WEBVTT 00:00:00.000 --> 00:00:02.419 When you think about AI today, it feels like 00:00:02.419 --> 00:00:05.719 it's just everywhere. It's in your phone, recognizing 00:00:05.719 --> 00:00:08.560 your face. It's in cars that are learning to 00:00:08.560 --> 00:00:11.140 drive themselves, these huge language models 00:00:11.140 --> 00:00:13.339 that can write poetry. It's completely woven 00:00:13.339 --> 00:00:16.239 into the fabric of our lives now. Right. But 00:00:16.239 --> 00:00:18.640 if you sort of peel back all those layers, you 00:00:18.640 --> 00:00:20.559 know, the big tech companies, the flashy product 00:00:20.559 --> 00:00:23.219 announcements, who's the person who actually 00:00:23.219 --> 00:00:26.260 built the engine? Who is the architect of the 00:00:26.260 --> 00:00:29.489 fundamental tech that... Let's a machine see. 00:00:30.230 --> 00:00:31.910 That's the question we're digging into today. 00:00:32.049 --> 00:00:35.030 And the person behind that curtain, our subject, 00:00:35.130 --> 00:00:37.560 is Jan -André Lecun. He's a French -American 00:00:37.560 --> 00:00:40.240 computer scientist. And his work really built 00:00:40.240 --> 00:00:42.840 the bridge between these like abstract academic 00:00:42.840 --> 00:00:46.219 ideas and AI that could actually be industrialized. 00:00:46.439 --> 00:00:48.479 He's known for deep learning, of course, but 00:00:48.479 --> 00:00:51.020 also computer vision, mobile robotics. He's foundational 00:00:51.020 --> 00:00:52.979 in all those areas. And you mentioned he's French 00:00:52.979 --> 00:00:55.420 -American. That dual citizenship is actually, 00:00:55.539 --> 00:00:58.280 I think, really interesting. Well, his career 00:00:58.280 --> 00:01:01.780 is this perfect blend of that European perdition 00:01:01.780 --> 00:01:05.359 of deep theoretical work with this very American 00:01:05.359 --> 00:01:07.430 drive for... massive industrial scale application. 00:01:08.290 --> 00:01:11.409 Even his name, LeCun, has this deep history. 00:01:11.469 --> 00:01:13.510 It comes from an old Breton word meaning John. 00:01:13.769 --> 00:01:16.760 So a career that. really mirrors his own identity. 00:01:16.959 --> 00:01:19.299 Exactly. But the big claim, the place you have 00:01:19.299 --> 00:01:22.319 to start with him is his status. He is one of 00:01:22.319 --> 00:01:24.459 the three figures who are, you know, universally 00:01:24.459 --> 00:01:27.340 called the godfathers of AI. Or more specifically, 00:01:27.519 --> 00:01:29.500 the godfathers of deep learning. Right. A more 00:01:29.500 --> 00:01:32.060 accurate term. And that trio, that's him, Jan 00:01:32.060 --> 00:01:35.019 LeCun, alongside Yoshua Bengio and Jeffrey Hinton. 00:01:35.400 --> 00:01:38.340 And their work, it didn't just improve AI a little 00:01:38.340 --> 00:01:40.439 bit. No, it was a complete paradigm shift. It 00:01:40.439 --> 00:01:42.579 was a revolution. They took neural networks from 00:01:42.579 --> 00:01:45.599 this, this kind of marginalized. almost forgotten 00:01:45.599 --> 00:01:48.620 corner of academia in the 80s and 90s, and they 00:01:48.620 --> 00:01:51.420 turned it into the single most important technological 00:01:51.420 --> 00:01:54.319 engine of our time. Okay, let's unpack this. 00:01:54.480 --> 00:01:57.239 So our mission for this deep dive is to trace 00:01:57.239 --> 00:02:00.299 the history of his most famous invention, the 00:02:00.299 --> 00:02:03.260 convolutional neural network, the CNN, and really 00:02:03.260 --> 00:02:06.299 understand how he took that idea from a sketch 00:02:06.299 --> 00:02:08.659 in a lab notebook to something that's driving 00:02:08.659 --> 00:02:12.050 entire global industries. We've got a great stack 00:02:12.050 --> 00:02:14.110 of source material for this. It covers his whole 00:02:14.110 --> 00:02:17.050 career, from his education, his time at Bell 00:02:17.050 --> 00:02:21.129 Labs, NYU, Meta, and a list of awards that is 00:02:21.129 --> 00:02:23.490 just, well, it's pretty staggering. And that 00:02:23.490 --> 00:02:26.110 path is so important for you, the listener, to 00:02:26.110 --> 00:02:28.189 understand if you want to get a real handle on 00:02:28.189 --> 00:02:30.770 AI's history and where it's going, his career 00:02:30.770 --> 00:02:33.069 is like a perfect roadmap for how a technology 00:02:33.069 --> 00:02:35.620 matures. What do you mean by that? Well, first 00:02:35.620 --> 00:02:37.419 you have the fundamental research phase, which 00:02:37.419 --> 00:02:39.259 for him was at a place like Bell Labs. Then you 00:02:39.259 --> 00:02:41.740 have the academic phase where you institutionalize 00:02:41.740 --> 00:02:43.599 that knowledge, which he did at NYU. And then 00:02:43.599 --> 00:02:46.240 the final stage. The final stage is scaling it 00:02:46.240 --> 00:02:48.780 up in a hyper competitive, massive scale environment, 00:02:48.939 --> 00:02:51.219 which is exactly what he did at Meta. So, you 00:02:51.219 --> 00:02:53.280 know, to understand his journey is to understand 00:02:53.280 --> 00:02:56.039 the last 30 years of technology itself. All right. 00:02:56.060 --> 00:02:58.409 Let's start at the very beginning. His early 00:02:58.409 --> 00:03:02.370 academic life. He was born July 8, 1960, in a 00:03:02.370 --> 00:03:04.810 suburb of Paris. In Soississou -Montmorency, 00:03:04.969 --> 00:03:07.509 to be exact. And his education was very technical, 00:03:07.629 --> 00:03:10.710 very rigorous from the start. He got his diplôme 00:03:10.710 --> 00:03:13.050 d 'ingénieur and engineering degree from EZ Paris 00:03:13.050 --> 00:03:16.590 in 1983. And then he went straight into the deep 00:03:16.590 --> 00:03:18.990 theory, getting his Ph .D. in computer science 00:03:18.990 --> 00:03:22.009 in 1987 from what's now part of Sorbonne University. 00:03:22.389 --> 00:03:25.250 And that Ph .D. thesis from 87, it's a really 00:03:25.250 --> 00:03:27.689 important historical document. It's called Modèle 00:03:27.689 --> 00:03:30.509 Connexioniste de l 'Apprentissage, which is just 00:03:30.509 --> 00:03:32.650 connectionist learning models. So even then, 00:03:32.650 --> 00:03:34.710 he was focused on how these networks learned. 00:03:34.810 --> 00:03:37.490 Right from the jump. And in that thesis, he actually 00:03:37.490 --> 00:03:39.949 proposed an early version of the back propagation 00:03:39.949 --> 00:03:42.530 algorithm. Which is, I mean, for anyone listening 00:03:42.530 --> 00:03:44.409 who isn't a computer scientist, that's the absolute. 00:03:44.430 --> 00:03:46.469 core mechanism for teaching a neural network, 00:03:46.629 --> 00:03:49.669 right? It is the core mechanism. The problem 00:03:49.669 --> 00:03:52.610 back then was all about efficiency. You have 00:03:52.610 --> 00:03:54.509 a network with all these layers of artificial 00:03:54.509 --> 00:03:57.129 neurons. You show it something, it makes a guess. 00:03:57.289 --> 00:03:59.509 And if it's wrong, you need a way to send that 00:03:59.509 --> 00:04:02.189 error signal all the way back through the network. 00:04:02.469 --> 00:04:04.729 So every single connection can be adjusted just 00:04:04.729 --> 00:04:07.990 a tiny bit. Exactly. And back propagation is 00:04:07.990 --> 00:04:10.990 the elegant mathematical way to do that efficiently. 00:04:11.909 --> 00:04:14.229 Before that, the methods were just too slow, 00:04:14.389 --> 00:04:16.350 too clunky. They couldn't handle networks that 00:04:16.350 --> 00:04:19.269 were more than a couple of layers deep. LeCun 00:04:19.269 --> 00:04:21.250 was right there, working on the mathematical 00:04:21.250 --> 00:04:23.629 tools that made deep learning even possible. 00:04:23.910 --> 00:04:26.550 So he's tackling this fundamental problem, and 00:04:26.550 --> 00:04:28.810 that work immediately connects him to the other 00:04:28.810 --> 00:04:32.009 godfathers. Right after his PhD in 87, he does 00:04:32.009 --> 00:04:34.269 a postdoc. At the University of Toronto. And 00:04:34.269 --> 00:04:36.930 his mentor there was Geoffrey Hinton. The man 00:04:36.930 --> 00:04:38.910 he'd later share the Turing Award with. It's 00:04:38.910 --> 00:04:41.600 an amazing piece of foreshadowing. That year 00:04:41.600 --> 00:04:44.040 solidified their intellectual bond. You have 00:04:44.040 --> 00:04:46.220 to remember, in the late 80s, these guys were 00:04:46.220 --> 00:04:48.959 not in the mainstream of AI research. Not at 00:04:48.959 --> 00:04:52.139 all. The dominant idea was still symbolic AI, 00:04:52.259 --> 00:04:56.100 logic -based systems. Right. This connectionist 00:04:56.100 --> 00:04:58.540 stuff, these neural networks, were seen as kind 00:04:58.540 --> 00:05:01.680 of a dead end by many. It was a small, tight 00:05:01.680 --> 00:05:03.839 -knit group of researchers who were keeping the 00:05:03.839 --> 00:05:06.379 flame alive during what we now call the AI winter, 00:05:06.500 --> 00:05:09.079 when all the funding and enthusiasm had just 00:05:09.079 --> 00:05:12.800 evaporated. And then, after that postdoc, he 00:05:12.800 --> 00:05:15.019 makes the move that changes everything. The one 00:05:15.019 --> 00:05:17.399 that proves this isn't just theory. Bell Labs. 00:05:18.040 --> 00:05:20.680 In 1988, he joins the Adaptive Systems Research 00:05:20.680 --> 00:05:23.699 Department at AT &T Bell Labs in New Jersey. 00:05:24.019 --> 00:05:26.060 And you really can't overstate what Bell Labs 00:05:26.060 --> 00:05:28.339 was at that time. It was the place to be, the 00:05:28.339 --> 00:05:30.360 world's greatest innovation factor. Without a 00:05:30.360 --> 00:05:32.480 doubt. And he spent eight years there from 88 00:05:32.480 --> 00:05:35.019 to 96. And this is where it all comes together. 00:05:35.180 --> 00:05:37.240 This is where he develops his signature invention. 00:05:37.699 --> 00:05:40.879 The Convolutional Neural Network. The CNN. The 00:05:40.879 --> 00:05:43.610 CNN. Specifically, the architecture that became 00:05:43.610 --> 00:05:46.569 famous as Lynette. And the whole idea was described 00:05:46.569 --> 00:05:49.769 as a biologically inspired model of image recognition. 00:05:50.009 --> 00:05:52.529 He was literally looking at how the animal visual 00:05:52.529 --> 00:05:54.889 cortex works. OK, so let's break that down because 00:05:54.889 --> 00:05:57.490 this is the really critical part. Why did he 00:05:57.490 --> 00:06:00.410 need a whole new type of network just for images? 00:06:00.850 --> 00:06:02.689 What was wrong with the neural networks they 00:06:02.689 --> 00:06:04.910 were already studying? That's the million dollar 00:06:04.910 --> 00:06:08.209 question. And the answer is scale. Just pure 00:06:08.209 --> 00:06:11.480 brute force scale. Think about. A simple black 00:06:11.480 --> 00:06:14.420 and white image, maybe it's 200 pixels by 200 00:06:14.420 --> 00:06:17.180 pixels. That's 40 ,000 inputs for your network 00:06:17.180 --> 00:06:19.279 right there. 40 ,000. Now imagine connecting 00:06:19.279 --> 00:06:23.120 every single one of those 40 ,000 pixels to every 00:06:23.120 --> 00:06:25.920 neuron in the next layer. Even if that next layer 00:06:25.920 --> 00:06:28.720 only has 100 neurons, you're suddenly looking 00:06:28.720 --> 00:06:31.959 at 4 million connections, 4 million different 00:06:31.959 --> 00:06:33.980 weights that the network has to learn. And on 00:06:33.980 --> 00:06:35.939 the computers of the late 80s, that's just... 00:06:36.300 --> 00:06:38.600 Completely impossible. The computational cost 00:06:38.600 --> 00:06:40.680 was astronomical. You didn't have the memory. 00:06:40.879 --> 00:06:43.199 And even if you did, the network would just overfit, 00:06:43.199 --> 00:06:45.439 basically memorize the training photos instead 00:06:45.439 --> 00:06:48.000 of learning the general concept of, say, what 00:06:48.000 --> 00:06:50.360 a cat looks like. So the data itself was the 00:06:50.360 --> 00:06:53.620 bottleneck. Precisely. And Lacoon's genius was 00:06:53.620 --> 00:06:56.720 to say, OK, let's not do that. Let's build a 00:06:56.720 --> 00:06:58.800 smarter architecture with some built in constraints 00:06:58.800 --> 00:07:02.000 inspired by how vision actually works. And that's 00:07:02.000 --> 00:07:04.459 where the three core ideas of CNN's come from. 00:07:04.579 --> 00:07:07.220 Which are. Local connectivity, shared weights, 00:07:07.439 --> 00:07:09.860 and pooling. Exactly. Let's take them one by 00:07:09.860 --> 00:07:13.360 one. Local connectivity just means that a neuron 00:07:13.360 --> 00:07:15.939 in one layer doesn't connect to the entire image. 00:07:16.220 --> 00:07:19.120 It only connects to a small local patch of pixels 00:07:19.120 --> 00:07:21.839 right in its neighborhood. So it's not trying 00:07:21.839 --> 00:07:23.439 to see the whole picture at once. It's just looking 00:07:23.439 --> 00:07:25.439 for a tiny feature, like an edge or a corner. 00:07:25.600 --> 00:07:28.459 Just a little clue. And then comes the real magic 00:07:28.459 --> 00:07:31.399 trick, shared weights. This means you use the 00:07:31.399 --> 00:07:33.439 exact same little feature detector, the same 00:07:33.439 --> 00:07:35.379 set of weights, which we call a filter, and you 00:07:35.379 --> 00:07:38.139 slide it across the entire image. Ah, I see. 00:07:38.220 --> 00:07:40.100 So if you have a filter that's good at finding 00:07:40.100 --> 00:07:42.879 a vertical line, you use that same filter to 00:07:42.879 --> 00:07:45.319 look for vertical lines in the top left, the 00:07:45.319 --> 00:07:47.779 bottom right, everywhere. That's the key insight. 00:07:48.180 --> 00:07:50.939 And what that does is, first, it dramatically, 00:07:51.279 --> 00:07:53.569 I mean... It dramatically reduces the number 00:07:53.569 --> 00:07:55.490 of parameters the network has to learn. Instead 00:07:55.490 --> 00:07:57.529 of millions of weights, you might only be learning 00:07:57.529 --> 00:08:00.550 a few dozen filters. And second, it gives the 00:08:00.550 --> 00:08:03.230 network a built -in superpower. Translational 00:08:03.230 --> 00:08:06.170 invariance. The ability to recognize an object 00:08:06.170 --> 00:08:08.730 no matter where it appears in the frame. A cat 00:08:08.730 --> 00:08:10.649 is a cat, whether it's on the left or the right. 00:08:10.850 --> 00:08:14.509 And the last piece, pooling. Pooling, or subsampling, 00:08:14.569 --> 00:08:18.040 is basically a way to shrink the data down. After 00:08:18.040 --> 00:08:20.660 you found all these little features, you summarize 00:08:20.660 --> 00:08:23.000 the information in a region, keeping the most 00:08:23.000 --> 00:08:25.079 important signal and throwing away some of the 00:08:25.079 --> 00:08:28.000 noise. It makes the network more robust to tiny 00:08:28.000 --> 00:08:30.220 shifts and distortions. So it's this brilliant 00:08:30.220 --> 00:08:33.019 hierarchical system. Find small local features, 00:08:33.220 --> 00:08:35.299 then summarize them, then use those summaries 00:08:35.299 --> 00:08:37.519 to find bigger features, and so on. That was 00:08:37.519 --> 00:08:39.460 the architectural breakthrough that made processing 00:08:39.460 --> 00:08:42.000 high -resolution images feasible for the very 00:08:42.000 --> 00:08:44.779 first time. And his work at Bell Labs went beyond 00:08:44.779 --> 00:08:47.600 just the architecture of Lynette. The sources 00:08:47.600 --> 00:08:49.960 also mentioned things like optimal brain damage 00:08:49.960 --> 00:08:52.340 regularization. Right, which is a very dramatic 00:08:52.340 --> 00:08:55.059 name. It does. It sounds a little scary. What 00:08:55.059 --> 00:08:57.440 was that trying to solve? It was another piece 00:08:57.440 --> 00:08:59.440 of the puzzle. Remember, they were working with 00:08:59.440 --> 00:09:02.820 small data sets in weak computers. Overfitting 00:09:02.820 --> 00:09:05.980 was still a huge problem. Optimal brain damage, 00:09:06.139 --> 00:09:09.580 or OBD, was a very clever way to make the network 00:09:09.580 --> 00:09:12.820 simpler and more general. How did it work? After 00:09:12.820 --> 00:09:15.519 you train the network, OBD would analyze all 00:09:15.519 --> 00:09:17.820 the connections and figure out which ones were 00:09:17.820 --> 00:09:20.139 least important for solving the problem. And 00:09:20.139 --> 00:09:22.559 then it would preen them. It would cut them out. 00:09:22.700 --> 00:09:24.379 Literally damaging the brain of the network. 00:09:24.600 --> 00:09:27.419 Exactly. It was a way to keep the essential knowledge 00:09:27.419 --> 00:09:29.559 while getting rid of the unnecessary complexity. 00:09:30.080 --> 00:09:32.879 It made the models much more efficient and much 00:09:32.879 --> 00:09:35.360 better at generalizing to new data they hadn't 00:09:35.360 --> 00:09:37.720 seen before. A vital tool for that low resource 00:09:37.720 --> 00:09:39.559 environment. And then there's something called 00:09:39.559 --> 00:09:42.759 graph transformer networks. Yeah, the GTNs. That 00:09:42.759 --> 00:09:45.759 was more focused on processing sequences of data, 00:09:45.860 --> 00:09:48.799 not just static images. Think about trying to 00:09:48.799 --> 00:09:51.139 read cursive handwriting, where one letter flows 00:09:51.139 --> 00:09:53.220 into the next. It's not a single image. It's 00:09:53.220 --> 00:09:57.200 a stream. GTNs were a way to apply these connectionist 00:09:57.200 --> 00:10:01.039 ideas to that kind of structured sequential data. 00:10:01.279 --> 00:10:03.019 Which all shows he was thinking about the whole 00:10:03.019 --> 00:10:06.600 ecosystem of problems, not just one thing. But 00:10:06.600 --> 00:10:08.889 here's where... For me, the story gets really 00:10:08.889 --> 00:10:10.950 incredible. This isn't just a theory that sat 00:10:10.950 --> 00:10:13.970 in a lab. Bell Labs immediately gave him a way 00:10:13.970 --> 00:10:16.649 to apply this on a massive industrial scale. 00:10:16.850 --> 00:10:19.190 This is the killer app. The proof that this wasn't 00:10:19.190 --> 00:10:21.929 just an academic curiosity. The application was 00:10:21.929 --> 00:10:24.659 handwriting recognition. specifically for reading 00:10:24.659 --> 00:10:27.039 the numbers on bank checks. And this wasn't some 00:10:27.039 --> 00:10:29.720 small pilot program. Lacan and his team built 00:10:29.720 --> 00:10:32.419 a system that was widely deployed by NCR and 00:10:32.419 --> 00:10:34.679 other big companies that made banking hardware. 00:10:34.940 --> 00:10:37.019 It's hard to imagine the scale of that now. Every 00:10:37.019 --> 00:10:39.379 day, millions and millions of paper checks were 00:10:39.379 --> 00:10:41.379 moving through the financial system. It was a 00:10:41.379 --> 00:10:45.279 huge, messy, complex problem. You have all these 00:10:45.279 --> 00:10:47.340 different handwriting styles, different pens, 00:10:47.460 --> 00:10:50.580 smudges, and the system had to be incredibly 00:10:50.580 --> 00:10:52.580 accurate. You can't make mistakes when you're 00:10:52.580 --> 00:10:54.860 reading the dollar amount on a check. And the 00:10:54.860 --> 00:10:57.879 sources claim that this system, built on Lynette, 00:10:58.059 --> 00:11:01.970 was reading, what was it, over 10%. Of all checks 00:11:01.970 --> 00:11:04.629 in the United States in the late 90s and early 00:11:04.629 --> 00:11:07.789 2000s. 10%. Just stop and think about that number. 00:11:07.870 --> 00:11:10.769 That's not a research project. That is a core 00:11:10.769 --> 00:11:12.889 piece of the national financial infrastructure. 00:11:13.730 --> 00:11:16.230 Billions of transactions processed by a deep 00:11:16.230 --> 00:11:18.250 learning network that the public had never even 00:11:18.250 --> 00:11:20.149 heard of. It's the ultimate proof of concept. 00:11:20.309 --> 00:11:22.990 It proved decades before the current AI boom 00:11:22.990 --> 00:11:25.429 that neural networks were accurate, reliable 00:11:25.429 --> 00:11:27.750 and scalable enough for mission critical industrial 00:11:27.750 --> 00:11:30.360 work. It answered the big question. Can you trust 00:11:30.360 --> 00:11:32.919 these things? And the banking industry implicitly 00:11:32.919 --> 00:11:35.340 said, absolutely yes. Which brings up a really 00:11:35.340 --> 00:11:38.360 fascinating question. If CNNs were so successful 00:11:38.360 --> 00:11:40.559 that they were reading 10 % of U .S. checks in 00:11:40.559 --> 00:11:43.639 the 90s, why did the AI winter continue? Why 00:11:43.639 --> 00:11:45.559 did it take another 15 years for the rest of 00:11:45.559 --> 00:11:47.539 the world to catch on to deep learning? That's 00:11:47.539 --> 00:11:49.500 the million -dollar question, and it's so important 00:11:49.500 --> 00:11:52.220 for understanding the history. Lacan had provided 00:11:52.220 --> 00:11:55.019 the blueprint. He had the algorithm. But two 00:11:55.019 --> 00:11:57.440 other essential ingredients were missing at scale. 00:11:57.700 --> 00:12:00.620 Hardware and data. Hardware and data. Training 00:12:00.620 --> 00:12:03.639 Lynette to read digits was possible on 90s computers. 00:12:04.039 --> 00:12:07.419 But trying to use those same ideas on much bigger, 00:12:07.440 --> 00:12:10.620 full -color images from the Internet or training 00:12:10.620 --> 00:12:12.879 networks that were 10 times deeper, it was just 00:12:12.879 --> 00:12:15.700 computationally out of reach. So we needed two 00:12:15.700 --> 00:12:19.039 things to happen. We needed massive labeled data 00:12:19.039 --> 00:12:21.019 sets, which we eventually got with things like 00:12:21.019 --> 00:12:24.340 ImageNet. And we needed cheap, massively parallel 00:12:24.340 --> 00:12:27.360 computers, which we got, ironically, from the 00:12:27.360 --> 00:12:30.580 video game industry. GPUs. Gravix Processing 00:12:30.580 --> 00:12:33.179 Unit. Exactly. The algorithm was ready and waiting 00:12:33.179 --> 00:12:36.159 in 1995. The rest of the ecosystem just took 00:12:36.159 --> 00:12:38.399 another decade and a half to catch up to LeCun's 00:12:38.399 --> 00:12:40.759 vision. That makes perfect sense. The blueprint 00:12:40.759 --> 00:12:43.139 for the car existed, but the factory and the 00:12:43.139 --> 00:12:45.360 highways hadn't been built yet. Yeah. So after 00:12:45.360 --> 00:12:48.639 that huge success in 1996, he makes a move within 00:12:48.639 --> 00:12:51.740 AT &T. He goes to AT &T Labs Research and becomes 00:12:51.740 --> 00:12:53.580 head of the Image Processing Research Department. 00:12:54.080 --> 00:12:56.820 And here, his focus shifts a little bit, but 00:12:56.820 --> 00:12:58.940 he's still working on these huge infrastructural 00:12:58.940 --> 00:13:02.120 problems. He and his collaborators, Lambo2 and 00:13:02.120 --> 00:13:04.600 Patrick Hafner, developed this technology called 00:13:04.600 --> 00:13:06.799 D -View. Which is an image compression technology. 00:13:07.039 --> 00:13:09.460 And a very, very smart one. It was designed specifically 00:13:09.460 --> 00:13:12.720 for scanned documents, especially ones from libraries 00:13:12.720 --> 00:13:15.240 and archives that have a mix of text, drawings, 00:13:15.399 --> 00:13:17.940 and photos. What was the big idea behind it? 00:13:18.019 --> 00:13:22.100 How is it better than, say, a JPEG? The genius 00:13:22.100 --> 00:13:25.019 of DZ was that it separated the image into layers. 00:13:25.799 --> 00:13:28.320 It would analyze the page and create a background 00:13:28.320 --> 00:13:30.919 layer for the paper texture, a high -resolution 00:13:30.919 --> 00:13:33.580 foreground layer for the crisp text and lines, 00:13:33.799 --> 00:13:36.580 and another layer for any photos. And then compress 00:13:36.580 --> 00:13:39.019 each layer separately using the best method for 00:13:39.019 --> 00:13:41.320 that type of content. Exactly. So you could aggressively 00:13:41.320 --> 00:13:43.240 compress the background and the photos, but you 00:13:43.240 --> 00:13:45.340 could keep the text perfectly sharp and readable, 00:13:45.519 --> 00:13:48.460 which is crucial for archives. It was so effective 00:13:48.460 --> 00:13:50.980 that the Internet Archive and other big digitization 00:13:50.980 --> 00:13:54.019 projects used it for millions of documents. So 00:13:54.019 --> 00:13:56.179 again, it's him. him tackling these fundamental 00:13:56.179 --> 00:13:59.879 problems of how we manage and access vast amounts 00:13:59.879 --> 00:14:02.559 of information at scale. First recognition with 00:14:02.559 --> 00:14:06.159 CNN's, now dissemination with DJVU. It's a consistent 00:14:06.159 --> 00:14:08.759 theme throughout his entire career. He's always 00:14:08.759 --> 00:14:11.120 thinking about the infrastructure. So after this 00:14:11.120 --> 00:14:13.240 incredibly productive period in corporate research, 00:14:13.320 --> 00:14:16.279 first at Bell Labs, then a short time at NEC 00:14:16.279 --> 00:14:19.279 Research Institute, Lacan makes a really big 00:14:19.279 --> 00:14:22.820 shift in 2003. He goes back to academia. He joins 00:14:22.820 --> 00:14:25.379 New York University, NYU, where he still is today. 00:14:25.720 --> 00:14:28.000 He's the Jacob T. Schwartz Chaired Professor 00:14:28.000 --> 00:14:30.340 of Computer Science and Neuroscience. And this 00:14:30.340 --> 00:14:32.620 feels like a very deliberate move. After proving 00:14:32.620 --> 00:14:35.600 the industrial viability of his ideas, he seems 00:14:35.600 --> 00:14:37.759 to be stepping back to focus on the next wave 00:14:37.759 --> 00:14:40.519 of fundamental research. Absolutely. It was a 00:14:40.519 --> 00:14:43.080 chance to dig deeper into the theory, to explore 00:14:43.080 --> 00:14:44.919 ideas that weren't quite ready for commercial 00:14:44.919 --> 00:14:47.480 application, and really importantly, to start 00:14:47.480 --> 00:14:49.539 building the institutions that would train the 00:14:49.539 --> 00:14:52.120 next generation of AI researchers. And the sources 00:14:52.120 --> 00:14:54.600 highlight a few key areas of his research at 00:14:54.600 --> 00:14:57.379 NYU. One of them is something called energy -based 00:14:57.379 --> 00:15:00.720 models. That sounds very abstract. It is, but 00:15:00.720 --> 00:15:02.840 it's a hugely important idea if you want to get 00:15:02.840 --> 00:15:05.940 to human -like intelligence. You see, most of 00:15:05.940 --> 00:15:08.899 the AI we use today, including the early CNNs, 00:15:08.899 --> 00:15:12.019 relies on supervised learning. You need massive 00:15:12.019 --> 00:15:14.899 data sets with labels. You need a million pictures 00:15:14.899 --> 00:15:17.840 of cats, each one labeled cat. Which is not how 00:15:17.840 --> 00:15:20.320 humans learn. Not at all. We learn mostly by 00:15:20.320 --> 00:15:22.720 just observing the world. Unsupervised learning. 00:15:23.059 --> 00:15:26.179 Energy -based models, or EBMs, are a framework 00:15:26.179 --> 00:15:28.059 for getting machines to do that. How do they 00:15:28.059 --> 00:15:31.059 work, conceptually? The basic idea is that the 00:15:31.059 --> 00:15:33.179 model assigns a score, which it calls energy, 00:15:33.440 --> 00:15:36.820 to any piece of data. A plausible, correct piece 00:15:36.820 --> 00:15:40.200 of data, like an image of a real cat, gets a 00:15:40.200 --> 00:15:43.330 low energy score. An impossible piece of data, 00:15:43.389 --> 00:15:46.049 like a garbled, nonsensical image, gets a very 00:15:46.049 --> 00:15:48.269 high energy score. Okay, so learning becomes 00:15:48.269 --> 00:15:50.850 about trying to find the low -energy configurations. 00:15:51.289 --> 00:15:53.570 Exactly. It's like a landscape with hills and 00:15:53.570 --> 00:15:56.370 valleys. The model learns to shape that landscape 00:15:56.370 --> 00:15:58.669 so that all the real -world data points settle 00:15:58.669 --> 00:16:00.929 into the deep valleys. It's a way for the machine 00:16:00.929 --> 00:16:03.129 to learn the underlying structure of reality, 00:16:03.450 --> 00:16:05.350 to figure out what's possible and what's not 00:16:05.350 --> 00:16:07.809 without needing a label for everything. That 00:16:07.809 --> 00:16:10.230 feels like a necessary step toward common sense. 00:16:10.590 --> 00:16:12.909 It is. It's the conceptual groundwork for the 00:16:12.909 --> 00:16:15.450 world models he's focused on today. It's about 00:16:15.450 --> 00:16:18.269 moving beyond just recognizing patterns and starting 00:16:18.269 --> 00:16:20.590 to build an actual understanding of how the world 00:16:20.590 --> 00:16:23.809 works. And alongside that theoretical work, he's 00:16:23.809 --> 00:16:26.129 also applying these ideas to the physical world, 00:16:26.250 --> 00:16:30.090 specifically to mobile robotics. The forces mention 00:16:30.090 --> 00:16:33.230 work on autonomous off -road driving. Yeah, and 00:16:33.230 --> 00:16:35.029 that's a great example. Driving on a perfectly 00:16:35.029 --> 00:16:37.809 marked highway is one thing. Driving through 00:16:37.809 --> 00:16:40.409 a forest or across a field where there are no 00:16:40.409 --> 00:16:43.190 lanes and the terrain is unpredictable, that's 00:16:43.190 --> 00:16:45.470 a much harder problem. It requires a much deeper 00:16:45.470 --> 00:16:48.230 level of visual understanding. It requires robustness. 00:16:48.269 --> 00:16:50.710 It shows he was always committed to taking these 00:16:50.710 --> 00:16:53.590 ideas out of the clean digital world and testing 00:16:53.590 --> 00:16:56.309 them in the messy, complicated physical world. 00:16:56.529 --> 00:16:57.990 And at the same time he's doing all this research, 00:16:58.110 --> 00:16:59.950 he's also becoming a major institution builder 00:16:59.950 --> 00:17:03.730 at NYU. In 2012, he becomes the founding director 00:17:03.730 --> 00:17:06.910 of the NYU Center for Data Science. A huge move. 00:17:07.130 --> 00:17:10.089 That helped formalize data science as its own 00:17:10.089 --> 00:17:12.670 serious academic discipline. It gave it a home. 00:17:12.829 --> 00:17:14.950 Before that, it was kind of scattered between 00:17:14.950 --> 00:17:17.769 computer science, statistics, and other departments. 00:17:17.970 --> 00:17:19.829 He helped put it at the center of the university. 00:17:20.150 --> 00:17:22.890 And that drive to build and shape the field went 00:17:22.890 --> 00:17:27.579 way beyond just NYU. In 2013, he and Yoshua Bengio 00:17:27.579 --> 00:17:30.099 co -found a new conference. The International 00:17:30.099 --> 00:17:32.779 Conference on Learning Representations, or ICLR. 00:17:32.980 --> 00:17:35.180 Which is now one of the top three most important 00:17:35.180 --> 00:17:37.460 AI conferences in the world. Easily. It's an 00:17:37.460 --> 00:17:39.720 absolute must -attend for anyone in the field. 00:17:39.839 --> 00:17:42.079 And it was born out of necessity. The field of 00:17:42.079 --> 00:17:44.799 deep learning was just exploding so fast that 00:17:44.799 --> 00:17:46.579 the older, broader machine learning conferences 00:17:46.579 --> 00:17:48.980 couldn't keep up. They needed their own dedicated 00:17:48.980 --> 00:17:51.380 venue. But ICLR was different. It wasn't just 00:17:51.380 --> 00:17:54.019 another conference. It had this radical new structure 00:17:54.019 --> 00:17:56.420 that looked... had been pushing for, an open 00:17:56.420 --> 00:17:59.720 review process. Yes, and this is so key to his 00:17:59.720 --> 00:18:02.519 philosophy. The traditional peer review process 00:18:02.519 --> 00:18:05.019 for scientific papers is slow and secretive. 00:18:05.059 --> 00:18:07.420 You submit a paper, it disappears for months, 00:18:07.480 --> 00:18:10.420 and you get anonymous reviews back. And LeCun 00:18:10.420 --> 00:18:12.980 argued that was holding the field back. He argued 00:18:12.980 --> 00:18:16.160 it was an obstacle to progress. With the ICLR 00:18:16.160 --> 00:18:19.039 model, Papers are published online first on a 00:18:19.039 --> 00:18:21.700 site like ArtFiev, and then the reviews and discussions 00:18:21.700 --> 00:18:24.180 happen out in the public. It's faster, it's more 00:18:24.180 --> 00:18:26.599 transparent, and it allows the entire community 00:18:26.599 --> 00:18:29.619 to participate in vetting new ideas. He was trying 00:18:29.619 --> 00:18:32.180 to build a scientific communication system that 00:18:32.180 --> 00:18:34.380 moved at the speed of the technology itself. 00:18:34.619 --> 00:18:37.250 Exactly. It's all about openness and acceleration. 00:18:37.630 --> 00:18:39.730 It's a theme you see again and again in his career. 00:18:39.890 --> 00:18:42.710 He also ran this legendary annual workshop in 00:18:42.710 --> 00:18:45.789 Snowbird, Utah for decades, just constantly building 00:18:45.789 --> 00:18:48.250 that community. Which all leads to this massive 00:18:48.250 --> 00:18:50.670 turning point for him and for the whole field. 00:18:50.950 --> 00:18:54.069 December 9th, 2013. A date that should be in 00:18:54.069 --> 00:18:56.650 the history books of AI. That's the day Jan LeCun 00:18:56.650 --> 00:18:58.990 became the first director of Meta AI Research, 00:18:59.190 --> 00:19:02.119 or FAIR. He joined what was then Facebook as 00:19:02.119 --> 00:19:04.460 their chief AI scientist. This was a seismic 00:19:04.460 --> 00:19:06.400 event. This was one of the godfathers of the 00:19:06.400 --> 00:19:08.599 field who had just built a major academic center, 00:19:08.740 --> 00:19:11.759 making a definitive move to big industry. And 00:19:11.759 --> 00:19:14.299 it signaled that the era of deep learning as 00:19:14.299 --> 00:19:17.279 a purely academic pursuit was over. This was 00:19:17.279 --> 00:19:19.700 the moment it moved into the hyperscale engine 00:19:19.700 --> 00:19:21.539 room of one of the biggest tech companies on 00:19:21.539 --> 00:19:23.779 the planet. So what was the crucial difference? 00:19:23.900 --> 00:19:26.220 What could he do at Meta that he just couldn't 00:19:26.220 --> 00:19:28.539 do at NYU, even with all his resources there? 00:19:28.839 --> 00:19:31.200 Three things that no university can ever provide 00:19:31.200 --> 00:19:34.700 at that level. Scale, hardware, and the feedback 00:19:34.700 --> 00:19:36.880 loop. Let's break those down. At Fifth Fire, 00:19:37.039 --> 00:19:39.119 he suddenly had access to data from billions 00:19:39.119 --> 00:19:41.880 of users flowing in real time. That's a scale 00:19:41.880 --> 00:19:43.700 of data that's just unimaginable in academia. 00:19:44.240 --> 00:19:47.319 Second, he had access to giant custom -built 00:19:47.319 --> 00:19:50.099 data centers full of GPUs designed for one purpose, 00:19:50.359 --> 00:19:53.359 training massive AI models. And third, the feedback 00:19:53.359 --> 00:19:56.299 loop. The ability to actually deploy a new model 00:19:56.299 --> 00:19:58.680 to millions of users overnight and see instantly 00:19:58.680 --> 00:20:01.140 how it performs in the real world. That cycle 00:20:01.140 --> 00:20:03.720 of idea, experiment, deployment, and feedback 00:20:03.720 --> 00:20:06.140 could happen in weeks at Meta, whereas in academia 00:20:06.140 --> 00:20:08.779 it might take years. So all of his foundational 00:20:08.779 --> 00:20:12.099 ideas, the CNNs, the energy -based models, they 00:20:12.099 --> 00:20:14.200 now had the perfect greenhouse to grow at an 00:20:14.200 --> 00:20:16.619 exponential rate. And they did. Everything from 00:20:16.619 --> 00:20:19.539 the news feed to ad targeting to content moderation 00:20:19.539 --> 00:20:23.079 to the language and vision models Meta uses today. 00:20:23.400 --> 00:20:25.839 It was all built on the foundation that Lacoon 00:20:25.839 --> 00:20:28.900 and the FAIR team established. It was the decade 00:20:28.900 --> 00:20:32.539 where deep learning became completely, undeniably 00:20:32.539 --> 00:20:35.099 industrialized. Which, of course, led to this 00:20:35.099 --> 00:20:38.279 incredible cascade of global recognition for 00:20:38.279 --> 00:20:41.059 him and his collaborators. We have to talk about 00:20:41.059 --> 00:20:44.630 the 2018 Turing Award. You do. I mean, the Turing 00:20:44.630 --> 00:20:47.089 is the Nobel Prize of computing. Getting that 00:20:47.089 --> 00:20:49.849 award was. It was the ultimate validation. It 00:20:49.849 --> 00:20:52.369 was the computing establishment officially declaring 00:20:52.369 --> 00:20:54.569 that the connectionists had won. He shared it, 00:20:54.630 --> 00:20:57.630 of course, with Bengio and Hinton. And the award 00:20:57.630 --> 00:21:00.009 citations specifically mentioned their conceptual 00:21:00.009 --> 00:21:02.250 and engineering breakthroughs. I love that they 00:21:02.250 --> 00:21:04.009 included engineering. It wasn't just for the 00:21:04.009 --> 00:21:06.569 ideas. It was for making the ideas work. Hinton 00:21:06.569 --> 00:21:08.589 for the core learning mechanisms, Bengio for 00:21:08.589 --> 00:21:10.990 sequences and generation, and Lacoon for the 00:21:10.990 --> 00:21:15.910 CNN architecture that finally unlocked. And it 00:21:15.910 --> 00:21:17.750 really was a celebration of their persistence, 00:21:18.009 --> 00:21:20.410 wasn't it? Sticking with these ideas through 00:21:20.410 --> 00:21:22.849 the AI winter when almost everyone else had given 00:21:22.849 --> 00:21:26.049 up. It was a total vindication. The Turing Award 00:21:26.049 --> 00:21:28.329 was the moment that deep learning officially 00:21:28.329 --> 00:21:30.849 moved from the fringe of computer science to 00:21:30.849 --> 00:21:33.250 its absolute center. And the awards just kept 00:21:33.250 --> 00:21:35.450 coming. And what's interesting is how they connect 00:21:35.450 --> 00:21:38.779 his work to the broader tech ecosystem. Take 00:21:38.779 --> 00:21:41.519 the 2022 Princess of Asturias award. He shared 00:21:41.519 --> 00:21:44.160 that one with Benjio and Hinton again, but also 00:21:44.160 --> 00:21:46.400 with Demis Hassabis from DeepMind. Which is a 00:21:46.400 --> 00:21:49.559 very significant pairing. It explicitly links 00:21:49.559 --> 00:21:52.220 the foundational academic work of the godfathers 00:21:52.220 --> 00:21:54.519 to the leader of one of the most successful industrial 00:21:54.519 --> 00:21:57.660 AI labs today. It draws a straight line from 00:21:57.660 --> 00:22:00.119 the theory to the application. And then the 2024 00:22:00.119 --> 00:22:03.019 VinFuture prize paints an even clearer picture. 00:22:03.380 --> 00:22:05.140 Because that one was shared with an even bigger 00:22:05.140 --> 00:22:08.200 group. the three godfathers, plus Fei -Fei Lai, 00:22:08.339 --> 00:22:11.660 and Jensen Huang, the CEO of NVIDIA. And that's 00:22:11.660 --> 00:22:14.819 the whole story right there. That award recognizes 00:22:14.819 --> 00:22:17.640 the complete system that was necessary for the 00:22:17.640 --> 00:22:20.700 deep learning revolution. You have the algorithms 00:22:20.700 --> 00:22:23.640 from Lacoon, Hinton, and Bengio. You have the 00:22:23.640 --> 00:22:26.019 massive dataset ImageNet from Fei -Fei Lai's 00:22:26.019 --> 00:22:29.319 team. And you have the hardware, the GPUs, from 00:22:29.319 --> 00:22:32.420 Jensen Huang's NVIDIA. Algorithm, data, and hardware. 00:22:32.859 --> 00:22:35.339 The holy trinity of modern AI. You can't have 00:22:35.339 --> 00:22:37.359 the revolution with just one or two of those. 00:22:37.440 --> 00:22:40.019 You need all three. And that prize was the ultimate 00:22:40.019 --> 00:22:42.599 recognition of that entire collaborative ecosystem. 00:22:42.940 --> 00:22:44.940 The recognition also came from the highest levels 00:22:44.940 --> 00:22:47.900 of government and engineering. The Queen Elizabeth 00:22:47.900 --> 00:22:51.119 Prize for Engineering in 2025. Again, framing 00:22:51.119 --> 00:22:53.759 his work not just as science, but as a world 00:22:53.759 --> 00:22:56.420 changing piece of engineering. And in 2023, he 00:22:56.420 --> 00:22:58.480 was named a Chevalier of the French Legion of 00:22:58.480 --> 00:23:00.430 Honor. which is the highest order of merit in 00:23:00.430 --> 00:23:03.130 France. It's his home country, recognizing that 00:23:03.130 --> 00:23:05.569 one of their own has fundamentally reshaped the 00:23:05.569 --> 00:23:08.309 global technological landscape. It's a beautiful 00:23:08.309 --> 00:23:10.289 moment that brings his French -American identity 00:23:10.289 --> 00:23:12.910 full circle. It all just paints this picture 00:23:12.910 --> 00:23:16.150 of a career that has had this profound, sustained, 00:23:16.450 --> 00:23:19.269 and universally recognized impact on the world. 00:23:19.410 --> 00:23:22.380 So after all that... After inventing the CNN, 00:23:22.700 --> 00:23:24.799 industrializing it at Bell Labs, building an 00:23:24.799 --> 00:23:28.500 academic powerhouse at NYU, guiding Meta's AI 00:23:28.500 --> 00:23:31.059 strategy for a decade, and winning every major 00:23:31.059 --> 00:23:34.299 award you can possibly win, what's left to do? 00:23:34.440 --> 00:23:36.579 That is the question. And according to our sources, 00:23:36.720 --> 00:23:39.059 we have an answer. The news from November 2025 00:23:39.059 --> 00:23:42.220 is that Yann LeCun is leaving Meta. After 10 00:23:42.220 --> 00:23:44.980 years as chief scientist, he's leaving to launch 00:23:44.980 --> 00:23:47.299 his own startup. And this isn't just him retiring 00:23:47.299 --> 00:23:49.539 or moving to a new company. This is a statement. 00:23:49.700 --> 00:23:52.059 He's pivoting to what he sees as the next great 00:23:52.059 --> 00:23:54.779 unsolved problem in AI. So what does this all 00:23:54.779 --> 00:23:57.349 mean? The focus of this new venture is very specific, 00:23:57.430 --> 00:23:59.390 and I think the wording is important. It's world 00:23:59.390 --> 00:24:01.450 model architectures and human -like artificial 00:24:01.450 --> 00:24:04.470 intelligence. That phrase, world models, is key. 00:24:04.609 --> 00:24:06.769 It signals a complete shift from the kind of 00:24:06.769 --> 00:24:09.109 AI that dominates today. How so? Think about 00:24:09.109 --> 00:24:12.009 what current AI, even the most advanced large 00:24:12.009 --> 00:24:15.289 language models, really does. At its core, it's 00:24:15.289 --> 00:24:17.589 a hyper -sophisticated pattern matching machine. 00:24:18.200 --> 00:24:21.000 It's incredibly good at recognizing what's in 00:24:21.000 --> 00:24:23.460 a picture or predicting the most statistically 00:24:23.460 --> 00:24:26.079 likely next word in a sentence. But it doesn't 00:24:26.079 --> 00:24:27.759 really understand what it's saying or seeing. 00:24:27.960 --> 00:24:31.200 Exactly. It lacks a model of how the world actually 00:24:31.200 --> 00:24:34.019 works. It doesn't grasp cause and effect or what 00:24:34.019 --> 00:24:36.460 you might call intuitive physics. It has no common 00:24:36.460 --> 00:24:39.220 sense. So give me an example. If you show a current 00:24:39.220 --> 00:24:42.349 AI. A video of a glass tipping over on a table. 00:24:42.450 --> 00:24:44.710 It could write a perfect caption. A glass is 00:24:44.710 --> 00:24:47.250 falling off a table. But a world model would 00:24:47.250 --> 00:24:49.549 do something more. It would anticipate. It would 00:24:49.549 --> 00:24:52.309 understand gravity and momentum and fragility. 00:24:52.589 --> 00:24:55.289 It would know, without having to be told, that 00:24:55.289 --> 00:24:57.730 the glass is going to hit the floor and shatter 00:24:57.730 --> 00:24:59.890 into a thousand pieces. That's about building 00:24:59.890 --> 00:25:02.750 an internal simulator of reality inside the AI. 00:25:03.200 --> 00:25:04.619 That's the perfect way to put it. It's moving 00:25:04.619 --> 00:25:06.660 from prediction of words to prediction of the 00:25:06.660 --> 00:25:09.420 future state of the world. It's what would allow 00:25:09.420 --> 00:25:12.619 an AI to plan, to reason about the consequences 00:25:12.619 --> 00:25:15.420 of actions, to handle situations it has never 00:25:15.420 --> 00:25:18.599 seen before. It's the leap from recognition to 00:25:18.599 --> 00:25:21.400 cognition. From system one, that fast, intuitive 00:25:21.400 --> 00:25:25.380 pattern matching. System two, the slower, more 00:25:25.380 --> 00:25:28.859 deliberate reasoning part of intelligence. That's 00:25:28.859 --> 00:25:30.920 the mountain he's setting out to climb now. And 00:25:30.920 --> 00:25:32.460 you can see his commitment to this direction. 00:25:32.490 --> 00:25:35.670 in his other ongoing work. He's still a co -director 00:25:35.670 --> 00:25:37.390 of the Learning and Machines and Brain program 00:25:37.390 --> 00:25:39.730 at CIFAR. Which is all about that intersection 00:25:39.730 --> 00:25:43.230 between neuroscience and AI, trying to learn 00:25:43.230 --> 00:25:46.369 from the ultimate world model, the human brain. 00:25:46.630 --> 00:25:48.890 And maybe most importantly, given his history, 00:25:49.049 --> 00:25:51.569 he's a scientific advisor to this French research 00:25:51.569 --> 00:25:55.190 group called QTAI. Yes, and QTAI is a very interesting 00:25:55.190 --> 00:25:57.890 organization. It's backed by some huge names 00:25:57.890 --> 00:26:02.940 in European tech and finance. Even Eric Schmidt. 00:26:03.200 --> 00:26:05.480 Right. And their entire mission, their founding 00:26:05.480 --> 00:26:07.700 principle, is that all of their research, all 00:26:07.700 --> 00:26:09.579 of their models will be completely open source. 00:26:09.839 --> 00:26:12.099 That feels like a direct challenge to the current 00:26:12.099 --> 00:26:15.140 trend in AI, where the most powerful models are 00:26:15.140 --> 00:26:17.880 kept proprietary and locked away inside a few 00:26:17.880 --> 00:26:20.339 giant tech companies. It's the ultimate expression 00:26:20.339 --> 00:26:23.900 of the philosophy he started with ICLR. He believes 00:26:23.900 --> 00:26:25.940 that this next generation of foundational AI 00:26:25.940 --> 00:26:28.940 is too important to be controlled by any one 00:26:28.940 --> 00:26:32.450 entity. His involvement with Kigatai shows a 00:26:32.450 --> 00:26:34.789 deep commitment to making sure that the tools 00:26:34.789 --> 00:26:37.569 to build human -like intelligence are available 00:26:37.569 --> 00:26:40.390 to everyone. So he's not just trying to invent 00:26:40.390 --> 00:26:43.170 the future of AI. He's trying to shape the culture 00:26:43.170 --> 00:26:45.289 of how it's built and shared. Absolutely. He's 00:26:45.289 --> 00:26:47.529 pushing the scientific frontier and the philosophical 00:26:47.529 --> 00:26:49.970 one at the same time. Okay, so to wrap this all 00:26:49.970 --> 00:26:52.630 up, to summarize this incredible journey, Yann 00:26:52.630 --> 00:26:55.369 LeCun's career is really the story of modern 00:26:55.369 --> 00:26:58.160 AI in miniature. It is. He starts in the mid 00:26:58.160 --> 00:27:00.599 -80s working on the core math of back propagation. 00:27:00.920 --> 00:27:03.339 Then at Bell Labs, he has his watershed moment 00:27:03.339 --> 00:27:06.240 creating the convolutional neural network. And 00:27:06.240 --> 00:27:08.640 he doesn't just create it. He proves its value 00:27:08.640 --> 00:27:10.720 in the real world with that system that read 00:27:10.720 --> 00:27:13.220 more than 10 % of all checks in the U .S., the 00:27:13.220 --> 00:27:16.359 first huge industrial success for deep learning. 00:27:16.619 --> 00:27:19.140 Then he moves to NYU, where he solidifies the 00:27:19.140 --> 00:27:21.640 academic field, helps create the Center for Data 00:27:21.640 --> 00:27:24.299 Science, and changes how research is shared with 00:27:24.299 --> 00:27:27.599 ICLR. And then the decade at Meta, taking all 00:27:27.599 --> 00:27:30.700 those ideas to planetary scale, and in the process, 00:27:30.900 --> 00:27:33.940 collecting every major award in science and engineering, 00:27:34.140 --> 00:27:37.039 from the Turing to the VinFuture Prize, which 00:27:37.039 --> 00:27:39.400 recognized that whole ecosystem of hardware, 00:27:39.480 --> 00:27:42.220 software, and data that he helped create. His 00:27:42.220 --> 00:27:44.539 career is the thread that connects those abstract 00:27:44.539 --> 00:27:47.440 connectionist ideas from the 80s to the powerful 00:27:47.440 --> 00:27:50.039 industrialized AI that shapes our world today. 00:27:50.380 --> 00:27:53.420 And now, the man who built so much of the present 00:27:53.420 --> 00:27:55.839 is leaving it all behind to focus purely on the 00:27:55.839 --> 00:27:58.200 future. The person who taught machines how to 00:27:58.200 --> 00:28:00.759 recognize things is now setting out to teach 00:28:00.759 --> 00:28:02.940 them how to understand things. That's it, exactly. 00:28:03.240 --> 00:28:05.140 And that brings us to the final thought we want 00:28:05.140 --> 00:28:07.700 to leave you with. Lacan's early career was all 00:28:07.700 --> 00:28:10.230 about solving the problem of perception. of recognition 00:28:10.230 --> 00:28:13.329 that the generation of ai he built the cnn's 00:28:13.329 --> 00:28:16.049 mastered the question of what what is in this 00:28:16.049 --> 00:28:18.609 image and now he's focused entirely on world 00:28:18.609 --> 00:28:21.130 models Right. So the question for the future 00:28:21.130 --> 00:28:23.670 is this. If that last generation of deep learning 00:28:23.670 --> 00:28:26.049 was about mastering the what, how profoundly 00:28:26.049 --> 00:28:28.430 will this next generation, the one focused on 00:28:28.430 --> 00:28:31.309 world models, master the why and the what if? 00:28:31.730 --> 00:28:34.609 If CNNs gave machines eyes, will world models 00:28:34.609 --> 00:28:36.750 give them imagination? That's the challenge that 00:28:36.750 --> 00:28:39.430 will define the next decade of AI, and Yann LeCun 00:28:39.430 --> 00:28:41.650 has put himself right at the heart of it all 00:28:41.650 --> 00:28:42.210 over again.