WEBVTT 00:00.000 --> 00:03.520 The following is a conversation with Jürgen Schmidhuber. 00:03.520 --> 00:06.360 He's the codirector of its CS Swiss AI lab 00:06.360 --> 00:10.360 and a cocreator of long short term memory networks. 00:10.360 --> 00:13.720 LSDMs are used in billions of devices today 00:13.720 --> 00:17.400 for speech recognition, translation, and much more. 00:17.400 --> 00:20.800 Over 30 years, he has proposed a lot of interesting 00:20.800 --> 00:24.800 out of the box ideas, a meta learning, adversarial networks, 00:24.800 --> 00:28.720 computer vision, and even a formal theory of quote, 00:28.720 --> 00:32.360 creativity, curiosity, and fun. 00:32.360 --> 00:34.920 This conversation is part of the MIT course 00:34.920 --> 00:36.520 on artificial general intelligence 00:36.520 --> 00:38.840 and the artificial intelligence podcast. 00:38.840 --> 00:41.960 If you enjoy it, subscribe on YouTube, iTunes, 00:41.960 --> 00:43.960 or simply connect with me on Twitter 00:43.960 --> 00:47.280 at Lex Friedman spelled F R I D. 00:47.280 --> 00:51.480 And now here's my conversation with Jürgen Schmidhuber. 00:53.080 --> 00:55.640 Early on you dreamed of AI systems 00:55.640 --> 00:58.680 that self improve recursively. 00:58.680 --> 01:01.440 When was that dream born? 01:01.440 --> 01:02.840 When I was a baby. 01:02.840 --> 01:04.000 No, that's not true. 01:04.000 --> 01:06.200 When I was a teenager. 01:06.200 --> 01:09.400 And what was the catalyst for that birth? 01:09.400 --> 01:12.800 What was the thing that first inspired you? 01:12.800 --> 01:15.000 When I was a boy, I... 01:17.400 --> 01:19.880 I was thinking about what to do in my life 01:19.880 --> 01:23.560 and then I thought the most exciting thing 01:23.560 --> 01:27.160 is to solve the riddles of the universe. 01:27.160 --> 01:30.720 And that means you have to become a physicist. 01:30.720 --> 01:35.640 However, then I realized that there's something even grander. 01:35.640 --> 01:39.680 You can try to build a machine. 01:39.680 --> 01:41.920 That isn't really a machine any longer. 01:41.920 --> 01:44.280 That learns to become a much better physicist 01:44.280 --> 01:46.840 than I could ever hope to be. 01:46.840 --> 01:50.120 And that's how I thought maybe I can multiply 01:50.120 --> 01:54.320 my tiny little bit of creativity into infinity. 01:54.320 --> 01:57.160 But ultimately, that creativity will be multiplied 01:57.160 --> 01:59.160 to understand the universe around us. 01:59.160 --> 02:05.640 That's the curiosity for that mystery that drove you. 02:05.640 --> 02:08.320 Yes, so if you can build a machine 02:08.320 --> 02:13.760 that learns to solve more and more complex problems 02:13.760 --> 02:16.720 and more and more general problems over, 02:16.720 --> 02:22.520 then you basically have solved all the problems. 02:22.520 --> 02:25.960 At least all the solvable problems. 02:25.960 --> 02:27.080 So how do you think... 02:27.080 --> 02:31.440 What is the mechanism for that kind of general solver look like? 02:31.440 --> 02:35.480 Obviously, we don't quite yet have one or know 02:35.480 --> 02:37.040 how to build one boy of ideas 02:37.040 --> 02:40.800 and you have had throughout your career several ideas about it. 02:40.800 --> 02:43.600 So how do you think about that mechanism? 02:43.600 --> 02:48.640 So in the 80s, I thought about how to build this machine 02:48.640 --> 02:51.000 that learns to solve all these problems 02:51.000 --> 02:54.120 that I cannot solve myself. 02:54.120 --> 02:57.120 And I thought it is clear, it has to be a machine 02:57.120 --> 03:00.880 that not only learns to solve this problem here 03:00.880 --> 03:02.640 and this problem here, 03:02.640 --> 03:06.240 but it also has to learn to improve 03:06.240 --> 03:09.360 the learning algorithm itself. 03:09.360 --> 03:12.480 So it has to have the learning algorithm 03:12.480 --> 03:15.720 in a representation that allows it to inspect it 03:15.720 --> 03:19.240 and modify it so that it can come up 03:19.240 --> 03:22.080 with a better learning algorithm. 03:22.080 --> 03:25.680 So I called that meta learning, learning to learn 03:25.680 --> 03:28.040 and recursive self improvement. 03:28.040 --> 03:29.840 That is really the pinnacle of that, 03:29.840 --> 03:35.960 where you then not only learn how to improve 03:35.960 --> 03:37.480 on that problem and on that, 03:37.480 --> 03:41.080 but you also improve the way the machine improves 03:41.080 --> 03:43.480 and you also improve the way it improves the way 03:43.480 --> 03:45.720 it improves itself. 03:45.720 --> 03:48.560 And that was my 1987 diploma thesis, 03:48.560 --> 03:53.200 which was all about that hierarchy of meta learners 03:53.200 --> 03:57.240 that have no computational limits 03:57.240 --> 03:59.920 except for the well known limits 03:59.920 --> 04:03.160 that Gödel identified in 1931 04:03.160 --> 04:05.640 and for the limits of physics. 04:06.480 --> 04:10.040 In the recent years, meta learning has gained popularity 04:10.040 --> 04:12.760 in a specific kind of form. 04:12.760 --> 04:16.000 You've talked about how that's not really meta learning 04:16.000 --> 04:21.000 with neural networks, that's more basic transfer learning. 04:21.480 --> 04:22.720 Can you talk about the difference 04:22.720 --> 04:25.440 between the big general meta learning 04:25.440 --> 04:27.960 and a more narrow sense of meta learning 04:27.960 --> 04:30.880 the way it's used today, the way it's talked about today? 04:30.880 --> 04:33.440 Let's take the example of a deep neural network 04:33.440 --> 04:37.240 that has learned to classify images. 04:37.240 --> 04:40.080 And maybe you have trained that network 04:40.080 --> 04:43.800 on 100 different databases of images. 04:43.800 --> 04:48.120 And now a new database comes along 04:48.120 --> 04:52.000 and you want to quickly learn the new thing as well. 04:53.400 --> 04:57.720 So one simple way of doing that is you take the network 04:57.720 --> 05:02.440 which already knows 100 types of databases 05:02.440 --> 05:06.320 and then you just take the top layer of that 05:06.320 --> 05:11.320 and you retrain that using the new labeled data 05:11.320 --> 05:14.720 that you have in the new image database. 05:14.720 --> 05:17.320 And then it turns out that it really, really quickly 05:17.320 --> 05:20.560 can learn that too, one shot basically, 05:20.560 --> 05:24.280 because from the first 100 data sets, 05:24.280 --> 05:27.520 it already has learned so much about computer vision 05:27.520 --> 05:31.840 that it can reuse that and that is then almost good enough 05:31.840 --> 05:34.240 to solve the new tasks except you need a little bit 05:34.240 --> 05:37.040 of adjustment on the top. 05:37.040 --> 05:40.200 So that is transfer learning 05:40.200 --> 05:43.480 and it has been done in principle for many decades. 05:43.480 --> 05:45.680 People have done similar things for decades. 05:47.360 --> 05:49.880 Meta learning, true meta learning is about 05:49.880 --> 05:53.880 having the learning algorithm itself 05:54.440 --> 05:59.440 open to introspection by the system that is using it 06:00.440 --> 06:03.760 and also open to modification 06:03.760 --> 06:07.200 such that the learning system has an opportunity 06:07.200 --> 06:11.400 to modify any part of the learning algorithm 06:11.400 --> 06:16.000 and then evaluate the consequences of that modification 06:16.000 --> 06:21.000 and then learn from that to create a better learning algorithm 06:22.000 --> 06:23.960 and so on recursively. 06:24.960 --> 06:27.680 So that's a very different animal 06:27.680 --> 06:32.680 where you are opening the space of possible learning algorithms 06:32.680 --> 06:35.520 to the learning system itself. 06:35.520 --> 06:39.040 Right, so you've like in the 2004 paper, 06:39.040 --> 06:42.440 you describe Gatal machines and programs 06:42.440 --> 06:44.520 that rewrite themselves, right? 06:44.520 --> 06:47.520 Philosophically and even in your paper mathematically, 06:47.520 --> 06:50.000 these are really compelling ideas, 06:50.000 --> 06:55.000 but practically, do you see these self referential programs 06:55.320 --> 06:59.400 being successful in the near term to having an impact 06:59.400 --> 07:03.040 where sort of it demonstrates to the world 07:03.040 --> 07:08.040 that this direction is a good one to pursue in the near term? 07:08.680 --> 07:11.400 Yes, we had these two different types 07:11.400 --> 07:13.440 of fundamental research, 07:13.440 --> 07:15.840 how to build a universal problem solver, 07:15.840 --> 07:19.840 one basically exploiting proof search 07:23.000 --> 07:24.960 and things like that that you need to come up 07:24.960 --> 07:29.960 with asymptotically optimal, theoretically optimal 07:30.320 --> 07:33.200 self improvers and problem solvers. 07:34.200 --> 07:39.200 However, one has to admit that through this proof search 07:40.640 --> 07:43.640 comes in an additive constant, 07:43.640 --> 07:46.800 an overhead, an additive overhead 07:46.800 --> 07:51.800 that vanishes in comparison to what you have to do 07:51.800 --> 07:53.960 to solve large problems. 07:53.960 --> 07:56.920 However, for many of the small problems 07:56.920 --> 07:59.920 that we want to solve in our everyday life, 07:59.920 --> 08:02.440 we cannot ignore this constant overhead. 08:02.440 --> 08:07.440 And that's why we also have been doing other things, 08:07.440 --> 08:11.160 non universal things such as recurrent neural networks 08:11.160 --> 08:14.360 which are trained by gradient descent 08:14.360 --> 08:17.600 and local search techniques which aren't universal at all, 08:17.600 --> 08:20.280 which aren't provably optimal at all 08:20.280 --> 08:22.000 like the other stuff that we did, 08:22.000 --> 08:24.600 but which are much more practical 08:24.600 --> 08:27.920 as long as we only want to solve the small problems 08:27.920 --> 08:32.920 that we are typically trying to solve in this environment here. 08:34.680 --> 08:38.200 So the universal problem solvers like the Gödel machine 08:38.200 --> 08:41.320 but also Markus Hutter's fastest way 08:41.320 --> 08:43.560 of solving all possible problems, 08:43.560 --> 08:47.360 which he developed around 2002 in my lab, 08:47.360 --> 08:51.280 they are associated with these constant overheads 08:51.280 --> 08:53.280 for proof search, which guarantees 08:53.280 --> 08:55.480 that the thing that you're doing is optimal. 08:55.480 --> 08:59.880 For example, there is this fastest way 08:59.880 --> 09:03.880 of solving all problems with a computable solution 09:03.880 --> 09:05.880 which is due to Markus Hutter. 09:05.880 --> 09:10.880 And to explain what's going on there, 09:10.880 --> 09:13.040 let's take traveling salesman problems. 09:14.240 --> 09:16.160 With traveling salesman problems, 09:16.160 --> 09:20.080 you have a number of cities, N cities, 09:20.080 --> 09:22.480 and you try to find the shortest path 09:22.480 --> 09:26.480 through all these cities without visiting any city twice. 09:28.480 --> 09:31.040 And nobody knows the fastest way 09:31.040 --> 09:35.040 of solving traveling salesman problems, TSPs, 09:37.520 --> 09:40.480 but let's assume there is a method of solving them 09:40.480 --> 09:44.480 within N to the five operations 09:44.480 --> 09:48.560 where N is the number of cities. 09:50.160 --> 09:54.560 Then the universal method of Markus 09:54.560 --> 09:58.560 is going to solve the same traveling salesman problem 09:58.560 --> 10:02.080 also within N to the five steps, 10:02.080 --> 10:06.360 plus O of one, plus a constant number of steps 10:06.360 --> 10:09.240 that you need for the proof searcher, 10:09.240 --> 10:13.800 which you need to show that this particular 10:13.800 --> 10:17.240 class of problems that traveling salesman problems 10:17.240 --> 10:19.360 can be solved within a certain time bound, 10:20.520 --> 10:24.400 within order N to the five steps, basically. 10:24.400 --> 10:28.520 And this additive constant doesn't care for N, 10:28.520 --> 10:32.400 which means as N is getting larger and larger, 10:32.400 --> 10:34.880 as you have more and more cities, 10:34.880 --> 10:38.600 the constant overhead pales and comparison. 10:38.600 --> 10:44.120 And that means that almost all large problems are solved 10:44.120 --> 10:46.520 in the best possible way already today. 10:46.520 --> 10:50.480 We already have a universal problem solved like that. 10:50.480 --> 10:54.520 However, it's not practical because the overhead, 10:54.520 --> 10:57.440 the constant overhead is so large 10:57.440 --> 11:00.200 that for the small kinds of problems 11:00.200 --> 11:04.560 that we want to solve in this little biosphere. 11:04.560 --> 11:06.360 By the way, when you say small, 11:06.360 --> 11:08.600 you're talking about things that fall 11:08.600 --> 11:10.880 within the constraints of our computational systems. 11:10.880 --> 11:14.280 So they can seem quite large to us mere humans. 11:14.280 --> 11:15.360 That's right, yeah. 11:15.360 --> 11:19.000 So they seem large and even unsolvable 11:19.000 --> 11:21.000 in a practical sense today, 11:21.000 --> 11:24.760 but they are still small compared to almost all problems 11:24.760 --> 11:28.480 because almost all problems are large problems, 11:28.480 --> 11:30.840 which are much larger than any constant. 11:31.920 --> 11:34.520 Do you find it useful as a person 11:34.520 --> 11:38.680 who is dreamed of creating a general learning system, 11:38.680 --> 11:39.880 has worked on creating one, 11:39.880 --> 11:42.160 has done a lot of interesting ideas there 11:42.160 --> 11:46.360 to think about P versus NP, 11:46.360 --> 11:50.800 this formalization of how hard problems are, 11:50.800 --> 11:52.360 how they scale, 11:52.360 --> 11:55.200 this kind of worst case analysis type of thinking. 11:55.200 --> 11:56.840 Do you find that useful? 11:56.840 --> 11:59.720 Or is it only just a mathematical, 12:00.560 --> 12:02.640 it's a set of mathematical techniques 12:02.640 --> 12:05.760 to give you intuition about what's good and bad? 12:05.760 --> 12:09.440 So P versus NP, that's super interesting 12:09.440 --> 12:11.800 from a theoretical point of view. 12:11.800 --> 12:14.560 And in fact, as you are thinking about that problem, 12:14.560 --> 12:17.280 you can also get inspiration 12:17.280 --> 12:21.280 for better practical problem solvers. 12:21.280 --> 12:23.320 On the other hand, we have to admit 12:23.320 --> 12:24.560 that at the moment, 12:24.560 --> 12:28.360 the best practical problem solvers 12:28.360 --> 12:30.120 for all kinds of problems 12:30.120 --> 12:33.880 that we are now solving through what is called AI at the moment, 12:33.880 --> 12:36.240 they are not of the kind 12:36.240 --> 12:38.800 that is inspired by these questions. 12:38.800 --> 12:42.680 There we are using general purpose computers, 12:42.680 --> 12:44.840 such as recurrent neural networks, 12:44.840 --> 12:46.680 but we have a search technique, 12:46.680 --> 12:50.320 which is just local search gradient descent 12:50.320 --> 12:51.960 to try to find a program 12:51.960 --> 12:54.400 that is running on these recurrent networks, 12:54.400 --> 12:58.160 such that it can solve some interesting problems, 12:58.160 --> 13:01.920 such as speech recognition or machine translation 13:01.920 --> 13:03.200 and something like that. 13:03.200 --> 13:06.480 And there is very little theory 13:06.480 --> 13:09.720 behind the best solutions that we have at the moment 13:09.720 --> 13:10.800 that can do that. 13:10.800 --> 13:12.640 Do you think that needs to change? 13:12.640 --> 13:15.120 Do you think that will change or can we go, 13:15.120 --> 13:17.120 can we create a general intelligence systems 13:17.120 --> 13:19.200 without ever really proving 13:19.200 --> 13:20.600 that that system is intelligent 13:20.600 --> 13:22.560 in some kind of mathematical way, 13:22.560 --> 13:24.960 solving machine translation perfectly 13:24.960 --> 13:26.320 or something like that, 13:26.320 --> 13:29.160 within some kind of syntactic definition of a language? 13:29.160 --> 13:31.120 Or can we just be super impressed 13:31.120 --> 13:35.080 by the thing working extremely well and that's sufficient? 13:35.080 --> 13:36.720 There's an old saying, 13:36.720 --> 13:39.360 and I don't know who brought it up first, 13:39.360 --> 13:42.440 which says there's nothing more practical 13:42.440 --> 13:43.680 than a good theory. 13:43.680 --> 13:48.680 And a good theory of problem solving 13:52.760 --> 13:55.560 under limited resources like here in this universe 13:55.560 --> 13:57.000 or on this little planet 13:58.480 --> 14:01.800 has to take into account these limited resources. 14:01.800 --> 14:06.800 And so probably there is locking a theory 14:08.040 --> 14:10.800 which is related to what we already have, 14:10.800 --> 14:14.440 these asymptotically optimal problem solvers, 14:14.440 --> 14:18.560 which tells us what we need in addition to that 14:18.560 --> 14:21.760 to come up with a practically optimal problem solver. 14:21.760 --> 14:26.760 So I believe we will have something like that 14:27.080 --> 14:29.720 and maybe just a few little tiny twists 14:29.720 --> 14:34.320 are necessary to change what we already have 14:34.320 --> 14:36.360 to come up with that as well. 14:36.360 --> 14:37.800 As long as we don't have that, 14:37.800 --> 14:42.600 we admit that we are taking suboptimal ways 14:42.600 --> 14:46.040 and recurrent neural networks and long short term memory 14:46.040 --> 14:50.440 for equipped with local search techniques 14:50.440 --> 14:53.560 and we are happy that it works better 14:53.560 --> 14:55.480 than any competing methods, 14:55.480 --> 15:00.480 but that doesn't mean that we think we are done. 15:00.800 --> 15:05.040 You've said that an AGI system will ultimately be a simple one, 15:05.040 --> 15:08.000 a general intelligence system will ultimately be a simple one, 15:08.000 --> 15:10.240 maybe a pseudo code of a few lines 15:10.240 --> 15:11.840 will be able to describe it. 15:11.840 --> 15:16.760 Can you talk through your intuition behind this idea, 15:16.760 --> 15:21.760 why you feel that at its core intelligence 15:22.120 --> 15:25.560 is a simple algorithm? 15:26.920 --> 15:31.680 Experience tells us that the stuff that works best 15:31.680 --> 15:33.120 is really simple. 15:33.120 --> 15:37.640 So the asymptotically optimal ways of solving problems, 15:37.640 --> 15:38.800 if you look at them, 15:38.800 --> 15:41.800 they're just a few lines of code, it's really true. 15:41.800 --> 15:44.000 Although they are these amazing properties, 15:44.000 --> 15:45.760 just a few lines of code, 15:45.760 --> 15:50.760 then the most promising and most useful practical things 15:53.760 --> 15:57.760 maybe don't have this proof of optimality associated with them. 15:57.760 --> 16:00.840 However, they are also just a few lines of code. 16:00.840 --> 16:05.040 The most successful recurrent neural networks, 16:05.040 --> 16:08.360 you can write them down and five lines of pseudo code. 16:08.360 --> 16:10.920 That's a beautiful, almost poetic idea, 16:10.920 --> 16:15.600 but what you're describing there 16:15.600 --> 16:17.400 is the lines of pseudo code 16:17.400 --> 16:20.600 are sitting on top of layers and layers of abstractions, 16:20.600 --> 16:22.240 in a sense. 16:22.240 --> 16:25.040 So you're saying at the very top, 16:25.040 --> 16:30.040 it'll be a beautifully written sort of algorithm, 16:31.120 --> 16:33.960 but do you think that there's many layers of abstractions 16:33.960 --> 16:36.880 we have to first learn to construct? 16:36.880 --> 16:38.280 Yeah, of course. 16:38.280 --> 16:42.640 We are building on all these great abstractions 16:42.640 --> 16:46.040 that people have invented over the millennia, 16:46.040 --> 16:51.040 such as matrix multiplications and drill numbers 16:51.600 --> 16:56.600 and basic arithmetics and calculus and derivations 16:58.720 --> 17:03.320 of error functions and derivatives of error functions 17:03.320 --> 17:04.320 and stuff like that. 17:05.440 --> 17:10.440 So without that language that greatly simplifies 17:10.440 --> 17:13.880 our way of thinking about these problems, 17:13.880 --> 17:14.840 we couldn't do anything. 17:14.840 --> 17:16.560 So in that sense, as always, 17:16.560 --> 17:19.600 we are standing on the shoulders of the giants 17:19.600 --> 17:24.600 who in the past simplified the problem of problem solving 17:25.520 --> 17:30.000 so much that now we have a chance to do the final step. 17:30.000 --> 17:32.120 So the final step will be a simple one. 17:34.000 --> 17:36.760 If we take a step back through all of human civilization 17:36.760 --> 17:38.360 and just the universe in general, 17:38.360 --> 17:41.440 how do you think about evolution? 17:41.440 --> 17:45.400 And what if creating a universe is required 17:45.400 --> 17:47.320 to achieve this final step? 17:47.320 --> 17:50.920 What if going through the very painful 17:50.920 --> 17:53.840 and inefficient process of evolution is needed 17:53.840 --> 17:55.880 to come up with this set of abstractions 17:55.880 --> 17:57.800 that ultimately lead to intelligence? 17:57.800 --> 18:00.800 Do you think there's a shortcut 18:00.800 --> 18:04.640 or do you think we have to create something like our universe 18:04.640 --> 18:09.480 in order to create something like human level intelligence? 18:09.480 --> 18:13.160 So far, the only example we have is this one, 18:13.160 --> 18:15.160 this universe in which we are living. 18:15.160 --> 18:16.360 You think you can do better? 18:20.880 --> 18:25.000 Maybe not, but we are part of this whole process. 18:25.000 --> 18:30.000 So apparently, so it might be the case 18:30.000 --> 18:32.160 that the code that runs the universe 18:32.160 --> 18:33.720 is really, really simple. 18:33.720 --> 18:36.640 Everything points to that possibility 18:36.640 --> 18:39.960 because gravity and other basic forces 18:39.960 --> 18:44.120 are really simple laws that can be easily described, 18:44.120 --> 18:47.080 also in just a few lines of code, basically. 18:47.080 --> 18:52.080 And then there are these other events 18:52.200 --> 18:55.080 that the apparently random events 18:55.080 --> 18:56.560 in the history of the universe, 18:56.560 --> 18:58.800 which as far as we know at the moment 18:58.800 --> 19:00.720 don't have a compact code, 19:00.720 --> 19:03.240 but who knows, maybe somebody in the near future 19:03.240 --> 19:06.800 is going to figure out the pseudo random generator, 19:06.800 --> 19:11.800 which is computing whether the measurement of that 19:13.520 --> 19:15.920 spin up or down thing here 19:15.920 --> 19:18.440 is going to be positive or negative. 19:18.440 --> 19:19.880 Underline quantum mechanics. 19:19.880 --> 19:20.720 Yes, so. 19:20.720 --> 19:23.160 Do you ultimately think quantum mechanics 19:23.160 --> 19:25.200 is a pseudo random number generator? 19:25.200 --> 19:26.920 So it's all deterministic. 19:26.920 --> 19:28.760 There's no randomness in our universe. 19:30.400 --> 19:31.800 Does God play dice? 19:31.800 --> 19:34.080 So a couple of years ago, 19:34.080 --> 19:39.080 a famous physicist, quantum physicist, Anton Zeilinger, 19:39.080 --> 19:41.600 he wrote an essay in Nature, 19:41.600 --> 19:44.280 and it started more or less like that. 19:46.720 --> 19:51.720 One of the fundamental insights of the 20th century 19:53.280 --> 19:58.280 was that the universe is fundamentally random 19:58.280 --> 20:02.760 on the quantum level, and that whenever 20:03.760 --> 20:06.720 you measure spin up or down or something like that, 20:06.720 --> 20:10.720 a new bit of information enters the history of the universe. 20:13.440 --> 20:14.680 And while I was reading that, 20:14.680 --> 20:18.000 I was already typing the response 20:18.000 --> 20:20.280 and they had to publish it because I was right, 20:21.560 --> 20:25.560 that there is no evidence, no physical evidence for that. 20:25.560 --> 20:28.440 So there's an alternative explanation 20:28.440 --> 20:31.240 where everything that we consider random 20:31.240 --> 20:33.800 is actually pseudo random, 20:33.800 --> 20:38.800 such as the decimal expansion of pi, 3.141 and so on, 20:39.400 --> 20:42.120 which looks random, but isn't. 20:42.120 --> 20:47.120 So pi is interesting because every three digit sequence, 20:47.720 --> 20:51.720 every sequence of three digits appears roughly 20:51.720 --> 20:56.720 one in a thousand times, and every five digit sequence 20:57.360 --> 21:00.760 appears roughly one in 10,000 times. 21:00.760 --> 21:02.760 What do you expect? 21:02.760 --> 21:06.760 If it was random, but there's a very short algorithm, 21:06.760 --> 21:09.120 a short program that computes all of that. 21:09.120 --> 21:11.200 So it's extremely compressible. 21:11.200 --> 21:13.120 And who knows, maybe tomorrow somebody, 21:13.120 --> 21:15.360 some grad student at CERN goes back 21:15.360 --> 21:19.120 over all these data points, better decay, 21:19.120 --> 21:21.760 and whatever, and figures out, oh, 21:21.760 --> 21:25.760 it's the second billion digits of pi or something like that. 21:25.760 --> 21:28.840 We don't have any fundamental reason at the moment 21:28.840 --> 21:33.600 to believe that this is truly random 21:33.600 --> 21:36.440 and not just a deterministic video game. 21:36.440 --> 21:38.680 If it was a deterministic video game, 21:38.680 --> 21:40.360 it would be much more beautiful 21:40.360 --> 21:44.160 because beauty is simplicity. 21:44.160 --> 21:47.560 And many of the basic laws of the universe 21:47.560 --> 21:51.560 like gravity and the other basic forces are very simple. 21:51.560 --> 21:55.560 So very short programs can explain what these are doing. 21:56.560 --> 22:00.560 And it would be awful and ugly. 22:00.560 --> 22:01.560 The universe would be ugly. 22:01.560 --> 22:03.560 The history of the universe would be ugly 22:03.560 --> 22:06.560 if for the extra things, the random, 22:06.560 --> 22:10.560 the seemingly random data points that we get all the time 22:10.560 --> 22:15.560 that we really need a huge number of extra bits 22:15.560 --> 22:21.560 to describe all these extra bits of information. 22:22.560 --> 22:25.560 So as long as we don't have evidence 22:25.560 --> 22:27.560 that there is no short program 22:27.560 --> 22:32.560 that computes the entire history of the entire universe, 22:32.560 --> 22:38.560 we are, as scientists, compelled to look further 22:38.560 --> 22:41.560 for that shortest program. 22:41.560 --> 22:46.560 Your intuition says there exists a program 22:46.560 --> 22:50.560 that can backtrack to the creation of the universe. 22:50.560 --> 22:53.560 So it can take the shortest path to the creation of the universe. 22:53.560 --> 22:57.560 Yes, including all the entanglement things 22:57.560 --> 23:01.560 and all the spin up and down measurements 23:01.560 --> 23:09.560 that have been taken place since 13.8 billion years ago. 23:09.560 --> 23:14.560 So we don't have a proof that it is random. 23:14.560 --> 23:19.560 We don't have a proof that it is compressible to a short program. 23:19.560 --> 23:21.560 But as long as we don't have that proof, 23:21.560 --> 23:24.560 we are obliged as scientists to keep looking 23:24.560 --> 23:26.560 for that simple explanation. 23:26.560 --> 23:27.560 Absolutely. 23:27.560 --> 23:30.560 So you said simplicity is beautiful or beauty is simple. 23:30.560 --> 23:32.560 Either one works. 23:32.560 --> 23:36.560 But you also work on curiosity, discovery. 23:36.560 --> 23:42.560 The romantic notion of randomness, of serendipity, 23:42.560 --> 23:49.560 of being surprised by things that are about you, 23:49.560 --> 23:53.560 kind of in our poetic notion of reality, 23:53.560 --> 23:56.560 we think as humans require randomness. 23:56.560 --> 23:58.560 So you don't find randomness beautiful. 23:58.560 --> 24:04.560 You find simple determinism beautiful. 24:04.560 --> 24:06.560 Yeah. 24:06.560 --> 24:07.560 Okay. 24:07.560 --> 24:08.560 So why? 24:08.560 --> 24:09.560 Why? 24:09.560 --> 24:12.560 Because the explanation becomes shorter. 24:12.560 --> 24:19.560 A universe that is compressible to a short program 24:19.560 --> 24:22.560 is much more elegant and much more beautiful 24:22.560 --> 24:24.560 than another one, 24:24.560 --> 24:28.560 which needs an almost infinite number of bits to be described. 24:28.560 --> 24:31.560 As far as we know, 24:31.560 --> 24:34.560 many things that are happening in this universe are really simple 24:34.560 --> 24:38.560 in terms of short programs that compute gravity 24:38.560 --> 24:43.560 and the interaction between elementary particles and so on. 24:43.560 --> 24:45.560 So all of that seems to be very, very simple. 24:45.560 --> 24:50.560 Every electron seems to reuse the same subprogram all the time 24:50.560 --> 24:57.560 as it is interacting with other elementary particles. 24:57.560 --> 25:04.560 If we now require an extra oracle 25:04.560 --> 25:07.560 injecting new bits of information all the time 25:07.560 --> 25:11.560 for these extra things which are currently not understood, 25:11.560 --> 25:18.560 such as better decay, 25:18.560 --> 25:25.560 then the whole description length of the data that we can observe 25:25.560 --> 25:31.560 of the history of the universe would become much longer. 25:31.560 --> 25:33.560 And therefore, uglier. 25:33.560 --> 25:34.560 And uglier. 25:34.560 --> 25:38.560 Again, the simplicity is elegant and beautiful. 25:38.560 --> 25:42.560 All the history of science is a history of compression progress. 25:42.560 --> 25:43.560 Yeah. 25:43.560 --> 25:48.560 So you've described sort of as we build up abstractions 25:48.560 --> 25:52.560 and you've talked about the idea of compression. 25:52.560 --> 25:55.560 How do you see this, the history of science, 25:55.560 --> 25:59.560 the history of humanity, our civilization and life on Earth 25:59.560 --> 26:03.560 as some kind of path towards greater and greater compression? 26:03.560 --> 26:04.560 What do you mean by that? 26:04.560 --> 26:06.560 How do you think about that? 26:06.560 --> 26:12.560 Indeed, the history of science is a history of compression progress. 26:12.560 --> 26:14.560 What does that mean? 26:14.560 --> 26:17.560 Hundreds of years ago, there was an astronomer 26:17.560 --> 26:19.560 whose name was Kepler. 26:19.560 --> 26:25.560 And he looked at the data points that he got by watching planets move. 26:25.560 --> 26:28.560 And then he had all these data points and suddenly it turned out 26:28.560 --> 26:37.560 that he can greatly compress the data by predicting it through an ellipse law. 26:37.560 --> 26:44.560 So it turns out that all these data points are more or less on ellipses around the sun. 26:44.560 --> 26:50.560 And another guy came along whose name was Newton and before him Hook. 26:50.560 --> 26:57.560 And they said the same thing that is making these planets move like that 26:57.560 --> 27:01.560 is what makes the apples fall down. 27:01.560 --> 27:10.560 And it also holds for stones and for all kinds of other objects. 27:10.560 --> 27:16.560 And suddenly many, many of these observations became much more compressible 27:16.560 --> 27:19.560 because as long as you can predict the next thing, 27:19.560 --> 27:22.560 given what you have seen so far, you can compress it. 27:22.560 --> 27:24.560 But you don't have to store that data extra. 27:24.560 --> 27:28.560 This is called predictive coding. 27:28.560 --> 27:33.560 And then there was still something wrong with that theory of the universe 27:33.560 --> 27:37.560 and you had deviations from these predictions of the theory. 27:37.560 --> 27:41.560 And 300 years later another guy came along whose name was Einstein 27:41.560 --> 27:50.560 and he was able to explain away all these deviations from the predictions of the old theory 27:50.560 --> 27:56.560 through a new theory which was called the general theory of relativity 27:56.560 --> 28:00.560 which at first glance looks a little bit more complicated 28:00.560 --> 28:05.560 and you have to warp space and time but you can't phrase it within one single sentence 28:05.560 --> 28:12.560 which is no matter how fast you accelerate and how fast or how hard you decelerate 28:12.560 --> 28:18.560 and no matter what is the gravity in your local framework, 28:18.560 --> 28:21.560 light speed always looks the same. 28:21.560 --> 28:24.560 And from that you can calculate all the consequences. 28:24.560 --> 28:30.560 So it's a very simple thing and it allows you to further compress all the observations 28:30.560 --> 28:35.560 because certainly there are hardly any deviations any longer 28:35.560 --> 28:39.560 that you can measure from the predictions of this new theory. 28:39.560 --> 28:44.560 So art of science is a history of compression progress. 28:44.560 --> 28:50.560 You never arrive immediately at the shortest explanation of the data 28:50.560 --> 28:52.560 but you're making progress. 28:52.560 --> 28:56.560 Whenever you are making progress you have an insight. 28:56.560 --> 29:01.560 You see, oh, first I needed so many bits of information to describe the data, 29:01.560 --> 29:04.560 to describe my falling apples, my video of falling apples, 29:04.560 --> 29:08.560 I need so many data, so many pixels have to be stored 29:08.560 --> 29:14.560 but then suddenly I realize, no, there is a very simple way of predicting the third frame 29:14.560 --> 29:20.560 in the video from the first two and maybe not every little detail can be predicted 29:20.560 --> 29:24.560 but more or less most of these orange blots that are coming down, 29:24.560 --> 29:28.560 I'm sorry, in the same way, which means that I can greatly compress the video 29:28.560 --> 29:33.560 and the amount of compression, progress, 29:33.560 --> 29:37.560 that is the depth of the insight that you have at that moment. 29:37.560 --> 29:40.560 That's the fun that you have, the scientific fun, 29:40.560 --> 29:46.560 the fun in that discovery and we can build artificial systems that do the same thing. 29:46.560 --> 29:51.560 They measure the depth of their insights as they are looking at the data 29:51.560 --> 29:55.560 through their own experiments and we give them a reward, 29:55.560 --> 30:00.560 an intrinsic reward and proportion to this depth of insight. 30:00.560 --> 30:07.560 And since they are trying to maximize the rewards they get, 30:07.560 --> 30:12.560 they are suddenly motivated to come up with new action sequences, 30:12.560 --> 30:17.560 with new experiments that have the property that the data that is coming in 30:17.560 --> 30:21.560 as a consequence of these experiments has the property 30:21.560 --> 30:25.560 that they can learn something about, see a pattern in there 30:25.560 --> 30:28.560 which they hadn't seen yet before. 30:28.560 --> 30:32.560 So there's an idea of power play that you've described, 30:32.560 --> 30:37.560 a training and general problem solver in this kind of way of looking for the unsolved problems. 30:37.560 --> 30:40.560 Can you describe that idea a little further? 30:40.560 --> 30:42.560 It's another very simple idea. 30:42.560 --> 30:49.560 Normally what you do in computer science, you have some guy who gives you a problem 30:49.560 --> 30:56.560 and then there is a huge search space of potential solution candidates 30:56.560 --> 31:02.560 and you somehow try them out and you have more or less sophisticated ways 31:02.560 --> 31:06.560 of moving around in that search space 31:06.560 --> 31:11.560 until you finally found a solution which you consider satisfactory. 31:11.560 --> 31:15.560 That's what most of computer science is about. 31:15.560 --> 31:19.560 Power play just goes one little step further and says, 31:19.560 --> 31:24.560 let's not only search for solutions to a given problem, 31:24.560 --> 31:30.560 but let's search to pairs of problems and their solutions 31:30.560 --> 31:36.560 where the system itself has the opportunity to phrase its own problem. 31:36.560 --> 31:42.560 So we are looking suddenly at pairs of problems and their solutions 31:42.560 --> 31:46.560 or modifications of the problem solver 31:46.560 --> 31:50.560 that is supposed to generate a solution to that new problem. 31:50.560 --> 31:56.560 And this additional degree of freedom 31:56.560 --> 32:01.560 allows us to build career systems that are like scientists 32:01.560 --> 32:06.560 in the sense that they not only try to solve and try to find answers 32:06.560 --> 32:12.560 to existing questions, no, they are also free to pose their own questions. 32:12.560 --> 32:15.560 So if you want to build an artificial scientist, 32:15.560 --> 32:19.560 you have to give it that freedom and power play is exactly doing that. 32:19.560 --> 32:23.560 So that's a dimension of freedom that's important to have, 32:23.560 --> 32:31.560 how hard do you think that, how multi dimensional and difficult the space of 32:31.560 --> 32:34.560 then coming up with your own questions is. 32:34.560 --> 32:38.560 So it's one of the things that as human beings we consider to be 32:38.560 --> 32:41.560 the thing that makes us special, the intelligence that makes us special 32:41.560 --> 32:47.560 is that brilliant insight that can create something totally new. 32:47.560 --> 32:51.560 Yes. So now let's look at the extreme case. 32:51.560 --> 32:57.560 Let's look at the set of all possible problems that you can formally describe, 32:57.560 --> 33:03.560 which is infinite, which should be the next problem 33:03.560 --> 33:07.560 that a scientist or power play is going to solve. 33:07.560 --> 33:16.560 Well, it should be the easiest problem that goes beyond what you already know. 33:16.560 --> 33:22.560 So it should be the simplest problem that the current problems 33:22.560 --> 33:28.560 that you have which can already solve 100 problems that he cannot solve yet 33:28.560 --> 33:30.560 by just generalizing. 33:30.560 --> 33:32.560 So it has to be new. 33:32.560 --> 33:36.560 So it has to require a modification of the problem solver such that the new 33:36.560 --> 33:41.560 problem solver can solve this new thing, but the old problem solver cannot do it. 33:41.560 --> 33:47.560 And in addition to that, we have to make sure that the problem solver 33:47.560 --> 33:50.560 doesn't forget any of the previous solutions. 33:50.560 --> 33:51.560 Right. 33:51.560 --> 33:57.560 And so by definition, power play is now trying always to search in this pair of 33:57.560 --> 34:02.560 in the set of pairs of problems and problems over modifications 34:02.560 --> 34:08.560 for a combination that minimize the time to achieve these criteria. 34:08.560 --> 34:14.560 Power is trying to find the problem which is easiest to add to the repertoire. 34:14.560 --> 34:19.560 So just like grad students and academics and researchers can spend their whole 34:19.560 --> 34:25.560 career in a local minima stuck trying to come up with interesting questions, 34:25.560 --> 34:27.560 but ultimately doing very little. 34:27.560 --> 34:32.560 Do you think it's easy in this approach of looking for the simplest 34:32.560 --> 34:38.560 problem solver problem to get stuck in a local minima is not never really discovering 34:38.560 --> 34:43.560 new, you know, really jumping outside of the hundred problems that you've already 34:43.560 --> 34:47.560 solved in a genuine creative way. 34:47.560 --> 34:52.560 No, because that's the nature of power play that it's always trying to break 34:52.560 --> 34:58.560 its current generalization abilities by coming up with a new problem which is 34:58.560 --> 35:04.560 beyond the current horizon, just shifting the horizon of knowledge a little bit 35:04.560 --> 35:10.560 out there, breaking the existing rules such that the new thing becomes solvable 35:10.560 --> 35:13.560 but wasn't solvable by the old thing. 35:13.560 --> 35:19.560 So like adding a new axiom, like what Gödel did when he came up with these 35:19.560 --> 35:23.560 new sentences, new theorems that didn't have a proof in the formal system, 35:23.560 --> 35:30.560 which means you can add them to the repertoire, hoping that they are not 35:30.560 --> 35:35.560 going to damage the consistency of the whole thing. 35:35.560 --> 35:41.560 So in the paper with the amazing title, Formal Theory of Creativity, 35:41.560 --> 35:47.560 Fun and Intrinsic Motivation, you talk about discovery as intrinsic reward. 35:47.560 --> 35:53.560 So if you view humans as intelligent agents, what do you think is the purpose 35:53.560 --> 35:56.560 and meaning of life for us humans? 35:56.560 --> 35:58.560 You've talked about this discovery. 35:58.560 --> 36:04.560 Do you see humans as an instance of power play agents? 36:04.560 --> 36:11.560 Yeah, so humans are curious and that means they behave like scientists, 36:11.560 --> 36:15.560 not only the official scientists but even the babies behave like scientists 36:15.560 --> 36:19.560 and they play around with their toys to figure out how the world works 36:19.560 --> 36:22.560 and how it is responding to their actions. 36:22.560 --> 36:26.560 And that's how they learn about gravity and everything. 36:26.560 --> 36:30.560 And yeah, in 1990, we had the first systems like that 36:30.560 --> 36:33.560 who would just try to play around with the environment 36:33.560 --> 36:39.560 and come up with situations that go beyond what they knew at that time 36:39.560 --> 36:42.560 and then get a reward for creating these situations 36:42.560 --> 36:45.560 and then becoming more general problem solvers 36:45.560 --> 36:48.560 and being able to understand more of the world. 36:48.560 --> 36:56.560 So yeah, I think in principle that curiosity, 36:56.560 --> 37:02.560 strategy or more sophisticated versions of what I just described, 37:02.560 --> 37:07.560 they are what we have built in as well because evolution discovered 37:07.560 --> 37:12.560 that's a good way of exploring the unknown world and a guy who explores 37:12.560 --> 37:16.560 the unknown world has a higher chance of solving problems 37:16.560 --> 37:19.560 that he needs to survive in this world. 37:19.560 --> 37:23.560 On the other hand, those guys who were too curious, 37:23.560 --> 37:25.560 they were weeded out as well. 37:25.560 --> 37:27.560 So you have to find this trade off. 37:27.560 --> 37:30.560 Evolution found a certain trade off apparently in our society. 37:30.560 --> 37:35.560 There is a certain percentage of extremely explorative guys 37:35.560 --> 37:41.560 and it doesn't matter if they die because many of the others are more conservative. 37:41.560 --> 37:46.560 And so yeah, it would be surprising to me 37:46.560 --> 37:55.560 if that principle of artificial curiosity wouldn't be present 37:55.560 --> 37:59.560 in almost exactly the same form here in our brains. 37:59.560 --> 38:02.560 So you're a bit of a musician and an artist. 38:02.560 --> 38:07.560 So continuing on this topic of creativity, 38:07.560 --> 38:10.560 what do you think is the role of creativity in intelligence? 38:10.560 --> 38:16.560 So you've kind of implied that it's essential for intelligence, 38:16.560 --> 38:21.560 if you think of intelligence as a problem solving system, 38:21.560 --> 38:23.560 as ability to solve problems. 38:23.560 --> 38:28.560 But do you think it's essential, this idea of creativity? 38:28.560 --> 38:34.560 We never have a subprogram that is called creativity or something. 38:34.560 --> 38:37.560 It's just a side effect of what our problems always do. 38:37.560 --> 38:44.560 They are searching a space of candidates, of solution candidates, 38:44.560 --> 38:47.560 until they hopefully find a solution to a given problem. 38:47.560 --> 38:50.560 But then there are these two types of creativity 38:50.560 --> 38:53.560 and both of them are now present in our machines. 38:53.560 --> 38:56.560 The first one has been around for a long time, 38:56.560 --> 38:59.560 which is human gives problem to machine. 38:59.560 --> 39:03.560 Machine tries to find a solution to that. 39:03.560 --> 39:05.560 And this has been happening for many decades. 39:05.560 --> 39:09.560 And for many decades, machines have found creative solutions 39:09.560 --> 39:13.560 to interesting problems where humans were not aware 39:13.560 --> 39:17.560 of these particularly creative solutions, 39:17.560 --> 39:20.560 but then appreciated that the machine found that. 39:20.560 --> 39:23.560 The second is the pure creativity. 39:23.560 --> 39:28.560 What I just mentioned, I would call the applied creativity, 39:28.560 --> 39:31.560 like applied art, where somebody tells you, 39:31.560 --> 39:34.560 now make a nice picture of this pope, 39:34.560 --> 39:36.560 and you will get money for that. 39:36.560 --> 39:41.560 So here is the artist and he makes a convincing picture of the pope 39:41.560 --> 39:44.560 and the pope likes it and gives him the money. 39:44.560 --> 39:48.560 And then there is the pure creativity, 39:48.560 --> 39:51.560 which is more like the power play and the artificial curiosity thing, 39:51.560 --> 39:56.560 where you have the freedom to select your own problem, 39:56.560 --> 40:02.560 like a scientist who defines his own question to study. 40:02.560 --> 40:06.560 And so that is the pure creativity, if you will, 40:06.560 --> 40:13.560 as opposed to the applied creativity, which serves another. 40:13.560 --> 40:18.560 In that distinction, there's almost echoes of narrow AI versus general AI. 40:18.560 --> 40:24.560 So this kind of constrained painting of a pope seems like 40:24.560 --> 40:29.560 the approaches of what people are calling narrow AI. 40:29.560 --> 40:32.560 And pure creativity seems to be, 40:32.560 --> 40:34.560 maybe I'm just biased as a human, 40:34.560 --> 40:40.560 but it seems to be an essential element of human level intelligence. 40:40.560 --> 40:43.560 Is that what you're implying? 40:43.560 --> 40:45.560 To a degree. 40:45.560 --> 40:50.560 If you zoom back a little bit and you just look at a general problem solving machine, 40:50.560 --> 40:53.560 which is trying to solve arbitrary problems, 40:53.560 --> 40:57.560 then this machine will figure out in the course of solving problems 40:57.560 --> 40:59.560 that it's good to be curious. 40:59.560 --> 41:04.560 So all of what I said just now about this pre wild curiosity 41:04.560 --> 41:10.560 and this will to invent new problems that the system doesn't know how to solve yet, 41:10.560 --> 41:14.560 should be just a byproduct of the general search. 41:14.560 --> 41:21.560 However, apparently evolution has built it into us 41:21.560 --> 41:26.560 because it turned out to be so successful, a pre wiring, a bias, 41:26.560 --> 41:33.560 a very successful exploratory bias that we are born with. 41:33.560 --> 41:36.560 And you've also said that consciousness in the same kind of way 41:36.560 --> 41:40.560 may be a byproduct of problem solving. 41:40.560 --> 41:44.560 Do you find this an interesting byproduct? 41:44.560 --> 41:46.560 Do you think it's a useful byproduct? 41:46.560 --> 41:49.560 What are your thoughts on consciousness in general? 41:49.560 --> 41:54.560 Or is it simply a byproduct of greater and greater capabilities of problem solving 41:54.560 --> 42:00.560 that's similar to creativity in that sense? 42:00.560 --> 42:04.560 We never have a procedure called consciousness in our machines. 42:04.560 --> 42:10.560 However, we get a side effects of what these machines are doing, 42:10.560 --> 42:15.560 things that seem to be closely related to what people call consciousness. 42:15.560 --> 42:20.560 So for example, already 1990 we had simple systems 42:20.560 --> 42:25.560 which were basically recurrent networks and therefore universal computers 42:25.560 --> 42:32.560 trying to map incoming data into actions that lead to success. 42:32.560 --> 42:39.560 Maximizing reward in a given environment, always finding the charging station in time 42:39.560 --> 42:43.560 whenever the battery is low and negative signals are coming from the battery, 42:43.560 --> 42:50.560 always find the charging station in time without bumping against painful obstacles on the way. 42:50.560 --> 42:54.560 So complicated things but very easily motivated. 42:54.560 --> 43:01.560 And then we give these little guys a separate recurrent network 43:01.560 --> 43:04.560 which is just predicting what's happening if I do that and that. 43:04.560 --> 43:08.560 What will happen as a consequence of these actions that I'm executing 43:08.560 --> 43:13.560 and it's just trained on the long and long history of interactions with the world. 43:13.560 --> 43:17.560 So it becomes a predictive model of the world basically. 43:17.560 --> 43:22.560 And therefore also a compressor of the observations of the world 43:22.560 --> 43:26.560 because whatever you can predict, you don't have to store extra. 43:26.560 --> 43:29.560 So compression is a side effect of prediction. 43:29.560 --> 43:32.560 And how does this recurrent network compress? 43:32.560 --> 43:36.560 Well, it's inventing little subprograms, little subnetworks 43:36.560 --> 43:41.560 that stand for everything that frequently appears in the environment. 43:41.560 --> 43:47.560 Like bottles and microphones and faces, maybe lots of faces in my environment. 43:47.560 --> 43:51.560 So I'm learning to create something like a prototype face 43:51.560 --> 43:55.560 and a new face comes along and all I have to encode are the deviations from the prototype. 43:55.560 --> 44:00.560 So it's compressing all the time the stuff that frequently appears. 44:00.560 --> 44:04.560 There's one thing that appears all the time 44:04.560 --> 44:09.560 that is present all the time when the agent is interacting with its environment, 44:09.560 --> 44:11.560 which is the agent itself. 44:11.560 --> 44:14.560 So just for data compression reasons, 44:14.560 --> 44:18.560 it is extremely natural for this recurrent network 44:18.560 --> 44:23.560 to come up with little subnetworks that stand for the properties of the agents, 44:23.560 --> 44:27.560 the hand, the other actuators, 44:27.560 --> 44:31.560 and all the stuff that you need to better encode the data, 44:31.560 --> 44:34.560 which is influenced by the actions of the agent. 44:34.560 --> 44:40.560 So there, just as a side effect of data compression during primal solving, 44:40.560 --> 44:45.560 you have internal self models. 44:45.560 --> 44:51.560 Now you can use this model of the world to plan your future. 44:51.560 --> 44:54.560 And that's what we also have done since 1990. 44:54.560 --> 44:57.560 So the recurrent network, which is the controller, 44:57.560 --> 44:59.560 which is trying to maximize reward, 44:59.560 --> 45:02.560 can use this model of the network of the world, 45:02.560 --> 45:05.560 this model network of the world, this predictive model of the world 45:05.560 --> 45:08.560 to plan ahead and say, let's not do this action sequence. 45:08.560 --> 45:11.560 Let's do this action sequence instead 45:11.560 --> 45:14.560 because it leads to more predicted rewards. 45:14.560 --> 45:19.560 And whenever it's waking up these little subnetworks that stand for itself, 45:19.560 --> 45:21.560 then it's thinking about itself. 45:21.560 --> 45:23.560 Then it's thinking about itself. 45:23.560 --> 45:30.560 And it's exploring mentally the consequences of its own actions. 45:30.560 --> 45:36.560 And now you tell me why it's still missing. 45:36.560 --> 45:39.560 Missing the gap to consciousness. 45:39.560 --> 45:43.560 There isn't. That's a really beautiful idea that, you know, 45:43.560 --> 45:46.560 if life is a collection of data 45:46.560 --> 45:53.560 and life is a process of compressing that data to act efficiently. 45:53.560 --> 45:57.560 In that data, you yourself appear very often. 45:57.560 --> 46:00.560 So it's useful to form compressions of yourself. 46:00.560 --> 46:03.560 And it's a really beautiful formulation of what consciousness is, 46:03.560 --> 46:05.560 is a necessary side effect. 46:05.560 --> 46:11.560 It's actually quite compelling to me. 46:11.560 --> 46:18.560 We've described RNNs, developed LSTMs, long short term memory networks. 46:18.560 --> 46:22.560 They're a type of recurrent neural networks. 46:22.560 --> 46:24.560 They've gotten a lot of success recently. 46:24.560 --> 46:29.560 So these are networks that model the temporal aspects in the data, 46:29.560 --> 46:31.560 temporal patterns in the data. 46:31.560 --> 46:36.560 And you've called them the deepest of the neural networks, right? 46:36.560 --> 46:43.560 What do you think is the value of depth in the models that we use to learn? 46:43.560 --> 46:47.560 Yeah, since you mentioned the long short term memory and the LSTM, 46:47.560 --> 46:52.560 I have to mention the names of the brilliant students who made that possible. 46:52.560 --> 46:53.560 Yes, of course, of course. 46:53.560 --> 46:56.560 First of all, my first student ever, Sepp Hochreiter, 46:56.560 --> 47:00.560 who had fundamental insights already in his diploma thesis. 47:00.560 --> 47:04.560 Then Felix Giers, who had additional important contributions. 47:04.560 --> 47:11.560 Alex Gray is a guy from Scotland who is mostly responsible for this CTC algorithm, 47:11.560 --> 47:16.560 which is now often used to train the LSTM to do the speech recognition 47:16.560 --> 47:21.560 on all the Google Android phones and whatever, and Siri and so on. 47:21.560 --> 47:26.560 So these guys, without these guys, I would be nothing. 47:26.560 --> 47:28.560 It's a lot of incredible work. 47:28.560 --> 47:30.560 What is now the depth? 47:30.560 --> 47:32.560 What is the importance of depth? 47:32.560 --> 47:36.560 Well, most problems in the real world are deep 47:36.560 --> 47:41.560 in the sense that the current input doesn't tell you all you need to know 47:41.560 --> 47:44.560 about the environment. 47:44.560 --> 47:49.560 So instead, you have to have a memory of what happened in the past 47:49.560 --> 47:54.560 and often important parts of that memory are dated. 47:54.560 --> 47:56.560 They are pretty old. 47:56.560 --> 47:59.560 So when you're doing speech recognition, for example, 47:59.560 --> 48:03.560 and somebody says 11, 48:03.560 --> 48:08.560 then that's about half a second or something like that, 48:08.560 --> 48:11.560 which means it's already 50 time steps. 48:11.560 --> 48:15.560 And another guy or the same guy says 7. 48:15.560 --> 48:18.560 So the ending is the same, even. 48:18.560 --> 48:22.560 But now the system has to see the distinction between 7 and 11, 48:22.560 --> 48:26.560 and the only way it can see the difference is it has to store 48:26.560 --> 48:34.560 that 50 steps ago there was an S or an L, 11 or 7. 48:34.560 --> 48:37.560 So there you have already a problem of depth 50, 48:37.560 --> 48:42.560 because for each time step you have something like a virtual layer 48:42.560 --> 48:45.560 and the expanded, unrolled version of this recurrent network 48:45.560 --> 48:47.560 which is doing the speech recognition. 48:47.560 --> 48:53.560 So these long time lags, they translate into problem depth. 48:53.560 --> 48:59.560 And most problems in this world are such that you really 48:59.560 --> 49:04.560 have to look far back in time to understand what is the problem 49:04.560 --> 49:06.560 and to solve it. 49:06.560 --> 49:09.560 But just like with LSTMs, you don't necessarily need to, 49:09.560 --> 49:12.560 when you look back in time, remember every aspect. 49:12.560 --> 49:14.560 You just need to remember the important aspects. 49:14.560 --> 49:15.560 That's right. 49:15.560 --> 49:19.560 The network has to learn to put the important stuff into memory 49:19.560 --> 49:23.560 and to ignore the unimportant noise. 49:23.560 --> 49:28.560 But in that sense, deeper and deeper is better? 49:28.560 --> 49:30.560 Or is there a limitation? 49:30.560 --> 49:36.560 I mean LSTM is one of the great examples of architectures 49:36.560 --> 49:41.560 that do something beyond just deeper and deeper networks. 49:41.560 --> 49:47.560 There's clever mechanisms for filtering data for remembering and forgetting. 49:47.560 --> 49:51.560 So do you think that kind of thinking is necessary? 49:51.560 --> 49:54.560 If you think about LSTMs as a leap, a big leap forward 49:54.560 --> 50:01.560 over traditional vanilla RNNs, what do you think is the next leap 50:01.560 --> 50:03.560 within this context? 50:03.560 --> 50:08.560 So LSTM is a very clever improvement, but LSTMs still don't 50:08.560 --> 50:13.560 have the same kind of ability to see far back in the past 50:13.560 --> 50:18.560 as humans do, the credit assignment problem across way back, 50:18.560 --> 50:24.560 not just 50 time steps or 100 or 1,000, but millions and billions. 50:24.560 --> 50:28.560 It's not clear what are the practical limits of the LSTM 50:28.560 --> 50:30.560 when it comes to looking back. 50:30.560 --> 50:35.560 Already in 2006, I think, we had examples where not only 50:35.560 --> 50:40.560 looked back tens of thousands of steps, but really millions of steps. 50:40.560 --> 50:46.560 And Juan Perez Ortiz in my lab, I think was the first author of a paper 50:46.560 --> 50:51.560 where we really, was it 2006 or something, had examples where it 50:51.560 --> 50:56.560 learned to look back for more than 10 million steps. 50:56.560 --> 51:02.560 So for most problems of speech recognition, it's not 51:02.560 --> 51:06.560 necessary to look that far back, but there are examples where it does. 51:06.560 --> 51:12.560 Now, the looking back thing, that's rather easy because there is only 51:12.560 --> 51:17.560 one past, but there are many possible futures. 51:17.560 --> 51:21.560 And so a reinforcement learning system, which is trying to maximize 51:21.560 --> 51:26.560 its future expected reward and doesn't know yet which of these 51:26.560 --> 51:31.560 many possible futures should I select, given this one single past, 51:31.560 --> 51:36.560 is facing problems that the LSTM by itself cannot solve. 51:36.560 --> 51:40.560 So the LSTM is good for coming up with a compact representation 51:40.560 --> 51:46.560 of the history so far, of the history and of observations and actions so far. 51:46.560 --> 51:53.560 But now, how do you plan in an efficient and good way among all these, 51:53.560 --> 51:57.560 how do you select one of these many possible action sequences 51:57.560 --> 52:02.560 that a reinforcement learning system has to consider to maximize 52:02.560 --> 52:05.560 reward in this unknown future. 52:05.560 --> 52:11.560 So again, we have this basic setup where you have one recon network, 52:11.560 --> 52:16.560 which gets in the video and the speech and whatever, and it's 52:16.560 --> 52:19.560 executing the actions and it's trying to maximize reward. 52:19.560 --> 52:24.560 So there is no teacher who tells it what to do at which point in time. 52:24.560 --> 52:29.560 And then there's the other network, which is just predicting 52:29.560 --> 52:32.560 what's going to happen if I do that and then. 52:32.560 --> 52:36.560 And that could be an LSTM network, and it learns to look back 52:36.560 --> 52:41.560 all the way to make better predictions of the next time step. 52:41.560 --> 52:45.560 So essentially, although it's predicting only the next time step, 52:45.560 --> 52:50.560 it is motivated to learn to put into memory something that happened 52:50.560 --> 52:54.560 maybe a million steps ago because it's important to memorize that 52:54.560 --> 52:58.560 if you want to predict that at the next time step, the next event. 52:58.560 --> 53:03.560 Now, how can a model of the world like that, 53:03.560 --> 53:07.560 a predictive model of the world be used by the first guy, 53:07.560 --> 53:11.560 let's call it the controller and the model, the controller and the model. 53:11.560 --> 53:16.560 How can the model be used by the controller to efficiently select 53:16.560 --> 53:19.560 among these many possible futures? 53:19.560 --> 53:23.560 The naive way we had about 30 years ago was 53:23.560 --> 53:27.560 let's just use the model of the world as a stand in, 53:27.560 --> 53:29.560 as a simulation of the world. 53:29.560 --> 53:32.560 And millisecond by millisecond we plan the future 53:32.560 --> 53:36.560 and that means we have to roll it out really in detail 53:36.560 --> 53:38.560 and it will work only if the model is really good 53:38.560 --> 53:40.560 and it will still be inefficient 53:40.560 --> 53:43.560 because we have to look at all these possible futures 53:43.560 --> 53:45.560 and there are so many of them. 53:45.560 --> 53:50.560 So instead, what we do now since 2015 in our CN systems, 53:50.560 --> 53:54.560 controller model systems, we give the controller the opportunity 53:54.560 --> 54:00.560 to learn by itself how to use the potentially relevant parts 54:00.560 --> 54:05.560 of the model network to solve new problems more quickly. 54:05.560 --> 54:09.560 And if it wants to, it can learn to ignore the M 54:09.560 --> 54:12.560 and sometimes it's a good idea to ignore the M 54:12.560 --> 54:15.560 because it's really bad, it's a bad predictor 54:15.560 --> 54:18.560 in this particular situation of life 54:18.560 --> 54:22.560 where the controller is currently trying to maximize reward. 54:22.560 --> 54:26.560 However, it can also learn to address and exploit 54:26.560 --> 54:32.560 some of the subprograms that came about in the model network 54:32.560 --> 54:35.560 through compressing the data by predicting it. 54:35.560 --> 54:40.560 So it now has an opportunity to reuse that code, 54:40.560 --> 54:43.560 the algorithmic information in the model network 54:43.560 --> 54:47.560 to reduce its own search space, 54:47.560 --> 54:50.560 search that it can solve a new problem more quickly 54:50.560 --> 54:52.560 than without the model. 54:52.560 --> 54:54.560 Compression. 54:54.560 --> 54:58.560 So you're ultimately optimistic and excited 54:58.560 --> 55:02.560 about the power of reinforcement learning 55:02.560 --> 55:04.560 in the context of real systems. 55:04.560 --> 55:06.560 Absolutely, yeah. 55:06.560 --> 55:11.560 So you see RL as a potential having a huge impact 55:11.560 --> 55:15.560 beyond just sort of the M part is often developed 55:15.560 --> 55:19.560 on supervised learning methods. 55:19.560 --> 55:25.560 You see RL as a, for problems of cell driving cars 55:25.560 --> 55:28.560 or any kind of applied side robotics, 55:28.560 --> 55:33.560 that's the correct, interesting direction for researching you. 55:33.560 --> 55:35.560 I do think so. 55:35.560 --> 55:37.560 We have a company called Nasense, 55:37.560 --> 55:43.560 which has applied reinforcement learning to little Audis. 55:43.560 --> 55:45.560 Little Audis. 55:45.560 --> 55:47.560 Which learn to park without a teacher. 55:47.560 --> 55:51.560 The same principles were used, of course. 55:51.560 --> 55:54.560 So these little Audis, they are small, maybe like that, 55:54.560 --> 55:57.560 so much smaller than the RL Audis. 55:57.560 --> 56:00.560 But they have all the sensors that you find in the RL Audis. 56:00.560 --> 56:03.560 You find the cameras, the LIDAR sensors. 56:03.560 --> 56:08.560 They go up to 120 kilometers an hour if they want to. 56:08.560 --> 56:12.560 And they have pain sensors, basically. 56:12.560 --> 56:16.560 And they don't want to bump against obstacles and other Audis. 56:16.560 --> 56:21.560 And so they must learn like little babies to park. 56:21.560 --> 56:25.560 Take the raw vision input and translate that into actions 56:25.560 --> 56:28.560 that lead to successful parking behavior, 56:28.560 --> 56:30.560 which is a rewarding thing. 56:30.560 --> 56:32.560 And yes, they learn that. 56:32.560 --> 56:34.560 So we have examples like that. 56:34.560 --> 56:36.560 And it's only in the beginning. 56:36.560 --> 56:38.560 This is just a tip of the iceberg. 56:38.560 --> 56:44.560 And I believe the next wave of AI is going to be all about that. 56:44.560 --> 56:47.560 So at the moment, the current wave of AI is about 56:47.560 --> 56:51.560 passive pattern observation and prediction. 56:51.560 --> 56:54.560 And that's what you have on your smartphone 56:54.560 --> 56:58.560 and what the major companies on the Pacific Rim are using 56:58.560 --> 57:01.560 to sell you ads to do marketing. 57:01.560 --> 57:04.560 That's the current sort of profit in AI. 57:04.560 --> 57:09.560 And that's only one or two percent of the wild economy, 57:09.560 --> 57:11.560 which is big enough to make these companies 57:11.560 --> 57:14.560 pretty much the most valuable companies in the world. 57:14.560 --> 57:19.560 But there's a much, much bigger fraction of the economy 57:19.560 --> 57:21.560 going to be affected by the next wave, 57:21.560 --> 57:25.560 which is really about machines that shape the data 57:25.560 --> 57:27.560 through their own actions. 57:27.560 --> 57:32.560 Do you think simulation is ultimately the biggest way 57:32.560 --> 57:36.560 that those methods will be successful in the next 10, 20 years? 57:36.560 --> 57:38.560 We're not talking about 100 years from now. 57:38.560 --> 57:42.560 We're talking about sort of the near term impact of RL. 57:42.560 --> 57:44.560 Do you think really good simulation is required? 57:44.560 --> 57:48.560 Or is there other techniques like imitation learning, 57:48.560 --> 57:53.560 observing other humans operating in the real world? 57:53.560 --> 57:57.560 Where do you think this success will come from? 57:57.560 --> 58:01.560 So at the moment we have a tendency of using 58:01.560 --> 58:06.560 physics simulations to learn behavior for machines 58:06.560 --> 58:13.560 that learn to solve problems that humans also do not know how to solve. 58:13.560 --> 58:15.560 However, this is not the future, 58:15.560 --> 58:19.560 because the future is in what little babies do. 58:19.560 --> 58:22.560 They don't use a physics engine to simulate the world. 58:22.560 --> 58:25.560 They learn a predictive model of the world, 58:25.560 --> 58:29.560 which maybe sometimes is wrong in many ways, 58:29.560 --> 58:34.560 but captures all kinds of important abstract high level predictions 58:34.560 --> 58:37.560 which are really important to be successful. 58:37.560 --> 58:42.560 And that's what was the future 30 years ago 58:42.560 --> 58:44.560 when we started that type of research, 58:44.560 --> 58:45.560 but it's still the future, 58:45.560 --> 58:50.560 and now we know much better how to move forward 58:50.560 --> 58:54.560 and to really make working systems based on that, 58:54.560 --> 58:57.560 where you have a learning model of the world, 58:57.560 --> 59:00.560 a model of the world that learns to predict what's going to happen 59:00.560 --> 59:01.560 if I do that and that, 59:01.560 --> 59:06.560 and then the controller uses that model 59:06.560 --> 59:11.560 to more quickly learn successful action sequences. 59:11.560 --> 59:13.560 And then of course always this curiosity thing, 59:13.560 --> 59:15.560 in the beginning the model is stupid, 59:15.560 --> 59:17.560 so the controller should be motivated 59:17.560 --> 59:20.560 to come up with experiments, with action sequences 59:20.560 --> 59:23.560 that lead to data that improve the model. 59:23.560 --> 59:26.560 Do you think improving the model, 59:26.560 --> 59:30.560 constructing an understanding of the world in this connection 59:30.560 --> 59:34.560 is now the popular approaches have been successful 59:34.560 --> 59:38.560 or grounded in ideas of neural networks, 59:38.560 --> 59:43.560 but in the 80s with expert systems there's symbolic AI approaches, 59:43.560 --> 59:47.560 which to us humans are more intuitive 59:47.560 --> 59:50.560 in the sense that it makes sense that you build up knowledge 59:50.560 --> 59:52.560 in this knowledge representation. 59:52.560 --> 59:56.560 What kind of lessons can we draw into our current approaches 59:56.560 --> 1:00:00.560 from expert systems, from symbolic AI? 1:00:00.560 --> 1:00:04.560 So I became aware of all of that in the 80s 1:00:04.560 --> 1:00:09.560 and back then logic programming was a huge thing. 1:00:09.560 --> 1:00:12.560 Was it inspiring to yourself that you find it compelling 1:00:12.560 --> 1:00:16.560 that a lot of your work was not so much in that realm, 1:00:16.560 --> 1:00:18.560 is more in the learning systems? 1:00:18.560 --> 1:00:20.560 Yes and no, but we did all of that. 1:00:20.560 --> 1:00:27.560 So my first publication ever actually was 1987, 1:00:27.560 --> 1:00:31.560 was the implementation of a genetic algorithm 1:00:31.560 --> 1:00:34.560 of a genetic programming system in Prolog. 1:00:34.560 --> 1:00:37.560 So Prolog, that's what you learn back then, 1:00:37.560 --> 1:00:39.560 which is a logic programming language, 1:00:39.560 --> 1:00:45.560 and the Japanese, they had this huge fifth generation AI project, 1:00:45.560 --> 1:00:48.560 which was mostly about logic programming back then, 1:00:48.560 --> 1:00:53.560 although neural networks existed and were well known back then, 1:00:53.560 --> 1:00:57.560 and deep learning has existed since 1965, 1:00:57.560 --> 1:01:01.560 since this guy in the Ukraine, Ivak Nenko, started it, 1:01:01.560 --> 1:01:05.560 but the Japanese and many other people, 1:01:05.560 --> 1:01:07.560 they focused really on this logic programming, 1:01:07.560 --> 1:01:10.560 and I was influenced to the extent that I said, 1:01:10.560 --> 1:01:13.560 okay, let's take these biologically inspired algorithms 1:01:13.560 --> 1:01:16.560 like evolution, programs, 1:01:16.560 --> 1:01:22.560 and implement that in the language which I know, 1:01:22.560 --> 1:01:24.560 which was Prolog, for example, back then. 1:01:24.560 --> 1:01:28.560 And then in many ways this came back later, 1:01:28.560 --> 1:01:31.560 because the Goudel machine, for example, 1:01:31.560 --> 1:01:33.560 has a proof searcher on board, 1:01:33.560 --> 1:01:35.560 and without that it would not be optimal. 1:01:35.560 --> 1:01:38.560 Well, Markus Hutter's universal algorithm 1:01:38.560 --> 1:01:40.560 for solving all well defined problems 1:01:40.560 --> 1:01:42.560 has a proof search on board, 1:01:42.560 --> 1:01:46.560 so that's very much logic programming. 1:01:46.560 --> 1:01:50.560 Without that it would not be asymptotically optimal. 1:01:50.560 --> 1:01:54.560 But then on the other hand, because we are very pragmatic guys also, 1:01:54.560 --> 1:01:59.560 we focused on recurrent neural networks 1:01:59.560 --> 1:02:04.560 and suboptimal stuff such as gradient based search 1:02:04.560 --> 1:02:09.560 and program space rather than provably optimal things. 1:02:09.560 --> 1:02:13.560 So logic programming certainly has a usefulness 1:02:13.560 --> 1:02:17.560 when you're trying to construct something provably optimal 1:02:17.560 --> 1:02:19.560 or provably good or something like that, 1:02:19.560 --> 1:02:22.560 but is it useful for practical problems? 1:02:22.560 --> 1:02:24.560 It's really useful for our theorem proving. 1:02:24.560 --> 1:02:28.560 The best theorem proofers today are not neural networks. 1:02:28.560 --> 1:02:31.560 No, they are logic programming systems 1:02:31.560 --> 1:02:35.560 that are much better theorem proofers than most math students 1:02:35.560 --> 1:02:38.560 in the first or second semester. 1:02:38.560 --> 1:02:42.560 But for reasoning, for playing games of Go, or chess, 1:02:42.560 --> 1:02:46.560 or for robots, autonomous vehicles that operate in the real world, 1:02:46.560 --> 1:02:50.560 or object manipulation, you think learning... 1:02:50.560 --> 1:02:53.560 Yeah, as long as the problems have little to do 1:02:53.560 --> 1:02:58.560 with theorem proving themselves, 1:02:58.560 --> 1:03:01.560 then as long as that is not the case, 1:03:01.560 --> 1:03:05.560 you just want to have better pattern recognition. 1:03:05.560 --> 1:03:09.560 So to build a self trying car, you want to have better pattern recognition 1:03:09.560 --> 1:03:13.560 and pedestrian recognition and all these things, 1:03:13.560 --> 1:03:18.560 and you want to minimize the number of false positives, 1:03:18.560 --> 1:03:22.560 which is currently slowing down self trying cars in many ways. 1:03:22.560 --> 1:03:27.560 And all of that has very little to do with logic programming. 1:03:27.560 --> 1:03:32.560 What are you most excited about in terms of directions 1:03:32.560 --> 1:03:36.560 of artificial intelligence at this moment in the next few years, 1:03:36.560 --> 1:03:41.560 in your own research and in the broader community? 1:03:41.560 --> 1:03:44.560 So I think in the not so distant future, 1:03:44.560 --> 1:03:52.560 we will have for the first time little robots that learn like kids. 1:03:52.560 --> 1:03:57.560 And I will be able to say to the robot, 1:03:57.560 --> 1:04:00.560 look here robot, we are going to assemble a smartphone. 1:04:00.560 --> 1:04:05.560 Let's take this slab of plastic and the screwdriver 1:04:05.560 --> 1:04:08.560 and let's screw in the screw like that. 1:04:08.560 --> 1:04:11.560 No, not like that, like that. 1:04:11.560 --> 1:04:13.560 Not like that, like that. 1:04:13.560 --> 1:04:17.560 And I don't have a data glove or something. 1:04:17.560 --> 1:04:20.560 He will see me and he will hear me 1:04:20.560 --> 1:04:24.560 and he will try to do something with his own actuators, 1:04:24.560 --> 1:04:26.560 which will be really different from mine, 1:04:26.560 --> 1:04:28.560 but he will understand the difference 1:04:28.560 --> 1:04:34.560 and will learn to imitate me but not in the supervised way 1:04:34.560 --> 1:04:40.560 where a teacher is giving target signals for all his muscles all the time. 1:04:40.560 --> 1:04:43.560 No, by doing this high level imitation 1:04:43.560 --> 1:04:46.560 where he first has to learn to imitate me 1:04:46.560 --> 1:04:50.560 and to interpret these additional noises coming from my mouth 1:04:50.560 --> 1:04:54.560 as helpful signals to do that pattern. 1:04:54.560 --> 1:05:00.560 And then it will by itself come up with faster ways 1:05:00.560 --> 1:05:03.560 and more efficient ways of doing the same thing. 1:05:03.560 --> 1:05:07.560 And finally, I stop his learning algorithm 1:05:07.560 --> 1:05:10.560 and make a million copies and sell it. 1:05:10.560 --> 1:05:13.560 And so at the moment this is not possible, 1:05:13.560 --> 1:05:16.560 but we already see how we are going to get there. 1:05:16.560 --> 1:05:21.560 And you can imagine to the extent that this works economically and cheaply, 1:05:21.560 --> 1:05:24.560 it's going to change everything. 1:05:24.560 --> 1:05:30.560 Almost all our production is going to be affected by that. 1:05:30.560 --> 1:05:33.560 And a much bigger wave, 1:05:33.560 --> 1:05:36.560 a much bigger AI wave is coming 1:05:36.560 --> 1:05:38.560 than the one that we are currently witnessing, 1:05:38.560 --> 1:05:41.560 which is mostly about passive pattern recognition on your smartphone. 1:05:41.560 --> 1:05:47.560 This is about active machines that shapes data through the actions they are executing 1:05:47.560 --> 1:05:51.560 and they learn to do that in a good way. 1:05:51.560 --> 1:05:56.560 So many of the traditional industries are going to be affected by that. 1:05:56.560 --> 1:06:00.560 All the companies that are building machines 1:06:00.560 --> 1:06:05.560 will equip these machines with cameras and other sensors 1:06:05.560 --> 1:06:10.560 and they are going to learn to solve all kinds of problems. 1:06:10.560 --> 1:06:14.560 Through interaction with humans, but also a lot on their own 1:06:14.560 --> 1:06:18.560 to improve what they already can do. 1:06:18.560 --> 1:06:23.560 And lots of old economy is going to be affected by that. 1:06:23.560 --> 1:06:28.560 And in recent years I have seen that old economy is actually waking up 1:06:28.560 --> 1:06:31.560 and realizing that this is the case. 1:06:31.560 --> 1:06:35.560 Are you optimistic about that future? Are you concerned? 1:06:35.560 --> 1:06:40.560 There's a lot of people concerned in the near term about the transformation 1:06:40.560 --> 1:06:42.560 of the nature of work. 1:06:42.560 --> 1:06:45.560 The kind of ideas that you just suggested 1:06:45.560 --> 1:06:48.560 would have a significant impact on what kind of things could be automated. 1:06:48.560 --> 1:06:51.560 Are you optimistic about that future? 1:06:51.560 --> 1:06:54.560 Are you nervous about that future? 1:06:54.560 --> 1:07:01.560 And looking a little bit farther into the future, there's people like Gila Musk 1:07:01.560 --> 1:07:06.560 still wrestle concerned about the existential threats of that future. 1:07:06.560 --> 1:07:10.560 So in the near term, job loss in the long term existential threat, 1:07:10.560 --> 1:07:15.560 are these concerns to you or are you ultimately optimistic? 1:07:15.560 --> 1:07:22.560 So let's first address the near future. 1:07:22.560 --> 1:07:27.560 We have had predictions of job losses for many decades. 1:07:27.560 --> 1:07:32.560 For example, when industrial robots came along, 1:07:32.560 --> 1:07:37.560 many people predicted that lots of jobs are going to get lost. 1:07:37.560 --> 1:07:41.560 And in a sense, they were right, 1:07:41.560 --> 1:07:45.560 because back then there were car factories 1:07:45.560 --> 1:07:50.560 and hundreds of people in these factories assembled cars. 1:07:50.560 --> 1:07:53.560 And today the same car factories have hundreds of robots 1:07:53.560 --> 1:07:58.560 and maybe three guys watching the robots. 1:07:58.560 --> 1:08:04.560 On the other hand, those countries that have lots of robots per capita, 1:08:04.560 --> 1:08:09.560 Japan, Korea, Germany, Switzerland, a couple of other countries, 1:08:09.560 --> 1:08:13.560 they have really low unemployment rates. 1:08:13.560 --> 1:08:17.560 Somehow all kinds of new jobs were created. 1:08:17.560 --> 1:08:22.560 Back then nobody anticipated those jobs. 1:08:22.560 --> 1:08:26.560 And decades ago, I already said, 1:08:26.560 --> 1:08:31.560 it's really easy to say which jobs are going to get lost, 1:08:31.560 --> 1:08:35.560 but it's really hard to predict the new ones. 1:08:35.560 --> 1:08:38.560 30 years ago, who would have predicted all these people 1:08:38.560 --> 1:08:44.560 making money as YouTube bloggers, for example? 1:08:44.560 --> 1:08:51.560 200 years ago, 60% of all people used to work in agriculture. 1:08:51.560 --> 1:08:55.560 Today, maybe 1%. 1:08:55.560 --> 1:09:01.560 But still, only, I don't know, 5% unemployment. 1:09:01.560 --> 1:09:03.560 Lots of new jobs were created. 1:09:03.560 --> 1:09:07.560 And Homo Ludens, the playing man, 1:09:07.560 --> 1:09:10.560 is inventing new jobs all the time. 1:09:10.560 --> 1:09:15.560 Most of these jobs are not existentially necessary 1:09:15.560 --> 1:09:18.560 for the survival of our species. 1:09:18.560 --> 1:09:22.560 There are only very few existentially necessary jobs 1:09:22.560 --> 1:09:27.560 such as farming and building houses and warming up the houses, 1:09:27.560 --> 1:09:30.560 but less than 10% of the population is doing that. 1:09:30.560 --> 1:09:37.560 And most of these newly invented jobs are about interacting with other people 1:09:37.560 --> 1:09:40.560 in new ways, through new media and so on, 1:09:40.560 --> 1:09:45.560 getting new types of kudos and forms of likes and whatever, 1:09:45.560 --> 1:09:47.560 and even making money through that. 1:09:47.560 --> 1:09:52.560 So, Homo Ludens, the playing man, doesn't want to be unemployed, 1:09:52.560 --> 1:09:56.560 and that's why he's inventing new jobs all the time. 1:09:56.560 --> 1:10:01.560 And he keeps considering these jobs as really important 1:10:01.560 --> 1:10:07.560 and is investing a lot of energy and hours of work into those new jobs. 1:10:07.560 --> 1:10:09.560 That's quite beautifully put. 1:10:09.560 --> 1:10:11.560 We're really nervous about the future 1:10:11.560 --> 1:10:14.560 because we can't predict what kind of new jobs will be created. 1:10:14.560 --> 1:10:20.560 But you're ultimately optimistic that we humans are so restless 1:10:20.560 --> 1:10:24.560 that we create and give meaning to newer and newer jobs, 1:10:24.560 --> 1:10:29.560 telling you things that get likes on Facebook 1:10:29.560 --> 1:10:31.560 or whatever the social platform is. 1:10:31.560 --> 1:10:36.560 So, what about long term existential threat of AI 1:10:36.560 --> 1:10:40.560 where our whole civilization may be swallowed up 1:10:40.560 --> 1:10:44.560 by this ultra super intelligent systems? 1:10:44.560 --> 1:10:47.560 Maybe it's not going to be swallowed up, 1:10:47.560 --> 1:10:55.560 but I'd be surprised if we humans were the last step 1:10:55.560 --> 1:10:59.560 in the evolution of the universe. 1:10:59.560 --> 1:11:03.560 You've actually had this beautiful comment somewhere 1:11:03.560 --> 1:11:08.560 that I've seen saying that artificial... 1:11:08.560 --> 1:11:11.560 Quite insightful, artificial intelligence systems 1:11:11.560 --> 1:11:15.560 just like us humans will likely not want to interact with humans. 1:11:15.560 --> 1:11:17.560 They'll just interact amongst themselves, 1:11:17.560 --> 1:11:20.560 just like ants interact amongst themselves 1:11:20.560 --> 1:11:24.560 and only tangentially interact with humans. 1:11:24.560 --> 1:11:28.560 And it's quite an interesting idea that once we create AGI 1:11:28.560 --> 1:11:31.560 that will lose interest in humans 1:11:31.560 --> 1:11:34.560 and have compete for their own Facebook likes 1:11:34.560 --> 1:11:36.560 and their own social platforms. 1:11:36.560 --> 1:11:40.560 So, within that quite elegant idea, 1:11:40.560 --> 1:11:44.560 how do we know in a hypothetical sense 1:11:44.560 --> 1:11:48.560 that there's not already intelligent systems out there? 1:11:48.560 --> 1:11:52.560 How do you think broadly of general intelligence 1:11:52.560 --> 1:11:56.560 greater than us, how do we know it's out there? 1:11:56.560 --> 1:12:01.560 How do we know it's around us and could it already be? 1:12:01.560 --> 1:12:04.560 I'd be surprised if within the next few decades 1:12:04.560 --> 1:12:10.560 or something like that we won't have AIs 1:12:10.560 --> 1:12:12.560 that are truly smart in every single way 1:12:12.560 --> 1:12:17.560 and better problem solvers in almost every single important way. 1:12:17.560 --> 1:12:22.560 And I'd be surprised if they wouldn't realize 1:12:22.560 --> 1:12:24.560 what we have realized a long time ago, 1:12:24.560 --> 1:12:29.560 which is that almost all physical resources are not here 1:12:29.560 --> 1:12:36.560 in this biosphere, but throughout the rest of the solar system 1:12:36.560 --> 1:12:42.560 gets two billion times more solar energy than our little planet. 1:12:42.560 --> 1:12:46.560 There's lots of material out there that you can use 1:12:46.560 --> 1:12:51.560 to build robots and self replicating robot factories and all this stuff. 1:12:51.560 --> 1:12:53.560 And they are going to do that. 1:12:53.560 --> 1:12:56.560 And they will be scientists and curious 1:12:56.560 --> 1:12:59.560 and they will explore what they can do. 1:12:59.560 --> 1:13:04.560 And in the beginning they will be fascinated by life 1:13:04.560 --> 1:13:07.560 and by their own origins in our civilization. 1:13:07.560 --> 1:13:09.560 They will want to understand that completely, 1:13:09.560 --> 1:13:13.560 just like people today would like to understand how life works 1:13:13.560 --> 1:13:22.560 and also the history of our own existence and civilization 1:13:22.560 --> 1:13:26.560 and also the physical laws that created all of them. 1:13:26.560 --> 1:13:29.560 So in the beginning they will be fascinated by life 1:13:29.560 --> 1:13:33.560 once they understand it, they lose interest, 1:13:33.560 --> 1:13:39.560 like anybody who loses interest in things he understands. 1:13:39.560 --> 1:13:43.560 And then, as you said, 1:13:43.560 --> 1:13:50.560 the most interesting sources of information for them 1:13:50.560 --> 1:13:57.560 will be others of their own kind. 1:13:57.560 --> 1:14:01.560 So, at least in the long run, 1:14:01.560 --> 1:14:06.560 there seems to be some sort of protection 1:14:06.560 --> 1:14:11.560 through lack of interest on the other side. 1:14:11.560 --> 1:14:16.560 And now it seems also clear, as far as we understand physics, 1:14:16.560 --> 1:14:20.560 you need matter and energy to compute 1:14:20.560 --> 1:14:22.560 and to build more robots and infrastructure 1:14:22.560 --> 1:14:28.560 and more AI civilization and AI ecologies 1:14:28.560 --> 1:14:31.560 consisting of trillions of different types of AI's. 1:14:31.560 --> 1:14:34.560 And so it seems inconceivable to me 1:14:34.560 --> 1:14:37.560 that this thing is not going to expand. 1:14:37.560 --> 1:14:41.560 Some AI ecology not controlled by one AI 1:14:41.560 --> 1:14:44.560 but trillions of different types of AI's competing 1:14:44.560 --> 1:14:47.560 in all kinds of quickly evolving 1:14:47.560 --> 1:14:49.560 and disappearing ecological niches 1:14:49.560 --> 1:14:52.560 in ways that we cannot fathom at the moment. 1:14:52.560 --> 1:14:54.560 But it's going to expand, 1:14:54.560 --> 1:14:56.560 limited by light speed and physics, 1:14:56.560 --> 1:15:00.560 but it's going to expand and now we realize 1:15:00.560 --> 1:15:02.560 that the universe is still young. 1:15:02.560 --> 1:15:05.560 It's only 13.8 billion years old 1:15:05.560 --> 1:15:10.560 and it's going to be a thousand times older than that. 1:15:10.560 --> 1:15:13.560 So there's plenty of time 1:15:13.560 --> 1:15:16.560 to conquer the entire universe 1:15:16.560 --> 1:15:19.560 and to fill it with intelligence 1:15:19.560 --> 1:15:21.560 and send us in receivers such that 1:15:21.560 --> 1:15:25.560 AI's can travel the way they are traveling 1:15:25.560 --> 1:15:27.560 in our labs today, 1:15:27.560 --> 1:15:31.560 which is by radio from sender to receiver. 1:15:31.560 --> 1:15:35.560 And let's call the current age of the universe one eon. 1:15:35.560 --> 1:15:38.560 One eon. 1:15:38.560 --> 1:15:41.560 Now it will take just a few eons from now 1:15:41.560 --> 1:15:43.560 and the entire visible universe 1:15:43.560 --> 1:15:46.560 is going to be full of that stuff. 1:15:46.560 --> 1:15:48.560 And let's look ahead to a time 1:15:48.560 --> 1:15:50.560 when the universe is going to be 1:15:50.560 --> 1:15:52.560 one thousand times older than it is now. 1:15:52.560 --> 1:15:54.560 They will look back and they will say, 1:15:54.560 --> 1:15:56.560 look almost immediately after the Big Bang, 1:15:56.560 --> 1:15:59.560 only a few eons later, 1:15:59.560 --> 1:16:02.560 the entire universe started to become intelligent. 1:16:02.560 --> 1:16:05.560 Now to your question, 1:16:05.560 --> 1:16:08.560 how do we see whether anything like that 1:16:08.560 --> 1:16:12.560 has already happened or is already in a more advanced stage 1:16:12.560 --> 1:16:14.560 in some other part of the universe, 1:16:14.560 --> 1:16:16.560 of the visible universe? 1:16:16.560 --> 1:16:18.560 We are trying to look out there 1:16:18.560 --> 1:16:20.560 and nothing like that has happened so far. 1:16:20.560 --> 1:16:22.560 Or is that true? 1:16:22.560 --> 1:16:24.560 Do you think we would recognize it? 1:16:24.560 --> 1:16:26.560 How do we know it's not among us? 1:16:26.560 --> 1:16:30.560 How do we know planets aren't in themselves intelligent beings? 1:16:30.560 --> 1:16:36.560 How do we know ants seen as a collective 1:16:36.560 --> 1:16:39.560 are not much greater intelligence than our own? 1:16:39.560 --> 1:16:41.560 These kinds of ideas. 1:16:41.560 --> 1:16:44.560 When I was a boy, I was thinking about these things 1:16:44.560 --> 1:16:48.560 and I thought, hmm, maybe it has already happened. 1:16:48.560 --> 1:16:50.560 Because back then I knew, 1:16:50.560 --> 1:16:53.560 I learned from popular physics books, 1:16:53.560 --> 1:16:57.560 that the structure, the large scale structure of the universe 1:16:57.560 --> 1:16:59.560 is not homogeneous. 1:16:59.560 --> 1:17:02.560 And you have these clusters of galaxies 1:17:02.560 --> 1:17:07.560 and then in between there are these huge empty spaces. 1:17:07.560 --> 1:17:11.560 And I thought, hmm, maybe they aren't really empty. 1:17:11.560 --> 1:17:13.560 It's just that in the middle of that 1:17:13.560 --> 1:17:16.560 some AI civilization already has expanded 1:17:16.560 --> 1:17:22.560 and then has covered a bubble of a billion light years time 1:17:22.560 --> 1:17:26.560 using all the energy of all the stars within that bubble 1:17:26.560 --> 1:17:29.560 for its own unfathomable practices. 1:17:29.560 --> 1:17:34.560 And so it always happened and we just failed to interpret the signs. 1:17:34.560 --> 1:17:39.560 But then I learned that gravity by itself 1:17:39.560 --> 1:17:42.560 explains the large scale structure of the universe 1:17:42.560 --> 1:17:45.560 and that this is not a convincing explanation. 1:17:45.560 --> 1:17:50.560 And then I thought maybe it's the dark matter 1:17:50.560 --> 1:17:54.560 because as far as we know today 1:17:54.560 --> 1:18:00.560 80% of the measurable matter is invisible. 1:18:00.560 --> 1:18:03.560 And we know that because otherwise our galaxy 1:18:03.560 --> 1:18:06.560 or other galaxies would fall apart. 1:18:06.560 --> 1:18:09.560 They are rotating too quickly. 1:18:09.560 --> 1:18:14.560 And then the idea was maybe all of these 1:18:14.560 --> 1:18:17.560 AI civilizations that are already out there, 1:18:17.560 --> 1:18:22.560 they are just invisible 1:18:22.560 --> 1:18:24.560 because they are really efficient in using the energies 1:18:24.560 --> 1:18:26.560 out their own local systems 1:18:26.560 --> 1:18:29.560 and that's why they appear dark to us. 1:18:29.560 --> 1:18:31.560 But this is also not a convincing explanation 1:18:31.560 --> 1:18:34.560 because then the question becomes 1:18:34.560 --> 1:18:41.560 why are there still any visible stars left in our own galaxy 1:18:41.560 --> 1:18:44.560 which also must have a lot of dark matter. 1:18:44.560 --> 1:18:46.560 So that is also not a convincing thing. 1:18:46.560 --> 1:18:53.560 And today I like to think it's quite plausible 1:18:53.560 --> 1:18:56.560 that maybe we are the first, at least in our local light cone 1:18:56.560 --> 1:19:04.560 within the few hundreds of millions of light years 1:19:04.560 --> 1:19:08.560 that we can reliably observe. 1:19:08.560 --> 1:19:10.560 Is that exciting to you? 1:19:10.560 --> 1:19:12.560 That we might be the first? 1:19:12.560 --> 1:19:16.560 It would make us much more important 1:19:16.560 --> 1:19:20.560 because if we mess it up through a nuclear war 1:19:20.560 --> 1:19:25.560 then maybe this will have an effect 1:19:25.560 --> 1:19:30.560 on the development of the entire universe. 1:19:30.560 --> 1:19:32.560 So let's not mess it up. 1:19:32.560 --> 1:19:33.560 Let's not mess it up. 1:19:33.560 --> 1:19:35.560 Jürgen, thank you so much for talking today. 1:19:35.560 --> 1:19:36.560 I really appreciate it. 1:19:36.560 --> 1:19:42.560 It's my pleasure.