WEBVTT 00:00:00.000 --> 00:00:02.220 Welcome back to The Deep Dive, the show built 00:00:02.220 --> 00:00:05.519 entirely around your specific curiosity. We take 00:00:05.519 --> 00:00:07.900 your stack of sources, articles, and research 00:00:07.900 --> 00:00:10.980 and distill the absolute highest impact knowledge 00:00:10.980 --> 00:00:14.099 just for you. Today, we are diving into a career 00:00:14.099 --> 00:00:16.839 that, you know, it didn't just witness the birth 00:00:16.839 --> 00:00:19.859 of the modern AI revolution. No, it practically 00:00:19.859 --> 00:00:22.820 architected it. And at the same time, it completely 00:00:22.820 --> 00:00:25.370 reshaped the way... millions of people, maybe 00:00:25.370 --> 00:00:27.329 even you, first learned about this whole field. 00:00:27.550 --> 00:00:29.269 We are, of course, talking about Andrew Meng. 00:00:29.589 --> 00:00:32.070 When you really analyze his professional trajectory, 00:00:32.530 --> 00:00:34.929 what immediately stands out is just the sheer 00:00:34.929 --> 00:00:37.490 scale. The scale and the breadth of his influence. 00:00:37.590 --> 00:00:40.549 It's wild. It goes from academia to industry 00:00:40.549 --> 00:00:42.729 leadership and, you know, all the way to global 00:00:42.729 --> 00:00:45.189 education. Right. He's a British -American computer 00:00:45.189 --> 00:00:47.590 scientist, a really distinguished academic, and 00:00:47.590 --> 00:00:49.950 a major technology entrepreneur. And his focus 00:00:49.950 --> 00:00:52.549 has been, I mean, relentlessly centered on machine 00:00:52.549 --> 00:00:54.960 learning and artificial intelligence since, what, 00:00:55.000 --> 00:00:57.619 the late 1990s? Okay, so let's try to unpack 00:00:57.619 --> 00:01:01.259 this. It's a truly unique and a really high impact 00:01:01.259 --> 00:01:05.180 career. Our mission today is to understand his 00:01:05.180 --> 00:01:07.640 influence through what we're sort of defining 00:01:07.640 --> 00:01:10.459 as three foundational pillars of his work. That's 00:01:10.459 --> 00:01:12.140 a perfect way to structure it. The first pillar, 00:01:12.239 --> 00:01:16.439 I'd say, is his leadership role in AI innovation 00:01:16.439 --> 00:01:19.310 at the very highest levels of industry. You mean 00:01:19.310 --> 00:01:21.590 like bringing deep learning out of the research 00:01:21.590 --> 00:01:24.730 lab and into massive real world production systems? 00:01:24.950 --> 00:01:27.730 Exactly. Then the second pillar is his pioneering. 00:01:27.849 --> 00:01:30.829 And I think you have to say. revolutionary commitment 00:01:30.829 --> 00:01:34.109 to democratizing knowledge through accessible, 00:01:34.329 --> 00:01:37.430 high -quality online education. And the third 00:01:37.430 --> 00:01:39.450 pillar is where he is now, right? His current 00:01:39.450 --> 00:01:41.989 phase. Driving the future of AI, not just through 00:01:41.989 --> 00:01:44.909 development, but through focused investment and 00:01:44.909 --> 00:01:47.450 targeted enterprise adoption. Yes. Basically 00:01:47.450 --> 00:01:49.329 making sure that these powerful tools actually 00:01:49.329 --> 00:01:51.469 get used effectively by businesses all over the 00:01:51.469 --> 00:01:53.870 world. When you just look at the logos he's associated 00:01:53.870 --> 00:01:56.489 with, the impact is immediately concrete. Oh, 00:01:56.489 --> 00:01:58.109 absolutely. We're talking about the co -founder 00:01:58.109 --> 00:02:00.099 and former head of Google Brain. The project 00:02:00.099 --> 00:02:02.379 that famously took deep learning and just plugged 00:02:02.379 --> 00:02:05.060 it into Google's massive distributed computing 00:02:05.060 --> 00:02:07.299 infrastructure. Then he was the former chief 00:02:07.299 --> 00:02:09.360 scientist at Baidu, where he was literally tasked 00:02:09.360 --> 00:02:11.460 with building and leading an artificial intelligence 00:02:11.460 --> 00:02:15.560 group from a small team to several thousand researchers 00:02:15.560 --> 00:02:17.500 and engineers. And then you swing over to the 00:02:17.500 --> 00:02:19.659 education side, and he's the co -founder of both 00:02:19.659 --> 00:02:22.599 Coursera and deeplearning .ai. I mean, think 00:02:22.599 --> 00:02:25.139 about that. How often do you find someone who 00:02:25.139 --> 00:02:28.240 has led high -impact research at two of the world's 00:02:28.240 --> 00:02:30.819 biggest tech companies and co -founded two of 00:02:30.819 --> 00:02:33.719 the biggest online learning platforms? It's incredibly 00:02:33.719 --> 00:02:36.199 rare. That resume just puts him right at the 00:02:36.199 --> 00:02:38.800 nexus of bleeding -edge research, industrial 00:02:38.800 --> 00:02:41.759 scale, and global teaching. It definitely does. 00:02:41.919 --> 00:02:44.280 And the sources all point out that his influence 00:02:44.280 --> 00:02:47.060 hasn't faded one bit. He's been named one of 00:02:47.060 --> 00:02:49.300 the Time 100 most influential people not just 00:02:49.300 --> 00:02:52.539 once. but twice. Which is a powerful indicator 00:02:52.539 --> 00:02:55.419 of enduring relevance. For sure. He first appeared 00:02:55.419 --> 00:02:58.039 on the list in 2013 and then again a decade later 00:02:58.039 --> 00:03:01.680 on the 2023 Time 100 AI Most Influential People 00:03:01.680 --> 00:03:04.659 list. So that signals his consistent impact throughout 00:03:04.659 --> 00:03:08.259 the entire boom cycle of AI from like the early 00:03:08.259 --> 00:03:11.120 days of deep learning all the way to the modern 00:03:11.120 --> 00:03:15.319 generative AI era we're in now. So how does a 00:03:15.319 --> 00:03:17.400 person accumulate this much foundational influence 00:03:17.400 --> 00:03:20.340 across so many different domains? Let's trace 00:03:20.340 --> 00:03:22.539 his academic foundation because it really seems 00:03:22.539 --> 00:03:24.860 to have established his early multidisciplinary 00:03:24.860 --> 00:03:27.740 approach. Okay, so he was born Andrew Yan Tak 00:03:27.740 --> 00:03:31.680 Nang in London back in 1976. His parents were 00:03:31.680 --> 00:03:34.139 immigrants from Hong Kong. And he spent his very 00:03:34.139 --> 00:03:37.150 early years there. In Hong Kong, before the family 00:03:37.150 --> 00:03:39.770 moved to Singapore in 1984. Right. And the source 00:03:39.770 --> 00:03:42.030 material highlights that he had this early aptitude 00:03:42.030 --> 00:03:44.629 for technical subjects. He apparently learned 00:03:44.629 --> 00:03:46.550 the basics of programming from books at like 00:03:46.550 --> 00:03:49.069 six years old. Six? That's incredible. Yeah. 00:03:49.169 --> 00:03:51.129 He attended Raffles Institution in Singapore, 00:03:51.310 --> 00:03:53.210 which is a very prestigious school. And while 00:03:53.210 --> 00:03:55.270 he was still in high school, his focus on analytical 00:03:55.270 --> 00:03:58.490 ability, it culminated in him winning a silver 00:03:58.490 --> 00:04:00.650 medal at the International Mathematical Olympiad. 00:04:00.969 --> 00:04:03.490 And that's not just some local competition. That 00:04:03.490 --> 00:04:06.229 is a global marker of, you know, extraordinary 00:04:06.229 --> 00:04:08.729 analytical talent and problem solving skills 00:04:08.729 --> 00:04:11.550 long before his career even focused on computer 00:04:11.550 --> 00:04:14.069 science. Exactly. But here's where it gets really 00:04:14.069 --> 00:04:17.149 interesting for, you know, the future architect 00:04:17.149 --> 00:04:21.430 of scalable AI systems, his undergraduate degree. 00:04:21.879 --> 00:04:24.740 He went to Carnegie Mellon University CMU, which 00:04:24.740 --> 00:04:27.620 is a total powerhouse in computer science. And 00:04:27.620 --> 00:04:30.660 he graduated in 1997 with what we can only call 00:04:30.660 --> 00:04:33.620 the triple threat degree, a triple major in computer 00:04:33.620 --> 00:04:36.500 science, statistics and economics. Wow. OK, that 00:04:36.500 --> 00:04:39.660 combination is. It's highly illuminating. It 00:04:39.660 --> 00:04:42.060 feels almost predictive of his entire career. 00:04:42.240 --> 00:04:44.420 I think so, too. I mean, computer science gives 00:04:44.420 --> 00:04:46.360 you the ability to build the systems. Right. 00:04:46.459 --> 00:04:48.980 Statistics provides the foundational mathematical 00:04:48.980 --> 00:04:51.620 rigor you need for machine learning models. And 00:04:51.620 --> 00:04:53.980 economics provides the framework for understanding 00:04:53.980 --> 00:04:56.959 market efficiency, business application, and 00:04:56.959 --> 00:04:59.600 crucially, how to scale these systems to impact 00:04:59.600 --> 00:05:02.540 millions of users or enterprises. So he didn't 00:05:02.540 --> 00:05:04.160 just learn how to build. He learned where to 00:05:04.160 --> 00:05:07.060 apply it for the biggest possible impact. And 00:05:07.060 --> 00:05:09.639 he was already applying that knowledge on a massive 00:05:09.639 --> 00:05:12.800 scale, even before he got his Ph .D., while he 00:05:12.800 --> 00:05:16.339 was earning his master's at MIT in 1998. In electrical 00:05:16.339 --> 00:05:18.339 engineering and computer science, right? Right. 00:05:18.560 --> 00:05:21.819 He built something that was, well, really foundational 00:05:21.819 --> 00:05:24.360 to how modern researchers access information 00:05:24.360 --> 00:05:27.660 even today. He developed the first publicly available, 00:05:27.939 --> 00:05:30.620 automatically indexed web search engine specifically 00:05:30.620 --> 00:05:33.680 for research papers. So a search engine just 00:05:33.680 --> 00:05:36.139 for academic paper. Exactly. The key was the 00:05:36.139 --> 00:05:38.439 automatic indexing and the specialization in 00:05:38.439 --> 00:05:41.060 machine learning literature. The sources note 00:05:41.060 --> 00:05:43.680 this was a direct precursor to databases you 00:05:43.680 --> 00:05:47.279 might know, like SiteCRX or Research Index. So 00:05:47.279 --> 00:05:49.540 even back in the late 90s, he was already thinking 00:05:49.540 --> 00:05:51.939 about how to organize and leverage specialized 00:05:51.939 --> 00:05:55.180 knowledge on a huge searchable scale. He was 00:05:55.180 --> 00:05:57.199 anticipating the information overload problem 00:05:57.199 --> 00:06:00.139 long before Google was what it is today. It shows 00:06:00.139 --> 00:06:02.720 this consistent thread. You know, a desire to 00:06:02.720 --> 00:06:04.600 lower the barriers to knowledge, whether it's 00:06:04.600 --> 00:06:06.740 for academic papers or educational courses. And 00:06:06.740 --> 00:06:08.839 that consistent thread led him to his doctoral 00:06:08.839 --> 00:06:10.899 work and the beginning of what you could call 00:06:10.899 --> 00:06:14.439 the Stanford years. Yep. He got his Ph .D. in 00:06:14.439 --> 00:06:17.720 computer science from UC Berkeley in 2002, and 00:06:17.720 --> 00:06:19.600 he studied under the highly respected Michael 00:06:19.600 --> 00:06:22.560 I. Jordan. His thesis title alone sounds incredibly 00:06:22.560 --> 00:06:26.160 complex. Shaping and Policy Search in Reinforcement 00:06:26.160 --> 00:06:29.230 Learning. For listeners who follow the field 00:06:29.230 --> 00:06:31.689 but maybe don't specialize in that area, what 00:06:31.689 --> 00:06:33.449 was the practical significance of that work? 00:06:33.689 --> 00:06:36.589 Okay. So reinforcement learning at its core is 00:06:36.589 --> 00:06:39.389 about training an agent like a robot or an algorithm 00:06:39.389 --> 00:06:41.769 to make a sequence of good decisions in some 00:06:41.769 --> 00:06:44.329 environment. Usually by rewarding it for the 00:06:44.329 --> 00:06:46.750 right behaviors. Right. But the challenge is 00:06:46.750 --> 00:06:49.029 often the learning process is incredibly slow 00:06:49.029 --> 00:06:51.930 because the agent might only get a reward way, 00:06:51.990 --> 00:06:55.420 way in the future. Eng's work on shaping introduced 00:06:55.420 --> 00:06:58.959 a new approach for providing incremental rewards. 00:06:59.079 --> 00:07:01.160 Like a series of smaller hints along the way. 00:07:01.300 --> 00:07:03.680 Exactly. To guide the learning agent more efficiently 00:07:03.680 --> 00:07:06.160 toward that final goal. It's about accelerating 00:07:06.160 --> 00:07:08.660 the training process by making the reward landscape 00:07:08.660 --> 00:07:10.920 easier to navigate. And what about the policy 00:07:10.920 --> 00:07:13.680 search part? So policy search is the method the 00:07:13.680 --> 00:07:16.439 agent uses to find that optimal set of actions. 00:07:17.129 --> 00:07:20.149 His thesis combined these methods, and it provided 00:07:20.149 --> 00:07:22.389 significant improvements in the efficiency of 00:07:22.389 --> 00:07:25.329 training complex systems like robots. It's still 00:07:25.329 --> 00:07:27.790 highly cited today because it addressed a core 00:07:27.790 --> 00:07:30.189 problem in making reinforcement learning actually 00:07:30.189 --> 00:07:33.379 practical. After getting his PhD, he joined Stanford 00:07:33.379 --> 00:07:36.860 as an assistant professor in 2002. He moves up 00:07:36.860 --> 00:07:39.139 pretty quickly to associate professor and eventually 00:07:39.139 --> 00:07:41.500 becomes director of the Stanford AI Laboratory, 00:07:41.800 --> 00:07:44.399 or SAIL. And his machine learning course there, 00:07:44.500 --> 00:07:47.600 CS229, it was immensely popular. We're talking 00:07:47.600 --> 00:07:49.439 sometimes enrolling over a thousand students. 00:07:49.620 --> 00:07:51.660 Which is just unheard of for an advanced computer 00:07:51.660 --> 00:07:53.779 science course. His teaching ability clearly 00:07:53.779 --> 00:07:56.579 resonated. But the truly defining technical decision 00:07:56.579 --> 00:07:58.939 during his academic tenure was actually about 00:07:58.939 --> 00:08:01.750 hardware. It was a strategic move that fundamentally 00:08:01.750 --> 00:08:03.990 accelerated the entire field of deep learning. 00:08:04.189 --> 00:08:06.850 You're talking about GPUs? I am. Eng's group 00:08:06.850 --> 00:08:09.189 at Stanford around 2008 was one of the very first 00:08:09.189 --> 00:08:11.290 in the United States to strongly advocate for 00:08:11.290 --> 00:08:13.990 and commit to using GPUs graphical processing 00:08:13.990 --> 00:08:16.889 units in deep learning research. Today, that 00:08:16.889 --> 00:08:20.089 just sounds obvious. I mean, GPUs are the lifeblood 00:08:20.089 --> 00:08:23.269 of modern AI training. But back then, why was 00:08:23.269 --> 00:08:26.009 this considered controversial and risky? Weren't 00:08:26.009 --> 00:08:28.829 GPUs just for, you know, making video games look 00:08:28.829 --> 00:08:31.709 good? And that's the key distinction. Traditional 00:08:31.709 --> 00:08:34.809 CPU architectures, the brains of most computers, 00:08:35.070 --> 00:08:38.070 are designed for sequential processing. They 00:08:38.070 --> 00:08:41.070 execute a few complex tasks very, very quickly. 00:08:41.250 --> 00:08:44.429 One after the other. But deep learning models, 00:08:44.590 --> 00:08:47.549 they primarily rely on massive amounts of repetitive 00:08:47.549 --> 00:08:50.659 matrix multiplication. Thousands and thousands 00:08:50.659 --> 00:08:52.799 of simple calculations that can be performed 00:08:52.799 --> 00:08:55.299 all at the same time. And that's what GPUs are 00:08:55.299 --> 00:08:58.220 good at, parallel processing. Exactly. GPUs are 00:08:58.220 --> 00:09:00.519 built with thousands of small parallel processing 00:09:00.519 --> 00:09:03.759 cores, making them far more adept at this. The 00:09:03.759 --> 00:09:06.200 rationale, as the sources note, was pure scaling. 00:09:06.559 --> 00:09:09.220 They needed a more efficient computation infrastructure 00:09:09.220 --> 00:09:11.860 to speed up their model training by orders of 00:09:11.860 --> 00:09:13.940 magnitude. The massive scaling issues that came 00:09:13.940 --> 00:09:16.179 with big data and deep learning just required 00:09:16.179 --> 00:09:18.809 a different kind of processing power. Precisely. 00:09:18.830 --> 00:09:22.090 And the risk wasn't just the expense or the novelty. 00:09:22.370 --> 00:09:24.570 General -purpose computing platforms built on 00:09:24.570 --> 00:09:27.450 GPUs were really difficult to program effectively 00:09:27.450 --> 00:09:30.789 for research outside of graphics. It required 00:09:30.789 --> 00:09:33.269 a huge intellectual investment from Enc's team 00:09:33.269 --> 00:09:36.169 to master this new architectural approach. But 00:09:36.169 --> 00:09:38.809 it paid off. It paid off spectacularly. It allowed 00:09:38.809 --> 00:09:41.389 them to tackle computational problems that were 00:09:41.389 --> 00:09:43.889 simply impossible for their peers using traditional 00:09:43.889 --> 00:09:47.600 CPU clusters. And after Eng's successful implementation, 00:09:48.120 --> 00:09:50.580 the GPU quickly became the industry standard. 00:09:50.799 --> 00:09:53.340 It proved he had foresight not just in software 00:09:53.340 --> 00:09:55.679 algorithms, but in the critical hardware infrastructure 00:09:55.679 --> 00:09:58.159 underneath. So he made a foundational hardware 00:09:58.159 --> 00:10:01.100 bet that just dramatically accelerated the entire 00:10:01.100 --> 00:10:04.470 field. But beyond that... His group at Stanford 00:10:04.470 --> 00:10:07.529 produced some incredible, tangible research projects 00:10:07.529 --> 00:10:10.070 that left these open source legacies. They absolutely 00:10:10.070 --> 00:10:12.309 did. There are three projects that really define 00:10:12.309 --> 00:10:14.590 his impact in pure robotics and machine learning 00:10:14.590 --> 00:10:16.230 research from that time. Okay, what's the first 00:10:16.230 --> 00:10:19.070 one? First, the Stanford Autonomous Helicopter 00:10:19.070 --> 00:10:22.200 Project. They successfully developed one of the 00:10:22.200 --> 00:10:25.019 world's most capable model -based autonomous 00:10:25.019 --> 00:10:27.399 helicopters. And this wasn't about just remote 00:10:27.399 --> 00:10:29.820 controlling it. No, this was complex machine 00:10:29.820 --> 00:10:31.879 learning algorithms teaching the system to fly 00:10:31.879 --> 00:10:35.039 precise acrobatic maneuvers all by itself. Correct. 00:10:35.100 --> 00:10:37.559 It was classic, high visibility, cutting edge 00:10:37.559 --> 00:10:39.899 robotics that really captured the public imagination. 00:10:40.279 --> 00:10:43.399 Okay. What was the second project? Second, and 00:10:43.399 --> 00:10:45.360 maybe even more influential for the industry, 00:10:45.480 --> 00:10:49.419 was the STAIR project. STAIR stood for Stanford 00:10:49.419 --> 00:10:52.240 Artificial Intelligence Robot. An ambitious name. 00:10:52.419 --> 00:10:54.720 Very. It was a long -term effort to build an 00:10:54.720 --> 00:10:57.559 AI robot that Nang envisioned could one day be 00:10:57.559 --> 00:11:00.059 in every home, performing generalized tasks. 00:11:00.539 --> 00:11:02.679 Now, while the robot itself was a vehicle for 00:11:02.679 --> 00:11:05.559 research, the major lasting outcome was the creation 00:11:05.559 --> 00:11:09.440 of the robot operating system. And ROS is now 00:11:09.440 --> 00:11:12.379 the standard open source software robotics platform 00:11:12.379 --> 00:11:14.980 used all over the world. That's another key moment 00:11:14.980 --> 00:11:17.159 of democratizing technology, right? Moving it 00:11:17.159 --> 00:11:19.100 out of the Stanford lab and into the hands of 00:11:19.100 --> 00:11:21.639 researchers and hobbyists everywhere. It standardized 00:11:21.639 --> 00:11:23.940 how robots communicate, how their components 00:11:23.940 --> 00:11:26.559 interact, and it significantly lowered the barrier 00:11:26.559 --> 00:11:29.059 to entry for robotics research and commercial 00:11:29.059 --> 00:11:31.759 deployment. So we have the helicopter, we have 00:11:31.759 --> 00:11:34.320 ROS. What's the third one? The third one is a 00:11:34.320 --> 00:11:36.820 bit more on the theoretical side. from his graduate 00:11:36.820 --> 00:11:39.059 school years at Berkeley. He was involved in 00:11:39.059 --> 00:11:42.919 a landmark publication introducing latent derelict 00:11:42.919 --> 00:11:46.539 allocation, or LDA. LDA. I've heard of that. 00:11:46.600 --> 00:11:49.320 It's a topic modeling algorithm. Yes, a foundational 00:11:49.320 --> 00:11:52.720 one. In simple terms, it's a technique used in 00:11:52.720 --> 00:11:55.120 natural language processing that lets a computer 00:11:55.120 --> 00:11:57.379 automatically discover the underlying themes 00:11:57.379 --> 00:12:00.240 or topics that are present in a huge collection 00:12:00.240 --> 00:12:02.840 of documents. So you could feed it a million 00:12:02.840 --> 00:12:04.820 news articles, and it could tell you that there 00:12:04.820 --> 00:12:07.740 are, say, 20 distinct topics being discussed, 00:12:07.799 --> 00:12:11.039 like sports, politics, finance, without you ever 00:12:11.039 --> 00:12:13.899 telling it what those topics are. Exactly. And 00:12:13.899 --> 00:12:16.000 that algorithm remains highly significant in 00:12:16.000 --> 00:12:18.679 how we process large texts, organize information, 00:12:19.179 --> 00:12:22.580 and structure big data analysis today. This early 00:12:22.580 --> 00:12:25.039 work just shows N's influence spanning the full 00:12:25.039 --> 00:12:27.879 spectrum. Practical robotics with the helicopter 00:12:27.879 --> 00:12:30.820 and ROS, technical infrastructure with the GPU 00:12:30.820 --> 00:12:33.500 -BET, and core machine learning theory with LDA. 00:12:33.929 --> 00:12:36.309 The natural progression for someone who figures 00:12:36.309 --> 00:12:38.789 out how to make algorithms learn faster and build 00:12:38.789 --> 00:12:41.750 smarter systems is, well, to leave the lab and 00:12:41.750 --> 00:12:43.730 scale that knowledge up to the massive industrial 00:12:43.730 --> 00:12:46.490 level. And that's precisely what Eng did, transitioning 00:12:46.490 --> 00:12:48.929 first to Google and then to Baidu. The first 00:12:48.929 --> 00:12:51.909 big move was to Google Brain, which he co -founded 00:12:51.909 --> 00:12:54.950 and directed from 2011 to 2012. He was working 00:12:54.950 --> 00:12:57.009 alongside other heavy hitters like Jeff Dean, 00:12:57.230 --> 00:13:00.419 Greg Corrado, and Rajat Manga. Right. The project's 00:13:00.419 --> 00:13:03.159 goal was essentially to rapidly apply deep learning 00:13:03.159 --> 00:13:05.639 to Google's massive distributed computing infrastructure, 00:13:05.899 --> 00:13:08.679 to turn pure research into core product functionality 00:13:08.679 --> 00:13:11.419 as fast as possible. And this is where the famous, 00:13:11.659 --> 00:13:14.580 and I think it's fair to say viral cat experiment 00:13:14.580 --> 00:13:16.840 happened. It's almost a pop culture reference 00:13:16.840 --> 00:13:19.559 now. But what was the technical significance 00:13:19.559 --> 00:13:22.399 that made it such a huge leap for AI? It was 00:13:22.399 --> 00:13:25.179 a massive proof point for something called unsupervised 00:13:25.179 --> 00:13:27.700 feature learning. at a scale that had never been 00:13:27.700 --> 00:13:30.259 attempted before. They used a neural network 00:13:30.259 --> 00:13:34.679 trained on a staggering 16 ,000 CPU cores. 16 00:13:34.679 --> 00:13:37.299 ,000. That in itself is an engineering feat. 00:13:37.500 --> 00:13:40.120 It is. But the crucial technical insight was 00:13:40.120 --> 00:13:42.899 that the network, using these deep learning algorithms, 00:13:43.159 --> 00:13:46.429 learned to recognize high -level concepts. specifically 00:13:46.429 --> 00:13:49.649 cats and human faces, just by observing 10 million 00:13:49.649 --> 00:13:52.690 unlabeled random still frames pulled from YouTube 00:13:52.690 --> 00:13:55.049 videos. So just to be clear, no one ever told 00:13:55.049 --> 00:13:57.610 the system this is a cat. It figured out the 00:13:57.610 --> 00:14:00.129 defining features of a cat all on its own. Correct. 00:14:00.210 --> 00:14:02.549 That is the essence of unsupervised feature learning. 00:14:02.690 --> 00:14:04.990 The system essentially self -organized all that 00:14:04.990 --> 00:14:07.950 raw pixel data into abstract features. It demonstrated 00:14:07.950 --> 00:14:10.309 the power of deep learning to extract meaning 00:14:10.309 --> 00:14:13.059 from massive unstructured data sets. And that 00:14:13.059 --> 00:14:15.240 showed that deep learning could potentially bypass 00:14:15.240 --> 00:14:18.879 the really arduous, expensive task of human labeling. 00:14:19.120 --> 00:14:21.259 Which had historically been the biggest bottleneck 00:14:21.259 --> 00:14:23.820 for large -scale AI deployment. This had immediate 00:14:23.820 --> 00:14:26.539 implications for Google's product lines. Absolutely. 00:14:26.700 --> 00:14:29.100 The technology developed by Google Brain was 00:14:29.100 --> 00:14:31.220 quickly integrated into Google's infrastructure 00:14:31.220 --> 00:14:34.799 and specifically into the Android operating system's 00:14:34.799 --> 00:14:37.620 speech recognition system. So the ability to 00:14:37.620 --> 00:14:40.340 recognize complex patterns in unstructured video 00:14:40.340 --> 00:14:43.509 data. like cat videos. Translated directly into 00:14:43.509 --> 00:14:46.389 the ability to recognize complex acoustic patterns 00:14:46.389 --> 00:14:49.990 in human speech. Exactly. It demonstrated a seamless 00:14:49.990 --> 00:14:53.269 transition from pure novel research to consumer 00:14:53.269 --> 00:14:56.029 -facing utility that millions of people started 00:14:56.029 --> 00:14:58.149 using every day on their phones. After proving 00:14:58.149 --> 00:15:00.169 the power of deep learning at Google, N made 00:15:00.169 --> 00:15:03.070 another significant shift. He moved east to join 00:15:03.070 --> 00:15:05.669 Baidu in China as chief scientist, a role he 00:15:05.669 --> 00:15:09.330 held from 2014 to 2017. This was a hugely strategic 00:15:09.330 --> 00:15:12.049 move. He recognized that China offered this unique 00:15:12.049 --> 00:15:14.769 combination of vast data sets, strong government 00:15:14.769 --> 00:15:17.789 support for AI research and a massive underserved 00:15:17.789 --> 00:15:21.190 market. So his primary task was to lead AI research 00:15:21.190 --> 00:15:23.289 and build the company's artificial intelligence 00:15:23.289 --> 00:15:25.929 group from the ground up. And he did. It quickly 00:15:25.929 --> 00:15:28.450 grew into a multidisciplinary team of several 00:15:28.450 --> 00:15:31.059 thousand people. This wasn't just a research 00:15:31.059 --> 00:15:33.740 role. It was about building a massive, dedicated 00:15:33.740 --> 00:15:36.919 AI corporation inside of Baidu. What were the 00:15:36.919 --> 00:15:39.179 key strategic developments that came out of his 00:15:39.179 --> 00:15:42.480 time leading Baidu's AI efforts? There were several, 00:15:42.539 --> 00:15:45.639 and they showcased an intent to apply AI across 00:15:45.639 --> 00:15:48.179 a bunch of diverse sectors, much like at Google. 00:15:48.700 --> 00:15:51.320 He established core research teams focusing on 00:15:51.320 --> 00:15:53.950 things like facial recognition. He also developed 00:15:53.950 --> 00:15:56.710 Melody, which was an AI chatbot designed for 00:15:56.710 --> 00:15:58.950 healthcare to help physicians collect patient 00:15:58.950 --> 00:16:01.750 history. Interesting. And critically, his teams 00:16:01.750 --> 00:16:04.389 developed the AI platform called Duero S. Tell 00:16:04.389 --> 00:16:06.110 us more about that. Why was it so important? 00:16:06.350 --> 00:16:09.250 Duero S was Baidu's conversational AI system. 00:16:09.490 --> 00:16:12.190 It was their strategic answer to Amazon's Alexa 00:16:12.190 --> 00:16:14.929 or Google Assistant. It was designed to be integrated 00:16:14.929 --> 00:16:17.490 into smart speakers, mobile devices, and eventually 00:16:17.490 --> 00:16:19.960 autonomous vehicles. And the sources note that 00:16:19.960 --> 00:16:21.799 the speed and the scope of these developments 00:16:21.799 --> 00:16:24.740 under Eng actually positioned Baidu ahead of 00:16:24.740 --> 00:16:27.240 Google in the public discourse around AI during 00:16:27.240 --> 00:16:29.379 that specific period. They did, particularly 00:16:29.379 --> 00:16:31.700 in crucial emerging markets and in conversational 00:16:31.700 --> 00:16:34.259 AI. It was a powerful statement. They were seen 00:16:34.259 --> 00:16:37.919 for a time as outpacing Google in innovation. 00:16:38.600 --> 00:16:41.740 And then resigned in March 2017, which marked 00:16:41.740 --> 00:16:44.059 the beginning of yet another major shift in his 00:16:44.059 --> 00:16:46.559 career. He pivoted back toward entrepreneurship 00:16:46.559 --> 00:16:50.200 and maybe most importantly, education. It's almost 00:16:50.200 --> 00:16:52.600 as if once he proved he could build AI at scale, 00:16:52.820 --> 00:16:55.519 his next goal was to scale the knowledge itself. 00:16:55.820 --> 00:16:57.899 That synthesis feels like the critical insight 00:16:57.899 --> 00:17:00.879 of his career. He saw AI not just as a technology 00:17:00.879 --> 00:17:03.360 to be, you know, monopolized by a few giants. 00:17:03.379 --> 00:17:06.539 But as a universal tool. And to ensure its universal 00:17:06.539 --> 00:17:09.079 application, the world needed universal education. 00:17:09.339 --> 00:17:11.559 That pivot to education is, you could argue, 00:17:11.720 --> 00:17:14.700 just as impactful as his contributions to the 00:17:14.700 --> 00:17:17.920 cat experiment or the GPU decision. He recognized 00:17:17.920 --> 00:17:19.980 really early on that the knowledge he possessed 00:17:19.980 --> 00:17:22.839 needed to be shared on a massive available scale. 00:17:23.119 --> 00:17:25.680 Yes, and the demand was already clearly signal. 00:17:26.569 --> 00:17:28.490 As we mentioned earlier, his machine learning 00:17:28.490 --> 00:17:31.970 course at Stanford CS229 was the most popular 00:17:31.970 --> 00:17:34.710 on campus, sometimes with over 1 ,000 students. 00:17:35.170 --> 00:17:37.950 That appetite for high -level technical content 00:17:37.950 --> 00:17:40.750 was the clear signal that set the stage for the 00:17:40.750 --> 00:17:44.710 massive open online course, or MOOC, movement. 00:17:44.970 --> 00:17:47.210 And the roots of Coursera actually trace back 00:17:47.210 --> 00:17:50.049 even further than most people realize. They go 00:17:50.049 --> 00:17:53.289 back to 2008 when Neng started the Stanford Engineering 00:17:53.289 --> 00:17:56.869 Everywhere program. What was the goal of that 00:17:56.869 --> 00:17:59.589 program? The explicit goal was to publish a number 00:17:59.589 --> 00:18:02.329 of Stanford courses online for free, including 00:18:02.329 --> 00:18:04.829 his really popular machine learning class. But 00:18:04.829 --> 00:18:07.250 Eng recognized the limitations of just, you know. 00:18:07.630 --> 00:18:09.630 posting lecture videos. He wanted to provide 00:18:09.630 --> 00:18:11.670 a much more complete course experience. Not just 00:18:11.670 --> 00:18:13.930 lectures, but structured materials, practice 00:18:13.930 --> 00:18:16.069 problems, and solutions. Right. And that distinction, 00:18:16.309 --> 00:18:18.950 aiming for a complete, structured course experience, 00:18:19.210 --> 00:18:21.569 not just passive video viewing, was critical. 00:18:21.809 --> 00:18:24.589 He was inspired by existing platforms, but wanted 00:18:24.589 --> 00:18:26.930 to synthesize the best elements of teaching and 00:18:26.930 --> 00:18:29.730 technology. The sources mention a few key inspirations. 00:18:29.890 --> 00:18:32.859 They do. He credited Sal Khan of Khan Academy 00:18:32.859 --> 00:18:36.299 as a huge inspiration for the accessibility and 00:18:36.299 --> 00:18:38.900 the way topics were broken down. He also looked 00:18:38.900 --> 00:18:41.019 very closely at the structured learning path 00:18:41.019 --> 00:18:43.660 of Lynda .com, which is now LinkedIn Learning. 00:18:43.779 --> 00:18:46.579 For professional development. Exactly. And critically, 00:18:46.759 --> 00:18:49.079 he looked at the collaborative, self -correcting 00:18:49.079 --> 00:18:51.480 design of the community forums on Stack Overflow 00:18:51.480 --> 00:18:54.240 for managing questions and peer learning at a 00:18:54.240 --> 00:18:57.059 massive scale. It's so interesting that he blended 00:18:57.059 --> 00:18:59.400 professional training models with open educational 00:18:59.400 --> 00:19:02.119 resources and community support. He was doing 00:19:02.119 --> 00:19:04.619 rapid iteration and testing on the teaching approach, 00:19:04.839 --> 00:19:07.460 not just the technical platform. Absolutely. 00:19:07.680 --> 00:19:09.380 The sources note that he even tested some of 00:19:09.380 --> 00:19:11.460 the original design practices for recording lessons 00:19:11.460 --> 00:19:13.640 and online interactivity with a local high school. 00:19:13.799 --> 00:19:15.819 He was basically A -B testing his educational 00:19:15.819 --> 00:19:18.000 delivery before committing to the massive online 00:19:18.000 --> 00:19:20.250 format. He was meticulous about it. And this 00:19:20.250 --> 00:19:23.430 rigor led to the pivotal trial run in late 2011. 00:19:23.630 --> 00:19:26.349 Yes. In October 2011, the applied version of 00:19:26.349 --> 00:19:30.289 a Stanford class, CS229A, was hosted on a specialized 00:19:30.289 --> 00:19:34.150 website, ml -class .org. The response was immediate 00:19:34.150 --> 00:19:37.700 and just overwhelming. Over 100 ,000 students 00:19:37.700 --> 00:19:40.220 registered for that first edition. And crucially, 00:19:40.259 --> 00:19:42.400 this wasn't just passive learning. The course 00:19:42.400 --> 00:19:44.759 included quizzes and graded programming assignments, 00:19:45.119 --> 00:19:48.099 making it one of the first truly successful massive 00:19:48.099 --> 00:19:51.539 open online courses that demanded active participation. 00:19:51.960 --> 00:19:54.180 And this launch didn't happen in a vacuum. It 00:19:54.180 --> 00:19:56.660 was part of a larger Stanford movement that really 00:19:56.660 --> 00:20:00.150 kicked off. The modern MOC era. Correct. That 00:20:00.150 --> 00:20:03.049 same year saw the launch of three highly visible 00:20:03.049 --> 00:20:06.250 Stanford MOCs in total. There was Eng's machine 00:20:06.250 --> 00:20:08.829 learning course, a databases course by Jennifer 00:20:08.829 --> 00:20:11.650 Windham, and an AI course by Sebastian Thrun 00:20:11.650 --> 00:20:14.430 and Peter Norvig. And Thrun's course, that led 00:20:14.430 --> 00:20:17.549 to the genesis of Udacity. Right. The core themes 00:20:17.549 --> 00:20:20.230 of these modern MOCs were scale and availability. 00:20:20.750 --> 00:20:23.089 The goal was to take Stanford caliber education 00:20:23.089 --> 00:20:25.509 out of the lecture hall and put it on the global 00:20:25.509 --> 00:20:28.289 stage. And across those initial three 10 -week 00:20:28.289 --> 00:20:30.650 courses, they awarded over 40 ,000 statements 00:20:30.650 --> 00:20:33.089 of accomplishment, proving that massive scale 00:20:33.089 --> 00:20:35.490 could coexist with measurable learning outcomes. 00:20:35.750 --> 00:20:38.390 That momentum led directly to the formal launch 00:20:38.390 --> 00:20:40.609 of the platform that defined the movement. In 00:20:40.609 --> 00:20:43.910 2012, Eng formally co -founded Coursera with 00:20:43.910 --> 00:20:46.329 fellow Stanford computer scientist Daphne Collar. 00:20:46.490 --> 00:20:49.769 Eng initially served as CEO. And that platform 00:20:49.769 --> 00:20:51.990 took the scale and availability themes global. 00:20:52.440 --> 00:20:54.359 partnering with universities all over the world. 00:20:54.460 --> 00:20:57.140 It quickly became one of the leading MOOC sites 00:20:57.140 --> 00:20:59.779 in the world. It's a direct product of Eng's 00:20:59.779 --> 00:21:02.680 triple background. The technical ability to build 00:21:02.680 --> 00:21:05.079 the platform, the statistical insight to track 00:21:05.079 --> 00:21:07.579 learning, and the economic foresight to see the 00:21:07.579 --> 00:21:10.059 global market for knowledge. And even as the 00:21:10.059 --> 00:21:12.420 platform grew and added thousands of courses, 00:21:12.779 --> 00:21:15.339 Eng's original material maintained an incredible 00:21:15.339 --> 00:21:18.720 level of popularity. Absolutely dominant. As 00:21:18.720 --> 00:21:21.740 of 2019 and 2020, years after the platform was 00:21:21.740 --> 00:21:23.880 established, three of the most popular courses 00:21:23.880 --> 00:21:26.359 on Coursera were still his. Machine Learning 00:21:26.359 --> 00:21:28.660 was number one, Neural Networks and Deep Learning 00:21:28.660 --> 00:21:31.640 was number two, and AI for Everyone was number 00:21:31.640 --> 00:21:34.380 five. So his material remains the essential entry 00:21:34.380 --> 00:21:36.759 point for millions of new learners. And he didn't 00:21:36.759 --> 00:21:40.019 stop with Coursera. After leaving Baidu in 2017, 00:21:40.420 --> 00:21:43.079 he doubled down on his educational mission with 00:21:43.079 --> 00:21:46.039 deeplearning .ai. Continuing that mission, focusing 00:21:46.039 --> 00:21:48.599 specifically on specialized AI and deep learning 00:21:48.599 --> 00:21:51.319 curricula. And when you tally the total impact, 00:21:51.599 --> 00:21:55.680 the numbers are just staggering. By 2023, Eng 00:21:55.680 --> 00:21:57.640 had taught an estimated 8 million individuals 00:21:57.640 --> 00:22:00.660 globally. through his courses on both platforms. 00:22:00.940 --> 00:22:03.680 Eight million. That's not just education. It's 00:22:03.680 --> 00:22:06.519 a massive self -generated workforce infusion 00:22:06.519 --> 00:22:09.740 of practical AI knowledge into the global economy. 00:22:10.039 --> 00:22:11.640 And I think the most strategically interesting 00:22:11.640 --> 00:22:15.119 course he launched after Baidu was in 2019. It's 00:22:15.119 --> 00:22:17.740 called AI for Everyone. And this one was different. 00:22:17.880 --> 00:22:20.119 It was specifically for non -technical learners 00:22:20.119 --> 00:22:22.660 bypassing the math and the code. Why do that? 00:22:22.839 --> 00:22:25.200 Because he realized that the ultimate bottleneck 00:22:25.200 --> 00:22:28.160 for AI adoption is no longer the technology itself. 00:22:28.670 --> 00:22:30.769 It's the gap between the technology and the business 00:22:30.769 --> 00:22:33.809 strategy. Exactly. AI for Everyone was designed 00:22:33.809 --> 00:22:36.150 to help C -suite leaders, managers, and policymakers 00:22:36.150 --> 00:22:39.509 grasp AI's impact on society, understand the 00:22:39.509 --> 00:22:41.670 business benefits and costs, and know how to 00:22:41.670 --> 00:22:44.329 strategically navigate the revolution. It was 00:22:44.329 --> 00:22:47.150 about creating fluent stakeholders who can properly 00:22:47.150 --> 00:22:50.170 scope and fund AI projects, not just fluent coders. 00:22:50.309 --> 00:22:53.250 It was an attempt to scale AI fluency vertically 00:22:53.250 --> 00:22:57.099 throughout the entire organization. Okay, let's 00:22:57.099 --> 00:22:59.779 transition now to Eng's current phase, which 00:22:59.779 --> 00:23:01.880 really combines the technical expertise of Google 00:23:01.880 --> 00:23:04.500 Brain with the economic and scaling foresight 00:23:04.500 --> 00:23:06.920 of his undergraduate degree. He's now operating 00:23:06.920 --> 00:23:09.039 as an entrepreneur, a venture capitalist, and 00:23:09.039 --> 00:23:11.630 an applied AI developer. This stage represents 00:23:11.630 --> 00:23:14.829 a crucial shift. He's moving from educating the 00:23:14.829 --> 00:23:17.130 workforce to funding and building the companies 00:23:17.130 --> 00:23:19.730 that will employ that workforce. In January 2018, 00:23:19.950 --> 00:23:22.609 he launched the AI Fund. Right. Raising a significant 00:23:22.609 --> 00:23:26.390 $175 million in committed capital specifically 00:23:26.390 --> 00:23:29.910 to invest in and, more accurately, to create 00:23:29.910 --> 00:23:32.710 new AI startups. So is it a traditional VC model 00:23:32.710 --> 00:23:35.230 or more of an incubator? It operates closer to 00:23:35.230 --> 00:23:37.789 an AI venture studio model. The idea is that 00:23:37.789 --> 00:23:40.299 they don't just wait for external pitches. identify 00:23:40.299 --> 00:23:43.119 high -impact problems, and then internally incubate 00:23:43.119 --> 00:23:45.599 companies to solve them, leveraging Eng's vast 00:23:45.599 --> 00:23:47.920 network and deep learning expertise from the 00:23:47.920 --> 00:23:50.460 very beginning. Which drastically minimizes the 00:23:50.460 --> 00:23:52.980 early execution risk. It makes the fund a powerful 00:23:52.980 --> 00:23:55.700 engine for applied AI innovation. But he wasn't 00:23:55.700 --> 00:23:58.680 just funding. He was also directly building through 00:23:58.680 --> 00:24:01.539 his primary venture, Landing AI. And Landing 00:24:01.539 --> 00:24:04.140 AI has a highly specific industrial mission. 00:24:04.380 --> 00:24:07.339 Yes. Its goal is to provide AI -powered software 00:24:07.339 --> 00:24:10.720 as a service products focused on democratizing 00:24:10.720 --> 00:24:13.880 AI technology and lowering the barrier for entrance 00:24:13.880 --> 00:24:16.730 to businesses and developers. That echoes his 00:24:16.730 --> 00:24:19.549 longstanding commitment to accessibility. But 00:24:19.549 --> 00:24:21.650 this time it's applied specifically to complex 00:24:21.650 --> 00:24:23.589 industrial sectors that, you know, traditionally 00:24:23.589 --> 00:24:26.089 lag in digital transformation. And they chose 00:24:26.089 --> 00:24:28.349 industrial manufacturing as their initial focus, 00:24:28.549 --> 00:24:32.329 specifically complex computer vision for quality 00:24:32.329 --> 00:24:34.710 control. So think about factory floors where 00:24:34.710 --> 00:24:37.250 tiny defects are really hard for humans to catch 00:24:37.250 --> 00:24:39.490 consistently over an eight hour shift. Right. 00:24:39.650 --> 00:24:42.950 Landing AI uses its platform landing lens to 00:24:42.950 --> 00:24:45.500 allow manufacturers to quickly deploy. AI models 00:24:45.500 --> 00:24:48.460 to spot these flaws. This focus is practical, 00:24:48.700 --> 00:24:51.579 it's measurable, and it avoids the often -hyped 00:24:51.579 --> 00:24:54.359 abstract AI theories. And he quickly secured 00:24:54.359 --> 00:24:57.240 major funding to execute that mission. Precisely. 00:24:57.240 --> 00:25:00.740 In November 2021, Landing AI secured a massive 00:25:00.740 --> 00:25:04.819 $57 million series around. And the capital was 00:25:04.819 --> 00:25:07.500 very targeted to help manufacturers specifically 00:25:07.500 --> 00:25:10.359 adopt these computer vision technologies, which 00:25:10.359 --> 00:25:13.039 requires not just software, but intensive on 00:25:13.039 --> 00:25:16.039 -site integration and specific training for often 00:25:16.039 --> 00:25:18.420 dirty, unpredictable industrial environments. 00:25:18.680 --> 00:25:20.900 His investment strategy through the AI Fund also 00:25:20.900 --> 00:25:23.039 highlights a commitment to global reach and high 00:25:23.039 --> 00:25:25.059 -impact sectors like healthcare. Absolutely. 00:25:25.099 --> 00:25:27.680 The fund made its first investment in India in 00:25:27.680 --> 00:25:31.359 October 2024, backing an AI healthcare startup 00:25:31.359 --> 00:25:34.519 called Jeevi. And what does Jeevy do? It focuses 00:25:34.519 --> 00:25:36.980 on using AI for critical functions, assisting 00:25:36.980 --> 00:25:39.319 with diagnoses, providing treatment recommendations, 00:25:39.640 --> 00:25:42.079 and handling complex, time -consuming administrative 00:25:42.079 --> 00:25:44.799 tasks. This investment not only signals confidence 00:25:44.799 --> 00:25:47.559 in India's growing AI sector, but also aligns 00:25:47.559 --> 00:25:49.880 perfectly with Eng's vision of applying AI to 00:25:49.880 --> 00:25:52.059 improve critical public services. And beyond 00:25:52.059 --> 00:25:54.339 his own ventures, his expertise is highly valued 00:25:54.339 --> 00:25:56.220 across a wide array of high -profile advisory 00:25:56.220 --> 00:25:58.680 and board roles. It really showcases his diverse 00:25:58.680 --> 00:26:00.720 interests. For sure. He serves as the chair of 00:26:00.720 --> 00:26:02.880 the board for Wobot Labs, a digital psychological 00:26:02.880 --> 00:26:05.950 clinic. Wobot uses a therapy chat bot powered 00:26:05.950 --> 00:26:08.109 by data science to provide cognitive behavioral 00:26:08.109 --> 00:26:10.869 therapy for conditions like depression. That's 00:26:10.869 --> 00:26:12.849 a high -impact application of conversational 00:26:12.849 --> 00:26:15.950 AI. Then, in autonomous driving, he was a member 00:26:15.950 --> 00:26:19.029 of the board for Drive .ai, a self -driving car 00:26:19.029 --> 00:26:21.569 company that was successfully acquired by Apple 00:26:21.569 --> 00:26:24.470 in 2019. Which is further evidence of his deep 00:26:24.470 --> 00:26:27.089 involvement at the highest levels of AI application 00:26:27.089 --> 00:26:30.450 in transport. And then cementing his status as 00:26:30.450 --> 00:26:33.029 a key figure in the global tech hierarchy, in 00:26:33.029 --> 00:26:35.890 April 2024, Amazon announced his appointment 00:26:35.890 --> 00:26:37.990 to its board of directors. Wow. So that places 00:26:37.990 --> 00:26:40.390 him at the strategic level, advising one of the 00:26:40.390 --> 00:26:42.789 world's largest companies as the generative AI 00:26:42.789 --> 00:26:45.670 race intensifies, guiding long term strategy 00:26:45.670 --> 00:26:48.289 for infrastructure, consumer services and the 00:26:48.289 --> 00:26:50.809 cloud. If we zoom out now to look at the entire 00:26:50.809 --> 00:26:53.369 arc of his career, we can gauge the full scale 00:26:53.369 --> 00:26:55.670 of his legacy and influence. It's a career built 00:26:55.670 --> 00:26:57.980 not just on developing AI technology. but on 00:26:57.980 --> 00:27:00.259 communicating it and democratizing access to 00:27:00.259 --> 00:27:03.700 it. The measurable output is just immense. He's 00:27:03.700 --> 00:27:06.740 the author or co -author of over 300 publications 00:27:06.740 --> 00:27:09.640 in robotics, machine learning, and related fields. 00:27:09.940 --> 00:27:13.019 His work is consistently cited and forms the 00:27:13.019 --> 00:27:15.480 basis for many modern AI curricula worldwide. 00:27:16.170 --> 00:27:19.029 And he consciously took the time to write highly 00:27:19.029 --> 00:27:21.809 practical material for the wider industry community, 00:27:22.089 --> 00:27:25.170 not just for his academic peers. Correct. That 00:27:25.170 --> 00:27:27.230 drive for practical application really shines 00:27:27.230 --> 00:27:29.509 through in his non -academic writings. He wrote 00:27:29.509 --> 00:27:31.750 the book Machine Learning Yearning, which he 00:27:31.750 --> 00:27:33.750 distributed for free. And it's a practical guide, 00:27:33.869 --> 00:27:35.910 right? Focusing on how to structure machine learning 00:27:35.910 --> 00:27:37.930 projects. Exactly. It wasn't about the theory. 00:27:38.069 --> 00:27:40.069 It was about the practical engineering decisions 00:27:40.069 --> 00:27:43.069 you need to make to ship an AI product. A core 00:27:43.069 --> 00:27:45.660 lesson from his time at Google and Baidu. He 00:27:45.660 --> 00:27:47.980 followed that up with the AI Transformation Playbook, 00:27:48.119 --> 00:27:50.460 which guides executive leadership at companies 00:27:50.460 --> 00:27:53.720 on how to integrate AI strategically. This high 00:27:53.720 --> 00:27:56.140 volume of high quality output spanning academia, 00:27:56.460 --> 00:27:58.920 industry, and education has earned him continuous 00:27:58.920 --> 00:28:01.740 recognition across two decades. The accolades 00:28:01.740 --> 00:28:04.059 are pretty significant. He received the Sloan 00:28:04.059 --> 00:28:08.680 Fellowship in 2007. In 2009, he won the IJCAI 00:28:08.680 --> 00:28:11.420 Computers and Thought Award. which is the highest 00:28:11.420 --> 00:28:14.960 award in AI for a researcher under 35. Cementing 00:28:14.960 --> 00:28:17.500 his status as a leading mind of his generation. 00:28:17.740 --> 00:28:20.220 Then, as we noted, Time 100 most influential 00:28:20.220 --> 00:28:24.279 people in 2013 and again in 2023. Fortunes 40 00:28:24.279 --> 00:28:27.779 under 40 in 2013. A World Economic Forum young 00:28:27.779 --> 00:28:31.019 global leader in 2015. So he's consistently recognized 00:28:31.019 --> 00:28:33.180 across different phases of the AI development 00:28:33.180 --> 00:28:35.779 cycle for his impact and leadership. But let's 00:28:35.779 --> 00:28:37.619 turn to his current vision for the future of 00:28:37.619 --> 00:28:40.359 AI, particularly his highly pragmatic perspective 00:28:40.359 --> 00:28:42.720 on regulation, which has been quite vocal in 00:28:42.720 --> 00:28:45.079 recent years. And he takes a remarkably grounded 00:28:45.079 --> 00:28:47.579 view on the existential threats that are often 00:28:47.579 --> 00:28:50.420 discussed in AI circles, specifically the doomsday 00:28:50.420 --> 00:28:53.099 scenarios. The evil killer robots. Right. He 00:28:53.099 --> 00:28:54.900 believes the industry and government are often 00:28:54.900 --> 00:28:57.539 distracted by evil killer robots, as he put it. 00:28:57.579 --> 00:29:00.019 The real threat in his... view is the much more 00:29:00.019 --> 00:29:02.200 immediate and tangible challenge to the labor 00:29:02.200 --> 00:29:07.150 market caused by these machines. Yes. That is 00:29:07.150 --> 00:29:09.750 the socioeconomic conversation, he argues, that 00:29:09.750 --> 00:29:12.089 everyone should be having and addressing through 00:29:12.089 --> 00:29:14.769 policy and educational changes. That reframes 00:29:14.769 --> 00:29:16.750 the debate from science fiction to socioeconomic 00:29:16.750 --> 00:29:19.609 reality. He is also a consistent advocate for 00:29:19.609 --> 00:29:21.950 technically accelerating the field itself. Yes. 00:29:22.150 --> 00:29:26.269 Since 2017, he has consistently advocated for 00:29:26.269 --> 00:29:28.750 the shift to high -performance computing, or 00:29:28.750 --> 00:29:32.490 HPC. It's a direct continuation of his pioneering 00:29:32.490 --> 00:29:36.619 2008 GPU advocacy. He sees computational power 00:29:36.619 --> 00:29:39.500 as the continuing bottleneck that must be overcome 00:29:39.500 --> 00:29:41.859 for the next breakthroughs. And regarding specific 00:29:41.859 --> 00:29:45.079 regulatory efforts aimed at AI safety, particularly 00:29:45.079 --> 00:29:47.559 those that might affect smaller players and open 00:29:47.559 --> 00:29:49.980 source models, he has been an outspoken critic. 00:29:50.240 --> 00:29:52.519 He has. He's argued that heavy -handed rules 00:29:52.519 --> 00:29:54.910 could be counterproductive. In a December 2023 00:29:54.910 --> 00:29:57.170 interview, he highlighted that potential regulations 00:29:57.170 --> 00:29:59.589 involving reporting mandates, licensing and liability 00:29:59.589 --> 00:30:02.009 risks could disproportionately burden smaller 00:30:02.009 --> 00:30:04.910 firms and effectively stifle innovation. His 00:30:04.910 --> 00:30:07.309 core argument is that compliance costs are steep. 00:30:07.529 --> 00:30:11.250 Very steep. And imposing them universally, especially 00:30:11.250 --> 00:30:14.109 on foundational open source technologies, would 00:30:14.109 --> 00:30:16.049 only strengthen the market dominance of a few 00:30:16.049 --> 00:30:18.589 large companies, the Googles and Betas of the 00:30:18.589 --> 00:30:21.730 world, while hindering the very democratization 00:30:21.730 --> 00:30:24.819 he champions. So the regulatory structure intended 00:30:24.819 --> 00:30:27.539 for safety might actually create a centralized, 00:30:27.880 --> 00:30:30.859 anti -competitive environment. Precisely. He 00:30:30.859 --> 00:30:33.259 advocates for regulations that are carefully 00:30:33.259 --> 00:30:36.380 designed to target specific, proven risks rather 00:30:36.380 --> 00:30:38.819 than hypothetical ones, and that prevent obstacles 00:30:38.819 --> 00:30:41.680 to the development of beneficial AI, especially 00:30:41.680 --> 00:30:44.539 open source models, which he views as the engine 00:30:44.539 --> 00:30:47.380 of broad innovation. We saw this entire tension 00:30:47.380 --> 00:30:49.940 play out dramatically in California with some 00:30:49.940 --> 00:30:53.269 proposed legislation earlier in 2020. That's 00:30:53.269 --> 00:30:56.769 the perfect case study. In June 2024, Eng expressed 00:30:56.769 --> 00:30:59.269 significant concern about a California bill that 00:30:59.269 --> 00:31:01.089 would have required developers of advanced models 00:31:01.089 --> 00:31:03.670 to implement safety mechanisms, including a mandatory 00:31:03.670 --> 00:31:06.049 kill switch. A kill switch. And liability for 00:31:06.049 --> 00:31:08.009 models that could potentially assist in dangerous 00:31:08.009 --> 00:31:10.509 activities. He described the bill as creating 00:31:10.509 --> 00:31:12.970 massive liabilities for science fiction risks 00:31:12.970 --> 00:31:16.170 and said that it stokes fear in anyone daring 00:31:16.170 --> 00:31:19.099 to innovate. The technical complexity and potential 00:31:19.099 --> 00:31:21.400 liability attached to requiring a kill switch 00:31:21.400 --> 00:31:24.500 for a widely distributed open source model were 00:31:24.500 --> 00:31:27.380 cited by many opponents, including Eng, as an 00:31:27.380 --> 00:31:30.000 impossible burden. The consensus among critics 00:31:30.000 --> 00:31:32.869 was that The bill disproportionately targeted 00:31:32.869 --> 00:31:35.569 open source developers and smaller AI companies 00:31:35.569 --> 00:31:38.269 who just don't have the legal and financial resources 00:31:38.269 --> 00:31:40.890 of the tech giants. Correct. And while the debate 00:31:40.890 --> 00:31:43.390 was significant and very high profile, the source 00:31:43.390 --> 00:31:45.690 material notes that the bill was ultimately vetoed 00:31:45.690 --> 00:31:48.210 by the governor in September 2024. Which aligned 00:31:48.210 --> 00:31:50.650 with Ang's stated concerns about regulation that 00:31:50.650 --> 00:31:52.910 targets perceived science fiction risks while 00:31:52.910 --> 00:31:55.730 imposing these heavy stifling burdenings on innovators. 00:31:55.930 --> 00:31:58.670 His entire advocacy really centers on empowering 00:31:58.670 --> 00:32:01.559 people to use AI tools. seeing this as essential 00:32:01.559 --> 00:32:04.259 for building beneficial AI applications all over 00:32:04.259 --> 00:32:06.819 the world. What an incredibly comprehensive and 00:32:06.819 --> 00:32:09.880 multifaceted career we've tracked today. We followed 00:32:09.880 --> 00:32:12.140 Andrew Nang from a math Olympian to the academic 00:32:12.140 --> 00:32:15.299 pioneer who championed GPUs, the industry leader 00:32:15.299 --> 00:32:18.059 who co -founded Google Brain and built Baidu's 00:32:18.059 --> 00:32:20.680 massive AI division, the educational revolutionary 00:32:20.680 --> 00:32:23.759 who co -founded Coursera and brought foundational 00:32:23.759 --> 00:32:27.259 AI knowledge to 8 million people, and now... 00:32:27.519 --> 00:32:29.859 The venture capitalist focused on democratizing 00:32:29.859 --> 00:32:33.420 applied AI for businesses through the AI fund 00:32:33.420 --> 00:32:36.180 and landing AI. His career trajectory, when you 00:32:36.180 --> 00:32:38.740 lay it all out, reveals one unifying, consistent 00:32:38.740 --> 00:32:40.900 philosophy that connects everything he's done. 00:32:41.079 --> 00:32:44.140 I think so, too. It's a deep -seated belief that 00:32:44.140 --> 00:32:47.259 high -level complex technology must be democratized. 00:32:47.359 --> 00:32:49.720 He did it first by open -sourcing foundational 00:32:49.720 --> 00:32:53.019 software like ROS to standardize robotics, then 00:32:53.019 --> 00:32:55.299 by mass distribution of knowledge through MOOCs 00:32:55.299 --> 00:32:57.900 to educate the world, and now he's lowering the 00:32:57.900 --> 00:32:59.759 barrier for business adoption with platforms 00:32:59.759 --> 00:33:02.250 like Landing AI. The goal has always been to 00:33:02.250 --> 00:33:04.690 remove the exclusivity of powerful technology. 00:33:04.970 --> 00:33:06.950 It's not just about building the smartest models. 00:33:07.089 --> 00:33:08.750 It's about giving everyone the organizational 00:33:08.750 --> 00:33:11.589 and educational tools to use them effectively 00:33:11.589 --> 00:33:14.690 and responsibly. Precisely. Which leaves us with 00:33:14.690 --> 00:33:17.230 a final thought for you to explore as you process 00:33:17.230 --> 00:33:19.710 all this knowledge. Given Andrew Nang's vocal 00:33:19.710 --> 00:33:22.809 opposition to AI safety regulation that he sees 00:33:22.809 --> 00:33:25.690 as unfairly burdening smaller firms and hindering 00:33:25.690 --> 00:33:28.059 open source development. and considering his 00:33:28.059 --> 00:33:30.559 decades -long commitment to expanding access 00:33:30.559 --> 00:33:33.599 to high -performance AI tools. How might his 00:33:33.599 --> 00:33:36.220 specific, influential vision shape the crucial 00:33:36.220 --> 00:33:38.599 balance between rapid innovation and responsible 00:33:38.599 --> 00:33:40.880 governance in the technological landscape over 00:33:40.880 --> 00:33:43.700 the next five years? Will the push for democratization 00:33:43.700 --> 00:33:46.440 eventually clash with the push for centralized 00:33:46.440 --> 00:33:49.299 control and safety? It's a tension that really 00:33:49.299 --> 00:33:51.519 defines the industry, and Andrew Nang is right 00:33:51.519 --> 00:33:53.579 in the middle of that conversation, guiding both 00:33:53.579 --> 00:33:55.720 the technology and the debate. Welcome to The 00:33:55.720 --> 00:33:58.099 Debate. Today, we're delving into the career 00:33:58.099 --> 00:34:02.759 and the stated concerns of Andrew Ng. He's a 00:34:02.759 --> 00:34:05.880 figure who has moved pretty seamlessly from foundational 00:34:05.880 --> 00:34:09.199 machine learning research to leading major labs 00:34:09.199 --> 00:34:11.539 at Google and Baidu, and then to creating this 00:34:11.539 --> 00:34:14.659 massive global education movement with Coursera 00:34:14.659 --> 00:34:19.079 and deep learning AI. His influence is just undeniable. 00:34:19.340 --> 00:34:22.940 Absolutely. And he's articulated two very different, 00:34:22.980 --> 00:34:25.820 but you could argue equally urgent, challenges 00:34:25.820 --> 00:34:28.340 for AI. On the one hand, you have this warning 00:34:28.340 --> 00:34:32.099 about a sort of societal disaster from job displacement. 00:34:32.260 --> 00:34:34.920 But on the other, he's now actively campaigning 00:34:34.920 --> 00:34:37.780 against what he sees as poorly formed policy 00:34:37.780 --> 00:34:41.579 that could choke innovation. Exactly. So that 00:34:41.579 --> 00:34:44.460 brings us to our core question. Which concern 00:34:44.460 --> 00:34:47.400 that Andrew Gee has articulated presents the 00:34:47.400 --> 00:34:49.719 more critical challenge to the future of AI? 00:34:50.119 --> 00:34:53.000 Is it the fundamental societal threat of job 00:34:53.000 --> 00:34:55.480 displacement, the whole future of work problem, 00:34:55.639 --> 00:34:58.679 or is it the immediate trajectory -altering risk 00:34:58.679 --> 00:35:01.619 of bad government regulations stifling open -source 00:35:01.619 --> 00:35:05.000 innovation? Hmm. I'm going to argue that the 00:35:05.000 --> 00:35:07.599 threat to human labor and the systemic conversation 00:35:07.599 --> 00:35:10.599 that needs to happen around it is the most substantive 00:35:10.599 --> 00:35:13.840 and foundational challenge he identifies. And 00:35:13.840 --> 00:35:16.900 I'll be arguing that Eng's recent very explicit 00:35:16.900 --> 00:35:20.159 advocacy against regulation actually highlights 00:35:20.159 --> 00:35:22.559 the more pressing and, frankly, more immediately 00:35:22.559 --> 00:35:25.659 disruptive risk to the entire innovation ecosystem. 00:35:26.000 --> 00:35:29.139 So my position is that Eng's most significant 00:35:29.139 --> 00:35:32.500 long -term contribution here is his laser focus 00:35:32.500 --> 00:35:37.219 on the profound social upheaval that AI adoption 00:35:37.219 --> 00:35:40.199 is causing. The source material captures his 00:35:40.199 --> 00:35:43.260 philosophy perfectly, I think. Eng explicitly 00:35:43.260 --> 00:35:45.780 calls the labor challenge the real threat. He 00:35:45.780 --> 00:35:48.320 emphasizes that the challenge to labor caused 00:35:48.320 --> 00:35:51.300 by these machines is a conversation that academia, 00:35:51.619 --> 00:35:54.940 industry, and government should have. Right. 00:35:55.019 --> 00:35:58.480 He sees this systemic issue. I mean, how do our 00:35:58.480 --> 00:36:01.460 economies even cope when cognitive tasks, not 00:36:01.460 --> 00:36:04.500 just manual ones, are automated? He sees that 00:36:04.500 --> 00:36:07.500 as the absolute priority. He's purposefully trying 00:36:07.500 --> 00:36:09.880 to distract us from what he calls the evil killer 00:36:09.880 --> 00:36:12.380 robots fair, and crucially his actions really 00:36:12.380 --> 00:36:15.539 back this up. His massive effort to democratize 00:36:15.539 --> 00:36:17.840 deep learning through online courses, teaching 00:36:17.840 --> 00:36:20.699 millions of students, that's a direct, sustained, 00:36:21.000 --> 00:36:23.039 strategic investment in preparing the global 00:36:23.039 --> 00:36:25.360 workforce for what he sees as an unavoidable 00:36:25.360 --> 00:36:28.079 transition. That level of effort tells me that's 00:36:28.079 --> 00:36:30.400 where he thinks the real threat lies. Okay, I 00:36:30.400 --> 00:36:33.389 see that. But I approach this from a perspective 00:36:33.389 --> 00:36:38.090 of, well, operational viability. While the labor 00:36:38.090 --> 00:36:41.690 issue is, yes, undeniably the largest theoretical 00:36:41.690 --> 00:36:45.190 consequence of AI success, the threat of poorly 00:36:45.190 --> 00:36:48.389 conceived regulation is the more immediate and 00:36:48.389 --> 00:36:52.030 devastating practical challenge. Regulation directly 00:36:52.030 --> 00:36:54.489 impacts the ability to build and deploy the very 00:36:54.489 --> 00:36:56.889 beneficial AI technologies that he champions, 00:36:57.190 --> 00:36:59.969 the same technologies we need to mitigate job 00:36:59.969 --> 00:37:06.440 displacement. I mean, Nung's recent public posture 00:37:06.440 --> 00:37:10.039 is one of immediate alarm. He's voiced strong 00:37:10.039 --> 00:37:13.079 concerns that regulations on basic or open source 00:37:13.079 --> 00:37:16.420 models could, and I'm quoting here, unfairly 00:37:16.420 --> 00:37:19.179 burden smaller firms and stifle innovation. When 00:37:19.179 --> 00:37:21.840 he opposed that proposed California bill, he 00:37:21.840 --> 00:37:24.539 wasn't just. you know, defending profits, he 00:37:24.539 --> 00:37:27.360 was warning that it created massive liabilities 00:37:27.360 --> 00:37:30.579 for science fiction risks and that it stokes 00:37:30.579 --> 00:37:33.360 fear in anyone daring to innovate. So the greatest 00:37:33.360 --> 00:37:35.940 threat right now is a political overreaction 00:37:35.940 --> 00:37:38.940 that could just halt progress entirely. If we 00:37:38.940 --> 00:37:41.599 can't innovate freely, that labor problem is 00:37:41.599 --> 00:37:44.219 guaranteed to get worse. But we're debating the 00:37:44.219 --> 00:37:47.349 scope of the real threat. And I think the language 00:37:47.349 --> 00:37:50.730 Njeng uses is really instructive. He explicitly 00:37:50.730 --> 00:37:53.630 frames the labor challenges fundamental. I mean, 00:37:53.630 --> 00:37:55.730 this isn't about a specific startup's profitability. 00:37:56.130 --> 00:37:58.710 This is about the stability of modern economies, 00:37:58.869 --> 00:38:01.389 the well -being of billions of workers. We're 00:38:01.389 --> 00:38:03.849 talking about an existential issue that demands 00:38:03.849 --> 00:38:06.630 this massive response from, you know, education 00:38:06.630 --> 00:38:09.750 policy, social safety nets. The scale of that 00:38:09.750 --> 00:38:12.409 action just far exceeds the concerns over a single 00:38:12.409 --> 00:38:15.130 state legislative effort. The focus on labor 00:38:15.130 --> 00:38:18.150 is proactive. The regulatory pushback is fundamentally 00:38:18.150 --> 00:38:21.489 a defensive reactive measure. Shouldn't we prioritize 00:38:21.489 --> 00:38:23.949 the foundational challenge over what is essentially 00:38:23.949 --> 00:38:26.809 operational friction? Yeah, but I think calling 00:38:26.809 --> 00:38:29.650 one foundational and the other just operational, 00:38:29.929 --> 00:38:32.949 I think that's a bit misleading, isn't it? The 00:38:32.949 --> 00:38:35.690 operational friction is foundational to his current 00:38:35.690 --> 00:38:38.849 goals. If we accept that AI innovation is the 00:38:38.849 --> 00:38:41.369 very mechanism we need to solve the labor crisis, 00:38:41.550 --> 00:38:45.059 creating new jobs, new tools, then anything that 00:38:45.059 --> 00:38:48.400 slows or stops that innovation is, by definition, 00:38:48.579 --> 00:38:51.900 accelerating the societal crisis. I'm not sure 00:38:51.900 --> 00:38:54.559 I follow. Well, the regulatory threat targets 00:38:54.559 --> 00:38:58.280 open -source AI and small startups. If the development 00:38:58.280 --> 00:39:01.260 environment becomes too punitive, only the giants, 00:39:01.480 --> 00:39:04.260 the very firms Eng often criticizes for having 00:39:04.260 --> 00:39:07.820 too much power, can afford the compliance. Eng 00:39:07.820 --> 00:39:09.719 knows that restricting these small, innovative 00:39:09.719 --> 00:39:12.820 firms limits the distribution and diversity of 00:39:12.820 --> 00:39:15.800 beneficial AI. So how can we fix the future of 00:39:15.800 --> 00:39:18.119 work if we eliminate the market's capacity to 00:39:18.119 --> 00:39:21.159 even generate solutions? I mean, if the societal 00:39:21.159 --> 00:39:23.760 threat is the real threat, why isn't he a leading 00:39:23.760 --> 00:39:26.820 advocate for, say, universal basic income or 00:39:26.820 --> 00:39:30.039 massive public works programs? His primary solution 00:39:30.039 --> 00:39:32.900 is private education, which addresses the labor 00:39:32.900 --> 00:39:36.079 threat symptomatically, not systematically. That's 00:39:36.079 --> 00:39:38.440 a fair point about private versus public solutions, 00:39:38.579 --> 00:39:41.539 but I think it misses the whole idea of democratization. 00:39:42.680 --> 00:39:45.760 Eng's response isn't about political restructuring. 00:39:45.800 --> 00:39:49.500 It's about empowerment. His career shift, you 00:39:49.500 --> 00:39:51.380 know, moving from Google Brain to co -founding 00:39:51.380 --> 00:39:54.239 Coursera and deep learning AI, that signifies 00:39:54.239 --> 00:39:57.059 a strategic choice to prioritize human capital 00:39:57.059 --> 00:40:00.280 on a scale we've never seen before. The long 00:40:00.280 --> 00:40:02.460 -term stability you get from preparing millions 00:40:02.460 --> 00:40:06.019 for this AI -powered world, I think that outweighs 00:40:06.019 --> 00:40:08.340 the risk of some temporary regulatory friction. 00:40:08.559 --> 00:40:12.000 You can't innovate in a society that's unraveling. 00:40:12.230 --> 00:40:14.690 I'm sorry, I have to push back on that term, 00:40:14.750 --> 00:40:18.489 regulatory friction. It is far more dangerous 00:40:18.489 --> 00:40:21.630 than just friction. And Gee's opposition to that 00:40:21.630 --> 00:40:23.809 California bill, the one mandating things like 00:40:23.809 --> 00:40:26.190 safety testing and kill switches for frontier 00:40:26.190 --> 00:40:28.989 models, that wasn't some academic objection. 00:40:29.269 --> 00:40:32.030 It was based on the practical fear that these 00:40:32.030 --> 00:40:34.909 policies impose compliance burdens and liabilities 00:40:34.909 --> 00:40:37.809 that are just impossible for small firms to meet. 00:40:38.780 --> 00:40:41.300 When he says massive liabilities for science 00:40:41.300 --> 00:40:43.679 fiction risks, what he means is that a developer 00:40:43.679 --> 00:40:46.000 of an open source model could be held liable 00:40:46.000 --> 00:40:49.219 for some catastrophic, highly improbable misuse 00:40:49.219 --> 00:40:51.400 scenario. You know, once that model is out in 00:40:51.400 --> 00:40:53.699 the world, that's a burden only a huge established 00:40:53.699 --> 00:40:56.239 firm can absorb. This directly counters his goal 00:40:56.239 --> 00:40:58.599 of lowering the barrier for entrance. These regulations 00:40:58.599 --> 00:41:01.420 don't just slow innovation. They're an insurmountable 00:41:01.420 --> 00:41:03.840 barrier for new entrance. They stifle the open 00:41:03.840 --> 00:41:06.239 source community, which is the very engine of 00:41:06.239 --> 00:41:09.239 democratized deployment. concern is an immediate 00:41:09.239 --> 00:41:11.639 threat to the mechanism of democratization itself. 00:41:11.980 --> 00:41:14.199 But the long -term efficacy of the education 00:41:14.199 --> 00:41:17.719 is what mitigates the generational risk. The 00:41:17.719 --> 00:41:20.440 whole democratization effort, evidenced by his 00:41:20.440 --> 00:41:22.960 MOOC success, I mean, reaching an estimated 8 00:41:22.960 --> 00:41:25.800 million students worldwide, that is an enduring 00:41:25.800 --> 00:41:28.860 policy solution. You look at the non -technical 00:41:28.860 --> 00:41:31.380 course, AI for Everyone. It was specifically 00:41:31.380 --> 00:41:34.000 designed to help people understand AI's impact. 00:41:34.590 --> 00:41:37.429 This shows a continuous, career -long commitment 00:41:37.429 --> 00:41:40.690 to managing the human side of AI. His actions 00:41:40.690 --> 00:41:43.409 across a decade prove that preparing society 00:41:43.409 --> 00:41:47.110 for the shock is his central mission. The friction 00:41:47.110 --> 00:41:49.630 from regulating startups while, yes, painful 00:41:49.630 --> 00:41:52.469 for the industry, is ultimately an industry problem. 00:41:52.730 --> 00:41:55.269 It's not the foundational existential threat 00:41:55.269 --> 00:41:57.269 to the global workforce that he spends most of 00:41:57.269 --> 00:41:59.630 his career trying to address. I think we need 00:41:59.630 --> 00:42:01.690 to look closer at his most recent professional 00:42:01.690 --> 00:42:04.090 shifts, though. That's a compelling argument 00:42:04.090 --> 00:42:06.510 about the scale of the education. But have you 00:42:06.510 --> 00:42:09.110 considered that his move to the AI fund, a huge 00:42:09.110 --> 00:42:12.150 venture capital vehicle, and landing AI fundamentally 00:42:12.150 --> 00:42:15.670 changes his priorities? NG is no longer primarily 00:42:15.670 --> 00:42:18.869 an educator. He's a VC investor. His investments 00:42:18.869 --> 00:42:21.829 in startups like that AI healthcare firm GV place 00:42:21.829 --> 00:42:24.210 him directly in the path of these new regulatory 00:42:24.210 --> 00:42:27.289 concerns. If a badly written bill creates immense 00:42:27.289 --> 00:42:29.969 liability, it could kill a promising beneficial 00:42:29.969 --> 00:42:32.719 company before it ever scales. When he publicly 00:42:32.719 --> 00:42:35.219 defends these interests, he's defending the producers 00:42:35.219 --> 00:42:38.380 of innovation. This shift suggests the immediacy 00:42:38.380 --> 00:42:40.480 of the threat to the development pipeline has 00:42:40.480 --> 00:42:43.119 escalated for him, demanding his active political 00:42:43.119 --> 00:42:45.960 engagement right now. Well, we've certainly illuminated 00:42:45.960 --> 00:42:49.039 two very distinct yet profoundly interconnected 00:42:49.039 --> 00:42:52.139 scales of threat here, all coming from the work 00:42:52.139 --> 00:42:54.960 and advocacy of Andrew Wang. Indeed. It feels 00:42:54.960 --> 00:42:57.159 like we're debating whether to address the long 00:42:57.159 --> 00:42:59.340 -term chronic illness, which is labor displacement. 00:43:00.010 --> 00:43:03.650 or the immediate acute injury, regulatory strangulation 00:43:03.650 --> 00:43:06.289 that risks preventing the cure from ever being 00:43:06.289 --> 00:43:08.610 developed in the first place. I still maintain 00:43:08.610 --> 00:43:11.010 that Nang's deepest concern remains systemic, 00:43:11.269 --> 00:43:13.590 the preparation of humanity for the radical shift 00:43:13.590 --> 00:43:16.750 in labor markets. He established early on that 00:43:16.750 --> 00:43:18.690 we have to address the future of work, positioning 00:43:18.690 --> 00:43:21.130 it as the necessary conversation for everyone. 00:43:21.590 --> 00:43:24.389 His educational empire is a direct, enduring, 00:43:24.610 --> 00:43:27.170 large -scale response to this profound, long 00:43:27.170 --> 00:43:30.000 -term societal issue. And I think it's the necessary 00:43:30.000 --> 00:43:33.099 prerequisite for any stable, sustainable AI deployment. 00:43:33.460 --> 00:43:36.179 Without societal stability, innovation just has 00:43:36.179 --> 00:43:38.940 no context to thrive in. And I still argue that 00:43:38.940 --> 00:43:42.440 while the societal challenge is grave, the more 00:43:42.440 --> 00:43:44.880 acute and immediate danger, according to his 00:43:44.880 --> 00:43:48.760 specific recent advocacy and his shift into venture 00:43:48.760 --> 00:43:52.420 capital, is the imposition of massive liabilities 00:43:52.420 --> 00:43:55.719 from regulation that stifles open source and 00:43:55.719 --> 00:43:58.949 small firms. This regulatory threat compromises 00:43:58.949 --> 00:44:01.329 the foundational development environment itself. 00:44:01.610 --> 00:44:04.489 It jeopardizes the very technologies and companies 00:44:04.489 --> 00:44:06.650 that could provide solutions to the labor challenges 00:44:06.650 --> 00:44:09.449 down the road. So one speaks to the long -term 00:44:09.449 --> 00:44:12.409 goal, the other speaks to the immediate viability 00:44:12.409 --> 00:44:16.150 of the entire AI ecosystem. This discussion really 00:44:16.150 --> 00:44:18.690 demonstrates the complexity inherent in managing 00:44:18.690 --> 00:44:21.619 a technology like AI. It shows how an influential 00:44:21.619 --> 00:44:24.800 figure like Nugg has to balance responding to 00:44:24.800 --> 00:44:27.320 these long -term societal needs with fighting 00:44:27.320 --> 00:44:30.059 immediate practical barriers to innovation. And 00:44:30.059 --> 00:44:33.360 it shows that the definition of the critical 00:44:33.360 --> 00:44:35.920 challenge really depends on your time horizon, 00:44:36.119 --> 00:44:38.579 whether you're focused on the economic stability 00:44:38.579 --> 00:44:41.500 of the next generation or the immediate viability 00:44:41.500 --> 00:44:44.219 of the tools being built right now. This tension 00:44:44.219 --> 00:44:46.780 reveals a lot about the delicate balance needed 00:44:46.780 --> 00:44:49.480 between innovation, regulation, and societal 00:44:49.480 --> 00:44:52.219 impact. And there's clearly so much more to explore 00:44:52.219 --> 00:44:52.960 in the material.