Pod Insights Lex Fridman Podcast François Chollet: Keras, Deep Learning, and the Progress of AI About this episode François Chollet is the creator of Keras, which is an open source deep learning library that is designed to enable fast, user-friendly experimentation with deep neural networks. It serves as an interface to several deep learning libraries, most popular of which is TensorFlow, and it was integrated into TensorFlow main codebase a while back. Aside from creating an exceptionally useful and popular library, François is also a world-class AI researcher and software engineer at Google, and is definitely an outspoken, if not controversial, personality in the AI world, especially in the realm of ideas around the future of artificial intelligence. This conversation is part of the Artificial Intelligence podcast. If you would like to get more information about this podcast go to https://lexfridman.com/ai or connect with @lexfridman on Twitter https://twitter.com/lexfridman, LinkedIn https://www.linkedin.com/in/lexfridman/, Facebook https://www.facebook.com/lexfridman, Medium https://medium.com/@lexfridman, or YouTube https://www.youtube.com/lexfridman where you can watch the video versions of these conversations. If you enjoy the podcast, please rate it 5 stars on iTunes or support it on Patreon https://www.patreon.com/lexfridman.
Transcript 17,952 words · 2 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Lex Fridman? The following is a conversation with Francois Chollet. He's the creator of Keras, which is an open source deep learning library that is designed to enable fast, user-friendly experimentation with deep neural networks. It serves as an interface to several deep learning libraries, most popular of which is TensorFlow, and it was integrated into the TensorFlow main codebase a while ago. Meaning, if you want to create, train, and use neural networks, probably the easiest and most popular option is to use Keras inside TensorFlow. Aside from creating an exceptionally useful and popular library, Francois is also a world-class AI researcher and software engineer at Google.
0:44 Lex Fridman? And he's definitely an outspoken, if not controversial, personality in the AI world, especially in the realm of ideas around the future of artificial intelligence. This is the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, give it five stars on iTunes, support it on Patreon, or simply connect with me on Twitter at LexFriedman, spelled F-R-I-D-M-A-N. And now, here's my conversation with Francois Chollet.
1:32 Lex Fridman? You're known for not sugarcoating your opinions and speaking your mind about ideas in AI, especially on Twitter. It's one of my favorite Twitter accounts. So what's one of the more controversial ideas you've expressed online and gotten some heat for? How do you pick?
1:50 Speaker 01 How do I pick? Yeah, no, I think if you have, if you go through the trouble of maintaining a Twitter account, you might as well speak your mind, you know? Otherwise, what's even the point of having a Twitter account? It's like having a nice car and just leaving it in the garage. Yeah, so what's one thing for which I got? A lot of pushback. Perhaps, you know, that time I wrote something about the idea of intelligence explosion. And I was questioning the idea and the reasoning behind this idea. And I got a lot of pushback on that. I guess I had a flag for it. So, yeah, so intelligence explosion, I'm sure you're familiar with the idea, but it's the idea that
2:34 Speaker 01 if you were to build general AI problem-solving algorithms. Well, the problem of building such an AI, that itself is a problem that could be solved by your AI. And maybe it could be solved better than what humans can do. So your AI could start tweaking its own algorithm, could start making a better version of itself, and so on iteratively, in a recursive fashion. And so you would end up with an AI with exponentially increasing intelligence.
3:07 Lex Fridman? That's right.
3:07 Speaker 01 And I was basically questioning this idea. First of all, because the notion of intelligence explosion uses an implicit definition of intelligence that doesn't sound quite right to me. It considers intelligence as a property of a brain that you can consider in isolation, like the height of a building, for instance.
3:32 Lex Fridman? Right.
3:33 Speaker 01 But that's not really what intelligence is. Intelligence emerges from the interaction between a brain, a body, like embodied intelligence, and an environment. And if you're missing one of these pieces, then you cannot really define intelligence anymore. So just tweaking a brain to make it smaller and smaller doesn't actually make any sense to me.
3:56 Lex Fridman? So first of all, you're crushing the dreams of many people, right? So there's a, let's look at like Sam Harris. Actually, a lot of physicists, Max Tegmark, people who think, you know, the universe is an information processing system. Our brain is kind of an information processing system. So what's the theoretical limit? Like, it doesn't make sense that there should be some...
4:21 Lex Fridman? It seems naive to think that our own brain is somehow the limit of the capabilities and disinformation. I'm playing devil's advocate here. This information processing system. And then if you just scale it, if you're able to build something that's on par with the brain, you just, the process that builds it just continues and it'll improve exponentially. So that, that's the logic that's used actually by almost everybody that is worried about superhuman intelligence.
4:53 Speaker 01 Yeah.
4:54 Lex Fridman? So you're, you're trying to make, so most people who are skeptical of that are kind of like, this doesn't, their thought process, this doesn't feel right. Like that's for me as well. So I'm more like, it doesn't,
5:08 Lex Fridman? the whole thing is shrouded in mystery where you, you can't really say anything concrete, but you could say this doesn't feel right. This doesn't feel like that's how the brain works. And you're trying to, with your blog post and now making it a little more explicit. So one idea is that the brain isn't exist alone. It exists within the environment. So you can't exponentially, you would have to somehow exponentially improve the environment and the brain together almost, yeah, in order to create something that's much smarter in some kind of, of course, we don't have a definition of intelligence.
5:47 Speaker 01 That's correct. That's correct. I, I don't think, you should look at very smart people today, even humans, not even talking about AIs. I don't think their brain and the performance of their brain is the bottleneck to their, to their expressed intelligence, to their achievements.
6:04 Speaker 01 You cannot just tweak one part of this system, like of this brain, body, environment system, and expect that capabilities, like what emerges out of this system to just, you know, explode exponentially. Because anytime you improve one part of a system with many intelligences like this, there's a new bottleneck that arises, right? And I don't think even today, for very smart people, their brain is not the bottleneck to the sort of problems they can solve, right? In fact, many very smart people today, you know, they're not actually solving any big scientific problems. They're not Einstein. They're like Einstein, but you know, the, the patent clerk days.
6:45 Speaker 01 Like Einstein became Einstein because this was a meeting of
6:52 Speaker 01 a genius with a big problem at the right time, right? But maybe this meeting could have never happened. And then Einstein, there's just been a pattern clerk, right? And in fact, many people today are probably like genius level smart, but you wouldn't know because they're not really expressing any of that.
7:11 Lex Fridman? Well, that's brilliant. So we can think of the world, Earth, but also the universe as just as a space of problems. So all these problems and tasks are roaming it of various difficulty. And there's agents, creatures like ourselves and animals and so on, that are also roaming it. And then you you get coupled with a problem and then you solve it. But without that coupling, you can't demonstrate your quote unquote intelligence.
7:39 Speaker 01 Exactly. Intelligence is the meeting of great problem solving capabilities with a great problem. And if you don't have the problem, you don't really express an intelligence. All you're left with is potential intelligence, like the performance of your brain or, you know, how high your IQ is, which in itself is just a number, right?
7:59 Lex Fridman? So you mentioned problem solving capacity.
8:03 Speaker 01 Yeah.
8:04 Lex Fridman? What what do you think of as problem solving capacity? What can you try to define intelligence? Like, what does it mean to be more or less intelligent? Is it completely coupled to a particular problem? Or is there something a little bit more universal?
8:22 Speaker 01 Yeah, I do believe all intelligence is specialized intelligence. Even human intelligence has some degree of generality. Well, all intelligent systems have some degree of generality. They're always specialized in in one category of problems. So the the human intelligence is specialized in the human experience. And that shows at various levels. That shows in some prior knowledge that's innate that we have at birth.
8:49 Speaker 01 Knowledge about things like agents, goal-driven behavior, visual priors about what makes an object, priors about time, and so on. That shows also in the way we learn. For instance, it's very, very easy for us to pick up language. It's very, very easy for us to learn certain things because we are basically hardcoded to learn them. And we are specialized in solving certain kinds of problems. And we are quite useless when it comes to other kinds of problems. For instance, we are not really designed to handle very long-term problems. We have no capability of seeing the very long term.
9:29 Speaker 01 We don't have very much working memory, you know.
9:35 Lex Fridman? So how do you think about long-term? Do you think long-term planning, or we're talking about scale of years, millennia? What do you mean by long-term we're not very good?
9:45 Speaker 01 Well, human intelligence is specialized in the human experience. And human experience is very short, like one lifetime is short. Even within one lifetime, we have a very hard time envisioning, you know, things on a scale of years. Like it's very difficult to project yourself at a scale of five years, a scale of 10 years, and so on.
10:03 Lex Fridman? Right.
10:04 Speaker 01 We can solve only fairly narrowly scoped problems. So when it comes to solving bigger problems, larger scale problems, we are not actually doing it on an individual level. So it's not actually our brain doing it. We have this thing called civilization, right? Which is itself a sort of problem-solving system, a sort of artificial intelligence system, right? And it's not running on one brain, it's running on a network of brains. In fact, it's running on much more than a network of brains. It's running on a lot of infrastructure, like books and computers and the internet and human institutions and so on.
10:42 Speaker 01 And that is capable of handling problems on a much greater scale than any individual human. If you look at some... Computer science, for instance, that's an institution that solves problems and it is superhuman, right? It operates on a greater scale, it can solve much bigger problems than an individual human could. And science itself, science as a system, as an institution, is a kind of artificially intelligent problem-solving algorithm that is superhuman.
11:15 Lex Fridman? Yeah, it's a, well, this computer science is like a theorem prover at a scale of thousands, maybe hundreds of thousands of human beings. At that scale, what do you think is an intelligent agent? So there's us humans at the individual level, there is millions, maybe billions of bacteria on our skin, there is, that's at the smaller scale, you can even go to the particle level as systems that behave, you can say intelligently in some ways, and then you can look at the Earth as a single organism, you can look at our galaxy, and even the universe as a single organism.
11:59 Lex Fridman? Do you think, how do you think about scale and defining intelligent systems? And we're here at Google, there is millions of devices doing computation in a distributed way. How do you think about intelligence as a scale?
12:12 Speaker 01 You can always characterize anything as a system.
12:16 Lex Fridman? Right.
12:17 Speaker 01 I think people who talk about things like intelligence explosion tend to focus on one agent is basically one brain, like one brain considered in isolation, like a brain, a jar, that's controlling a body in a very like top-to-bottom kind of fashion, and that body is pursuing goals into an environment. So it's a very hierarchical view, you have the brain at the top of the pyramid, then you have the body just plainly receiving orders, and then the body is manipulating objects in the environment and so on. So everything is subordinate to this one thing, this epicenter, which is the brain. But in real life, intelligent agents don't really work like this, right? There is no strong delimitation between the brain and the body to start with. You have to look not just at the brain, but at the
13:02 Speaker 01 Nervous system, but then the nervous system and the body are naturally two separate entities. So you have to look at an entire animal as one agent. But then you start realizing as you observe an animal over a, a, a, a, any length of time, that a lot of the intelligence of an animal is actually externalized. That's especially true for humans. A lot of our intelligence is externalized. When you write down some notes, that is externalized intelligence. When you write a computer program, you are externalizing cognition. So it's externalized in books, it's externalized in, in computers, the internet, in other humans.
13:40 Speaker 01 It's externalized in language and so on. So it's, there is no like hard delimitation of what makes an intelligent agent. It's all about context.
13:52 Lex Fridman? Okay, but AlphaGo is better at Go than the best human player.
13:59 Lex Fridman? You know, there's levels of skill here. So do you think there's such a ability as, such a concept as a intelligence explosion in a specific task? And then, well, yeah, do you think it's possible to have a category of tasks on which you do have something like an exponential growth of ability to solve that particular problem?
14:24 Speaker 01 I think if you consider a specific vertical, it's probably possible to some extent.
14:32 Speaker 01 I also don't think we have to speculate about it because we have real world examples of recursively self-improving intelligent systems, right? As far as in science, is a problem solving system, a knowledge generation system, like a system that experiences the world in some sense and then gradually understands it and can act on it. And that system is superhuman and it is clearly recursively self-improving. Because science feeds into technology, technology can be used to build better tools, better computers, better instrumentation and so on, which in turn can make science faster, right?
15:13 Speaker 01 So science is probably the closest thing we have today to a recursively self-improving superhuman AI. And you can just observe, you know, Is scientific progress today exploding? Which, you know, itself is an interesting question. You can use that as a basis to try to understand what will happen with a superhuman AI that has a science-like behavior.
15:38 Lex Fridman? Let me linger on it a little bit more. What is your intuition why an intelligence explosion is not possible? Like, taking the scientific, all the semi-scientific revolutions, why can't we slightly accelerate that process?
15:55 Speaker 01 So you can absolutely accelerate any problem-solving process. So recursive search improvement is absolutely a real thing. But what happens with a recursively self-improving system is typically not explosion, because no system exists in isolation. And so tweaking one part of the system means that suddenly another part of the system becomes a bottleneck. And if you look at science, for instance, which is clearly recursively self-improving, clearly a problem-solving system, scientific progress is not actually exploding. If you look at science, what you see is the picture of a system that is consuming an exponentially increasing amount of resources.
16:37 Speaker 01 But it's having a linear output in terms of scientific progress. And maybe that will seem like a very strong claim. Many people are actually saying that, you know, scientific progress is exponential. But when they're claiming this, they're actually looking at indicators of resource consumptions, resource consumption by science. For instance, the number of papers being published, the number of patents being filed and so on, which are just completely correlated with how many people are working on science today.
17:14 Lex Fridman? Yeah.
17:15 Speaker 01 Right? So it's actually an indicator of resource consumption. But what you should look at is the output, is progress in terms of the knowledge that science generates, in terms of the scope and significance of the problems that we solve. And some people have actually been trying to measure that. Like Michael Nielsen, for instance. He had a very nice paper, I think it was last year, about it.
17:42 Speaker 01 So his approach to measure scientific progress was to look at the timeline of scientific discoveries over the past, you know, 100, 150 years. And for each major discovery, ask a panel of experts to rate the significance of the discovery. And if the output of sciences institution were exponential, you will expect the temporal density of significance to go up exponentially. Maybe because there's a faster rate of discoveries, maybe because the discoveries are, you know, increasingly more important. And what actually happens if you if you plot this temporal density of significance measured
18:27 Speaker 01 in this way is that you see very much a flat graph. You see a flat graph across all disciplines, across physics, biology, medicine, and so on. And it actually makes a lot of sense if you think about it, because think about the progress of physics 110 years ago, right? It was a time of crazy change. Think about the progress of technology, you know, 170 years ago, when we started letting, you know, replacing horses with cars, when we started having electricity and so on. It was a time of incredible change. And today is also a time of very, very fast change. But it would be an unfair characterization to say that today, technology and science are moving way faster than they did 50 years ago, 100 years ago.
19:11 Speaker 01 And if you do try to rigorously plot the temporal density of the significance, yeah, of significance idea, of significance, sorry, you do see very flat curves. That's fascinating. And you can check out the paper that Michael Nielsen had about this idea. And so the way I interpret it is as you make progress.
19:39 Speaker 01 In a given field or in a given subfield of science, it becomes exponentially more difficult to make further progress. Like the very first person to work on information theory. If you enter a new field and it's still the very early years, there's a lot of low-hanging fruit you can pick.
19:58 Lex Fridman? That's right, yeah.
19:58 Speaker 01 But the next generation of researchers is gonna have to dig much harder, actually, to make smaller discoveries. A probably larger number of smaller discoveries. And to achieve the same amount of impact, you're gonna need a much greater headcount. And that's exactly the picture you're seeing with science, that the number of scientists and engineers is in fact increasing exponentially. The amount of computational resources that are available to science is increasing exponentially and so on. So the resource consumption of science is exponential, but the output in terms of progress, in terms of significance, is linear. And the reason why is because, and even though science is recursively self-improving,
20:43 Speaker 01 meaning that scientific progress turns into technological progress, which in turn helps science. If you look at computers, for instance, our products of science and computers are tremendously useful in speeding up science. The internet, same thing. The internet is a technology that's made possible by very recent scientific advances. And itself, because it enables scientists to network, to communicate, to exchange papers and ideas much faster, it is a way to speed up scientific progress. So even though you're looking at a recursively self-improving system, it is consuming exponentially more resources to produce the same amount of problem-solving advantage.
21:26 Lex Fridman? So that's a fascinating way to paint it. And certainly, that holds for the deep learning community, right? If you look at the temporal, what did you call it? The temporal density of significant ideas. If you look at in deep learning, I think, I'd have to think about that. But if you really look at significant ideas in deep learning, they might even be decreasing.
21:49 Speaker 01 So I do believe the per paper significance is decreasing. But the amount of papers is still today exponentially increasing. So I think if you look at an aggregate, my guess is that you would see a linear progress. If you were to sum. To sum the significance of all papers,
22:13 Speaker 01 you would see roughly linear progress. And in my opinion, it is not a coincidence that you're seeing linear progress in science despite exponential resource consumption. I think the resource consumption is dynamically adjusting itself to maintain linear progress because we, as a community, expect linear progress. Meaning that if we start investing less and seeing less progress, it means that suddenly there are some lower hanging fruits that become available and someone's going to step up and pick them.
22:47 Lex Fridman? Right.
22:48 Speaker 01 So it's very much like a market for discoveries and ideas.
22:54 Lex Fridman? But there's another fundamental part which you're highlighting, which as a hypothesis as science or like the space of ideas, any one path you travel down, it gets exponentially more difficult to get new, to develop new ideas.
23:11 Speaker 01 Yes.
23:11 Lex Fridman? And your sense is that's going to hold across our mysterious universe.
23:18 Speaker 01 Yes, well, exponential progress triggers exponential friction. So that if you tweak one part of the system, suddenly some other part becomes a bottleneck. For instance, let's say you develop some device that measures its own acceleration and then it has some engine and it outputs even more acceleration in proportion of its own acceleration and you drop it somewhere. It's not going to reach infinite speed because it exists in a certain context.
23:46 Speaker 01 So the air around it is going to generate friction. It's going to block it at some top speed. And even if you were to consider a broader context and leave the bottleneck there, like the bottleneck of friction.
24:00 Speaker 01 Then some other part of the system would start stepping in and creating exponential friction, maybe the speed of flight or whatever. And it's definitely holds true when you look at the problem solving algorithm that is being run by science as an institution, science as a system. As you make more and more progress, despite having this recursive self-improvement component, you are encountering exponential friction. Like the more researchers you have working on different ideas, the more overhead you have in terms of communication across researchers. If you look at, you were mentioning quantum mechanics, right? Well, if you want to start making significant discoveries today,
24:44 Speaker 01 significant progress in quantum mechanics, there is an amount of knowledge you have to ingest, which is huge. So there's a very large overhead to even start to contribute. There is a large amount of overhead to synchronize across researchers and so on. And of course, the significant practical experiments are going to require exponentially expensive equipment because the easier ones have already been run, right?
25:14 Lex Fridman? So in your senses, there is no way escaping, there's no way of escaping this kind of friction with artificial intelligent systems?
25:26 Speaker 01 Yeah, no, I think science is a very good way to model what would happen with a superhuman recursive self-improving AI.
25:34 Lex Fridman? That's your sense, I mean.
25:36 Speaker 01 That's my intuition. It's not, it's not like a mathematical proof of anything. That's not my point. Like, I'm not, I'm not trying to prove anything. I'm just trying to make an argument to question the narrative of intelligence explosion, which is quite a dominant narrative. And you do get a lot of pushback if you go against it. Because, so for many people, right? AI is not just a subfield of computer science. It's more like a belief system, like this belief that the world is headed towards an event, the singularity, past which, you know.
26:11 Speaker 01 AI will go exponential very much and the world will be transformed and humans will become obsolete. And if you go against this narrative, because it is not really a scientific argument, but more of a belief system, it is part of the identity of many people. If you go against this narrative, it's like you're attacking the identity of people who believe in it. It's almost like saying God doesn't exist or something. So you do get a lot of pushback if you try to question these ideas.
26:41 Lex Fridman? First of all, I believe most people, they might not be as eloquent or explicit as you're being, but most people in computer science and most people who actually have built anything that you could call AI, quote unquote, would agree with you. They might not be describing in the same kind of way. It's more, so the pushback you're getting is from people who get attached to the narrative from, not from a place of science, but from a place of imagination.
27:10 Speaker 01 That's correct.
27:11 Lex Fridman? So why do you think that's so appealing? Because the usual dreams that people have when you create a super intelligence system past the singularity, that what people imagine is somehow always destructive. Do you have, if you were put on your psychology hat, what's, why is it so appealing to imagine the ways that all of human civilization will be destroyed?
27:37 Speaker 01 I think it's a good story. You know, it's a good story and very interestingly, it mirrors religious stories, right? Religious mythology. If you look at the mythology of most civilizations, it's about the world being headed towards some final events in which the world will be destroyed and some new world order will arise that will be mostly spiritual, like the apocalypse followed by a paradise probably.
28:05 Lex Fridman? Yeah.
28:06 Speaker 01 It's a very appealing story on a fundamental level and we all need stories. We all need stories to structure the way we see the world, especially at timescales that are beyond our ability to make predictions.
28:20 Lex Fridman? So on a more serious, non-exponential explosion.
28:27 Lex Fridman? Question, do you think there will be a time when we'll create something like human-level intelligence or intelligent systems that will make you sit back and be just surprised at damn how smart this thing is? That doesn't require exponential growth or exponential improvement, but what's your sense of a timeline and so on that where you'll be really surprised at certain capabilities? And we'll talk about limitations and deep learning and so on. When do you think in your lifetime you'll be really damn surprised?
29:03 Speaker 01 Around 2013, 2014, I was many times surprised by the capabilities of deep learning, actually. That was before we had assessed exactly what deep learning could do and could not do. And it felt like a time of immense potential. And then we started narrowing it down, but I was very surprised. So we'd say it has already happened.
29:23 Lex Fridman? Was there a moment, there must have been a day in there, where your surprise was almost bordering on the belief of the narrative that we just discussed? Was there a moment, because you've written quite eloquently about the limits of deep learning, was there a moment that you thought that maybe deep learning is limitless?
29:45 Speaker 01 No, I don't think I've ever believed this. What was really shocking is that it worked, right?
29:52 Lex Fridman? It worked at all, yeah.
29:53 Speaker 01 Yeah. But there's a big jump between being able to do really good computer vision and human-level intelligence. So I don't think at any point I wasn't under the impression that the results we got in computer vision meant that we were very close to human-level intelligence. I don't think we're very close to human-level intelligence. I do believe that there's no reason why we won't achieve it at some point. I also believe that, you know...
30:22 Speaker 01 The problem with talking about human-level intelligence is that implicitly you're considering like an axis of intelligence with different levels. But that's not really how intelligence works. Intelligence is very multidimensional. And so there's the question of capabilities, but there's also the question of being human-like. And it's two very different things. Like you can build potentially very advanced intelligent agents that are not human-like at all. And you can also build very human-like agents. And these are two very different things, right?
30:54 Lex Fridman? Right. Let's go from the philosophical to the practical. Can you give me a history of Keras and all the major deep learning frameworks that you kind of remember in relation to Keras? And in general, TensorFlow, Theano, the old days. Can you give a brief overview, Wikipedia-style history and your role in it before we return to AGI discussions?
31:16 Speaker 01 Yeah, that's a broad topic. So I started working on Keras. It was a name Keras at the time. I actually picked the name like just the day I was going to release it. So I started working on it in February 2015. And so at the time, there weren't too many people working on deep learning, maybe like fewer than 10,000. The software tooling was not very developed.
31:42 Speaker 01 So the main deep learning library was Caffe, which was mostly C++.
31:47 Lex Fridman? Why do you say Caffe was the main one?
31:49 Speaker 01 Caffe was vastly more popular than Theano in late 2014, early 2015. Caffe was the one library that everyone was using for computer vision.
32:00 Lex Fridman? And computer vision was the most popular problem.
32:03 Speaker 01 Absolutely. Like ConvNets was like the subfield of deep learning that everyone was working on. So myself, so in late 2014, I was actually interested in RNNs, in recurrent neural networks, which was a very niche topic at the time, right? It really took off around 2016. And so I was looking for good tools. I had used Torch 7, I had used Theano, used Theano a lot in Kaggle competitions.
32:36 Speaker 01 I had used Caffe and there was no like good solution for RNNs at the time. Like there was no reusable open source implementation of an LSTM, for instance. So I decided to build my own. And at first, the pitch for that was it was going to be mostly around LSTM, recurrent neural networks. It was going to be in Python. An important decision at the time that was kind of not obvious is that the models would be defined via Python code, which was kind of like going against the mainstream at the time because Caffe,
33:14 Speaker 01 Python, and so on, like all the big libraries were actually going with the approach of using static configuration files in YAML to define models. So some libraries were using code to define models, like Torch 7, obviously, but that was not Python. Lazagne was like a Tiano-based, very early library that was, I think, developed. I don't remember exactly, probably late 2014.
33:37 Lex Fridman? It's Python as well.
33:38 Speaker 01 It's Python as well. It was, it was like on top of Tiano. And so I started working on something. And the, and the value proposition at the time was that not only that the, what I think was the first reusable open source implementation of LSTM,
33:57 Speaker 01 you could combine RNNs and convnets with the same library, which was not really possible before. Like Caffe was only doing convnets.
34:07 Speaker 01 And it was kind of easy to use because so before I was using Tiano, I was actually using Scikit-learn and I loved Scikit-learn for its usability. So I drew a lot of inspiration from Scikit-learn when I met Karas. It's almost like Scikit-learn for neural networks.
34:21 Lex Fridman? Yeah, the fit function.
34:23 Speaker 01 Exactly, the fit function, like reducing a complex training loop to a single function call, right? And of course, you know, some people will say this is hiding a lot of details, but that's exactly the point, right? The magic is the point.
34:37 Lex Fridman? Right.
34:37 Speaker 01 So it's magical, but in a good way. It's magical in the sense that it's delightful, right?
34:42 Lex Fridman? Yeah, I'm actually quite surprised. I didn't know that it was born out of desire to implement RNNs and LSTMs.
34:49 Speaker 01 It was.
34:49 Lex Fridman? That's fascinating. So you were actually one of the first people to really try to attempt. To get the major architectures together. And it's also interesting, you made me realize that that was a design decision at all is defining the model in code. Just, I'm putting myself in your shoes, whether the YAML, especially if Caffe was the most popular.
35:10 Speaker 01 It was the most popular by far at the time.
35:13 Lex Fridman? If I were, yeah, I don't, I didn't like the YAML thing, but it makes more sense that you will put in a configuration file the definition of a model. That's an interesting gutsy move to stick with defining it in code, just if you look back.
35:28 Speaker 01 Other libraries were doing it as well, but it was definitely the more niche option.
35:33 Lex Fridman? Yeah, okay, Keras and then?
35:35 Speaker 01 Keras, so I released Keras in March 2015 and it got users pretty much from the start. So the deep learning committee was very, very small at the time. Lots of people were starting to be interested in the STM, so it was gonna release it at the right time because it was offering an easy to use STM implementation, exactly at the time where lots of you started to be intrigued by the capabilities of RNN, RNN S1 LP, so it grew from there.
36:03 Speaker 01 Then I joined Google about six months later, and that was actually completely unrelated to Keras, so actually joined a research team working on image classification, mostly like computer vision. So I was doing computer vision research at Google initially. And immediately when I joined Google, I was exposed to the early internal version of TensorFlow. And the way it appeared to me at the time, and that was definitely the way it was at the time, is that this was an improved version of Tiano. So I immediately knew I had to port Keras to this new TensorFlow thing.
36:43 Speaker 01 And I was actually very busy as a new Googler, so I had not time to work on that. But then in November, I think it was November 2015, TensorFlow got released and it was kind of like my Wake up call that hey, I had to actually go and make it happen. So in December, I ported Keras to run on top of TensorFlow. But it was not exactly port. It was more like a refactoring where I was abstracting away all the backend functionality into one module. So that the same code base could run on top of multiple backends. So on top of TensorFlow or Theano.
37:24 Speaker 01 And for the next year, Theano stayed as the default option. It was easier to use, somewhat less buggy. It was much faster, especially when it came to finance. But eventually TensorFlow overtook it.
37:44 Lex Fridman? And TensorFlow, the early TensorFlow, has similar architectural decisions as Theano. So it was a natural transition.
37:54 Speaker 01 Yeah, absolutely.
37:55 Lex Fridman? So what, I mean, that still Keras is a side, almost found project, right?
38:02 Speaker 01 Yeah, so it was not my job assignment. It was not, I was doing it on the side. So and even though it grew to have a lot of users for a deep learning library at the time, like Stroud 2016. But I wasn't doing it as my main job. So things started changing in, I think it must have been maybe October 2016, so one year later. So Rajat, who was the lead on TensorFlow, basically showed up one day in our building. Where I was doing like, so I was doing research and things like, so I did a lot of computer vision research. Also collaborations with Christian Ziegedy and deep learning for theorem proving.
38:46 Speaker 01 It was a really interesting research topic.
38:51 Speaker 01 And so Rajat was saying, hey, we saw Keras, we like it. We saw that you're at Google. Why don't you come over for like a quarter and, and, and work with us? And I was like, yeah, that sounds like a great opportunity. Let's do it. And so I started working on integrating the Keras API into TensorFlow more tightly. So what followed up is a sort of like temporary TensorFlow only version of Keras that was in tensorflow.com trib for a while and finally moved to TensorFlow Core. And, you know, I've never actually gotten back to my old team doing research.
39:34 Lex Fridman? Well, it's, it's kind of funny that somebody like you who dreams of, or at least sees the, the power of AI systems that reason and theorem proving we'll talk about, has also created a system that makes the, the most basic kind of Lego building that is deep learning, super accessible, super easy, so beautifully so. It's, it's a funny irony that you're, there's both, you're responsible for both things. But so TensorFlow 2.0 is kind of, there's a sprint. I don't know how long it'll take, but there's a sprint towards the finish. What do you look, what are you working on these days? What are you excited about?
40:19 Lex Fridman? What are you excited about in 2.0? Eager execution. There's so many things that just make it a lot easier to work. What are you excited about? And what's also really hard? What are the problems you have to kind of solve?
40:32 Speaker 01 So I've spent the past year and a half working on TensorFlow 2. And it's been a long journey. I'm actually extremely excited about it. I think it's a great product. It's a delightful product compared to TensorFlow 1. We've made a huge progress. So on the Keras side, what I'm really excited about is that, so, you know, previously Keras has been this very easy to use, high level interface to do deep learning. But if you wanted to.
41:06 Speaker 01 You know, if you wanted a lot of flexibility, the Keras framework was probably not the optimal way to do things compared to just writing everything from scratch.
41:17 Lex Fridman? Right.
41:19 Speaker 01 So in some way, the framework was getting in the way. And in TensorFlow 2, you don't have this at all, actually. You have the usability of the high-level interface, but you have the flexibility of this low-level interface, and you have this spectrum of workflows where you can get more or less usability and flexibility, the trade-offs depending on your needs, right? You can write everything from scratch, and you get a lot of help doing so by subclassing models and writing custom train loops using eager execution. It's very flexible, it's very easy to debug, it's very powerful.
41:59 Speaker 01 But all of this integrates seamlessly with higher-level features, up to the classic Keras workflows, which are very psychically unlike and are ideal for a data scientist, machine learning engineer type of profile. So now you can have the same framework, offering the same set of APIs that enable a spectrum of workflows that are more or less low-level, more or less high-level, that are suitable for profiles ranging from researchers to data scientists and everything in between.
42:32 Lex Fridman? Yeah, so that's super exciting. I mean, it's not just that, it's connected to all kinds of tooling. You can go on mobile, you can go with TensorFlow Lite, you can go in the cloud or serving and so on, and all is connected together. Some of the best software written ever is often done by one person, sometimes two. So with Google, you're now seeing Keras having to be integrated, and TensorFlow, I'm sure, has a ton of engineers working on. And there's, I'm sure, a lot of tricky design decisions to be made. How does that process usually happen, from at least your perspective? What are the debates like?
43:16 Lex Fridman? Is there a lot of thinking, considering different options and so on?
43:24 Speaker 01 Yes. So a lot of the time...
43:28 Speaker 01 I spent at Google is actually discussing design discussions, writing design docs, participating in design review meetings and so on. This is as important as actually writing a code.
43:42 Lex Fridman? Right.
43:42 Speaker 01 So there's a lot of thought and a lot of care that is
43:48 Speaker 01 taken in coming up with these decisions and taking into account all of our users, because TensorFlow has this extremely diverse user base. It's not just one user segment where everyone has the same needs. We have small-scale production users, large-scale production users. We have startups, we have researchers. It's all over the place and we have to cater to all of their needs.
44:14 Lex Fridman? If I just look at the standard debates of C++ or Python, there's some heated debates. Do you have those at Google? I mean, they're not heated in terms of emotionally, but there's probably multiple ways to do it right. So how do you arrive through those design meetings at the best way to do it, especially in deep learning, where the field is evolving as you're doing it? Is there some magic to it? There's some magic to the process?
44:43 Speaker 01 I don't know if there's magic to the process, but there definitely is a process. So making design decisions is about satisfying a set of constraints, but also trying to do so in the simplest way possible, because this is what can be maintained, this is what can be expanded in the future. So you don't want to naively satisfy the constraints by just, for each capability you need available, you're going to come up with one argument in your API and so on. You want to design APIs that are modular and hierarchical, so that they have an API surface that is as small as possible. And you want this modular hierarchical
45:27 Speaker 01 architecture to reflect the way that domain experts think about the problem.
45:33 Speaker 01 Because like as a domain expert, when you're reading about a new API, you're reading a tutorial or some docs pages. You already have a way that you're thinking about the problem.
45:45 Lex Fridman? Right.
45:45 Speaker 01 You already have like a certain concepts in mind and you're thinking about how they relate together. And when you're reading docs, you're trying to build as quickly as possible a mapping between the concepts featured in your API and the concepts in your mind. So you're trying to map your mental model as a domain expert to the way things work in the API. So you need an API and an underlying implementation that are reflecting the way people think about these things.
46:17 Lex Fridman? So in minimizing the time it takes to do the mapping?
46:19 Speaker 01 Yes, minimizing the time, the cognitive load there is in ingesting this new knowledge about your API. An API should not be self referential or referring to implementation details. It should only be referring to domain specific concepts that people already understand.
46:40 Lex Fridman? Brilliant. So what's the future of Keras and TensorFlow look like? What does TensorFlow 3.0 look like?
46:47 Speaker 01 So that's kind of too far in the future for me to answer, especially since I'm not even the one making these decisions.
46:54 Lex Fridman? Okay.
46:56 Speaker 01 But so from my perspective, which is, you know, just one perspective among many different perspectives on the TensorFlow team.
47:04 Speaker 01 I'm really excited by developing even higher level APIs, higher level than Keras. I'm really excited by hyperparameter tuning, by automated machine learning, AutoML.
47:18 Speaker 01 I think the future is not just, you know, defining a model like
47:23 Speaker 01 like you were assembling Lego blocks and then collect fit on it. It's more like. An automagical model that would just look at your data and optimize the objective you're after, right? So that's what I'm looking into.
47:40 Lex Fridman? Yeah, so you put the baby into a room with the problem and come back a few hours later with a fully solved problem.
47:48 Speaker 01 Exactly. It's not like a box of Legos. It's more like the combination of a kid that's really good at Legos and a box of Legos. And just building the thing on its own.
47:58 Lex Fridman? Very nice. So that's an exciting feature. And I think there's a huge amount of applications and revolutions to be had under the constraints of the discussion we previously had. But what do you think are the current limits of deep learning? If we look specifically at these function approximators that try to generalize from data. You've talked about local versus extreme generalization. You mentioned that neural networks don't generalize well, and humans do. So there's this gap.
48:34 Lex Fridman? And you've also mentioned that extreme generalization requires something like reasoning to fill those gaps. So how can we start trying to build systems like that?
48:44 Speaker 01 Right, yeah, so this is by design, right? Deep learning models are like huge parametric models,
48:54 Speaker 01 differentiable, so continuous, that go from an input space to an output space. And they're trained with gradient descent. So they're trained pretty much point by point. They are learning a continuous geometric morphing from an input vector space to an output vector space. And because this is done point by point, a deep neural network can only make sense of points in experience space that are very close to things that it has already seen in its training data. At best, it can do interpolation across points.
49:30 Speaker 01 But that means, you know, that means in order to train your network, you need a dense sampling of the input cross output space. Almost a point by point sampling, which can be very expensive if you're dealing with complex, real world problems like autonomous driving, for instance, or robotics. It's doable if you're looking at the subset of the visual space. But even then, it's still fairly expensive. You still need millions of examples. And it's only going to be able to make sense of things that are very close to what is seen before. And in contrast to that, well, of course, you have human intelligence. But even if you're not looking at human intelligence, you can look at very simple rules, algorithms.
50:13 Speaker 01 If you have a symbolic rule, it can actually apply to a very, very large set of inputs because it is abstract. It is not obtained by doing a point by point mapping, right? For instance, if you try to learn a sorting algorithm using a deep neural network, well, you're very much limited to learning point by point
50:37 Speaker 01 what the sorted representation of this specific list is like. But instead, you could have a very, very simple sorting algorithm written in a few lines. Maybe it's just, you know, two nested loops. And it can process any list at all because it is abstract, because it is a set of rules. So deep learning is really like point by point geometric morphings, morphing, trained with gradient descent. And meanwhile, abstract rules can generalize much better. And I think the future is really to combine the two.
51:13 Lex Fridman? So how do we, you think, combine the two? How do we combine good point by point functions with programs?
51:23 Lex Fridman? Which is what symbolic AI type systems.
51:26 Speaker 01 Yeah.
51:26 Lex Fridman? At which levels the combination happen. I mean, obviously, we're jumping into the realm of where there's no good answers. It's just kind of ideas and intuitions and so on.
51:37 Speaker 01 Well, if you look at the really successful AI systems today, I think there are already hybrid systems that are combining symbolic AI with deep learning. For instance, successful robotics systems are already mostly model-based, rule-based. Things like planning algorithms and so on. At the same time, they're using deep learning as perception modules. Sometimes they're using deep learning as a way to inject a fuzzy intuition into a rule-based process. If you look at a system like a self-driving car, it's not just one big N2N neural network, you know, that wouldn't work at all. Precisely because in order to train that, you would need a dense sampling
52:20 Speaker 01 of experience base when it comes to driving, which is completely unrealistic, obviously. Instead, the self-driving car is mostly
52:31 Speaker 01 symbolic, you know, it's software, it's programmed by hand. So it's mostly based on explicit models, in this case mostly 3D models of the of the environment around the car. But it's interfacing with the real world using deep learning modules, right?
52:48 Lex Fridman? Right, so the deep learning there serves as a way to convert the raw sensory information to something usable by symbolic systems. Okay, well, let's linger on that a little more. So dense sampling from input to output. You said it's obviously very difficult. Is it possible?
53:07 Speaker 01 In the case of self-driving, you mean?
53:08 Lex Fridman? Let's say self-driving, right? Self-driving for many people.
53:14 Lex Fridman? Let's not even talk about self-driving. Let's talk about steering, so staying inside the lane.
53:22 Speaker 01 Lane-following, yeah, it's definitely a problem you can solve with an end-to-end deep learning model, but that's like one small subset.
53:27 Lex Fridman? Hold on a second. I don't know why you're jumping from the extreme so easily, because I disagree with you on that. I think, well, it's not obvious to me that you can solve lane-following.
53:40 Speaker 01 No, it's not obvious. I think it's doable. I think in general, you know, there is no hard limitations to what you can learn with a deep neural network, as long as the search space is rich enough, is flexible enough, and as long as you have this dense sampling of the input cross-output space. The problem is that this dense sampling could mean anything from 10,000 examples to like trillions and trillions.
54:09 Lex Fridman? So that's my question. So what's your intuition? And if you could just give it a chance and think, what kind of problems can be solved by getting a huge amount of data and thereby creating a dense mapping? So let's think about natural language dialogue, the Turing test. Do you think the Turing test can be solved with a neural network alone?
54:38 Speaker 01 Well, the Turing test is all about tricking people into believing they're talking to a human, and I don't think that's actually very difficult, because it's more about exploiting human perception and not so much about intelligence. There's a big difference between mimicking intelligent behavior and actual intelligent behavior.
54:59 Lex Fridman? So, OK, let's look at maybe the Alexa Prize and so on, the different formulations of the natural language conversation that are less about mimicking and more about maintaining a fun conversation that lasts for 20 minutes. That's a little less about mimicking, and that's more about, I mean, it's still mimicking, but it's more about being able to carry forward a conversation with all the tangents that happen in dialogue and so on. Do you think that problem is learnable?
55:26 Lex Fridman? With this kind of, with a neural network that does the point-to-point mapping.
55:31 Speaker 01 So I think it would be very, very challenging to do this with deep learning. I don't think it's out of the question either. I wouldn't rule it out.
55:40 Lex Fridman? The space of problems that can be solved with a large neural network. What's your sense about the space of those problems? So useful problems for us.
55:49 Speaker 01 In theory, it's it's infinite, right? You can solve any problem. In practice, well, deep learning is a great fit for perception problems. In general, any
56:02 Speaker 01 any problem which is naturally amenable to explicit handcrafted rules or rules that you can generate by exhaustive search over some program space.
56:13 Speaker 01 So perception, artificial intuition, as long as you have a sufficient training data.
56:20 Lex Fridman? And that's the question. I mean, perception, there's interpretation and understanding of the scene, which seems to be outside the reach of current perception systems. So do you think larger networks will be able to start to understand the physics? And the physics of the scene, the three-dimensional structure and relationships of objects in the scene and so on? Or really, that's where symbolic AI has to step in?
56:46 Speaker 01 Well, it's it's always possible
56:50 Speaker 01 to solve these problems. With with with deep learning is just extremely inefficient. A model would be, an explicit rule-based abstract model would be a far far better, more compressed representation of physics than learning just this mapping between in this situation, this thing happens. If you change the situation slightly, then this other thing happens and so on.
57:11 Lex Fridman? Do you think it's possible to automatically generate the programs that that would require that kind of reasoning? Or does it have to? So where expert systems fail, there's so many facts about the world had to be hand-coded in. Do you think it's possible to learn those logical statements that are true about the world and their relationships?
57:35 Lex Fridman? Do you think, I mean, that's kind of what theorem proving at a basic level is trying to do, right?
57:39 Speaker 01 Yeah, except it's much harder to formulate statements about the world compared to formulating mathematical statements. Statements about the world, you know, tend to be subjective.
57:51 Speaker 01 So can you, can you learn? Rule-based models?
57:56 Lex Fridman? Yes.
57:56 Speaker 01 Yes, definitely. That's the, that's the field of program synthesis. However, today, we just don't really know how to do it. So it's, it's very much a graph search or tree search problem. And so we are limited to, you know, the sort of a tree session graph search algorithms that we have today. Personally, I think genetic algorithms are very promising.
58:19 Lex Fridman? So almost like genetic programming?
58:21 Speaker 01 Genetic programming, exactly.
58:22 Lex Fridman? Can you discuss the field of program synthesis? Like what?
58:28 Lex Fridman? How many people are working and thinking about it? What, where we are in the history of program synthesis? And what are your hopes for it?
58:37 Speaker 01 Well, if it were deep learning, this is like the 90s. So meaning that we already have, we already have existing solutions. We are starting to have some basic understanding of what this is about. But it's still a field that is in its infancy. There are very few people are working on it. There are very few real-world applications.
59:01 Speaker 01 So the one real-world application I'm aware of is a flash fill in Excel. It's a way to automatically learn very simple programs to format cells in an Excel spreadsheet from, from a few examples. For instance, learning a way to format a date, things like that.
59:19 Lex Fridman? Oh, that's fascinating.
59:20 Speaker 01 Yeah.
59:21 Lex Fridman? You know, okay, that's, that's fascinating topic. I always wonder when I provide a few samples to Excel, what it's able to figure out? Like just giving it a few dates. What are you able to figure out from the pattern I just gave you? That's a fascinating question and it's fascinating whether that's learnable. Patterns and you're saying they're working on that?
59:41 Speaker 01 Yeah.
59:42 Lex Fridman? How big is the toolbox currently? Are we completely in the dark? So if you said the nineties.
59:47 Speaker 01 In terms of progress this? No, so I would say, so maybe nineties even even too optimistic. Because by the nineties, you know, we already understood backprop. We already understood, you know, the engine of deep learning. Even though we couldn't really see its potential quite. Today, I don't think we have found the engine of program synthesis.
1:00:07 Lex Fridman? So we're in the winter before backprop.
1:00:09 Speaker 01 Yeah. In a way, yes. So I do believe program synthesis in general, discrete search over rule based models is going to be a cornerstone of AI research in the next century. Right? And that doesn't mean we're going to drop deep learning. Deep learning is immensely useful, like being able to learn this is a very flexible, adaptable, parametric models, which got Anderson. That's that's actually immensely useful. Like all it's doing is pattern recognition. But being good at pattern recognition, given lots of data is is just extremely powerful. So we are still going to be working on deep learning. We're going to be working on program synthesis. We're going to be combining the two in increasingly automated ways.
1:00:53 Lex Fridman? So let's talk a little about about data. You've tweeted
1:00:59 Lex Fridman? about 10,000 deep learning papers have been written about hard coding priors, about a specific task and a neural network architecture works better than a lack of a prior. Basically summarizing all these efforts, they put a name to an architecture. But really what they're doing is hard coding some priors that improve the performance of the system. But we get straight to the point is is probably true. So you say that you can always buy performance, buy in quotes, performance by either training on more data, better data, or by injecting task information to the architecture of the preprocessing. However, this isn't informative about the generalization power of the techniques used, the fundamental ability to generalize.
1:01:41 Lex Fridman? Do you think we can go far by coming up with better methods for this kind of cheating? For better methods of large scale annotation of data? So building better prior.
1:01:51 Speaker 01 If you if you made it, it's not cheating anymore.
1:01:54 Lex Fridman? Right. I'm joking about the cheating, but large scale. So basically I'm asking about.
1:02:03 Lex Fridman? Something that hasn't, from my perspective, been researched too much is exponential improvement in annotation of data. Do you often think about?
1:02:15 Speaker 01 I think it's actually been researched quite a bit. You just don't see publications about it because, you know, people who publish papers are going to publish about known benchmarks. Sometimes they're going to raise a new benchmark. People who actually have real-world, large-scale, depending problems, they're going to spend a lot of resources into data annotation and good data annotation pipelines, but you don't see any papers about it.
1:02:36 Lex Fridman? That's interesting. So do you think, certainly the resources, but do you think there's innovation happening?
1:02:41 Speaker 01 Oh, yeah, definitely. To clarify the point in the tweet. So, machine learning in general is the science of generalization. You want to generate knowledge that can be reused across different data sets, across different tasks. And if instead you're looking at one data set and then you are hard coding knowledge about this task into your architecture, this is no more useful than training a network and then saying, "Oh, I found these weight values perform well," right?
1:03:19 Speaker 01 So David Ha, I don't know if you know David, he had a paper the other day about weight-agnostic neural networks. And this is a very interesting paper because it really illustrates the fact that an architecture, even without weights, an architecture is a knowledge about a task. It encodes knowledge. And when it comes to architectures that are uncrafted by researchers, in some cases, it is very, very clear that all they are doing is artificially re-encoding the template that corresponds to the proper way to solve a task, including a given data set.
1:04:02 Speaker 01 For instance, I don't know if you've looked at the Baby data set, which is about natural language question answering. It is generated by an algorithm. So this is a question and answer pairs that are generated by an algorithm. The algorithm is so long as certain templates. Turns out if you craft a network that literally encodes this template, you can solve this data set with nearly 100% accuracy. But that doesn't actually tell you anything about how to solve question answering in general, which is the point.
1:04:34 Lex Fridman? Right. The question is just to linger on it, whether it's from the data side or from the size of the network. I don't know if you've read the blog post by Ray Sutton, The Bitless Lesson?
1:04:44 Speaker 01 Yeah.
1:04:44 Lex Fridman? Where he says, "The biggest lesson that we can read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective." So as opposed to figuring out methods that can generalize effectively, do you think we can
1:05:02 Lex Fridman? get pretty far by just having something that leverages computation and the improvement of computation?
1:05:08 Speaker 01 Yeah, so I think Rich is making a very good point, which is that a lot of these papers, which are actually all about manually hard coding prior knowledge about a task into some system, doesn't have to be deep learning architecture, but into some system, right?
1:05:25 Speaker 01 You know, these papers are not actually making any impact. Instead, what's making really long-term impact is very simple, very general systems that are really agnostic to all these tricks, because these tricks do not generalize. And of course, the one
1:05:43 Speaker 01 general and simple thing that you should focus on is that which leverages computation. Because computation, the availability of large-scale computation has been, you know, increasing exponentially following Moore's law. So if your algorithm is all about exploiting this, then your algorithm is certainly exponentially improving, right? So I think Rich is definitely right.
1:06:09 Speaker 01 However, you know, he's right about the past 70 years. He's like assessing the past 70 years. I am not sure that this assessment will still hold true for the next 70 years. It might to some extent. I suspect it will not because the truth of his assessment is a function of the context, right, in which in which this research took place. And the context is changing, like Moore's law might not be applicable anymore, for instance, in the future. And I do believe that, you know, when you, when you, when you tweak one aspect of a system, when you explore one aspect of a system, some other aspect starts becoming the bottleneck.
1:06:53 Speaker 01 Let's say you have unlimited computation. Well, then data is the bottleneck. And I think we are already starting to be in a regime where our systems are so large in scale and so data hungry, the data today and the quality of data and the scale of data is the bottleneck. And in this environment, the bitter lesson from Rich is not going to be true anymore. Right? So I think we are going to move from a focus on a scale of a competition scale to focus on data efficiency.
1:07:26 Lex Fridman? Data efficiency. So that's getting to the question of symbolic AI. But to linger on the deep learning approaches, do you have hope for either unsupervised learning or reinforcement learning, which are ways of being more data efficient in terms of the amount of data they need that require human annotation?
1:07:48 Speaker 01 So unsupervised learning and reinforcement learning are frameworks for learning, but they are not like any specific technique. So usually when people say reinforcement learning, what they really mean is deep reinforcement learning, which is like one approach, which is actually very questionable.
1:08:04 Lex Fridman? The question I was asking was unsupervised learning with deep neural networks and deep reinforcement learning.
1:08:11 Speaker 01 Well, these are not really data efficient because you're still leveraging, you know, these huge parametric models, train point by point with gradient descent. It is more efficient in terms of the number of annotations, the density of annotations you need. So the idea being to learn the latent space around which the data is organized and then map the sparse annotations into it. And sure, I mean, that's clearly a very good idea.
1:08:40 Speaker 01 It's not really a topic I would be working on, but it's clearly a good idea.
1:08:44 Lex Fridman? So it would get us to solve some problems that...
1:08:48 Speaker 01 It would get us to incremental improvements in label data efficiency.
1:08:55 Lex Fridman? Do you have concerns about short-term or long-term threats from AI, from artificial intelligence?
1:09:04 Speaker 01 Yes, definitely to some extent.
1:09:07 Lex Fridman? And what's the shape of those concerns?
1:09:09 Speaker 01 This is actually something I've briefly written about. But the capabilities of deep learning technology can be used in many ways that are concerning from mass surveillance with things like facial recognition. In general, you know, tracking lots of data about everyone and then being able to making sense of this data, to do identification, to do prediction.
1:09:39 Speaker 01 That's concerning. That's something that's being very aggressively pursued by totalitarian states like China.
1:09:48 Speaker 01 One thing I am very much concerned about is that
1:09:53 Speaker 01 our lives are increasingly online, are increasingly digital, made of information, made of information consumption and information production, our digital footprint, I would say. And if you absorb all of this data and you're in control of where you consume information, you know, social networks and so on. Recomендation engines. Then you can build a sort of reinforcement loop for human behavior. You can observe the state of your mind at time T.
1:10:35 Speaker 01 You can predict how you would react to different pieces of content. How to get you to move your mind, you know, in a certain direction. And then you can feed, and feed you the specific piece of content that would move you in a specific direction. And you can do this at scale.
1:10:56 Speaker 01 You know, at scale in terms of doing it continuously in real time. You can also do it at scale in terms of scaling this to many, many people, to entire populations. So potentially artificial intelligence, even in its current state, if you combine it with the internet, with the fact that we have, all of our lives are moving to digital devices and digital information consumption and creation. What you get is the possibility to achieve mass manipulation of behavior and mass psychological control. And this is a very real possibility.
1:11:35 Lex Fridman? Yeah, so you're talking about any kind of recommender system.
1:11:38 Speaker 01 Yeah.
1:11:39 Lex Fridman? Let's look at the YouTube algorithm.
1:11:42 Lex Fridman? Facebook, anything that recommends content you should watch next.
1:11:45 Speaker 01 Yeah.
1:11:46 Lex Fridman? And it's fascinating to think that there's some aspects of human behavior that you can, you know, say a problem of, is this person hold Republican beliefs or Democratic beliefs? And this is a trivial, that's an objective function. And you can optimize and you can measure and you can turn everybody into a Republican or everybody into a Democrat.
1:12:12 Speaker 01 Absolutely, yeah. I do believe it's true. So the human mind is very.
1:12:19 Speaker 01 If you look at the human mind as a kind of computer program, it has a very large exploit surface, right? It has many, many vulnerabilities. Exploit surfaces, yeah. Where you can control it. For instance, when it comes to your political beliefs, this is very much tied to your identity. So for instance, if I'm in control of your news feed on your favorite social media platforms, this is actually where you're getting your news from. And I can, of course, I can choose to only show you news that will make you see the world in a specific way, right? But I can also, you know, create incentives for you to post about some political beliefs.
1:13:01 Speaker 01 And then when I get you to express a statement, if it's a statement that me as the controller, I want to reinforce, I can just show it to people who will agree and they will like it. And that will reinforce the statement in your mind. If this is a statement I want you to, this is a belief I want you to abandon, I can, on the other hand, show it to opponents, right? Will attack you. And then because they attack you, at the very least, next time, you will think twice about posting it. But maybe you will even, you know, stop believing this because you got pushback, right? So there are many ways in which
1:13:44 Speaker 01 social media platforms can potentially control your opinions. And today, the, so all of these things are already being controlled by AI algorithms. These algorithms do not have any explicit political goal today, while potentially they could, like, if some totalitarian government takes over, you know, social media platforms and decides that, you know, now we're gonna use this not just for mass surveillance, but also for mass opinion control and behavior control. Very bad things could happen. But what's really fascinating and actually quite concerning is that even without an explicit intent to manipulate,
1:14:28 Speaker 01 you're already seeing very dangerous dynamics in terms of how these content recommendation algorithms behave. Because right now, the The goal, the objective function of these algorithms is to maximize engagement, right, which seems fairly innocuous at first, right? However, it is not, because content that will maximally engage people, you know, get people to react in an emotional way, get people to click on something, it is very often content that, you know, is not healthy to the public discourse.
1:15:11 Speaker 01 For instance, fake news are far more likely to get you to click on them than real news, simply because they are not constrained to reality. So they can be as atrocious, as surprising, as good stories as you want, because they're artificial, right?
1:15:32 Lex Fridman? Yeah. To me, that's an exciting world, because so much good can come. So there's an opportunity to educate people. You can balance people's worldview with other ideas. So there's so many objective functions. The space of objective functions that create better civilizations is large, arguably infinite. But there's also a large space that creates division and destruction, civil war, a lot of bad stuff. And the worry is, naturally, probably that space is bigger, first of all.
1:16:16 Lex Fridman? And if we don't explicitly think about what kind of effects are going to be observed from different objective functions, then we're going to get into trouble. But the question is, how do we, how do we get into rooms and have discussions? So inside Google, inside Facebook, inside Twitter, and think about, okay, how can we drive up engagement and at the same time create a good society? Is it even possible to have that kind of philosophical discussion?
1:16:48 Speaker 01 I think you can definitely try. So from my perspective, I would feel rather uncomfortable with companies that are in control of these news-filled algorithms, with them making explicit decisions.
1:17:04 Speaker 01 To manipulate people's opinions or behaviors, even if the intent is good. Because that's a very totalitarian mindset. So instead, what I would like to see, that's probably never going to happen, because it's not very realistic, but that's actually something I really care about. I would like all these algorithms to present configuration settings to their users, so that the users can actually make the decision about how they want to be impacted by these information recommendation, content recommendation algorithms. For instance, as a user of something like YouTube or Twitter, maybe I want to maximize learning about a specific topic, right?
1:17:47 Speaker 01 So I want the algorithm to feed my curiosity, right? Which is in itself a very interesting problem. So instead of maximizing my engagement, it will maximize how fast and how much I'm learning. And it will also take into account the accuracy, hopefully, of the information I'm learning. So, yeah, the user should be able to determine exactly how these algorithms are affecting their lives. I don't want actually any entity making decisions about in which direction they're going to try to manipulate me, right? I want, I want technology. So AI, these algorithms are increasingly going to be
1:18:32 Speaker 01 our interface to a world that is increasingly made of information. And I want, I want everyone to be in control of this interface, to interface with the world on their own terms. So if someone wants these algorithms to serve, you know, their own personal growth goals, they should be able to configure these algorithms in such a way.
1:18:58 Lex Fridman? Yeah, but so, I know it's painful to have explicit decisions, but there is underlying explicit decisions, which is some of the most beautiful fundamental philosophy that, that we have before us, which is...
1:19:16 Lex Fridman? Personal growth. If I want to watch videos from which I can learn, what does that mean? So if I have a checkbox that wants to emphasize learning, there's still an algorithm with explicit decisions in it that would promote learning. What does that mean for me? Like, for example, I've watched a documentary on flat earth theory, I guess. It was very, like, I learned a lot. I'm really glad I watched it. It was a friend recommended it to me. Because I don't have such an allergic reaction to crazy people as my fellow colleagues do. But it was very, it was very eye opening. And for others, it might not be.
1:19:59 Lex Fridman? For others, they might just get turned off from that. Same with Republican, Democrat. And what? It's a non-trivial problem. And first of all, if it's done well, I don't think it's something that wouldn't happen, that YouTube wouldn't be promoting, or Twitter wouldn't be. It's just a really difficult problem. How do we do, how to give people control?
1:20:21 Speaker 01 Well, it's mostly an interface design problem. The way I see it, you want to create technology that's like a mentor, or a coach, or an assistant. So that it's not your boss, right? You are in control of it. You are telling it what to do for you. And if you feel like it's manipulating you, it's not actually, it's not actually doing what you want. You should be able to switch to a different algorithm.
1:20:50 Lex Fridman? Right. So that, fine tune control, you kind of learn, you're trusting the human collaboration. I mean, that's how I see autonomous vehicles too, is giving as much information as possible, and you learn that dance yourself. Yeah, Adobe, I don't know if you use Adobe products for like photos?
1:21:07 Speaker 01 Yeah, Photoshop.
1:21:08 Lex Fridman? Yeah. They're trying to see if they can inject YouTube into their interface, but basically allow you to, show you all these videos that, because everybody's confused about what to do with features. So basically, teach people by linking to, in that way, it's an assistant that shows, uses videos as a basic element of information.
1:21:29 Speaker 01 Yeah.
1:21:30 Lex Fridman? Okay, so what practically should people do to try to, to try to fight against abuses of these algorithms, or algorithms that manipulate us?
1:21:44 Speaker 01 Honestly, it's a very, very difficult problem because to start with, there is very little public awareness of these issues.
1:21:52 Speaker 01 Very few people would think there's, you know, anything wrong with their unused algorithm, even though there is actually something wrong already, which is that it's trying to maximize engagement most of the time, which has very negative side effects, right? So ideally, so the very first thing is to stop trying to purely maximize engagement. Try to propagate content based on popularity, right? Instead, take into account the goals and the profiles of each user. So you will, you will be, one example is, for instance, when I look at topic recommendations
1:22:37 Speaker 01 on Twitter, like, you know, they have this news tab with search recommendations. It's always the worst garbage because it's content that appeals to the smallest command denominator to all Twitter users because they're trying to optimize, they're purely trying to optimize popularity, they're purely trying to optimize engagement, but that's not what I want. So they should put me in control of some setting so that I define what's the objective function that Twitter is going to be following to show me this content. So, and honestly, so this is all about interface design. And we are not, it's not realistic to give users control of a bunch of knobs that define the algorithm.
1:23:20 Speaker 01 Instead, we should purely put them in charge of defining the objective function. Like, let the user tell us what they want to achieve, how they want this algorithm to impact their lives.
1:23:32 Lex Fridman? So do you think it is that? Or do they provide individual article by article reward structure where you give a signal, I'm glad I saw this or I'm glad I didn't?
1:23:41 Speaker 01 So like a Spotify type
1:23:43 Lex Fridman? Yeah.
1:23:44 Speaker 01 a feedback mechanism. It works to some extent. I'm kind of skeptical about it because the only way the algorithm the algorithm will attempt to relate your choices with the choices of everyone else, which might, you know, if you have an average profile, that works fine. I'm sure Spotify accommodations work fine if you just like mainstream stuff. If you don't, it can be, it's not optimal at all, actually.
1:24:11 Lex Fridman? It'll be an inefficient search for the for the part of the Spotify world that represents you.
1:24:17 Speaker 01 So it's a tough problem, but do note that even a feedback system like what Spotify has does not give me control over what the algorithm is trying to optimize for.
1:24:33 Lex Fridman? Well, public awareness, which is what we're doing now, is a good place to start. Do you have concerns about long term existential threats of artificial intelligence?
1:24:45 Speaker 01 Well, as I was saying, our world is increasingly made of information. AI algorithms are increasingly going to be our interface to this world of information. And somebody will be in control of these algorithms. And that puts us in any kind of a bad situation, right? It has risks.
1:25:05 Speaker 01 It has risks coming from potentially large companies wanting to optimize their their own goals, maybe profit, maybe something else, also from governments who might want to use these algorithms as a means of control of their population.
1:25:21 Lex Fridman? Do you think there's existential threat that could arise from that?
1:25:24 Speaker 01 So existential threat. So maybe you're referring to the singularity narrative where robots just take over?
1:25:33 Lex Fridman? Well, I don't, not Terminator robots, and I don't believe it has to be a singularity. We're just talking to, just like you said, the algorithm controlling masses of populations. The existential threat being hurt ourselves, much like a nuclear war would hurt ourselves. That kind of thing. I don't think that requires a singularity, that requires a loss of control over AI algorithms.
1:26:00 Speaker 01 Yes. So I do agree there are concerning trends. Honestly, I wouldn't want to make any long-term predictions. I don't think today we really have the capability to see what the dangers of AI are going to be in 50 years, in 100 years. I do see that we are already faced with concrete and present dangers surrounding the negative side effects of content recommendation systems, of news feed algorithms, concerning algorithmic bias as well. So we are delegating more and more decision processes to algorithms.
1:26:42 Speaker 01 Some of these algorithms are uncrafted, some are learned from data, but we are delegating control. Sometimes it's a good thing, sometimes not so much. And there is in general very little supervision of this process, right? So we are still in this period of very fast change, even chaos, where society is restructuring itself, turning into an information society, which itself is turning into an increasingly automated information passing society. And well, yeah, I think the best we can do today is try to raise awareness around some of these issues. And I think we're actually making good progress.
1:27:25 Speaker 01 If you look at algorithmic bias, for instance, three years ago, even two years ago, very, very few people were talking about it. And now all the big companies are talking about it. They are often not in a very serious way, but at least it is part of the public discourse. You see people in Congress talking about it. So, and it all started from raising awareness.
1:27:49 Lex Fridman? Right. So in terms of alignment problem, trying to teach, as we allow algorithms, just even recommender systems on Twitter.
1:28:00 Lex Fridman? Encoding human values and morals, decisions that touch on ethics, how hard do you think that problem is? How do we have loss functions in neural networks that have some fuzzy components of human morals?
1:28:19 Speaker 01 Well, I think this is really all about objective function engineering, which is probably going to be increasingly a topic of concern in the future. Like for now, we are just using very naive loss functions because the hard part is not actually what you're trying to minimize, it's everything else. But as the everything else is going to be increasingly automated, we're going to be focusing our human attention on increasingly high-level components, like what's actually driving the whole learning system, like the objective function. So, loss function engineering is going to be... Loss function engineer is probably going to be a job title in the future, you know?
1:28:57 Lex Fridman? And then the tooling you're creating with Keras essentially takes care of all the details underneath and basically the human expert is needed for exactly that.
1:29:09 Speaker 01 That's the idea. Keras is the interface between the data you're collecting and the business goals. And your job as an engineer is going to be to express your business goals and your understanding of your business or your product, your system, as a kind of loss function or a kind of set of constraints.
1:29:28 Lex Fridman? Does the possibility of creating an AGI system excite you or scare you or bore you?
1:29:36 Speaker 01 So, intelligence can never really be general. At best, it can have some degree of generality, like human intelligence. It also always has some specialization in the same way that human intelligence is specialized in a certain category of problems, is specialized in the human experience. And when people talk about AGI, I'm never quite sure if they are talking about...
1:29:58 Speaker 01 Very, very smart AI, so smart that it's even smarter than humans. Or they're talking about human-like intelligence, because these are different things.
1:30:06 Lex Fridman? Let's say presumably I'm impressing you today with my humanness. So imagine that I was in fact a robot. So what does that mean? I'm impressing you with natural language processing. Maybe if you weren't able to see me, maybe this is a phone call.
1:30:24 Speaker 01 Yeah.
1:30:25 Lex Fridman? That kind of system.
1:30:26 Speaker 01 Okay, so.
1:30:27 Lex Fridman? Companion.
1:30:28 Speaker 01 So that's very much about building human-like AI. And you're asking me, you know, is this, is this an exciting perspective?
1:30:35 Lex Fridman? Yes.
1:30:36 Speaker 01 I think so, yes. Not so much because of what artificial human-like intelligence could do, but you know, from an intellectual perspective, I think if you could build truly human-like intelligence, that means you could actually understand human intelligence, which is fascinating, right? Yeah, human-like intelligence is going to require emotions. It's going to require consciousness, which is not things that would normally be required by an intelligent system. If you look at, you know, we were mentioning earlier like science as a superhuman problem-solving agent or system, it does not have consciousness, it doesn't have emotions in general.
1:31:19 Speaker 01 So emotions, I see consciousness as being on the same spectrum as emotions. It is a component of the subjective experience that is meant very much to guide behavior generation, right? It's meant to guide your behavior. In general, human intelligence and animal intelligence has evolved for the purpose of behavior generation, right? Including in a social context. So that's why we actually need emotions. That's why we need consciousness. An artificial intelligence system developed in a different context may well never need them. They will, they will never be conscious, like science.
1:31:59 Lex Fridman? Well, on that point, I would argue it's possible to imagine that there's echoes of consciousness in science when viewed as an organism, that science is consciousness.
1:32:12 Speaker 01 So, I mean, how would you go about testing this hypothesis? How do you probe the subjective experience of an abstract system like science?
1:32:24 Lex Fridman? Well, the point of probing any subjective experience is impossible because I'm not science. I'm Lex, so I can't probe another entities. The another, it's no more than bacteria on my skin.
1:32:37 Speaker 01 You're Lex, I can ask you questions about your subjective experience and you can answer me. And that's how I know you're conscious.
1:32:45 Lex Fridman? Yes, but that's because we speak the same language. You perhaps, we have to speak the language of science in order to ask.
1:32:52 Speaker 01 Honestly, I don't think consciousness, just like emotions or pain and pleasure, is not something that inevitably arises from any sort of sufficiently intelligent information processing. It is a feature of the mind, and if you've not implemented it explicitly, it is not there.
1:33:11 Lex Fridman? So you think it's a, it's a, it's an emergent feature of a particular architecture? So do you think?
1:33:17 Speaker 01 It's, it's, it's a feature in the same sense. So again, the subjective experience is all about guiding behavior. If you, if, if the problems you're trying to solve don't really involve embodied agents, maybe in a social context, generating behavior and pursuing goals like this. And if you get science, that's not actually what's, what's happening, even though it is, it is a form of artificial AI in this, artificial intelligence, in the sense that it is solving problems. It is accumulating knowledge, accumulating solutions and so on. So if you're not explicitly implementing a subjective experience, implementing a certain emotions
1:33:59 Speaker 01 and implementing consciousness, it's not going to just spontaneously emerge.
1:34:04 Lex Fridman? Yeah, but so for a system like, human-like intelligent system that has consciousness, do you think it needs to have a body?
1:34:13 Speaker 01 Yes, definitely. I mean, it doesn't have to be a physical body, right? And there's not that much difference between a realistic simulation and the real world.
1:34:20 Lex Fridman? Oh, so there has to be something you have to preserve, kind of thing?
1:34:23 Speaker 01 Yes, but human-like intelligence can only arise in a, in a human-like context. Intelligence.
1:34:29 Lex Fridman? You need other humans in order for you to demonstrate that you have human-like intelligence, essentially.
1:34:36 Speaker 01 Yes.
1:34:37 Lex Fridman? So what kind of tests and demonstration would be sufficient for you to demonstrate human-like intelligence?
1:34:48 Speaker 01 Yeah.
1:34:48 Lex Fridman? Just out of curiosity, you talked about in terms of theorem proving and program synthesis, I think you've written about that there's no good benchmarks for this.
1:34:57 Speaker 01 Yeah.
1:34:57 Lex Fridman? That's one of the problems. So let's, let's talk programs, program synthesis.
1:35:03 Speaker 01 So.
1:35:03 Lex Fridman? What do you imagine is a good, I think it's related questions for human-like intelligence and for program synthesis. What's a good benchmark for either or both?
1:35:12 Speaker 01 Right, so I mean, you're actually asking, asking two questions, which is, one is about quantifying intelligence and comparing the intelligence of an artificial system to the intelligence for human. And the other is about a degree to which this intelligence is human-like. It's actually two different questions. So if you look, you mentioned earlier the Turing test.
1:35:36 Lex Fridman? Right.
1:35:36 Speaker 01 Well, I actually don't like the Turing test because it's very lazy. It's, it's all about completely bypassing the problem of defining and measuring intelligence.
1:35:45 Lex Fridman? Right.
1:35:45 Speaker 01 And instead delegating to a human judge or a panel of human judges. So it's, it's, it's a total cop out, right? If you want to, to measure how human-like an agent is, I think you have to make it interact with other humans. Maybe it's, it's not necessarily a good idea to have these other humans be the judges. Maybe you should just observe behavior and compare it to what a human would actually have done. When it comes to measuring how smart, how clever an agent is and comparing that to the, to, to the degree of, of human intelligence. So we're already talking about two things, right?
1:36:30 Speaker 01 The degree, kind of like the magnitude, the magnitude of an intelligence and its direction, right? Like the norm of a vector.
1:36:38 Lex Fridman? Right.
1:36:38 Speaker 01 And its direction.
1:36:40 Lex Fridman? Yeah.
1:36:40 Speaker 01 And the direction is like human likeness and the magnitude, the norm is intelligence. You could call it intelligence, right?
1:36:51 Lex Fridman? So the, the direction, your sense, the, the space of directions that are human-like is very narrow.
1:36:58 Speaker 01 Yeah. So the, the way you would measure.
1:37:03 Speaker 01 The magnitude of intelligence in a system in a way that also enables you to compare it
1:37:10 Speaker 01 to that of a human. Well, if you look at different benchmarks for intelligence today, they are all too focused on skill at a given task. Let's scale at playing chess, scale at playing Go, scale at playing Dota. And I think that's not the right way to go about it because you can always be too human at one specific task. The reason why our skill at playing Go or at juggling or anything is impressive is because we are expressing this skill within a certain set of constraints. If you remove the constraints, the constraints that we have one lifetime, that we have this body and so on. If you remove the context,
1:37:54 Speaker 01 if you have unlimited train data, if you can have access to, you know, for instance, if you look at juggling, if you have no restriction on the hardware, then achieving arbitrary levels of skill is not very interesting and says nothing about the amount of intelligence you've achieved. So if you want to measure intelligence, you need to rigorously define what intelligence is, which in itself, you know, it's a very challenging problem.
1:38:21 Lex Fridman? And do you think that's possible?
1:38:22 Speaker 01 To define intelligence, yes, absolutely. I mean, you can provide, many people have provided, you know, some definition. I have my own definition.
1:38:30 Lex Fridman? Where does your definition begin if it doesn't end?
1:38:33 Speaker 01 Well, I think intelligence is essentially the efficiency with which you turn experience into generalizable programs. So what that means is it's the efficiency with which you turn a sampling of experience base into the ability to process a larger chunk of experience space. So measuring skill can be one proxy. Of course, many, many different tasks can be one proxy for measure intelligence. But if you want to only measure skill, you should control for two things. You should control for
1:39:19 Speaker 01 The amount of experience that your system has and the priors that your system has. But if you control, if you look at two agents and you give them the same priors and you give them the same amount of experience, there is one of the agents that is gonna learn programs, representations, something, a model, that will perform well on the larger chunk of experience space than the other. And that is the smarter agent.
1:39:49 Lex Fridman? Yeah, so if you fix the experience which generate better programs, better meaning more generalizable. That's really interesting. And that's a very nice, clean definition of...
1:39:59 Speaker 01 Oh, by the way, in this definition, it is already very obvious that intelligence has to be specialized because you're talking about experience space. And you're talking about segments of experience space. You're talking about priors and you're talking about experience. All of these things define the context in which intelligence emerges.
1:40:21 Speaker 01 And you can never look at the totality of experience space, right? So intelligence has to be specialized.
1:40:29 Lex Fridman? But it can be sufficiently large, the experience space, even though specialized. There's a certain point when the experience space is large enough to where it might as well be general. It feels general, it looks general.
1:40:41 Speaker 01 Sure, I mean, it's very relative. Like, for instance, many people would say human intelligence is general. In fact, it is quite specialized. You know, we can definitely build systems that start from the same innate priors as what humans have at birth. Because we already understand fairly well what sort of priors we have as humans. Like, many people have worked on this problem, most notably Elizabeth Spelker from Harvard, I don't know if you know her. She's worked a lot on what she calls a core knowledge. And it is very much about trying to determine and describe what priors we are born with.
1:41:19 Lex Fridman? Like, language skills and so on, all that kind of stuff.
1:41:21 Speaker 01 Exactly.
1:41:24 Speaker 01 So we have some pretty good understanding of what priors we are born with. So we could... So I've actually been working on a benchmark for the past couple of years, you know, on and off. I hope to be able to release it at some point. The idea is to measure the intelligence of systems by constraining for priors, constraining for amount of experience, and by assuming the same priors as what humans are born with, so that you can actually compare these scores to human intelligence, and you can actually have humans pass the same test in a way that's fair.
1:42:00 Lex Fridman? Yeah.
1:42:01 Speaker 01 And so importantly, such a benchmark should be such that
1:42:08 Speaker 01 any amount of practicing does not increase your score. So try to picture a game where no matter how much you play this game, that does not change your skill at the game. Can you picture that?
1:42:25 Lex Fridman? As a person who deeply appreciates practice, I cannot actually. I cannot, I cannot, I...
1:42:33 Speaker 01 There's actually a very simple trick. So in order to come up with a task, so the only thing you can measure is skill at a task.
1:42:41 Lex Fridman? Yes.
1:42:41 Speaker 01 All tasks are going to involve priors.
1:42:44 Lex Fridman? Yes.
1:42:45 Speaker 01 The trick is to know what they are and to describe that. And then you make sure that this is the same set of priors as what humans start with. So you create a task that assumes these priors, that exactly documents these priors. So that's the priors are made explicit, and there are no other priors involved. And then you generate a certain number of samples in experience base for this task, right? And this, for one task, assuming that the task is new for the agent passing it, that's one test of this.
1:43:20 Speaker 01 Definition of intelligence that we set up. And now you can scale that to many different tasks. That all, you know, each task should be new to the agent passnet. And also should be human interpretable and understandable. So that you can actually have a human pass the same test. And then you can compare the score of your machine and the score of your human.
1:43:39 Lex Fridman? Which could be a lot of, it could even start a task like MNIST. Just as long as you start with the same set of priors.
1:43:46 Speaker 01 Yeah, so the problem with MNIST, humans are already trained to recognize digits. Right? And, but let's say, let's say you are considering objects that are not digits. Some completely arbitrary patterns. While humans already come with visual priors about how to process that. So in order to make the game fair, you would have to
1:44:11 Speaker 01 isolate these priors and describe them and then express them as computational rules.
1:44:15 Lex Fridman? Having worked a lot with vision science people, that's exceptionally difficult. A lot of progress has been made. There's been a lot of good tests. And basically reducing all of human vision into some good priors. And we're still probably far away from that perfectly. But as a start for a benchmark, that's an exciting possibility.
1:44:32 Speaker 01 Yeah, so Elizabeth Pelker actually lists
1:44:38 Speaker 01 objectness as one of the core knowledge priors.
1:44:41 Lex Fridman? Objectness, cool.
1:44:43 Speaker 01 Objectness, yeah. So we have priors about objectness, like about the visual space, about time, about agents, about goal-oriented behavior. We have many different priors. But what's interesting is that sure, we have, you know, this, this pretty diverse and, and, and rich set of priors. But it's also not that diverse, right? We are not born into this world with a ton of knowledge about the world. With only a small set of.
1:45:15 Speaker 01 Core knowledge, right?
1:45:16 Lex Fridman? Yeah, it's hard. Do you have a sense of how it feels to us humans that that set is not that large, but just even the nature of time that we kind of integrate pretty effectively through all of our perception, all of our reasoning, maybe, how, you know, do you have a sense of how easy it is to encode those priors? Maybe it requires building a universe and then the human brain in order to encode those priors? Or do you have a hope that it's can be listed like an axiomatically?
1:45:47 Speaker 01 I don't think so. So you have to keep in mind that any knowledge about the world that we are born with is something that has to have been encoded into our DNA by evolution at some point.
1:45:59 Lex Fridman? Right.
1:46:00 Speaker 01 And DNA is a very, very low bandwidth medium, like it's extremely long and expensive to encode anything into DNA because first of all, you need some sort of evolutionary pressure to guide this writing process. And then, you know, the higher level of information you're trying to write, the longer it's going to take. And the thing in the environment that you're trying to encode knowledge about has to be stable over this duration.
1:46:33 Lex Fridman? Yes.
1:46:34 Speaker 01 So you can only encode into DNA things that constitute an evolutionary advantage. So this is actually a very small subset of all possible knowledge about the world. You can only encode things that are stable, that are true, over very, very long periods of time, typically millions of years. For instance, we might have some visual prior about the shape of snakes, right? But all the what makes a face, what's the difference between a face and an ant face? But consider this interesting question. Do we have any innate sense of the visual difference between a male face and a female face? What do you think? For human, I mean.
1:47:18 Lex Fridman? I would have to look back into evolutionary history when the genders emerged, but yeah, most.
1:47:25 Speaker 01 I mean, the faces of humans are quite different from the faces of great apes. Great apes, right?
1:47:31 Lex Fridman? Yeah.
1:47:32 Speaker 01 Like you couldn't tell, you couldn't tell the face of a female chimpanzee from the face of a male chimpanzee, probably.
1:47:40 Lex Fridman? Yeah, and I don't think us humans have all that ability.
1:47:43 Speaker 01 So we do have innate knowledge of what makes a face, but it's actually impossible for us to have any DNA encoded knowledge of the difference between a female human face and a male human face. Because that knowledge, that information came up into the world actually very recently. If you look at the slowness of the process of encoding knowledge into DNA.
1:48:13 Lex Fridman? Yeah, so that's interesting. That's a really powerful argument. The DNA is a low bandwidth and it takes a long time to encode.
1:48:19 Speaker 01 Yeah.
1:48:19 Lex Fridman? That naturally creates a very efficient encoding.
1:48:22 Speaker 01 And one important consequence of this is that, so yes, we are born into this world with a bunch of knowledge, sometimes high level knowledge about the world, like the shape, the rough shape of a snake, of the rough shape of face. But importantly, because this knowledge takes so long to write, almost all of this innate knowledge is shared with our cousins, with great apes, right? So it is not actually this innate knowledge that makes us special.
1:48:54 Lex Fridman? But to throw it right back at you from earlier on in our discussion, it's that encoding might also include the entirety of the environment of Earth.
1:49:07 Speaker 01 To some extent, so it can include things that are important to survival and production, so for which there is some evolutionary pressure, and things that are stable, constant over very, very, very long time periods. And honestly, it's not that much information. There's also, besides the bandwidth constraint and constraints of the writing process, there's also memory constraints. Like DNA, the part of DNA that deals with the human brain, it's actually fairly small. It's like, you know, on the order of megabytes, right? There's not that much high level knowledge about the world you can encode.
1:49:48 Lex Fridman? That's quite brilliant and hopeful for a benchmark of that you're referring to of encoding priors. I actually look forward to I'm skeptical whether you can do it in the next couple of years, but hopefully.
1:50:02 Speaker 01 I've been working. So honestly, it's a very simple benchmark and it's not like a big Brex or anything. It's more like a fun, a fun side project, right?
1:50:10 Lex Fridman? But these fun, so is ImageNet. That these fun side projects could launch entire groups of efforts towards towards creating reasoning systems and so on. And I think.
1:50:22 Speaker 01 Yeah, that's the goal. It's trying to measure strong generalization, to measure the strength of abstraction in our minds. Well, in our minds and in a, in a, in a artificial intelligent agent.
1:50:34 Lex Fridman? And if there's anything true about this science organism is its individual cells love competition. So and benchmarks encourage competition. So that's a, that's an exciting possibility. If you, do you think an AI winter is coming? And how do we prevent it?
1:50:52 Speaker 01 Not really. So an AI winter is something that would occur when there's a big mismatch between how we are selling the capabilities of AI and the, and the actual capabilities of AI. And today, so deep learning is creating a lot of value. And it will keep creating a lot of value in the sense that these models are applicable to very wide range of problems that are relevant today. And we are only just getting started with applying these algorithms to every problem they could be solving. So deep learning will keep creating a lot of value for the time being. What's concerning, however, is that there's a lot of hype around deep learning and around AI. There are lots of people are overselling the capabilities
1:51:37 Speaker 01 of these systems, not just the capabilities, but also overselling the the fact that they might be more or less, you know, brain like. Like given a kind of a mystical aspect of these technologies and also overselling the pace of progress, which you know, it might look fast in the sense that we have this exponentially increasing number of papers. But again, this is just a simple consequence of the fact that we have ever more people coming into the field. It doesn't mean the progress is actually exponentially fast. Like, let's say you're trying to raise money for your startup or your research lab.
1:52:19 Speaker 01 You might want to tell, you know, a grandiose story to investors about how deep learning is just like the brain and how it can solve all these incredible problems like self-driving and robotics and so on. And maybe you can tell them that the field is progressing so fast and we are going to have AGI within 15 years or even 10 years. And all, none of this is true. And every time you're like saying these things and an investor or you know, a decision maker believes them. Well, you're, this is like the equivalent of taking on credit card debt. But for, for trust, right?
1:53:00 Speaker 01 And maybe this, this will, you know, uh, uh, this will, this will be what enables you to raise a lot of money. But ultimately you are creating damage, you are damaging the field.
1:53:12 Lex Fridman? That's the concern is that, that, that's what happens with the other AI winters. Is the concern is you actually tweeted about this with autonomous vehicles, right? There's almost every single company now have promised that they will have full autonomous vehicles by 2021, 2022.
1:53:28 Speaker 01 That's a good example of, of the, the consequences of overhyping the capabilities of AI and the pace of progress.
1:53:36 Lex Fridman? So because I work especially a lot recently in this area, I have a deep concern of what happens when all of these companies, after they've invested billions, have a meeting and say, how much did we actually? First of all, do we have an autonomous vehicle? The answer will definitely be no. And second would be, wait a minute, we've invested one, two, three, four billion dollars into this and we made no profit. And the reaction to that. Maybe going very hard in another direction.
1:54:05 Speaker 01 Yes.
1:54:05 Lex Fridman? That might impact even other industries.
1:54:07 Speaker 01 And that's what we call an AI winter is when there is backlash where no one believes any of these promises anymore because they've turned out to be big lies the first time around.
1:54:16 Lex Fridman? Yeah.
1:54:17 Speaker 01 And this will definitely happen to some extent for autonomous vehicles. Because the public and decision makers have been convinced that, you know, around 2015, they've been convinced by these people who are trying to raise money for their startup and so on, that L5 driving was coming in maybe 2016, maybe 2017, maybe 2018. Now, when 2019, we're still waiting for it. And so I don't believe we are going to have a full on AI winter because we have these technologies that are producing a tremendous amount of real value, right?
1:54:54 Lex Fridman? Yes.
1:54:54 Speaker 01 But there is also too much hype, so there will be some backlash, especially there will be backlash.
1:55:02 Speaker 01 So, you know, some startups are trying to sell the dream of AGI, right? And the fact that AGI is going to create infinite value, like AGI is like a freelance. Like, if you can, if you can develop an AI system that passes a certain threshold of IQ or something, then suddenly you have infinite value, right?
1:55:21 Lex Fridman? Yes.
1:55:23 Speaker 01 And, well, there are actually lots of investors buying into this idea. And, you know, they will wait maybe, maybe 10, 15 years and nothing will happen. And, and the next time around, well, maybe, maybe there will be a new generation of investors. No one will care. You know, a human memory is fairly short after all.
1:55:44 Lex Fridman? I don't know about you, but, because I've spoken about AGI sometimes poetically, like, I get a lot of emails from people giving me, they're usually like a large manifestos of, they've, they say to me that they have.
1:56:02 Lex Fridman? Created an AGI system where they know how to do it. And there's a long write up of how to do it.
1:56:06 Speaker 01 I get a lot of these emails, yeah.
1:56:07 Lex Fridman? They're little bits feel like it's generated by an AI system, actually. But there's usually no.
1:56:15 Speaker 01 Maybe that's recursively self improving AI. It's you have a transformer generating crank papers about AGI.
1:56:23 Lex Fridman? So what the question is about, because you've been such a good, you have a good radar for crank papers. How do we know they're not on to something? How do I?
1:56:36 Lex Fridman? So when you start to talk about AGI or anything like the reasoning benchmarks and so on. So something that doesn't have a benchmark, it's really difficult to know. I mean, I talked to Jeff Hawkins, who's really looking at neuroscience approaches to how. And there's some, there's echoes of really interesting ideas in at least Jeff's case, which he's showing. How do you usually think about this? Like preventing yourself from being too narrow minded and elitist about, you know, deep learning. It has to work on these particular benchmarks, otherwise it's trash.
1:57:13 Speaker 01 Well, you know, the thing is, intelligence does not exist in the abstract. Intelligence has to be applied. So if you don't have a benchmark, if you're not doing improvement and some benchmark, maybe it's a new benchmark. Right? Maybe it's not something I've been looking at before. But you do need a problem that you're trying to solve. You're not going to come up with a solution without a problem.
1:57:38 Lex Fridman? So you general intelligence, I mean, you've clearly highlighted generalization. If you want to claim that you have an intelligent system, it should come with a benchmark.
1:57:48 Speaker 01 It should, yes, it should display capabilities of some kind. It should, it should show that it can create some form of value, even if it's a very artificial form of value. And that's also the reason why you don't actually need to care about telling which papers have actually submitted potential and which do not. Because if.
1:58:12 Speaker 01 If there is a new technique that's actually creating value, you know, this is going to be brought to light very quickly because it's actually making a difference. So it's the difference between something that is ineffective and something that is actually useful. And ultimately usefulness is our guide, not just in this field, but if you look at science in general. Maybe there are many, many people over the years that have had some really interesting theories of everything, but they were just completely useless. And you don't actually need to tell the interesting theories from the useless theories. All you need is to see, you know, is this actually having an effect on something else? You know, is this actually useful? Is this making an impact or not?
1:58:54 Lex Fridman? That's beautifully put. I mean, the same applies to quantum mechanics, to string theory, to the holographic principle.
1:59:00 Speaker 01 We are doing deep learning because it works. You know, that's like before it started working, people, you know, considered people working on neural networks as cranks very much. Like, you know, no one was working on this anymore. And now it's working, which is what makes it valuable. It's not about being right, right? It's about being effective.
1:59:19 Lex Fridman? And nevertheless, the individual entities of this scientific mechanism, just like Yoshua Banjo, Yann Le Cun, they, while being called cranks, stuck with it, right?
1:59:30 Speaker 01 Yeah.
1:59:30 Lex Fridman? And so us individual agents, even if everyone's laughing at us, should stick with it.
1:59:35 Speaker 01 If you believe you have something, you should stick with it and see it through.
1:59:40 Lex Fridman? That's a beautiful inspirational message to end on. Francois, thank you so much for talking today. That was amazing.
1:59:45 Speaker 01 Thank you.
Previous Vijay Kumar: Flying Robots Next Colin Angle: iRobot Page updated October 26, 2026 · transcript and summary are machine-generated
