these fun, so is ImageNet. 1:49:56.480 --> 1:50:04.080 These fun side projects could launch entire groups of efforts towards creating reasoning 1:50:04.080 --> 1:50:04.960 systems and so on. 1:50:04.960 --> 1:50:05.440 And I think... 1:50:05.440 --> 1:50:06.160 Yeah, that's the goal. 1:50:06.160 --> 1:50:12.080 It's trying to measure strong generalization, to measure the strength of abstraction in 1:50:12.080 --> 1:50:16.960 our minds, well, in our minds and in artificial intelligence agencies. 1:50:16.960 --> 1:50:24.800 And if there's anything true about this science organism is its individual cells love competition. 1:50:24.800 --> 1:50:26.800 So and benchmarks encourage competition. 1:50:26.800 --> 1:50:29.520 So that's an exciting possibility. 1:50:29.520 --> 1:50:32.640 If you, do you think an AI winter is coming? 1:50:33.520 --> 1:50:34.640 And how do we prevent it? 1:50:35.440 --> 1:50:36.080 Not really. 1:50:36.080 --> 1:50:42.160 So an AI winter is something that would occur when there's a big mismatch between how we 1:50:42.160 --> 1:50:47.280 are selling the capabilities of AI and the actual capabilities of AI. 1:50:47.280 --> 1:50:50.560 And today, some deep learning is creating a lot of value. 1:50:50.560 --> 1:50:56.240 And it will keep creating a lot of value in the sense that these models are applicable 1:50:56.240 --> 1:51:00.000 to a very wide range of problems that are relevant today. 1:51:00.000 --> 1:51:05.120 And we are only just getting started with applying these algorithms to every problem 1:51:05.120 --> 1:51:06.320 they could be solving. 1:51:06.320 --> 1:51:10.160 So deep learning will keep creating a lot of value for the time being. 1:51:10.160 --> 1:51:15.920 What's concerning, however, is that there's a lot of hype around deep learning and around 1:51:15.920 --> 1:51:16.240 AI. 1:51:16.240 --> 1:51:22.000 There are lots of people are overselling the capabilities of these systems, not just 1:51:22.000 --> 1:51:27.760 the capabilities, but also overselling the fact that they might be more or less, you 1:51:27.760 --> 1:51:36.640 know, brain like, like given the kind of a mystical aspect, these technologies and also 1:51:36.640 --> 1:51:43.840 overselling the pace of progress, which, you know, it might look fast in the sense that 1:51:43.840 --> 1:51:46.480 we have this exponentially increasing number of papers. 1:51:47.760 --> 1:51:52.960 But again, that's just a simple consequence of the fact that we have ever more people 1:51:52.960 --> 1:51:53.840 coming into the field. 1:51:54.400 --> 1:51:57.440 It doesn't mean the progress is actually exponentially fast. 1:51:58.640 --> 1:52:02.720 Let's say you're trying to raise money for your startup or your research lab. 1:52:02.720 --> 1:52:09.120 You might want to tell, you know, a grandiose story to investors about how deep learning 1:52:09.120 --> 1:52:14.240 is just like the brain and how it can solve all these incredible problems like self driving 1:52:14.240 --> 1:52:15.760 and robotics and so on. 1:52:15.760 --> 1:52:19.440 And maybe you can tell them that the field is progressing so fast and we are going to 1:52:19.440 --> 1:52:23.040 have AGI within 15 years or even 10 years. 1:52:23.040 --> 1:52:25.920 And none of this is true. 1:52:25.920 --> 1:52:32.800 And every time you're like saying these things and an investor or, you know, a decision maker 1:52:32.800 --> 1:52:41.680 believes them, well, this is like the equivalent of taking on credit card debt, but for trust, 1:52:41.680 --> 1:52:42.480 right? 1:52:42.480 --> 1:52:50.160 And maybe this will, you know, this will be what enables you to raise a lot of money, 1:52:50.160 --> 1:52:54.320 but ultimately you are creating damage, you are damaging the field. 1:52:54.320 --> 1:53:00.160 So that's the concern is that that debt, that's what happens with the other AI winters is 1:53:00.160 --> 1:53:04.160 the concern is you actually tweeted about this with autonomous vehicles, right? 1:53:04.160 --> 1:53:08.960 There's almost every single company now have promised that they will have full autonomous 1:53:08.960 --> 1:53:11.760 vehicles by 2021, 2022. 1:53:11.760 --> 1:53:18.080 That's a good example of the consequences of over hyping the capabilities of AI and 1:53:18.080 --> 1:53:19.280 the pace of progress. 1:53:19.280 --> 1:53:25.200 So because I work especially a lot recently in this area, I have a deep concern of what 1:53:25.200 --> 1:53:30.400 happens when all of these companies after I've invested billions have a meeting and 1:53:30.400 --> 1:53:33.600 say, how much do we actually, first of all, do we have an autonomous vehicle? 1:53:33.600 --> 1:53:35.280 The answer will definitely be no. 1:53:35.840 --> 1:53:40.560 And second will be, wait a minute, we've invested one, two, three, four billion dollars 1:53:40.560 --> 1:53:43.120 into this and we made no profit. 1:53:43.120 --> 1:53:49.200 And the reaction to that may be going very hard in other directions that might impact 1:53:49.200 --> 1:53:50.400 even other industries. 1:53:50.400 --> 1:53:55.520 And that's what we call an AI winter is when there is backlash where no one believes any 1:53:55.520 --> 1:53:59.360 of these promises anymore because they've turned that to be big lies the first time 1:53:59.360 --> 1:54:00.240 around. 1:54:00.240 --> 1:54:06.000 And this will definitely happen to some extent for autonomous vehicles because the public 1:54:06.000 --> 1:54:13.360 and decision makers have been convinced that around 2015, they've been convinced by these 1:54:13.360 --> 1:54:19.600 people who are trying to raise money for their startups and so on, that L5 driving was coming 1:54:19.600 --> 1:54:22.880 in maybe 2016, maybe 2017, maybe 2018. 1:54:22.880 --> 1:54:26.080 Now we're in 2019, we're still waiting for it. 1:54:27.600 --> 1:54:32.800 And so I don't believe we are going to have a full on AI winter because we have these 1:54:32.800 --> 1:54:36.640 technologies that are producing a tremendous amount of real value. 1:54:37.680 --> 1:54:39.920 But there is also too much hype. 1:54:39.920 --> 1:54:43.520 So there will be some backlash, especially there will be backlash. 1:54:44.960 --> 1:54:53.040 So some startups are trying to sell the dream of AGI and the fact that AGI is going to create 1:54:53.040 --> 1:54:53.760 infinite value. 1:54:53.760 --> 1:54:55.680 Like AGI is like a free lunch. 1:54:55.680 --> 1:55:02.800 Like if you can develop an AI system that passes a certain threshold of IQ or something, 1:55:02.800 --> 1:55:04.400 then suddenly you have infinite value. 1:55:04.400 --> 1:55:14.160 And well, there are actually lots of investors buying into this idea and they will wait maybe 1:55:14.160 --> 1:55:17.760 10, 15 years and nothing will happen. 1:55:17.760 --> 1:55:22.560 And the next time around, well, maybe there will be a new generation of investors. 1:55:22.560 --> 1:55:23.360 No one will care. 1:55:24.800 --> 1:55:27.280 Human memory is fairly short after all. 1:55:27.280 --> 1:55:34.320 I don't know about you, but because I've spoken about AGI sometimes poetically, I get a lot 1:55:34.320 --> 1:55:42.000 of emails from people giving me, they're usually like a large manifestos of they've, they say 1:55:42.000 --> 1:55:47.200 to me that they have created an AGI system or they know how to do it. 1:55:47.200 --> 1:55:48.880 And there's a long write up of how to do it. 1:55:48.880 --> 1:55:50.560 I get a lot of these emails, yeah. 1:55:50.560 --> 1:55:57.760 They're a little bit feel like it's generated by an AI system actually, but there's usually 1:55:57.760 --> 1:56:06.640 no diagram, you have a transformer generating crank papers about AGI. 1:56:06.640 --> 1:56:12.160 So the question is about, because you've been such a good, you have a good radar for crank 1:56:12.160 --> 1:56:16.720 papers, how do we know they're not onto something? 1:56:16.720 --> 1:56:24.240 How do I, so when you start to talk about AGI or anything like the reasoning benchmarks 1:56:24.240 --> 1:56:28.160 and so on, so something that doesn't have a benchmark, it's really difficult to know. 1:56:29.120 --> 1:56:34.560 I mean, I talked to Jeff Hawkins, who's really looking at neuroscience approaches to how, 1:56:35.200 --> 1:56:41.520 and there's some, there's echoes of really interesting ideas in at least Jeff's case, 1:56:41.520 --> 1:56:42.320 which he's showing. 1:56:43.280 --> 1:56:45.040 How do you usually think about this? 1:56:46.640 --> 1:56:52.880 Like preventing yourself from being too narrow minded and elitist about deep learning, it 1:56:52.880 --> 1:56:56.720 has to work on these particular benchmarks, otherwise it's trash. 1:56:56.720 --> 1:57:05.280 Well, you know, the thing is, intelligence does not exist in the abstract. 1:57:05.280 --> 1:57:07.200 Intelligence has to be applied. 1:57:07.200 --> 1:57:11.040 So if you don't have a benchmark, if you have an improvement in some benchmark, maybe it's 1:57:11.040 --> 1:57:12.400 a new benchmark, right? 1:57:12.400 --> 1:57:16.640 Maybe it's not something we've been looking at before, but you do need a problem that 1:57:16.640 --> 1:57:17.360 you're trying to solve. 1:57:17.360 --> 1:57:20.000 You're not going to come up with a solution without a problem. 1:57:20.000 --> 1:57:25.520 So you, general intelligence, I mean, you've clearly highlighted generalization. 1:57:26.320 --> 1:57:31.200 If you want to claim that you have an intelligence system, it should come with a benchmark. 1:57:31.200 --> 1:57:35.760 It should, yes, it should display capabilities of some kind. 1:57:35.760 --> 1:57:41.840 It should show that it can create some form of value, even if it's a very artificial form 1:57:41.840 --> 1:57:42.800 of value. 1:57:42.800 --> 1:57:48.800 And that's also the reason why you don't actually need to care about telling which papers have 1:57:48.800 --> 1:57:52.000 actually some hidden potential and which do not. 1:57:53.120 --> 1:57:59.200 Because if there is a new technique that's actually creating value, this is going to 1:57:59.200 --> 1:58:02.480 be brought to light very quickly because it's actually making a difference. 1:58:02.480 --> 1:58:08.160 So it's the difference between something that is ineffectual and something that is actually 1:58:08.160 --> 1:58:08.800 useful. 1:58:08.800 --> 1:58:14.080 And ultimately usefulness is our guide, not just in this field, but if you look at science 1:58:14.080 --> 1:58:18.720 in general, maybe there are many, many people over the years that have had some really interesting 1:58:19.440 --> 1:58:22.800 theories of everything, but they were just completely useless. 1:58:22.800 --> 1:58:27.280 And you don't actually need to tell the interesting theories from the useless theories. 1:58:28.000 --> 1:58:34.080 All you need is to see, is this actually having an effect on something else? 1:58:34.080 --> 1:58:35.360 Is this actually useful? 1:58:35.360 --> 1:58:36.800 Is this making an impact or not? 1:58:37.600 --> 1:58:38.640 That's beautifully put. 1:58:38.640 --> 1:58:43.680 I mean, the same applies to quantum mechanics, to string theory, to the holographic principle. 1:58:43.680 --> 1:58:45.280 We are doing deep learning because it works. 1:58:46.960 --> 1:58:52.720 Before it started working, people considered people working on neural networks as cranks 1:58:52.720 --> 1:58:53.120 very much. 1:58:54.560 --> 1:58:56.320 No one was working on this anymore. 1:58:56.320 --> 1:58:59.120 And now it's working, which is what makes it valuable. 1:58:59.120 --> 1:59:00.320 It's not about being right. 1:59:01.120 --> 1:59:02.560 It's about being effective. 1:59:02.560 --> 1:59:08.080 And nevertheless, the individual entities of this scientific mechanism, just like Yoshua 1:59:08.080 --> 1:59:12.480 Banjo or Jan Lekun, they, while being called cranks, stuck with it. 1:59:12.480 --> 1:59:12.880 Right? 1:59:12.880 --> 1:59:13.280 Yeah. 1:59:13.280 --> 1:59:17.840 And so us individual agents, even if everyone's laughing at us, just stick with it. 1:59:18.880 --> 1:59:21.840 If you believe you have something, you should stick with it and see it through. 1:59:23.520 --> 1:59:25.920 That's a beautiful inspirational message to end on. 1:59:25.920 --> 1:59:27.600 Francois, thank you so much for talking today. 1:59:27.600 --> 1:59:28.640 That was amazing. 1:59:28.640 --> 1:59:44.000 Thank you.