WTF is Artificial Intelligence Really? | Yann LeCun x Nikhil Kamath | People by WTF Ep #4

1. 0:00 – 1:50

   Yann’s Intro

   1. NKNikhil Kamath0:00

      [upbeat music] I thought we could use today to figure out, A, what is AI? How did we get here? What likely next is. [upbeat music] As an Indian twenty-year-old who wants to build a business in AI, a career in AI, what do we do?

   2. YLYann LeCun0:52

      Today?

   3. NKNikhil Kamath0:53

      Today.

   4. YLYann LeCun0:53

      Like, right now?

   5. NKNikhil Kamath0:53

      Yeah. [upbeat music] Hi, Yann. Good morning.

   6. YLYann LeCun1:05

      And you too.

   7. NKNikhil Kamath1:06

      Uh, thank you for doing this.

   8. YLYann LeCun1:08

      Pleasure.

   9. NKNikhil Kamath1:09

      The very first thing we like to do is get to know you a bit more, uh, how you came to be what you are today. Uh, could you tell us a little bit about where you were born, where you grew up, leading up to today?

   10. YLYann LeCun1:23

       So, I, I grew up near Paris, uh, in the suburbs. Um, my dad was an engineer, and I learned almost everything from him. [chuckles] Um, and, um, [clears throat] um, always was interested in, in science and technology since I was a little kid. And, [clears throat] and always saw myself as, uh, perhaps becoming an engineer. I had no idea how you became a scientist, uh, but I became interested in this afterwards.

   11. NKNikhil Kamath1:49

       What

2. 1:50 – 3:15

   Difference between an Engineer and a Scientist

   1. NKNikhil Kamath1:50

      is the difference between an engineer and a scientist?

   2. YLYann LeCun1:52

      Well, um, it's very difficult to, [chuckles] to define, and, uh, very often you have to be a little bit of both.

   3. NKNikhil Kamath1:58

      Mm-hmm.

   4. YLYann LeCun1:59

      [clears throat] But, uh, um, scientists, you try to understand the world. Um, engineer, you try to create new things, and very often, if you want to understand the world, you need to create new things. The progress of science very much is linked with progress in technology that allows to collect data. You know, the invention of the telescope allowed the discovery of planets, and that planets are, um, rotating around the sun and things like this, right? The microscope opened the door to all kinds of things. So, um, so technology enables science, and for... The problem that really has been my obsession [chuckles] uh, for a long time, is, uh, discovering the mysteries of, uncovering the mysteries of intelligence. Um, and as, as an engineer, I think the, the only way to do this is to build a machine that is intelligent, right? So there's both an aspect of, a scientific aspect of understanding intelligence, what it is, um, at a theoretical level and more practical, uh, side of things. And then, um, of course, the consequences of building intelligent machines could be, could have, you know, could be really important for humanity.

   5. NKNikhil Kamath3:12

      And school in Paris, studying what?

3. 3:15 – 4:05

   Yann’s interest in AI and Mathematics

   1. YLYann LeCun3:15

      So I, I studied electrical engineering.

   2. NKNikhil Kamath3:18

      Mm-hmm.

   3. YLYann LeCun3:18

      Um, but as I progressed in my studies, I became more and more interested in sort of, uh, more fundamental questions in mathematics, physics, and, and AI.

   4. NKNikhil Kamath3:29

      Mm-hmm.

   5. YLYann LeCun3:29

      Um, I, I did not study computer science. [chuckles]

   6. NKNikhil Kamath3:32

      Right.

   7. YLYann LeCun3:33

      Uh, of course, there is always computers involved when you studied electrical engineering, even in the 1980s, and sev-- late '70s, actually, when I started. Um, but, um, but I got to do a few independent projects with mathematics professors on, on the questions of AI and, and, and things like that, and, uh, I really got hooked into research. I, I was, uh, um... You know, my, my, uh, m- my, my favorite activity is to, to build new things, invent new things-

   8. NKNikhil Kamath4:03

      Mm-hmm

   9. YLYann LeCun4:03

      ... and then understand new things

4. 4:05 – 5:22

   Godfather of AI | Yann’s feelings about it

   1. YLYann LeCun4:05

      in a new way.

   2. NKNikhil Kamath4:06

      When somebody says godfather of AI, the term, how does it make you feel? What do you think about it?

   3. YLYann LeCun4:15

      Uh, I mean, I don't particularly [chuckles] like this term. You know, I, I live in New Jersey. Godfather in New Jersey means you are, you belong to the mafia, [chuckles] right? I mean, it, i- science is never, uh, a sort of individual, uh, pursuit. I- you, you, you make progress by, by the collision of ideas from m- multiple people, and you, you, you, you, you make... You do make hypothesis, and then you try to show that your hypothesis is, is correct by demonstrating that the idea you have, the mental model of what should work, um, is correct, by demonstrating that, uh, that it works, um, or doing some theory and things like that. Um, and, um, it, it's not an isolated, uh, activity, so there's always a lot of people who have contributed to, uh, to progress. But then, because of the nature of how [chuckles] the world works, we only remember just a few people. Um, uh, I think a lot of the credit should go to a lot more people. It's just that we don't have a good, you know, memory for attributing credit to, to a lot of people.

5. 5:22 – 6:00

   Teaching & fame at NYU

   1. NKNikhil Kamath5:22

      So how does it feel to be a teacher today, Yann? When you were at NYU, are you the celebrity at NYU?

   2. YLYann LeCun5:30

      Um, let's say over the last, uh, several [chuckles] years, uh, students come up to me at the end of the class and want to take selfies. [chuckles]

   3. NKNikhil Kamath5:39

      Yeah.

   4. YLYann LeCun5:40

      So, so there's a little bit of that. I, I think, um, if you are in the same room with someone, I think it's important to sort of, uh, make the session interactive, because otherwise you can just watch a video. Um, so, so that's why, uh, yeah, that, that, that's what I try to do, really sort of engage with

6. 6:00 – 7:46

   Heroes in Science

   1. YLYann LeCun6:00

      the students.

   2. NKNikhil Kamath6:01

      Do you suspect being a hero in academia, in research?... is much like being a hero in sport or entrepreneurship, or do you think it's harder?

   3. YLYann LeCun6:13

      Okay, there's something I'm, I'm happy about. The fact that there can be heroes, you know, from, in science, um, academia or not.

   4. NKNikhil Kamath6:22

      One can argue there was Newton and Einstein and all these people, right?

   5. YLYann LeCun6:25

      Well, Newton was not really kind of a public figure, I, I think.

   6. NKNikhil Kamath6:29

      Mm-hmm.

   7. YLYann LeCun6:30

      Um, I mean, he was at Cambridge.

   8. NKNikhil Kamath6:31

      Einstein was.

   9. YLYann LeCun6:31

      But Einstein certainly was, yeah. Um, and to some extent, you know, other, some other scientists also were, were sort of minor celebrities. [chuckles] Um, so I mean, I think, uh, some of that comes from, you know, scientific production, but frankly, there's a lot of people who have made scientific contributions that are completely unknown, um, and which I find it a little sad. But, um, uh, I, I think a lot of people who have become prominent in, in, in science and technology is not just because of the science they've, they've produced, but also because of their public stance and, um, out there. You know, one, one, one thing that perhaps differentiates me from other scientists who are a little quieter, is that I'm very present on s- social networks, and I give public talks, and I have strong opinions about not just technical issues, but also, uh, policy issues to some extent. Uh, so I, I, that, that I think amplifies a little bit the-

   10. NKNikhil Kamath7:33

       Yeah

   11. YLYann LeCun7:33

       ... the popularity or unpopularity for, in certain circles, I'm seen like a complete idiot. [chuckles]

   12. NKNikhil Kamath7:40

       I've watched a lot of your interviews over the last fortnight, last month, in fact.

7. 7:46 – 10:18

   Three problems with the world - Yann’s lens

   1. NKNikhil Kamath7:47

      If you were to state three problems with the world from Yann's lens, what would they be?

   2. YLYann LeCun7:55

      Um, so as, [chuckles] as a scientist, you try to establish causal models of the world, right? So there are effects that we're seeing, and then the question is, what is it caused by? And almost, for al- for almost every problem that we have, the, the, uh, the cause is really a lack of knowledge or, or intelligence by humans. We're making mistakes. We're making mi- mistakes because we're not smart enough to figure out we have a problem, because we're not smart enough to figure out solutions. We're not smart enough to organize ourselves to find solutions, right? So things like, I mean, climate change is a huge issue, right?

   3. NKNikhil Kamath8:37

      Mm-hmm.

   4. YLYann LeCun8:37

      Um, and, um, you know, there might... There are, you know, political issues with that and, um, questions of organizing, uh, the world, the governments, et cetera, but also, uh, potentially, there are technological solutions to, to climate change, and, and I wish we were s- you know, smarter, so, so that we could find solutions faster. Um-

   5. NKNikhil Kamath9:00

      So are you saying humans don't know why we do what we do, and that's the problem?

   6. YLYann LeCun9:05

      No, I think the mistakes we're making is because, um, uh, if we were s- if we were a little smarter, if we had a bet- better mental model of, of how the world works, uh, and that's the central question in AI as well [chuckles] , uh, I think we could, we could solve our, our problems better. We would take decisions that are more rational. Um, and what I s- the, the, the, the, the big issue I see in the world today is, is, is people who, um, are not interested in finding the facts, are not interested in educating themselves. Um, or maybe they are, but they don't have the means to do this. Uh, they don't have access to information and knowledge. So I think the best thing we can do, and maybe that's why, you know, I became a professor, [chuckles] is to, uh, make people smarter. And to some extent, that's the best reason also to work on AI, because AI is going to amplify human intelligence, I mean, the, the overall intelligence of humanity, if you want. So I, I, I think that, uh, that's the key to solving a lot of the problems that we have.

   7. NKNikhil Kamath10:11

      So just to preface this conversation, uh, I'm an idiot. When it comes to anything around

8. 10:18 – 13:13

   What is AI and how did we get here?

   1. NKNikhil Kamath10:18

      AI or technology, there isn't much that I know, and I've tried to learn, uh, over the past, well, over the very recent past, and, uh, I have a lot of curiosity for it, but I don't know enough about it. A lot of the people watching us today are wannabe entrepreneurs, primarily based out of India. A lot of us have heard conjecture around AI. We have heard about the edge cases, both on the positive side and the negative side. I thought we could use today to figure out, for all of us, A, what is AI? How did we get here, and what likely next? If I were to break today down into three parts. Should we start with what is AI?

   2. YLYann LeCun11:14

      Okay, um, that's a good question. [chuckles] What is intelligence, even? Um, so in the history of AI, I think the problem of what is AI feels a little bit like the, the story of the blind men with the elephant, right? [chuckles] Uh, that there are v- very different aspects to intelligence, and, uh, over the history of AI, people have addressed one view of what intelligence is, and, and, and basically ignored all the other, all the other aspects. So, um, one of the early aspects of, uh, of intelligence that people addressed with AI in the 1950s was, you know, intelligence is about reasoning. Um, how do we reason? Um, so how do we reason logically? Um, how do we search for solutions to a, a new problem? And in the '50s, pe- people figured out, um-... Well, we have a, a problem. Let's say that's become a standard problem in, uh, in AI. It's, or computer science now. Um, you know, I give you a bunch of cities, and I ask you, you have to go through every single city, and what's the shortest path, the, the shortest circuit to go around the city? That's called a traveling salesman problem. Um, and they say, like, every reasoning problem can be formulated in terms of searching for a solution to a problem. There's a space of possible solution. There is something that tells you whether you found a good solution or not, or, or some number that tells you the length of the path, and you just have to search for the shortest path, right? And, and to some extent, you could reduce every reasoning problem to a problem of, of this type. In mathematics, we call this optimization. Okay? So you, you, you have, um, a problem. You can, you can evaluate whether your problem is solved or not with a number that indicates, you know, it's low if your length of your path is small, and it's high if it's longer, and you search for a solution that minimizes that, uh, that, the, the length of the path.

   3. NKNikhil Kamath13:10

      So is finding solutions related to intelligence? If

9. 13:13 – 15:00

   What is intelligence? | The Elephant Analogy

   1. NKNikhil Kamath13:13

      you were to ask me, "What is intelligence?" I would be, like, dumbfounded in trying to define it-

   2. YLYann LeCun13:19

      Yeah

   3. NKNikhil Kamath13:19

      ... in a sentence.

   4. YLYann LeCun13:20

      Right. I mean, the, so that comes back to the elephant analogy. [chuckles]

   5. NKNikhil Kamath13:25

      Can you explain the elephant analogy?

   6. YLYann LeCun13:26

      Well, so you know the, the blind, the, the blind man and the elephant, right? So you, [chuckles] the first blind man who goes to the side of the elephant and say, "That sounds like a wall-- looks like a wall." And then one goes to a leg, "That looks like a tree." Um, and another one, you know, touches the trunk. "That's a, that's a pipe." And nobody has a complete picture of what an elephant is, right? And you, you see it from, from the various angles. So this aspect of intelligence as being a search for a solution to a particular problem is, you know, a small piece of the elephant. It's, uh, it's one aspect of intelligence, but it's not, it's not the entire thing. Uh, but in the fifties, um, one branch of, of AI was basically only concerned by this. Um, and, and that branch was essentially dominant until, until the nineteen nineties. Um, that, that AI, uh, consists in searching for a solution, for plans, you know, if you want to, you know, stack a bunch of objects on top of each other, and some objects are bigger than others, you have to sort of organize the order in which you're gonna stack the objects. You know, you search for a sequence of actions to arrive at, at a goal. That's called planning.

   7. NKNikhil Kamath14:34

      Mm-hmm.

   8. YLYann LeCun14:35

      Or even, um, let's say, uh, you have a robot arm, and you have to grab an object, but there is, there's, you know, obstacles in front of it. You have to plan a trajectory for the arm to grab, to grab the object. Um, so all of that is planning. That's part of this searching for a solution to a problem. Um, but that part of AI, which again, was started in the fifties and was dominant until the nineties, uh, completely ignored

10. 15:00 – 16:20

    AI - perception & understanding

    1. YLYann LeCun15:00

       things like perception. Like, how do we understand the world?

    2. NKNikhil Kamath15:03

       Mm-hmm.

    3. YLYann LeCun15:04

       Um, how do we recognize an object? Um, how do we separate an object from its background, so we can identify it? Um, um, and, uh, you know, how, how, how do we think, um, not in terms of logic or, or, or search, but perhaps in more abstract, uh, abstract terms? And so that was essentially ignored, but, um, there was another branch of AI, also started in the fifties, um, that said, "Well, let's try to reproduce the mechanisms of intelligence that, that we see in animals and humans." And i- animals and humans have brains. Um, the brains basically organize themselves. They, they learn, right? They're, they're not spontaneously smart, and the intelligence is sort of a em- emerging, emergent phenomenon of networks of very simple elements in large numbers that are connected with each other. Um, so i- in the fifties or forties, people started discovering that intelligence and, and memory comes from the strength of the connections between neurons in a sort of simplified manner, and the way the brain learns is by modifying the strengths of the connections between neurons. So, so some people came up with

11. 16:20 – 17:30

    The two branches of AI - solving & learning

    1. YLYann LeCun16:20

       sort of theoretical models and, and actually electronic circuits that reproduce this. You know, can we build a neural network?

    2. NKNikhil Kamath16:25

       So intelligence, you're saying, was largely the ability to solve a certain problem?

    3. YLYann LeCun16:30

       So that's the first view, right?

    4. NKNikhil Kamath16:32

       Mm.

    5. YLYann LeCun16:32

       To solve particular problems that are, that, that were given. The second one is the ability to learn.

    6. NKNikhil Kamath16:38

       Right.

    7. YLYann LeCun16:38

       Okay? And that created those two branches of AI.

    8. NKNikhil Kamath16:41

       Right.

    9. YLYann LeCun16:41

       Um, so the, the, the one that started with the ability to learn, um, there was some success in the late fifties, early sixties, and it died in the late sixties because, uh, the type of learning procedures, uh, for, for those neural networks that people devised in the, in the sixties turned out to be extremely limited. You-- There was no way you could use this to produce truly intelligent machines, but it had a lot of consequences in, uh, various parts of engineering, uh, uh, a field of engineering called pattern recognition. Um, where-

    10. NKNikhil Kamath17:12

        So you're saying now that intelligence is the ability of a system to learn as well? The-

    11. YLYann LeCun17:16

        To learn, and, and the simplest situation in which you need machines to learn is for perception, interpreting, um, images, interpreting sounds.

    12. NKNikhil Kamath17:28

        And what did computers

12. 17:30 – 20:15

    Emergence of classical computer science | Heuristic programming

    1. NKNikhil Kamath17:30

       use to do that?

    2. YLYann LeCun17:31

       So for that, w-- it's, it's basically what caused the emergence of what we could call Classical computer science.

    3. NKNikhil Kamath17:38

       Mm-hmm.

    4. YLYann LeCun17:39

       Okay, you write a program, and that program basically internally searches for a solution and has some way of checking whether the solution it, it proposes is good or not. Um, uh, people had a name for this in the sixties. They called this, uh, Heuristic programming because you can't... You can never exhaustively search all solutions for a good one, 'cause the number of solution is ridiculously large. You know, at chess, for example, right, you, you can play a certain number of moves, but then for every move that you play, your opponent can play a certain number of moves. And then for every of those moves, you can play a certain number of moves. So you get this ex- exponential explosion of the number of possible-... trajectories, basically, or sequences of moves, and, uh, you cannot possibly explore all of them until the end of the game to figure out which move to, uh, uh, to play first. So, so you have to use what, what, you know, what's called heuristics to, to basically not search the entire, uh, graph or tree of possibilities.

    5. NKNikhil Kamath18:35

       So we'll put up a graph explaining this, but what you're saying in heuristics AI is you would have a user who would put in an input, there would be a bunch of rules, and you would use like a tree search or a expert AI, which would run a function like, if this, then that, if not, then this, to try and get to an end state.

    6. YLYann LeCun18:55

       Yeah. So something that would-

    7. NKNikhil Kamath18:57

       But a defined end state.

    8. YLYann LeCun18:59

       Uh, be defined and, and the, the, the, the program would be completely written by a person, um, and, uh, and the, the difference between a good and a bad [chuckles] system would be in how, uh, smart the system is in, in searching for a good solution without doing exhaustive search. Okay? That's the heuristics, uh, part of it. Um, a slightly different approach is the, the one that's based on logic, right? So you have rules and facts. What other facts can you deduce from the, from, from the existing facts and the rules, which would be logical formula and things like this. That was, you know, pretty dominant in the nineteen eighties. Um, and that led to a, um, an area of, uh, of AI called expert systems or world-based systems, um, that to some extent is very connected with this idea of search. Okay, and then in parallel to this, there is the bottom-up, uh, approach. You know, let's try to reproduce the, uh, to some extent, get inspiration from the, the basic mechanisms of intelligence in biology, implement, um, allow machines to learn and basically organize themselves, um, with the idea that-

    9. NKNikhil Kamath20:10

       How would you do that?

    10. YLYann LeCun20:11

        So, so it's based on the same idea that, uh, a neuroscientist

13. 20:15 – 26:36

    Is A.I. inspired from biology?

    1. YLYann LeCun20:15

       figured out what's going on in the brain, which is that the learning mechanism in the brain proceeds by modification of the strength of the connections between neurons, right? Um, and, and, and people had imagined that, you know, this type of learning could, uh, actually be reproduced in machines. So first, it was the, uh, idea that you could, you know, that neurons were simple, simple computational elements. Um, um, and there were proposals around those lines in the nineteen forties by mathematicians like McCulloch and Pitts, and, uh, people like that. And then in the fifties, um, and early sixties, people proposed a very simple algorithm to change the strength of the connections between neurons so that they could learn a task. Um, so the first machine of this type was called a perceptron, and it was proposed in nineteen fifty-seven. It's a very simple thing, and it's very simple to understand. Um, let's say you want to train a system to recognize simple shapes, um, uh, images. Okay, what is an image for a computer or for an artificial system? It's a, it's an array of numbers. It, um... we know that today because we're familiar with digital cameras and pixels, right? So, um, let, let's take a, a black and white camera. A pixel, um, is, if the pixel is black, it's a zero; if it's white, it's a one. Okay? So it can take only two values, uh, black or white. Um, if you want to build this with nineteen fifties technology, you would put an, an array of photo sensors, photo cells, right, [chuckles] with, with a lens in front of them, and you would show an image, very low resolution, maybe twenty by twenty pixels or something like this, or even lower. Um, so now that gives you an array of numbers that you can feed to a computer. But what they did in fif- nineteen fifties, computers were incredibly expensive, so they actually built electronic circuits. So the, the pixels were voltages, um, coming out of the photo sensors. Um, and then you want to train a system to recognize simple shapes, let's say distinguish the shape of a C from the shape of a D, uh, drawn on this, uh, on this array. Um, so you show an example of a C, and then you let the system produce an output. This output will also be a voltage, and the way the output is gonna be computed is a weighted sum of the, of the values that come in, of the pixels that are one or zero. The weights are connections to a simulated neuron, which is just an electronic circuit that computes, you know, if it's a, if it's a one or a zero, I'm gonna multiply this one or the zero by a, a weight, which is like a resistor that you can change the value of, okay? And then, um, all of the pixels with their weight are gonna be summed up. If the weighted sum is larger than the threshold, it's a C. If it's lower than that threshold, it's a D. All right?

    2. NKNikhil Kamath23:14

       What era was this? Which year did you say?

    3. YLYann LeCun23:16

       Nineteen fifty-seven.

    4. NKNikhil Kamath23:18

       Mm.

    5. YLYann LeCun23:19

       Um, so now how do you train this? So training consists in changing the value of those weights. You can have positive or negative weights.

    6. NKNikhil Kamath23:25

       Mm.

    7. YLYann LeCun23:25

       Um, and what you do is, you show a C, and the s- system computes the weighted sum. So for a C, you want the weighted sum to be large, larger than zero, let's say. Okay, and let's say it's smaller than zero, so the system made a mistake. So you tell it, "No, it should be larger." Okay, [chuckles] you press a button, basically, and you tell it: I really want the output to be s- to be bigger. So what the system does is that it changes all the weights that get a one, so it increases them a little bit. If you increase all the weights that get a one, the weighted sum increases, right? And if you keep re-- if you keep doing this, changing the weights just a little bit every time, eventually the weighted sum is gonna go above zero, and then the system will recognize this, uh, as a C.

    8. NKNikhil Kamath24:12

       And what did we use this for back in the fifties and sixties?

    9. YLYann LeCun24:15

       So nothing really very practical-

    10. NKNikhil Kamath24:17

        Mm-hmm.

    11. YLYann LeCun24:17

        -uh, other than-... recognizing simple shapes, okay?

    12. NKNikhil Kamath24:21

        Mm-hmm.

    13. YLYann LeCun24:22

        Um, so you, you, you repeat showing a C and a D, and for, for the C, you say increase the weighted sum, for the D, you say decrease the weighted sum. So decrease the weights that have a one, increase the weights that have a zero, and then eventually the system settles on the configuration of weights, so that when you show a C, it's above the threshold, when you show a D, it's, it's be- it's below the threshold, so it can distinguish the two. And what it's going to do is, you know, give a positive weight to the pixels that only appear for the C, and a negative weight to the pixels that only a- appear for the D, and that will sort of discriminate between those two. So-

    14. NKNikhil Kamath24:55

        So we had, we had heuristics AI, expert AI, trying to mimic biology, all of this in the '50s and '60s?

    15. YLYann LeCun25:01

        In the '50s. Yeah, starting in the '50s. And then, you know, two different branches basically competing with each other.

    16. NKNikhil Kamath25:08

        Mm-hmm.

    17. YLYann LeCun25:09

        And, and they tried to kind of, um... So one person, a prominent figure in, um, in AI, uh, in the pioneering days, is, uh, Marvin Minsky.

    18. NKNikhil Kamath25:20

        Mm-hmm.

    19. YLYann LeCun25:20

        He was a professor at MIT.

    20. NKNikhil Kamath25:22

        He has a, Marvin... There is a, I remember reading about this. There's a Marvin clause or debate or something like that, right?

    21. YLYann LeCun25:30

        Um, yeah, [chuckles] well, he was, uh... He had pretty strong opinions about things, so there was a lot of discussions. [chuckles] Um, and he, he's interesting because he started his PhD in the '50s, trying to build neural nets, and then completely changed his mind-

    22. NKNikhil Kamath25:44

        Mm-hmm

    23. YLYann LeCun25:45

        ... and, and, and became, uh, basically a big advocate for the other approach, the, the more logic-based and search approach. And in the late '60s or mid '60s, he wrote a book, co-wrote a book with, uh, Seymour Papert, who was a mathematician at MIT, um, which title was Perceptron. And the whole book was to do some theory about Perceptron and to show that the cap- the capabilities of the Perceptron was limited. So the people who were working on neural net at the time kept working on neural net, but they changed the name of what they were doing. They called it, uh, statistical pattern recognition, which sounds much more serious, or adaptive filter theory, which also sounds very serious. And those had enormous applications in the real world.

    24. NKNikhil Kamath26:29

        Yeah.

    25. YLYann LeCun26:29

        Adaptive-

    26. NKNikhil Kamath26:29

        In my world, it's always been, uh... I work in finance, and hedge funds

14. 26:36 – 28:36

    Is building authentic models for finance possible through AI?

    1. NKNikhil Kamath26:36

       and fund managers have always been attempting to pump a lot of data into a neural network to recognize patterns.

    2. YLYann LeCun26:44

       Right.

    3. NKNikhil Kamath26:44

       Is it the same thing that we're talking about-

    4. YLYann LeCun26:46

       Yeah

    5. NKNikhil Kamath26:46

       ... an evolution from the '50s?

    6. YLYann LeCun26:48

       Yeah, absolutely. Um, I mean, the process I described of changing coefficients, you know, up or down to get the output you, you want, uh, you could think of this as a iterative process, very similar to linear regression-

    7. NKNikhil Kamath27:02

       Mm-hmm

    8. YLYann LeCun27:02

       ... which, if you work in finance, you probably know about. [chuckles] Uh, so it, it, it's-

    9. NKNikhil Kamath27:06

       But what I've realized, Yann, is it, even today-

    10. YLYann LeCun27:10

        Uh-huh

    11. NKNikhil Kamath27:10

        ... it's very easy to tweak data that you have collected retrospectively to make something appear like it makes sense, but financial activity tends to be so random, that I don't know if you can build a model based on that.

    12. YLYann LeCun27:26

        Right. So the, um, well, that, that addresses a, a, a bigger issue. When you, when you train a system this way, right? So the, the, the generic principle, which is called supervised learning, is, uh, you give an input to the system, it produces an output. If the output is not the one you want, uh, you-

    13. NKNikhil Kamath27:44

        Mm-hmm

    14. YLYann LeCun27:44

        ... adjust the coefficient so that the output gets closer to the one you want, okay? And there are efficient ways to figure out how to tweak the parameters so that the output gets closer to the one you want. And if you keep doing this on hundreds, thousands, millions, billions of examples, eventually the system, if it's powerful enough, will figure it out. Uh, now, the problem with the perceptron is that the type of functions, input/output functions, that was accessible to perceptron was very limited.

    15. NKNikhil Kamath28:09

        Mm-hmm.

    16. YLYann LeCun28:10

        So there was no way you could take a natural image, uh, you know, a photo, uh, uh, and, and train the system to tell you whether there is a, a dog or a cat or a table in it.

    17. NKNikhil Kamath28:20

        Right.

    18. YLYann LeCun28:20

        That was just not possible. The system was not able to, um, was not powerful enough to really compute this kind of complex function. Uh, this is what neural nets and deep learning changed in the 1980s. And the what-

    19. NKNikhil Kamath28:34

        Just before you get-

    20. YLYann LeCun28:34

        Yes

    21. NKNikhil Kamath28:34

        ... into neural nets,

15. 28:36 – 30:18

    Different parts of A.I | GOFAI, Machine learning

    1. NKNikhil Kamath28:37

       if I'm trying to paint the entirety of the picture, would you say there is intelligence on top, artificial intelligence, and below that is machine learning, and neural nets are a part of machine learning?

    2. YLYann LeCun28:48

       Yeah, so in terms of fields and subfields-

    3. NKNikhil Kamath28:50

       Yeah

    4. YLYann LeCun28:50

       ... AI is more of a problem than a solution.

    5. NKNikhil Kamath28:52

       Yeah.

    6. YLYann LeCun28:52

       Okay, it's a field of investigation, and then there is different techniques you can use for that, right? So there is something that jokingly is referred to as Good Old Fashioned AI, GOFAI-

    7. NKNikhil Kamath29:03

       Mm-hmm

    8. YLYann LeCun29:03

       ... uh, which is using logic, and search, and Heuristic programming, and things like this, which is, this is what you will find in sort of tent- standard textbooks on, on AI.

    9. NKNikhil Kamath29:11

       Mm-hmm.

    10. YLYann LeCun29:11

        Then there is machine learning. So there, the idea is, you don't completely program a machine to do something, you just, you train it from data. That means you need data. Within this, there is a subcategory called deep learning, and this is what, the reason why we hear so much about AI in the last dozen years, is because of deep learning.

    11. NKNikhil Kamath29:28

        Mm-hmm.

    12. YLYann LeCun29:29

        Uh, and neural nets is really the ancestor of deep learning. Deep learning is a new name for it, if you want.

    13. NKNikhil Kamath29:33

        Mm-hmm.

    14. YLYann LeCun29:34

        Um, um, and then, and then there is, you know, application areas, um, so-

    15. NKNikhil Kamath29:39

        Below that.

    16. YLYann LeCun29:40

        Below that. So... And, and they can use combinations of those techniques, right? So, so big applications are computer vision-

    17. NKNikhil Kamath29:46

        Mm-hmm

    18. YLYann LeCun29:46

        ... interpreting images, uh, speech recognition, natural language understanding-

    19. NKNikhil Kamath29:52

        Mm-hmm

    20. YLYann LeCun29:52

        ... um, and maybe speech synthesis also can be viewed as part of this, although it, it's more connected with signal processing. Uh, and then, you know, various other applications, so in, you know, time series prediction or financial modeling and things like this, you know, could be seen as part of this as well.

    21. NKNikhil Kamath30:08

        So if you had to... So I'm, I'm breaking it down. AI has GOFAI under it, which is traditional in nature, like you explained, then machine learning.

    22. YLYann LeCun30:16

        Yes.

    23. NKNikhil Kamath30:17

        Can you

16. 30:18 – 31:14

    What is GOFAI?

    1. NKNikhil Kamath30:18

       define GOFAI in a simple, uh, to-... one line definition.

    2. YLYann LeCun30:22

       So GOFAI is the, the descendant of, uh, the, the, what I was describing earlier as searching for solutions, right? The, this idea that reas- it's all about reasoning, reasoning is all about search, uh, you know, looking for a solution to a problem, and having a way of characterizing whether you have found a solution.

    3. NKNikhil Kamath30:40

       So you mean the rule-based thing, an input and an output based on what applies, the board rule applies, like that?

    4. YLYann LeCun30:47

       Um, the, the, um... Yeah, I mean, any, any rule-based system, anything that uses logical inference-

    5. NKNikhil Kamath30:53

       Mm-hmm.

    6. YLYann LeCun30:54

       -um, deducing facts from rules and previous facts-

    7. NKNikhil Kamath30:57

       Mm-hmm.

    8. YLYann LeCun30:58

       -uh, searching for a solution, like finding the shortest path in a, you know-

    9. NKNikhil Kamath31:02

       Mm-hmm

    10. YLYann LeCun31:02

        ... in a graph or something. Um, tho- those are good old-fashioned AI, uh-

    11. NKNikhil Kamath31:06

        And under machine learning, what are the different types of ML?

    12. YLYann LeCun31:09

        So, so there is, uh, so-called traditional machine learning. I'm not

17. 31:14 – 32:22

    Different types of Machine Learning

    1. YLYann LeCun31:14

       sure that de- deserves the term, and this is basically derived from, uh, statistical estimation. So things like linear regression-

    2. NKNikhil Kamath31:21

       Mm-hmm

    3. YLYann LeCun31:22

       ... would be part of it. And then there are other methods that are slightly more, uh, sophisticated, uh, uh, boosting, um, uh, classification trees, support vector machines, kernel methods. Um, I mean, there's, there's a bunch of methods of this type, and Bayesian inference, that are part of machine learning in the sense that they, they obey that model of, uh, you know, you, you, you build a program, but the program is really not finished. It's got a bunch of tunable parameters-

    4. NKNikhil Kamath31:50

       Mm-hmm

    5. YLYann LeCun31:50

       ... and the input/output function is determined by the value of those parameters. And so you train the system from data, uh, using this iterative, uh, adjustment techniques that I described before. Show examples, if the answer is incorrect, adjust the parameters so that it, it, it comes closer to the answer you want.

    6. NKNikhil Kamath32:06

       So machine learning is supervised in a way?

    7. YLYann LeCun32:09

       So that's supervised learning.

    8. NKNikhil Kamath32:10

       Mm-hmm.

    9. YLYann LeCun32:10

       Okay? You, you tell the system, "Here is an output, here is the, the desired, desired output."

    10. NKNikhil Kamath32:16

        Mm-hmm.

    11. YLYann LeCun32:17

        Um, but there are other forms of learning. So one, one

18. 32:22 – 33:19

    What is Reinforcement learning?

    1. YLYann LeCun32:22

       different form is, uh, reinforcement learning.

    2. NKNikhil Kamath32:24

       Mm-hmm.

    3. YLYann LeCun32:25

       So in reinforcement learning, you don't tell the system the correct answer, you just tell it whether the answer it produced was good or bad. You give, you give it a single number that tells it your answer was good or was bad.

    4. NKNikhil Kamath32:37

       And what happens next? Say, I'm a reinforcement learning engine, and you tell me an answer was good or bad. What do I do next?

    5. YLYann LeCun32:44

       Well, so if your answer was good, uh, you don't do much.

    6. NKNikhil Kamath32:48

       Mm-hmm.

    7. YLYann LeCun32:49

       If your answer was bad, then, uh, you have to figure out which answer, among all the possible answers that could have produced, which one would be a better one.

    8. NKNikhil Kamath32:58

       Mm-hmm.

    9. YLYann LeCun32:58

       So maybe you try another answer, and you say, "What about this one? Is it better or is it worse?" Uh, if the environment tells you it was better, then you kind of de-emphasize the first one and emphasize that one by tuning the parameters inside of a neural net or something like that, some sort of learning, uh, learning machine.

    10. NKNikhil Kamath33:17

        So what is self-supervised

19. 33:19 – 35:14

    What is Self supervised learning? Up & Coming

    1. NKNikhil Kamath33:19

       learning?

    2. YLYann LeCun33:20

       Okay, so self-supervised learning is what has become very prominent over the last five, six years, and, um, is, is really the, the, the main component or the, the main contribution to the success of things like chatbot and natural, uh, language understanding systems, and-

    3. NKNikhil Kamath33:36

       They don't fall under reinforce- reinforcement learning?

    4. YLYann LeCun33:39

       No, i- it's more similar to supervised learning, but the difference is that instead of having a clear input and output, and training the system to produce the output from the input, um, you basically only have things that can either be input or output. Let me take an example. Um, you take a, a piece of text, and you corrupt that text in some way-

    5. NKNikhil Kamath34:05

       Mm.

    6. YLYann LeCun34:05

       -so by removing some words.

    7. NKNikhil Kamath34:07

       Mm-hmm.

    8. YLYann LeCun34:07

       Right? So now you have a partially masked, uh, text, where some words are missing, and you train a machine to predict the words that are missing.

    9. NKNikhil Kamath34:15

       Mm-hmm.

    10. YLYann LeCun34:16

        So the technique you would use for this is supervised learning, because you tell the system, "Here is the correct word that you should predict at that location."

    11. NKNikhil Kamath34:22

        Mm-hmm.

    12. YLYann LeCun34:23

        Um, and the system can use all the words that it can see to predict the words that it cannot see.

    13. NKNikhil Kamath34:27

        And this is an example for supervised learning?

    14. YLYann LeCun34:29

        Self-supervised learning. It's self-supervised because the, there is no differentiation between input and output. It's really kind of the same thing.

    15. NKNikhil Kamath34:36

        Mm-hmm.

    16. YLYann LeCun34:37

        And if the input is, for example, an image, the, the way you, um, you would train a self-supervised learning system is that you would corrupt or transform the image in some way, and then you would train the system to recover the original image from the corrupted or transformed version of it. Okay, so there is no supervision. You don't need someone to go through a few million images and labeling them, is it a cat, or a dog, or a table, or a chair? Um, it's, it's, it's a task of basically understanding the, the input, uh, the internal structure of the input by being able to filling in the—to fill in the blanks.

    17. NKNikhil Kamath35:12

        Forgive me for asking maybe a really stupid

20. 35:14 – 38:00

    Is AI telling you what you want to hear?

    1. NKNikhil Kamath35:14

       question. I'm trying to picture this. Let's say I have X amount of data. I have ten lines that say, "Cats are black, dogs are white," whatever, ten lines. I remove a part of it, and then I tell the model to fill it in.

    2. YLYann LeCun35:30

       Yep.

    3. NKNikhil Kamath35:31

       Are you saying at that point of time, I also tell the model the answer, saying, "This should be the answer?"

    4. YLYann LeCun35:36

       Yeah. You, you, you tell it, "Here is the answer that I removed." Like, "Can you predict this missing... "

    5. NKNikhil Kamath35:43

       Can you arrive at the answer which I removed, and I'm telling you that this was the answer?

    6. YLYann LeCun35:47

       Right. But you can only use the thing that you can see, so you don't see the answer on the input, you have to predict it. But I'm telling you when, during training, I tell you what it is, and so the system can adjust its parameter to, its parameters in a supervised fashion. So the, the only difference, the difference is not in the algorithms themselves-

    7. NKNikhil Kamath36:06

       Mm-hmm

    8. YLYann LeCun36:06

       ... it's basically supervised learning, but it's in the structure of the system and the way the data is, uh, is, is used and, and, and produced. You, you don't need to basically have, uh, you know, someone going through millions of images and telling you, uh, this is a cat or a dog or a table. Um, you just show an image of a dog, a cat or a table, and you-... corrupt it, partially change it, change the colors maybe or something, um, and then ask the system to recover the original one from the corrupted one. Okay. So that's, that's, uh, one particular form of self-supervised learning, and this is what's been incredibly successful for natural language understanding. ChatGPT. So things like, so chatbots are, uh, or LLMs, large language models- Mm-hmm. -are a special case of that. Mm-hmm. Where you train a system to predict a word, but you only allow it to look at the words, the, the words that precede it. Mm-hmm. Um, you know, that are to the left of it. Mm-hmm. Um, and that requires kind of building the neural net in a particular way so that the connections that, that predict one word only look at the, the words that precede. So then you don't need to corrupt the input. You just show an input, and through the structure of the system, the system can only predict... It can only— It's trained to predict the next word from, from, from the context. And these are all examples of neural networks in a way? These are all, underlying this, are particular way of connecting neural network, neurons with each other, simulated neurons, right? Uh, or, or simple elements that compute a very simple mathematical function, something like a weighted sum, and what's adjustable are the weights. Or in the case of, uh, transformer architectures, which are, are very, uh, popular at the moment, um, uh, they consist in basically c-comparing every input to each other and, and producing weights. Uh, I, I could explain this as sort of more complicated, but-

21. 38:00 – 40:24

    What is a transformer?

    1. YLYann LeCun38:00

       What is a transformer, Yann? So, okay, so there are several architectural components, uh, which, from which you can build a neural net. So let me start with, um, very simple idea. Let's say you want to build a neural net that recognizes images, okay? So again, an image is a, an, an array of numbers indicating the brightness of every pixel, right? Um, you can build a neural network with a single layer. So let's say you want to distinguish, um, ten categories, okay? Cats, dogs, tables, and chairs, and cars, and whatever. Um, or let's say simpler, you want to recognize the ten digits, okay, zero to nine. Uh, someone drawing a digit, it's drawn on a sixteen by sixteen pixel area, so you have two hundred and fifty- six, two hundred and fifty-six inputs, and you have ten outputs. Okay, you can have a single, sing... What's called a single-layer neural net, uh, which basically each output is a weighted sum of, of the pixels, and you try to train those weights in such a way that when you show a zero, the output zero is the most active, and the other ones are less active, and, and so forth for, uh, all the categories. Okay, that may work for simple shapes like, like printed, uh- Mm-hmm ... digits. It won't work for handwriting because there's so much variability in the characters that you cannot reduce the classification to a simple weighted sum. Okay, so the breakthrough that occurred in the nineteen eighties was to, um, stack multiple layers of neurons. So each neuron computes a weighted sum and then passes this weighted sum through essentially a threshold function. So if the weighted sum is below a threshold, the, the neuron stays inactive, the output is zero. And if it's above a threshold, it's active. Okay, there's various ways to do this. Um, but it's nonlinear, and that's very important. So you stack two layers, where, um, the, the middle layer you could think of as detecting sort of basic motifs on the inputs, and then the second layer sort of integrate those motifs to figure out, "Okay, this is a, a C because it's got two endpoints." You know, the, the, uh, the shape of the C kind of stops there, and I can detect that, so if there's two of them, that's a C. And a D doesn't, but a C has two corners. Maybe I can detect that. The systems potentially learns to do this from end to end, and, and the way it works is through an algorithm called backpropagation. Um, and what

22. 40:24 – 42:58

    What is a back propagation algorithm?

    1. YLYann LeCun40:24

       this backpropagation algorithm does is that when you show an image of a C, and you tell the system, "This is a C, so activate this output neuron, does not, and do not activate the other ones," it knows how to adjust the parameters so that the output gets closer to the one you want. Um, and that's done by propagating signals backwards to, um, to basically figure out the sensitivity of each output to, um, to, to each weight, so that you can change the weights in such a way that the good output increases and the bad outputs decrease, right? Um, so that's backpropagation. That, um, um, algorithms to do this, so the backpropagation algorithm popped up in the nineteen eighties. Um, conceptually, it existed before, but people didn't realize they could use it for machine learning. And there was a wave of interest in neural nets starting in the mid-nin, mid, mid-eighties, uh, lasting, uh, ten, fifteen years, um, to kind of explore this idea of multilayer networks. And this, this was crucial because it, it lifted some of the limitations that Minsky and Papert in the sixties said were, y- you know, the perceptron was the, was subjected to. Um, so a big wave of interest. But then people realized that to train those neural nets, um, you need a lot of data, and this was before the internet. Mm-hmm. There was not much data. You need fast computers, and computers were not that fast. [chuckles] Mm-hmm. Um, so, so people kind of lost interest a little bit in this. But one thing that I worked on in the late eighties, early nineties, is, um, if you want a system of this type to recognize images, you kinda have to connect the neurons to each other in a particular way that facilitates the system sort of paying attention, you know, being able to detect motifs, for example, right? Uh, local motifs. So, um, uh, I got inspiration from biology again, um, a classical work in neuroscience that, that went back to the nineteen sixties, to basically organize the way the neurons are connected to each other into layers, uh, so that they, they bias towards kind of finding good solutions for image recognition. So that's called a convolutional neural network or ConvNet.

    2. NKNikhil Kamath42:39

       ... um, so just, just to come back to this, where you are. Like, so you broke down machine learning. I'm sorry I keep going back-

    3. YLYann LeCun42:46

       Sure.

    4. NKNikhil Kamath42:46

       -or I'll get confused.

    5. YLYann LeCun42:47

       Yes.

    6. NKNikhil Kamath42:47

       So under machine learning, the really popular pathway right now, let's say, self-supervised, which has ChatGPT and a bunch of other things.

23. 42:58 – 48:06

    What’s happening in the reinforcement learning space?

    1. NKNikhil Kamath42:58

       What's happening in the reinforcement learning space?

    2. YLYann LeCun43:01

       So not so much anymore. Um, there was a big wave of interest in reinforcement learning, um, about, you know, a dozen years ago, and companies like DeepMind-

    3. NKNikhil Kamath43:12

       Mm-hmm

    4. YLYann LeCun43:13

       -set themselves up with the idea that reinforcement learning was going to be the, the key element towards building truly intelligent machines.

    5. NKNikhil Kamath43:23

       Mm-hmm.

    6. YLYann LeCun43:23

       And that's-

    7. NKNikhil Kamath43:24

       Can you again, like, define reinforcement learning once more in a line?

    8. YLYann LeCun43:27

       So reinforcement learning is a situation where you don't tell the system what the correct answer is, you just tell it whether the answer it produced was good or bad.

    9. NKNikhil Kamath43:34

       Right. Okay.

    10. YLYann LeCun43:35

        Okay, so there are many possible answers. It's very inefficient-

    11. NKNikhil Kamath43:38

        Mm-hmm

    12. YLYann LeCun43:38

        -because the system has to try many things before it gets the correct answer.

    13. NKNikhil Kamath43:41

        Mm-hmm.

    14. YLYann LeCun43:42

        Um, and so it's very inefficient. It requires many, many, many trials, and so it works really well for games. You know, you, you... It's very efficient if you want to train a system to play chess or Go or-

    15. NKNikhil Kamath43:53

        Mm

    16. YLYann LeCun43:53

        -or things like that, poker. Reinforcement learning is great because you can have the system play millions of games against itself-

    17. NKNikhil Kamath44:00

        Mm-hmm

    18. YLYann LeCun44:00

        -or copies of itself. Um, and it can adjust it, you know, it, it wins or loses a game, so it can, you know, it knows which policy, which flavor of the neural net won the game and sort of reinforces that and de-emphasizes the one that lost. And so the system basically can train itself, right?

    19. NKNikhil Kamath44:19

        And what did you say a transformer was?

    20. YLYann LeCun44:21

        Okay, so I was go- coming to this-

    21. NKNikhil Kamath44:23

        Mm-hmm

    22. YLYann LeCun44:23

        -you know, through convolutional net, right. So there is this, uh, um, particular way of connecting similarity neurons with each other to bias it towards doing a good job for certain types of data. And, uh, convolutional nets are really good for, uh, data that comes from the natural world, whether it's an image or an audio signal, um, which are things that are, um, where, where nearby values in the array of numbers that come to you in an image or audio signal, nearby values are generally very similar to each other. So if you take a picture, any picture, natural image, and you take two neighboring pixels, they're very likely to have the same, the same color or the same intensity. Now, what I'm talking about here is the fact that the, you know, natural data, like images and audio and, and just about any natural signal, has some natural underlying structure to it. Um, and if you build a neural net in a particular way that, that can take advantage of this structure, it, it will learn faster, it will learn with fewer samples. So we started doing, doing experiments with this in the late '80s and, and, and build those convolutional nets. They are inspired by the architecture of the, of the visual cortex, really. Um, and there is some mathematical-

    23. NKNikhil Kamath45:38

        Mm

    24. YLYann LeCun45:38

        -justification for it.

    25. NKNikhil Kamath45:39

        Mm-hmm.

    26. YLYann LeCun45:40

        Uh, but the, the basic idea is that, um, each neuron in a convolutional net only looks at a small area of the image, and, and you have multiple neurons looking at multiple areas of the image, and they all do the same thing. They all have the same weights. Um, it's a basic concept which connects with a mathematical concept called convolutions, and so that's why those things are called convolutional nets. Okay, so that's what's called an architectural component or a module. A convolution is something that has an interesting property, which is that if you show it an input, it's gonna produce a particular output. If you shift the input, the output will be shifted, but otherwise unchanged. And that's a very interesting property for audio signals, images, and various other natural signals. Okay, now, a transformer-

    27. NKNikhil Kamath46:23

        Mm-hmm

    28. YLYann LeCun46:24

        -is a different way of arranging the, the, the neurons, uh, if you will, in such a way that, um, the, the inputs are a number of different items. We call them tokens. They're really vectors, which means list of numbers, okay? And the property of a, uh, a, the layer or the block of a transformer is that if you permute the inputs, the output will be permuted similarly, but otherwise unchanged. Um-

    29. NKNikhil Kamath46:54

        When you say otherwise unchanged, you mean?

    30. YLYann LeCun46:56

        You mean, I... What I mean is that, um, if you, you give a bunch of tokens, you run through the transformer, you will get a bunch of output tokens, okay? The same number generally as the number of input tokens. They'll be different vectors. Um, if you now take the first half and the second half of, of your sequence of input tokens and you flip them, what you will get is the same result that you got previously, but it will be flipped exactly the same way. Okay, so the input/output function is, uh, technically, we call this equivariant to permutation.
24. 48:06 – 49:08

    What is a convolutional neural network ?

    1. NKNikhil Kamath48:06

       Yann? I'm sorry. I'm gonna ask you, like, to simplify every single term.

    2. YLYann LeCun48:10

       Oh, absolutely. So a convolution is this component for a convolutional neural net. So, uh, the idea of it is that you have a neuron that looks at, uh, at a part of the input, and then you have another neuron that looks at another part of the input, but it computes the same function as the first neuron. And then you replicate that same neuron for every location on the input, so that, um, you can think of each of those neurons as detecting a particular motif on a part of the input, and all the neurons detecting the same motif at different parts of the input. So that now, if you take an input and, and you shift it-... the, you're gonna get the same output shifted because, you know, you're gonna have the same neurons looking, detecting the same motif, just at different locations. So that, that's what gives you this shift equivariant. Um, that's a convolution. Mathematically, there's something called a convolution that-

    3. NKNikhil Kamath49:01

       Mm.

    4. YLYann LeCun49:01

       -mathematicians invented a long time ago, and that's basically what this, what this does.

    5. NKNikhil Kamath49:06

       When you say neuron

25. 49:08 – 50:00

    What is a Neuron - the Machine Learning perspective

    1. NKNikhil Kamath49:08

       in all of this, can you explain the basis of just that term? What is it?

    2. YLYann LeCun49:13

       So, uh, we, we, we use that term, it's an abuse of language-

    3. NKNikhil Kamath49:17

       Mm.

    4. YLYann LeCun49:17

       -because those neurons are not really neurons like in the brain.

    5. NKNikhil Kamath49:21

       Mm, mm.

    6. YLYann LeCun49:21

       They're, they are to real neurons as an airplane wing is to a bird wing, [chuckles] okay?

    7. NKNikhil Kamath49:26

       Mm.

    8. YLYann LeCun49:27

       So it performs the same-- it has the same concept. And what a neuron, uh, does in a neural net is computing a weighted sum of its inputs, and then comparing that weighted sum to a threshold, activating the output if it's above the threshold, and, and producing zero if it's below the threshold. That's the, the basic neuron. Now, there are variations of this, and in a transformer, it's a slightly different type of mathematics. You're kind of comparing vectors to each other and-

    9. NKNikhil Kamath49:52

       Mm-hmm

    10. YLYann LeCun49:52

        ... things like that. But, uh, but that's kind of the basic functionality of a neuron. It's a combination of a linear operation, where you have coefficients

26. 50:00 – 58:00

    What is a neural network language model & how does it work?

    1. YLYann LeCun50:00

       that you can change the value of through training, and then a nonlinear, uh, uh, function, a threshold, or something like that, that, uh, uh, you know, detects something or not.

    2. NKNikhil Kamath50:11

       Right. We looked online, and while we were researching, we could not find a good definition for neural network language model and how it works in simple terms.

    3. YLYann LeCun50:22

       Okay. Um, so the idea of a language model goes back to the nineteen forties.

    4. NKNikhil Kamath50:28

       Mm-hmm.

    5. YLYann LeCun50:28

       A gentleman called Claude Shannon, he's a very famous, uh, mathematician who used to work at Bell Labs, where I used to work, um, although he wasn't there anymore when I joined. Uh, and he, he came up with a theory called information theory, and then was fascinated by the idea that you could discover the structure in data, right? So i- uh, he invented something where you take a text, um, and, and you say, "I'm give you a se-- I'm giving you a sequence of letters, and I'm asking you, what is the next letter that comes afterwards?"

    6. NKNikhil Kamath50:59

       Mm.

    7. YLYann LeCun50:59

       Okay? So let's take a, you know, a, a English word, um, or whatever in, uh, in a sort of, uh, let's say, a wo..., you know, Roman language. If you have a series of letters and the last one is a Q, it's very likely that the next letter is a U. You almost never have a Q without a U behind it, unless it's an Arabic word or something that's been transliterated.

    8. NKNikhil Kamath51:24

       Right.

    9. YLYann LeCun51:25

       Um, so for every, every letter that you observe, you can, you can build a, a table of the probability that the next letter will be an A, a B, a C, a Q, or U.

    10. NKNikhil Kamath51:37

        This is where the word generative comes from?

    11. YLYann LeCun51:40

        Yeah, so it's generative because if you have this table of conditional, what we call conditional probabilities, right? Given the previous letter, what is the next letter-- what is the probability of the next letter? You can use this to generate text.

    12. NKNikhil Kamath51:50

        Mm-hmm.

    13. YLYann LeCun51:50

        You, you start with a letter, let's say Q, okay? And then you look through the table of probability. What's the next letter that is most likely?

    14. NKNikhil Kamath51:59

        Mm-hmm.

    15. YLYann LeCun51:59

        You just pick that one. That's gonna be U.

    16. NKNikhil Kamath52:00

        Mm-hmm.

    17. YLYann LeCun52:01

        Um, or you don't pick that one, you, you, you pick the next letter with the probability that is... You know, you flip a coin, or you generate a random number in the computer, and then you, you produce the, uh, you know, the following letter according to the probabilities that you measured on real text. Um, and you keep doing this, and the system is gonna just generate letters. Um, it's not gonna look like words.

    18. NKNikhil Kamath52:24

        Mm-hmm.

    19. YLYann LeCun52:24

        Uh, it's probably not even gonna be pronounceable.

    20. NKNikhil Kamath52:27

        Right.

    21. YLYann LeCun52:27

        Um, but that- if instead of a context of one letter, you take a context of two letters, then it becomes kind of more readable.

    22. NKNikhil Kamath52:35

        Mm-hmm.

    23. YLYann LeCun52:36

        It's still not words, right? If you take a context of three letters, then it, it becomes, you know, even nicer. And as you increase the size of the context, that determines the, the probability of the next letter, it becomes more and more readable. But you have an issue there, which is that, um, uh, the, the size of the table you need, if you have—if you look at the first letter, and, and, and, and you have to figure out what's the probability for the next letter, you need a table of twenty-six rows and twenty-six columns. For each first letter, what is the probability for every possible second letter, right? So it's table twenty-six by twenty-six. Now, if the context has two letters, now the number of rows in your table is twenty-six squared, because you have twenty-six squared sequences of two, possible sequences of two letters, right? And for each of those, you need twenty-six probabilities. So it's twenty-six cube, the size of your table. As you add characters, um, the table increases to twenty-six to the power n, when n is the length of the, of, of the, uh, of the sequence. So that's called an n-gram model, and that's a language model. You can do this at the level of characters. It's more difficult to do this at the level of a word, because you might have a hundred po-- thousand possible words, right?

    24. NKNikhil Kamath53:49

        Mm-hmm.

    25. YLYann LeCun53:49

        So now your table is gigantic.

    26. NKNikhil Kamath53:51

        Mm-hmm.

    27. YLYann LeCun53:51

        So you can train, uh, a, a word model, uh, or, or a language model by just filling up this table of probabilities by training on a large corpus of text. Uh, but it becomes impractical, uh, above a certain length of context, number of-

    28. NKNikhil Kamath54:08

        Because of the amount of wo- compute and work required.

    29. YLYann LeCun54:11

        It, it's, it's also, it's the memory of storing-

    30. NKNikhil Kamath54:13

        Mm
27. 58:00 – 59:55

    The AI tree | LLMs

    1. NKNikhil Kamath58:00

       today and what everybody's so excited about. Machine learning has different things, different neural networks under it. There's a reinforcement, one like DeepMind, there is a self-supervised, generative ChatGPT, because using it as a placeholder, as it's the most popular one right now. And-

    2. YLYann LeCun58:21

       LLM.

    3. NKNikhil Kamath58:21

       Huh?

    4. YLYann LeCun58:22

       LLM. Autoregressive LLM, really, that's what it should be called. [chuckles]

    5. NKNikhil Kamath58:25

       Autoregressive LLM.

    6. YLYann LeCun58:27

       Yeah. I mean, the, the, the proper organization is, yeah, there is AI at the top.

    7. NKNikhil Kamath58:30

       Mm-hmm.

    8. YLYann LeCun58:31

       Uh, machine learning is a particular way of approaching the AI problem.

    9. NKNikhil Kamath58:37

       Mm-hmm.

    10. YLYann LeCun58:37

        Under this is deep learning, which is really the, the, the foundation of pretty much-

    11. NKNikhil Kamath58:43

        Mm-hmm

    12. YLYann LeCun58:43

        ... all of AI today.

    13. NKNikhil Kamath58:44

        Mm-hmm.

    14. YLYann LeCun58:44

        Um, so basically, neural networks with multiple layers, right? So the idea of this goes back to the 1980s and back propagation. That's still the, the, the basic foundation of everything we do. Under this, there is several families of architectures, convolutional nets, transformers, combinations thereof. Um, then there is, under transformers, there is, uh, several, um, flavors of it, some of which can be applied to image recognition or audio, some of which can be applied to representing natural language but not generating it.

    15. NKNikhil Kamath59:15

        Mm-hmm.

    16. YLYann LeCun59:15

        And then there's a subcategory, large language models, which are autoregressive transformers. So transformers have a particular architecture that, uh, allow them to predict the next word, and then you can u- you can use it to just generate word, because, you know, given a sequence of word, it's been trained to produce the next word. So given a text, you have it produce the next word, and then you shift the input by one. So now, the word it generated is part of its input, and you can ask it to generate the second word, shift that, third word, shift that, fourth word. That's autoregressive prediction. It's the same concept as autoregressive models in finance and econometrics-

    17. NKNikhil Kamath59:54

        Yeah

    18. YLYann LeCun59:54

        ... and stuff like that. Same stuff.

28. 59:55 – 1:01:40

    The next challenge of AI

    1. NKNikhil Kamath59:55

       And these work best for text, but not for pictures, videos, or any of that?

    2. YLYann LeCun1:00:02

       That's right. And the reason it works for text and not for other things, is because text is discrete, so there is a finite number of possible things that can happen, right? There's a finite number of words in the dictionary.

    3. NKNikhil Kamath1:00:13

       Mm-hmm.

    4. YLYann LeCun1:00:14

       There's a, you know, um, so if you can discretize your signal, then you can use those autoregressive prediction systems. And the, the, you know, the main, the main issue is that, um, you're never gonna be able to make an exact prediction.

    5. NKNikhil Kamath1:00:30

       Mm-hmm.

    6. YLYann LeCun1:00:31

       And so the system is gonna have to learn, um, some sort of probability distribution, or at least, you know, produce scores that are different-

    7. NKNikhil Kamath1:00:40

       Mm

    8. YLYann LeCun1:00:40

       ... for, for, uh, for, for different potential outputs. So you can output a list of probabilities if you have a finite number of possibilities, which is the case for language. Um, but if you want to predict what is gonna happen in a video, the number of possible frames, video frames, is essentially infinite.... Right? You, you have, you know, let's say, [coughs] a million pixels, right? An image, thousand by thousand pixels. The pixels are in color, so you have three values. So that's three million values, um, that you have to produce, and, and we don't know how to represent a probability distribution over the set of all possible images with three million pixels.

    9. NKNikhil Kamath1:01:24

       But this is what everybody's very excited about.

    10. YLYann LeCun1:01:27

        This is what a lot of us consider the next challenge in AI. So basically, you have systems that can learn how the world works, uh, by watching videos. We don't-

    11. NKNikhil Kamath1:01:38

        And if you were to say videos, learn from videos

29. 1:01:40 – 1:03:20

    Pictures/ Videos - what's happening there?

    1. NKNikhil Kamath1:01:40

       and pictures, which will be the next phase, where does that fall in this entire equation? Does it come under where LLM sit today?

    2. YLYann LeCun1:01:49

       No, it's completely different from LLM, which is why I've been, uh, pretty vocal about the fact that LLMs are not the path to human-level intelligence. Um, LLMs work for discrete worlds. They don't work for continuous high-dimensional worlds, which is the, the case for video. And this is why LLMs do not understand the physical world, um, and, and cannot be used in their current form to really understand the physical world. And so we have-- I mean, LLMs are amazing in their ability to manipulate language-

    3. NKNikhil Kamath1:02:20

       Mm-hmm.

    4. YLYann LeCun1:02:21

       ... but they can make very, very stupid mistakes that reveal they really don't understand how the world works, right? The underlying world. And, um, this is why we have systems that can pass the bar exam or, or write an essay for you. But we don't have domestic robots, we don't have self-driving cars or completely autonomous Level Five self-driving cars. We, we, you know, we don't have systems that really understand b- very basic things that your cat can understand. So I've been-

    5. NKNikhil Kamath1:02:48

       Yeah

    6. YLYann LeCun1:02:48

       ... you know, kind of vocal saying that-

    7. NKNikhil Kamath1:02:50

       Yeah

    8. YLYann LeCun1:02:50

       ... you know, the smartest LLMs are not as smart as your house cat.

    9. NKNikhil Kamath1:02:52

       Yeah.

    10. YLYann LeCun1:02:53

        And it's really true. Uh, so, so the, the challenge for the next few years is to build AI systems that lift the limitation of—limitations of LLM. So systems that understand the physical world, uh, have persistent memory, which LLMs really don't have at the moment, uh-

    11. NKNikhil Kamath1:03:11

        Persistent memory.

    12. YLYann LeCun1:03:12

        Persistent memory, which means, you know, they can remember things, right? Store facts in a, in a memory, and then retrieve them when it's, uh, interesting.

30. 1:03:20 – 1:04:45

    LLM’s limited memory | Types of memory

    1. YLYann LeCun1:03:20

       Um-

    2. NKNikhil Kamath1:03:21

       Can't LLMs remember stuff now? Memories.

    3. YLYann LeCun1:03:23

       The only memory that an LLM... uh, the only two—there's two types of memory that an LLM has. The first type is in the parameters, in the coefficients that are adjusted during training, right? So they will learn something. They, they don't, they... It's not really kind of storing, uh, a piece of information. If you train a LLM on a bunch of novels, it cannot regurgitate the novels. But it will remember something about the statistics of the words in that novel, and it, it might be able to answer questions, you know, general questions about, about the story and things like this, but it's not gonna be able to regurgitate all the words, right? Um-

    4. NKNikhil Kamath1:04:00

       Just-

    5. YLYann LeCun1:04:00

       ... kind of like humans, right? You read a novel, you can't, you can't remember all the words-

    6. NKNikhil Kamath1:04:03

       Mm-hmm

    7. YLYann LeCun1:04:03

       ... you, you, unless you spend a lot of efforts, uh, trying to do this. Uh, so that's the first type of memory, and then the second memory is the context, the, the prompt that you type.

    8. NKNikhil Kamath1:04:14

       Mm-hmm.

    9. YLYann LeCun1:04:15

       And since the system can generate word, and, and, and those words are, or those tokens are injected in its input, it can use this as some sort of working memory, but it's a very limited, uh, form of, of memory. What you want is a memory that would be more similar to what we have in our brains, what mammalians have, called the hippocampus. Um, the hippocampus is a kind of a brain structure in the center of the brain, inhibits all the cortex, and if you don't have a hippocampus, you can't remember things for more than about ninety seconds.

31. 1:04:45 – 1:10:26

    AI’s path to human like learning

    1. NKNikhil Kamath1:04:45

       And if you were to draw a path from intelligence that we described on top, all the way down to self-supervised learning, how do you suspect that path will look towards us getting to the point where we are learning from videos and images and more human-like intelligence?

    2. YLYann LeCun1:05:01

       So the, the path that, um, I've been trying to, to plot, um, is, uh, discovering new architectures different from those autoregressive architecture used for LLM-

    3. NKNikhil Kamath1:05:14

       Mm-hmm

    4. YLYann LeCun1:05:14

       ... that would be applicable to, to video, so that self-supervised learning could be used to train those systems. And this type of self-supervised learning basically would be: here is a piece of a video, predict what comes next. Um, and, and if a system can do a good job at predicting what's gonna happen next in a video, that means it probably has understood a lot about the underlying structure of the world. Similarly to a large language model learns a lot about-

    5. NKNikhil Kamath1:05:39

       Mm-hmm

    6. YLYann LeCun1:05:39

       ... you know, language by just being trained to predict the next word, right? So-

    7. NKNikhil Kamath1:05:42

       Not like I will understand, but if you had to give us a line on how that architecture might look.

    8. YLYann LeCun1:05:47

       Okay, so here is the issue.

    9. NKNikhil Kamath1:05:48

       Mm.

    10. YLYann LeCun1:05:48

        Because a- as I told you, um, those autoregressive architecture work for text because text is discrete.

    11. NKNikhil Kamath1:05:55

        Mm-hmm.

    12. YLYann LeCun1:05:55

        And you can never predict, uh, what, what comes next. We can produce a probability, probability distribution over what comes next. You cannot do this for images and video because it's just too complicated mathematically, and you can show that it's intractable and blah, blah, blah. So predicting all the pixels in a video that follow, uh, a pa-- a particular video segment, basically is not possible.

    13. NKNikhil Kamath1:06:18

        Mm-hmm.

    14. YLYann LeCun1:06:19

        Or not possible to a degree that would be useful for the problem that we're interested in.

    15. NKNikhil Kamath1:06:23

        Mm-hmm.

    16. YLYann LeCun1:06:23

        You know, what we want is a system that has the ability that, uh, the, the ability to predict what's gonna happen in the world, uh, because that's a good way to, for a system to be able to plan. If I can plan that, if I approach my hand, uh, you know, uh, to this glass, um, and I, I close my hand, and I lift it up, you know, I'm gonna grab the glass, and I can drink. Um, I can plan a sequence of actions to arrive at a particular result, right?

    17. NKNikhil Kamath1:06:50

        Mm-hmm.

    18. YLYann LeCun1:06:50

        So I have a good model of the world that says: The state of the world at time T is this, the glass is on the table. The action I'm gonna take is close my hand around it, okay, um, and lift. What is gonna be the state of the world at time T plus-... three seconds after I close my hand and lifted my arm. And the state of the world is gonna be, I'm gonna have that glass in my hand. Um, so if you have this kind of world model-

    19. NKNikhil Kamath1:07:16

        Mm-hmm.

    20. YLYann LeCun1:07:17

        um, state of the world, action, next state of the world, then you can imagine, you can, you can predict, you, you can predict the outcome of a sequence of actions. You can imagine taking a sequence of actions, and then predict in your mind what the outcome will be. You can predict if this outcome is something that satisfies a goal that you want to accomplish, like drink a, p- a little bit of water, uh, take a sip. And what you can do is, through search, so now we are connecting with old AI, search a sequence of actions that will actually satisfy this goal. Um, so this is the type of, um, reasoning and planning that psychologists, uh, call system two.

    21. NKNikhil Kamath1:07:58

        Mm-hmm.

    22. YLYann LeCun1:07:58

        Okay? Daniel Kahneman is a, a late, um, Nobel Prize-winning, uh, psychologist. Um, and he, he, he makes this distinction between system one and system two, where system one is actions you can take without thinking, sub- subconscious. It's just reaction- reactive, and the system two is what you have to deliberately plan, uh, and think about to be able to, uh, produce an action or a sequence of actions.

    23. NKNikhil Kamath1:08:23

        So Yann, will memor- memory eventually be the answer? Because as humans from biology, we learn through memory, right?

    24. YLYann LeCun1:08:31

        Well, uh, uh, it depends what type of memory. I mean, we also have multiple types of memory. We have the hippocampus that I-

    25. NKNikhil Kamath1:08:36

        Mm-hmm

    26. YLYann LeCun1:08:36

        ... I mentioned. So hippocampus i- is used to store long-term memories, like, um, you know, things that happened to you when you were a child-

    27. NKNikhil Kamath1:08:45

        Mm-hmm

    28. YLYann LeCun1:08:45

        ... and things like that.

    29. NKNikhil Kamath1:08:46

        Mm-hmm.

    30. YLYann LeCun1:08:47

        Uh, basic facts about the world, like, you know, when your mom was born or something.
32. 1:10:26 – 1:11:58

    What is JEPA ?

    1. YLYann LeCun1:10:26

       we came up with a different way of doing things that are called JEPA.

    2. NKNikhil Kamath1:10:28

       Mm-hmm.

    3. YLYann LeCun1:10:28

       So it's a different architecture-

    4. NKNikhil Kamath1:10:30

       Mm-hmm

    5. YLYann LeCun1:10:30

       ... and that means Joint Embedding Predictive Architecture. And what it means is-

    6. NKNikhil Kamath1:10:34

       I watched this for a long time on your Lex Fridman interview-

    7. YLYann LeCun1:10:37

       Right, okay

    8. NKNikhil Kamath1:10:37

       ... when you spoke about JEPA, and I still don't get it. [laughing]

    9. YLYann LeCun1:10:40

       [chuckles] Okay, here's a basic idea.

    10. NKNikhil Kamath1:10:42

        Yeah.

    11. YLYann LeCun1:10:42

        Tell me if you don't understand, 'cause I can-

    12. NKNikhil Kamath1:10:44

        Yeah

    13. YLYann LeCun1:10:44

        ... explain it in different ways. Uh, instead of taking a piece of video and training a big neural net to predict all the pixels of the continuation of that video-

    14. NKNikhil Kamath1:10:53

        Mm-hmm

    15. YLYann LeCun1:10:54

        ... you, y- you, you take the, the video, and you run it through an encoder-

    16. NKNikhil Kamath1:10:59

        Mm-hmm

    17. YLYann LeCun1:10:59

        ... which is gonna be a big neural net that's gonna produce an abstract representation of the video. Okay? And then you take the remainder of the video, the, the, the future, uh, you know, the second half of that video, run it through the same encoder, and then you, you train your prediction system, which is similar perhaps to an LLM.

    18. NKNikhil Kamath1:11:16

        Much like LLM, where you delete a part of the data to train the model?

    19. YLYann LeCun1:11:19

        That's right. So, you know, an LLM, you, you take a, a piece of text, and you train it to predict the remainder of the text, right?

    20. NKNikhil Kamath1:11:24

        Mm-hmm.

    21. YLYann LeCun1:11:24

        And you do this word by word, but you can do the... Well, you can predict multiple words. So here, we're gonna do the same thing. We're gonna take a, a video, and then train the system to predict the remainder of the video. But instead of predicting all the pixels in the video, we're going to run those videos through encoders, which are going to compute abstract representations of the video, and we're gonna do the prediction in that space of representations. So instead of predicting pixels, we predict abstract representations of those pixels, where all the things that are basically unpredictable have been eliminated from the representation. So-

    22. NKNikhil Kamath1:11:57

        Is that a bit like also

33. 1:11:58 – 1:14:10

    How far in the future can you predict through JEPA ?

    1. NKNikhil Kamath1:11:58

       predicting tomorrow? 'Cause if I were to-

    2. YLYann LeCun1:12:01

       Yeah

    3. NKNikhil Kamath1:12:01

       ... video my life up until now and run it through the enco- encoder, it will give me some kind of representation of tomorrow?

    4. YLYann LeCun1:12:08

       Well, yes, but at an abstract level, right?

    5. NKNikhil Kamath1:12:10

       Mm-hmm.

    6. YLYann LeCun1:12:11

       So you can predict, um... You're based in Bangalore, I, I-

    7. NKNikhil Kamath1:12:15

       Yeah

    8. YLYann LeCun1:12:15

       ... I heard.

    9. NKNikhil Kamath1:12:15

       Yeah.

    10. YLYann LeCun1:12:15

        So at some point, you're gonna fly back to Bangalore.

    11. NKNikhil Kamath1:12:17

        Yeah.

    12. YLYann LeCun1:12:18

        Uh, and you can predict how long it's gonna take to go back to Bangalore, but you cannot predict all the details of what will happen, uh, during your, your journey-

    13. NKNikhil Kamath1:12:29

        Right

    14. YLYann LeCun1:12:29

        ... back to Bangalore, exactly how long it's gonna take, given traffic.

    15. NKNikhil Kamath1:12:32

        How far can you extrapolate what will happen three months from now if I have data, video data of the last ten years of my life?

    16. YLYann LeCun1:12:39

        So here's the trick, the, uh, so the interesting question: You can predict very long term-

    17. NKNikhil Kamath1:12:44

        Mm-hmm

    18. YLYann LeCun1:12:45

        ... but the, the longer in the future you can predict, the more abstract the representation level at which you, you can make the prediction.

    19. NKNikhil Kamath1:12:52

        Let me ask you a question. If you were to extrapolate fifty years forward, all of our lives, you figure out how to build this architecture, and it's implemented, and it's working, where video of our life up until now has been programmed into it, and we are trying to predict fifty years forward, what do you suspect you will see? Climate change and world war?

    20. YLYann LeCun1:13:13

        ... [chuckles] So what I see is, okay, uh, so there is a, a plan for the next few years-

    21. NKNikhil Kamath1:13:18

        Mm.

    22. YLYann LeCun1:13:18

        -to build systems that can understand the world from video.

    23. NKNikhil Kamath1:13:20

        Right.

    24. YLYann LeCun1:13:20

        Um, perhaps what they'll be able to learn are those world models-

    25. NKNikhil Kamath1:13:24

        Mm.

    26. YLYann LeCun1:13:24

        -which are action conditions, so they, they will be able to imagine what the consequence of a, an action or a sequence of action will be.

    27. NKNikhil Kamath1:13:31

        Mm-hmm.

    28. YLYann LeCun1:13:31

        They'll be perhaps able to plan complex sequences of action hierarchically-

    29. NKNikhil Kamath1:13:36

        Mm.

    30. YLYann LeCun1:13:36

        -because those world models will be hierarchical. They will have world models that can predict really short term-
34. 1:14:10 – 1:16:30

    AI’s future prediction - Utopian or Dystopian ?

    1. YLYann LeCun1:14:10

       satisfy certain criteria that, that you have. So if we can build systems-

    2. NKNikhil Kamath1:14:13

       If AI, AI were to predict the future, would it be utopian or dystopian?

    3. YLYann LeCun1:14:17

       It would be utopian, uh, because it would be just a, a, a, an alternative way for-

    4. NKNikhil Kamath1:14:23

       Mm-hmm

    5. YLYann LeCun1:14:23

       ... predicting the future than our brains, and for planning action sequences to satisfy certain conditions, to achieve goals that is, a alternative to using our brains.

    6. NKNikhil Kamath1:14:35

       Mm.

    7. YLYann LeCun1:14:35

       Perhaps accumulating more knowledge, eh, to be able to do this, and perhaps having abilities that humans don't have because of the limitations of our brain, right? Computers can calculate and stuff like that, right? So, so the, the future is that if we succeed in this plan, which may succeed within the next five or ten years, you know, five to ten years, we'll have systems that, as time goes by, we can build up to become as intelligent as humans, perhaps.

    8. NKNikhil Kamath1:15:02

       Mm.

    9. YLYann LeCun1:15:02

       So reach, uh, human level intelligence within a decade. That may be optimistic, all right?

    10. NKNikhil Kamath1:15:07

        Mm-hmm. Mm-hmm.

    11. YLYann LeCun1:15:07

        Um, five to ten years would be if everything goes great, all the plans that we're, we've been making will succeed. We're not going to encounter unexpected obstacles, but that is almost certainly not going to happen.

    12. NKNikhil Kamath1:15:20

        You don't like that, right? Like, AGI and h- human-level intelligence, you think is far, far away or unlikely?

    13. YLYann LeCun1:15:27

        No, I, I don't think it's that far away.

    14. NKNikhil Kamath1:15:29

        Mm.

    15. YLYann LeCun1:15:29

        I, I don't think my opinion about how far it is are very different from what you will hear from Sam Altman or Demis Hassabis-

    16. NKNikhil Kamath1:15:34

        Mm

    17. YLYann LeCun1:15:34

        ... or things like this. Um, it's, you know, quite possibly within a decade, but it's not gonna, it's not gonna happen next year. It's not gonna happen in two years.

    18. NKNikhil Kamath1:15:42

        Mm-hmm.

    19. YLYann LeCun1:15:42

        It's gonna, it's gonna take longer, and so you don't want to extrapolate the capabilities of LLM and, and say, "We're just going to scale up LLM, train them on w- with bigger computers or more data, and, you know, human-level intelligence, intelligence is going to emerge." This... It's not going to work this way. We're going to have to have those new architectures, those JEPA systems that learn from, uh, from the real world, um, and can plan hierarchically, uh, can, can plan a sequence of action, eh, you know, as opposed to just producing one word after the other, essentially, without thinking. So system two, instead of system one. LLMs are system one. The architecture I'm describing, which I call objective-driven AI, is system two. Yeah.

    20. NKNikhil Kamath1:16:21

        I'd love to come, like, do a course at your college and learn, if you'll have me as a student.

    21. YLYann LeCun1:16:26

        [chuckles]

    22. NKNikhil Kamath1:16:27

        I don't know if I qualify. I'll have to go back and finish high school,

35. 1:16:30 – 1:18:50

    The LLM Loop | What needs to change

    1. NKNikhil Kamath1:16:30

       but-

    2. YLYann LeCun1:16:30

       [chuckles]

    3. NKNikhil Kamath1:16:31

       ... would, would love it. Just to finish the LLM loop, so because it's in the news and everybody's talking about-

    4. YLYann LeCun1:16:39

       Sure

    5. NKNikhil Kamath1:16:39

       ... LLMs. So you define a problem, you find a large dataset. Most of the time goes in cleaning the data. You choose a model, you train the model, and then you execute the model. Uh-

    6. YLYann LeCun1:16:51

       Before that, you fine-tune the model.

    7. NKNikhil Kamath1:16:53

       Before that, you fine-tune the model. Yes. What will change here?

    8. YLYann LeCun1:16:59

       Um, so there's still going to be a need for collecting data and, uh, and filtering data to, to pre- to, to keep high-quality data and basically get rid of junk.

    9. NKNikhil Kamath1:17:10

       Mm-hmm.

    10. YLYann LeCun1:17:10

        Um, that's actually a pretty expensive part of the whole thing.

    11. NKNikhil Kamath1:17:13

        Mm-hmm.

    12. YLYann LeCun1:17:14

        But I think what's gonna need to happen in that respect is that currently the, eh, the, the, you know, LLMs are trained with a combination of publicly available data and licensed data, basically.

    13. NKNikhil Kamath1:17:26

        Mm-hmm.

    14. YLYann LeCun1:17:27

        Um, but it's mostly publicly available data, you know, publicly available text on the Internet.

    15. NKNikhil Kamath1:17:31

        Right.

    16. YLYann LeCun1:17:31

        And it's extremely biased in many ways, that, um, the, the, uh, a lot of it is in English. Um, uh, you know, this, you know, significant amount of, of data in, in commonly spoken languages like Hindi, but not so much in all twenty-two official languages of India-

    17. NKNikhil Kamath1:17:50

        Yeah

    18. YLYann LeCun1:17:50

        ... and certainly not in all the seven hundred dialects or, or whatever the number is, particularly since most of those dialects are not written, [chuckles] so-

    19. NKNikhil Kamath1:17:58

        Mm

    20. YLYann LeCun1:17:58

        ... are only spoken. So, um, what we need in the future is, uh, datasets that are more encompassing so that the, the systems that are trained with it understand all the world's languages, all the world's cultures, all the value systems-

    21. NKNikhil Kamath1:18:13

        Mm-hmm

    22. YLYann LeCun1:18:13

        ... you know, everything. And no single entity, I think, it would be able to, to do this.

    23. NKNikhil Kamath1:18:17

        Mm-hmm.

    24. YLYann LeCun1:18:18

        Um, which is why I think the future of AI is, uh, AI is going to become a, a, a kind of common infrastructure which people will use as a repository of all human knowledge, and this cannot be built by a single entity. It's gonna, it's gonna have to be a collaborative, uh, project-

    25. NKNikhil Kamath1:18:39

        Right

    26. YLYann LeCun1:18:39

        ... with training being distributed all around the world-

    27. NKNikhil Kamath1:18:43

        Mm-hmm

    28. YLYann LeCun1:18:43

        ... so that you can have models that is trained on all data around the world, but you don't have to copy the data anywhere.

    29. NKNikhil Kamath1:18:49

        And a

36. 1:18:50 – 1:21:09

    Building data centers in India - Yann’s thoughts

    1. NKNikhil Kamath1:18:50

       private digression. I, I was reviewing a data center business to invest into. Uh, a lot of people tell me that compute as a commodity will soon be sold outside of the data center and not inherently in it. Is it a good place to focus energy and time on, like building data centers out of India? I'm taking the sovereign AI model, where every country will probably fight to retain their data a bit more than they're doing currently.

    2. YLYann LeCun1:19:16

       Yeah. So in, in that kind of future, which also I, I-... alluded to with the distributed training of models, uh, having local computing infrastructure, I think is very important. So yes, I think that's kind of crucial. It's crucial for two reasons. One is, uh, hav- having local ability to, to train models, okay?

    3. NKNikhil Kamath1:19:34

       Mm-hmm.

    4. YLYann LeCun1:19:34

       Uh, and the second one is, um, having very low-cost access to inference for AI systems. Because if you want AI systems to be used by, I don't know, eight hundred million Indians, [chuckles] right? Um, uh, I know there are more Indians than this, but most pe- you know, uh, not everybody will use AI systems. But, um, it's a lot of computing infrastructure. It's actually much bigger than the infrastructure for, for learning. Um, and there is this scenario for which there is a lot more innovation than, than training. Training is dominated by NVIDIA at the moment.

    5. NKNikhil Kamath1:20:09

       Mm-hmm.

    6. YLYann LeCun1:20:09

       There's going to be other players-

    7. NKNikhil Kamath1:20:10

       Mm.

    8. YLYann LeCun1:20:10

       ... but they, they have a hard time, um, uh, competing because of the software stack, basically. Uh, their hardware may be really good, but the software stack is, uh, is a challenge. For inference, though, there is a lot more innovation there.

    9. NKNikhil Kamath1:20:21

       Mm-hmm.

    10. YLYann LeCun1:20:22

        And, and that innovation is bringing down the cost. Uh, I think the cost of inference for LLM has gone down by a factor of one hundred in two years. I mean, it should-- it's amazing, right? It's way faster than Moore's Law. Uh, and I think there is still a lot of room for improvement, and you need that because you basically need the inference for a million tokens to be a few rupees. Um, so, so that, that's a big future if you want to deploy-

    11. NKNikhil Kamath1:20:46

        [chuckles]

    12. YLYann LeCun1:20:47

        ... AI assistance widely, uh, in India. Um-

    13. NKNikhil Kamath1:20:50

        I wanna, I wanna use the time, Yann, 'cause I realize we are running out, to bring it into the Indian context. Are people watching this, like I said, are entrepreneurs in play or people trying to be entrepreneurs? As an Indian twenty-year-old

37. 1:21:09 – 1:26:18

    What should a 25 y/o build in the AI space? | Careers in the AI space

    1. NKNikhil Kamath1:21:10

       who wants to build a business in AI, a career in AI, what do we do? Like, as we sit today.

    2. YLYann LeCun1:21:18

       A twenty-year-old, uh, today, I will cross my finger so that when I graduate at twenty-two, uh, there will be good PhD programs, you know, uh, in India.

    3. NKNikhil Kamath1:21:29

       Outside of the academic lens. I, I mean more-

    4. YLYann LeCun1:21:32

       No, no, but that, that's, that's what I need to train myself to innovate. You know, doing a PhD or gradu- graduate studies, it trains, it trains you to invent new things and, and also make sure that the methodology you use, um, prevent you, you from, from fooling yourself into thinking you're being, inno-- an innovator, but you're not. Okay?

    5. NKNikhil Kamath1:21:52

       What if I'm-

    6. YLYann LeCun1:21:52

       So you, you learn this. You-

    7. NKNikhil Kamath1:21:53

       What if I'm an entrepreneur, a twenty-five-year-old entrepreneur?

    8. YLYann LeCun1:21:56

       You still want to do a PhD if you're an entrepreneur-

    9. NKNikhil Kamath1:21:58

       Okay

    10. YLYann LeCun1:21:58

        ... or at least a master's.

    11. NKNikhil Kamath1:21:59

        Yeah.

    12. YLYann LeCun1:21:59

        Because you want to really sort of learn deep... I mean, you might be doing this by yourself. You don't have to, but it's useful because you, you learn more about, you know, what exists out there, what's possible, what's not possible, what, uh, uh, you get more, uh, legitimacy in hiring talented people. I mean, there's a, a lot of advantages, you know, p- particularly in a complex, uh, deeply technical, uh, uh, area, like, like AI. Um, you might succeed if you don't, you know, that's not the issue, but, but it gives you kind of a different perspective. Okay, but now, you, you know, you're doing your PhD, you're cr- you're doing your start-up, it might be easier to raise money if you've published a few papers where you've invented something new.

    13. NKNikhil Kamath1:22:38

        Mm.

    14. YLYann LeCun1:22:38

        And you say, "Well, this is a new technique that really may make a difference."

    15. NKNikhil Kamath1:22:42

        Yeah.

    16. YLYann LeCun1:22:42

        You know, you go see an investor.

    17. NKNikhil Kamath1:22:43

        What if, what if I were to even, like, go one step further? Let's say intelligence. I'm, I'm going to leave the AGI side of it. Let's say narrow intelligence.

    18. YLYann LeCun1:22:51

        Uh-huh.

    19. NKNikhil Kamath1:22:51

        Self-driving cars, robots, all of that. What should I build in? If I have to pick a subset where I can use narrow intelligence through any of the models that we spoke about, what would I start, which has a capitalistic leg to it?

    20. YLYann LeCun1:23:08

        Okay. So today?

    21. NKNikhil Kamath1:23:10

        Today.

    22. YLYann LeCun1:23:11

        Like right now?

    23. NKNikhil Kamath1:23:11

        Yeah.

    24. YLYann LeCun1:23:12

        Uh, the, the most likely business model, uh, that has to do with AI is taking a open source foundation model-

    25. NKNikhil Kamath1:23:22

        Mm

    26. YLYann LeCun1:23:22

        ... like Llama-

    27. NKNikhil Kamath1:23:23

        Mm-hmm

    28. YLYann LeCun1:23:23

        ... which is the-

    29. NKNikhil Kamath1:23:24

        Yeah

    30. YLYann LeCun1:23:24

        ... Meta open source system, which is used everywhere now, right? Every, almost every start-up uses, u- uses it, uh, even large companies.
38. 1:26:18 – 1:29:13

    What should an investor invest in | Yann’s future prediction

    1. NKNikhil Kamath1:26:18

       would a investor benefit from investing into AI? Would it be NVIDIA, Llama, Meta-

    2. YLYann LeCun1:26:26

       [chuckles]

    3. NKNikhil Kamath1:26:26

       ... ChatGPT, OpenAI?

    4. YLYann LeCun1:26:28

       Okay, uh, so I think the first order of thing-

    5. NKNikhil Kamath1:26:32

       Uh-huh

    6. YLYann LeCun1:26:32

       -is imagine what the future is going to be five years from now.

    7. NKNikhil Kamath1:26:35

       Mm-hmm.

    8. YLYann LeCun1:26:36

       And basically, that's going be, that's going to be the-

    9. NKNikhil Kamath1:26:39

       I suspect you would do a much better job at imagining the future than I, [chuckles] Yann.

    10. YLYann LeCun1:26:43

        [laughs]

    11. NKNikhil Kamath1:26:44

        Can you pick, uh... Can you depict a future five years from now?

    12. YLYann LeCun1:26:47

        So five years from now, the world is going to be dominated by open source platforms. Uh, for the same reason that the world of, you know, uh, embedded devices and operating system is dominated by Linux.

    13. NKNikhil Kamath1:26:59

        Mm-hmm.

    14. YLYann LeCun1:26:59

        Uh, you know, the entire world runs on Linux, and it wasn't the case twenty years ago-

    15. NKNikhil Kamath1:27:03

        Mm-hmm

    16. YLYann LeCun1:27:04

        ... uh, twenty-five years ago.

    17. NKNikhil Kamath1:27:05

        Yeah.

    18. YLYann LeCun1:27:05

        Um, and it's become so because open source platforms are more portable, they're more flexible, they're more secure, they're, you know, they, they, they're cheaper to-

    19. NKNikhil Kamath1:27:14

        I shouldn't take-

    20. YLYann LeCun1:27:14

        -deploy.

    21. NKNikhil Kamath1:27:14

        I shouldn't take credit for this, but we have somebody called Kailash, who's our CTO, who's a big proponent of this, and everything we do is open source. We have a fund which gives grants to open source companies and stuff like that.

    22. YLYann LeCun1:27:26

        Right. Okay, so the world is going to be open source.

    23. NKNikhil Kamath1:27:28

        Yeah.

    24. YLYann LeCun1:27:28

        We're going to have open source AI platforms. Uh, in a few years, they'll probably be trained in a distributed fashion, so they're not going to be kind of completely controlled by a single company.

    25. NKNikhil Kamath1:27:37

        Mm-hmm.

    26. YLYann LeCun1:27:38

        Uh, the proprietary engines, I think, are not going to be nearly as important as they are today because the open source platforms are catching up in terms of performance. And then what we know is that a fine-tune open source, uh, uh, engine like Llama-

    27. NKNikhil Kamath1:27:52

        Mm-hmm

    28. YLYann LeCun1:27:53

        ... always works better than a non fine-tune, generic, uh, uh, you know, top-performing model. So-

    29. NKNikhil Kamath1:27:59

        But if everything is open sourced, it'll also be democratized for a investor to invest into, then what is the differentiation?

    30. YLYann LeCun1:28:07

        Well, it, it enables the ecosystem.
39. 1:29:13 – 1:32:09

    What happens to human intelligence in an AI world?

    1. NKNikhil Kamath1:29:13

       this changing? What becomes... Forget computers and AI for a second. For humans, what is intelligence in that world?

    2. YLYann LeCun1:29:22

       So people's intelligence will be moving to a different set of tasks than the one we are trying to do today. Um, because a lot of what we're trying to do today will be done by AI systems, and so we will focus on other tasks. So things like not doing things, but deciding what to do, or figuring out what to do. Okay? Those are two different things. Like, think about the difference between, uh, a low-level employee in a company-

    3. NKNikhil Kamath1:29:50

       Mm-hmm

    4. YLYann LeCun1:29:50

       ... that is told what to do and just does it, and then, you know, a high-level manager in the company that has to figure out, like, strategy and think about, like, what to do, and then tell people, you know, what to do. Okay? We're going to-- We're all going to be a boss. We're all going to be like those, uh, high-level managers. We're going to tell our AI assistants what to do, but we're not going to have to do it ourselves necessarily, okay? So the type-

    5. NKNikhil Kamath1:30:14

       But we need lesser people to tell something more efficient than us what to do than we need today for them to actually do the task, right?

    6. YLYann LeCun1:30:23

       Yeah.

    7. NKNikhil Kamath1:30:23

       So what happens to everyone else?

    8. YLYann LeCun1:30:26

       Well, I think everyone is going to be in that situation, i- is going to have access to AI assistants and, um, and, and be able to delegate a lot of, uh, tasks, uh, you know, mostly in the virtual world, but eventually in the real world. Uh, we're going to have, at some point, domestic robots and, and, uh, [chuckles] self-driving cars and things like this. Once we figure out how to get the system to learn the, how the real world, uh, works, uh, from video. Uh, but, um, uh, so, so the, the, the type of tasks on which we are going to be able to concentrate ourselves are going to be more abstract. The same way, you know, nobody needs to do, like, super fast mental arithmetics anymore, we have calculators.

    9. NKNikhil Kamath1:31:04

       Mm-hmm. Mm-hmm.

    10. YLYann LeCun1:31:05

        Uh, or, or, you know, solve integrals or differential equations. We, we have to learn the basics of how we do this, but we have-- you know, we can use, uh, computer tools to do this-

    11. NKNikhil Kamath1:31:14

        Mm

    12. YLYann LeCun1:31:14

        ... right? Um, so, so it's going to lift the abstraction level at which we can place ourselves-

    13. NKNikhil Kamath1:31:20

        Mm-hmm

    14. YLYann LeCun1:31:20

        ... and basically enable us to be more creative, uh, be more productive. Okay? And, and there are a lot of things that-... you and I have learned to do that our descendants would not have to learn to do, uh, because that will be taken care of-

    15. NKNikhil Kamath1:31:38

        Go to school

    16. YLYann LeCun1:31:38

        -by machines.

    17. NKNikhil Kamath1:31:39

        Like go to school?

    18. YLYann LeCun1:31:41

        No, no, we'll still go to school.

    19. NKNikhil Kamath1:31:42

        [chuckles]

    20. YLYann LeCun1:31:42

        We'll have to educate ourselves.

    21. NKNikhil Kamath1:31:43

        Mm.

    22. YLYann LeCun1:31:43

        We'll have to... There's still going to be the, the, you know, the, the competition between humans to kind of do something better than the others, or something different-

    23. NKNikhil Kamath1:31:51

        Mm

    24. YLYann LeCun1:31:51

        ... more creative.

    25. NKNikhil Kamath1:31:51

        Always, right? Innately, we-

    26. YLYann LeCun1:31:53

        Always

    27. NKNikhil Kamath1:31:53

        -want to compete with our peer groups.

    28. YLYann LeCun1:31:54

        Yeah, yeah. So we're not gonna run out of jobs.

    29. NKNikhil Kamath1:31:56

        Mm.

    30. YLYann LeCun1:31:56

        Economists that I talk to tell me-
40. 1:32:09 – 1:34:00

    What is intelligence? | Final Verdict

    1. NKNikhil Kamath1:32:09

       trying to define what is intelligence, really. I had written down, "Intelligence is a collection of information and the ability to absorb new skills."

    2. YLYann LeCun1:32:18

       It's a collection of skills and an ability to learn new skills really quickly, or an ability to solve problems without learning.

    3. NKNikhil Kamath1:32:27

       Mm-hmm.

    4. YLYann LeCun1:32:27

       This is called zero shot in the, in the-

    5. NKNikhil Kamath1:32:28

       Mm

    6. YLYann LeCun1:32:29

       ... AI business. You know, you're faced with a new problem, and you can think about it for a while, and you may not- you may never have faced similar problem before, but you can solve it by just, you know, thinking and, and using your mental model of, uh, of the situation. That's called zero shot. You're not learning a new skill, you're just solving a problem from scratch. Uh, so the combination of those three things, you know, having already a num- number of skills that you know, experience with solving problems and accomplishing tasks, being able to learn new tasks really quickly with a few trials. Um, and then the next step is being a- able to solve new problems, zero shot, without having to learn anything new.

    7. NKNikhil Kamath1:33:10

       Mm-hmm.

    8. YLYann LeCun1:33:11

       Um, that's the combination of those three thing- things really is, is intelligence.

    9. NKNikhil Kamath1:33:16

       Now, thank you, Yann, so much for doing this. Uh, I'm gonna try and figure out how I can do, like, a course under you wherever you're teaching. Maybe you can recommend me, uh, to the college to give me a seat so I can attend some lectures. But I'd love to-

    10. YLYann LeCun1:33:30

        We can do better.

    11. NKNikhil Kamath1:33:31

        Uh? [chuckles]

    12. YLYann LeCun1:33:31

        Uh, the 2021 edition of my deep learning course-

    13. NKNikhil Kamath1:33:35

        Yeah

    14. YLYann LeCun1:33:35

        ... is freely available on the internet for free. It's all on YouTube, all the problems, all the-

    15. NKNikhil Kamath1:33:41

        Uh

    16. YLYann LeCun1:33:41

        ... um, the, uh, homework and everything. [laughs] If you have, you know, any free time.

    17. NKNikhil Kamath1:33:44

        I, I feel like I'm going back to old school. I feel like being in front of you and learning first-person has a innate value of its own. So I'll, I'll try and do this and that.

    18. YLYann LeCun1:33:54

        [laughs] Wonderful.

    19. NKNikhil Kamath1:33:55

        Thank you so much, Yann, for doing this. Thank you.

    20. YLYann LeCun1:33:57

        Pleasure. A real pleasure. [upbeat music]

41. 1:34:00 – 1:36:05

    The end or just the beginning?

    1. NKNikhil Kamath1:34:04

       Thank you. That was fun.

    2. YLYann LeCun1:34:05

       That was fun, yeah.

    3. NKNikhil Kamath1:34:06

       You didn't get bored?

    4. YLYann LeCun1:34:10

       No. [laughs]

    5. NKNikhil Kamath1:34:11

       [laughs]

    6. YLYann LeCun1:34:12

       You asked... You asked a professor to speak, like, you know, that's the job. [laughs]

    7. NKNikhil Kamath1:34:18

       But I, I guess, like, when you're talking to people who know so much lesser than you, it can't be fun all the time.

    8. YLYann LeCun1:34:24

       It's, it's an art. Uh, I mean, I-

    9. NKNikhil Kamath1:34:26

       [laughs]

    10. YLYann LeCun1:34:26

        ... I don't, I don't claim to be, I don't claim to be particularly good at it, but-

    11. NKNikhil Kamath1:34:30

        Yeah

    12. YLYann LeCun1:34:31

        ... I, I try hard. So, you know, trying to kind of simplify concepts and stuff like that.

    13. NKNikhil Kamath1:34:34

        Yeah.

    14. YLYann LeCun1:34:35

        So-

    15. NKNikhil Kamath1:34:35

        But I think we needed it, 'cause so many Indians are speaking about AI, but so few of us actually understand what went behind where we are today.

    16. YLYann LeCun1:34:44

        Oh, that's true across the world. It's not just India. [chuckles]

    17. NKNikhil Kamath1:34:48

        Yeah.

    18. YLYann LeCun1:34:48

        In fact, I think it's kind of the opposite in India. There's, like, way more people who are, who are kind of educating themselves, like, particularly among the young people.

    19. NKNikhil Kamath1:34:55

        So we wanted to focus today on that, like, just to get to telling our people... A lot of young people watch this, young, bright people-

    20. YLYann LeCun1:35:03

        Yeah

    21. NKNikhil Kamath1:35:03

        ... how we got to be where we are today. So most questions around that.

    22. YLYann LeCun1:35:08

        Yeah, I think it's important because it, it, it helps convince people that, uh, they can do it regardless of what-

    23. NKNikhil Kamath1:35:17

        Yeah.

    24. YLYann LeCun1:35:17

        Like, you know, I went to... I studied engineering in France, but I didn't go to, like, one of the top, you know, equivalent Ivy League or anything. I went to, like, a regular school.

    25. NKNikhil Kamath1:35:26

        Yeah.

    26. YLYann LeCun1:35:26

        Right? Um, [laughs] um, and I didn't do my PhD with a famous person and, you know, all that stuff. And I was in France, I was writing papers in French that were terrible, nobody read, but, you know, kind of managed to [chuckles] do something. And, and people are, are sometimes telling me, like, "You're... You know, you helped convince me that I could do something impactful, even though I didn't go to Harvard or MIT or Stanford."

    27. NKNikhil Kamath1:35:54

        Yeah.

    28. YLYann LeCun1:35:54

        Thank you.

    29. NKNikhil Kamath1:35:54

        It was a good conversation. [upbeat music]

Episode duration: 1:36:05