#839: Dr. Fei-Fei Li, The Godmother of AI — Asking Audacious Questions, Civilizational Technology, and Finding Your North Star ( #839)

0:00:04 Hello, boys and girls, ladies and germs. This is Tim Ferriss. Welcome to another episode of
0:00:09 The Tim Ferriss Show, where it’s my job to deconstruct world-class performers. I interview
0:00:14 them to tease out the habits, routines, frameworks, etc., that you can apply to your own lives.
0:00:19 My guest today is Dr. Fei-Fei Li. She is the inaugural Sequoia professor in the computer
0:00:25 science department at Stanford University. She’s been called the godmother of AI. She’s
0:00:31 a founding co-director of Stanford’s Human-Centered AI Institute and the co-founder and CEO of World
0:00:36 Labs, a generative AI company focusing on spatial intelligence. She’s also the author of The
0:00:43 World’s I See, Curiosity, Exploration, and Discovery at the Dawn of AI, her memoir, and one of Barack
0:00:50 Obama’s recommended books on AI and a Financial Times Best Book of 2023. Her story is incredible.
0:00:57 It is one of beating the odds on so many different levels. And let’s get straight to it. Dr. Fei-Fei Li.
0:01:03 Optimal minimal. At this altitude, I can run flat out for a half mile before my hands start shaking.
0:01:07 Can I ask you a personal question? Now would’ve seen a perfect time.
0:01:13 What if I did the opposite? I’m a cybernetic organism living tissue over a metal endoskeleton.
0:01:25 Dr. Fei-Fei Li, it is nice to see you. Thanks for making the time.
0:01:27 Hi, Tim. Very nice to be here. Very excited.
0:01:34 And we were chatting a little bit before we started recording about how miraculous and I suppose
0:01:39 unfortunate it is that somehow we managed to spend three years on the same campus and didn’t bump into
0:01:45 each other. I know. And now I’m wondering which college you were at and which clubs.
0:01:50 Oh, yeah. I was Forbes. I was in Forbes College. No, I was Forbes too.
0:01:55 Okay. This is for people who don’t know what the hell we’re talking about. There are these
0:02:01 residential colleges where students are split up when they come into the school. And Forbes was way out
0:02:10 there in the sticks right next to a fast food spot like 7-Eleven called Wawa and next to the commuter
0:02:16 train. And then there’s something called eating clubs at Princeton. People can look them up, but they’re
0:02:21 effectively co-ed fraternity slash sororities where you also eat unless you want to make your own meals.
0:02:23 And I was in Terrace.
0:02:30 I was not any of that. But for those of you wondering why we didn’t meet, we should say we were very studious
0:02:32 students who are only in the libraries.
0:02:40 Yeah. We were very studious. I actually made my, whatever it was, $6 an hour at guest library working
0:02:40 up in the attic.
0:02:45 Tim, I work in the same library. I don’t understand why we did not meet.
0:02:49 Hilarious. Okay. So, well, now we’re meeting. Now we’re, I mean, we’ve met.
0:02:52 Did you change name or something? Maybe we did meet.
0:02:58 I did. Didn’t change my name, but here we are. We’ve reunited. That’s wild that we didn’t bump into
0:03:04 each other. I was also gone for a period of time because I went to Princeton in Beijing and went to
0:03:11 the, what was it? Capital University of Business and Economics after that. So I was gone for a good
0:03:18 period of time and then took a year off before graduating with the class of 2000. Still,
0:03:25 we had a lot of overlap, but let’s hop into the conversation. And this is a very perhaps
0:03:31 typical way to start. But in your case, I think it’s a good place to start, which is just with the
0:03:39 basics chronologically. Where did you grow up and could you describe your upbringing? Because based on
0:03:45 my reading, your parents were pretty atypical for Chinese parents in my experience, certainly.
0:03:47 Yes. And you know a lot.
0:03:49 Yeah. Could you speak to that, please?
0:03:56 Yeah. I would say my childhood and leading up to the formative years is a tale of two cities.
0:04:03 I grew up in a town in China called Chengdu. I was born in Beijing, but most of my childhood was
0:04:12 spent in Chengdu where it’s very famous for panda bears. And at the age of 15, my mom and I joined my dad
0:04:22 in a town called Persephone, New Jersey. So I went from a relatively typical middle-class Chinese kid
0:04:29 to become a new immigrant in a completely different world of all places, New Jersey,
0:04:37 and to learn a new language, to learn a new culture, to embrace a new country. And then from there on,
0:04:44 I went to Princeton as a physics major, but I did take some of the classes you took.
0:04:50 And then I went to Caltech as a PhD student to study AI, and the rest is history.
0:04:55 I want to hear about both your parents, but I want to hear a little bit about your dad,
0:05:04 because he seems like, based on my reading, a very whimsical, sort of creative soul, which is a sharp
0:05:12 contrast in some ways to, for instance, I had Bo Xiao on the podcast, amazing entrepreneur. And his father
0:05:18 was, I suppose, what some folks might think of when they imagine, not a tiger mom, but like a tiger dad.
0:05:26 So in the case of Bo’s upbringing, his father is very strict, but if he, meaning Bo, won a math competition,
0:05:30 then he would get extra love, and he would be allowed to have certain treats and things like that.
0:05:33 Could you just describe your parents a little bit?
0:05:38 Yeah, so first of all, clearly you read my book. Thank you for that.
0:05:45 It is true. As a child, you don’t realize that. As I was just going through my own science
0:05:50 memory, I was writing it. The more I wrote about it, the more I realized, oh my God, I really did not
0:06:00 have a typical dad. My dad loved and still loves nature. He’s just a curious mind. He finds humor and
0:06:10 fun in unserious things, you know, like he loves bugs, insects. He loves taking me as a kid. Growing up in
0:06:19 the 1980s in China, there isn’t much abundance in terms of material resources. So my city, Chengdu,
0:06:25 was expanding. So we lived in apartment complexes at the edge of the city, even though my dad and my mom
0:06:32 worked in the middle of the city. So on the weekends, my dad and I would just play in the fields where
0:06:40 there’s still rice fields. There’s water buffaloes. I had a puppy. And my dad would just, really all my
0:06:48 memory is just like finding bugs, really. And then sometimes my dad and I will follow some, I don’t
0:06:54 know, we took an art class. I took a kid’s art class. I will go to the mountains, neighboring mountains
0:07:05 to draw. But my entire childhood memory of my dad is, is just a very unserious parent who had no interest
0:07:14 in my grades or what I’m doing in class. Did I achieve anything? Did I bring back any like competition
0:07:21 awards? Nothing to do with that. Even when I came to New Jersey with my parents,
0:07:30 life became extremely tough. It was immigrant life. We were in a lot of poverty. And even that,
0:07:37 my memory is that he has so much fun in yard sales. I would just go to yard sales. And those are our
0:07:45 every, every weekend. It was just, yay, let’s go to yard sales and just use that as a treasure hunt
0:07:52 almost. So he’s a very curious and childlike mind in that way.
0:07:57 So Matt, I’m asking about your parents in part because I know you’re a parent and ultimately I’m
0:08:03 going to want to ask how you think about parenting. That will come up at some point. But since listeners
0:08:07 will certainly be asking themselves this question, and we’re not going to get into any geopolitics
0:08:11 because there are plenty of people who want to get into that and fight over that, which we’re not
0:08:17 going to do. But why did your parents leave China? Like what was the catalyst or what were the reasons
0:08:24 behind leaving what you knew or leaving what they knew and coming to a very different foreign country?
0:08:31 I mean, you’re going from Chengdu, which is a city to suburban New Jersey, which is, as I think
0:08:36 you’ve described it, felt very empty, right? And then you have the language barriers and the financial
0:08:38 barriers. There’s so many things. Why the move?
0:08:48 I’ll give you two answers. Early teenage Fei Fei would say, I have no idea because my dad left when I was
0:08:55 12 and my mom and I joined him when I was 15. And those years, you’re a teenager, right? Like there’s
0:09:03 so many strange things in your head. And all I knew is that, you know, they said, let’s go to America.
0:09:11 And I had no idea. I really did not know what happened. There was this vague sense of there’s
0:09:19 opportunities and freedom. Education is very different. And I had a hunch that I was not a
0:09:27 typical kid in the sense that, you know, I was a girl and I loved physics. I loved fighter jets of
0:09:38 all things. I can tell you all the fighter jets I love from F-117 to F-16 to, you know, to all the
0:09:46 different things that I loved. So that’s all I knew. In hindsight, as a grownup Fei Fei, I
0:09:56 appreciated my parents. They’re very brave people because I don’t know this age myself would just
0:10:04 pick up and leave a country I’m familiar with and go to, I don’t know, a completely different country
0:10:11 that I speak zero language and I have zero connectivity to. And mind you, that’s pre-internet,
0:10:16 pre-AI age. So when you are going to a different country, you might as well go to a different
0:10:22 planet. You’re cut off. Yeah. So I think they’re very brave. The grownup Fei Fei realized that they
0:10:30 wanted me to have an opportunity that they think will be unprecedented for my education.
0:10:38 and they turned out that’s kind of true. Yeah. Well, certainly looking at your bio, I mean,
0:10:47 it’s mind boggling to imagine all of the different sliding door events and different paths you could
0:10:52 have taken. So we’re going to hop pretty closely along chronologically, but we’re going to ultimately
0:10:58 get to a lot of the meat and potatoes of the conversation. But I want to touch on maybe some
0:11:04 other formative figures. And I would like to hear about your mother as well, because just with the
0:11:09 context of your dad, it’s like, okay, that seems fascinating and very unusual, particularly if you’ve
0:11:13 spent any time in China, especially during that period of time.
0:11:19 He is very unusual that way. Yeah, very unusual. So then people might wonder, well, where does the
0:11:24 drive come from? Where does the technical focus come from? And I’d love to hear your answer to that
0:11:32 and also hear you explain who Bob Sabella was, if I’m pronouncing that correctly.
0:11:38 Yes. Yeah. Yeah. There are two questions. Mostly, is my mom the one who puts in the drive and the
0:11:46 technical passion and what role did Bob play in my life. So first one, first of all, my mom has zero
0:11:53 technical genes. She really has no, I sometimes still laugh at her. She cannot do math. Let’s put
0:11:59 it this way. So I think the technical passion is just, I was born with it. My dad is more technical,
0:12:07 but he’s, you know, he loves bugs more than insects, more than equations for sure. So I think that’s,
0:12:15 you know, as an educator for so many decades now myself, and also as a parent, you have to respect
0:12:24 the wonders of nature. There is this inner love and fire and passion and curiosity that comes with
0:12:30 the package. But my mom is much more disciplined person. She’s still not a tiger mom in the sense,
0:12:38 I don’t remember my mom ever going after me on grades, or she really did not. Both my parents
0:12:46 never, ever cared about me bringing any awards home. Maybe I did, maybe I didn’t, but I can tell you
0:12:54 in our house, there’s zero wall hangings of anything, which actually carry to today. Even for myself,
0:13:02 my own house, my own office have zero of those decorations of achievements or awards. It’s just,
0:13:09 my mom did not care about that. But she did care about me being a focus person. If I want to do
0:13:15 something, she doesn’t want me to play while doing homework. And that kind of thing would bother her.
0:13:23 She would say, just finish your homework, say by 6pm. If you don’t finish your homework, you’re not allowed
0:13:28 to do more homework, you have to deal with the consequences. So she instilled some discipline. But that’s
0:13:34 about it. She’s tougher than my dad. She is very rebellious. She had an unfinished dream herself. She was
0:13:42 very academic when she was a kid herself. And cultural revolution really crushed all her dreams. So she became
0:13:50 a more rebellious person in that sense that I think I did observe and experience as a daughter.
0:13:59 So maybe part of immigration is even part of that. She has this, many years later, she would say,
0:14:06 I had no plan coming to New Jersey, but I think I’m going to survive. I just believe I’m going to survive.
0:14:13 And I’m going to make sure Fei-Fei survives. I think that is her strength, her stubbornness,
0:14:15 and her rebelliousness.
0:14:18 When does Bob enter the picture? And who is Bob?
0:14:25 Bob Sabella was a high school math teacher in Persephone High School. He was my own math teacher,
0:14:33 as well as many, many students. He entered my life. So it’s kind of bordering sophomore to junior year
0:14:42 in Persephone High School when I started taking AP calculus. But he quickly became the most influential
0:14:50 person in my formative years as a new American immigrant, as a teenager, because he became my mentor,
0:15:01 my friend, and eventually his entire family became my American family. And he became my friend when I was a
0:15:10 very lonely ESL, English as second language student. I was excelling in math, but I think it’s more because
0:15:18 I was lonely and he was very friendly. He treated me more like a friend who talks about books we love,
0:15:25 talk about the culture, talks about science fiction, and also listened to me as a very, I wouldn’t say
0:15:33 confused. But a teenager undergoing a lot of life’s turmoil in my unique circumstance. And that
0:15:42 unconditional support made me very close to him and his family. And one thing he did to me that I did
0:15:51 not appreciate till later is that when Persephone High School couldn’t offer a full calculus BC class,
0:15:59 because it just didn’t have that. He just sacrificed his lunch hour, his only lunch hour, to teach me calculus BC.
0:16:08 So it was a one-to-one class. And I’m sure that contributed to me, an immigrant kid, getting to Princeton
0:16:17 eventually. But later, as I became teacher myself, it’s exhausting to teach all day long. And the fact that
0:16:26 on top of that, he would use his lunch hours to do that extra class for me is just such a gift that I now
0:16:29 appreciate more than I was as a teenager.
0:16:35 Yeah. Thank God for the teachers who go the extra mile. It’s just incredible, especially when you
0:16:39 get a bit older and you have more context and you can look back and realize.
0:16:47 I really think these public teachers in America are the unsung heroes of our society because
0:16:53 they are dealing with kids of all backgrounds. They’re dealing with the changing times.
0:17:00 the kind of stories Bob would share with me in terms of how he went extra miles, not just with me,
0:17:07 but with many students. Because Persephone is a heavily immigrant town. So his students are
0:17:14 from all over the world and how he helped them and their family. It’s just, those are the stories
0:17:19 that people don’t write about. And that’s part of the reason I wrote the book was to celebrate a teacher
0:17:21 like that. Yeah.
0:17:28 Just a quick thanks to our sponsors and we’ll be right back to the show.
0:17:34 I have been fascinated by the microbiome and probiotics as well as prebiotics for decades,
0:17:40 but products never quite live up to the hype. Now things are starting to change. And that includes
0:17:48 this episode’s sponsor, Seeds DSO-1 Daily Symbiotic. I’ve always been very skeptical of most probiotics,
0:17:53 but after incorporating two capsules of Seeds DSO-1 into my morning routine, I have noticed improved
0:18:00 digestion and improved overall health. So why is Seeds DSO-1 so effective? For one, it is a two-in-one
0:18:07 probiotic and prebiotic formulated with 24 clinically and scientifically studied strains. But if the probiotic
0:18:11 strains don’t make it to the right place, they’re not as effective. So Seeds developed a proprietary
0:18:17 capsule-in-capsule delivery system that survives digestion and delivers a precision release of the
0:18:23 live and viable probiotics to the colon. And now you can get 20% off your first month with code 20Tim
0:18:34 at Seed.com slash Tim using code 20Tim all put together. One more time, Seed.com slash Tim, code 20Tim.
0:18:40 As many of you know, for the last few years, I’ve been sleeping on a Midnight Luxe mattress from
0:18:45 today’s sponsor, Helix Sleep. I also have one in the guest bedroom downstairs and feedback from friends
0:18:50 has always been fantastic. It’s something they comment on without any prompting from me whatsoever.
0:18:57 I also recently had a chance to test the Helix Sunset Elite. The Sunset Elite delivers exceptional
0:19:02 comfort while putting the right support in the right spots. It is made with five tailored foam layers,
0:19:06 including a base layer with full perimeter zoned lumbar support right where I need it,
0:19:11 and middle layers with premium foam and microcoils that create a soft contouring feel.
0:19:18 Helix offers a 100-night sleep trial, fast, free shipping, and a 15-year warranty. So check it all out.
0:19:26 And now you can get 20% off anything on their website, so site-wide. So just go to HelixSleep.com
0:19:33 slash Tim. One more time, HelixSleep.com slash Tim. With Helix, better sleep starts now.
0:19:40 I have so much I want to cover, and I know we’re going to run out of time before we run out of
0:19:44 topics. So I want to spend more time on Bob, and at the same time, I want to keep the
0:19:51 conversation moving. So we’re going to do that, and I’ll just perhaps hit on a few things and then
0:19:59 dig into a number of questions. But certainly at Princeton, you, but also your entire family,
0:20:03 had to survive. So you were involved with operating a dry cleaning shop in New Jersey
0:20:11 as one option, right? You ran that for seven years. So through that, I mean, you gained perspective on,
0:20:16 it feels like you’ve gained perspective on many different levels that have then helped inform
0:20:21 what you’ve done professionally, right? So you learn to think about not just people who are
0:20:26 protected in an ivory tower, but people all the way down and across in society. So from every swath
0:20:33 of society. Your mother also, although she was not technical, she imbued in you this discipline and also
0:20:41 seems to have had a very broad appreciation and knowledge of literature and international literature.
0:20:48 So now you have this global perspective, presumably at the time in Chinese, and you end up at Princeton.
0:20:56 And I know we’re going to be hopping around quite a bit, but I’m curious to know how ImageNet came
0:21:01 about. You can introduce this any way you like. You can tell people what it is and what it became and
0:21:05 why it’s important and then talk about how it started, or you can just talk about how it started.
0:21:15 Yeah. So let me just explain what ImageNet is. ImageNet on the surface was built in between 2007
0:21:22 and 2009 when I was an assistant professor at Princeton, and then I moved to Stanford. So during
0:21:30 this transitional time, my student and I built this, at that time, the field of AI’s largest training and
0:21:37 benchmarking dataset for computer vision or visual intelligence. The significance today, after almost
0:21:46 20 years of ImageNet was, it was the inflection point of big data. Before ImageNet, AI as a field was not
0:21:52 working on big data. And because of that, and a couple of other reasons, which I’ll get into,
0:21:58 AI was stagnating. The public thinks that was the AI winter, even though as a researcher,
0:22:04 young researcher at that time, it was the most exciting field for me, but I get it that it wasn’t
0:22:13 showing breakthroughs that the public needs. But ImageNet, together with two other modern computing
0:22:20 ingredients, one is called neural network algorithm. The other one is modern chips called GPU,
0:22:31 Graphic Processing Unit. These three things converged in a seminal work, milestone work in 2012 called
0:22:39 ImageNet Classification, Deep Convolutional Neural Network Approach. That was a paper that a group of
0:22:48 scientists did to show that the combination of large data by ImageNet, fast parallel computing by GPUs,
0:22:56 and a neural network algorithm could achieve AI performances in the field of image recognition
0:23:03 in a way that’s historically unprecedented. And that particular milestone is, many people call it the
0:23:11 birth of modern AI. And my work image that was one third of that, if you count the elements. And I think
0:23:18 that was the significance, I feel very, really very lucky and privileged that my own work was pivotal in
0:23:27 bringing modern AI to life. But the journey to ImageNet was longer than that. The journey to ImageNet started in
0:23:33 Princeton, in Princeton when I was an undergrad. You were in the East Asian study department. I was hiding in
0:23:41 Jadwin Hall, which is our physics department. I loved physics since I was a young kid. I don’t know how
0:23:51 somehow my dad’s love of bugs and insects and nature translated in my head into just the curiosity for the universe. So I loved,
0:23:59 you know, looking to the stars, I loved the speed of fighter jets, and then the intricate engineering of that eventually
0:24:08 translated into the love of the discipline that asks the most audacious question of our civilization, such as
0:24:15 what is the smallest matter? What is the definition of space-time? How big is the universe? What is the
0:24:24 beginning of the universe? In that early teenagehood love, I loved Einstein. I loved his work, and I wanted to go to
0:24:32 Princeton for that. But it turned out what physics taught me was not just the math and physics. It was
0:24:41 really this passion to ask audacious questions. So by the end of my undergrad years, I wanted my own audacious
0:24:47 question. You know, I wasn’t satisfied with just pursuing somebody else’s audacious question. And through
0:24:55 reading books and all that, I realized my passion was not the physical matters. It was more about intelligence.
0:25:03 I was really, really enamored by the question of what is intelligence and how do we make intelligent machines?
0:25:12 So at that time, I swear I did not know it was called AI. I just knew that I wanted to pursue the study of
0:25:19 intelligence and intelligent machines. And then I applied to grad school and I went to Caltech.
0:25:27 Caltech was my PhD. I started in the turn of the century, 2000. And I think I consider that moment I became
0:25:36 a budding AI scientist. That was my formal training as a computer scientist in AI. Then
0:25:45 my physics training continued in the sense that physics taught me to ask audacious questions
0:25:52 and turn them into a North Star. And in scientific terms, that North Star became a hypothesis.
0:26:02 And it was very important for me to define my North Star. And my first North Star for the following years to come
0:26:10 was solving the problem of visual intelligence is how we can make machines see the world. And it’s
0:26:18 not just by seeing the RGB colors or the shades of light. It’s about making sense of what’s seen,
0:26:24 which is, you know, I’m looking at you, Tim. I see you. I see a beautiful painting behind you.
0:26:30 I see you’re sitting on a chair like that is seeing. Seeing is making sense of what this world is.
0:26:38 So that became my North Star question. And that hypothesis that I had is I have to solve object
0:26:45 recognition. And then that was in my entire PhD was the battle with object recognition. There were many,
0:26:53 many mathematical models we have done. And there are many questions. But me and my field was struggling.
0:26:59 We can write papers, no problem, but we did not have a breakthrough. And then luckily for me,
0:27:08 Princeton called me back as a faculty in 2007. It was one of my happiest moments of my life. I feel so
0:27:16 validated my alma mater would consider giving me a faculty job. So I happily moved back to Princeton
0:27:24 as a faculty this time. And I continue to be a Forbes member, actually. So at Princeton,
0:27:33 there was an epiphany is that I realized there was a hypothesis that everybody missed. And that hypothesis
0:27:39 was big data. This is the point that I’m so, so curious about. And I just want to pause for a second.
0:27:44 Also, for people who are interested in some of the history of Princeton, it’s pretty crazy. They should
0:27:50 look up the history of the Princeton Institute for advanced study. And I remember taking some of those
0:27:55 East Asian studies classes that you referred to in classrooms where Einstein taught. And it’s just the
0:28:02 aura, the veneer. You want to believe that you can feel it just permeating the entire campus. And it’s
0:28:08 fun. In that respect, it’s very fun. But I’m going to read something from a Wired piece that discussed
0:28:15 you at length. And as you mentioned, big data before and after in terms of its integration into
0:28:21 the type of research that you’re describing. And as it was written, and please feel free to fact check
0:28:26 this or push back on it. But in Wired, they said, the problem was, a researcher might write one algorithm
0:28:33 to identify dogs and another to identify cats. And then you, it says, you know, Lee, began to wonder if the
0:28:38 problem wasn’t the model, but the data. She thought that if a child learns to see by experiencing the visual
0:28:44 world, by observing countless objects and scenes in her early years, maybe a computer can learn in a similar
0:28:51 way. And I want you to expand on that for sure. And the question for me is like, why did you see it?
0:28:59 Why didn’t it happen sooner? We’re all students of history. One thing I actually don’t like about
0:29:07 the telling of scientific history is there’s too much focus on single genius. Yes, agreed.
0:29:14 We know Newton discovered the modern laws of physics, but yes, he is a genius not to take away any of that
0:29:20 from Newton. But science is a lineage, and science is actually a nonlinear lineage. For example, why did
0:29:28 I see, why was I inspired by this hypothesis of big data? Because many other scientists inspire me. In my
0:29:35 book, I talked about this particular lineage of work by Professor Irv Biederman, who was a psychologist,
0:29:41 who was not interested in AI, but he was interested in understanding minds. And I was reading his paper,
0:29:50 and he particularly was talking about the massive number of visual objects that young children was able
0:29:58 to learn in early ages, right? So that piece of work itself is not image that, but without reading that
0:30:04 piece of work, I would not have formulated my hypothesis. So while I’m proud of what I have done,
0:30:14 one, my book especially wanted to tell the history of AI in a way that so many unsung heroes, so many
0:30:25 generations of scientists, so many cross-disciplinary ideas pollinate each other. So I was lucky at that
0:30:31 time as someone who is passionate about the problem, but also someone who benefited from all this research.
0:30:39 research. So yes, something happened in my brain, but I would really attribute to many things happen across
0:30:48 so many people’s work throughout their lifetime devotion to science that we got to the point of ImageNet.
0:30:54 I’m so glad that you’re underscoring this because if you really dig as a, I don’t consider myself a
0:31:00 scientist, but I love reading about the history of science. There’s so many inputs, so many influences,
0:31:08 so many interdependencies, and the simplicity of the single hero’s journey is appealing, and it’s
0:31:10 simplicity, but it’s almost never true.
0:31:18 It probably is never true. Even my biggest hero Einstein, right? Anybody who knows me, anybody who
0:31:25 read my book knows how much I revere him, and I just love everything he’s done. The special relativity
0:31:34 equation is a continuation of Lawrence’s transform. So even Einstein, he builds upon so many other people’s
0:31:40 work. So I think it’s really important, especially, I’m sure we’ll talk about it. I’m here calling you
0:31:46 in the middle of Silicon Valley, and we’re in the middle of an AI hype. And obviously, I’m very proud
0:31:55 of my field. But I think that when the media or whatever tells the story of AI, it almost always just
0:32:00 talks about a few geniuses. And it’s just not true. It’s generations of computer scientists,
0:32:05 cognitive scientists, and engineers who made this field happen.
0:32:12 Yeah, for sure. I mean, everyone knows Watson and Crick, for instance, but without Rosalind Franklin
0:32:17 and her X-ray crystallography, it doesn’t happen. It doesn’t happen. It just doesn’t happen point
0:32:23 blank. We’re going to hop to modern day in a second. But with ImageNet, I would love for you to speak
0:32:30 to some of the decisions, let’s say decisions, or moments that were just formative and making that
0:32:38 successful. Because for instance, if you’re going to try to allow a machine to, and I’m using very simple
0:32:46 terms because I’m not technical enough to do otherwise, to learn to identify objects closer to the path that
0:32:53 a child would take, you have to label a lot of images, right? So I was reading about how
0:33:01 Mechanical Turk came into play. And then there’s a competitive aspect that seems to have driven some
0:33:08 of the watershed moments. Could you just speak to some of the elements or decisions that made it successful?
0:33:15 A lot of people ask me this question because after ImageNet, many, many people have attempted to make
0:33:23 data sets, but still only very few are successful. So what made ImageNet successful?
0:33:29 I think one of the success was timing is that we truly were the first people who see the impact of big
0:33:37 data. So that very categorical or qualitative change itself is a part of the success. But it’s also,
0:33:46 as you were asking, the hypothesis of big data is not just size. A lot of people actually misunderstands
0:33:53 ImageNet’s significance, as well as other data sets’ significance. Coming with the data set
0:34:03 is a scientific hypothesis of what is the question to ask. For example, in visual recognition, you can make a
0:34:11 data set of discerning RGB, and that would not be as impactful of a data set that is organized around
0:34:19 objects. We can go down the rabbit hole of why. Not because RGB is easier, per se. It’s because you
0:34:26 have to ask the scientific question in the right way. So another example is, instead of making a data set of
0:34:34 objects, why don’t you make a data set of cities? You know, that’s even more complicated than objects. But then
0:34:41 that’s dialing too complicated. So every scientific quest, you have to have the right hypothesis and
0:34:49 asking the right question. So that’s one part of the success is we defined visual object categorization
0:34:58 as the right hypothesis. That was one rightness, I guess. Another rightness is that people just think,
0:35:06 Oh, it’s easy. You just collect a lot of data. Well, first of all, it’s laborious. But even aside from being laborious,
0:35:13 how do you define the quality? You could say, well, if quality is big enough, we don’t care about quality.
0:35:23 But how do you dial between what is big, what is great, what is good? And how do you trade off? That is a deeply scientific
0:35:32 question that we have to do a lot of research on. And then another decision that is a set of decision that
0:35:39 is really hard is, what defines quality in terms of image? Is it every image has higher resolution?
0:35:50 Is it it’s photo realistic? Is it because it’s everyday image that look very cluttered? Is it all product
0:35:57 shots that look clean? And these are questions that if you’re too far away, you wouldn’t even think about
0:36:04 asking. But as a scientist, as we were formulating the deep question of object recognition, we have to ask
0:36:12 this in so many dimensions. And then you mentioned Amazon Mechanical Turk. That is actually a consequence
0:36:22 of desperation. Because when we formulated this hypothesis, our conclusion is we need at least
0:36:34 tens of millions of high quality images across every possible diverse dimension, whether it’s user photos,
0:36:43 or is it product shots, or is it stock photography, like, and then we need also high quality labels.
0:36:52 Once we make that decision, we realize this has to be human filtered from billions of images.
0:36:59 So with that, we became very desperate. We’re like, how are we going to do that? You know, I did try to hire
0:37:05 our Princeton undergrads. And as you know, Princeton undergrads are very smart, but…
0:37:08 They have a very high opinion of the value of their time.
0:37:15 Yes, and they’re expensive. But even if I had all the money in the world, which we didn’t,
0:37:22 it would have taken so long. So we were very, very stuck for very, very long. We thought we had other
0:37:31 shortcuts. But the truth is, human labeling is a gold standard. We want to train machines that are measured
0:37:38 against human capabilities. So we cannot shortcut that at that time. So we had to go to what we
0:37:46 eventually found out is called crowd engineering, crowdsourcing. And that was a very new technology,
0:37:55 technology was barely a year old or so by Amazon. They created a lot online marketplace for people to
0:38:05 do small tasks to earn money when these tasks can be uploaded on the internet. I remembered when I
0:38:12 heard about Amazon Mechanical Turk, I logged into my Amazon account. I checked the first task I checked out
0:38:21 to do just to try was labeling wine bottles or transcribing wine bottle labels. So you, you know,
0:38:28 the task will give you a picture of a wine bottle and you have to say this is 1999 Bordeaux and all that.
0:38:36 So people upload these kinds of micro tasks and then online workers, like someone in their leisure time,
0:38:42 like me, if I had leisure time, I would just go sign up and get paid to do that.
0:38:50 And we realized that was, again, out of desperation, that was a massive parallel processing with online
0:38:58 global population to do this for us. And that’s how we labeled billions of images and distilled it
0:39:02 down to 15 million high quality image that images.
0:39:06 It’s just so wild when you look at these stories. I just finished a book on
0:39:11 Genentech. And there were all these little technical inflection points that also allowed things to
0:39:18 happen. So if it had been five years earlier, or maybe three years earlier, right, without Mechanical Turk,
0:39:24 boy, like it presents a challenge. But also, as you pointed out, I mean, in science, it’s one thing to get
0:39:30 answers, but you need the input on the front end with a proper hypothesis or a good question.
0:39:38 And even with Mechanical Turk, if you’re only focused on the mechanics of employing that,
0:39:43 you can get yourself into trouble. Because if humans are incentivized to, let’s just say,
0:39:49 I think this was the example I read about, identify pandas in photographs and they’re paid for identifying
0:39:54 pandas. Well, what’s to stop them from identifying a panda in every photo, whether they exist in the photos
0:39:59 or not. So you have to follow the incentives as well. How did you solve for that?
0:40:08 Yeah, I know. This is where, you know, my student and I had, I cannot tell you how many hours and
0:40:15 hours of conversation we have about controlling the quality. We have to solve for that in multiple steps.
0:40:21 We need to first filter out online workers who are serious about doing the work. So for example,
0:40:28 we have to have some upfront quizzes so that they understand what a panda is. They read the question.
0:40:35 And then once they get into, they qualify for that, we ask them to label pandas. But there are some
0:40:42 images we have pre, we know the correct answer. Some are true pandas, some of them are not true pandas.
0:40:51 So the labelers don’t know. So in a way we implicitly monitor the quality of the work by knowing where
0:40:59 the gold standard answers are. So these are the kind of computational tactics we have to use to ensure
0:41:06 the quality of labeling. Amazing. Yeah, just incredible. All right. And I’ll actually just put a
0:41:11 recommendation out there for a book, “Pattern Breakers” by a friend of mine, Mike Maples, Jr. He taught me
0:41:17 the ropes initially of angel investing. But in terms of identifying inflection points and converging
0:41:21 technological, in some cases, converging technological trends that for the first
0:41:26 time makes something possible, which then opens an opportunity for something with the right prepared
0:41:31 mind, in your case, and those of your collaborators and the people you built upon for something like
0:41:37 ImageNet, “Pattern Breakers” is a really good read for folks. So let’s hop to modern day then for a
0:41:43 moment. And I would love to ask you because you’ve been called the godmother of AI in our alumni magazine,
0:41:50 in fact, and elsewhere. But you’ve had such a, not just technical, but historical viewpoint, meaning
0:41:56 you’ve over a broad timeline, you’ve been brought by AI standards, been able to watch the development and
0:42:06 forking and perils and promise of this technology. What are people missing? What do you think is eating
0:42:11 up all the oxygen in the room? What are people missing? Whether it’s things they should know or
0:42:16 things they should be skeptical of or otherwise? Especially I’m here calling you from the heart
0:42:27 of Silicon Valley. And I think people are missing the importance of people in AI. And there’s multiple
0:42:34 facades or dimensions to this statement is that AI is absolutely a civilizational technology.
0:42:41 I define civilizational technology in the sense that because of the power of this technology,
0:42:50 it’ll have, or already having a profound impact in the economic, social, cultural, political,
0:42:58 downstream effects of our society. So I just heard, this is unverified, but I just heard that 50%
0:43:11 of the US GDP growth last year is attributed to AI growth. So apparently this number is 4% for US GDP
0:43:20 have grown 4%. If you take away AI, it’s only 2%. That’s what it means. So that’s civilizational
0:43:26 from an economic point of view. It’s obviously redefining our culture, right? Think about,
0:43:34 you’re talking about the word sucking oxygen out of the room, everywhere from Hollywood, to Wall Street,
0:43:42 to Silicon Valley, to political campaign, to TikTok, to YouTube, to Insta.
0:43:47 Taxis in Japan. I was just there and the videos playing on the back of the headset and the taxi,
0:43:49 we’re all talking about AI. It’s everywhere.
0:43:55 It’s culturally impactful, not only impactful, it’s shifting our culture. It’s going to shift
0:44:03 every education. Every parent today is wondering what should their kids study to have a better future.
0:44:10 Every grandparent is saying, “I’m so glad I’m born early. I don’t have to deal with AI, but still worry
0:44:17 about their grandchildren’s future.” So AI is a civilizational technology. But what I think it’s
0:44:24 missing right now is that Silicon Valley is very eager to talk about tech and the growth that comes
0:44:32 with the tech. Politicians are just eager to talk about whatever gets the vote, I guess. But really,
0:44:39 at the end of the day, people are at the heart of everything. People made AI, people will be using AI,
0:44:47 people will be impacted by AI, and people should have a say in AI. And no matter how AI advances,
0:44:56 people’s self-dignity as individuals, as a community, as society, should not be taken away. And that’s what I worry
0:45:03 about because I think there’s so much more anxiety that because the sense of dignity and sense of agency,
0:45:11 sense of being part of the future is slipping in some people. And I think we need to change that.
0:45:24 It has been a wild year for money in the markets, but managing your cash doesn’t have to be a guessing
0:45:29 game. Wealthfront has a simple solution to help you cut through all the chaos and manage your money
0:45:36 with confidence. With the Wealthfront cash account, your cash can earn 3.5% variable annual percentage
0:45:42 yield from partner banks. You get instant fee-free withdrawals to eligible accounts 24/7. Your money is
0:45:47 always accessible when you need it. And when you’re ready to invest, transferring funds into one of
0:45:53 Wealthfront’s expert-built investing portfolios is easy peasy. For a limited time, Wealthfront is
0:45:59 offering my listeners, that’s you guys, an additional 0.65% boost over the base rate for three months,
0:46:08 which means you can get up to 4.15% APY on your first $150,000 in deposits. More than 1 million people
0:46:15 trust Wealthfront to help build their wealth. Go to wealthfront.com/tim to receive the boosted APY offer
0:46:22 and start earning up to 4.15% variable APY today. This is a paid endorsement of Wealthfront. Wealthfront
0:46:30 brokerage isn’t a bank. The base APY is as of November 7th, 2025 and subject to change. For more information,
0:46:32 we’ll see the episode description.
0:46:38 Hey folks, Tim here. My best-selling card game, Coyote, which I made with the amazing team at Exploding Kittens,
0:46:45 just won the Pop Insider Best Geeky Game of 2025 and also Best Stocking Filler in the Made for Moms
0:46:54 Toy Awards 2025. It is on sale everywhere. It’s cheap, it’s fast to learn, has 4.8 stars out of 5. People are
0:47:00 loving it. CoyoteGame.com will take you to all the retailers, but you can find it everywhere. It is a game
0:47:08 of thinking fast and laughing faster. Thinks charades meets hot potato meets a bunch of brain fun. It’s good for your head.
0:47:14 It’s perfect for families with kids age 10 plus or adults who are kids at heart or don’t take themselves too seriously.
0:47:21 A lot of adults love this game. And as I said, it’s available everywhere. Amazon, Walmart, Target, 8,000 plus retail locations,
0:47:28 you name it. So please check it out. I loved making it. People are really enjoying it. It has 300 or 400 million plus social views
0:47:36 of gameplay online and try it. Enjoy it this holiday season. Check it out. CoyoteGame.com. One more time,
0:47:41 that’s CoyoteGame.com or anywhere you buy your games. Now, back to the episode.
0:47:52 I’ve heard you say that you’re an optimist because you’re a mother and both optimism and pessimism to an
0:47:58 extreme can bias us in ways that are unhelpful or create blind spots. And I’m curious who
0:48:04 if you try to put your most objective hat on, which is difficult for any human, but if you try to do
0:48:12 that, do you think people are too worried, not worried enough, or worrying about the wrong things
0:48:19 for people who are not the CEOs and builders and engineers behind AI? Because you’re right,
0:48:24 of course. I mean, everybody will agree with this, that a lot of people are very worried. And I’m just
0:48:28 wondering if it’s ill-placed, because I don’t really, if you talk to some of the VCs who are
0:48:34 the biggest investors, of course, they have this sort of, in my view, sort of beyond all possibilities,
0:48:41 techno-optimist view of the future where AI solves everything. And it’s hard to believe there’s a free
0:48:47 lunch there. And then you have the doomers, the doom and gloom, where suddenly it’s Skynet next year,
0:48:51 and we’re all slaves to robots, we’re eliminated, turned into paperclips. And reality is probably in
0:48:57 between those two. So do you think people are worrying about the right things, or have they
0:49:04 lost the plot in some way? First of all, I call myself a pragmatic optimist. I’m not a utopian,
0:49:11 so I’m actually the boring kind. I don’t believe in the extreme on both sides. I travel around the
0:49:20 world. Just last month, I was in Middle East, I was in Europe, I was in UK, I was in Canada, I came back home
0:49:32 in America. I think people in America, and people in Western Europe are more worried about AI than,
0:49:43 say, people in Middle East, in Asia. We don’t have to litigate on why they’re more worried. But just to come
0:49:52 closer to home, just talk about US. I wish I have a megaphone to tell people in the US that you’re known
0:50:02 to be one of the most innovative people. Our country has innovated so many great things for humanity, for
0:50:13 socialization. We have a society that is free and vibrant. And we have a political system that we still
0:50:27 have so much say in how we want to build our country. I do wish that our country has more optimism and
0:50:37 positivity towards the future of using AI than what is being heard now. I think people like me,
0:50:44 technologists living in Silicon Valley, has a lot of responsibility in the right kind of public
0:50:53 communication. So there’s a lot of things that was not communicated in an effective way. But I do hope
0:51:07 that we can instill more sense of hope and self-agency into everybody in our country because I think there’s
0:51:16 so much upside of using AI in the right way. And I want not just people in Silicon Valley or in Manhattan,
0:51:26 but I want people in rural communities, in the traditional industries, in everywhere, 50 states to be able to
0:51:29 embrace and benefit from AI.
0:51:34 Why are you building what you’re building? What is WorldLabs? Why decide to do this?
0:51:42 I actually answer this question very often to every member of my team. I built WorldLabs. There are two levels of this
0:51:49 answer from a technology point of view. WorldLabs is building the next generation AI focusing on spatial
0:51:56 intelligence. Because spatial intelligence, just like language intelligence, is fundamental in unlocking
0:52:05 incredible capabilities in machines so that it can help humans to create better, to manufacture better,
0:52:12 to design better, to build better robots. So spatial intelligence is a linchpin technology.
0:52:19 But one level up, why am I still a technologist? It’s because I believe
0:52:30 humanity is the only species that builds civilizations. Animals build colonies or herds, but we build civilizations.
0:52:38 We build civilizations because we want to be better and better. We want to do good, even though along the
0:52:45 way we do a lot of bad things. But there is a desire of having better lives, having better community,
0:52:53 having better society, live more healthily, have more prosperity. That desire is where civilization is built
0:53:02 upon. And because I believe that humanity can do that, I believe science and technology is the most
0:53:09 powerful tool, one of the most powerful tools in building civilizations. And I want to contribute to
0:53:16 that. That’s why I’m still a scientist and a technologist, and I’m building WorldLabs for that.
0:53:25 Can you explain to people what spatial intelligence is and what the product is,
0:53:28 so to speak, at least as it stands right now that you’re building?
0:53:38 Spatial intelligence is a capability that humans have, which goes beyond language. It’s when you pack a
0:53:49 sandwich in a bag, when you take a run or a hike in a mountain, when you paint your bedroom. Everything that has to
0:53:58 do with seeing and turning that scene into understanding of the 3D world, understanding of the
0:54:05 environment. And then in turn, you can interact with it. You can change it. You can enjoy it. You can
0:54:14 make things out of it. That whole loop between seeing and doing is supported by the capability of
0:54:20 spatial intelligence, right? The fact that you can pack a sandwich means you know what the bread looks
0:54:26 like. You know how to put the knife in between. You know how to put the lettuce leaf on the bread. You
0:54:34 know how to like put the sandwich into a Ziploc bag. Every part of this is spatial intelligence. And does
0:54:41 today’s AI have that? It’s getting better. But compared to language intelligence, AI is still very early in that
0:54:51 ability to see, to reason, and also to do in world, in both virtual 3D world as well as real 3D world.
0:55:00 So that’s what WorldLabs is doing. We are creating a frontier model that can have intelligent capability
0:55:11 in the model to create world, to reason around the world, and to enable, for example, creators or designers
0:55:18 robots to interact with the world. That’s spatial intelligence. Could you expand on the, you know,
0:55:24 designers or creatives or robots interacting with the world? So does that mean that you could,
0:55:29 and my team has been playing with some of the tools, so thank you for that. What does that mean? If you
0:55:36 could paint a picture for, let’s say, a year from now, two years from now, how might someone use this or
0:55:42 how might a robot use this? I was just talking to someone a couple of weeks ago, and it was really
0:55:50 inspiring, is that high school theaters are very low budget, right? Like, okay, sometimes I go to San
0:55:59 Francisco opera or musicals, and the sets that’s built for theater are just so beautiful. But it’s very hard
0:56:08 hard for high school to have that budget to do that. Imagine that you can take today’s WorldLabs model,
0:56:19 we call it marble, and then you create a set in, I don’t know, in medieval French town. And then you put
0:56:31 that in the background and use that digital form to help transport the actors and action into that world.
0:56:38 And of course, depending on the auxiliary technology, whether you’re on a computer or eventually people
0:56:46 can use a headset or whatever, you can have that immersive feeling of being in a medieval French town.
0:56:53 That would be an amazing creative tool for a lot of creators. That was an example someone
0:57:00 and I was talking about it a couple of weeks ago, but we already see creators all over the world. Some of
0:57:10 them are VFX creators. Some of them are interior design creators. Some of them are gaming creators. Some of them
0:57:18 are educators who want to build some worlds that transport their students into different experiences
0:57:26 are already starting to use our model because they find it very powerful at their fingertips to be able to
0:57:33 create 3D worlds that they can use to immerse either their characters or themselves into.
0:57:40 And just process-wise, if someone’s wondering how this works, let’s just say it’s a public school teacher,
0:57:48 let’s just say, who’s hoping to inspire and teach their students going the extra mile. What does it look
0:57:52 like for someone to use this? Are they typing in text, describing the world they’d like to create,
0:57:58 uploading assets or photos, almost like an image board? How does it work if someone’s non-technical?
0:58:05 Yes. So they don’t need to be technical at all. They open our page on desktop or in their phone,
0:58:11 but desktop is more fun because it has more features. And then they can type, you know,
0:58:19 a French medieval town, or they can actually go to anywhere. They can use Mid Journey or Nano Banana to
0:58:26 create a photo of a French medieval town, or they can get an actual photo about that. And then they upload it,
0:58:35 we call it prompt. And then after a few minutes, our model gives you a 3D world that is, say,
0:58:43 a part of the town. It does have a limit in its range. And then that 3D world is generally 3D because
0:58:50 you can just use the mouse to drag and turn around and walk around and see that world. And then downstream,
0:58:59 if you want to use it, you have many ways to use it. You can actually create a movie out of it by, like,
0:59:06 using one of our tools on the website to just put cameras and you can make a particular movie out of it.
0:59:10 If you’re a game developer, I was just going to say, it sounds a lot like a gaming engine.
0:59:18 Yes, you can put a lot of characters in it. If you’re a VFX professional, we have a lot of VFX
0:59:24 professionals. They can actually take this and put it in the workflow of their movie shooting and have
0:59:32 real actors shooting movies. We’ve also have psychology researchers using that immersive world,
0:59:39 in particular psychiatric studies. We could also use that as the simulation for robotic training
0:59:45 because a lot of robotic training needs a lot of data, and then use that for generating a lot of
0:59:51 different data. So is it almost like a flight simulator for robots before they go into the real world?
0:59:59 That’s part of the goal. We are still early, so the flight simulator is not complete yet, but that’s part of the journey.
1:00:05 You mentioned psychiatric studies. I think that’s what you just mentioned. What might that look like?
1:00:11 Yeah, so we actually got this researcher who called us, and they’re studying people who have
1:00:18 psychological disorders, like obsessive compulsive disorder, where they’re triggered by certain
1:00:25 environments, and they want to study the trigger and also just study how the treatment. But how do you
1:00:34 trigger someone who, let’s say, is particularly have issue with, let’s say, a strawberry field? I’m just making it up.
1:00:40 Yeah. I mean, you can take them to a strawberry field, but what about you want to know if it’s
1:00:48 a strawberry field in the summer, or strawberry field at night, or it’s strawberry, or it’s many
1:00:54 strawberry? Like, how do you do this? Suddenly, this researcher realized we give them the cheapest
1:01:01 possible way of varying all kinds of dimensions, and they can test this out and do their studies.
1:01:05 That’s really interesting. Yeah, I could see it being applied to, it might be called exposure
1:01:11 therapy, but in terms of now that you’re describing it, I could see how it could be added into, I mean,
1:01:15 pretty much everything, right? I mean, if you think about how humans operate in the real world.
1:01:22 Yes. Yeah. And the boundary between real world and digital world is less and less,
1:01:30 thinner and thinner, because we live in many screens, we live in the real world, we do things in virtual
1:01:36 world, we do things in real world, we’ll create machines that can do things in real world and
1:01:41 virtual world. So there’s a lot we do in digital and physical spaces.
1:01:51 Who are some scientists or researchers who you pay attention to, who are not necessarily
1:01:56 kind of the big brand names and marquee lights that are already very public in the world? Is there
1:02:00 anybody who stands out where you’re like, you know, there’s some really tremendous people doing good work?
1:02:06 Well, that’s part of the reason I wrote the book, especially in the middle chapters where I wrote
1:02:13 about the journey of doing ImageNet that combines cognitive science with computer science. And I
1:02:20 actually talk about psychologists and neuroscientists and developmental psychologists. You know, some of
1:02:26 them are still with us, some of them are not. For example, the late Anne Treisman, Irv Biederman,
1:02:32 they all passed away in the last few years, but they were giants in cognitive science,
1:02:38 whose work has informed computer science and eventually AI. You know, there are still
1:02:45 lots of scientists around the world, many of them are in the US, who are thinkers in developmental
1:02:52 psychology, in AI, I follow their work. I think that the world of science, just to name some names,
1:03:00 right, Liz Belke in Harvard, Alison Gopnik in Berkeley. I love Rodney Brook, who was a former
1:03:08 MIT professor in robotics. And there’s just a lot of them. I don’t mean to just single them out,
1:03:14 but you’re asking me for names that are not in the news of AI.
1:03:21 Yeah. That’s perfect. Thank you. I would also love to get your perspective on what might be,
1:03:28 this is a very strong word, but seemingly inevitable in terms of developments in the
1:03:33 near intermediate future. And I’ll give you an example of what I mean. In 2008, 2009, I became
1:03:38 involved with Shopify, the company back when they had like 10 employees. And there were a few things
1:03:43 happening around that time. And you could ask questions, you know, in the next 10 years or 20
1:03:48 years, will there be more broadband access or less? More. Okay. Will there be more e-commerce or less?
1:03:54 There’ll be more. Okay. And when you have four or five of those that seem over a long enough time
1:03:59 horizon, absolute yeses, it begins to paint a picture of where things are going.
1:04:05 Are there any things that in the next handful of years you think are perhaps underappreciated
1:04:11 as near inevitabilities? You want me to talk about underappreciated? I mean,
1:04:15 I don’t know if they’re overappreciated, but they’re definitely appreciated. The need for power
1:04:25 is appreciated. The trend of more AI, not less AI is appreciated. The long-term trend of robots coming
1:04:33 is appreciated. So these are appreciated. What’s underappreciated is spatial intelligence is
1:04:38 underappreciated in the sense that everybody’s still now talking about large language models,
1:04:46 but really world modeling of pixels of 3D worlds is underappreciated because like you were saying,
1:04:53 it powers so many things from storytelling to entertainment to experiences to robotic simulation.
1:05:01 I think AI and education is underappreciated because what we are going to see is that
1:05:11 AI can accelerate the learning for those who want to learn, which will have downstream implication
1:05:18 in our school system as well as in just human capital landscape. Like how do we assess
1:05:26 qualified workers? You know, it used to be which school you graduate from with which degree that will be
1:05:36 changing with AI being at the fingertip of so many people. That’s underappreciated. I think AI’s impact in our
1:05:47 economic structure, including labor market, is underappreciated. The nuance is underappreciated. I think this whole
1:05:59 rhetoric of either total utopia post-scarcity is hyperbolic or like everybody’s job will be gone. It’s hyperbolic.
1:06:10 But the messy middle is how from knowledge worker to blue collar to hospitality, to all these changes that’s
1:06:20 what’s happening. It’s underappreciated by our policy workers, by our scholars, by just overall society.
1:06:26 Well, what are some of the nuances from the job perspective? Maybe this ties into what I promised
1:06:31 earlier I was going to ask you, which is what you are telling or will tell at own other ages,
1:06:37 your children are recommending. Let’s just say, I don’t know how old they are. But if we assume that
1:06:42 they, just for the sake of discussion, of the age where they’re trying to decide what they should study,
1:06:49 where they should focus, things of that nature, how would you think about answering that, even provisionally?
1:06:59 I think the ability to learn is even more important because when there was less tools, fewer tools to
1:07:06 learn, it’s easier to just follow tracks. You go through elementary school, middle school, high school,
1:07:15 college, and then get some training vocationally. And that’s kind of a path. And with that is a set of
1:07:24 structured credentials from degrees and all that. But AI has really changed it. For example, my startup.
1:07:31 When we interviewed a software engineer, honestly, how much I personally feel the degree they have
1:07:40 matters less to us now. It’s more about what have you learned? What tools do you use? How quickly can you
1:07:48 superpower yourself in using these tools? And a lot of these are AI tools. What’s your mindset towards using
1:08:01 these tools matter more to me? At this point in 2025, hiring at World Labs, I would not hire any software engineer who
1:08:09 does not embrace AI collaborative software tools. It’s not because I believe AI software tools are perfect.
1:08:19 It’s because I believe that shows, first of all, the ability of the person to grow with the fast
1:08:27 growing toolkits, the open mindedness, and also the end result is if you’re able to use these tools, you’re able to
1:08:35 learn, you can superpower yourself better. So that is definitely shifting. So coming back to your question, what do you
1:08:45 what do you tell young people, tell children? I think the timeless value of learning to learn, the ability to learn
1:08:48 is even more important now.
1:08:57 Yeah, it strikes me as we’re talking that it’s only going to get increasingly easier for the ambitious to
1:09:04 act as superpowered autodidacts. We’ve already seen this with certainly YouTube has a nice track record.
1:09:09 Now you can either entertain yourself to death and avoid doing things that help with self-growth and
1:09:14 development or you can supercharge it. And similarly with AI, right, you flash forward, we don’t even need
1:09:21 to flash forward. But it’s how does a teacher audit that their students are doing the work they’re supposed
1:09:26 to be doing? On so many levels, it’s getting to the point. There are some exceptions, but of near impossibility.
1:09:32 And students can either avoid all work or they can supercharge their own work. But the output might look
1:09:38 very similar, at least for a period of time. So schooling is going to change a lot. It’s very, very interesting.
1:09:48 I actually think, Tim, if the school evaluation is structured in a way that whatever AI gives and
1:09:54 whatever the student gives is the same, there’s something wrong with the structure of the evaluation.
1:09:57 Okay. Can you say more about that? That’s interesting.
1:10:05 So, for example, English essay. This is not me. This is me hearing a story that I so agree with. I’ll
1:10:14 retell the story is that as a high school freshman English class teacher, someone told me the story of their
1:10:21 kid’s school. On the first day of school, the teacher actually said to the class, “I want to show you
1:10:32 how I would score AI.” So the teacher gave an essay topic, show the students this is what the best AI gave me,
1:10:38 And I’m going to show you how I think this is good, this is bad, how this is suboptimal, and I’ll give it
1:10:47 a B minus. Now I will tell you, this is my bar. If you’re so lazy that you ask AI to write your essay,
1:10:55 this is what you’re going to get. You can use AI. That’s totally fine. But if you can do the work, learn,
1:11:04 think, be the best human creator you can, and work on top of that, you can get to A. You can get to A pluses.
1:11:12 And that would be, in my opinion, the right way to structure the evaluation. It’s not to hit humans
1:11:19 against the AI and then try to police the use or not use of AI. It’s to show where the tools,
1:11:23 the bar of the tools are, and where the bar of the human learner should be.
1:11:29 I’m gonna sit with that example and try to think of more examples. It’s very interesting. And boy,
1:11:33 oh boy, I’ve been shocked by how quickly the models improve. But yes, that’s like as a thought
1:11:39 experiment. I’m gonna chew on that. I know we only have a few minutes left. Fefe, I wanted to ask you
1:11:45 a question I ask a lot, which is if you could put a quote or a message, something on a billboard,
1:11:50 something to get in front of millions, billions of people. Just assume they all understand it.
1:11:55 Could be an image, could be a question, could be a quote, anything at all. A saying, mantra,
1:12:00 doesn’t matter, could be almost anything. What would you or what might you put on that billboard?
1:12:02 What is your North Star?
1:12:09 What is your North Star? This is, of course, critically important. And coming back to
1:12:14 how you define that or find that for yourself. I mean, you were talking about audacious questions.
1:12:20 And then that leading to a North Star hypothesis. Is there another way that you would
1:12:24 encourage people on top of that to think about finding their North Star?
1:12:32 I believe that’s how that makes us so human and makes us to be so fully alive, is that
1:12:42 we as a species can live beyond the chasing of just basic needs, right? But dreams, missions,
1:12:50 and goals, and passion. And everybody’s North Star is different. And that’s fine. Not everybody has to
1:12:56 have AI as their North Star. But finding that goes to the heart of education again. And I don’t mean
1:13:04 formal classroom education. It’s just the journey of education. A lot of that is the ability to learn who
1:13:12 you are and to learn how to formulate your North Star and how to chase after that.
1:13:16 Last question. Did your parents ever explain to you why they named you Feifei?
1:13:24 Yes, it’s because when my mom was going through labor, my dad was characteristically late to
1:13:30 the hospital. And along the way, he caught a bird. He let it go, but he did catch a bird. I don’t know,
1:13:38 he was just distracted. It was in Beijing, in the city of Beijing. My dad was bicycling to my mom’s
1:13:46 hospital. And that inspired him to call me Feifei. Feifei. Oh, wait, sorry. For those who don’t speak
1:13:51 Chinese, I forgot you do speak Chinese. But for those who don’t speak Chinese, fei means flying.
1:13:54 So yeah, so be inspired by a bird.
1:13:59 You know, really quick, I’ll just say, because it’s kind of funny. My first Chinese name that I had
1:14:07 was Feitingchang, which is because I was very blunt and honest. So Tingchang. But Feitingchang. But
1:14:13 when I was first starting, my tones in China were not polished. And people thought I was saying that my
1:14:21 name was Feijichang, which is airport. So I petitioned my teachers and we changed my name to
1:14:22 something less confusing.
1:14:24 What’s your new name?
1:14:33 Feiyuchang. Feiyuchang. It’s like, it’s like, but it’s without the
1:14:40 at the bottom. Oh, wow. Fancy name. That’s way more sophisticated than mine.
1:14:46 Well, I get to script it with my Chinese teachers. So I have an unfair advantage.
1:14:50 Dr. Lee, thank you so much for the time. We will link to the show notes for
1:14:53 everybody at tim.blog slash podcast. They’ll be able to find you easily.
1:15:00 And everybody should check out worldlabs.ai and we’ll put every other link, your social and so on
1:15:03 in the show links. But thank you for the time. I really appreciate it.
1:15:06 I enjoyed our conversation. Yeah, likewise. Bye.
1:15:14 Hey guys, this is Tim again. Just one more thing before you take off. And that is Five Bullet Friday.
1:15:18 Would you enjoy getting a short email from me every Friday that provides a little fun
1:15:23 week before the weekend. Between one and a half and two million people subscribe to my free newsletter,
1:15:29 my super short newsletter called Five Bullet Friday. Easy to sign up, easy to cancel. It is
1:15:35 basically a half page that I send out every Friday to share the coolest things I’ve found or discovered
1:15:40 or have started exploring over that week. It’s kind of like my diary of cool things. It often includes articles
1:15:47 I’m reading books. I’m reading albums, perhaps gadgets, gizmos, all sorts of tech tricks and so
1:15:53 on that get sent to me by my friends, including a lot of podcast guests. And these strange esoteric
1:16:00 things end up in my field. And then I test them and then I share them with you. So if that sounds fun,
1:16:04 again, it’s very short, a little tiny bite of goodness before you head off for the weekend,
1:16:09 something to think about. If you’d like to try it out, just go to Tim dot blog slash Friday,
1:16:14 type that into your browser, Tim dot blog slash Friday, drop in your email, and you’ll get the
1:16:20 very next one. Thanks for listening. As many of you know, for the last few years, I’ve been sleeping on
1:16:25 a midnight luxe mattress from today’s sponsor, Helix Sleep. I also have one in the guest bedroom
1:16:31 downstairs and feedback from friends has always been fantastic. It’s something they comment on without
1:16:37 any prompting from me whatsoever. I also recently had a chance to test the Helix Sunset Elite. The
1:16:42 Sunset Elite delivers exceptional comfort while putting the right support in the right spots. It is made with
1:16:48 five tailored foam layers, including a base layer with full perimeter zoned lumbar support, right where
1:16:55 I need it, and middle layers with premium foam and micro coils that create a soft contouring feel. Helix offers
1:17:02 a 100-night sleep trial, fast, free shipping, and a 15-year warranty. So check it all out. And now you can
1:17:10 get 20% off anything on their website, so site-wide. So just go to helixsleep.com/tim. One more time,
1:17:15 helixsleep.com/tim. With Helix, better sleep starts now.
1:17:23 I have been fascinated by the microbiome and probiotics as well as prebiotics for decades,
1:17:29 but products never quite live up to the hype. Now things are starting to change, and that includes
1:17:36 this episode’s sponsor, Seeds DS01 Daily Symbiotic. I’ve always been very skeptical of most probiotics,
1:17:42 but after incorporating two capsules of Seeds DS01 into my morning routine, I have noticed improved
1:17:49 digestion, and improved overall health. So why is Seeds DS01 so effective? For one, it is a two-in-one
1:17:55 probiotic and prebiotic formulated with 24 clinically and scientifically studied strains. But if the
1:17:59 probiotic strains don’t make it to the right place, they are not as effective. So Seeds developed a
1:18:05 proprietary capsule-in-capsule delivery system that survives digestion and delivers a precision release
1:18:11 of the live and viable probiotics to the colon. And now you can get 20% off your first month with code
1:18:29 20Tim at Seed.com/Tim using code 20Tim all put together. One more time, Seed.com/Tim code 20Tim.