WEBVTT 00:00.000 --> 00:03.720 The following is a conversation with Francois Chalet. 00:03.720 --> 00:07.300 He's the creator of Keras, which is an open source deep learning 00:07.300 --> 00:10.560 library that is designed to enable fast, user friendly 00:10.560 --> 00:13.600 experimentation 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 --> 00:24.120 a while ago. 00:24.120 --> 00:27.560 Meaning, if you want to create, train, and use 00:27.560 --> 00:31.040 neural networks, probably the easiest and most popular option 00:31.040 --> 00:34.840 is to use Keras inside TensorFlow. 00:34.840 --> 00:37.760 Aside from creating an exceptionally useful and popular 00:37.760 --> 00:41.920 library, Francois is also a world class AI researcher 00:41.920 --> 00:44.560 and software engineer at Google. 00:44.560 --> 00:48.080 And he's definitely an outspoken, if not controversial, 00:48.080 --> 00:51.480 personality in the AI world, especially 00:51.480 --> 00:53.720 in the realm of ideas around the future 00:53.720 --> 00:55.920 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 --> 01:04.160 give us five stars on iTunes, support on Patreon, 01:04.160 --> 01:06.080 or simply connect with me on Twitter 01:06.080 --> 01:09.960 at Lex Freedman, spelled F R I D M A N. 01:09.960 --> 01:14.880 And now, here's my conversation with Francois Chalet. 01:14.880 --> 01:17.320 You're known for not sugarcoating your opinions 01:17.320 --> 01:19.640 and speaking your mind about ideas in AI, especially 01:19.640 --> 01:21.120 on Twitter. 01:21.120 --> 01:22.800 That's one of my favorite Twitter accounts. 01:22.800 --> 01:26.360 So what's one of the more controversial ideas 01:26.360 --> 01:30.440 you've expressed online and gotten some heat for? 01:30.440 --> 01:33.080 How do you pick? 01:33.080 --> 01:33.920 How do I pick? 01:33.920 --> 01:38.280 Yeah, no, I think if you go through the trouble of maintaining 01:38.280 --> 01:41.880 Twitter accounts, you might as well speak your mind. 01:41.880 --> 01:44.640 Otherwise, what's even the point of doing Twitter accounts, 01:44.640 --> 01:48.600 like getting an eye scar and just leaving it in the garage? 01:48.600 --> 01:50.360 Yeah, so that's one thing for which 01:50.360 --> 01:53.640 I got a lot of pushback. 01:53.640 --> 01:56.720 Perhaps that time, I wrote something 01:56.720 --> 02:00.960 about the idea of intelligence explosion. 02:00.960 --> 02:05.680 And I was questioning the idea and the reasoning behind this 02:05.680 --> 02:06.880 idea. 02:06.880 --> 02:09.720 And I got a lot of pushback on that. 02:09.720 --> 02:11.840 I got a lot of flak for it. 02:11.840 --> 02:14.360 So yeah, so intelligence explosion, I'm sure you're familiar 02:14.360 --> 02:15.800 with the idea, but it's the idea 02:15.800 --> 02:21.360 that if you were to build general AI problems 02:21.360 --> 02:27.600 solving algorithms, well, the problem of building such an AI, 02:27.600 --> 02:30.640 that itself is a problem that could be solved by your AI. 02:30.640 --> 02:33.840 And maybe it could be solved better than what humans can do. 02:33.840 --> 02:36.920 So your AI could start tweaking its own algorithm, 02:36.920 --> 02:39.640 could start making a better version of itself. 02:39.640 --> 02:43.320 And so on, iteratively, in a recursive fashion, 02:43.320 --> 02:47.360 and so you would end up with an AI 02:47.360 --> 02:50.920 with exponentially increasing intelligence. 02:50.920 --> 02:55.880 And I was basically questioning this idea. 02:55.880 --> 02:59.080 First of all, because the notion of intelligence explosion 02:59.080 --> 03:02.240 uses an implicit definition of intelligence 03:02.240 --> 03:05.400 that doesn't sound quite right to me. 03:05.400 --> 03:11.200 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 --> 03:26.720 between a brain, a body, like embodied intelligence, 03:26.720 --> 03:28.320 and an environment. 03:28.320 --> 03:30.720 And if you're missing one of these pieces, 03:30.720 --> 03:33.840 then you cannot really define intelligence anymore. 03:33.840 --> 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 --> 03:43.000 So first of all, you're crushing the dreams of many people. 03:43.000 --> 03:46.000 So let's look at Sam Harris. 03:46.000 --> 03:48.680 Actually, a lot of physicists, Max Tegmark, 03:48.680 --> 03:53.600 people who think the universe is an information processing 03:53.600 --> 03:54.640 system. 03:54.640 --> 03:57.680 Our brain is kind of an information processing system. 03:57.680 --> 04:00.040 So what's the theoretical limit? 04:00.040 --> 04:04.840 It doesn't make sense that there should be some, 04:04.840 --> 04:08.080 it seems naive to think that our own brain is somehow 04:08.080 --> 04:11.600 the limit of the capabilities and this information. 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 --> 04:18.040 And then if you just scale it, if you're 04:18.040 --> 04:20.880 able to build something that's on par with the brain, 04:20.880 --> 04:24.000 you just, the process that builds it just continues 04:24.000 --> 04:26.360 and it will improve exponentially. 04:26.360 --> 04:30.120 So that's the logic that's used actually 04:30.120 --> 04:33.920 by almost everybody that is worried 04:33.920 --> 04:36.880 about super human intelligence. 04:36.880 --> 04:39.800 Yeah, so you're trying to make, so most people 04:39.800 --> 04:42.320 who are skeptical of that are kind of like, 04:42.320 --> 04:44.360 this doesn't, their thought process, 04:44.360 --> 04:46.520 this doesn't feel right. 04:46.520 --> 04:47.680 Like that's for me as well. 04:47.680 --> 04:52.320 So I'm more like, it doesn't, the whole thing is shrouded 04:52.320 --> 04:55.840 in mystery 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 --> 05:00.680 This doesn't feel like that's how the brain works. 05:00.680 --> 05:02.400 And you're trying to, with your blog post 05:02.400 --> 05:05.680 and now making it a little more explicit. 05:05.680 --> 05:10.280 So one idea is that the brain isn't, 05:10.280 --> 05:13.840 exists alone, it exists within the environment. 05:13.840 --> 05:17.520 So you can't exponentially, you would have to somehow 05:17.520 --> 05:19.360 exponentially improve the environment 05:19.360 --> 05:22.280 and the brain together, almost yet in order 05:22.280 --> 05:26.280 to create something that's much smarter 05:26.280 --> 05:29.120 in some kind of, of course we don't have 05:29.120 --> 05:30.560 a definition of intelligence. 05:30.560 --> 05:31.880 That's correct, that's correct. 05:31.880 --> 05:34.560 I don't think, you should look at very smart people 05:34.560 --> 05:37.840 to the even humans, not even talking about AI's. 05:37.840 --> 05:40.000 I don't think their brain and the performance 05:40.000 --> 05:42.520 of their brain is the bottleneck 05:42.520 --> 05:45.760 to their expressed intelligence, to their achievements. 05:47.160 --> 05:50.480 You cannot just tweak one part of this system, 05:50.480 --> 05:53.360 like of this brain, body, environment system 05:53.360 --> 05:56.480 and expect the capabilities, like what emerges 05:56.480 --> 05:59.000 out of this system to just, you know, 05:59.000 --> 06:00.800 explode exponentially. 06:00.800 --> 06:04.720 Because anytime you improve one part of a system 06:04.720 --> 06:07.280 with many interdependencies like this, 06:07.280 --> 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 --> 06:14.960 their brain is not the bottleneck 06:14.960 --> 06:17.560 to the sort of problems they can solve, right? 06:17.560 --> 06:21.480 In fact, many very smart people today, you know, 06:21.480 --> 06:23.760 they're not actually solving any big scientific problems. 06:23.760 --> 06:24.800 They're not Einstein. 06:24.800 --> 06:26.560 They're like Einstein, but, you know, 06:26.560 --> 06:28.280 the patent clerk days. 06:28.280 --> 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, there's 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 --> 06:58.520 Well, that's brilliant. 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 of 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 --> 07:25.440 Yeah, exactly. Intelligence is the meaning of 07:25.440 --> 07:28.760 great problem solving capabilities 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 in intelligence. 07:32.280 --> 07:34.760 All you're left with is potential intelligence, 07:34.760 --> 07:36.920 like the performance of your brain or, you know, 07:36.920 --> 07:41.920 how high your IQ is, 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 --> 07:51.040 What do you think of as problem solving capacity? 07:51.040 --> 07:55.200 What, can you try to define intelligence? 07:56.680 --> 08:00.040 Like, what does it mean to be more or less intelligent? 08:00.040 --> 08:03.040 Is it completely coupled to a particular problem? 08:03.040 --> 08:05.760 Or is there something a little bit more universal? 08:05.760 --> 08:07.480 Yeah, I do believe all intelligence 08:07.480 --> 08:09.120 is specialized intelligence. 08:09.120 --> 08:12.280 Even human intelligence has some degree of generality. 08:12.280 --> 08:15.400 Well, all intelligence systems have some degree of generality, 08:15.400 --> 08:19.480 but they're always specialized in one category of problems. 08:19.480 --> 08:21.920 So the human intelligence is specialized 08:21.920 --> 08:25.560 in the human experience and that shows at various levels, 08:25.560 --> 08:29.320 that shows in some prior knowledge, 08:29.320 --> 08:32.040 that's innate, that we have at birth, 08:32.040 --> 08:35.360 knowledge about things like agents, 08:35.360 --> 08:40.440 goal driven behavior, visual priors about what makes an object, 08:40.440 --> 08:43.520 priors about time, and so on. 08:43.520 --> 08:45.360 That shows also in the way we learn, 08:45.360 --> 08:48.920 for instance, it's very easy for us to pick up language, 08:48.920 --> 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 problems 08:58.280 --> 09:01.440 and we are quite useless 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 --> 09:12.840 We have no capability of seeing the very long term. 09:12.840 --> 09:17.840 We don't have very much working memory, you know? 09:17.840 --> 09:19.960 So how do you think about long term? 09:19.960 --> 09:21.280 Do you think long term planning, 09:21.280 --> 09:24.760 we're talking about scale of years, millennia, 09:24.760 --> 09:27.960 what do you mean by long term, we're not very good? 09:27.960 --> 09:30.600 Well, human intelligence is specialized in the human experience 09:30.600 --> 09:34.120 and human experience is very short, like one lifetime is short. 09:34.120 --> 09:38.600 Even within one lifetime, we have a very hard time envisioning, 09:38.600 --> 09:41.080 you know, things on a scale of years. 09:41.080 --> 09:43.920 Like it's very difficult to project yourself at the scale of five, 09:43.920 --> 09:46.080 at the scale of 10 years and so on. 09:46.080 --> 09:49.960 Right. We can solve only fairly narrowly scoped problems. 09:49.960 --> 09:53.720 So when it comes to solving bigger problems, larger scale problems, 09:53.720 --> 09:56.320 we are not actually doing it on an individual level. 09:56.320 --> 09:59.240 So it's not actually our brain doing it. 09:59.240 --> 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 artificial intelligence system, right? 10:10.000 --> 10:14.080 And it's not running on one brain, it's running on a network of brains. 10:14.080 --> 10:16.760 In fact, it's running on much more than a network of brains. 10:16.760 --> 10:21.960 It's running on a lot of infrastructure, like books and computers 10:21.960 --> 10:25.760 and the internet and human institutions and so on. 10:25.760 --> 10:31.640 And that is capable of handling problems on a much greater scale 10:31.640 --> 10:33.720 than any individual human. 10:33.720 --> 10:37.560 If you look at computer science, for instance, 10:37.560 --> 10:42.480 that's an institution that solves problems and it is super human, right? 10:42.480 --> 10:46.840 It operates on a greater scale, it can solve much bigger problems 10:46.840 --> 10:49.040 than an individual human could. 10:49.040 --> 10:52.120 And science itself, science as a system, as an institution, 10:52.120 --> 10:57.640 is a kind of artificially intelligent problem solving algorithm 10:57.640 --> 10:59.360 that is super human. 10:59.360 --> 11:06.080 Yeah, it's a computer science is like a theorem prover 11:06.080 --> 11:10.360 at a scale of thousands, maybe hundreds of thousands of human beings. 11:10.360 --> 11:14.640 At that scale, what do you think is an intelligent agent? 11:14.640 --> 11:18.280 So there's us humans at the individual level. 11:18.280 --> 11:23.880 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 --> 11:31.880 You can even go to the particle level as systems that behave. 11:31.880 --> 11:35.400 You can say intelligently in some ways. 11:35.400 --> 11:37.840 And then you can look at the Earth as a single organism. 11:37.840 --> 11:42.160 You can look at our galaxy and even the universe as a single organism. 11:42.160 --> 11:46.320 Do you think, how do you think about scale and defining intelligent systems? 11:46.320 --> 11:51.840 And we're here at Google, there is millions of devices doing computation 11:51.840 --> 11:53.400 in a distributed way. 11:53.400 --> 11:55.880 How do you think about intelligence versus scale? 11:55.880 --> 12:00.640 You can always characterize anything as a system. 12:00.640 --> 12:05.320 I think people who talk about things like intelligence explosion 12:05.320 --> 12:08.760 tend to focus on one agent is basically one brain, 12:08.760 --> 12:11.920 like one brain considered in isolation, like a brain, a jar 12:11.920 --> 12:16.280 that's controlling a body in a very 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 --> 12:22.840 You have the brain at the top of the pyramid, 12:22.840 --> 12:25.960 then you have the body just plainly receiving orders, 12:25.960 --> 12:28.920 then the body is manipulating objects in an environment and so on. 12:28.920 --> 12:33.680 So everything is subordinate to this one thing, this epicenter, 12:33.680 --> 12:34.760 which is the brain. 12:34.760 --> 12:39.240 But in real life, intelligent agents don't really work like this. 12:39.240 --> 12:43.400 There is no strong delimitation between the brain and the body to start with. 12:43.400 --> 12:46.520 You have to look not just at the brain, but at the nervous system. 12:46.520 --> 12:50.760 But then the nervous system and the body 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 --> 13:00.200 But then you start realizing as you observe an animal over any length of time 13:00.200 --> 13:04.600 that a lot of the intelligence of an animal 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 --> 13:11.960 When you write down some notes, there is externalized intelligence. 13:11.960 --> 13:16.000 When you write a computer program, you are externalizing cognition. 13:16.000 --> 13:17.320 So it's externalized in books. 13:17.320 --> 13:23.040 It's externalized in computers, the internet, in other humans. 13:23.040 --> 13:25.400 It's externalized in language and so on. 13:25.400 --> 13:32.640 So there is no hard delimitation of what makes an intelligent agent. 13:32.640 --> 13:34.920 It's all about context. 13:34.920 --> 13:42.440 OK, but AlphaGo is better at Go than the best human player. 13:42.440 --> 13:44.960 There's levels of skill here. 13:44.960 --> 13:52.680 So do you think there is such a concept as an intelligence explosion 13:52.680 --> 13:54.720 in a specific task? 13:54.720 --> 14:00.080 And then, well, yeah, do you think it's possible to have a category of tasks 14:00.080 --> 14:05.000 on which you do have something like an exponential growth of ability 14:05.000 --> 14:07.400 to solve that particular problem? 14:07.400 --> 14:15.280 I think if you consider a specific vertical, it's probably possible to some extent. 14:15.280 --> 14:18.320 I also don't think we have to speculate about it 14:18.320 --> 14:24.760 because we have real world examples of free classivity self improving 14:24.760 --> 14:26.880 intelligent systems. 14:26.880 --> 14:32.560 For instance, science is a problem solving system, a knowledge generation system, 14:32.560 --> 14:36.240 like a system that experiences the world in some sense 14:36.240 --> 14:40.120 and then gradually understands it and can act on it. 14:40.120 --> 14:45.560 And that system is superhuman and it is clearly recursively self improving 14:45.560 --> 14:47.520 because science fits into technology. 14:47.520 --> 14:51.120 Technology can be used to build better tools, better computers, 14:51.120 --> 14:56.720 better instrumentation and so on, which in turn can make science faster. 14:56.720 --> 15:00.520 So science is probably the closest thing we have today 15:00.520 --> 15:04.720 to a real civility self improving superhuman AI. 15:04.720 --> 15:10.280 And you can just observe, is science, is scientific progress today exploding, 15:10.280 --> 15:12.760 which itself is an interesting question. 15:12.760 --> 15:15.800 You can use that as a basis to try to understand what 15:15.800 --> 15:20.960 will happen with a superhuman AI that has science like behavior. 15:20.960 --> 15:23.320 Let me linger on it a little bit more. 15:23.320 --> 15:28.520 What is your intuition why an intelligence explosion is not possible? 15:28.520 --> 15:34.400 Like taking the scientific, all the semi scientific revolutions. 15:34.400 --> 15:38.080 Why can't we slightly accelerate that process? 15:38.080 --> 15:43.160 So you can absolutely accelerate any problem solving process. 15:43.160 --> 15:48.640 So recursively, recursive self improvement is absolutely a real thing. 15:48.640 --> 15:51.880 But what happens with a recursively self improving system 15:51.880 --> 15:56.480 is typically not explosion because no system exists in isolation. 15:56.480 --> 16:00.840 And so tweaking one part of the system means that suddenly another part of the system 16:00.840 --> 16:02.120 becomes a bottleneck. 16:02.120 --> 16:06.760 And if you look at science, for instance, which is clearly a recursively self improving, 16:06.760 --> 16:11.960 clearly a problem solving system, scientific progress is not actually exploding. 16:11.960 --> 16:17.840 If you look at science, what you see is the picture of a system that is consuming 16:17.840 --> 16:20.440 an exponentially increasing amount of resources. 16:20.440 --> 16:26.000 But it's having a linear output in terms of scientific progress. 16:26.000 --> 16:28.960 And maybe that will seem like a very strong claim. 16:28.960 --> 16:34.520 Many people are actually saying that scientific progress is exponential. 16:34.520 --> 16:40.000 But when they're claiming this, they're actually looking at indicators of resource 16:40.000 --> 16:43.080 consumption by science. 16:43.080 --> 16:49.200 For instance, the number of papers being published, the number of patterns being 16:49.200 --> 16:55.760 filed, and so on, which are just completely correlated with how many people are working 16:55.760 --> 16:57.640 on science today. 16:57.640 --> 17:00.720 So it's actually an indicator of resource consumption. 17:00.720 --> 17:06.760 But what you should look at is the output is progress in terms of the knowledge that 17:06.760 --> 17:12.840 science generates in terms of the scope and significance of the problems that we solve. 17:12.840 --> 17:16.920 And some people have actually been trying to measure that. 17:16.920 --> 17:22.800 Like Michael Nielsen, for instance, he had a very nice paper, I think that was last 17:22.800 --> 17:25.280 year about it. 17:25.280 --> 17:32.760 So his approach to measure scientific progress was to look at the timeline of scientific 17:32.760 --> 17:37.400 discoveries over the past 100, 150 years. 17:37.400 --> 17:46.120 And for each major discovery, ask a panel of experts to rate the significance of the 17:46.120 --> 17:47.120 discovery. 17:47.120 --> 17:54.440 And if the output of sciences in the institution were exponential, you would expect the temporal 17:54.440 --> 18:01.080 density of significance to go up exponentially, maybe because there's a faster rate of discoveries, 18:01.080 --> 18:05.120 maybe because the discoveries are increasingly more important. 18:05.120 --> 18:10.360 And what actually happens if you plot this temporal density of significance measured 18:10.360 --> 18:14.600 in this way, is that you see very much a flat graph. 18:14.600 --> 18:20.040 You see a flat graph across all disciplines, across physics, biology, medicine and so on. 18:20.040 --> 18:24.400 And it actually makes a lot of sense if you think about it, because think about the progress 18:24.400 --> 18:28.120 of physics 110 years ago. 18:28.120 --> 18:30.240 It was a time of crazy change. 18:30.240 --> 18:36.640 Think about the progress of technology 170 years ago, when we started replacing horses, 18:36.640 --> 18:40.080 with cars, when we started having electricity and so on. 18:40.080 --> 18:41.640 It was a time of incredible change. 18:41.640 --> 18:44.800 And today is also a time of very, very fast change. 18:44.800 --> 18:50.480 But it would be an unfair characterization to say that today, technology and science 18:50.480 --> 18:54.600 are moving way faster than they did 50 years ago or 100 years ago. 18:54.600 --> 19:08.800 And if you do try to rigorously plot the temporal density of the significance, you do see very 19:08.800 --> 19:16.240 flat curves and you can check out the paper that Michael Nielsen had about this idea. 19:16.240 --> 19:25.280 And so the way I interpret it is as you make progress in a given field or in a given subfield 19:25.280 --> 19:30.640 of science, it becomes exponentially more difficult to make further progress, like the 19:30.640 --> 19:35.120 very first person to work on information theory. 19:35.120 --> 19:40.320 If you enter a new field and it's still the very early years, there's a lot of low hanging 19:40.320 --> 19:42.200 fruit you can pick. 19:42.200 --> 19:48.240 But the next generation of researchers is going to have to dig much harder, actually, 19:48.240 --> 19:52.800 to make smaller discoveries, probably larger numbers, smaller discoveries. 19:52.800 --> 19:57.640 And to achieve the same amount of impact, you're going to need a much greater head count. 19:57.640 --> 20:02.840 And that's exactly the picture you're seeing with science, is that the number of scientists 20:02.840 --> 20:06.680 and engineers is, in fact, increasing exponentially. 20:06.680 --> 20:11.520 The amount of computational resources that are available to science is increasing exponentially 20:11.520 --> 20:12.520 and so on. 20:12.520 --> 20:18.240 So the resource consumption of science is exponential, but the output in terms of progress, 20:18.240 --> 20:21.160 in terms of significance, is linear. 20:21.160 --> 20:26.200 And the reason why is because, and even though science is rigorously self improving, meaning 20:26.200 --> 20:33.000 that scientific progress turns into technological progress, which in turn helps science. 20:33.000 --> 20:39.240 If you look at computers, for instance, our products of science and computers are tremendously 20:39.240 --> 20:41.600 useful in spinning up science. 20:41.600 --> 20:42.600 The internet, same thing. 20:42.600 --> 20:47.680 The internet is a technology that's made possible by very recent scientific advances. 20:47.680 --> 20:53.960 And itself, because it enables scientists to network, to communicate, to exchange papers 20:53.960 --> 20:57.480 and ideas much faster, it is a way to speed up scientific progress. 20:57.480 --> 21:02.800 So even though you're looking at a recursively self improving system, it is consuming exponentially 21:02.800 --> 21:09.240 more resources to produce the same amount of problem solving, in fact. 21:09.240 --> 21:11.200 So that's a fascinating way to paint it. 21:11.200 --> 21:14.960 And certainly that holds for the deep learning community, right? 21:14.960 --> 21:18.040 If you look at the temporal, what did you call it? 21:18.040 --> 21:21.260 The temporal density of significant ideas. 21:21.260 --> 21:27.440 If you look at in deep learning, I think, I'd have to think about that, but if you really 21:27.440 --> 21:32.480 look at significant ideas in deep learning, they might even be decreasing. 21:32.480 --> 21:39.720 So I do believe the per paper significance is decreasing. 21:39.720 --> 21:43.480 But the amount of papers is still today, exponentially increasing. 21:43.480 --> 21:49.480 So I think if you look at an aggregate, my guess is that you would see a linear progress. 21:49.480 --> 21:58.720 If you were to sum the significance of all papers, you would see a roughly linear progress. 21:58.720 --> 22:05.680 And in my opinion, it is not a coincidence that you're seeing linear progress in science 22:05.680 --> 22:07.640 despite exponential resource consumption. 22:07.640 --> 22:15.840 I think the resource consumption is dynamically adjusting itself to maintain linear progress 22:15.840 --> 22:21.360 because we as a community expect linear progress, meaning that if we start investing less and 22:21.360 --> 22:26.160 seeing less progress, it means that suddenly there are some lower hanging fruits that become 22:26.160 --> 22:31.320 available and someone's going to step up and pick them. 22:31.320 --> 22:37.200 So it's very much like a market for discoveries and ideas. 22:37.200 --> 22:41.640 But there's another fundamental part which you're highlighting, which as a hypothesis 22:41.640 --> 22:49.440 as science or the space of ideas, any one path you travel down, it gets exponentially 22:49.440 --> 22:54.800 more difficult to develop new ideas. 22:54.800 --> 23:01.080 And your sense is that's going to hold across our mysterious universe. 23:01.080 --> 23:02.080 Yes. 23:02.080 --> 23:06.800 Well, exponential progress triggers exponential friction so that if you tweak one part of 23:06.800 --> 23:10.200 the system, suddenly some other part becomes a bottleneck. 23:10.200 --> 23:17.440 For instance, let's say we develop some device that measures its own acceleration and then 23:17.440 --> 23:22.240 it has some engine and it outputs even more acceleration in proportion of its own acceleration 23:22.240 --> 23:23.240 and you drop it somewhere. 23:23.240 --> 23:29.120 It's not going to reach infinite speed because it exists in a certain context. 23:29.120 --> 23:32.960 So the error on this is going to generate friction and it's going to block it at some 23:32.960 --> 23:34.440 top speed. 23:34.440 --> 23:39.880 And even if you were to consider a broader context and lift the bottleneck there, like 23:39.880 --> 23:46.200 the bottleneck of friction, then some other part of the system would start stepping in 23:46.200 --> 23:50.040 and creating exponential friction, maybe the speed of flight or whatever. 23:50.040 --> 23:55.400 And this definitely holds true when you look at the problem solving algorithm that is being 23:55.400 --> 23:59.780 run by science as an institution, science as a system. 23:59.780 --> 24:06.880 As you make more and more progress, despite having this recursive self improvement component, 24:06.880 --> 24:11.840 you are encountering exponential friction, like the more researchers you have working 24:11.840 --> 24:18.200 on different ideas, the more overhead you have in terms of communication across researchers. 24:18.200 --> 24:23.160 If you look at, you were mentioning quantum mechanics, right? 24:23.160 --> 24:28.480 Well if you want to start making significant discoveries today, significant progress in 24:28.480 --> 24:34.200 quantum mechanics, there is an amount of knowledge you have to ingest, which is huge. 24:34.200 --> 24:40.000 But there is a very large overhead to even start to contribute, there is a large amount 24:40.000 --> 24:44.240 of overhead to synchronize across researchers and so on. 24:44.240 --> 24:50.720 And of course, the significant practical experiments are going to require exponentially 24:50.720 --> 24:57.920 expensive equipment because the easier ones have already been run, right? 24:57.920 --> 25:08.520 So in your senses, there is no way of escaping this kind of friction with artificial intelligence 25:08.520 --> 25:09.520 systems. 25:09.520 --> 25:15.360 Yeah, no, I think science is a very good way to model what would happen with a superhuman 25:15.360 --> 25:17.880 recursive research improving AI. 25:17.880 --> 25:20.960 That's my intuition. 25:20.960 --> 25:26.680 It's not like a mathematical proof of anything, that's not my point, like I'm not trying 25:26.680 --> 25:31.440 to prove anything, I'm just trying to make an argument to question the narrative of intelligence 25:31.440 --> 25:35.600 explosion, which is quite a dominant narrative and you do get a lot of pushback if you go 25:35.600 --> 25:36.920 against it. 25:36.920 --> 25:43.280 Because so for many people, right, AI is not just a subfield of computer science, it's 25:43.280 --> 25:49.560 more like a belief system, like this belief that the world is headed towards an event, 25:49.560 --> 25:58.000 the singularity, past which, you know, AI will become, will go exponential very much 25:58.000 --> 26:02.160 and the world will be transformed and humans will become obsolete. 26:02.160 --> 26:07.880 And if you go against this narrative, because it is not really a scientific argument but 26:07.880 --> 26:12.240 more of a belief system, it is part of the identity of many people. 26:12.240 --> 26:15.680 If you go against this narrative, it's like you're attacking the identity of people who 26:15.680 --> 26:16.680 believe in it. 26:16.680 --> 26:22.880 It's almost like saying God doesn't exist or something, so you do get a lot of pushback 26:22.880 --> 26:25.200 if you try to question his ideas. 26:25.200 --> 26:29.880 First of all, I believe most people, they might not be as eloquent or explicit as you're 26:29.880 --> 26:34.400 being, but most people in computer science are most people who actually have built anything 26:34.400 --> 26:39.160 that you could call AI, quote unquote, would agree with you. 26:39.160 --> 26:43.880 They might not be describing in the same kind of way, it's more, so the pushback you're 26:43.880 --> 26:51.120 getting is from people who get attached to the narrative from, not from a place of science, 26:51.120 --> 26:53.520 but from a place of imagination. 26:53.520 --> 26:54.520 That's correct. 26:54.520 --> 26:55.520 That's correct. 26:55.520 --> 26:57.240 So why do you think that's so appealing? 26:57.240 --> 27:03.880 Because the usual dreams that people have when you create a superintelligence system 27:03.880 --> 27:09.520 past the singularity, that what people imagine is somehow always destructive. 27:09.520 --> 27:13.760 Do you have, if you were put on your psychology hat, what's, why is it so? 27:13.760 --> 27:20.200 Why is it so appealing to imagine the ways that all of human civilization will be destroyed? 27:20.200 --> 27:22.200 I think it's a good story. 27:22.200 --> 27:23.200 You know, it's a good story. 27:23.200 --> 27:30.680 And very interestingly, it mirrors religious stories, right, religious mythology. 27:30.680 --> 27:36.960 If you look at the mythology of most civilizations, it's about the world being headed towards 27:36.960 --> 27:42.240 some final events in which the world will be destroyed and some new world order will 27:42.240 --> 27:49.640 arise that will be mostly spiritual, like the apocalypse followed by a paradise, probably. 27:49.640 --> 27:52.880 It's a very appealing story on a fundamental level. 27:52.880 --> 27:54.640 And we all need stories. 27:54.640 --> 27:59.920 We all need stories to structure in the way we see the world, especially at timescales 27:59.920 --> 28:04.600 that are beyond our ability to make predictions. 28:04.600 --> 28:14.920 So on a more serious non exponential explosion question, do you think there will be a time 28:14.920 --> 28:21.880 when we'll create something like human level intelligence or intelligence systems that 28:21.880 --> 28:28.720 will make you sit back and be just surprised at damn how smart this thing is? 28:28.720 --> 28:32.360 That doesn't require exponential growth or an exponential improvement. 28:32.360 --> 28:39.840 But what's your sense of the timeline and so on, that you'll be really surprised at 28:39.840 --> 28:40.840 certain capabilities? 28:40.840 --> 28:44.360 And we'll talk about limitations and deep learning, so do you think in your lifetime 28:44.360 --> 28:46.760 you'll be really damn surprised? 28:46.760 --> 28:53.960 Around 2013, 2014, I was many times surprised by the capabilities of deep learning, actually. 28:53.960 --> 28:57.880 That was before we had assessed exactly what deep learning could do and could not do and 28:57.880 --> 29:00.680 it felt like a time of immense potential. 29:00.680 --> 29:03.120 And then we started narrowing it down. 29:03.120 --> 29:07.240 But I was very surprised, so I would say it has already happened. 29:07.240 --> 29:13.640 Was there a moment, there must have been a day in there where your surprise was almost 29:13.640 --> 29:19.640 bordering on the belief of the narrative that we just discussed? 29:19.640 --> 29:23.200 Was there a moment, because you've written quite eloquently about the limits of deep 29:23.200 --> 29:28.600 learning, was there a moment that you thought that maybe deep learning is limitless? 29:28.600 --> 29:32.520 No, I don't think I've ever believed this. 29:32.520 --> 29:35.120 What was really shocking is that it worked. 29:35.120 --> 29:37.800 It worked at all, yeah. 29:37.800 --> 29:43.880 But there's a big jump between being able to do really good computer vision and human 29:43.880 --> 29:45.040 level intelligence. 29:45.040 --> 29:50.840 So I don't think at any point, I wasn't an impression that the results we got in computer 29:50.840 --> 29:54.040 vision meant that we were very close to human level intelligence. 29:54.040 --> 29:56.000 I don't think we're very close to human level intelligence. 29:56.000 --> 30:01.720 I do believe that there's no reason why we won't achieve it at some point. 30:01.720 --> 30:10.280 I also believe that the problem with talking about human level intelligence is that implicitly 30:10.280 --> 30:13.920 you're considering an axis of intelligence with different levels. 30:13.920 --> 30:17.200 But that's not really how intelligence works. 30:17.200 --> 30:19.600 Intelligence is very multidimensional. 30:19.600 --> 30:24.440 And so there's the question of capabilities, but there's also the question of being human 30:24.440 --> 30:29.640 like, and it's two very different things, like you can build potentially very advanced 30:29.640 --> 30:32.760 intelligent agents that are not human like at all. 30:32.760 --> 30:35.320 And you can also build very human like agents. 30:35.320 --> 30:37.920 And these are two very different things, right? 30:37.920 --> 30:38.920 Right. 30:38.920 --> 30:42.360 Let's go from the philosophical to the practical. 30:42.360 --> 30:46.560 Can you give me a history of Keras and all the major deep learning frameworks that you 30:46.560 --> 30:51.600 kind of remember in relation to Keras and in general, TensorFlow, Theano, the old days. 30:51.600 --> 30:57.440 Can you give a brief overview, Wikipedia style history, and your role in it before we return 30:57.440 --> 30:58.840 to AGI discussions? 30:58.840 --> 31:00.840 Yeah, that's a broad topic. 31:00.840 --> 31:06.800 So I started working on Keras, it was a name Keras at the time, I actually picked the 31:06.800 --> 31:09.920 name like just the day I was going to release it. 31:09.920 --> 31:15.040 So I started working on it in February 2015. 31:15.040 --> 31:18.440 And so at the time, there weren't too many people working on deep learning, maybe like 31:18.440 --> 31:25.480 fewer than 10,000, the software tooling was not really developed. 31:25.480 --> 31:30.960 So the main deep learning library was Cafe, which was mostly C++. 31:30.960 --> 31:33.040 Why do you say Cafe was the main one? 31:33.040 --> 31:39.120 Cafe was vastly more popular than Theano in late 2014, early 2015. 31:39.120 --> 31:43.480 Cafe was the one library that everyone was using for computer vision. 31:43.480 --> 31:46.240 And computer vision was the most popular problem. 31:46.240 --> 31:47.240 Absolutely. 31:47.240 --> 31:53.280 Like, Covenant was like the subfield of deep learning that everyone was working on. 31:53.280 --> 32:01.840 So myself, so in late 2014, I was actually interested in RNNs, in recurrent neural networks, 32:01.840 --> 32:08.800 which was a very niche topic at the time, right, it really took off around 2016. 32:08.800 --> 32:11.520 And so I was looking for good tools. 32:11.520 --> 32:19.480 I had used Torch 7, I had used Theano, used Theano a lot in Kaggle competitions, I had 32:19.480 --> 32:21.240 used Cafe. 32:21.240 --> 32:27.880 And there was no like good solution for RNNs at the time, like there was no reusable open 32:27.880 --> 32:30.280 source implementation of an LSTM, for instance. 32:30.280 --> 32:33.200 So I decided to build my own. 32:33.200 --> 32:39.600 And at first, the pitch for that was it was going to be mostly around LSTM recurrent neural 32:39.600 --> 32:40.600 networks. 32:40.600 --> 32:46.000 So in Python, an important decision at the time that was kind of nonobvious is that the 32:46.000 --> 32:54.520 models would be defined via Python code, which was kind of like going against the mainstream 32:54.520 --> 33:00.320 at the time, because Cafe, Pylon 2 and so on, like all the big libraries were actually 33:00.320 --> 33:05.840 going with you, approaching static configuration files in YAML to define models. 33:05.840 --> 33:10.560 So some libraries were using code to define models like Torch 7, obviously, but that was 33:10.560 --> 33:11.560 not. 33:11.560 --> 33:17.840 Python Lasagne was like a Theano based very early library that was, I think, developed. 33:17.840 --> 33:18.840 I don't remember exactly. 33:18.840 --> 33:19.840 Probably late 2014. 33:19.840 --> 33:20.840 It's Python as well. 33:20.840 --> 33:21.840 It's Python as well. 33:21.840 --> 33:25.040 It was like on top of Theano. 33:25.040 --> 33:32.760 And so I started working on something and the value proposition at the time was that not 33:32.760 --> 33:40.920 only that what I think was the first reusable open source implementation of LSTM, you could 33:40.920 --> 33:47.080 combine RNNs and covenants with the same library, which is not really possible before. 33:47.080 --> 33:50.760 Like Cafe was only doing covenants. 33:50.760 --> 33:52.880 And it was kind of easy to use. 33:52.880 --> 33:55.760 Because so before I was using Theano, I was actually using Psykitlin. 33:55.760 --> 33:58.480 And I loved Psykitlin for its usability. 33:58.480 --> 34:02.440 So I drew a lot of inspiration from Psykitlin when I met Keras. 34:02.440 --> 34:05.680 It's almost like Psykitlin for neural networks. 34:05.680 --> 34:06.680 The fit function. 34:06.680 --> 34:07.680 Exactly. 34:07.680 --> 34:08.680 The fit function. 34:08.680 --> 34:13.000 Like reducing a complex string loop to a single function call. 34:13.000 --> 34:17.480 And of course, some people will say, this is hiding a lot of details, but that's exactly 34:17.480 --> 34:18.480 the point. 34:18.480 --> 34:20.360 The magic is the point. 34:20.360 --> 34:25.280 So it's magical, but in a good way, it's magical in the sense that it's delightful. 34:25.280 --> 34:27.600 I'm actually quite surprised. 34:27.600 --> 34:31.920 I didn't know that it was born out of desire to implement RNNs and LSTMs. 34:31.920 --> 34:32.920 It was. 34:32.920 --> 34:33.920 That's fascinating. 34:33.920 --> 34:39.160 So you were actually one of the first people to really try to attempt to get the major 34:39.160 --> 34:41.160 architecture together. 34:41.160 --> 34:45.160 And it's also interesting, I mean, you realize that that was a design decision at all is 34:45.160 --> 34:47.480 defining the model and code. 34:47.480 --> 34:52.320 Just I'm putting myself in your shoes, whether the YAML, especially if Cafe was the most 34:52.320 --> 34:53.320 popular. 34:53.320 --> 34:54.760 It was the most popular by far. 34:54.760 --> 35:01.880 If I was if I were, yeah, I don't, I didn't like the YAML thing, but it makes more sense 35:01.880 --> 35:05.760 that you will put in a configuration file, the definition of a model. 35:05.760 --> 35:10.160 That's an interesting gutsy move to stick with defining it in code. 35:10.160 --> 35:14.800 Just if you look back, other libraries, we're doing it as well, but it was definitely the 35:14.800 --> 35:16.200 more niche option. 35:16.200 --> 35:17.200 Yeah. 35:17.200 --> 35:18.200 Okay. 35:18.200 --> 35:19.200 Keras and then Keras. 35:19.200 --> 35:24.220 So I released Keras in March, 2015, and it got users pretty much from the start. 35:24.220 --> 35:27.480 So the deep learning community was very, very small at the time. 35:27.480 --> 35:30.640 Lots of people were starting to be interested in LSTMs. 35:30.640 --> 35:34.760 So it was going to release at the right time because it was offering an easy to use LSTM 35:34.760 --> 35:35.760 implementation. 35:35.760 --> 35:40.840 Exactly at the time where lots of you started to be intrigued by the capabilities of RNN, 35:40.840 --> 35:42.340 RNNs 1LP. 35:42.340 --> 35:47.000 So it grew from there. 35:47.000 --> 35:53.760 Then I joined Google about six months later, and that was actually completely unrelated 35:53.760 --> 35:54.760 to Keras. 35:54.760 --> 36:00.720 Keras actually joined a research team working on image classification mostly like computer 36:00.720 --> 36:01.720 vision. 36:01.720 --> 36:03.840 So I was doing computer vision research at Google initially. 36:03.840 --> 36:11.440 And immediately when I joined Google, I was exposed to the early internal version of TensorFlow. 36:11.440 --> 36:15.400 And the way it appeared to me at the time, and it was definitely the way it was at the 36:15.400 --> 36:20.880 time, is that this was an improved version of Tiano. 36:20.880 --> 36:27.040 So I immediately knew I had to port Keras to this new TensorFlow thing. 36:27.040 --> 36:31.760 And I was actually very busy as a new Googler. 36:31.760 --> 36:34.600 So I had not time to work on that. 36:34.600 --> 36:41.360 But then in November, I think it was November 2015, TensorFlow got released. 36:41.360 --> 36:47.440 And it was kind of like my wake up call that, hey, I had to actually go and make it happen. 36:47.440 --> 36:53.360 So in December, I ported Keras to run on TensorFlow, but it was not exactly a port. 36:53.360 --> 36:59.360 It was more like a refactoring where I was abstracting away all the backend functionality 36:59.360 --> 37:05.200 into one module so that the same code base could run on top of multiple backends. 37:05.200 --> 37:07.560 So on top of TensorFlow or Tiano. 37:07.560 --> 37:21.000 And for the next year, Tiano stayed as the default option, it was easier to use, it was 37:21.000 --> 37:23.440 much faster, especially when it came to on it. 37:23.440 --> 37:27.560 But eventually, TensorFlow overtook it. 37:27.560 --> 37:34.000 And TensorFlow, the early TensorFlow has similar architectural decisions as Tiano. 37:34.000 --> 37:38.360 So it was a natural transition. 37:38.360 --> 37:45.360 So what, I mean, that still carries as a side, almost one project, right? 37:45.360 --> 37:50.280 Yeah, so it was not my job assignment, it was not. 37:50.280 --> 37:52.360 I was doing it on the side. 37:52.360 --> 37:57.840 And even though it grew to have a lot of uses for deep learning library at the time, like 37:57.840 --> 38:02.560 Stroud 2016, but I wasn't doing it as my main job. 38:02.560 --> 38:10.680 So things started changing in, I think it must have been maybe October 2016, so one year 38:10.680 --> 38:11.680 later. 38:11.680 --> 38:18.440 So Rajat, who has the lead in TensorFlow, basically showed up one day in our building 38:18.440 --> 38:23.040 where I was doing like, so I was doing research and things like, so I did a lot of computer 38:23.040 --> 38:29.040 vision research, also collaborations with Christian Zegedi and Deep Learning for Theraim 38:29.040 --> 38:34.720 Proving, that was a really interesting research topic. 38:34.720 --> 38:42.600 And so Rajat was saying, hey, we saw Keras, we like it, we saw that you had Google, why 38:42.600 --> 38:46.960 don't you come over for like a quarter and work with us? 38:46.960 --> 38:50.560 And I was like, yeah, that sounds like a great opportunity, let's do it. 38:50.560 --> 38:57.520 And so I started working on integrating the Keras API into TensorFlow more tightly. 38:57.520 --> 39:06.000 So what followed up is a sort of temporary TensorFlow only version of Keras that was 39:06.000 --> 39:12.560 in TensorFlow.contrib for a while, and finally moved to TensorFlow Core. 39:12.560 --> 39:17.320 And I've never actually gotten back to my old team doing research. 39:17.320 --> 39:27.360 Well, it's kind of funny that somebody like you who dreams of or at least sees the power 39:27.360 --> 39:33.800 of AI systems that reason and Theraim Proving will talk about has also created a system 39:33.800 --> 39:41.600 that makes the most basic kind of Lego building that is deep learning, super accessible, super 39:41.600 --> 39:43.840 easy, so beautifully so. 39:43.840 --> 39:50.280 It's a funny irony that you're both, you're responsible for both things. 39:50.280 --> 39:55.360 So TensorFlow 2.0 is kind of, there's a sprint, I don't know how long it'll take, but there's 39:55.360 --> 39:57.080 a sprint towards the finish. 39:57.080 --> 40:01.120 What do you look, what are you working on these days? 40:01.120 --> 40:02.120 What are you excited about? 40:02.120 --> 40:05.040 What are you excited about in 2.0? 40:05.040 --> 40:09.880 Eager execution, there's so many things that just make it a lot easier to work. 40:09.880 --> 40:11.640 What are you excited about? 40:11.640 --> 40:13.800 And what's also really hard? 40:13.800 --> 40:15.880 What are the problems you have to kind of solve? 40:15.880 --> 40:22.880 So I've spent the past year and a half working on TensorFlow 2.0 and it's been a long journey. 40:22.880 --> 40:25.040 I'm actually extremely excited about it. 40:25.040 --> 40:26.560 I think it's a great product. 40:26.560 --> 40:29.440 It's a delightful product compared to TensorFlow 1.0. 40:29.440 --> 40:32.800 We've made huge progress. 40:32.800 --> 40:40.640 So on the Keras side, what I'm really excited about is that, so previously Keras has been 40:40.640 --> 40:50.880 this very easy to use high level interface to do deep learning, but if you wanted to, 40:50.880 --> 40:57.760 if you wanted a lot of flexibility, the Keras framework was probably not the optimal way 40:57.760 --> 41:02.160 to do things compared to just writing everything from scratch. 41:02.160 --> 41:05.040 So in some way, the framework was getting in the way. 41:05.040 --> 41:08.280 And in TensorFlow 2.0, you don't have this at all, actually. 41:08.280 --> 41:13.600 You have the usability of the high level interface, but you have the flexibility of this lower 41:13.600 --> 41:20.520 level interface, and you have this spectrum of workflows where you can get more or less 41:20.520 --> 41:26.960 usability and flexibility, the tradeoffs, depending on your needs. 41:26.960 --> 41:33.800 You can write everything from scratch and you get a lot of help doing so by subclassing 41:33.800 --> 41:38.520 models and writing some train loops using eager execution. 41:38.520 --> 41:39.520 It's very flexible. 41:39.520 --> 41:40.520 It's very easy to debug. 41:40.520 --> 41:42.400 It's very powerful. 41:42.400 --> 41:48.600 But all of this integrates seamlessly with higher level features up to the classic Keras 41:48.600 --> 41:56.440 workflows, which are very psychedelic and ideal for a data scientist, machine learning 41:56.440 --> 41:58.320 engineer type of profile. 41:58.320 --> 42:04.320 So now you can have the same framework offering the same set of APIs that enable a spectrum 42:04.320 --> 42:11.000 of workflows that are lower level, more or less high level, that are suitable for profiles 42:11.000 --> 42:15.400 ranging from researchers to data scientists and everything in between. 42:15.400 --> 42:16.400 Yeah. 42:16.400 --> 42:17.400 So that's super exciting. 42:17.400 --> 42:18.600 I mean, it's not just that. 42:18.600 --> 42:21.560 It's connected to all kinds of tooling. 42:21.560 --> 42:26.760 You can go on mobile, you can go with TensorFlow Lite, you can go in the cloud or serving 42:26.760 --> 42:29.240 and so on, it all is connected together. 42:29.240 --> 42:37.440 Some of the best software written ever is often done by one person, sometimes two. 42:37.440 --> 42:42.920 So with a Google, you're now seeing sort of Keras having to be integrated in TensorFlow. 42:42.920 --> 42:46.520 I'm sure it has a ton of engineers working on. 42:46.520 --> 42:52.320 So I'm sure there are a lot of tricky design decisions to be made. 42:52.320 --> 42:54.600 How does that process usually happen? 42:54.600 --> 43:00.800 At least your perspective, what are the debates like? 43:00.800 --> 43:07.160 Is there a lot of thinking considering different options and so on? 43:07.160 --> 43:08.160 Yes. 43:08.160 --> 43:17.920 So a lot of the time I spend at Google is actually discussing design discussions, writing design 43:17.920 --> 43:22.200 docs, participating in design review meetings and so on. 43:22.200 --> 43:25.520 This is as important as actually writing a code. 43:25.520 --> 43:34.080 So there's a lot of thought and a lot of care that is taken in coming up with these decisions 43:34.080 --> 43:39.920 and taking into account all of our users because TensorFlow has this extremely diverse user 43:39.920 --> 43:40.920 base. 43:40.920 --> 43:45.560 It's not like just one user segment where everyone has the same needs. 43:45.560 --> 43:49.640 We have small scale production users, large scale production users. 43:49.640 --> 43:56.520 We have startups, we have researchers, it's all over the place, and we have to cater to 43:56.520 --> 43:57.520 all of their needs. 43:57.520 --> 44:04.160 If I just look at the standard debates of C++ or Python, there's some heated debates. 44:04.160 --> 44:05.680 Do you have those at Google? 44:05.680 --> 44:10.560 I mean, they're not heated in terms of emotionally, but there's probably multiple ways to do it, 44:10.560 --> 44:11.560 right? 44:11.560 --> 44:16.080 So how do you arrive through those design meetings at the best way to do it, especially in deep 44:16.080 --> 44:21.960 learning where the field is evolving as you're doing it? 44:21.960 --> 44:23.440 Is there some magic to it? 44:23.440 --> 44:25.240 Is there some magic to the process? 44:25.240 --> 44:30.800 I don't know if there's magic to the process, but there definitely is a process. 44:30.800 --> 44:37.240 So making design decisions is about satisfying a set of constraints, but also trying to do 44:37.240 --> 44:42.720 so in the simplest way possible because this is what can be maintained, this is what can 44:42.720 --> 44:45.080 be expanded in the future. 44:45.080 --> 44:51.200 So you don't want to naively satisfy the constraints by just, you know, for each capability you 44:51.200 --> 44:54.760 need available, you're going to come up with one argument in your API and so on. 44:54.760 --> 45:03.920 You want to design APIs that are modular and hierarchical so that they have an API surface 45:03.920 --> 45:07.520 that is as small as possible, right? 45:07.520 --> 45:14.800 And you want this modular hierarchical architecture to reflect the way that domain experts think 45:14.800 --> 45:19.960 about the problem because as a domain expert, when you're reading about a new API, you're 45:19.960 --> 45:27.120 reading a tutorial or some docs, pages, you already have a way that you're thinking about 45:27.120 --> 45:28.120 the problem. 45:28.120 --> 45:35.600 You already have certain concepts in mind and you're thinking about how they relate together 45:35.600 --> 45:41.280 and when you're reading docs, you're trying to build as quickly as possible a mapping 45:41.280 --> 45:47.240 between the concepts featured in your API and the concepts in your mind so you're trying 45:47.240 --> 45:53.720 to map your mental model as a domain expert to the way things work in the API. 45:53.720 --> 45:59.320 So you need an API and an underlying implementation that are reflecting the way people think about 45:59.320 --> 46:00.320 these things. 46:00.320 --> 46:02.960 So in minimizing the time it takes to do the mapping? 46:02.960 --> 46:03.960 Yes. 46:03.960 --> 46:10.000 Minimizing the time, the cognitive load there is in ingesting this new knowledge about your 46:10.000 --> 46:11.000 API. 46:11.000 --> 46:16.080 An API should not be self referential or referring to implementation details, it should only 46:16.080 --> 46:22.360 be referring to domain specific concepts that people already understand. 46:22.360 --> 46:24.560 Brilliant. 46:24.560 --> 46:27.640 So what's the future of Keras and TensorFlow look like? 46:27.640 --> 46:30.680 What does TensorFlow 3.0 look like? 46:30.680 --> 46:36.440 So that's kind of too far in the future for me to answer, especially since I'm not even 46:36.440 --> 46:39.480 the one making these decisions. 46:39.480 --> 46:44.840 But so from my perspective, which is just one perspective among many different perspectives 46:44.840 --> 46:52.600 on the TensorFlow team, I'm really excited by developing even higher level APIs, higher 46:52.600 --> 46:53.600 level than Keras. 46:53.600 --> 47:01.040 I'm really excited by hyperparameter tuning, by automated machine learning, AutoML. 47:01.040 --> 47:07.480 I think the future is not just defining a model like you were assembling Lego blocks 47:07.480 --> 47:14.280 and then colleague fit on it, it's more like an automagical model that would just look 47:14.280 --> 47:19.120 at your data and optimize the objective you're after. 47:19.120 --> 47:22.440 So that's what I'm looking into. 47:22.440 --> 47:23.440 Yes. 47:23.440 --> 47:30.120 So you put the baby into a room with the problem and come back a few hours later with a fully 47:30.120 --> 47:31.120 solved problem. 47:31.120 --> 47:32.120 Exactly. 47:32.120 --> 47:36.520 It's not like a box of Legos, it's more like the combination of a kid that's really good 47:36.520 --> 47:41.560 at Legos, and a box of Legos, and just building the thing on the song. 47:41.560 --> 47:42.760 Very nice. 47:42.760 --> 47:44.080 So that's an exciting feature. 47:44.080 --> 47:50.680 I think there's a huge amount of applications and revolutions to be had under the constraints 47:50.680 --> 47:52.800 of the discussion we previously had. 47:52.800 --> 47:57.520 But what do you think are the current limits of deep learning? 47:57.520 --> 48:05.200 If we look specifically at these function approximators that tries to generalize from 48:05.200 --> 48:06.200 data? 48:06.200 --> 48:11.800 If you've talked about local versus extreme generalization, you mentioned that neural 48:11.800 --> 48:17.840 networks don't generalize well and humans do, so there's this gap. 48:17.840 --> 48:22.840 And you've also mentioned that extreme generalization requires something like reasoning to fill those 48:22.840 --> 48:24.040 gaps. 48:24.040 --> 48:27.120 So how can we start trying to build systems like that? 48:27.120 --> 48:28.120 Right. 48:28.120 --> 48:29.120 Yes. 48:29.120 --> 48:30.640 So this is by design, right? 48:30.640 --> 48:39.600 And deep learning models are huge, parametric models, differentiable, so continuous, that 48:39.600 --> 48:42.840 go from an input space to an output space. 48:42.840 --> 48:46.560 And they're trained with gradient descent, so they're trained pretty much point by point. 48:46.560 --> 48:53.560 They're learning a continuous geometric morphing from an input vector space to an output vector 48:53.560 --> 48:55.640 space, right? 48:55.640 --> 49:02.920 And because this is done point by point, a deep neural network can only make sense of 49:02.920 --> 49:08.160 points in experience space that are very close to things that it has already seen in string 49:08.160 --> 49:09.160 data. 49:09.160 --> 49:14.040 At best, it can do interpolation across points. 49:14.040 --> 49:20.560 But that means in order to train your network, you need a dense sampling of the input cross 49:20.560 --> 49:27.040 output space, almost a point by point sampling, which can be very expensive if you're dealing 49:27.040 --> 49:33.760 with complex real world problems like autonomous driving, for instance, or robotics. 49:33.760 --> 49:37.240 It's doable if you're looking at the subset of the visual space. 49:37.240 --> 49:41.200 But even then, it's still fairly expensive, you still need millions of examples. 49:41.200 --> 49:45.600 And it's only going to be able to make sense of things that are very close to ways that's 49:45.600 --> 49:47.000 seen before. 49:47.000 --> 49:50.720 And in contrast to that, well, of course, you have human intelligence, but even if you're 49:50.720 --> 49:56.840 not looking at human intelligence, you can look at very simple rules, algorithms. 49:56.840 --> 50:03.080 If you have a symbolic rule, it can actually apply to a very, very large set of inputs 50:03.080 --> 50:04.920 because it is abstract. 50:04.920 --> 50:10.760 It is not obtained by doing a point by point mapping, right? 50:10.760 --> 50:15.640 For instance, if you try to learn a sorting algorithm using a deep neural network, well, 50:15.640 --> 50:21.800 you're very much limited to learning point by point what the sorted representation of 50:21.800 --> 50:24.520 this specific list is like. 50:24.520 --> 50:32.120 But instead, you could have a very, very simple sorting algorithm written in a few lines. 50:32.120 --> 50:35.720 Maybe it's just two nested loops. 50:35.720 --> 50:42.320 And it can process any list at all because it is abstract, because it is a set of rules. 50:42.320 --> 50:47.440 So deep learning is really like point by point geometric morphings, morphings trained with 50:47.440 --> 50:48.880 God and Descent. 50:48.880 --> 50:54.200 And meanwhile, abstract rules can generalize much better. 50:54.200 --> 50:56.400 And I think the future is really to combine the two. 50:56.400 --> 50:59.720 So how do we, do you think, combine the two? 50:59.720 --> 51:08.040 How do we combine good point by point functions with programs, which is what the symbolic AI 51:08.040 --> 51:09.040 type systems? 51:09.040 --> 51:10.040 Yeah. 51:10.040 --> 51:11.600 At which levels the combination happened. 51:11.600 --> 51:17.480 I mean, obviously, we're jumping into the realm of where there's no good answers. 51:17.480 --> 51:20.120 It's just kind of ideas and intuitions and so on. 51:20.120 --> 51:21.120 Yeah. 51:21.120 --> 51:25.200 Well, if you look at the really successful AI systems today, I think there are already 51:25.200 --> 51:29.600 hybrid systems that are combining symbolic AI with deep learning. 51:29.600 --> 51:36.120 For instance, successful robotics systems are already mostly model based, rule based 51:36.120 --> 51:39.560 things like planning algorithms and so on. 51:39.560 --> 51:44.320 At the same time, they're using deep learning as perception modules. 51:44.320 --> 51:49.120 Sometimes they're using deep learning as a way to inject fuzzy intuition into a rule 51:49.120 --> 51:51.000 based process. 51:51.000 --> 51:56.720 If you look at a system like a self driving car, it's not just one big end to end neural 51:56.720 --> 52:00.920 network that wouldn't work at all, precisely because in order to train that, you would 52:00.920 --> 52:06.960 need a dense sampling of experience space when it comes to driving, which is completely 52:06.960 --> 52:08.480 unrealistic, obviously. 52:08.480 --> 52:18.560 Instead, the self driving car is mostly symbolic, it's software, it's programmed by hand. 52:18.560 --> 52:25.760 It's mostly based on explicit models, in this case, mostly 3D models of the environment 52:25.760 --> 52:31.600 around the car, but it's interfacing with the real world, using deep learning modules. 52:31.600 --> 52:36.480 The deep learning there serves as a way to convert the raw sensory information to something 52:36.480 --> 52:38.600 usable by symbolic systems. 52:38.600 --> 52:42.440 Okay, well, let's linger on that a little more. 52:42.440 --> 52:48.400 So dense sampling from input to output, you said it's obviously very difficult. 52:48.400 --> 52:49.400 Is it possible? 52:49.400 --> 52:51.960 In the case of self driving, you mean? 52:51.960 --> 52:53.240 Let's say self driving, right? 52:53.240 --> 52:57.760 Self driving for many people. 52:57.760 --> 53:03.320 Let's not even talk about self driving, let's talk about steering, so staying inside the 53:03.320 --> 53:05.320 lane. 53:05.320 --> 53:09.200 It's definitely a problem you can solve with an end to end deep learning model, but that's 53:09.200 --> 53:10.200 like one small subset. 53:10.200 --> 53:14.600 Hold on a second, I don't know how you're jumping from the extreme so easily, because 53:14.600 --> 53:17.800 I disagree with you on that. 53:17.800 --> 53:23.240 I think, well, it's not obvious to me that you can solve lane following. 53:23.240 --> 53:25.720 No, it's not obvious, I think it's doable. 53:25.720 --> 53:33.800 I think in general, there is no hard limitations to what you can learn with a deep neural network, 53:33.800 --> 53:42.160 as long as the search space is rich enough, is flexible enough, and as long as you have 53:42.160 --> 53:47.640 this dense sampling of the input cross output space, the problem is that this dense sampling 53:47.640 --> 53:52.920 could mean anything from 10,000 examples to trillions and trillions. 53:52.920 --> 53:54.440 So that's my question. 53:54.440 --> 53:56.360 So what's your intuition? 53:56.360 --> 54:01.800 And if you could just give it a chance and think what kind of problems can be solved 54:01.800 --> 54:08.080 by getting a huge amounts of data and thereby creating a dense mapping. 54:08.080 --> 54:14.040 So let's think about natural language dialogue, the Turing test. 54:14.040 --> 54:20.080 Do you think the Turing test can be solved with a neural network alone? 54:20.080 --> 54:26.480 Well, the Turing test is all about tricking people into believing they're talking to a 54:26.480 --> 54:27.480 human. 54:27.480 --> 54:35.720 It's actually very difficult because it's more about exploiting human perception and 54:35.720 --> 54:37.680 not so much about intelligence. 54:37.680 --> 54:41.520 There's a big difference between mimicking into Asian behavior and actually into Asian 54:41.520 --> 54:42.520 behavior. 54:42.520 --> 54:46.680 So, okay, let's look at maybe the Alexa prize and so on, the different formulations of the 54:46.680 --> 54:51.720 natural language conversation that are less about mimicking and more about maintaining 54:51.720 --> 54:54.920 a fun conversation that lasts for 20 minutes. 54:54.920 --> 54:59.240 It's a little less about mimicking and that's more about, I mean, it's still mimicking, 54:59.240 --> 55:03.200 but it's more about being able to carry forward a conversation with all the tangents that 55:03.200 --> 55:05.120 happen in dialogue and so on. 55:05.120 --> 55:12.480 Do you think that problem is learnable with this kind of neural network that does the 55:12.480 --> 55:14.600 point to point mapping? 55:14.600 --> 55:17.800 So I think it would be very, very challenging 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 --> 55:23.440 I wouldn't read it out. 55:23.440 --> 55:27.080 The space of problems that can be solved with a large neural network. 55:27.080 --> 55:31.280 What's your sense about the space of those problems? 55:31.280 --> 55:32.680 Useful problems for us. 55:32.680 --> 55:33.960 In theory, it's infinite. 55:33.960 --> 55:36.320 You can solve any problem. 55:36.320 --> 55:45.400 In practice, while deep learning is a great fit for perception problems, in general, any 55:45.400 --> 55:52.120 problem which is naturally amenable to explicit handcrafted rules or rules that you can generate 55:52.120 --> 55:56.160 by exhaustive search over some program space. 55:56.160 --> 56:03.400 So perception, artificial intuition, as long as you have a sufficient training data set. 56:03.400 --> 56:04.400 And that's the question. 56:04.400 --> 56:08.800 I mean, perception, there's interpretation and understanding of the scene, which seems 56:08.800 --> 56:13.040 to be outside the reach of current perception systems. 56:13.040 --> 56:19.240 So do you think larger networks will be able to start to understand the physics and the 56:19.240 --> 56:23.960 physics of the scene, the three dimensional structure and relationships of objects in 56:23.960 --> 56:25.720 the scene, and so on? 56:25.720 --> 56:28.880 Or really, that's where symbolic at has to step in? 56:28.880 --> 56:37.680 Well, it's always possible to solve these problems with deep learning is just extremely 56:37.680 --> 56:38.680 inefficient. 56:38.680 --> 56:45.240 A model would be an explicit rule based abstract model would be a far better, more compressed 56:45.240 --> 56:50.280 representation of physics than learning just this mapping between in this situation, this 56:50.280 --> 56:51.280 thing happens. 56:51.280 --> 56:54.520 If you change the situation slightly, then this other thing happens and so on. 56:54.520 --> 57:00.840 Do you think it's possible to automatically generate the programs that would require that 57:00.840 --> 57:01.840 kind of reasoning? 57:01.840 --> 57:07.120 Or does it have to, so where expert systems fail, there's so many facts about the world 57:07.120 --> 57:08.640 had to be hand coded in. 57:08.640 --> 57:15.360 Do you think it's possible to learn those logical statements that are true about the 57:15.360 --> 57:17.120 world and their relationships? 57:17.120 --> 57:22.640 I mean, that's kind of what they're improving at a basic level is trying to do, right? 57:22.640 --> 57:28.360 Yeah, except it's much harder to formulate statements about the world compared to fermenting 57:28.360 --> 57:30.680 mathematical statements. 57:30.680 --> 57:34.320 Statements about the world tend to be subjective. 57:34.320 --> 57:39.320 So can you learn rule based models? 57:39.320 --> 57:40.320 Yes. 57:40.320 --> 57:41.320 Yes, definitely. 57:41.320 --> 57:43.720 That's the field of program synthesis. 57:43.720 --> 57:48.080 However, today we just don't really know how to do it. 57:48.080 --> 57:52.640 So it's very much a grass search or tree search problem. 57:52.640 --> 57:58.080 And so we are limited to the sort of a tree session grass search algorithms that we have 57:58.080 --> 57:59.080 today. 57:59.080 --> 58:02.080 Personally, I think genetic algorithms are very promising. 58:02.080 --> 58:04.640 So it's almost like genetic programming. 58:04.640 --> 58:05.760 Genetic programming, exactly. 58:05.760 --> 58:12.200 Can you discuss the field of program synthesis, like what, how many people are working and 58:12.200 --> 58:13.840 thinking about it? 58:13.840 --> 58:20.360 What, where we are in the history of program synthesis and what are your hopes for it? 58:20.360 --> 58:24.760 Well, if it were deep learning, this is like the 90s. 58:24.760 --> 58:29.320 So meaning that we already have existing solutions. 58:29.320 --> 58:35.720 We are starting to have some basic understanding of what this is about. 58:35.720 --> 58:38.120 But it's still a field that is in its infancy. 58:38.120 --> 58:40.560 There are very few people working on it. 58:40.560 --> 58:44.520 There are very few real world applications. 58:44.520 --> 58:51.960 So the one real world application I'm aware of is Flash Fill in Excel. 58:51.960 --> 58:58.240 It's a way to automatically learn very simple programs to format cells in an Excel spreadsheet 58:58.240 --> 58:59.840 from a few examples. 58:59.840 --> 59:02.840 For instance, learning a way to format a date, things like that. 59:02.840 --> 59:03.840 Oh, that's fascinating. 59:03.840 --> 59:04.840 Yeah. 59:04.840 --> 59:06.280 You know, okay, that's that's fascinating topic. 59:06.280 --> 59:12.880 I was wondering when I provide a few samples to Excel, what it's able to figure out, like 59:12.880 --> 59:18.280 just giving it a few dates, what are you able to figure out from the pattern I just gave 59:18.280 --> 59:19.280 you? 59:19.280 --> 59:20.280 That's a fascinating question. 59:20.280 --> 59:24.240 It's fascinating whether that's learnable patterns and you're saying they're working 59:24.240 --> 59:25.240 on that. 59:25.240 --> 59:26.240 Yeah. 59:26.240 --> 59:27.240 How big is the toolbox currently? 59:27.240 --> 59:28.240 Yeah. 59:28.240 --> 59:29.240 Are we completely in the dark? 59:29.240 --> 59:30.240 So if you set the 90s. 59:30.240 --> 59:32.240 In terms of program synthesis? 59:32.240 --> 59:33.240 No. 59:33.240 --> 59:40.520 So I would say, so maybe 90s is even too optimistic because by the 90s, you know, we already understood 59:40.520 --> 59:41.520 backprop. 59:41.520 --> 59:44.720 We already understood, you know, the engine of deep learning, even though we couldn't 59:44.720 --> 59:50.440 really see its potential quite today, I don't think we found the engine of program synthesis. 59:50.440 --> 59:52.960 So we're in the winter before backprop. 59:52.960 --> 59:53.960 Yeah. 59:53.960 --> 59:55.760 In a way, yes. 59:55.760 --> 1:00:02.400 So I do believe program synthesis, in general, discrete search over rule based models is going 1:00:02.400 --> 1:00:06.960 to be a cornerstone of AI research in the next century, right? 1:00:06.960 --> 1:00:10.240 And that doesn't mean we're going to drop deep learning. 1:00:10.240 --> 1:00:11.960 Deep learning is immensely useful. 1:00:11.960 --> 1:00:19.480 Like being able to learn this is a very flexible, adaptable, parametric models, that's actually 1:00:19.480 --> 1:00:20.480 immensely useful. 1:00:20.480 --> 1:00:24.960 Like all it's doing, it's pattern cognition, but being good at pattern cognition, given 1:00:24.960 --> 1:00:27.880 lots of data is just extremely powerful. 1:00:27.880 --> 1:00:31.000 So we are still going to be working on deep learning and we're going to be working on 1:00:31.000 --> 1:00:32.000 program synthesis. 1:00:32.000 --> 1:00:36.520 We're going to be combining the two in increasingly automated ways. 1:00:36.520 --> 1:00:38.640 So let's talk a little bit about data. 1:00:38.640 --> 1:00:46.120 You've tweeted about 10,000 deep learning papers have been written about hard coding 1:00:46.120 --> 1:00:50.280 priors, about a specific task in a neural network architecture, it works better than 1:00:50.280 --> 1:00:52.760 a lack of a prior. 1:00:52.760 --> 1:00:57.480 By summarizing all these efforts, they put a name to an architecture, but really what 1:00:57.480 --> 1:01:01.680 they're doing is hard coding some priors that improve the performance of the system. 1:01:01.680 --> 1:01:07.000 But we get straight to the point, it's probably true. 1:01:07.000 --> 1:01:12.080 So you say that you can always buy performance, buy in quotes performance by either training 1:01:12.080 --> 1:01:17.520 on more data, better data, or by injecting task information to the architecture of the 1:01:17.520 --> 1:01:18.520 preprocessing. 1:01:18.520 --> 1:01:22.720 However, this is informative about the generalization power the techniques use, the fundamentals 1:01:22.720 --> 1:01:23.720 of ability to generalize. 1:01:23.720 --> 1:01:30.040 Do you think we can go far by coming up with better methods for this kind of cheating, 1:01:30.040 --> 1:01:35.320 for better methods of large scale annotation of data, so building better priors? 1:01:35.320 --> 1:01:37.400 If you've made it, it's not cheating anymore. 1:01:37.400 --> 1:01:38.400 Right. 1:01:38.400 --> 1:01:46.480 I'm joking about the cheating, but large scale, so basically I'm asking about something 1:01:46.480 --> 1:01:54.300 that hasn't, from my perspective, been researched too much is exponential improvement in annotation 1:01:54.300 --> 1:01:56.800 of data. 1:01:56.800 --> 1:01:58.120 You often think about... 1:01:58.120 --> 1:02:00.880 I think it's actually been researched quite a bit. 1:02:00.880 --> 1:02:06.120 You just don't see publications about it, because people who publish papers are going 1:02:06.120 --> 1:02:10.000 to publish about known benchmarks, sometimes they're going to read a new benchmark. 1:02:10.000 --> 1:02:14.360 People who actually have real world large scale defining problems, they're going to spend 1:02:14.360 --> 1:02:18.800 a lot of resources into data annotation and good data annotation pipelines, but you don't 1:02:18.800 --> 1:02:19.800 see any papers about it. 1:02:19.800 --> 1:02:20.800 That's interesting. 1:02:20.800 --> 1:02:24.600 Do you think there are certain resources, but do you think there's innovation happening? 1:02:24.600 --> 1:02:25.920 Oh, yeah. 1:02:25.920 --> 1:02:33.960 To clarify at the point in the twist, machine learning in general is the science of generalization. 1:02:33.960 --> 1:02:41.080 You want to generate knowledge that can be reused across different datasets, across different 1:02:41.080 --> 1:02:42.680 tasks. 1:02:42.680 --> 1:02:49.320 If instead you're looking at one dataset, and then you are hard coding knowledge about 1:02:49.320 --> 1:02:55.920 this task into your architecture, this is no more useful than training a network and 1:02:55.920 --> 1:03:03.160 then saying, oh, I found these weight values perform well. 1:03:03.160 --> 1:03:08.720 David Ha, I don't know if you know David, he had a paper the other day about weight 1:03:08.720 --> 1:03:13.840 agnostic neural networks, and this is very interesting paper because it really illustrates 1:03:13.840 --> 1:03:20.800 the fact that an architecture, even without weight, an architecture is a knowledge about 1:03:20.800 --> 1:03:21.800 a task. 1:03:21.800 --> 1:03:24.280 It encodes knowledge. 1:03:24.280 --> 1:03:31.560 When it comes to architectures that are uncrafted by researchers, in some cases, it is very, 1:03:31.560 --> 1:03:39.400 very clear that all they are doing is artificially reencoding the template that corresponds 1:03:39.400 --> 1:03:45.240 to the proper way to solve the task and coding in a given dataset. 1:03:45.240 --> 1:03:52.120 For instance, if you've looked at the baby dataset, which is about natural language 1:03:52.120 --> 1:03:55.800 question answering, it is generated by an algorithm. 1:03:55.800 --> 1:03:59.320 This is a question under pairs that are generated by an algorithm. 1:03:59.320 --> 1:04:01.680 The algorithm is solving a certain template. 1:04:01.680 --> 1:04:06.760 Turns out, if you craft a network that literally encodes this template, you can solve this 1:04:06.760 --> 1:04:13.160 dataset with nearly 100% accuracy, but that doesn't actually tell you anything about how 1:04:13.160 --> 1:04:17.760 to solve question answering in general, which is the point. 1:04:17.760 --> 1:04:21.560 The question is just the linger on it, whether it's from the data side or from the size of 1:04:21.560 --> 1:04:22.560 the network. 1:04:22.560 --> 1:04:27.960 I don't know if you've read the blog post by Ray Sutton, the bitter lesson, where he 1:04:27.960 --> 1:04:33.480 says the biggest lesson that we can read from 70 years of AI research is that general methods 1:04:33.480 --> 1:04:38.120 that leverage computation are ultimately the most effective. 1:04:38.120 --> 1:04:45.520 As opposed to figuring out methods that can generalize effectively, do you think we can 1:04:45.520 --> 1:04:50.720 get pretty far by just having something that leverages computation and the improvement of 1:04:50.720 --> 1:04:51.720 computation? 1:04:51.720 --> 1:04:52.720 Yes. 1:04:52.720 --> 1:04:56.880 I think Rich is making a very good point, which is that a lot of these papers, which 1:04:56.880 --> 1:05:03.760 are actually all about manually hard coding prior knowledge about a task into some system, 1:05:03.760 --> 1:05:08.720 doesn't have to be deeply architected into some system, right? 1:05:08.720 --> 1:05:11.560 These papers are not actually making any impact. 1:05:11.560 --> 1:05:18.680 Instead, what's making really long term impact is very simple, very general systems that 1:05:18.680 --> 1:05:23.560 are really agnostic to all these tricks, because these tricks do not generalize. 1:05:23.560 --> 1:05:31.680 And of course, the one general and simple thing that you should focus on is that which 1:05:31.680 --> 1:05:37.360 leverages computation, because computation, the availability of large scale computation 1:05:37.360 --> 1:05:40.720 has been increasing exponentially, following Morse law. 1:05:40.720 --> 1:05:46.160 So if your algorithm is all about exploiting this, then your algorithm is suddenly exponentially 1:05:46.160 --> 1:05:47.640 improving, right? 1:05:47.640 --> 1:05:51.800 So I think Rich is definitely right. 1:05:51.800 --> 1:05:59.520 However, he's right about the past 70 years, he's like assessing the past 70 years. 1:05:59.520 --> 1:06:05.440 I am not sure that this assessment will still hold true for the next 70 years. 1:06:05.440 --> 1:06:12.040 It might, to some extent, I suspect it will not, because the truth of his assessment is 1:06:12.040 --> 1:06:17.040 a function of the context, right, in which this research took place. 1:06:17.040 --> 1:06:22.560 And the context is changing, like Morse law might not be applicable anymore, for instance, 1:06:22.560 --> 1:06:24.080 in the future. 1:06:24.080 --> 1:06:32.320 And I do believe that when you tweak one aspect of a system, when you exploit one aspect 1:06:32.320 --> 1:06:36.680 of a system, some other aspect starts becoming the bottleneck. 1:06:36.680 --> 1:06:41.640 Let's say you have unlimited computation, well, then data is the bottleneck. 1:06:41.640 --> 1:06:46.560 And I think we are already starting to be in a regime where our systems are so large 1:06:46.560 --> 1:06:50.960 in scale and so data ingrained, the data today, and the quality of data, and the scale of 1:06:50.960 --> 1:06:53.280 data is the bottleneck. 1:06:53.280 --> 1:07:00.960 And in this environment, the beta lesson from Rich is not going to be true anymore, right? 1:07:00.960 --> 1:07:08.000 So I think we are going to move from a focus on a scale of a competition scale to focus 1:07:08.000 --> 1:07:10.080 on data efficiency. 1:07:10.080 --> 1:07:11.080 Data efficiency. 1:07:11.080 --> 1:07:13.240 So that's getting to the question of symbolic AI. 1:07:13.240 --> 1:07:19.120 But the linger on the deep learning approaches, do you have hope for either unsupervised learning 1:07:19.120 --> 1:07:28.280 or reinforcement learning, which are ways of being more data efficient in terms of the 1:07:28.280 --> 1:07:31.720 amount of data they need that require human annotation? 1:07:31.720 --> 1:07:36.320 So unsupervised learning and reinforcement learning are frameworks for learning, but 1:07:36.320 --> 1:07:39.080 they are not like any specific technique. 1:07:39.080 --> 1:07:42.800 So usually when people say reinforcement learning, what they really mean is deep reinforcement 1:07:42.800 --> 1:07:47.440 learning, which is like one approach which is actually very questionable. 1:07:47.440 --> 1:07:53.440 The question I was asking was unsupervised learning with deep neural networks and deeper 1:07:53.440 --> 1:07:54.440 reinforcement learning. 1:07:54.440 --> 1:07:58.840 Well, these are not really data efficient because you're still leveraging these huge 1:07:58.840 --> 1:08:03.760 parametric models, point by point with gradient descent. 1:08:03.760 --> 1:08:09.000 It is more efficient in terms of the number of annotations, the density of annotations 1:08:09.000 --> 1:08:10.000 you need. 1:08:10.000 --> 1:08:16.680 The idea being to learn the latent space around which the data is organized and then map the 1:08:16.680 --> 1:08:18.960 sparse annotations into it. 1:08:18.960 --> 1:08:23.640 And sure, I mean, that's clearly a very good idea. 1:08:23.640 --> 1:08:27.960 It's not really a topic I would be working on, but it's clearly a good idea. 1:08:27.960 --> 1:08:32.040 So it would get us to solve some problems that... 1:08:32.040 --> 1:08:38.280 It will get us to incremental improvements in labeled data efficiency. 1:08:38.280 --> 1:08:46.640 Do you have concerns about short term or long term threats from AI, from artificial intelligence? 1:08:46.640 --> 1:08:50.720 Yes, definitely to some extent. 1:08:50.720 --> 1:08:52.360 And what's the shape of those concerns? 1:08:52.360 --> 1:08:57.200 This is actually something I've briefly written about. 1:08:57.200 --> 1:09:06.160 But the capabilities of deep learning technology can be used in many ways that are concerning 1:09:06.160 --> 1:09:13.920 from mass surveillance with things like facial recognition, in general, tracking lots of 1:09:13.920 --> 1:09:20.040 data about everyone and then being able to making sense of this data, to do identification, 1:09:20.040 --> 1:09:22.520 to do prediction. 1:09:22.520 --> 1:09:23.520 That's concerning. 1:09:23.520 --> 1:09:31.680 That's something that's being very aggressively pursued by totalitarian states like China. 1:09:31.680 --> 1:09:40.760 One thing I am very much concerned about is that our lives are increasingly online, are 1:09:40.760 --> 1:09:45.960 increasingly digital, made of information, made of information consumption and information 1:09:45.960 --> 1:09:52.160 production or digital footprint, I would say. 1:09:52.160 --> 1:10:01.200 And if you absorb all of this data and you are in control of where you consume information, 1:10:01.200 --> 1:10:10.160 social networks and so on, recommendation engines, then you can build a sort of reinforcement 1:10:10.160 --> 1:10:13.920 loop for human behavior. 1:10:13.920 --> 1:10:18.440 You can observe the state of your mind at time t. 1:10:18.440 --> 1:10:25.040 You can predict how you would react to different pieces of content, how to get you to move 1:10:25.040 --> 1:10:33.280 your mind in a certain direction, then you can feed the specific piece of content that 1:10:33.280 --> 1:10:35.920 would move you in a specific direction. 1:10:35.920 --> 1:10:45.000 And you can do this at scale in terms of doing it continuously in real time. 1:10:45.000 --> 1:10:50.560 You can also do it at scale in terms of scaling this to many, many people, to entire populations. 1:10:50.560 --> 1:10:57.800 So potentially, artificial intelligence, even in its current state, if you combine it with 1:10:57.800 --> 1:11:04.120 the internet, with the fact that we have all of our lives are moving to digital devices 1:11:04.120 --> 1:11:11.800 and digital information consumption and creation, what you get is the possibility to achieve 1:11:11.800 --> 1:11:16.960 mass manipulation of behavior and mass psychological control. 1:11:16.960 --> 1:11:18.360 And this is a very real possibility. 1:11:18.360 --> 1:11:22.240 Yeah, so you're talking about any kind of recommender system. 1:11:22.240 --> 1:11:28.160 Let's look at the YouTube algorithm, Facebook, anything that recommends content you should 1:11:28.160 --> 1:11:35.480 watch next, and it's fascinating to think that there's some aspects of human behavior 1:11:35.480 --> 1:11:45.520 that you can say a problem of, is this person hold Republican beliefs or Democratic beliefs? 1:11:45.520 --> 1:11:52.720 And it's a trivial, that's an objective function, and you can optimize and you can measure and 1:11:52.720 --> 1:11:55.720 you can turn everybody into a Republican or everybody into a Democrat. 1:11:55.720 --> 1:11:56.720 Absolutely, yeah. 1:11:56.720 --> 1:11:57.960 I do believe it's true. 1:11:57.960 --> 1:12:02.520 So the human mind is very... 1:12:02.520 --> 1:12:06.760 If you look at the human mind as a kind of computer program, it has a very large exploit 1:12:06.760 --> 1:12:07.760 surface, right? 1:12:07.760 --> 1:12:08.760 It has many, many vulnerabilities. 1:12:08.760 --> 1:12:09.760 Exploit surfaces, yeah. 1:12:09.760 --> 1:12:16.920 Where you can control it, for instance, when it comes to your political beliefs, this is 1:12:16.920 --> 1:12:19.360 very much tied to your identity. 1:12:19.360 --> 1:12:26.080 So for instance, if I'm in control of your news feed on your favorite social media platforms, 1:12:26.080 --> 1:12:29.680 this is actually where you're getting your news from. 1:12:29.680 --> 1:12:35.560 And of course, I can choose to only show you news that will make you see the world in a 1:12:35.560 --> 1:12:37.200 specific way, right? 1:12:37.200 --> 1:12:44.720 But I can also create incentives for you to post about some political beliefs. 1:12:44.720 --> 1:12:52.720 And then when I get you to express a statement, if it's a statement that me as a controller, 1:12:52.720 --> 1:12:53.720 I want to reinforce. 1:12:53.720 --> 1:12:57.080 I can just show it to people who will agree and they will like it. 1:12:57.080 --> 1:12:59.400 And that will reinforce the statement in your mind. 1:12:59.400 --> 1:13:06.280 If this is a statement I want you to, this is a belief I want you to abandon, I can, 1:13:06.280 --> 1:13:10.800 on the other hand, show it to opponents, right, will attack you. 1:13:10.800 --> 1:13:16.440 And because they attack you at the very least, next time you will think twice about posting 1:13:16.440 --> 1:13:17.440 it. 1:13:17.440 --> 1:13:22.920 But maybe you will even, you know, stop believing this because you got pushed back, right? 1:13:22.920 --> 1:13:30.560 So there are many ways in which social media platforms can potentially control your opinions. 1:13:30.560 --> 1:13:38.320 And today, the, so all of these things are already being controlled by algorithms. 1:13:38.320 --> 1:13:43.080 These algorithms do not have any explicit political goal today. 1:13:43.080 --> 1:13:50.960 Well, potentially they could, like if some totalitarian government takes over, you know, 1:13:50.960 --> 1:13:55.280 social media platforms and decides that, you know, now we're going to use this not just 1:13:55.280 --> 1:13:59.960 for my surveillance, but also for my opinion control and behavior control, very bad things 1:13:59.960 --> 1:14:02.000 could happen. 1:14:02.000 --> 1:14:08.680 But what's really fascinating and actually quite concerning is that even without an 1:14:08.680 --> 1:14:15.480 explicit intent to manipulate, you're already seeing very dangerous dynamics in terms of 1:14:15.480 --> 1:14:19.960 how this content recommendation algorithms behave. 1:14:19.960 --> 1:14:26.920 Because right now, the goal, the objective function of these algorithms is to maximize 1:14:26.920 --> 1:14:32.600 engagement, right, which seems fairly innocuous at first, right? 1:14:32.600 --> 1:14:40.400 However, it is not because content that will maximally engage people, you know, get people 1:14:40.400 --> 1:14:44.480 to react in an emotional way, get people to click on something. 1:14:44.480 --> 1:14:54.480 It is very often content that, you know, is not healthy to the public discourse. 1:14:54.480 --> 1:15:01.560 For instance, fake news are far more likely to get you to click on them than real news, 1:15:01.560 --> 1:15:07.080 simply because they are not constrained to reality. 1:15:07.080 --> 1:15:14.120 So they can be as outrageous, as surprising as good stories as you want, because they 1:15:14.120 --> 1:15:15.120 are artificial, right? 1:15:15.120 --> 1:15:16.120 Yeah. 1:15:16.120 --> 1:15:19.640 To me, that's an exciting world because so much good can come. 1:15:19.640 --> 1:15:24.680 So there's an opportunity to educate people. 1:15:24.680 --> 1:15:31.200 You can balance people's worldview with other ideas. 1:15:31.200 --> 1:15:33.880 So there's so many objective functions. 1:15:33.880 --> 1:15:41.080 The space of objective functions that create better civilizations is large, arguably infinite. 1:15:41.080 --> 1:15:51.720 But there's also a large space that creates division and destruction, civil war, a lot 1:15:51.720 --> 1:15:53.360 of bad stuff. 1:15:53.360 --> 1:15:59.480 And the worry is, naturally, probably that space is bigger, first of all. 1:15:59.480 --> 1:16:06.920 And if we don't explicitly think about what kind of effects are going to be observed from 1:16:06.920 --> 1:16:10.280 different objective functions, then we're going to get into trouble. 1:16:10.280 --> 1:16:16.400 Because the question is, how do we get into rooms and have discussions? 1:16:16.400 --> 1:16:22.200 So inside Google, inside Facebook, inside Twitter, and think about, okay, how can we 1:16:22.200 --> 1:16:28.240 drive up engagement and at the same time create a good society? 1:16:28.240 --> 1:16:31.760 Is it even possible to have that kind of philosophical discussion? 1:16:31.760 --> 1:16:33.200 I think you can definitely try. 1:16:33.200 --> 1:16:40.160 So from my perspective, I would feel rather uncomfortable with companies that are in control 1:16:40.160 --> 1:16:49.760 of these new algorithms, with them making explicit decisions to manipulate people's opinions 1:16:49.760 --> 1:16:55.360 or behaviors, even if the intent is good, because that's a very totalitarian mindset. 1:16:55.360 --> 1:16:59.840 So instead, what I would like to see is probably never going to happen, because it's not super 1:16:59.840 --> 1:17:02.560 realistic, but that's actually something I really care about. 1:17:02.560 --> 1:17:10.680 I would like all these algorithms to present configuration settings to their users, so 1:17:10.680 --> 1:17:17.960 that the users can actually make the decision about how they want to be impacted by these 1:17:17.960 --> 1:17:22.080 information recommendation, content recommendation algorithms. 1:17:22.080 --> 1:17:27.120 For instance, as a user of something like YouTube or Twitter, maybe I want to maximize 1:17:27.120 --> 1:17:30.480 learning about a specific topic. 1:17:30.480 --> 1:17:38.720 So I want the algorithm to feed my curiosity, which is in itself a very interesting problem. 1:17:38.720 --> 1:17:44.840 So instead of maximizing my engagement, it will maximize how fast and how much I'm learning, 1:17:44.840 --> 1:17:50.880 and it will also take into account the accuracy, hopefully, of the information I'm learning. 1:17:50.880 --> 1:17:57.800 So yeah, the user should be able to determine exactly how these algorithms are affecting 1:17:57.800 --> 1:17:58.800 their lives. 1:17:58.800 --> 1:18:08.240 I don't want actually any entity making decisions about in which direction they're going to 1:18:08.240 --> 1:18:09.480 try to manipulate me. 1:18:09.480 --> 1:18:11.840 I want technology. 1:18:11.840 --> 1:18:18.520 So AI, these algorithms are increasingly going to be our interface to a world that is increasingly 1:18:18.520 --> 1:18:20.280 made of information. 1:18:20.280 --> 1:18:27.440 And I want everyone to be in control of this interface, to interface with the world on 1:18:27.440 --> 1:18:29.160 their own terms. 1:18:29.160 --> 1:18:38.040 So if someone wants these algorithms to serve their own personal growth goals, they should 1:18:38.040 --> 1:18:41.920 be able to configure these algorithms in such a way. 1:18:41.920 --> 1:18:50.400 Yeah, but so I know it's painful to have explicit decisions, but there is underlying explicit 1:18:50.400 --> 1:18:57.240 decisions, which is some of the most beautiful fundamental philosophy that we have before 1:18:57.240 --> 1:19:01.200 us, which is personal growth. 1:19:01.200 --> 1:19:08.080 If I want to watch videos from which I can learn, what does that mean? 1:19:08.080 --> 1:19:13.600 So if I have a checkbox that wants to emphasize learning, there's still an algorithm with 1:19:13.600 --> 1:19:18.000 explicit decisions in it that would promote learning. 1:19:18.000 --> 1:19:19.000 What does that mean for me? 1:19:19.000 --> 1:19:25.440 Like, for example, I've watched a documentary on Flat Earth theory, I guess. 1:19:25.440 --> 1:19:28.200 It was very, like, I learned a lot. 1:19:28.200 --> 1:19:29.880 I'm really glad I watched it. 1:19:29.880 --> 1:19:35.480 It was a friend recommended it to me, because I don't have such an allergic reaction to 1:19:35.480 --> 1:19:37.800 crazy people as my fellow colleagues do. 1:19:37.800 --> 1:19:42.320 But it was very eye opening, and for others, it might not be. 1:19:42.320 --> 1:19:47.640 From others, they might just get turned off from the same with the Republican and Democrat. 1:19:47.640 --> 1:19:50.480 And it's a non trivial problem. 1:19:50.480 --> 1:19:56.440 And first of all, if it's done well, I don't think it's something that wouldn't happen 1:19:56.440 --> 1:20:00.160 that the YouTube wouldn't be promoting or Twitter wouldn't be. 1:20:00.160 --> 1:20:02.400 It's just a really difficult problem. 1:20:02.400 --> 1:20:05.080 How do we do, how do give people control? 1:20:05.080 --> 1:20:09.000 Well, it's mostly an interface design problem. 1:20:09.000 --> 1:20:16.280 The way I see it, you want to create technology that's like a mentor or a coach or an assistant 1:20:16.280 --> 1:20:22.680 so that it's not your boss, right, you are in control of it. 1:20:22.680 --> 1:20:25.920 You are telling it what to do for you. 1:20:25.920 --> 1:20:30.760 And if you feel like it's manipulating you, it's not actually, it's not actually doing 1:20:30.760 --> 1:20:31.920 what you want. 1:20:31.920 --> 1:20:35.040 You should be able to switch to a different algorithm, you know. 1:20:35.040 --> 1:20:39.720 So that fine tune control, you kind of learn, you're trusting the human collaboration. 1:20:39.720 --> 1:20:44.440 I mean, that's how I see autonomous vehicles, too, is giving as much information as possible 1:20:44.440 --> 1:20:46.560 and you learn that dance yourself. 1:20:46.560 --> 1:20:51.040 Yeah, Adobe, I don't know if you use Adobe product for like Photoshop. 1:20:51.040 --> 1:20:56.600 Yeah, they're trying to see if they can inject YouTube into their interface, but basically 1:20:56.600 --> 1:21:01.920 allow you to show you all these videos that, because everybody's confused about what to 1:21:01.920 --> 1:21:03.360 do with features. 1:21:03.360 --> 1:21:09.720 So basically teach people by linking to, in that way, it's an assistant that shows, uses 1:21:09.720 --> 1:21:12.960 videos as a basic element of information. 1:21:12.960 --> 1:21:23.080 Okay, so what practically should people do to try to, to try to fight against abuses of 1:21:23.080 --> 1:21:26.880 these algorithms or algorithms that manipulate us? 1:21:26.880 --> 1:21:31.080 Honestly, it's a very, very difficult problem because to start with, there is very little 1:21:31.080 --> 1:21:34.120 public awareness of these issues. 1:21:34.120 --> 1:21:39.960 Very few people would think that, you know, anything wrong with their new algorithm, even 1:21:39.960 --> 1:21:44.440 though there is actually something wrong already, which is that it's trying to maximize engagement 1:21:44.440 --> 1:21:50.000 most of the time, which has very negative side effects, right? 1:21:50.000 --> 1:21:59.760 So ideally, so the very first thing is to stop trying to purely maximize engagement, try 1:21:59.760 --> 1:22:11.000 to propagate content based on popularity, right, instead take into account the goals 1:22:11.000 --> 1:22:13.640 and the profiles of each user. 1:22:13.640 --> 1:22:20.200 So you will, you will be, one example is, for instance, when I look at topic recommendations 1:22:20.200 --> 1:22:25.640 on Twitter, it's like, you know, they have this news tab with switch recommendations. 1:22:25.640 --> 1:22:33.480 That's always the worst garbage because it's content that appeals to the smallest command 1:22:33.480 --> 1:22:37.560 denominator to all Twitter users because they're trying to optimize, they're purely 1:22:37.560 --> 1:22:41.680 trying to obtain us popularity, they're purely trying to optimize engagement, but that's 1:22:41.680 --> 1:22:43.080 not what I want. 1:22:43.080 --> 1:22:50.440 So they should put me in control of some setting so that I define what's the objective function 1:22:50.440 --> 1:22:54.280 that Twitter is going to be following to show me this content. 1:22:54.280 --> 1:22:59.320 And honestly, so this is all about interface design, and we are not, it's not realistic 1:22:59.320 --> 1:23:04.760 to give users control of a bunch of knobs that define an algorithm, instead, we should 1:23:04.760 --> 1:23:11.200 purely put them in charge of defining the objective function, like let the user tell 1:23:11.200 --> 1:23:15.320 us what they want to achieve, how they want this algorithm to impact their lives. 1:23:15.320 --> 1:23:20.200 So do you think it is that or do they provide individual article by article reward structure 1:23:20.200 --> 1:23:24.760 where you give a signal, I'm glad I saw this or I'm glad I didn't? 1:23:24.760 --> 1:23:31.520 So like a Spotify type feedback mechanism, it works to some extent, I'm kind of skeptical 1:23:31.520 --> 1:23:38.920 about it because the only way the algorithm, the algorithm will attempt to relate your choices 1:23:38.920 --> 1:23:45.040 with the choices of everyone else, which might, you know, if you have an average profile that 1:23:45.040 --> 1:23:49.680 works fine, I'm sure Spotify accommodations work fine if you just like mainstream stuff. 1:23:49.680 --> 1:23:54.040 But if you don't, it can be, it's not optimal at all, actually. 1:23:54.040 --> 1:24:00.880 It'll be in an efficient search for the part of the Spotify world that represents you. 1:24:00.880 --> 1:24:09.000 So it's a tough problem, but do note that even a feedback system like what Spotify has 1:24:09.000 --> 1:24:15.680 does not give me control over why the algorithm is trying to optimize for. 1:24:15.680 --> 1:24:21.440 Well, public awareness, which is what we're doing now, is a good place to start. 1:24:21.440 --> 1:24:27.760 Do you have concerns about long term existential threats of artificial intelligence? 1:24:27.760 --> 1:24:34.800 Well, as I was saying, our world is increasingly made of information, AI algorithms are increasingly 1:24:34.800 --> 1:24:40.280 going to be our interface to this world of information, and somebody will be in control 1:24:40.280 --> 1:24:46.000 of these algorithms, and that puts us in any kind of bad situation, right? 1:24:46.000 --> 1:24:48.120 It has risks. 1:24:48.120 --> 1:24:55.000 It has risks coming from potentially large companies wanting to optimize their own goals, 1:24:55.000 --> 1:25:01.760 maybe profit, maybe something else, also from governments who might want to use these algorithms 1:25:01.760 --> 1:25:04.720 as a means of control of the entire population. 1:25:04.720 --> 1:25:07.560 Do you think there's existential threat that could arise from that? 1:25:07.560 --> 1:25:15.840 So existential threat, so maybe you're referring to the singularity narrative where robots 1:25:15.840 --> 1:25:16.840 just take over? 1:25:16.840 --> 1:25:22.040 Well, I don't not terminate a robot, and I don't believe it has to be a singularity. 1:25:22.040 --> 1:25:30.000 We're just talking to, just like you said, the algorithm controlling masses of populations, 1:25:30.000 --> 1:25:37.840 the existential threat being hurt ourselves much like a nuclear war would hurt ourselves, 1:25:37.840 --> 1:25:38.840 that kind of thing. 1:25:38.840 --> 1:25:44.600 I don't think that requires a singularity, that requires a loss of control over AI algorithms. 1:25:44.600 --> 1:25:47.920 So I do agree there are concerning trends. 1:25:47.920 --> 1:25:53.600 Honestly, I wouldn't want to make any long term predictions. 1:25:53.600 --> 1:25:59.560 I don't think today we really have the capability to see what the dangers of AI are going to 1:25:59.560 --> 1:26:02.240 be in 50 years, in 100 years. 1:26:02.240 --> 1:26:11.480 I do see that we are already faced with concrete and present dangers surrounding the negative 1:26:11.480 --> 1:26:17.280 side effects of content recombination systems of new seed algorithms concerning algorithmic 1:26:17.280 --> 1:26:19.520 bias as well. 1:26:19.520 --> 1:26:26.000 So we are delegating more and more decision processes to algorithms. 1:26:26.000 --> 1:26:30.160 Some of these algorithms are uncrafted, some are learned from data. 1:26:30.160 --> 1:26:34.040 But we are delegating control. 1:26:34.040 --> 1:26:37.240 Sometimes it's a good thing, sometimes not so much. 1:26:37.240 --> 1:26:41.720 And there is in general very little supervision of this process. 1:26:41.720 --> 1:26:50.160 So we are still in this period of very fast change, even chaos, where society is restructuring 1:26:50.160 --> 1:26:56.160 itself, turning into an information society, which itself is turning into an increasingly 1:26:56.160 --> 1:26:59.240 automated information processing society. 1:26:59.240 --> 1:27:05.760 And well, yeah, I think the best we can do today is try to raise awareness around some 1:27:05.760 --> 1:27:06.760 of these issues. 1:27:06.760 --> 1:27:13.000 And I think we are actually making good progress if you look at algorithmic bias, for instance. 1:27:13.000 --> 1:27:17.240 Three years ago, even two years ago, very, very few people were talking about it. 1:27:17.240 --> 1:27:22.400 And now all the big companies are talking about it, often not in a very serious way, 1:27:22.400 --> 1:27:24.600 but at least it is part of the public discourse. 1:27:24.600 --> 1:27:27.360 You see people in Congress talking about it. 1:27:27.360 --> 1:27:32.840 And it all started from raising awareness. 1:27:32.840 --> 1:27:40.200 So in terms of alignment problem, trying to teach as we allow algorithms, just even recommend 1:27:40.200 --> 1:27:50.280 their systems on Twitter, encoding human values and morals, decisions that touch on ethics. 1:27:50.280 --> 1:27:52.640 How hard do you think that problem is? 1:27:52.640 --> 1:27:59.800 How do we have lost functions in neural networks that have some component, some fuzzy components 1:27:59.800 --> 1:28:01.280 of human morals? 1:28:01.280 --> 1:28:07.400 Well, I think this is really all about objective function engineering, which is probably going 1:28:07.400 --> 1:28:10.680 to be increasingly a topic of concern in the future. 1:28:10.680 --> 1:28:16.160 Like for now, we are just using very naive loss functions because the hard part is not 1:28:16.160 --> 1:28:19.240 actually what you're trying to minimize, it's everything else. 1:28:19.240 --> 1:28:25.280 But as the everything else is going to be increasingly automated, we're going to be 1:28:25.280 --> 1:28:30.920 focusing our human attention on increasingly high level components, like what's actually 1:28:30.920 --> 1:28:34.040 driving the whole learning system, like the objective function. 1:28:34.040 --> 1:28:38.360 So loss function engineering is going to be, loss function engineer is probably going to 1:28:38.360 --> 1:28:40.760 be a job title in the future. 1:28:40.760 --> 1:28:46.200 And then the tooling you're creating with Keras essentially takes care of all the details 1:28:46.200 --> 1:28:52.960 underneath and basically the human expert is needed for exactly that. 1:28:52.960 --> 1:28:59.240 Keras is the interface between the data you're collecting and the business goals. 1:28:59.240 --> 1:29:04.280 And your job as an engineer is going to be to express your business goals and your understanding 1:29:04.280 --> 1:29:10.440 of your business or your product, your system as a kind of loss function or a kind of set 1:29:10.440 --> 1:29:11.440 of constraints. 1:29:11.440 --> 1:29:19.560 Does the possibility of creating an AGI system excite you or scare you or bore you? 1:29:19.560 --> 1:29:23.600 So intelligence can never really be general, you know, at best it can have some degree 1:29:23.600 --> 1:29:26.600 of generality, like human intelligence. 1:29:26.600 --> 1:29:30.720 And it's also always as some specialization in the same way that human intelligence is 1:29:30.720 --> 1:29:35.680 specialized in a certain category of problems, is specialized in the human experience. 1:29:35.680 --> 1:29:41.440 And when people talk about AGI, I'm never quite sure if they're talking about very, 1:29:41.440 --> 1:29:46.200 very smart AI, so smart that it's even smarter than humans, or they're talking about human 1:29:46.200 --> 1:29:49.880 like intelligence, because these are different things. 1:29:49.880 --> 1:29:54.840 Let's say, presumably I'm oppressing you today with my humanness. 1:29:54.840 --> 1:29:59.400 So imagine that I was in fact a robot. 1:29:59.400 --> 1:30:02.400 So what does that mean? 1:30:02.400 --> 1:30:05.160 I'm oppressing you with natural language processing. 1:30:05.160 --> 1:30:08.320 Maybe if you weren't able to see me, maybe this is a phone call. 1:30:08.320 --> 1:30:09.320 That kind of system. 1:30:09.320 --> 1:30:10.320 Okay. 1:30:10.320 --> 1:30:11.320 So companion. 1:30:11.320 --> 1:30:15.200 So that's very much about building human like AI. 1:30:15.200 --> 1:30:18.200 And you're asking me, you know, is this an exciting perspective? 1:30:18.200 --> 1:30:19.200 Yes. 1:30:19.200 --> 1:30:21.960 I think so, yes. 1:30:21.960 --> 1:30:29.640 Not so much because of what artificial human like intelligence could do, but, you know, 1:30:29.640 --> 1:30:34.240 from an intellectual perspective, I think if you could build truly human like intelligence, 1:30:34.240 --> 1:30:40.160 that means you could actually understand human intelligence, which is fascinating, right? 1:30:40.160 --> 1:30:44.480 Human like intelligence is going to require emotions, it's going to require consciousness, 1:30:44.480 --> 1:30:48.640 which is not things that would normally be required by an intelligent system. 1:30:48.640 --> 1:30:55.560 If you look at, you know, we were mentioning earlier like science as a superhuman problem 1:30:55.560 --> 1:31:02.240 solving agent or system, it does not have consciousness, it doesn't have emotions. 1:31:02.240 --> 1:31:07.760 In general, so emotions, I see consciousness as being on the same spectrum as emotions. 1:31:07.760 --> 1:31:17.560 It is a component of the subjective experience that is meant very much to guide behavior 1:31:17.560 --> 1:31:20.880 generation, right, it's meant to guide your behavior. 1:31:20.880 --> 1:31:27.080 In general, human intelligence and animal intelligence has evolved for the purpose of 1:31:27.080 --> 1:31:30.760 behavior generation, right, including in a social context. 1:31:30.760 --> 1:31:32.600 So that's why we actually need emotions. 1:31:32.600 --> 1:31:35.080 That's why we need consciousness. 1:31:35.080 --> 1:31:39.280 An artificial intelligence system developed in a different context may well never need 1:31:39.280 --> 1:31:43.280 them, may well never be conscious like science. 1:31:43.280 --> 1:31:50.160 But on that point, I would argue it's possible to imagine that there's echoes of consciousness 1:31:50.160 --> 1:31:55.640 in science when viewed as an organism, that science is consciousness. 1:31:55.640 --> 1:31:59.320 So I mean, how would you go about testing this hypothesis? 1:31:59.320 --> 1:32:07.240 How do you probe the subjective experience of an abstract system like science? 1:32:07.240 --> 1:32:12.280 Well the point of probing any subjective experience is impossible, because I'm not science, I'm 1:32:12.280 --> 1:32:13.280 a science. 1:32:13.280 --> 1:32:20.720 So I can't probe another entity's, another, it's no more than bacteria on my skin. 1:32:20.720 --> 1:32:25.360 Your legs, I can ask you questions about your subjective experience and you can answer me. 1:32:25.360 --> 1:32:27.720 And that's how I know you're conscious. 1:32:27.720 --> 1:32:32.080 Yes, but that's because we speak the same language. 1:32:32.080 --> 1:32:35.800 You perhaps, we have to speak the language of science and we have to ask it. 1:32:35.800 --> 1:32:41.120 Honestly, I don't think consciousness, just like emotions of pain and pleasure, is not 1:32:41.120 --> 1:32:47.120 something that inevitably arises from any sort of sufficiently intelligent information 1:32:47.120 --> 1:32:48.120 processing. 1:32:48.120 --> 1:32:54.080 It is a feature of the mind and if you've not implemented it explicitly, it is not there. 1:32:54.080 --> 1:32:59.120 So you think it's an emergent feature of a particular architecture. 1:32:59.120 --> 1:33:00.120 So do you think? 1:33:00.120 --> 1:33:02.080 It's a feature in the same sense. 1:33:02.080 --> 1:33:09.800 So again, the subjective experience is all about guiding behavior. 1:33:09.800 --> 1:33:15.560 If the problems you're trying to solve don't really involve embedded agents, maybe in a 1:33:15.560 --> 1:33:19.800 social context, generating behavior and pursuing goals like this. 1:33:19.800 --> 1:33:23.280 And if you look at science, that's not really what's happening, even though it is, it is 1:33:23.280 --> 1:33:29.600 a form of artificial air in this artificial intelligence in the sense that it is solving 1:33:29.600 --> 1:33:35.240 problems, it is committing knowledge, committing solutions and so on. 1:33:35.240 --> 1:33:41.120 So if you're not explicitly implementing a subjective experience, implementing certain 1:33:41.120 --> 1:33:47.120 emotions and implementing consciousness, it's not going to just spontaneously emerge. 1:33:47.120 --> 1:33:48.360 Yeah. 1:33:48.360 --> 1:33:53.640 But so for a system like human like intelligent system that has consciousness, do you think 1:33:53.640 --> 1:33:55.240 it needs to have a body? 1:33:55.240 --> 1:33:56.240 Yes, definitely. 1:33:56.240 --> 1:33:59.920 I mean, it doesn't have to be a physical body, right? 1:33:59.920 --> 1:34:03.680 And there's not that much difference between a realistic simulation in the real world. 1:34:03.680 --> 1:34:06.560 So there has to be something you have to preserve kind of thing. 1:34:06.560 --> 1:34:07.560 Yes. 1:34:07.560 --> 1:34:12.400 But human like intelligence can only arise in a human like context. 1:34:12.400 --> 1:34:13.400 Intelligence needs to be tired. 1:34:13.400 --> 1:34:20.480 You need other humans in order for you to demonstrate that you have human like intelligence, essentially. 1:34:20.480 --> 1:34:29.240 So what kind of tests and demonstration would be sufficient for you to demonstrate human 1:34:29.240 --> 1:34:30.480 like intelligence? 1:34:30.480 --> 1:34:31.480 Yeah. 1:34:31.480 --> 1:34:37.080 And just out of curiosity, you talked about in terms of theorem proving and program synthesis, 1:34:37.080 --> 1:34:40.480 I think you've written about that there's no good benchmarks for this. 1:34:40.480 --> 1:34:41.480 Yeah. 1:34:41.480 --> 1:34:42.480 That's one of the problems. 1:34:42.480 --> 1:34:46.560 So let's talk programs, program synthesis. 1:34:46.560 --> 1:34:51.440 So what do you imagine is a good, I think it's related questions for human like intelligence 1:34:51.440 --> 1:34:53.720 and for program synthesis. 1:34:53.720 --> 1:34:56.160 What's a good benchmark for either or both? 1:34:56.160 --> 1:34:57.160 Right. 1:34:57.160 --> 1:34:59.400 So I mean, you're actually asking two questions. 1:34:59.400 --> 1:35:06.520 Which is one is about quantifying intelligence and comparing the intelligence of an artificial 1:35:06.520 --> 1:35:08.800 system to the intelligence for human. 1:35:08.800 --> 1:35:13.520 And the other is about a degree to which this intelligence is human like. 1:35:13.520 --> 1:35:16.800 It's actually two different questions. 1:35:16.800 --> 1:35:19.320 So if you look, you mentioned earlier the Turing test. 1:35:19.320 --> 1:35:20.320 Right. 1:35:20.320 --> 1:35:24.080 Well, I actually don't like the Turing test because it's very lazy. 1:35:24.080 --> 1:35:28.960 It's all about completely bypassing the problem of defining and measuring intelligence. 1:35:28.960 --> 1:35:34.400 And instead delegating to a human judge or a panel of human judges. 1:35:34.400 --> 1:35:38.400 So it's a total cobalt, right? 1:35:38.400 --> 1:35:45.640 If you want to measure how human like an agent is, I think you have to make it interact 1:35:45.640 --> 1:35:47.920 with other humans. 1:35:47.920 --> 1:35:54.120 Maybe it's not necessarily a good idea to have these other humans be the judges. 1:35:54.120 --> 1:36:00.800 Maybe you should just observe BFU and compare it to what the human would actually have done. 1:36:00.800 --> 1:36:09.160 When it comes to measuring how smart, how clever an agent is and comparing that to the 1:36:09.160 --> 1:36:11.240 degree of human intelligence. 1:36:11.240 --> 1:36:13.680 So we're already talking about two things, right? 1:36:13.680 --> 1:36:20.600 The degree, kind of like the magnitude of an intelligence and its direction, right? 1:36:20.600 --> 1:36:23.560 Like the norm of a vector and its direction. 1:36:23.560 --> 1:36:27.200 And the direction is like human likeness. 1:36:27.200 --> 1:36:32.880 And the magnitude, the norm is intelligence. 1:36:32.880 --> 1:36:34.280 You could call it intelligence, right? 1:36:34.280 --> 1:36:42.440 So the direction, your sense, the space of directions that are human like is very narrow. 1:36:42.440 --> 1:36:49.880 So the way you would measure the magnitude of intelligence in a system in a way that 1:36:49.880 --> 1:36:54.960 also enables you to compare it to that of a human. 1:36:54.960 --> 1:37:02.000 Well, if you look at different benchmarks for intelligence today, they're all too focused 1:37:02.000 --> 1:37:04.480 on skill at a given task. 1:37:04.480 --> 1:37:11.080 That's skill at playing chess, skill at playing Go, skill at playing Dota. 1:37:11.080 --> 1:37:17.560 And I think that's not the right way to go about it because you can always be the human 1:37:17.560 --> 1:37:20.240 at one specific task. 1:37:20.240 --> 1:37:25.320 The reason why our skill at playing Go or at juggling or anything is impressive is because 1:37:25.320 --> 1:37:29.480 we are expressing this skill within a certain set of constraints. 1:37:29.480 --> 1:37:33.840 If you remove the constraints, the constraints that we have one lifetime, that we have this 1:37:33.840 --> 1:37:40.120 body and so on, if you remove the context, if you have unlimited train data, if you 1:37:40.120 --> 1:37:44.840 can have access to, you know, for instance, if you look at juggling, if you have no restriction 1:37:44.840 --> 1:37:50.040 on the hardware, then achieving arbitrary levels of skill is not very interesting and 1:37:50.040 --> 1:37:53.960 says nothing about the amount of intelligence you've achieved. 1:37:53.960 --> 1:37:59.320 So if you want to measure intelligence, you need to rigorously define what intelligence 1:37:59.320 --> 1:38:04.360 is, which in itself, you know, it's a very challenging problem. 1:38:04.360 --> 1:38:05.960 And do you think that's possible? 1:38:05.960 --> 1:38:06.960 To define intelligence? 1:38:06.960 --> 1:38:07.960 Yes, absolutely. 1:38:07.960 --> 1:38:11.680 I mean, you can provide, many people have provided, you know, some definition. 1:38:11.680 --> 1:38:13.640 I have my own definition. 1:38:13.640 --> 1:38:16.520 Where does your definition begin if it doesn't end? 1:38:16.520 --> 1:38:25.560 Well, I think intelligence is essentially the efficiency with which you turn experience 1:38:25.560 --> 1:38:29.960 into generalizable programs. 1:38:29.960 --> 1:38:35.280 So what that means is it's the efficiency with which you turn a sampling of experience 1:38:35.280 --> 1:38:46.200 space into the ability to process a larger chunk of experience space. 1:38:46.200 --> 1:38:53.480 So measuring skill can be one proxy because many, many different tasks can be one proxy 1:38:53.480 --> 1:38:54.680 for measure intelligence. 1:38:54.680 --> 1:38:58.880 But if you want to only measure skill, you should control for two things. 1:38:58.880 --> 1:39:07.920 You should control for the amount of experience that your system has and the priors that your 1:39:07.920 --> 1:39:08.920 system has. 1:39:08.920 --> 1:39:14.120 But if you control, if you look at two agents and you give them the same priors and you 1:39:14.120 --> 1:39:21.480 give them the same amount of experience, there is one of the agents that is going to learn 1:39:21.480 --> 1:39:27.720 programs, representation, something, a model that will perform well on the larger chunk 1:39:27.720 --> 1:39:29.760 of experience space than the other. 1:39:29.760 --> 1:39:31.920 And that is the smaller agent. 1:39:31.920 --> 1:39:32.920 Yeah. 1:39:32.920 --> 1:39:39.920 So if you fix the experience, which generate better programs, better meaning, more generalizable, 1:39:39.920 --> 1:39:40.920 that's really interesting. 1:39:40.920 --> 1:39:42.760 That's a very nice, clean definition of... 1:39:42.760 --> 1:39:49.560 By the way, in this definition, it is already very obvious that intelligence has to be specialized 1:39:49.560 --> 1:39:53.600 because you're talking about experience space and you're talking about segments of experience 1:39:53.600 --> 1:39:54.600 space. 1:39:54.600 --> 1:39:59.680 You're talking about priors and you're talking about experience, all of these things define 1:39:59.680 --> 1:40:04.840 the context in which intelligence emerges. 1:40:04.840 --> 1:40:10.040 And you can never look at the totality of experience space. 1:40:10.040 --> 1:40:12.520 So intelligence has to be specialized. 1:40:12.520 --> 1:40:16.760 But it can be sufficiently large, the experience space, even though specialized is a certain 1:40:16.760 --> 1:40:22.200 point when the experience space is large enough to where it might as well be general. 1:40:22.200 --> 1:40:23.200 It feels general. 1:40:23.200 --> 1:40:24.200 It looks general. 1:40:24.200 --> 1:40:25.200 I mean, it's very relative. 1:40:25.200 --> 1:40:29.560 For instance, many people would say human intelligence is general. 1:40:29.560 --> 1:40:32.960 In fact, it is quite specialized. 1:40:32.960 --> 1:40:37.960 We can definitely build systems that start from the same innate priors as what humans 1:40:37.960 --> 1:40:43.720 have at birth because we already understand fairly well what sort of priors we have as 1:40:43.720 --> 1:40:44.720 humans. 1:40:44.720 --> 1:40:50.680 Like many people have worked on this problem, most notably, Elzebeth Spelke from Harvard, 1:40:50.680 --> 1:40:56.240 and if you know her, she's worked a lot on what she calls a core knowledge. 1:40:56.240 --> 1:41:02.560 And it is very much about trying to determine and describe what priors we are born with. 1:41:02.560 --> 1:41:06.080 Like language skills and so on and all that kind of stuff. 1:41:06.080 --> 1:41:07.080 Exactly. 1:41:07.080 --> 1:41:11.520 So we have some pretty good understanding of what priors we are born with. 1:41:11.520 --> 1:41:13.960 So we could... 1:41:13.960 --> 1:41:18.720 So I've actually been working on a benchmark for the past couple of years, on and off. 1:41:18.720 --> 1:41:21.440 I hope to be able to release it at some point. 1:41:21.440 --> 1:41:29.120 The idea is to measure the intelligence of systems by considering for priors, considering 1:41:29.120 --> 1:41:34.840 for amount of experience, and by assuming the same priors as what humans are born with 1:41:34.840 --> 1:41:40.160 so that you can actually compare these scores to human intelligence and you can actually 1:41:40.160 --> 1:41:44.440 have humans pass the same test in a way that's fair. 1:41:44.440 --> 1:41:54.720 And so importantly, such a benchmark should be such that any amount of practicing does 1:41:54.720 --> 1:41:56.800 not increase your score. 1:41:56.800 --> 1:42:04.120 So try to picture a game where no matter how much you play this game, it does not change 1:42:04.120 --> 1:42:05.400 your skill at the game. 1:42:05.400 --> 1:42:08.600 Can you picture that? 1:42:08.600 --> 1:42:14.840 As a person who deeply appreciates practice, I cannot actually... 1:42:14.840 --> 1:42:19.040 There's actually a very simple trick. 1:42:19.040 --> 1:42:24.760 So in order to come up with a task, so the only thing you can measure is skill at a task. 1:42:24.760 --> 1:42:28.280 All tasks are going to involve priors. 1:42:28.280 --> 1:42:32.480 The trick is to know what they are and to describe that. 1:42:32.480 --> 1:42:36.040 And then you make sure that this is the same set of priors as what humans start with. 1:42:36.040 --> 1:42:41.080 So you create a task that assumes these priors, that exactly documents these priors, so that 1:42:41.080 --> 1:42:44.720 the priors are made explicit and there are no other priors involved. 1:42:44.720 --> 1:42:52.240 And then you generate a certain number of samples in experience space for this task. 1:42:52.240 --> 1:42:59.480 And this, for one task, assuming that the task is new for the agent passing it, that's 1:42:59.480 --> 1:43:07.560 one test of this definition of intelligence that we set up. 1:43:07.560 --> 1:43:12.360 And now you can scale that to many different tasks, that each task should be new to the 1:43:12.360 --> 1:43:13.360 agent passing it. 1:43:13.360 --> 1:43:18.680 And also should be human interpretable and understandable, so that you can actually have 1:43:18.680 --> 1:43:21.960 a human pass the same test and then you can compare the score of your machine and the score 1:43:21.960 --> 1:43:22.960 of your human. 1:43:22.960 --> 1:43:23.960 Which could be a lot. 1:43:23.960 --> 1:43:28.580 It could even start a task like MNIST, just as long as you start with the same set of 1:43:28.580 --> 1:43:29.580 priors. 1:43:29.580 --> 1:43:35.880 Yeah, so the problem with MNIST, humans are already trained to recognize digits. 1:43:35.880 --> 1:43:44.240 But let's say we're considering objects that are not digits, some complete arbitrary patterns. 1:43:44.240 --> 1:43:50.120 Well, humans already come with visual priors about how to process that. 1:43:50.120 --> 1:43:55.760 So in order to make the game fair, you would have to isolate these priors and describe 1:43:55.760 --> 1:43:58.720 them and then express them as computational rules. 1:43:58.720 --> 1:44:03.760 Having worked a lot with vision science people has exceptionally difficult, a lot of progress 1:44:03.760 --> 1:44:07.720 has been made, there's been a lot of good tests, and basically reducing all of human 1:44:07.720 --> 1:44:09.360 vision into some good priors. 1:44:09.360 --> 1:44:14.640 We're still probably far away from that perfectly, but as a start for a benchmark, that's an 1:44:14.640 --> 1:44:15.640 exciting possibility. 1:44:15.640 --> 1:44:25.320 Yeah, so Elisabeth Belke actually lists objectness as one of the core knowledge priors. 1:44:25.320 --> 1:44:26.320 Objectness. 1:44:26.320 --> 1:44:27.320 Cool. 1:44:27.320 --> 1:44:28.320 Objectness. 1:44:28.320 --> 1:44:29.320 Yeah. 1:44:29.320 --> 1:44:33.000 So we have priors about objectness, like about the visual space, about time, about agents, 1:44:33.000 --> 1:44:34.600 about goal oriented behavior. 1:44:34.600 --> 1:44:42.680 We have many different priors, but what's interesting is that, sure, we have this pretty 1:44:42.680 --> 1:44:48.520 diverse and rich set of priors, but it's also not that diverse, right? 1:44:48.520 --> 1:44:52.560 We are not born into this world with a ton of knowledge about the world. 1:44:52.560 --> 1:44:59.240 There is only a small set of core knowledge, right? 1:44:59.240 --> 1:45:00.240 Yeah. 1:45:00.240 --> 1:45:07.120 So do you have a sense of how it feels to us humans that that set is not that large, 1:45:07.120 --> 1:45:11.920 but just even the nature of time that we kind of integrate pretty effectively through all 1:45:11.920 --> 1:45:17.680 of our perception, all of our reasoning, maybe how, you know, do you have a sense of 1:45:17.680 --> 1:45:19.880 how easy it is to encode those priors? 1:45:19.880 --> 1:45:26.000 Maybe it requires building a universe, and then the human brain in order to encode those 1:45:26.000 --> 1:45:27.000 priors. 1:45:27.000 --> 1:45:30.680 Or do you have a hope that it can be listed like an XAMAT? 1:45:30.680 --> 1:45:31.680 I don't think so. 1:45:31.680 --> 1:45:36.480 So you have to keep in mind that any knowledge about the world that we are born with is something 1:45:36.480 --> 1:45:43.280 that has to have been encoded into our DNA by evolution at some point. 1:45:43.280 --> 1:45:50.720 And DNA is a very, very low bandwidth medium, like it's extremely long and expensive to 1:45:50.720 --> 1:45:57.120 encode anything into DNA, because first of all, you need some sort of evolutionary pressure 1:45:57.120 --> 1:45:59.400 to guide this writing process. 1:45:59.400 --> 1:46:05.720 And then, you know, the higher level of information you're trying to write, the longer it's going 1:46:05.720 --> 1:46:13.960 to be, and the thing in the environment that you're trying to encode knowledge about has 1:46:13.960 --> 1:46:17.240 to be stable over this duration. 1:46:17.240 --> 1:46:22.840 So you can only encode into DNA things that constitute an evolutionary advantage. 1:46:22.840 --> 1:46:27.120 So this is actually a very small subset of all possible knowledge about the world. 1:46:27.120 --> 1:46:33.360 You can only encode things that are stable, that are true over very, very long periods 1:46:33.360 --> 1:46:35.480 of time, typically millions of years. 1:46:35.480 --> 1:46:40.520 For instance, we might have some visual prior about the shape of snakes, right? 1:46:40.520 --> 1:46:43.800 But what makes a face? 1:46:43.800 --> 1:46:46.440 What's the difference between a face and a nonface? 1:46:46.440 --> 1:46:49.840 But consider this interesting question. 1:46:49.840 --> 1:46:57.800 Do we have any innate sense of the visual difference between a male face and a female 1:46:57.800 --> 1:46:58.800 face? 1:46:58.800 --> 1:46:59.800 What do you think? 1:46:59.800 --> 1:47:01.320 For a human, I mean. 1:47:01.320 --> 1:47:05.920 I would have to look back into evolutionary history when the genders emerged. 1:47:05.920 --> 1:47:11.280 But yeah, most, I mean, the faces of humans are quite different from the faces of great 1:47:11.280 --> 1:47:14.000 apes, great apes, right? 1:47:14.000 --> 1:47:15.000 Yeah. 1:47:15.000 --> 1:47:16.000 That's interesting. 1:47:16.000 --> 1:47:17.000 But yeah. 1:47:17.000 --> 1:47:23.200 You couldn't tell the face of a female chimpanzee from the face of a male chimpanzee, probably. 1:47:23.200 --> 1:47:24.200 Yeah. 1:47:24.200 --> 1:47:26.720 And I don't think most humans evolve all that ability. 1:47:26.720 --> 1:47:33.160 We do have innate knowledge of what makes a face, but it's actually impossible for us 1:47:33.160 --> 1:47:39.200 to have any DNA encoding knowledge of the difference between a female human face and 1:47:39.200 --> 1:47:40.680 a male human face. 1:47:40.680 --> 1:47:50.800 Because that knowledge, that information came up into the world actually very recently. 1:47:50.800 --> 1:47:56.920 If you look at the slowness of the process of encoding knowledge into DNA. 1:47:56.920 --> 1:47:57.920 Yeah. 1:47:57.920 --> 1:47:58.920 So that's interesting. 1:47:58.920 --> 1:48:01.640 That's a really powerful argument that DNA is a low bandwidth and it takes a long time 1:48:01.640 --> 1:48:05.480 to encode that naturally creates a very efficient encoding. 1:48:05.480 --> 1:48:12.400 But one important consequence of this is that, so yes, we are born into this world with a 1:48:12.400 --> 1:48:17.440 bunch of knowledge, sometimes very high level knowledge about the world like the rough shape 1:48:17.440 --> 1:48:20.800 of a snake, of the rough shape of a face. 1:48:20.800 --> 1:48:27.040 But importantly, because this knowledge takes so long to write, almost all of this innate 1:48:27.040 --> 1:48:33.360 knowledge is shared with our cousins, with great apes, right? 1:48:33.360 --> 1:48:37.600 So it is not actually this innate knowledge that makes us special. 1:48:37.600 --> 1:48:44.120 But to throw it right back at you from the earlier on in our discussion, that encoding 1:48:44.120 --> 1:48:50.600 might also include the entirety of the environment of Earth. 1:48:50.600 --> 1:48:56.520 To sum up, so it can include things that are important to survival and production. 1:48:56.520 --> 1:49:01.840 So for which there is some evolutionary pressure and things that are stable, constant over 1:49:01.840 --> 1:49:05.240 very, very, very long time periods. 1:49:05.240 --> 1:49:07.440 And honestly, it's not that much information. 1:49:07.440 --> 1:49:15.600 There's also, besides the bandwidths, constraints and constraints of the writing process, there's 1:49:15.600 --> 1:49:22.600 also memory constraints like DNA, the part of DNA that deals with the human brain, it's 1:49:22.600 --> 1:49:23.600 actually fairly small. 1:49:23.600 --> 1:49:26.360 It's like, you know, on the order of megabytes, right? 1:49:26.360 --> 1:49:31.880 There's not that much high level knowledge about the world you can encode. 1:49:31.880 --> 1:49:39.400 That's quite brilliant and hopeful for a benchmark that you're referring to of encoding priors. 1:49:39.400 --> 1:49:43.680 I actually look forward to, I'm skeptical that you can do it in the next couple of years, 1:49:43.680 --> 1:49:44.680 but hopefully... 1:49:44.680 --> 1:49:45.960 I've been working on it. 1:49:45.960 --> 1:49:50.120 So honestly, it's a very simple benchmark and it's not like a big breakthrough or anything. 1:49:50.120 --> 1:49:53.920 It's more like a fun side project, right? 1:49:53.920 --> 1:49:56.720 So is ImageNet. 1:49:56.720 --> 1:50:04.120 These fun side projects could launch entire groups of efforts towards creating reasoning 1:50:04.120 --> 1:50:05.120 systems and so on. 1:50:05.120 --> 1:50:06.120 And I think... 1:50:06.120 --> 1:50:07.120 Yeah, that's the goal. 1:50:07.120 --> 1:50:12.160 It's trying to measure strong generalization, to measure the strength of abstraction in 1:50:12.160 --> 1:50:17.160 our minds, well, in our minds and in an artificially intelligent agency. 1:50:17.160 --> 1:50:24.960 And if there's anything true about this science organism, it's individual cells love competition. 1:50:24.960 --> 1:50:27.000 So benchmarks encourage competition. 1:50:27.000 --> 1:50:29.680 So that's an exciting possibility. 1:50:29.680 --> 1:50:30.680 If you... 1:50:30.680 --> 1:50:35.720 Do you think an AI winter is coming and how do we prevent it? 1:50:35.720 --> 1:50:36.720 Not really. 1:50:36.720 --> 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 --> 1:50:47.560 are selling the capabilities of AI and the actual capabilities of AI. 1:50:47.560 --> 1:50:52.000 And today, deep learning is creating a lot of value and it will keep creating a lot of 1:50:52.000 --> 1:50:59.360 value in the sense that these models are applicable to a very wide range of problems that are 1:50:59.360 --> 1:51:00.360 even today. 1:51:00.360 --> 1:51:05.320 And we are only just getting started with applying algorithms to every problem they 1:51:05.320 --> 1:51:06.520 could be solving. 1:51:06.520 --> 1:51:10.440 So deep learning will keep creating a lot of value for the time being. 1:51:10.440 --> 1:51:16.000 What's concerning, however, is that there's a lot of hype around deep learning and around 1:51:16.000 --> 1:51:17.000 AI. 1:51:17.000 --> 1:51:22.840 A lot of people are overselling the capabilities of these systems, not just the capabilities 1:51:22.840 --> 1:51:31.520 but also overselling the fact that they might be more or less brain like, like given a kind 1:51:31.520 --> 1:51:40.480 of a mystical aspect, these technologies, and also overselling the pace of progress, 1:51:40.480 --> 1:51:46.000 which it might look fast in the sense that we have this exponentially increasing number 1:51:46.000 --> 1:51:48.080 of papers. 1:51:48.080 --> 1:51:53.000 But again, that's just a simple consequence of the fact that we have ever more people 1:51:53.000 --> 1:51:54.000 coming into the field. 1:51:54.000 --> 1:51:58.000 It doesn't mean the progress is actually exponentially fast. 1:51:58.000 --> 1:52:02.960 Like, let's say you're trying to raise money for your startup or your research lab. 1:52:02.960 --> 1:52:09.120 You might want to tell, you know, a grand yos 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 --> 1:52:19.040 and robotics and so on, and maybe you can tell them that the field is progressing so fast 1:52:19.040 --> 1:52:27.000 and we're going to have AI within 15 years or even 10 years, and none of this is true. 1:52:27.000 --> 1:52:33.320 And every time you're like saying these things and an investor or, you know, a decision maker 1:52:33.320 --> 1:52:43.400 believes them, well, this is like the equivalent of taking on credit card debt, but for trust. 1:52:43.400 --> 1:52:50.920 And maybe this will, you know, this will be what enables you to raise a lot of money, 1:52:50.920 --> 1:52:55.160 but ultimately you are creating damage, you are damaging the field. 1:52:55.160 --> 1:53:01.240 That's the concern is that debt, that's what happens with the other AI winters is the concern 1:53:01.240 --> 1:53:04.440 is you actually tweeted about this with autonomous vehicles, right? 1:53:04.440 --> 1:53:08.960 There's almost every single company now have promised that they will have full autonomous 1:53:08.960 --> 1:53:12.000 vehicles by 2021, 2022. 1:53:12.000 --> 1:53:18.280 That's a good example of the consequences of overhyping the capabilities of AI and the 1:53:18.280 --> 1:53:19.280 pace of progress. 1:53:19.280 --> 1:53:25.160 So because I work especially a lot recently in this area, I have a deep concern of what 1:53:25.160 --> 1:53:30.480 happens when all of these companies after every invested billions have a meeting and 1:53:30.480 --> 1:53:33.720 say, how much do we actually, first of all, do we have an autonomous vehicle? 1:53:33.720 --> 1:53:36.360 The answer will definitely be no. 1:53:36.360 --> 1:53:40.680 And second will be, wait a minute, we've invested one, two, three, four billion dollars 1:53:40.680 --> 1:53:43.400 into this and we made no profit. 1:53:43.400 --> 1:53:49.280 And the reaction to that may be going very hard in other directions that might impact 1:53:49.280 --> 1:53:50.840 you that even other industries. 1:53:50.840 --> 1:53:55.320 And that's what we call in the air winter is when there is backlash where no one believes 1:53:55.320 --> 1:54:00.600 any of these promises anymore because they've turned out to be big lies the first time around. 1:54:00.600 --> 1:54:06.120 And this will definitely happen to some extent for autonomous vehicles because the public 1:54:06.120 --> 1:54:13.440 and decision makers have been convinced that around 2015, they've been convinced by these 1:54:13.440 --> 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 --> 1:54:23.120 in maybe 2016, maybe 2017, May 2018. 1:54:23.120 --> 1:54:28.040 Now in 2019, we're still waiting for it. 1:54:28.040 --> 1:54:32.880 And so I don't believe we are going to have a full on AI winter because we have these 1:54:32.880 --> 1:54:39.480 technologies that are producing a tremendous amount of real value, but there is also too 1:54:39.480 --> 1:54:40.480 much hype. 1:54:40.480 --> 1:54:45.240 So there will be some backlash, especially there will be backlash. 1:54:45.240 --> 1:54:53.080 So some startups are trying to sell the dream of AGI and the fact that AGI is going to create 1:54:53.080 --> 1:54:54.080 infinite value. 1:54:54.080 --> 1:55:01.240 AGI is like a freelance, like if you can develop an AI system that passes a certain threshold 1:55:01.240 --> 1:55:06.440 of IQ or something, then suddenly you have infinite value. 1:55:06.440 --> 1:55:11.640 And well, there are actually lots of investors buying into this idea. 1:55:11.640 --> 1:55:18.920 And they will wait maybe 10, 15 years and nothing will happen. 1:55:18.920 --> 1:55:22.800 And the next time around, well, maybe there will be a new generation of investors, no 1:55:22.800 --> 1:55:24.040 one will care. 1:55:24.040 --> 1:55:27.160 Human memory is very short after all. 1:55:27.160 --> 1:55:34.440 I don't know about you, but because I've spoken about AGI sometimes poetically, I get a lot 1:55:34.440 --> 1:55:42.360 of emails from people giving me, they're usually like a large manifestos of, they say to me 1:55:42.360 --> 1:55:48.320 that they have created an AGI system or they know how to do it and there's a long write 1:55:48.320 --> 1:55:49.320 up of how to do it. 1:55:49.320 --> 1:55:51.400 I get a lot of these emails. 1:55:51.400 --> 1:55:57.840 They're a little bit feel like it's generated by an AI system actually, but there's usually 1:55:57.840 --> 1:55:58.840 no backup. 1:55:58.840 --> 1:56:04.920 Maybe that's recursively self improving AI, it's you have a transformer generating crank 1:56:04.920 --> 1:56:06.880 papers about a GI. 1:56:06.880 --> 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 --> 1:56:16.960 papers, how do we know they're not onto something? 1:56:16.960 --> 1:56:24.320 How do I, so when you start to talk about AGI or anything like the reasoning benchmarks 1:56:24.320 --> 1:56:28.720 and so on, so something that doesn't have a benchmark, it's really difficult to know. 1:56:28.720 --> 1:56:35.480 I mean, I talked to Jeff Hawkins who's really looking at neuroscience approaches to how, 1:56:35.480 --> 1:56:41.800 and there's some, there's echoes of really interesting ideas in at least Jeff's case, 1:56:41.800 --> 1:56:43.520 which he's showing. 1:56:43.520 --> 1:56:45.840 How do you usually think about this? 1:56:45.840 --> 1:56:52.920 Like preventing yourself from being too narrow minded and elitist about deep learning. 1:56:52.920 --> 1:56:57.040 It has to work on these particular benchmarks, otherwise it's trash. 1:56:57.040 --> 1:57:05.880 Well, the thing is intelligence does not exist in the abstract. 1:57:05.880 --> 1:57:07.440 Intelligence has to be applied. 1:57:07.440 --> 1:57:11.040 So if you don't have a benchmark, if you don't have an improvement on some benchmark, maybe 1:57:11.040 --> 1:57:12.680 it's a new benchmark. 1:57:12.680 --> 1:57:16.760 Maybe it's not something we've been looking at before, but you do need a problem that 1:57:16.760 --> 1:57:17.760 you're trying to solve. 1:57:17.760 --> 1:57:21.040 You're not going to come up with a solution without a problem. 1:57:21.040 --> 1:57:26.760 So you, general intelligence, I mean, you've clearly highlighted generalization. 1:57:26.760 --> 1:57:31.320 If you want to claim that you have an intelligence system, it should come with a benchmark. 1:57:31.320 --> 1:57:35.960 It should, yes, it should display capabilities of some kind. 1:57:35.960 --> 1:57:41.920 It should show that it can create some form of value, even if it's a very artificial form 1:57:41.920 --> 1:57:43.160 of value. 1:57:43.160 --> 1:57:48.840 And that's also the reason why you don't actually need to care about telling which papers have 1:57:48.840 --> 1:57:53.520 actually some hidden potential and which do not. 1:57:53.520 --> 1:57:58.880 Because if there is a new technique, it's actually creating value. 1:57:58.880 --> 1:58:02.640 This is going to be brought to light very quickly because it's actually making a difference. 1:58:02.640 --> 1:58:08.240 So it's a difference between something that is ineffectual and something that is actually 1:58:08.240 --> 1:58:09.240 useful. 1:58:09.240 --> 1:58:14.120 And ultimately, usefulness is our guide, not just in this field, but if you look at science 1:58:14.120 --> 1:58:19.560 in general, maybe there are many, many people over the years that have had some really interesting 1:58:19.560 --> 1:58:23.120 theories of everything, but they were just completely useless. 1:58:23.120 --> 1:58:28.240 And you don't actually need to tell the interesting theories from the useless theories. 1:58:28.240 --> 1:58:34.120 All you need is to see, you know, is this actually having an effect on something else? 1:58:34.120 --> 1:58:35.600 You know, is this actually useful? 1:58:35.600 --> 1:58:37.960 Is this making an impact or not? 1:58:37.960 --> 1:58:42.480 As beautifully put, I mean, the same applies to quantum mechanics, to string theory, to 1:58:42.480 --> 1:58:43.480 the holographic principle. 1:58:43.480 --> 1:58:46.080 We are doing deep learning because it works. 1:58:46.080 --> 1:58:52.720 Before it started working, people considered people working on neural networks as cranks 1:58:52.720 --> 1:58:53.720 very much. 1:58:53.720 --> 1:58:56.560 Like, you know, no one was working on this anymore. 1:58:56.560 --> 1:58:59.400 And now it's working, which is what makes it valuable. 1:58:59.400 --> 1:59:02.840 It's not about being right, it's about being effective. 1:59:02.840 --> 1:59:08.120 And nevertheless, the individual entities of this scientific mechanism, just like Yoshio 1:59:08.120 --> 1:59:13.160 Banjo or Yanlacun, they, while being called cranks, stuck with it, right? 1:59:13.160 --> 1:59:19.080 And so, us individual agents, even if everyone's laughing at us, should stick with it. 1:59:19.080 --> 1:59:23.840 If you believe you have something, you should stick with it and see it through. 1:59:23.840 --> 1:59:25.920 That's a beautiful, inspirational message to end on. 1:59:25.920 --> 1:59:27.800 Francois, thank you so much for talking today. 1:59:27.800 --> 1:59:28.800 That was amazing. 1:59:28.800 --> 1:59:35.800 Thank you.