Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11

Lex Fridman 0:00 The following is a conversation with Jurgen Schmidhuber. He's the co-director of atia Swiss AI lab and the co-creator of long short-term memory networks. LSTMs are used in billions of devices today for speech recognition, translation, and much more. Over 30 years, he has proposed a lot of interesting out of the box ideas on meta-learning, adversarial networks, computer vision, and even a formal theory of, quote, "Creativity, curiosity, and fun." This conversation is part of the MIT course on Artificial General Intelligence and the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, iTunes, or simply connect with me on Twitter @lexfridspelled F-R-I-D. And now here's my conversation with Jurgen Schmidhuber. Early on, you dreamed of AI systems that self-improve recursively. When was that dream born?

Jürgen Schmidhuber 1:01 When I was a baby. No, that's not true.

Lex Fridman 1:04 (laughs)

Jürgen Schmidhuber 1:04 When I was a teenager.

Lex Fridman 1:06 And what was the catalyst for that birth? What was the thing that first inspired you?

Jürgen Schmidhuber 1:13 When I was a boy, I'm, I was thinking about what to do in my life and then I thought the most exciting thing is to solve the riddles of the universe and, and that means you have to become a physicist. However, then I realized that there's something even grander, you can try to build a machine that isn't really a machine any longer that learns to become a much better physicist than I could ever hope to be. And that's how I thought maybe I can multiply my tiny little bit of creativity into infinity.

Lex Fridman 1:54 But ultimately, that creativity will be multiplied to understand the universe around us? That's, that's the, the curiosity for that mystery that, that drove you?

Jürgen Schmidhuber 2:05 Yes. Uh, so if you can build a machine that learns to solve more and more complex problems and more and more general problem solver, then you basically ha-have, um, solved all the problems. At least all the solvable problems.

Lex Fridman 2:26 So how do you think, what is the mechanism for that kind of general solver look like? Because obviously we don't quite yet have one or know how to build one, but we have ideas and you have had throughout your career several ideas about it. So how do you think about that mechanism?

Jürgen Schmidhuber 2:43 So in the '80s, I thought about how to build this machine that learns to solve all these problems that I cannot solve myself and I thought it is clear it has to be a machine that not only learns to solve this problem here and this problem here, but it also has to learn to improve the learning algorithm itself.

Lex Fridman 3:08 Right.

Jürgen Schmidhuber 3:09 So it has to have the learning algorithm in, um, representation that allows it to inspect it and modify it, such that it can come up with a better learning algorithm. So I call that meta-learning, learning to learn and recursive self-improvement, um, that is really the pinnacle of that where you then not only learn, um, how to improve on that problem and on that, but you also improve the way the machine improves, and you also improve the way it improves the way it improves itself. And that was my 1987 diploma thesis which was all about that, hierarchy of meta-learners that have no computational limits except for the well-known limits, uh, that Goedel identified in 1931, and, uh, for the limits are physics.

Lex Fridman 4:05 Mm-hmm. In the recent years, meta-learning has gained popularity in a v- in a specific kind of form. You've talked about how that's not really meta-learning w-with neural networks, that's more m- basic transfer learning. Can you talk about the difference between the big general meta-learning-

Jürgen Schmidhuber 4:25 Mm-hmm.

Lex Fridman 4:25 ... and a more narrow sense of meta-learning the way it's used today, the way it's talked about today?

Jürgen Schmidhuber 4:30 Let's take the example of a deep neural network that has, uh, learned to classify images and maybe you have trained that, um, network on 100 different databases of images.

Lex Fridman 4:43 Mm-hmm.

Jürgen Schmidhuber 4:45 And now a new database comes along and you want to quickly learn the new thing as well. So one simple way of doing that is you take the network which already knows 100 types of databases and then you just take the top layer of that and you retrain that, uh, using the new label data that you have in the new image database. And then it turns out that it really, really quickly can learn that too.

Lex Fridman 5:18 Mm-hmm.

Jürgen Schmidhuber 5:19 One shot basically.

Lex Fridman 5:20 Mm-hmm.

Jürgen Schmidhuber 5:20 Because from the first 100 datasets, it already has learned so much about, about computer vision that it can reuse that and that is then almost good enough to solve the new task except you need a little bit of, um, adjustment on the top. So that is transfer learning and it has been done in principle for many decades. People have done similar things for decades. Meta-learning, true meta-learning is about having the learning algorithm itself open to introspection by the system that is using it.... and also open to modification, such that the learning system has an opportunity to modify any part of the learning algorithm and then evaluate the consequences of that modification and then learn from that, uh, to create a better learning algorithm and so on recursively. So that's a very different animal where you are opening the space of possible learning algorithms to the learning system itself.

Lex Fridman 6:35 Right. So you've, uh, like in the 2004 paper, you described, uh, Gödel machines and programs that rewrite themselves.

Jürgen Schmidhuber 6:43 Yeah.

Lex Fridman 6:44 Right? Philosophically and even in your paper mathematically, these are really compelling ideas. But practically, do you see these self-referential programs being successful in the near term to having an impact where sort of it demonstrates to the world that th- this direction is a, is a good one to pursue in the near term?

Jürgen Schmidhuber 7:08 Yes. We had these two different types of, um, fundamental research, um, how to build a universal problem solver. One basically exploiting proof search and things like that that you need to come up with asymptotically optimal, theoretically optimal self-improvers and problem solvers. However, one has to admit that through this proof search comes in an additive constant, an overhead, an additive overhead that vanishes in comparison to, uh, what you have to do to solve large problems.

Lex Fridman 7:54 Mm-hmm.

Jürgen Schmidhuber 7:55 However, for many of the small problems that we want to solve in our everyday life, we cannot ignore this constant overhead and that's why we also have been, um, doing other things, non-universal things such as recurrent neural networks which are trained by gradient descent-

Lex Fridman 8:15 Mm-hmm.

Jürgen Schmidhuber 8:15 ... and local search techniques which aren't universal at all, which aren't provably optimal at all like the other stuff that we did, but which are much more practical as long as we only want to solve the small problems that we are typically trying to solve in this environment here. Yeah, so the universal problem solvers-

Lex Fridman 8:37 Yeah.

Jürgen Schmidhuber 8:37 ... uh, like the Gödel machine but also Marcus Hutter's, um, fastest way of solving all possible problems which he developed around 2002, uh, two in my lab. Uh, they are associated with these c- constant overheads, uh, for proof search which guarantees that the thing you're doing is optimal. For example, there is this fastest way of solving all problems with a computable solution which is due to Marcus, Marcus Hutter.

Lex Fridman 9:07 Mm-hmm.

Jürgen Schmidhuber 9:08 And, uh, and to explain what's going on there, let's take traveling salesman problems.

Lex Fridman 9:14 Mm-hmm.

Jürgen Schmidhuber 9:15 With traveling salesman problems you have a number of cities, N cities, and you try to find the shortest path-

Lex Fridman 9:23 Mm-hmm.

Jürgen Schmidhuber 9:23 ... through all these cities without visiting any city twice. And nobody knows the fastest way of solving traveling salesman problems, TSPs, but let's assume there is a method of solving them within N to the 5 operations where N is the number of cities. Then, uh, the universal method of Marcus is going to solve the same traveling salesman problem also within N to the 5 steps.

Lex Fridman 10:01 Mm-hmm.

Jürgen Schmidhuber 10:02 Plus-

Lex Fridman 10:03 A constant.

Jürgen Schmidhuber 10:03 ... O of one plus a constant number of steps that you need for the proof searcher which, uh, you need to, um, show that this particular class of problems, the traveling salesman, salesman problems can be solved within a certain time bound, um, within order N to the 5 steps basically. And this, uh, additive constant doesn't care for N which means as N is getting larger and larger, as you have more and more cities, the constant overhead pales in comparison. And that means that almost all large problems are solved in the best possible way already today. We already have a universal problem solver like that. However, it's not practical because the overhead, the constant overhead is so large that for the small kinds of problems that we want to solve in this little biosphere...

Lex Fridman 11:04 By the way, when you say small, you're talking about things that fall within the constraints of our computational systems. So they can, they can seem quite large to us mere humans, right?

Jürgen Schmidhuber 11:14 That's right, yeah. So they seem large and even unsolvable in a practical sense today, but they are still small compared to almost all problems because almost all problems are large problems which are much larger than any constant.

Lex Fridman 11:32 Do you find it useful as a person who has dreamed of creating a general learning system, has worked on creating one, has done a lot of interesting ideas there, to think about P versus NP, this, uh, formalization of how hard problems are, how they scale, this kind of worst-case analysis type of thinking?

Jürgen Schmidhuber 11:55 Mm-hmm.

Lex Fridman 11:55 Do you find that useful or is it only just a mathematical... it, it's a set of mathematical techniques to give you intuition about what's good and bad?

Jürgen Schmidhuber 12:05 Mm-hmm. So P versus NP, that's, uh, super interesting from a theoretical point of view and in fact as you are thinking about that problem, you can also get inspiration for better practical problem solvers.... on the other hand, we have to admit that at the moment, the best practical problem-solvers for all kinds of problems that we are now solving through what is called AI at the moment-

Lex Fridman 12:33 Mm-hmm.

Jürgen Schmidhuber 12:33 ... they are not of the kind that is inspired by these questions.

Lex Fridman 12:38 Yes.

Jürgen Schmidhuber 12:38 No. There we are using, um, general purpose computers such as recurrent neural networks, but we have a search technique which is just local search, gradient descent-

Lex Fridman 12:50 Mm-hmm.

Jürgen Schmidhuber 12:50 ... to try to find a program that is running on these recurrent networks such that it can solve some interesting problems such, such as speech recognition or machine translation and something like that. And there is very little theory behind the best solutions that we have at the moment that can do that.

Lex Fridman 13:10 Do you think that needs to change? Do you think that will change or can we go, can we create a general intelligent systems without ever really proving that that system is intelligent in some kinda mathematical way? Solving machine translation perfectly or something like that, within some kinda syntactic definition or language or can we just be super impressed by the thing working extremely well and that's sufficient?

Jürgen Schmidhuber 13:35 There's an old saying, and I don't know who brought it up first, uh, which says there is nothing more practical than a good theory.

Lex Fridman 13:44 (laughs)

Jürgen Schmidhuber 13:45 (laughs) And, um-

Lex Fridman 13:46 Yeah.

Jürgen Schmidhuber 13:47 ... and a good theory of problem-solving under limited resources like here in this universe or on this little planet, has to take into account these limited resources. And so probably there is lurking a theory which is related to what we already have, these asymptotically optimal problem-solvers, which, which, uh, tells us what we need in addition to that to come up with a practically optimal problem-solver. So I believe we will have something like that, and maybe just a few little tiny twists unnecessary to, to change what we already have to come up with that as well. As long as we don't have that, we, mm, admit that we are taking suboptimal ways and recurrent neural networks and long short-term memory for, uh, equipped with local search techniques, and we are happy that it works better than any competing method but, um, that doesn't mean that we, we think we are done.

Lex Fridman 15:00 You said that an AGI system will ultimately be a simple one. Uh, a general intelligence system will ultimately be a simple one, maybe a pseudocode of a few lines will be able to describe it. Can you talk through your intuition behind this idea, why you feel that a s- at its core, intelligence is a simple algorithm?

Jürgen Schmidhuber 15:27 Experience tells us that the stuff that works best is really simple. So the asymptotically optimal ways of solving problems, if you look at them, they're just a few lines of code. It's really true. Although they have these amazing properties, just a few lines of code. Then the most, mm, promising and most useful practical things maybe don't have this proof of optimality associated with them. However, they are also just a few lines of code. The most successful, mm, recurrent neural networks, you can write them down in five lines of pseudocode.

Lex Fridman 16:08 Th- that's a beautiful, almost poetic idea, but w- what you're describing there is the s- the lines of pseudocode are sitting on top of layers and layers of abstractions in a sense.

Jürgen Schmidhuber 16:21 Mm-hmm.

Lex Fridman 16:22 So y- you're saying at the very top-

Jürgen Schmidhuber 16:24 Mm-hmm.

Lex Fridman 16:25 ... it will be a beautifully written sort of, uh, algorithm.

Jürgen Schmidhuber 16:30 Mm-hmm.

Lex Fridman 16:31 But do you think that there's many layers of abstractions we have to first learn to construct?

Jürgen Schmidhuber 16:37 Yeah. Of course, we are building on all these, um, great abstractions that people have invented over the millennia such as matrix multiplications and real numbers and basic arithmetics and calculus and derivations of, um, arrow functions and derivatives of arrow functions and stuff like that. So without that language, that greatly simplifies our way of thinking about these problems, we couldn't do anything. So in that sense as always, we are standing on the shoulders of the giants who, in the past, um, simplified the problem of problem-solving so much that now we have a chance to do the final step.

Lex Fridman 17:29 (laughs) So the final step will be a simple one. Uh, w- if we, if we take a step back through all of human civilization and just the universe in general (laughs) , uh, w- how do you think about evolution and what if creating a universe is required to achieve this final step?

Jürgen Schmidhuber 17:46 Mm-hmm.

Lex Fridman 17:47 What if going through the very painful and inefficient process of evolution is needed-

Jürgen Schmidhuber 17:53 Mm-hmm.

Lex Fridman 17:53 ... to come up with this set of abstractions that ultimately lead to intelligence? Do you think there's a shortcut or do you think we have to create something like our universe in order to create something like human level intelligence?

Jürgen Schmidhuber 18:07 Hmm. So far the only example we have, uh-

Lex Fridman 18:11 (laughs)

Jürgen Schmidhuber 18:12 ... is this one.

Lex Fridman 18:13 Yeah.

Jürgen Schmidhuber 18:13 This universe, um, in which we are living.

Lex Fridman 18:15 Do you think it can do better?

Jürgen Schmidhuber 18:20 Maybe not.

Lex Fridman 18:22 (laughs)

Jürgen Schmidhuber 18:22 But, um, we are part of this whole process.

Lex Fridman 18:24 Right.

Jürgen Schmidhuber 18:25 So...... apparently, so it might be the case that the code that runs the universe is really, really simple. Everything points to that possibility because gravity and other basic forces are really simple laws that can be easily described also in just a few lines of code basically. And, uh, and then there are these other, um, events that... the apparently random events in the history of the universe which, as far as we know at the moment, don't have a compact code, but who knows? Maybe somebody in the near future is going to figure out the pseudo-random generator which is, um, which, um, is computing whether the measurement of that, um, spin up or down thing here is, um, going to be positive or negative.

Lex Fridman 19:18 Underlying quantum mechanics?

Jürgen Schmidhuber 19:20 Yes. So-

Lex Fridman 19:20 Do you ultimately think quantum mechanics is a, a pseudo-random number gen- so it's all deterministic? There's no randomness in our universe? Does God play dice?

Jürgen Schmidhuber 19:32 So a couple of years ago, a famous physicist, quantum physicist, um, Anton Zeilinger, he wrote an essay in nature and it started more or less like that. One of the fundamental insights of the, of the 20th century was that the universe is fundamentally random on the quantum level. And that whenever you measure spin up or down or something like that, a new bit of information enters the history of the universe. And while I was reading that, I was already typing, uh, the response and they had to publish it because I was right.

Lex Fridman 20:22 (laughs)

Jürgen Schmidhuber 20:22 That there is no evidence, no physical evidence, uh, for that. So there's an alternative explanation where everything that we consider random is actually pseudo-random, such as the decimal expansion of pi, 3.141 and so on, which looks random but isn't. So pi is interesting because every three-digit sequence, ev- every sequence of, uh, three digits appears roughly one in a thousand times, and every five digit sequence appears roughly one in 10,000 times what you would, would expect if it was ran- random. But there's a very short algorithm, a short program that computes all of that. So it's extremely compressible. And who knows? Maybe tomorrow somebody, some grad student at CERN goes back over all these, um, data points, beta decay and whatever, and figures out, oh, it's, uh, the s- the second billion digits of pi or something like that.

Lex Fridman 21:28 (laughs)

Jürgen Schmidhuber 21:28 We don't have any fundamental reason at the moment to believe that this is truly random and not just a deterministic video game. If it was a deterministic video game, it would be much more beautiful because beauty is simplicity, and many of the basic laws of the universe, like gravity and, um, the other basic forces are very simple. So very short programs can explain what these are doing. And, um, and it would be awful and ugly. The universe would be ugly. The history of the universe would be ugly if for the extra things, the random, the seemingly random data points that we get all the time, that we really need a huge number of extra bits to describe all these, um, these extra bits of information. So as long as we don't have evidence that there is no short program that computes the entire history of the entire universe, we are, as scientists, compelled to look further for that, um, shortest program.

Lex Fridman 22:43 Your intuition says there exists a shortest... a program that can backtrack to the, to the creation of the universe-

Jürgen Schmidhuber 22:52 Yeah. So-

Lex Fridman 22:52 ... to compute the shortest path to the creation of the universe?

Jürgen Schmidhuber 22:54 Yes. Including all the, um, entanglement things and all the, um, spin up and down measurements, uh, that have been taken place, um, since 13.8 billion years ago and so on. Yeah. So we don't have a proof that it is, uh, random. We don't have a proof that it is compressible to a short program. But as long as we don't have that proof, we are obliged as scientists to keep looking for that simple explanation.

Lex Fridman 23:27 Absolutely. So you said simplicity is beautiful or beauty is simple, either one works, but you also work on curiosity, discovery, you know, the romantic notion of randomness, of serendipity, of, of, um, being surprised by things that are about you. Kind of in our poetic notion of reality, we think as humans require randomness. So you don't find randomness beautiful? You, you s- you find simple determinism beautiful?

Jürgen Schmidhuber 24:05 Yeah.

Lex Fridman 24:05 (laughs) Okay.

Jürgen Schmidhuber 24:07 So why?

Lex Fridman 24:08 Why?

Jürgen Schmidhuber 24:09 Because the explanation becomes shorter. A universe that is compressible to a short program is much more elegant and much more beautiful than another one which needs an almost infinite number of bits to be described. As far as we know...... many things that are happening in this universe are really simple in terms of, um, short programs that compute gravity and, uh, the interaction between elementary particles and so on. So all of that seems to be very, very simple. Every electron seems to reuse the same sub-program all the time-

Lex Fridman 24:50 Mm-hmm.

Jürgen Schmidhuber 24:51 ... as it is interacting with other elementary particles. (smacks lips) If we now required an extra oracle injecting new bits of information all the time for these extra things which are currently not understood such as beta decay...

Lex Fridman 25:16 Mm-hmm.

Jürgen Schmidhuber 25:18 ... then, um, the whole description length of the data that we can observe of, of the history of the universe would become, uh, much longer, and therefore uglier.

Lex Fridman 25:33 (laughs) And uglier. Again, s- simplicity is elegant and beautiful.

Jürgen Schmidhuber 25:38 All the history of science is a history of compression progress.

Lex Fridman 25:42 Yeah, so you, you've (smacks lips) described sort of as we build up abstractions and you've t- talked about the idea of, uh, compression, how do you see this, the history of science, the history of humanity or civilization and life on Earth as some kind of, uh, path towards greater and greater compression? What do you mean by that? How do you think about that?

Jürgen Schmidhuber 26:06 Hmm. Indeed, the history of science is a history of compression progress. What does that mean? Hundreds of years ago, there was an astronomer whose name was Kepler, and he looked at the data points that he got by watching planets move, and then he had all these data points and suddenly it turned out that he can greatly compress the data by predicting it through an ellipse law. So it turns out that all these data points are more or less on ellipses around the sun. And another guy came along whose name was Newton, and before him Hooke, and they said the same thing that is making these planets move like that is what makes the apples fall down, and it also holds for stones and for all kinds of other objects. And suddenly, many, many of these compressions, o- of these observations became much more compressible because as long as you can predict the next thing given what you have seen so far, you can compress it, but y- you don't have to store that data extra. This is called predictive coding. And then there was still something wrong with that theory of the universe, and you had deviations from these predictions of the theory, and 300 years later another guy came along whose name was Einstein, and he, um, he was able to explain away all these deviations from the, um, predictions-

Lex Fridman 27:49 Mm-hmm.

Jürgen Schmidhuber 27:49 ... of the old theory through a new theory, uh, which was called the General Theory of Relativity, which at first glance looks a little bit more complicated and you have to warp space and time, but you can phrase it within one single sentence which is, no matter how fast you accelerate and how fast or hard you, um, decelerate, and, um, no matter what is the, um, gravity in your local framework, light speed always looks the same. And from, from that you can calculate all the consequences, so it's a very simple thing, and it allows you to further compress all the observations because suddenly there are hardly any deviations any longer that you can measure from the predictions of this new theory. So all of science, um, is a history of compression progress. You never arrive immediately at the shortest, uh, explanation of the data, but you're making progress.

Lex Fridman 28:52 Mm-hmm.

Jürgen Schmidhuber 28:52 Whenever you are making progress, you have an insight. You see, oh, first I needed so many bits of information to describe the data, to describe my falling apples, my video of falling apples, I need so many data-

Lex Fridman 29:05 Mm-hmm.

Jürgen Schmidhuber 29:05 ... and so many pixels have to be stored, but then suddenly I realize, no, there is a very simple way of predicting the third frame in the video from the first two, and, um, and maybe not every little detail can be predicted, but more or less most of these orange blaks, uh, blobs that are coming down, they accelerate in the same way, which means that I can greatly compress the video. And the amount of compression, progress, that is the depth of the insight that you have

Jürgen Schmidhuber 29:36 at that moment, that's the fun that you have, the scientific fun, the fun in that discovery, and we can build artificial systems that do the same thing, that measure the depth of their insights as they are looking at the data which is coming in through their own experiments, and we give them a reward, an intrinsic reward in proportion to this depth of insight.

Lex Fridman 30:00 Mm-hmm.

Jürgen Schmidhuber 30:00 W- and since they are trying to maximize the, um, rewards they get, they are suddenly motivated to come up with new action sequences, with new experiments that have the property that the data that is coming in as a consequence of these experiments has the property that they can learn something about, see a pattern in there which they hadn't seen yet before.

Lex Fridman 30:28 S- there's an idea of power play that you've described-... uh, uh, training a general problem solver in this kind of way of looking for the unsolved problems.

Jürgen Schmidhuber 30:37 Yeah.

Lex Fridman 30:38 Can you describe that idea a little further?

Jürgen Schmidhuber 30:40 Yeah. It's another very simple idea. So normally, what you do in computer science, you have, um, you have some guy who gives you a problem, and then there is a- a huge, uh, search space of potential solution candidates. And you somehow try them out and, um, you have more or less sophisticated ways of, mm, moving around in that search space until you finally found a solution, uh, which you consider satisfactory.

Lex Fridman 31:11 Mm-hmm.

Jürgen Schmidhuber 31:13 That's what most of computer science is about. Power play just goes one little step further and says, "Let's not only search for solutions to a given problem, but let's search two pairs of problems and their solutions where the system itself has the opportunity to phrase its own problem." So we are looking suddenly at pairs of problems and their solutions or, uh, modifications of the problem solver that is supposed to generate a solution to that, um, new problem. And- and this additional, um, degree of freedom allows us to build courier systems that are like scientists in the sense that they not only try to solve and try to find answers to existing questions, no, they are also free to, um, pose their own questions. So if you want to build an artificial scientist, we have to give it that freedom and power play is exactly doing that.

Lex Fridman 32:19 So that's- that's a dimension of freedom that's important to have, but how do you th- how hard do you think that, how multi-dimensional and difficult the space of then coming up with your own questions is?

Jürgen Schmidhuber 32:34 Yeah.

Lex Fridman 32:34 So it's w- as, it's one of the things that as human beings we, uh, consider to be the thing that makes us special-

Jürgen Schmidhuber 32:39 Mm.

Lex Fridman 32:39 ... the intelligence that makes us special is that brilliant insight-

Jürgen Schmidhuber 32:44 Yeah.

Lex Fridman 32:45 ... that can create something totally new.

Jürgen Schmidhuber 32:47 Yes. So now let's look at the extreme case. Let's look at the set of all possible problems that you can formally describe-

Lex Fridman 32:57 Mm-hmm.

Jürgen Schmidhuber 32:58 ... which is infinite, which should be the next problem that a scientist or power play is going to solve. Well, it should be the easiest problem that goes beyond what you already know. So it should be the simplest problem that the current problem solver that you have, which can already solve 100 problems, that he cannot solve yet by just generalizing. So it has to be new, so it has to require a modification of the problem solver such that the new problem solver can solve this new thing, but the old problem solver cannot do it. And in addition to that, we have to make sure that the problem solver doesn't forget any of the previous solutions.

Lex Fridman 33:50 Right.

Jürgen Schmidhuber 33:51 And so by definition, power play is now trying always to search in this pair of-

Lex Fridman 33:56 Mm-hmm.

Jürgen Schmidhuber 33:56 ... in- in- in the set of pairs of problems and problem solver modifications for a combination that, uh, minimize the time-

Lex Fridman 34:06 Mm-hmm.

Jürgen Schmidhuber 34:06 ... to achieve these criteria. So it's always trying to find the problem which is easiest to add to the repertoire.

Lex Fridman 34:14 So just like grad students and, uh, academics and researchers can spend their whole career in a local minima-

Jürgen Schmidhuber 34:21 Mm-hmm.

Lex Fridman 34:22 ... stuck trying to, uh, come up with interesting questions-

Jürgen Schmidhuber 34:25 Mm-hmm.

Lex Fridman 34:25 ... but ultimately doing very little, do you think it's easy w- in this approach of looking for the simplest unsolvable problem to get stuck in a local minima and not never really discovering new, uh, you know, really jumping outside of the 100 problems that you've already solved-

Jürgen Schmidhuber 34:43 Mm-hmm.

Lex Fridman 34:44 ... in- in a genuine creative way?

Jürgen Schmidhuber 34:47 No, because that's the nature of power play, that it's always trying to break its current generalization abilities by coming s- up with a new problem which is beyond the current horizon.

Lex Fridman 35:00 I was gonna say.

Jürgen Schmidhuber 35:00 So just shifting the horizon of knowledge a little bit out there, breaking the existing rules such that the new thing becomes solvable but wasn't solvable by the old thing-

Lex Fridman 35:13 Mm-hmm.

Jürgen Schmidhuber 35:13 ... so like adding a new axiom, um, like what Godel did, uh, when he- he came up with these new sentences-

Lex Fridman 35:20 Mm-hmm.

Jürgen Schmidhuber 35:20 ... new theorems that didn't have a proof in the formal system, which means you can add them to the repertoire hoping that- that, uh, they, um, are not going to damage the consistency of the whole thing.

Lex Fridman 35:35 So in the, uh, paper with the amazing title Formal Theory of Creativity, Fun and Intrinsic Motivation you talk about discovery as intrinsic reward.

Jürgen Schmidhuber 35:47 Mm.

Lex Fridman 35:47 So if you view humans as intelligent agents, what do you think is the purpose and meaning of life for us humans? Is, you've talked about this discovery, uh, do you see humans a- as an instance of power play agents?

Jürgen Schmidhuber 36:04 Yeah. Uh, so humans, uh, are curious and, um, that means they behave like scientists, not only the official scientists but even the babies-

Lex Fridman 36:14 Mm-hmm.

Jürgen Schmidhuber 36:14 ... behave like scientists and they play around with their toys to figure out how the world works and how it is responding to their actions, and that's how they learn about gravity and everything.And, yeah, in 1990, we had the first systems like that. We would just try to, to play around with the environment and, uh, come up with situations that, um, go beyond what they knew at that time, and then get a reward for creating these situations and then becoming, um, more general problem solvers and being able to understand more of the world. So, yeah, I think in principle that, um, that, that curiosity, um, strategy or sophis- more sophisticated versions of what I just described, they are what we have built in as well because evolution discovered that's a good way of exploring-

Lex Fridman 37:10 Mm-hmm.

Jürgen Schmidhuber 37:10 ... the unknown world.

Lex Fridman 37:11 Beautiful.

Jürgen Schmidhuber 37:11 And a guy who explores the unknown world has a, a higher chance of, um, solving problems that he needs to survive in this world. On the other hand, those guys who were too curious, they were weeded out as well. So, you have to find this trade-off. Evolution found a certain trade-off. Apparently in our society, there are, um, is a certain percentage of extremely explorative guys and it doesn't matter if they die because many of the others are more conservative.

Lex Fridman 37:41 Mm-hmm.

Jürgen Schmidhuber 37:42 And, um, and, and so yeah, it, it would be surprising to me if, um, if that principle of artificial curiosity wouldn't be present in almost exactly the same form here.

Lex Fridman 37:58 In our brains. So, you're a bit of a musician and an artist. So, continuing in this topic of creativity, what do you think is the role of creativity in intelligence? So, you've kind of implied that it's essential for intelligence, if you think of intelligence as a problem-solving system, as ability to solve problems. But do you think it's essential-

Jürgen Schmidhuber 38:25 (laughs)

Lex Fridman 38:25 ... this idea of creativity?

Jürgen Schmidhuber 38:29 We never have a program, a sub-program that is called creativity or something. It's just a side effect of what our problem solvers do. They are searching a space of problems, or a space of, uh, candidates, of solution candidates until they hopefully find a solution to a given problem.

Lex Fridman 38:47 Mm-hmm.

Jürgen Schmidhuber 38:48 But then there are these two types of creativity and, uh, both of them are now present in our machines. Um, the first one has been around for a long time, which is human gives problem to machine, machine tries to, um, find a solution to that. And this has been happening for many decades. And for many decades, machines have found creative solutions-

Lex Fridman 39:09 Mm-hmm.

Jürgen Schmidhuber 39:09 ... to interesting problems where humans, um, were not aware of these, um, particularly creative solutions, but then appreciated that the machine found that.

Lex Fridman 39:20 Mm-hmm.

Jürgen Schmidhuber 39:21 The, the second is the pure creativity. That I would call, what I just mentioned, I would call the applied creativity.

Lex Fridman 39:28 Mm-hmm.

Jürgen Schmidhuber 39:29 Like applied art where somebody tells you, "Now, make a nice picture of, of this pope and you will get money for that." Okay. So, here is the artist and he makes a convincing picture of the pope and the pope likes it and gives him the money. And then there is the pure creativ- creativity which is more like the power play and the artificial curiosity thing where you have the freedom to select your own problem, like a scientist who defines his own question to study. And so, that is the pure creativity if you will.

Lex Fridman 40:07 And that-

Jürgen Schmidhuber 40:08 As opposed to the applied creativity which serves another.

Lex Fridman 40:14 And in that distinction, there's almost echoes of narrow AI versus general AI. So this kind of constrained painting of a pope seems like the, the approaches of what people are calling narrow AI.

Jürgen Schmidhuber 40:29 Mm-hmm.

Lex Fridman 40:29 And pure creativity seems to be... maybe I'm just biased as a human, but it seems to be an essential element of human-level intelligence. Is that what you're implying to a degree?

Jürgen Schmidhuber 40:46 If you zoom back a little bit and you just look at a general problem-solving machine which is trying to solve arbitrary problems, then this machine will figure out in the course of solving problems that it's good to be curious. So, all of what I said just now about this pre-wired curiosity-

Lex Fridman 41:04 Mm-hmm.

Jürgen Schmidhuber 41:05 ... and this will to invent new problems, um, that the system doesn't know how to solve yet should be just a byproduct of the general search. However, apparently evolution has built it into us because it turned out to be so successful, uh, uh, uh, pre-wiring a bias, a very successful exploratory bias that, um, that we are born with.

Lex Fridman 41:33 And you've also said that consciousness in the same kind of way may be a byproduct of, of problem-solving.

Jürgen Schmidhuber 41:40 Yeah.

Lex Fridman 41:41 Do you think, do you find it's an interesting byproduct? Do you think it's a useful byproduct? What are, what are your thoughts on consciousness in general? Or is it simply a byproduct of greater and greater capabilities of problem-solving that's, um, that's similar to creativity in that sense?

Jürgen Schmidhuber 42:01 Yeah. We never have a procedure called consciousness in our machines. However, we get as side effects of what these machines are doing things that seem to be closely related to what people call consciousness.

Lex Fridman 42:15 Mm-hmm.

Jürgen Schmidhuber 42:16 So for example, uh, in 1990, we had simple systems which were basically, um, recurrent networks and therefore universal computers trying to, mm, map incoming data into-... actions that lead to success. Uh, maximizing reward in a given environment, always finding the charging station in time whenever the battery is low and negative signals are coming from the battery-

Lex Fridman 42:43 Mm-hmm.

Jürgen Schmidhuber 42:44 ... always find the charging station in time without bumping against painful obstacles on the way. So complicated things, but very easily motivated.

Lex Fridman 42:53 Mm-hmm.

Jürgen Schmidhuber 42:54 And then, uh, we give these little guys a separate recurrent neural network, which is just predicting what's happening if I do that and that-

Lex Fridman 43:04 Mm-hmm.

Jürgen Schmidhuber 43:04 ... what will happen as a consequence of these actions that I'm executing, and it's just trained on the long and long history of interactions with the world. So it becomes a predictive model of the world, basically.

Lex Fridman 43:17 Mm-hmm.

Jürgen Schmidhuber 43:18 And therefore also a compressor of the observations of the world because whatever you can predict, you don't have to store extra. So compression is a side effect of prediction.

Lex Fridman 43:30 Mm-hmm.

Jürgen Schmidhuber 43:30 And how does this recurrent network compress? Well, it's inventing little subprograms, little subnet- networks that stand for everything that frequently appears in the environment-

Lex Fridman 43:41 Mm-hmm.

Jürgen Schmidhuber 43:42 ... like bottles, and microphones, and faces, maybe lots of faces in my, um, uh, environment. So I'm learning to create something like a prototype face, and a new face comes along, and all I have to encode are the deviations from the prototype. So it's compressing all the time the stuff that frequently appears. There's one thing that appears all the time, that is present all the time when the agent is interacting with its environment, which is the agent itself. So just for data compression reasons, it is extremely natural for this recurrent network to come up with little subnetworks that stand for the properties of the agents, the hand, uh, the, the, the other actuators, and all the stuff that you need to better encode the data, which is influenced by the actions of the agent. So there, just as a side effect of data compression during problem-solving, you have internal self-models. Now, you can use this model of the world to plan your future, and that's what you also have done since 1990. So the recurrent network, which is the controller, which is trying to maximize reward, can use this model of the network, uh, of the world, this model network of the world, this predictive model of the world, to plan ahead and say-

Lex Fridman 45:07 Mm-hmm.

Jürgen Schmidhuber 45:07 ... "Let's not do this action sequence. Let's do this action sequence instead because it leads to more predicted rewards."

Lex Fridman 45:14 Mm-hmm.

Jürgen Schmidhuber 45:14 And whenever it's waking up these little subnetworks that stand for itself, then it's thinking about itself. Then it's thinking about itself and it's exploring mentally the consequences of its own actions. And, and now you tell me what is still missing, um, uh-

Lex Fridman 45:37 Missing the next, the, the gap to consciousness.

Jürgen Schmidhuber 45:39 Yeah.

Lex Fridman 45:40 Uh, there, there isn't. That's a really beautiful idea that, um, you know, if, if life is a collection of data and, and life is a process of compressing that data to act, uh, efficiently, uh, in that data, you yourself appear very often. (laughs) So it's useful to, uh, form compressions of yourself. I mean, it's a really beautiful formulation of what consciousness is, is a necessary side effect. It's, uh, actually quite c- compelling

Lex Fridman 46:10 to me. You've described RNNs, developed, uh, LSTMs, long short-term memory networks that are, they're a type of recurrent neural networks that have gotten a lot of success recently. So these are networks that model the temporal aspects in the data, t- temporal patterns in the data, and you've called them the deepest of the neural networks, right? So what do you think is the value of depth in the models that we use to learn?

Jürgen Schmidhuber 46:43 Yeah. Since you mentioned the long short-term memory and the LSTM, um, I have to mention the names of the brilliant students who made that possible.

Lex Fridman 46:51 Yes, of course, of course.

Jürgen Schmidhuber 46:53 Um, first of all, my first student ever, Sepp Hochreiter, who had fundamental insights already in his diploma thesis. Then Felix Gears, who had additional important contributions. Alex Gray is a guy from Scotland who, um, uh, is mostly responsible for this, uh, CTC algorithm which is now often used to, to train, uh, the LSTM to do the speech recognition on all the Google Android phones and whatever, um, and CRV and so on. So, um, uh, these guys, without these guys, um-

Lex Fridman 47:25 Yeah.

Jürgen Schmidhuber 47:25 ... I would be nothing.

Lex Fridman 47:26 It's a lot of incredible work.

Jürgen Schmidhuber 47:29 What is now the depth? Uh, what is the importance of depth? Well, um, most problems in the real world are deep in the sense that, um, the current input doesn't tell you all you need to know about the environment.

Lex Fridman 47:44 Mm-hmm.

Jürgen Schmidhuber 47:45 So instead, um, you have to have a memory of what happened in the past. And often important parts of that memory are dated. They are pretty old.

Lex Fridman 47:56 Mm-hmm.

Jürgen Schmidhuber 47:56 So, um, when you're doing speech recognition, for example, and somebody says, "11,"-

Lex Fridman 48:02 Mm-hmm.

Jürgen Schmidhuber 48:04 ... then that's about half a second or something like that-

Lex Fridman 48:08 Mm-hmm.

Jürgen Schmidhuber 48:09 ... which means it's already 50 timesteps.

Lex Fridman 48:11 Mm-hmm.

Jürgen Schmidhuber 48:12 And another guy or the same guy sa- says, "Seven," so the ending is the same, -even.

Lex Fridman 48:18 Mm-hmm.

Jürgen Schmidhuber 48:18 But now the system has to see the distinction between seven and 11, and the only way it can see the difference is it has to store that, uh, 50 steps ago, there was an S or an L, 11 or a 7.... so there, you have already a problem of depth 50 because for each time step, you have something like a virtual, uh, layer in the expanded unrolled version of this recurrent network which is doing the speech recognition. So these long time lags, they translate into problem depth and most problems in this world are such that you really, um, have to look far back in time to understand what is the problem and to solve it.

Lex Fridman 49:06 But just like with LSTMs, you don't necessarily need to, when you look back in time, remember every aspect. You just need to r- remember the important aspects.

Jürgen Schmidhuber 49:14 That's right. The network has to learn to put the important stuff in- into memory and to ignore the unimportant noise.

Lex Fridman 49:23 So, but in that sense, deeper and deeper is better or is there a limitation? Is- is there... I mean, LSTM is- is one of the great examples of architectures that, uh, do something beyond just deeper and deeper networks. Uh, there's clever mechanisms for filtering data, for remembering and forgetting. Uh, so do you think th- that kind of thinking is necessary? If you think about LSTMs as a leap, a big leap forward over traditional vanilla RNNs, what do you think is the next leap-

Jürgen Schmidhuber 49:59 Hmm.

Lex Fridman 50:01 ... i- within this context? So LSTM is a very clever improvement, but LSTMs still don't have the same kind of ability to see far back in the futu- in the- in the past as us humans do, the credit assignment problem across way back-

Jürgen Schmidhuber 50:18 Hmm.

Lex Fridman 50:19 ... not just 50 time steps or 100 or 1,000, but millions and billions?

Jürgen Schmidhuber 50:24 It's not clear what are the practical limits of the LSTM when it comes to looking back. Already in 2006, I think, we had examples where it not only looked back tens or thousands of steps, but really millions of steps and, um, Juan Perez, um, Ortiz in my lab, I think was the first author of a paper where, um, we really, was it 2006 or something?

Lex Fridman 50:50 Mm-hmm.

Jürgen Schmidhuber 50:50 Had, uh, examples where it learned to look back, um, for more than 10 million steps.

Lex Fridman 50:56 Wow.

Jürgen Schmidhuber 50:57 So for most problems of speech recognition, it's not necessary to look that far back, but there are examples where it does. Now, the looking back thing, that's rather easy because there's only one past, but there are many possible futures.

Lex Fridman 51:17 Hm.

Jürgen Schmidhuber 51:17 And so a reinforcement learning system which is trying to maximize its future expected reward and doesn't know yet which of these many possible futures should I select given this one single past is facing problems that the LSTM by itself cannot solve. So the LSTM is good for coming up with a compact representation of the history so far, of the history and of observations and actions so far-

Lex Fridman 51:46 Mm-hmm.

Jürgen Schmidhuber 51:46 ... but now how do you plan in an efficient and good way among all these... How do you select one of these many possible action sequences that a reinforcement learning system has to consider to maximize reward in this unknown future? So again, it... we have this basic, um, setup where you have one recurrent network which gets in the video and the speech and whatever, and is executing the actions and is trying to maximize reward. So there is no teacher who tells it what to do at which point in time. And then there's the other network which is just predicting what's going to happen if I do them then, and that could be an LSTM network and it learns to look back all the way to make better predictions-

Lex Fridman 52:39 Mm-hmm.

Jürgen Schmidhuber 52:40 ... of the next time step. So essentially, although it's m- predicting only the next time step, it is motivated to learn to put into memory something that happened maybe a million steps ago because it's important, uh, to memorize that if you want to predict that, uh, at the next time step, the next event, you know? Now, um, how can a model of the world like that, a predictive model of the world be used by the first guy, let's call it the controller-

Lex Fridman 53:09 Mm-hmm.

Jürgen Schmidhuber 53:09 ... and the model, the controller and the model, how can the model be used by the controller to efficiently select amou- among these many possible futures? The naive way we had, um, about 30 years ago was let's just use the model of the world as a stand-in, as a simulation of the world and millisecond by millisecond, we plan the future-

Lex Fridman 53:32 Mm-hmm.

Jürgen Schmidhuber 53:32 ... and, um, that means we have to roll it out really in detail and it will work only if the model is really good and it will still be inefficient because we have to look at all these possible futures and- and there are so many of them. So instead, what we do now since 2015 in our CM systems, Controller Model systems, we give the controller the opportunity to learn by itself how to use the potentially relevant parts of the M, of the model network-

Lex Fridman 54:02 Mm-hmm.

Jürgen Schmidhuber 54:02 ... to solve new problems more quickly and if it wants to, it can learn to ignore the M and sometimes it's a good idea to ignore the- the M because it's really bad. It's a bad predictor in this particular, um, situation of life, uh, where the control is currently trying to maximize reward. However, it can also learn to address and exploit some of the subprograms that, uh, came about in the model network through compressing the data by predicting it.... so it now has an opportunity to reuse that code, the algorithmic information in the model network, to reduce its own search space, such that it can solve a new problem more quickly than without the model.

Lex Fridman 54:53 Compression. So, you're ultimately optimistic and excited about the power of RL, of reinforcement learning, in the context of real systems?

Jürgen Schmidhuber 55:05 Absolutely, yeah.

Lex Fridman 55:07 So you see RL as a potential having a huge impact beyond just sort of the M part that is often developed on, you know, supervised learning methods? You see RL a- a- as a, uh, for problems of self-driving cars or any kind of applied cyber robotics? That's the correct, interesting direction for research in your view?

Jürgen Schmidhuber 55:34 I do think so. We have a company called Nescens-

Lex Fridman 55:37 Nescens.

Jürgen Schmidhuber 55:38 ... which, um, has applied reinforcement learning to little Audis.

Lex Fridman 55:44 Little Audis.

Jürgen Schmidhuber 55:45 Which learn to park without a teacher. The same principles were used, of course. So these little Audis, they are small, maybe like that, so much smaller than the real Audis, but they have all the sensors, uh, that you find in the real Audis, you find the cameras, the LiDAR sensors. They go up to 120-20 kilometers an hour if you, if, if they want to.

Lex Fridman 56:08 (laughs)

Jürgen Schmidhuber 56:09 And, um, and they have, um, pain sensors basically and they don't want to bump against obstacles and other Audis, and so they, um, must learn like little babies to park. Take the raw vision input and translate that into actions that lead to successful parking behavior, which is a rewarding thing. And yes, they learn that.

Lex Fridman 56:32 They learn, successfully.

Jürgen Schmidhuber 56:33 Um, so we have examples like that, and it's only in the beginning. Um, you know, this is just the tip of the iceberg, and I believe the next wave of AI is going to be all about that. Uh, so at the moment, the current wave of AI is about passive pattern observation and, uh, prediction and, um, and that's what you have on your smartphone, and what the major companies on the Pacific Rim are using to sell you ads, to do marketing.

Lex Fridman 57:02 (laughs) Yeah.

Jürgen Schmidhuber 57:02 That's the current, uh-

Lex Fridman 57:03 Yes.

Jürgen Schmidhuber 57:03 ... source of profit in AI, and that's only one or 2% of the world economy, um, which is big enough to make these companies pretty much the most valuable companies in the world. But there's a much, much bigger fraction of the economy going to be affected by the next wave, which is really about machines that shape the data through their own actions. And-

Lex Fridman 57:28 Do you think simulation is ultimately the biggest, uh, way that, that th- those methods will be successful in the next 10, 20 years? We're not talking about 100 years from now.

Jürgen Schmidhuber 57:38 Yeah.

Lex Fridman 57:38 We're talking about sort of the near term impact of RL. Do you think really good simulation is required? Or is there other t- techniques like imitation learning, you know, observing other, m- humans-

Jürgen Schmidhuber 57:52 Yeah.

Lex Fridman 57:52 ... operating in the real world? Where, where do you think, uh, the success will come from?

Jürgen Schmidhuber 57:57 So at the moment, we have a tendency of using physics simulations to learn behavior for machines that, um, learn to solve problems that, uh, humans also do not know how to solve.

Lex Fridman 58:12 Mm-hmm.

Jürgen Schmidhuber 58:14 However, this is not the future, because the future is in what little babies do. They don't use a physics engine to simulate the world. No. They learn a predictive model of the world which, um, maybe sometimes is wrong in many ways, but captures all kinds of important abstract high level predictions which are really important to, uh, be successful. And, um, and that's what is, what was the future 30 years ago when we started that type of research, but it's still the future and now we know much better how to go there, um, uh, to, to move there, to move forward, uh, and to, uh, really make working systems based on that where you have a learning model of the world, a model of the world that learns to predict what's going to happen if I do that and that. And then, uh, the controller uses that model to, um, more quickly learn successful action sequences. And then of course always this curiosity thing. In the beginning the model is stupid so the controller should be motivated to come up with experiments, with action sequences that lead to data that improve the model.

Lex Fridman 59:24 Uh, do you think, uh, improving the model, constructing an understanding of the world in this connection is the, in, uh, now the p- the popular approaches have been successful are y- you know, grounded in ideas of neural networks. But in the '80s with expert systems there's symbolic AI approaches which, uh, to us humans are more intuitive in the sense that it makes sense that you build up knowledge and this k- knowledge representation. What kind of lessons can we draw in our current approaches-

Jürgen Schmidhuber 59:55 Mm-hmm.

Lex Fridman 59:55 ... in, for, from expert systems, from symbolic AI?

Jürgen Schmidhuber 1:00:00 So I, um, b- became aware of all of that in the '80s, and back then a logic program, logic programming was a huge thing.

Lex Fridman 1:00:09 Was it inspiring to you yourself? Did you find it compelling? Because most, a lot of your work was, uh, not so much in that realm, right? It was more in the learning systems.

Jürgen Schmidhuber 1:00:18 Yes and no, but we did all of that. Uh-

Lex Fridman 1:00:20 You did.

Jürgen Schmidhuber 1:00:20 ... so we, we... My first, um, publication ever actually was, um-... 1987 was, uh, the implementation of, um, genetic algorithm of a genetic programming system in Prolog. So, Prolog, that's what you learned back then, which is a logic programming language. And the Japanese, they had this huge fifth generation AI project, which was mostly about, uh, logic pro- programming back then. Although neural networks ex- existed and were well-known back then, and deep learning has existed since, um, 1965, um, since this guy in the Ukraine, um, Ivakhnenko, started it. But, um, uh, the Japanese and many other people, uh, they focused really on this logic programming, and I was influenced to the extent that I said, "Okay, let's take these biologically inspired algorithms like evolution-"

Lex Fridman 1:01:15 Mm-hmm.

Jürgen Schmidhuber 1:01:15 "... uh, programs, uh, and, um, and, and, mm, and implement that in the language which I know, which was Prolog, for example, back then."

Lex Fridman 1:01:25 Yeah.

Jürgen Schmidhuber 1:01:25 And then, um, in, in many ways this came back later because, uh, the Gödel machine, for example, has a proof searcher on board, and without that, it would not be optimal. While Marcus Hutter's, uh, universal algorithm for solving all well-defined problems has a proof search on board. So, that's very much logic programming.

Lex Fridman 1:01:45 Mm-hmm.

Jürgen Schmidhuber 1:01:46 Without that, it would not be asymptotically optimum. But then on the other hand, because we are very pragmatic guys also, um, we, we focused on recurrent neural networks and, and, and suboptimal, uh, stuff such as gradient-based search and program space, rather than provably optimal things.

Lex Fridman 1:02:09 The logic programming does it ha- certainly has a usefulness in, uh, when you're trying to construct something provably optimal or provably good or s- something like that, but is it useful for, for practical problems?

Jürgen Schmidhuber 1:02:22 It's really useful for theorem proving.

Lex Fridman 1:02:24 Theorem. (laughs)

Jürgen Schmidhuber 1:02:24 The best theorem provers today are not, uh, neural networks.

Lex Fridman 1:02:27 Right.

Jürgen Schmidhuber 1:02:28 No. They are, uh, logic programming systems, and they are much better theorem provers than most, uh, math students in their first or second semester.

Lex Fridman 1:02:37 Mm-hmm. But for reasoning, to, for playing games of Go or chess, or for robots, autonomous vehicles that operate in the real world-

Jürgen Schmidhuber 1:02:46 Yeah.

Lex Fridman 1:02:47 ... or, uh, object manipulation-

Jürgen Schmidhuber 1:02:49 Yeah.

Lex Fridman 1:02:49 ... you think learning...

Jürgen Schmidhuber 1:02:51 Yeah, as long as the problems have little to do, um, with, um, with, um, theorem proving-

Lex Fridman 1:02:57 Yeah.

Jürgen Schmidhuber 1:02:57 ... uh, themselves, then, um, as long as that is not the case, um, you, you just want to have better pattern recognition. So, to build a self-driving car, you want to have better pattern recognition and, um, and, uh, pedestrian recognition and all these things, and you want to, uh, you mini- you want to minimize the number of false positives, which is currently slowing down self-driving cars in many ways. And, um, and all of that has very little to do with logic programming. Yeah.

Lex Fridman 1:03:27 What are you most excited about in terms of directions of artificial intelligence at this moment in the next few years in your own research and in the broader community?

Jürgen Schmidhuber 1:03:41 So, I think in the not-so-distant future, we will have for the first time little robots that learn like kids. And I will be able to say to the robot, um, "Look here, robot, we are going to assemble a smartphone."

Lex Fridman 1:04:00 Mm-hmm.

Jürgen Schmidhuber 1:04:00 "Let's take this slab of plastic, um, and the screwdriver, and let's screw in the screw like that," you know? No, not like that, like that.

Lex Fridman 1:04:11 Mm-hmm.

Jürgen Schmidhuber 1:04:11 Not like that, like that. And I don't have a data glove or something. He will see me, and he will hear me, and he will try to do something with h- his own actuators, which will be really different from mine. But he will understand the difference and will learn to imitate me, but not in the supervised way-

Lex Fridman 1:04:33 Mm-hmm.

Jürgen Schmidhuber 1:04:34 ... um, where a teacher is giving, um, target signals for all his muscles all the time. No, by doing this high-level imitation where he first has to learn to imitate me, and then to interpret these additional noises coming from my mouth, um, as helping, helpful signals to-

Lex Fridman 1:04:52 Mm-hmm.

Jürgen Schmidhuber 1:04:53 ... to do that better. And then it will, mm, by itself come up with, uh, faster ways and more efficient ways of doing the same thing. And finally, I stop his learning algorithm and make a million copies and sell it.

Lex Fridman 1:05:10 (laughs)

Jürgen Schmidhuber 1:05:11 And so at the moment, this is not possible, but we already see how we are going to get there. And you can imagine to the extent that this works economically and cheaply, it's going to change everything. Almost all of production is going to be affected by that. And a much bigger wave, a much bigger AI wave is coming than the one that we are currently witnessing, which is mostly about passive pattern recognition on your smartphone. This is about active machines that shapes the data through the actions they are executing, and they learn to do that in a good way. So, many of the traditional industries are going to be affected by that. All the companies that are building machines will equip these machines with cameras and other sensors, and they are going to learn to solve all kinds of problems through interaction with humans, but also a lot on their own, uh, to improve what they already can do. And lots of old economy is going to be affected by that. And in recent years, I have seen that old economy is actually waking up and realizing that this is the case.... and, um-

Lex Fridman 1:06:32 Are you optimistic about that future? Are you concerned? Uh, there's a lot of people concerned on, in the near term about the transformation of the nature of work. The kind of ideas that you just suggested would have a significant impact of what kind of things could be automated. Are you optimistic about that future, or are you nervous about that future? And looking a little bit farther into the future, there's people like Elon Musk, uh, Stuart Russell, concerned about the existential threats-

Jürgen Schmidhuber 1:07:04 Mm.

Lex Fridman 1:07:05 ... of that future.

Jürgen Schmidhuber 1:07:06 Mm.

Lex Fridman 1:07:06 So, in the near term, job loss, in the long term, existential threat. Are these concerns to you, or are you ultimately optimistic?

Jürgen Schmidhuber 1:07:15 So, let's first address the near future. We have had predictions of job losses for many decades. For example, when industrial robots came along-

Lex Fridman 1:07:31 Mm-hmm.

Jürgen Schmidhuber 1:07:33 ... many pe- many people predicted lots of jobs are going to get lost. And, in a sense, they were right.

Lex Fridman 1:07:42 Mm-hmm.

Jürgen Schmidhuber 1:07:42 Because back then, there were, um, car factories and hundreds of people, and these factories assembled cars. And today, the same car factories have hundreds of robots and maybe three guys watching the robots. On the other hand, those countries that have lots of, uh, robots per capita, Japan, Korea, Germany, Switzerland, couple of other countries, they have really low unemployment rates. Somehow, all kinds of new jobs were created. Back then, nobody anticipated those jobs. And, um, decades ago, I already said, it's really easy to, um, say which jobs are going to get lost, but it's really hard to predict the new ones. Thirty years ago, who would have predicted all these people making money as, uh, YouTube bloggers, for example?

Lex Fridman 1:08:45 Mm.

Jürgen Schmidhuber 1:08:45 200 years ago, 60% of all people used to work in agriculture.

Lex Fridman 1:08:52 Mm-hmm.

Jürgen Schmidhuber 1:08:53 Today, maybe 1%. But still, only, I don't know, 5% unemployment. Lots of new jobs were created, and Homo Ludens, the, the playing man, is inventing new jobs all the time. Most of these jobs are not existentially necessary-

Lex Fridman 1:09:15 Mm.

Jürgen Schmidhuber 1:09:15 ... for the survival of our species. There are only very few existentially necessary jobs, such as farming and building houses, and, and warming up the houses, but less than 10% of the population is doing that. And most of these newly, um, invented jobs are about, um, interacting with other people in new ways, through new media and so on, getting new ki- types of kudos and forms of likes and whatever, and even making money through that. So, Homo Ludens, the playing man, doesn't want to be unemployed, and that's why he's inventing new jobs all the time, and he keeps considering these jobs as really important and is investing a lot of energy and hours of work into, into those new jobs.

Lex Fridman 1:10:08 That's, uh, quite beautifully put. We're really nervous about the future because we can't predict what kind of new jobs will be created-

Jürgen Schmidhuber 1:10:14 Mm-hmm.

Lex Fridman 1:10:15 ... but you're ul- uh, ultimately optimistic that we, uh, humans are so restless that we create and give meaning to newer and newer jobs, totally new, uh, cl- likes on Fa- things that get likes on Facebook or whatever the social platform is. So, what about long-term existential threat of AI, where our whole civilization may be swallowed up by this ultra super intelligent systems?

Jürgen Schmidhuber 1:10:45 Maybe it's not going to be swallowed up, but, um, I'd be surprised if, uh, we were, uh, we humans were the last step in the evolution of the universe. And, um-

Lex Fridman 1:11:00 You- you've actually had this beautiful comment somewhere that, uh, I've seen saying that, uh, artificial... (laughs) Uh, quite insightful. So, artificial general intelligence systems, just like us humans, will likely not want to interact with humans. They'll just interact amongst themselves, just like ants interact amongst themselves, and only, uh, tangentially interact with humans.

Jürgen Schmidhuber 1:11:24 Mm-hmm.

Lex Fridman 1:11:25 And it's- it's quite an interesting idea that once we create AGI, they will lose interest in humans and- and have- compete for their own Facebook likes on their own social platforms.

Jürgen Schmidhuber 1:11:36 Mm-hmm.

Lex Fridman 1:11:36 So, within that, uh, quite elegant idea, how do we know, in a hypothetical sense, that there's not already intelligent systems out there? How do you think broadly of general intelligence greater than us? How would we know it's out there?

Jürgen Schmidhuber 1:11:56 Mm.

Lex Fridman 1:11:56 How would we know it's around us, and could it already be?

Jürgen Schmidhuber 1:12:02 I'd be surprised if n- with- within the next few decades or something like that, we, um, we won't have AIs that are truly smart in every single way and better problem solvers in almost every single important way. And I'd be surprised if they wouldn't realize what we have realized a long time ago, which is that, um, almost all physical resources are not here in this biosphere, but further out.... the rest of the solar system gets two billion times more solar energy than our little planet. There's lots of material out there that you can use to build robots and self-replicating robot factories and all this stuff. And they are going to do that, and they will be scientists, um, and curious, and they will explore what they can do. And in the beginning, they will be fascinated by life and by their own origins in our civilization. They will want to understand that completely, just like, uh, people today would like to understand how life works and, um, and also, um, the history of our own, um, existence and civilization, but then also the physical laws that created all of that. So they, um, in the beginning, they will be fascinated by life. Once they understand it, they lose interest, um, like anybody who loses interest in things he understands. And then, as you said, um, the most interesting sources of, um, information for them will be others of their own kind.

Lex Fridman 1:13:53 Mm-hmm.

Jürgen Schmidhuber 1:13:58 So at least in the long run, there seems to be some sort of protection through lack of interest on the other side. And, um, and now it seems also clear as far as we understand physics, you need matter and energy to compute and to build more robots and infrastructure and more AI civilization and A-AI ecologies consisting of trillions of different types of AIs. And so it seems inconceivable to, to me that this thing is not going to expand. Some AI ecology not controlled by one AI, but, but by trillions of different types of AIs competing in all kinds of quickly evolving and disappearing ecological niches in ways that we cannot fathom at the moment. But it's going to expand, limited by light speed and physics, but it's going to expand and, and now we realize that the universe is still young. It's only 13.8 billion years old, and it's going to be a thousand times older than that. So there's plenty of time to conquer the entire universe and to fill it with intelligence and senders and receivers such that AIs can travel the way they are traveling in our labs today, which is by radio from sender to receiver. And let's call the current age of the universe one eon. One eon. Now, it will take just a few eons from now and the entire visible universe is going to be full of that stuff. And let's look ahead to a time when the universe is going to be 1,000 times older than it is now. They will look back and they will say, "Look, almost immediately after the Big Bang, only a few eons later, the entire universe started to become intelligent." Now to your question, how do we see whether anything like that has already happened or is already in a more advanced stage in some other part of the universe, of the visible universe? We are trying to look out there and nothing like that has happened so far, or is that true?

Lex Fridman 1:16:22 Do you think we would recognize it? Well, how do we know it's not among us?

Jürgen Schmidhuber 1:16:25 Yeah.

Lex Fridman 1:16:25 How do we know planets aren't, in themselves, intelligent beings?

Jürgen Schmidhuber 1:16:30 Yeah.

Lex Fridman 1:16:30 How do we know, uh, ants, um, seen as a collective are not much greater intelligence-

Jürgen Schmidhuber 1:16:39 Yeah.

Lex Fridman 1:16:39 ... than our own? These kinds of ideas.

Jürgen Schmidhuber 1:16:41 Yeah. When I was a boy, I was thinking about these things and I thought, hmm, maybe it has already happened because back then I know, I knew, I learned from popular physics books that, um, the structure, the large-scale structure of the universe is not homogeneous. And you have these clusters of galaxies, and then in between there are these huge empty spaces. And I thought, hmm, maybe they aren't really empty. It's just that in the middle of that, some AI civilization already has expanded and then has covered a bubble of a billion light years diameter and is using all the energy, uh, of all the stars within that bubble for its own unfathomable purposes. And so it already has happened, and we just, um, fail to interpret the signs. But then I learned, uh, that gravity by itself explains the large-scale structure of the universe and that this is not a convincing explanation. And then I thought maybe, maybe it's the dark matter because as far as we know today, 80% of the measurable matter is invisible. And we know that because otherwise our galaxy or other galaxies would fall apart. They would... They are rotating too quickly. And, um, then the idea was maybe all of these AI civilizations that are already out there, they, they, they are just invisible because they're really efficient in using the energies of their own, uh, local systems and that's why they appear dark to us. But this is also not a convincing explanation because then the question becomes why is there... Are there still any, uh, visible stars left in our own galaxy, which also must have a lot of dark matter? So that is also not a convincing thing. And today, (laughs) I like to, um, think it's quite plausible that maybe we are the first, at least in our local light cone, within, um, the few hundreds of millions of light years that we can reliably observe.

Lex Fridman 1:19:08 Observe. Is that exciting to you, that we might be the first?

Jürgen Schmidhuber 1:19:12 And, um, it would make us much more important because if we mess it up through a nuclear war, then, then maybe this will have an effect on the, on the, on the development on, of the entire universe.

Lex Fridman 1:19:31 So let's not mess it up.

Jürgen Schmidhuber 1:19:32 Let's not mess it up.

Lex Fridman 1:19:33 Jürgen, thank you so much for talking today. I really appreciate it.

Jürgen Schmidhuber 1:19:37 It's my pleasure.