WEBVTT 00:00:00.000 --> 00:00:02.140 Welcome to the Deep Dive, where we take complex 00:00:02.140 --> 00:00:04.679 source material from dense academic papers to 00:00:04.679 --> 00:00:07.120 breaking corporate news and filter it down into 00:00:07.120 --> 00:00:09.160 the core knowledge you need to be truly informed. 00:00:10.000 --> 00:00:13.419 Today, we are undertaking a, well, a massive 00:00:13.419 --> 00:00:16.160 deep dive into a figure who stands at the absolute 00:00:16.160 --> 00:00:20.140 epicenter of the modern AI revolution, Dr. Fei 00:00:20.140 --> 00:00:22.730 -Fei Li. It's an essential subject for anyone 00:00:22.730 --> 00:00:24.710 trying to understand the trajectory of machine 00:00:24.710 --> 00:00:28.570 learning. Dr. Lee is a singular figure who simultaneously 00:00:28.570 --> 00:00:31.250 authored the foundational technical blueprint 00:00:31.250 --> 00:00:33.649 for modern computer vision. She literally gave 00:00:33.649 --> 00:00:36.350 AI the power of sight. Exactly. While at the 00:00:36.350 --> 00:00:38.429 same time remaining probably the most urgent 00:00:38.429 --> 00:00:40.810 voice calling for ethical guardrails and diversity 00:00:40.810 --> 00:00:42.960 in the field. Absolutely. Okay, let's unpack 00:00:42.960 --> 00:00:46.020 this. Our mission for this deep dive is to synthesize 00:00:46.020 --> 00:00:48.420 the two main pillars of her unparalleled influence. 00:00:48.759 --> 00:00:51.500 First, we really need to graph the sheer magnitude 00:00:51.500 --> 00:00:53.840 of her work on ImageNet, you know, the research 00:00:53.840 --> 00:00:56.140 project that fundamentally sparked the deep learning 00:00:56.140 --> 00:00:59.759 boom. And second, we will explore her unwavering, 00:00:59.759 --> 00:01:03.039 persistent commitment to ensuring that this powerful 00:01:03.039 --> 00:01:05.459 technology, the technology she helped create, 00:01:05.760 --> 00:01:09.760 is developed in a human -centered, diverse, and... 00:01:10.219 --> 00:01:12.980 uh ultimately benevolent way what's so fascinating 00:01:12.980 --> 00:01:15.810 here is that her career path doesn't just stick 00:01:15.810 --> 00:01:18.590 to one track. We are looking at a computer scientist 00:01:18.590 --> 00:01:20.769 who has achieved the highest academic honors. 00:01:21.069 --> 00:01:23.549 I mean, she holds the Sequoia Capital Professorship 00:01:23.549 --> 00:01:26.010 at Stanford. Right. But who moves seamlessly 00:01:26.010 --> 00:01:28.689 between executive roles in big tech, founding 00:01:28.689 --> 00:01:31.049 crucial nonprofits to address systemic inequality. 00:01:31.370 --> 00:01:34.329 And now. And now leading a billion dollar entrepreneurial 00:01:34.329 --> 00:01:37.109 venture focused on the next generation of AI 00:01:37.109 --> 00:01:39.849 perception. It's a spectacular range of influence. 00:01:40.049 --> 00:01:41.909 It really is. We will literally trace her path 00:01:41.909 --> 00:01:44.049 from her early life as an immigrant working in 00:01:44.049 --> 00:01:45.799 a family. dry cleaning shop in New Jersey to 00:01:45.799 --> 00:01:48.120 her current role, where she is influencing global 00:01:48.120 --> 00:01:51.140 policy decisions at the United Nations and setting 00:01:51.140 --> 00:01:53.239 the technical standard for spatial intelligence. 00:01:53.500 --> 00:01:56.659 And that's the key, I think. Her impact spans 00:01:56.659 --> 00:01:59.140 the most technical aspects of deep learning, 00:01:59.319 --> 00:02:02.280 the very data that fuels the models, and the 00:02:02.280 --> 00:02:05.120 most ethical aspects, the governance and diversity 00:02:05.120 --> 00:02:07.879 of the field itself. So that path of perseverance 00:02:07.879 --> 00:02:11.580 starts in Section 1. early life, and the foundation 00:02:11.580 --> 00:02:14.539 of persistence. Before she was a world -class 00:02:14.539 --> 00:02:17.120 scientist, she was navigating the incredibly 00:02:17.120 --> 00:02:20.099 demanding reality of immigration and family support. 00:02:20.560 --> 00:02:22.840 It's the kind of background story that, you know, 00:02:22.840 --> 00:02:25.520 it profoundly shapes a scientist's output. She 00:02:25.520 --> 00:02:28.659 was born in Beijing in 1976 and spent her formative 00:02:28.659 --> 00:02:31.139 years growing up in Chengdu in Sichuan province. 00:02:31.340 --> 00:02:33.900 When she was 12, her father immigrated to the 00:02:33.900 --> 00:02:36.979 U .S. And then four years later, at age 16, she 00:02:36.979 --> 00:02:38.819 and her mother joined him in Parsippany, New 00:02:38.819 --> 00:02:41.379 Jersey. And this wasn't an easy transition from 00:02:41.379 --> 00:02:43.680 what the sources say. While attending Parsippany 00:02:43.680 --> 00:02:45.759 High School, she quickly took on responsibilities 00:02:45.759 --> 00:02:48.479 outside of her academic life. She worked weekends 00:02:48.479 --> 00:02:50.500 at her family's dry cleaning shop. And that's 00:02:50.500 --> 00:02:53.560 a crucial detail. This wasn't just like a part 00:02:53.560 --> 00:02:56.580 time job for pocket money. It was central to 00:02:56.580 --> 00:02:59.360 the family's economic stability. The sources 00:02:59.360 --> 00:03:03.539 really highlight that the sheer effort and organizational 00:03:03.539 --> 00:03:06.219 discipline required to maintain this commitment 00:03:06.219 --> 00:03:09.780 while excelling in a rigorous American high school 00:03:09.780 --> 00:03:12.500 environment. Well, it laid a clear foundation 00:03:12.500 --> 00:03:15.099 for her later career. It certainly speaks to 00:03:15.099 --> 00:03:18.259 an incredible tenacity. And that work ethic didn't 00:03:18.259 --> 00:03:20.120 stop when she got into Princeton University, 00:03:20.379 --> 00:03:23.240 which is already a huge achievement in itself. 00:03:23.439 --> 00:03:26.719 No, it didn't. Not at all. She pursued a Bachelor 00:03:26.719 --> 00:03:29.080 of Arts with a major in physics, one of the most 00:03:29.080 --> 00:03:31.560 intellectually demanding fields you can pick. 00:03:31.659 --> 00:03:33.960 Right. But she still returned home most weekends 00:03:33.960 --> 00:03:37.180 to help run that dry cleaning business. And if 00:03:37.180 --> 00:03:39.060 that weren't enough, the sources specifically 00:03:39.060 --> 00:03:41.860 note that she also worked as a dishwasher to 00:03:41.860 --> 00:03:43.979 supplement the family income while she was studying. 00:03:44.430 --> 00:03:47.150 Just think about that internal drive. You're 00:03:47.150 --> 00:03:48.969 wrestling with concepts like quantum mechanics 00:03:48.969 --> 00:03:51.229 and astrophysics during the week. And then on 00:03:51.229 --> 00:03:53.710 the weekends, you're focused on this high intensity, 00:03:53.990 --> 00:03:57.729 practical, physically demanding labor of running 00:03:57.729 --> 00:03:59.830 a business or working in a kitchen. That must 00:03:59.830 --> 00:04:02.229 instill a unique approach to problem solving. 00:04:02.560 --> 00:04:05.520 a blend of abstract rigor and, you know, necessary 00:04:05.520 --> 00:04:08.620 hands -on practicality. It absolutely does. And 00:04:08.620 --> 00:04:10.699 that dedication is reflected not just in her 00:04:10.699 --> 00:04:13.039 perseverance, but in her intellectual pivot. 00:04:13.460 --> 00:04:15.960 While she was a physics major, her senior thesis 00:04:15.960 --> 00:04:19.420 was titled Auditory Binaural Corellogram Difference, 00:04:19.639 --> 00:04:22.800 a new computational model for Huggins' dichotic 00:04:22.800 --> 00:04:25.439 pitch. Okay, this is a critical point we need 00:04:25.439 --> 00:04:27.800 to dwell on. She's studying physics, but her 00:04:27.800 --> 00:04:30.180 core project involves computational modeling 00:04:30.180 --> 00:04:33.480 of human perception. Exactly. Specifically, how 00:04:33.480 --> 00:04:37.160 the brain processes auditory information, differentiating 00:04:37.160 --> 00:04:39.399 between the sounds entering each ear. That's 00:04:39.399 --> 00:04:41.399 the key through line of her entire scientific 00:04:41.399 --> 00:04:43.660 methodology. It's not just about building machines. 00:04:43.699 --> 00:04:45.540 It's about reverse engineering human capacity. 00:04:46.160 --> 00:04:48.980 She was focused on psychophysics, the relationship 00:04:48.980 --> 00:04:51.779 between physical stimuli and the sensations and 00:04:51.779 --> 00:04:54.339 perceptions they produce. Got it. This early 00:04:54.339 --> 00:04:56.720 work using computation to understand the brain's 00:04:56.720 --> 00:04:58.860 mechanisms for processing sensory input was, 00:04:58.959 --> 00:05:01.800 well, it was the intellectual precursor to ImageNet. 00:05:01.939 --> 00:05:04.420 So she wasn't just randomly interested in different 00:05:04.420 --> 00:05:07.740 senses. She was developing a consistent methodological 00:05:07.740 --> 00:05:11.509 interest. computational modeling of how organisms 00:05:11.509 --> 00:05:14.329 perceive the world. Precisely. This methodology 00:05:14.329 --> 00:05:17.250 took her directly to Caltech, where she secured 00:05:17.250 --> 00:05:19.750 some really prestigious support, including the 00:05:19.750 --> 00:05:21.750 National Science Foundation Graduate Research 00:05:21.750 --> 00:05:24.670 Fellowship and the Paul and Daisy Soros Fellowships 00:05:24.670 --> 00:05:27.329 for New Americans. And that allowed her to pursue 00:05:27.329 --> 00:05:29.910 graduate studies in electrical engineering. She 00:05:29.910 --> 00:05:32.670 received her Master of Science in 2001 and her 00:05:32.670 --> 00:05:35.870 Ph .D. in 2005. And her doctoral work formally 00:05:35.870 --> 00:05:38.949 merged this computational approach with a sense 00:05:38.949 --> 00:05:41.209 that would define her career, vision. Right. 00:05:41.290 --> 00:05:44.310 Her dissertation, Visual Recognition, Computational 00:05:44.310 --> 00:05:47.290 Models in Human Psychophysics, under the supervision 00:05:47.290 --> 00:05:50.129 of Pietro Perona and Christoph Koch, is a landmark. 00:05:50.649 --> 00:05:53.230 Koch is one of the world's most renowned neuroscientists 00:05:53.230 --> 00:05:55.629 specializing in consciousness, while Perona is 00:05:55.629 --> 00:05:58.189 a leading figure in computer vision. So her choice 00:05:58.189 --> 00:06:01.230 of advisors and thesis topic, it explicitly put 00:06:01.230 --> 00:06:03.629 her at the intersection of engineering AI and 00:06:03.629 --> 00:06:05.930 the fundamental understanding of human vision 00:06:05.930 --> 00:06:08.670 psychophysics. That's it. She was asking, what 00:06:08.670 --> 00:06:11.009 does it take for a machine to see the world with 00:06:11.009 --> 00:06:13.870 the complexity that a human child does? That 00:06:13.870 --> 00:06:15.790 exact question. What does it take for a machine 00:06:15.790 --> 00:06:18.670 to see? brings us directly to the technical turning 00:06:18.670 --> 00:06:22.089 point of her career and arguably the entire field 00:06:22.089 --> 00:06:25.490 of AI ImageNet. We really cannot overstate how 00:06:25.490 --> 00:06:27.990 important this piece of research is. This was 00:06:27.990 --> 00:06:30.029 the pivot point for the modern deep learning 00:06:30.029 --> 00:06:33.110 revolution. It's the cornerstone. And to understand 00:06:33.110 --> 00:06:35.170 its revolutionary nature, you have to step back 00:06:35.170 --> 00:06:38.009 to the mid -2000s. Computer vision wasn't a... 00:06:38.269 --> 00:06:41.290 in a deep slump. Models were good at specific, 00:06:41.449 --> 00:06:44.389 constrained tasks, but they couldn't generalize. 00:06:44.509 --> 00:06:46.550 Right. So if you trained a computer to recognize 00:06:46.550 --> 00:06:49.790 a cat on one set of images, it would often feel 00:06:49.790 --> 00:06:52.129 completely when shown a new set. Totally. It 00:06:52.129 --> 00:06:54.470 was brittle. So what was the fundamental technical 00:06:54.470 --> 00:06:57.189 bottleneck there? It was a problem of scale and 00:06:57.189 --> 00:06:59.459 scope in the training data. The gold standard 00:06:59.459 --> 00:07:01.680 for classification competitions at the time was 00:07:01.680 --> 00:07:04.579 the Pascal Visual Object Classes Challenge, or 00:07:04.579 --> 00:07:07.300 Pascal VOC. And while that was valuable, Pascal 00:07:07.300 --> 00:07:10.910 VOC only offered around, what? 20 object categories 00:07:10.910 --> 00:07:13.129 and maybe a few thousand images per category. 00:07:13.410 --> 00:07:15.329 So the models being trained were essentially 00:07:15.329 --> 00:07:18.009 memorizing these tiny snapshots of the world. 00:07:18.110 --> 00:07:20.490 They had no idea of the sheer variability and 00:07:20.490 --> 00:07:22.649 complexity that exists in reality. None at all. 00:07:22.709 --> 00:07:24.889 So if you only show a machine 20 things, it can 00:07:24.889 --> 00:07:26.870 only recognize 20 things. And that's where her 00:07:26.870 --> 00:07:29.029 psychophysics background provided the intellectual 00:07:29.029 --> 00:07:31.949 leap. Right. She realized the fundamental difference 00:07:31.949 --> 00:07:34.709 between human vision and machine vision was not 00:07:34.709 --> 00:07:37.670 the algorithm. It was the volume and organization 00:07:37.670 --> 00:07:41.290 of the input. Drawing on cognitive psychologist 00:07:41.290 --> 00:07:43.629 Irving Biederman's research, which estimated 00:07:43.629 --> 00:07:46.110 that humans recognize around 30 ,000 distinct 00:07:46.110 --> 00:07:49.389 object categories, she set an audacious goal 00:07:49.389 --> 00:07:52.589 in 2007. What was the goal? To build a database 00:07:52.589 --> 00:07:56.370 of 14 million high -resolution images across 00:07:56.370 --> 00:08:00.089 22 ,000 different categories. 14 million images 00:08:00.089 --> 00:08:03.029 in 22 ,000 categories. I mean, that scale was 00:08:03.029 --> 00:08:05.670 truly unheard of and, as the sources note, met 00:08:05.670 --> 00:08:08.860 with intense skepticism. How do you even organize 00:08:08.860 --> 00:08:11.339 22 ,000 categories in a way that's useful for 00:08:11.339 --> 00:08:13.759 a computer? Well, that was the second genius 00:08:13.759 --> 00:08:16.319 move. Instead of creating the categories from 00:08:16.319 --> 00:08:18.779 scratch, which would have been impossible, they 00:08:18.779 --> 00:08:22.220 leveraged WordNet. WordNet is a massive linguistic 00:08:22.220 --> 00:08:25.579 database that groups English words into sets 00:08:25.579 --> 00:08:28.860 of synonyms called synsets, which represent distinct 00:08:28.860 --> 00:08:32.320 concepts. So they used WordNet's existing hierarchical 00:08:32.320 --> 00:08:35.210 structure. The structure that naturally groups 00:08:35.210 --> 00:08:37.990 concepts like mammal containing dog, which contains 00:08:37.990 --> 00:08:40.669 poodle, to organize their visual categories. 00:08:40.990 --> 00:08:43.149 Exactly right. So they were essentially building 00:08:43.149 --> 00:08:46.250 a visual dictionary mapped onto a linguistic 00:08:46.250 --> 00:08:49.509 hierarchy. It gave the visual data a sense of 00:08:49.509 --> 00:08:52.389 relational structure that went far beyond a simple 00:08:52.389 --> 00:08:54.789 flat list of labels. So it gave the system semantic 00:08:54.789 --> 00:08:57.450 context. Precisely. But then came the massive 00:08:57.450 --> 00:08:59.990 logistical challenge of annotating 14 million 00:08:59.990 --> 00:09:02.289 images. I mean, if you hire a PhD student, they 00:09:02.289 --> 00:09:04.330 might label a few hundred images a day. Which 00:09:04.330 --> 00:09:06.509 would take thousands of years. Exactly. This 00:09:06.509 --> 00:09:08.669 is where that practical dry cleaning shop discipline 00:09:08.669 --> 00:09:11.879 kicks in. recognizing the need for an efficient 00:09:11.879 --> 00:09:14.740 system of mass production. The solution was Amazon 00:09:14.740 --> 00:09:17.120 Mechanical Turk, a crowdsourcing marketplace. 00:09:17.580 --> 00:09:20.360 They broke down the labeling task into micropayments, 00:09:20.480 --> 00:09:23.320 a few cents per image, and utilized thousands 00:09:23.320 --> 00:09:26.639 of anonymous workers globally to verify, click, 00:09:26.820 --> 00:09:30.500 and label those 14 million images. It was an 00:09:30.500 --> 00:09:32.960 unprecedented feat of data engineering, combining 00:09:32.960 --> 00:09:35.580 linguistic structure, computational theory, and 00:09:35.580 --> 00:09:38.259 global crowdsourcing. And this process provided 00:09:38.259 --> 00:09:41.240 the enormous messy, structured data set that 00:09:41.240 --> 00:09:44.440 the field desperately needed. But the true inflection 00:09:44.440 --> 00:09:46.600 point came with the competition that used this 00:09:46.600 --> 00:09:49.460 data set. That was the ImageNet Large Scale Visual 00:09:49.460 --> 00:09:53.159 Recognition Challenge, or ILS VRC, which ran 00:09:53.159 --> 00:09:57.000 annually from 2010 to 2017. ILS VRC became the 00:09:57.000 --> 00:09:58.980 proving ground for every new machine learning 00:09:58.980 --> 00:10:01.340 technique. Researchers knew if they could win 00:10:01.340 --> 00:10:03.200 this competition, they had a breakthrough model. 00:10:03.519 --> 00:10:05.759 And what did ImageNet allow researchers to finally 00:10:05.759 --> 00:10:07.639 prove? It allowed them to prove the power of 00:10:07.639 --> 00:10:10.419 deep convolutional neural networks, or DCNNs. 00:10:10.539 --> 00:10:13.039 Before ImageNet, researchers were forced to manually 00:10:13.039 --> 00:10:15.059 engineer features. They had to tell the computer, 00:10:15.179 --> 00:10:17.620 look for an edge here, or look for a corner there. 00:10:17.940 --> 00:10:20.519 But in 2012, the breakthrough moment came with 00:10:20.519 --> 00:10:22.860 a model known as AlexNet. The famous moment when 00:10:22.860 --> 00:10:24.919 the error rate just plummeted. That's right. 00:10:25.019 --> 00:10:28.059 In 2010 and 2011, the error rate for image classification 00:10:28.059 --> 00:10:32.179 was around 25%. AlexNet, trained on the massive 00:10:32.179 --> 00:10:35.059 ImageNet dataset, dropped the error rate to 15 00:10:35.059 --> 00:10:38.580 .3%. This wasn't just an incremental improvement. 00:10:38.860 --> 00:10:40.919 This was the moment deep learning went from an 00:10:40.919 --> 00:10:43.940 academic curiosity to a field -defining technology. 00:10:44.360 --> 00:10:47.240 Because the computer, finally, had enough data 00:10:47.240 --> 00:10:49.340 to learn its own features, rather than being 00:10:49.340 --> 00:10:51.740 told what to look for. That's it. So what does 00:10:51.740 --> 00:10:54.200 this all mean? ImageNet provided the foundational 00:10:54.200 --> 00:10:56.799 fuel that enabled the massive performance leap 00:10:56.799 --> 00:10:59.860 of DCNNs, accelerating the timeline of AI development 00:10:59.860 --> 00:11:03.440 by, what, decades? Arguably, yes. It made things 00:11:03.440 --> 00:11:05.299 possible that were previously science fiction. 00:11:05.659 --> 00:11:08.179 Autonomous vehicles require real -time classification 00:11:08.179 --> 00:11:11.399 of thousands of objects in complex scenes. Medical 00:11:11.399 --> 00:11:13.960 imaging diagnostics rely on recognizing subtle 00:11:13.960 --> 00:11:17.240 patterns in vast datasets of scans. Facial recognition. 00:11:17.820 --> 00:11:21.139 Facial recognition and, yes, the subsequent ethical 00:11:21.139 --> 00:11:23.720 debates around bias, all of it flows directly 00:11:23.720 --> 00:11:26.600 from ImageNet. It cemented her place not just 00:11:26.600 --> 00:11:28.960 as a great researcher, but as the architect of 00:11:28.960 --> 00:11:32.000 the modern AI data infrastructure. Her work essentially 00:11:32.000 --> 00:11:35.019 dictated the scale and ambition of all AI research 00:11:35.019 --> 00:11:37.600 that followed. From this massive academic breakthrough, 00:11:37.740 --> 00:11:40.059 her career naturally transitioned into leadership 00:11:40.059 --> 00:11:42.539 and real -world application, which brings us 00:11:42.539 --> 00:11:45.299 to academic leadership and industry interlude. 00:11:45.870 --> 00:11:49.549 Following her PhD, she had a really rapid ascent 00:11:49.549 --> 00:11:51.789 through the top universities. She did. Starting 00:11:51.789 --> 00:11:53.389 as an assistant professor at the University of 00:11:53.389 --> 00:11:55.309 Illinois Urbana -Champaign and then moving to 00:11:55.309 --> 00:11:57.210 Princeton, she eventually joined Stanford in 00:11:57.210 --> 00:12:00.409 2009. Her tenure track was swift, and she quickly 00:12:00.409 --> 00:12:02.669 became a central figure at the nexus of technology 00:12:02.669 --> 00:12:05.009 and research in Silicon Valley. And she took 00:12:05.009 --> 00:12:07.909 on a huge administrative responsibility by serving 00:12:07.909 --> 00:12:10.129 as the director of the Stanford Artificial Intelligence 00:12:10.129 --> 00:12:13.929 Lab, or SAIL, from 2013 to 2018. What was the 00:12:13.929 --> 00:12:16.350 significance of her leadership there? Well, SAIL 00:12:16.350 --> 00:12:18.629 is one of the world's most prestigious AI research 00:12:18.629 --> 00:12:21.649 centers. During her directorship, she was instrumental 00:12:21.649 --> 00:12:24.269 in fostering an environment that embraced the 00:12:24.269 --> 00:12:27.070 deep learning revolution she had initiated. It 00:12:27.070 --> 00:12:29.070 was a period of intense intellectual ferment. 00:12:29.269 --> 00:12:31.570 So she was nurturing the next generation of researchers 00:12:31.570 --> 00:12:34.750 who would go on to lead major AI efforts globally. 00:12:35.190 --> 00:12:37.850 Absolutely. Her impact was felt not just in papers 00:12:37.850 --> 00:12:40.049 published, but in the talent she helped cultivate. 00:12:40.389 --> 00:12:42.509 But then came the strategic decision to take 00:12:42.509 --> 00:12:46.159 a sabbatical in 2017 to join Google. This was 00:12:46.159 --> 00:12:49.000 a significant move for a tenured Stanford professor. 00:12:49.340 --> 00:12:51.559 It was a massive statement about the influence 00:12:51.559 --> 00:12:53.740 shifting toward industry and her willingness 00:12:53.740 --> 00:12:56.700 to meet that influence head on. She served as 00:12:56.700 --> 00:12:59.580 chief scientist of AML and vice president at 00:12:59.580 --> 00:13:02.659 Google Cloud from early 2017 through late 2018. 00:13:03.000 --> 00:13:05.419 And her mandate at Google Cloud wasn't just pure 00:13:05.419 --> 00:13:07.360 research. It was about the democratization of 00:13:07.360 --> 00:13:10.090 AI. That's a key distinction. Her team's focus 00:13:10.090 --> 00:13:14.250 was explicitly democratizing AI technology and 00:13:14.250 --> 00:13:16.509 lowering the barrier for entrance to businesses 00:13:16.509 --> 00:13:19.149 and developers. They understood that deep learning 00:13:19.149 --> 00:13:21.929 required specialized knowledge, knowing how to 00:13:21.929 --> 00:13:24.710 tune models, select architectures, manage data 00:13:24.710 --> 00:13:27.330 pipelines. This was still too restrictive for 00:13:27.330 --> 00:13:29.789 most businesses. Can you give a practical example 00:13:29.789 --> 00:13:32.730 of how they achieved this democratization? What 00:13:32.730 --> 00:13:35.710 did a product like AutoML actually do? So AutoML 00:13:35.710 --> 00:13:38.740 was the flagship effort. Prior to this, if you 00:13:38.740 --> 00:13:41.019 were a developer trying to build a custom image 00:13:41.019 --> 00:13:44.620 classifier for, say, sorting inventory, you needed 00:13:44.620 --> 00:13:47.220 a deep understanding of deep learning, specifically 00:13:47.220 --> 00:13:49.779 how to select and tune thousands of variables 00:13:49.779 --> 00:13:52.500 or hypoparameters. Right. AutoML essentially 00:13:52.500 --> 00:13:54.840 automated the selection, training, and tuning 00:13:54.840 --> 00:13:57.039 of these models. This meant a small business 00:13:57.039 --> 00:13:59.639 developer in any sector didn't need a PhD in 00:13:59.639 --> 00:14:02.019 deep learning to deploy a highly functional classification 00:14:02.019 --> 00:14:04.360 model. They could leverage Google's infrastructure 00:14:04.360 --> 00:14:07.679 to build custom AI tools with far... less specialized 00:14:07.679 --> 00:14:11.039 expertise. So she was taking the power unleashed 00:14:11.039 --> 00:14:13.799 by ImageNet and building the tools to put it 00:14:13.799 --> 00:14:16.320 into the hands of the masses. That is consistent 00:14:16.320 --> 00:14:19.559 theme, isn't it? Taking monumental academic breakthroughs 00:14:19.559 --> 00:14:21.539 and making them practically accessible. That 00:14:21.539 --> 00:14:24.139 continuity is crucial. Her mission wasn't simply 00:14:24.139 --> 00:14:26.039 to build the biggest models. It was to ensure 00:14:26.039 --> 00:14:28.379 the technology was broadly applied and understood. 00:14:29.210 --> 00:14:31.889 Upon her return to Stainwood in the fall of 2018, 00:14:32.269 --> 00:14:35.230 she brought that ethos back into academia and 00:14:35.230 --> 00:14:37.669 formalized it by co -founding the Human -Centered 00:14:37.669 --> 00:14:41.450 AI Institute. That's HAI. What is the institutional 00:14:41.450 --> 00:14:44.110 mission of HAI and why did she feel the need 00:14:44.110 --> 00:14:46.029 to build it? She is the founding co -director 00:14:46.029 --> 00:14:49.070 alongside former Stanford Provost Dr. John Etchemendy. 00:14:49.480 --> 00:14:52.559 The institution's aim is to advance AI research, 00:14:52.840 --> 00:14:55.539 education, policy and practice with the express 00:14:55.539 --> 00:14:58.860 goal of improving the human condition. It wasn't 00:14:58.860 --> 00:15:00.779 enough to study the technology. They needed to 00:15:00.779 --> 00:15:03.460 study the technology's impact on society, politics 00:15:03.460 --> 00:15:06.299 and the economy. The institute is explicitly 00:15:06.299 --> 00:15:08.799 interdisciplinary, bringing together computer 00:15:08.799 --> 00:15:11.460 scientists, ethicists, legal scholars, social 00:15:11.460 --> 00:15:14.580 scientists. It is the formal architectural expression 00:15:14.580 --> 00:15:19.100 of her belief that AI must serve humanity. positive 00:15:19.100 --> 00:15:21.879 impact naturally transitions into Section 4, 00:15:22.039 --> 00:15:25.580 the push for ethical, human -centered AI. This 00:15:25.580 --> 00:15:27.580 is where her role transcends the technical and 00:15:27.580 --> 00:15:30.440 becomes genuinely societal. She's not just building 00:15:30.440 --> 00:15:32.120 algorithms, she's building the future talent 00:15:32.120 --> 00:15:34.240 pipeline and setting moral boundaries. And her 00:15:34.240 --> 00:15:36.940 focus on diversity and inclusion is not an afterthought. 00:15:37.080 --> 00:15:39.460 It is structurally integrated into her mission. 00:15:39.799 --> 00:15:42.759 She co -founded and chairs the non -profit organization 00:15:42.759 --> 00:15:48.600 AI4AL in 2017. The mission is unambiguous, to 00:15:48.600 --> 00:15:51.279 educate and prepare the next generation of AI 00:15:51.279 --> 00:15:53.960 technologists, thinkers, and leaders by promoting 00:15:53.960 --> 00:15:56.159 diversity and inclusion. And this effort started 00:15:56.159 --> 00:15:58.320 locally before it scaled nationally, right? It 00:15:58.320 --> 00:16:00.700 grew out of a much earlier targeted program she 00:16:00.700 --> 00:16:03.720 co -founded in 2015 called Sailors, the Stanford 00:16:03.720 --> 00:16:06.809 AI Lab Outreach Summers. She co -directed this 00:16:06.809 --> 00:16:09.730 program with her former PhD student, Olga Rusikovsky. 00:16:10.049 --> 00:16:12.289 This program focused intensely on introducing 00:16:12.289 --> 00:16:14.570 ninth grade high school girls to AI education 00:16:14.570 --> 00:16:17.090 and research. Starting at ninth grade is so strategic. 00:16:17.330 --> 00:16:19.389 That's a critical age for students to decide 00:16:19.389 --> 00:16:21.899 whether they see themselves in STEM fields. Absolutely. 00:16:22.080 --> 00:16:24.779 The idea was to intervene early enough to break 00:16:24.779 --> 00:16:27.580 the typical pipeline leakage, showing young women 00:16:27.580 --> 00:16:29.940 specifically that they could be creators and 00:16:29.940 --> 00:16:34.100 leaders in this field. AI4AL then expanded this 00:16:34.100 --> 00:16:36.519 model nationally, scaling it through collaborations 00:16:36.519 --> 00:16:39.279 with major figures and institutions, including 00:16:39.279 --> 00:16:42.100 Melinda French -Gates and Jensen Hom of NVIDIA. 00:16:42.620 --> 00:16:45.379 And by 2018, it had expanded to major institutions 00:16:45.379 --> 00:16:47.860 like Princeton, Carnegie Mellon and UC Berkeley. 00:16:48.000 --> 00:16:50.899 That's right. So why the urgent focus on diversity 00:16:50.899 --> 00:16:54.340 for a purely technical field? Why is the inclusion 00:16:54.340 --> 00:16:56.559 of different perspectives so critical from her 00:16:56.559 --> 00:16:59.039 point of view? She emphasizes that the systems 00:16:59.039 --> 00:17:01.100 we are building, the very systems that influence 00:17:01.100 --> 00:17:04.000 everything from loan applications to hiring decisions 00:17:04.000 --> 00:17:07.640 to medical diagnostics, are trained on data created 00:17:07.640 --> 00:17:10.460 by humans. And those systems are built by a very 00:17:10.460 --> 00:17:12.740 narrow slice of humanity. So if the teams building 00:17:12.740 --> 00:17:15.299 the AI are not diverse, the models they create 00:17:15.299 --> 00:17:18.059 will inevitably inherit and amplify the existing 00:17:18.059 --> 00:17:20.200 biases embedded in the data and in the world. 00:17:20.480 --> 00:17:22.960 You're pointing to the concept of bias in data 00:17:22.960 --> 00:17:26.259 sets, even data sets as groundbreaking as ImageNet, 00:17:26.420 --> 00:17:29.119 which, while revolutionary, required constant 00:17:29.119 --> 00:17:31.579 refinement to address implicit biases concerning 00:17:31.579 --> 00:17:34.420 underrepresented populations or geographically 00:17:34.420 --> 00:17:37.480 constrained data. Exactly. And she stresses that 00:17:37.480 --> 00:17:39.500 we are at a turning point where AI is gaining 00:17:39.500 --> 00:17:42.759 unprecedented influence. To ensure its positive 00:17:42.759 --> 00:17:45.319 impact, we have to seize this moment to support 00:17:45.319 --> 00:17:48.140 structural changes, extending from early education 00:17:48.140 --> 00:17:53.769 and mentorship to change. So it's about ensuring 00:17:53.769 --> 00:17:57.289 that the creators of AI reflect the complex global 00:17:57.289 --> 00:17:59.990 population that AI is meant to serve. That's 00:17:59.990 --> 00:18:02.750 the core idea. This dedication to ethical structure 00:18:02.750 --> 00:18:04.890 was put to the sharpest test during her time 00:18:04.890 --> 00:18:07.509 at Google, specifically concerning Project Maven. 00:18:07.609 --> 00:18:09.230 Let's get into the context of that decision. 00:18:09.470 --> 00:18:12.029 Right. So in September 2017, while she was leading 00:18:12.029 --> 00:18:14.410 AI efforts at Google Cloud, the company secured 00:18:14.410 --> 00:18:16.730 Project Maven, a contract from the U .S. Department 00:18:16.730 --> 00:18:20.269 of Defense. The project used AI to analyze drone 00:18:20.269 --> 00:18:22.970 footage, primarily for tasks like automatically 00:18:22.970 --> 00:18:25.650 identifying vehicles and infrastructure. And 00:18:25.650 --> 00:18:28.410 this immediately triggered internal revolt among 00:18:28.410 --> 00:18:30.849 Google employees, raising fundamental questions 00:18:30.849 --> 00:18:34.230 about the militarization of AI. It did. The company 00:18:34.230 --> 00:18:36.970 tried to frame it as non -offensive, merely analytical 00:18:36.970 --> 00:18:39.920 work. However, the fear among employees and the 00:18:39.920 --> 00:18:42.559 public was clear. Was Google helping accelerate 00:18:42.559 --> 00:18:44.759 the development of autonomous weapons systems? 00:18:45.220 --> 00:18:48.500 This is where leaked internal emails showed her 00:18:48.500 --> 00:18:51.259 private communications, and they were very revealing 00:18:51.259 --> 00:18:53.279 about her internal dilemma. It's interesting. 00:18:53.359 --> 00:18:55.819 I wonder how effective it is to set a moral boundary 00:18:55.819 --> 00:18:58.960 for yourself if you're also, as chief scientist, 00:18:59.180 --> 00:19:01.880 overseeing the creation and democratization of 00:19:01.880 --> 00:19:04.279 the very foundational tools like sophisticated 00:19:04.279 --> 00:19:07.059 image classification and deep learning frameworks. 00:19:07.109 --> 00:19:09.589 works that make autonomous weapon systems possible 00:19:09.589 --> 00:19:12.210 for anyone, including other governments or contractors. 00:19:12.529 --> 00:19:14.910 Did the democratizing effort at Google Cloud 00:19:14.910 --> 00:19:17.769 potentially contradict her ethical stance? That's 00:19:17.769 --> 00:19:19.950 the core tension in her position, and it really 00:19:19.950 --> 00:19:22.390 reflects the complexity of the modern AI landscape. 00:19:22.730 --> 00:19:25.349 The email showed she was enthusiastic about the 00:19:25.349 --> 00:19:27.869 Google Cloud commercial role democratizing AI 00:19:27.869 --> 00:19:30.829 for good, but she specifically warned against 00:19:30.829 --> 00:19:33.789 mentioning the AI component in relation to Project 00:19:33.789 --> 00:19:37.339 Maven. Why? Why that distinction? It comes down 00:19:37.339 --> 00:19:40.180 to public perception. She recognized that military 00:19:40.180 --> 00:19:42.920 A .I. in the public mind is inexorably linked 00:19:42.920 --> 00:19:45.519 to autonomous weapons, which she views as crossing 00:19:45.519 --> 00:19:48.700 a clear moral line. So why did she specifically 00:19:48.700 --> 00:19:51.240 single out the public perception of autonomous 00:19:51.240 --> 00:19:53.940 weapons rather than other military applications 00:19:53.940 --> 00:19:56.799 like logistical analysis or intelligence gathering? 00:19:56.980 --> 00:19:59.259 I think it comes down to agency and human control. 00:19:59.849 --> 00:20:02.730 The debate over autonomous weapons systems, killer 00:20:02.730 --> 00:20:05.690 robots, is fundamentally about removing the human 00:20:05.690 --> 00:20:08.430 from the decision loop of lethal force. For her, 00:20:08.470 --> 00:20:10.509 that is the ultimate failure of human -centered 00:20:10.509 --> 00:20:12.990 AI. It's the point where AI is deployed to harm 00:20:12.990 --> 00:20:16.130 humans without human final arbitration. So when 00:20:16.130 --> 00:20:18.390 the internal emails were publicized, she issued 00:20:18.390 --> 00:20:20.809 a public statement clarifying her stance, stating, 00:20:22.009 --> 00:20:24.609 I believe in human -centered AI to benefit people 00:20:24.609 --> 00:20:27.609 in positive and benevolent ways. It is deeply 00:20:27.609 --> 00:20:29.930 against my principles to work on any project 00:20:29.930 --> 00:20:32.630 that I think is to weaponize AI. That public 00:20:32.630 --> 00:20:34.730 commitment drew a clear line in the sand for 00:20:34.730 --> 00:20:37.190 the industry. It did. And while Google internally 00:20:37.190 --> 00:20:39.789 defended the contract, they ultimately did not 00:20:39.789 --> 00:20:42.230 seek renewal of the Project Maven contract in 00:20:42.230 --> 00:20:45.329 June 2018, just before she returned to Stanford. 00:20:45.880 --> 00:20:48.839 This episode wasn't just a personal choice. It 00:20:48.839 --> 00:20:51.339 was a high -profile, industry -shaping moment 00:20:51.339 --> 00:20:53.980 that solidified her reputation as a formidable 00:20:53.980 --> 00:20:56.700 ethical advocate willing to stake her career 00:20:56.700 --> 00:20:59.539 on her principles. Moving from academia and policy, 00:20:59.799 --> 00:21:02.059 it seems Dr. Lay has now turned her incredible 00:21:02.059 --> 00:21:04.880 energy toward market creation. Section 5 covers 00:21:04.880 --> 00:21:07.140 current ventures and global governance, showing 00:21:07.140 --> 00:21:09.480 how this leading academic is also taking a bold 00:21:09.480 --> 00:21:11.779 entrepreneurial path. This is perhaps the most 00:21:11.779 --> 00:21:14.480 surprising dimension of her recent career. She 00:21:14.480 --> 00:21:16.759 is currently on a partial academic leave from 00:21:16.759 --> 00:21:20.019 Stanford from early 2024 through the end of 2025, 00:21:20.299 --> 00:21:22.859 specifically to focus on her entrepreneurial 00:21:22.859 --> 00:21:25.799 endeavors. She is putting her scientific philosophy 00:21:25.799 --> 00:21:28.859 directly into commercial practice. And her startup, 00:21:29.059 --> 00:21:32.240 World Labs, is one of the most successful ventures 00:21:32.240 --> 00:21:34.839 we have seen launched recently. It's explosive 00:21:34.839 --> 00:21:38.859 growth. World Labs was co -founded in 2024. They 00:21:38.859 --> 00:21:42.059 managed to raise an astronomical $230 million 00:21:42.059 --> 00:21:45.759 in seed funding. And even more incredibly, the 00:21:45.759 --> 00:21:49.559 company was valued at over $1 billion, the benchmark 00:21:49.559 --> 00:21:52.299 for unicorn status, within just four months of 00:21:52.299 --> 00:21:54.930 its launch. This speed highlights the market's 00:21:54.930 --> 00:21:57.170 intense anticipation for her next technical move. 00:21:57.390 --> 00:21:59.950 So after pioneering computer vision, what is 00:21:59.950 --> 00:22:01.829 the next frontier that World Labs is tackling? 00:22:01.990 --> 00:22:04.990 What exactly is spatial intelligence? Well, if 00:22:04.990 --> 00:22:07.309 ImageNet focused on 2D classification, recognizing 00:22:07.309 --> 00:22:10.269 a cat or sign in a static image, spatial intelligence 00:22:10.269 --> 00:22:12.650 is about understanding and reasoning about the 00:22:12.650 --> 00:22:15.210 three -dimensional dynamic physical world. It 00:22:15.210 --> 00:22:17.450 requires integrating perception with action and 00:22:17.450 --> 00:22:19.910 context. How does this differ technically from 00:22:19.910 --> 00:22:22.980 current AI that uses 3D models? It's an order 00:22:22.980 --> 00:22:26.079 of magnitude more complex. Simple 3D models only 00:22:26.079 --> 00:22:29.259 map geometry. Spatial intelligence goes beyond 00:22:29.259 --> 00:22:31.660 identifying what an object is and where it is 00:22:31.660 --> 00:22:33.779 to understanding its physical properties, its 00:22:33.779 --> 00:22:36.400 potential interactions, and its temporal relationship 00:22:36.400 --> 00:22:38.559 to everything else. Can you elaborate on the 00:22:38.559 --> 00:22:40.759 technical inputs required? It must be more than 00:22:40.759 --> 00:22:43.900 just camera footage. It absolutely is. This requires 00:22:43.900 --> 00:22:46.920 integrating multiple modalities, not just standard 00:22:46.920 --> 00:22:49.839 optical cameras, but depth maps from sensors 00:22:49.839 --> 00:22:53.099 like LiDAR and structured light. The AI needs 00:22:53.099 --> 00:22:55.640 to not only see the object but understand its 00:22:55.640 --> 00:22:58.400 mass, its material properties. Is it glass, wood, 00:22:58.619 --> 00:23:01.200 or fabric? And how it will react if you push 00:23:01.200 --> 00:23:04.099 it. So it needs to predict physics. It requires 00:23:04.099 --> 00:23:06.339 four -dimensional reasoning, incorporating the 00:23:06.339 --> 00:23:09.109 element of time. The AI needs to predict physics, 00:23:09.269 --> 00:23:11.549 yes. So instead of merely classifying a chair, 00:23:11.769 --> 00:23:14.430 the AI system understands that the chair affords 00:23:14.430 --> 00:23:17.309 sitting, that it can be moved, that it will fall 00:23:17.309 --> 00:23:19.890 if pushed off a ledge, and that it occupies specific 00:23:19.890 --> 00:23:22.170 volume and space relative to a moving person. 00:23:22.640 --> 00:23:25.200 That is precisely the goal. The sources indicate 00:23:25.200 --> 00:23:28.019 World Labs aims to enable robotic systems to 00:23:28.019 --> 00:23:30.819 perform complex everyday tasks based on natural 00:23:30.819 --> 00:23:33.720 verbal instructions. She described the effort 00:23:33.720 --> 00:23:36.960 as aiming for more human like reasoning, merging 00:23:36.960 --> 00:23:39.559 high level cognition with physical embodiment 00:23:39.559 --> 00:23:42.380 and utility. It's the essential technical leap 00:23:42.380 --> 00:23:45.539 needed for truly useful general purpose robotics. 00:23:45.980 --> 00:23:48.240 This blend of cutting edge entrepreneurship and 00:23:48.240 --> 00:23:51.099 foundational research is impressive, but she 00:23:51.099 --> 00:23:53.390 hasn't abandoned in global governance either. 00:23:53.569 --> 00:23:56.410 She's simultaneously working at the highest international 00:23:56.410 --> 00:23:59.349 level. Her influence spans the public and private 00:23:59.349 --> 00:24:02.589 sectors. In August 2023, she was appointed to 00:24:02.589 --> 00:24:04.670 the United Nations Scientific Advisory Board, 00:24:04.869 --> 00:24:07.630 established by Secretary General Antonio Guterres. 00:24:07.890 --> 00:24:10.230 What is the specific mandate of this UN board? 00:24:10.529 --> 00:24:12.690 The board's role is critical. It offers independent 00:24:12.690 --> 00:24:14.690 perspectives on emerging trends that intersect 00:24:14.690 --> 00:24:17.730 science, technology, ethics, governance and sustainable 00:24:17.730 --> 00:24:20.950 development. As AI and biotech advance so rapidly, 00:24:21.210 --> 00:24:23.789 the UN needs a core group of top scientists to 00:24:23.789 --> 00:24:25.930 translate these technical changes into usable 00:24:25.930 --> 00:24:28.660 policy. advice from member states. So Dr. Lai 00:24:28.660 --> 00:24:31.839 is essentially advising the world body on how 00:24:31.839 --> 00:24:34.259 to responsibly handle the very technology she 00:24:34.259 --> 00:24:36.839 is pioneering. That's right. Given her dual roles, 00:24:36.980 --> 00:24:39.559 building a billion -dollar company and advising 00:24:39.559 --> 00:24:43.039 the UN, what is her most consistent policy stance 00:24:43.039 --> 00:24:46.779 regarding AI governance? Her primary policy concern 00:24:46.779 --> 00:24:49.440 revolves around the profound imbalance in investment. 00:24:49.849 --> 00:24:52.430 She advocates strongly for greater public funding 00:24:52.430 --> 00:24:55.849 for scientific AI uses and risk assessment. She's 00:24:55.849 --> 00:24:57.690 noted that the immense private sector investment, 00:24:57.849 --> 00:25:00.269 exemplified by the hundreds of millions raised 00:25:00.269 --> 00:25:03.869 by world labs and similar ventures, it just dwarfs 00:25:03.869 --> 00:25:06.250 the public money available for foundational independent 00:25:06.250 --> 00:25:09.009 research into safety, alignment, and ethical 00:25:09.009 --> 00:25:11.869 oversight. So the engines of innovation are running 00:25:11.869 --> 00:25:14.049 at full speed in the private sector, but the 00:25:14.049 --> 00:25:16.210 engines of safety and policy are starved for 00:25:16.210 --> 00:25:18.269 resources in the public and academic sectors. 00:25:18.450 --> 00:25:20.809 Exactly. And this imbalance is a risk to global 00:25:20.809 --> 00:25:23.250 stability. Furthermore, her governance philosophy 00:25:23.250 --> 00:25:26.910 is intensely pragmatic. In February 2025, she 00:25:26.910 --> 00:25:28.829 addressed the Artificial Intelligence Action 00:25:28.829 --> 00:25:31.990 Summit in Paris, making a strong appeal to global 00:25:31.990 --> 00:25:34.349 policymakers. What was the essence of that appeal? 00:25:34.700 --> 00:25:37.160 She urged policymakers that AI governance must 00:25:37.160 --> 00:25:39.460 be based on science rather than on science fiction. 00:25:39.720 --> 00:25:41.839 She was criticizing the tendency to regulate 00:25:41.839 --> 00:25:44.720 based on sensational, often exaggerated existential 00:25:44.720 --> 00:25:48.039 threats rather than on the measurable objective 00:25:48.039 --> 00:25:50.740 capabilities and current limitations of the technology. 00:25:51.079 --> 00:25:52.960 She called for a far more rigorous scientific 00:25:52.960 --> 00:25:55.960 approach to objectively assessing AI's capacity 00:25:55.960 --> 00:25:59.019 and potential risks. She did. If we want effective 00:25:59.019 --> 00:26:01.920 regulation, we must first truly understand the 00:26:01.920 --> 00:26:04.279 science underpinning the capabilities. are trying 00:26:04.279 --> 00:26:07.079 to manage. That is a crucial distinction. It 00:26:07.079 --> 00:26:09.400 argues that emotional responses should not replace 00:26:09.400 --> 00:26:12.420 data -driven risk assessment when defining regulations 00:26:12.420 --> 00:26:14.920 that will shape the future of global technology. 00:26:15.279 --> 00:26:18.839 What an extraordinary deep dive. The life and 00:26:18.839 --> 00:26:21.440 career of Dr. Fei -Fei Li offer a compelling 00:26:21.440 --> 00:26:24.079 narrative that connects intense technical rigor 00:26:24.079 --> 00:26:27.339 with unwavering ethical responsibility. Let's 00:26:27.339 --> 00:26:29.099 bring it all back together for you, the listener. 00:26:29.400 --> 00:26:31.640 We've traced a continuous thread in her work. 00:26:31.759 --> 00:26:34.359 She began by asking a fundamental question rooted 00:26:34.359 --> 00:26:37.039 in human psychophysics. What does it take to 00:26:37.039 --> 00:26:39.279 perceive the world? Her answer was ImageNet. 00:26:39.480 --> 00:26:42.039 The first core takeaway is the magnitude of the 00:26:42.039 --> 00:26:45.460 data set she engineered. 14 million images mapped 00:26:45.460 --> 00:26:48.480 onto the WordNet hierarchy. This structure and 00:26:48.480 --> 00:26:50.740 scale solved the critical bottleneck in data 00:26:50.740 --> 00:26:53.380 availability, enabling the deep learning revolution 00:26:53.380 --> 00:26:56.460 and giving AI sight. Then the second pillar, 00:26:56.640 --> 00:26:58.579 woven throughout her career in leadership at 00:26:58.579 --> 00:27:01.279 SAIL and the founding of HAI, is the relentless 00:27:01.279 --> 00:27:04.519 push for diversity and human relevance. Her nonprofit 00:27:04.519 --> 00:27:08.339 work with AI4AL, scaling from the Precursor Sailors 00:27:08.339 --> 00:27:11.000 program, is dedicated to diversifying the talent 00:27:11.000 --> 00:27:12.900 pool of creators, recognizing that structural 00:27:12.900 --> 00:27:15.400 inclusion is essential to combat systemic bias 00:27:15.400 --> 00:27:17.980 in the resulting AI systems. And finally, her 00:27:17.980 --> 00:27:20.259 current focus is defining the next stage of AI 00:27:20.259 --> 00:27:23.460 capability, spatial intelligence. Through world 00:27:23.460 --> 00:27:25.880 labs, she's moving AI beyond 2D classification 00:27:25.880 --> 00:27:29.000 into 4D reasoning, enabling systems to understand 00:27:29.000 --> 00:27:30.859 the physical world in terms of action, physics, 00:27:31.000 --> 00:27:34.880 and context. highly commercial, is still bound 00:27:34.880 --> 00:27:36.779 by the human -centered principles she defended 00:27:36.779 --> 00:27:38.779 during the pivotal Project Maven controversy, 00:27:39.119 --> 00:27:41.440 and which she now promotes at the United Nations. 00:27:41.640 --> 00:27:43.759 And if we connect this to the bigger picture. 00:27:44.349 --> 00:27:47.210 Her career trajectory shows a consistent effort 00:27:47.210 --> 00:27:49.569 to ensure that the monumental technological capability 00:27:49.569 --> 00:27:52.509 she pioneered is paired with a clear moral compass. 00:27:52.769 --> 00:27:56.170 She gave AI vision with ImageNet, and every subsequent 00:27:56.170 --> 00:27:58.170 move, whether in policy or entrepreneurship, 00:27:58.509 --> 00:28:01.269 has been dedicated to ensuring that AI also gains 00:28:01.269 --> 00:28:03.930 a conscience. That dual mission site and conscience 00:28:03.930 --> 00:28:06.650 is what makes her a true architect of the AI 00:28:06.650 --> 00:28:09.950 age. Now, for our final provocative thought for 00:28:09.950 --> 00:28:13.220 you, the listener. Dr. Lee argues that AI governance 00:28:13.220 --> 00:28:15.619 should be based on science rather than science 00:28:15.619 --> 00:28:18.519 fiction. Given that her work in spatial intelligence 00:28:18.519 --> 00:28:21.180 demands objective, measurable understanding of 00:28:21.180 --> 00:28:24.019 complex 4D systems, what specific scientific 00:28:24.019 --> 00:28:26.900 metrics, not philosophical fears, but quantifiable, 00:28:27.119 --> 00:28:29.799 repeatable data points could be used to objectively 00:28:29.799 --> 00:28:32.579 measure the safety, functional limitations, and 00:28:32.579 --> 00:28:34.640 potential biases of spatial intelligence as it 00:28:34.640 --> 00:28:36.960 begins to navigate and act within our complex 00:28:36.960 --> 00:28:39.289 physical world? Something to mull over as we 00:28:39.289 --> 00:28:41.529 move into a future where AI systems are no longer 00:28:41.529 --> 00:28:43.930 just observing, but actively performing tasks 00:28:43.930 --> 00:28:44.789 all around us.