a16z Podcast - The Worlds She Sees with Godmother of AI, Fei-Fei Li

Episode Date: November 13, 2023

Fei-Fei Li, PhD,  Professor in the Computer Science Department at Stanford University, and Co-Director of Stanford’s Human-Centered AI Institute, joins Bio + Health founding partner Vijay Pande.In this candid conversation, Li unfolds her transformation from a young immigrant to an influential figure in AI. The conversation explores the birth of ImageNet, a pivotal step that bridged the gap between visual intelligence and accessible AI technology. They delve into the notion of a 'Dignity Economy,' hinting at a future where technology serves to elevate human experience rather than undermine it. Li also touches on the delicate balance between relentless innovation and life's humble pursuits. This episode peels back the layers on the human side of AI, offering a rare glimpse into the personal and professional realms of a pioneer shaping the AI landscape.Check out her new book, The Worlds I See, here: https://us.macmillan.com/books/9781250897930/theworldsiseeCheck out other episodes form our sister podcast, Bio Eats World: https://a16z.com/podcasts/bio-eats-world/ Stay Updated: Find a16z on Twitter: https://twitter.com/a16zFind a16z on LinkedIn: https://www.linkedin.com/company/a16zSubscribe on your favorite podcast app: https://a16z.simplecast.com/Follow our host: https://twitter.com/stephsmithioPlease note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.

That just kind of changed me. To this day, physics is still, you know, first love. Learning is organic, learning is messy, learning is big data, learning is reinforcement, you know, try an error. It took me two years of agony. It's hard to resist Stanford and the Silicon Valley when AI was taking off.

Not a single university today in America can train a chat GPD model, not even probably GPD sweet. The quest for artificial intelligence has spanned for decades, with the field really kicking off in the 1950s. Each era was marked by unique breakthroughs, and despite the recent flourish of advancements, the foundational work has long been underway by a cast of characters dedicated to pushing this frontier. One notable figure, some even referring to her as the grandmother of AI, is Faye Bailey, a professor of the computer science department at Stanford University and co-director of Stanford's human-centered AI Institute.

And in today's episode, we have the pleasure of her joining Bio and Health founding partner Vijay Ponday to discuss her journey through the years, from uprooting as a teenager from Chengdu, China to Percipany, New Jersey, and barely speaking English at the time, all the way to building ImageNet, an instrumental project taking computer vision projects from tens of thousands of images to tens of millions. Throughout the episode, you'll also hear many of Faye's very human stories, like running a dry cleaning shop while she was at Princeton, to her mother's role in allowing her to pursue physics, all of which helps to shape her vision for the very technology that she helped bring to life. Faye Faye also just released her new book, The World's

I See, that dives even deeper into her history and relationship with technology. You can find the link to that book in our show notes, and I can tell you as a daughter of two immigrants, this story really resonated. I hope it does for you too. And this episode comes from our sister podcast, Bioweets World. So make sure to go check that out too. Let's dive in. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16C fund. Please note that A16Z and its affiliates may also maintain investments in the

companies discussed in this podcast. For more details, including a link to our investments, please see A16C.com slash disclosures. Tafay, thank you so much for joining us on BioEats World. Yeah, well, MNJ, it's always a pleasure to have a conversation with you. Thank you. Thank you. Well, so you've written this really, really.

beautiful book that outlines a lot of your life story and history and then connects to where we are now. And perhaps most importantly, it talks about where the world could be going. And maybe the question at our hands is, which world do we build and how do we build it? Well, thanks for talking about my upcoming book called The Worlds I See. I am an immigrant. I came to this country when I was 15-year-old and landed in Percipity, New Jersey. But before that, I lived in a city in China called Chengdu. I was, not surprisingly, a STEM kid. I loved science and nature. Partially, I think, very inspired by my own father, who is a very curious person about nature. Yeah, so what was STEM like in China at that time?

Our teachers were rigorous, so we did have more or less good foundational training in terms of math. But what really, I think, was a light bulb that went on for me was, I think it was around seventh grade that physics was introduced as a subject in middle school. And that just kind of changed me. Like to this day, physics is still, you know, first love. I mean, to the point I majored in physics at Princeton. And I had good teachers. And I just basically fell in love with that subject. I'm in the same camp with you there. Do you think it's just a coincidence or do you think there's a connection between the love of physics and creating AI? I'll tell you the connection for me, and I'm curious what was the connection for you.

The connection for me is physics, training, or love for physics, to me is about audacious questions. Is you dare to ask what's the smallest particle of the universe or what is the end of time and space? You go to the most extreme of what nature's mystery is, and that audacity, to me, what I love about physics. And it was very natural for me that the audacity of can you make a thinking machine. What is intelligence? The tool set is computational and mathematical. And that goes without saying. But that's the connection for me. What is it for you? I think you hit on tune out, I would add one more. So very much the big audacious goals and as a kid learning about special relativity or general relativity. It's just amazing that you can just with math and a very few.

simple postulates get to crazy ideas like equals MC squared without actually all that much math is kind of amazing. What was it like shifting from China to come to the U.S. and immigration? That must have been, for many, it's typically a bit of a culture shock. A bit? That's an understatement. But yes, 15-year-old. Think about a 15-year-old teenager. You can hardly get them to brush their teeth and let alone uproot them, right? So, it was more than a shock. I landed in Percipany, New Jersey high school, and I barely speak English. Like, I can barely recognize the sign of the restrooms to know which ones to go to.

Terrifying. It was. It was terrifying. My parents didn't speak English. They dropped everything and decided to leave. And so it was kind of a rebirth because instead of doing my physics and loving, reading, special relativity in Chinese, I suddenly have to go back to almost ABC, not quite ABC, but I was in all ESL classes. I remember ESL English, and they did put me in a regular math, and everything

I understand was only loving numbers and symbols, and I did it all perfectly, and everything has words. I didn't know how to answer at all. Well, it's also fascinating because I can only imagine or assume that you must have thought a lot about learning because you had to learn everything again. So much of AI is, especially modern AI is not expert systems. It's learning how to learn. And you were forced to learn how to learn. That's a really interesting observation, Vij. I guess I've done this twice now. Once it's relearning everything as a teenager and then becoming a mother and observing little humans learning. And you're right. And learning is organic. Learning is not. There is some rule base.

You teach little kids to follow certain rules, you know. But learning is organic, learning is messy, learning is big data, learning is reinforcement to try an error. Yes, you're right. Also, the other part in the book that comes out very clearly, you're talking about your math teacher, Mr. Sebella. Could you tell us a little bit about him and his sort of impact on your life? Yeah, I can only use the word profound because imagine you are just this teenager, barely speaking English, dropped in the middle of a public high school, and parents were all trying to survive. Most of my off-school hours were in kitchens of Chinese restaurants. And then I

really was lonely, was somewhat scared, nervous, and really pretty much clueless. And then I stumbled upon this teacher who kind of took me under his wing. Mr. Sabella is not your typical thinking about how warm and radiant that kind of teacher. He is a tough love teacher. And somehow we had so much chemistry that I just started frequented his office and asking him questions about just even how to navigate this world. What is really memorable for me is my senior year, I got placed out of all the math curriculum that the school district could offer.

And it's not anything fancy, Vijay. It was just they don't do multivariate calculus. So Mr. Sabella just created a one-person class for me for his lunch break hours. And I was doing multivirate calculus with him. As most people love, most places don't offer multivariate calculus at that time, right? So that was already you doing deep into university math in high school, which itself was probably very impressive. Honestly, Vijay, I had no idea. Honestly, this is such a new country.

my cohort is Chinese restaurant workers, and my parents don't know any of this. So if Mr. Sabella, if he didn't go that extra miles for me, I wouldn't even have known he went that extra mile, right? What would you say the role of your parents were in terms of guiding you towards going into science and computer science and all these things? On one hand, very little, on the other hand profound, actually, very little in the sense of My dad is quasi-stemmed. He had a little bit of engineering education, but he came from that generation, really cultural revolution generation that just didn't really have much schooling. And my mom, for very sad reasons, was not even allowed to go to school. So they didn't have much of that. So they're also very hands-off. They never check my homework or anything. And we landed in America. They can barely survive themselves. They were in cashier jobs, long. hours, my mom has terrible health, but all that means I was just on my own. In the meantime,

with all that backdrop of life's challenges, they were relentlessly convicted for making sure I go for my passion. You know how the typical stereotype, they never came home and say, Faefei, you need to be a doctor or lawyer, whatever that is. They majored in physics. I literally remembered that friends and neighbors talked to my mom and said, you have one girl, and she's a girl. Why are you allowing her to major in the most useless major for a girl? And my mom had a simple, seemingly non-perfound, but very profound answer,

which is what she likes it. What was your time like at Princeton? And my understanding is that, you know, while you did do physics, computation, and neuroscience did come into your thinking. Yeah. Okay. So I was living in a tale of two towns. So Monday to Friday, I was living at Princeton student life in the dorm. But starting Friday evening to weekend, I was running my dry cleaner shop in Pennsylvania, New Jersey.

So what happened is that as soon as I got into Princeton, the job, the job, were very unstable for my parents. My mom's health was deteriorating, and I feel like we need to do something. So we borrowed money, including from my math teacher, Mr. Sabella, and bought a tiny little dry cleaner shop. And in Silicon Valley Spirit, I had my startup, and I was the founding CEO. I hired my two employees, and I started running that dry cleaner. So on one hand, And Princeton was the Mecca. I was just in heaven, right? Literally, Vijay, first day of physics 105, the professor said, Palmer Hall, I instead had sat here. And I was just totally in heaven. But I think just the experience you've been through, I think just probably also

built such an entrepreneurial spirit. And I think you must have taken everything much more seriously because you weren't playing around. You weren't messing around, right? It was survival. not afford it. It was survival and I remembered I had to pass heavy machinery license to run a dry cleaner shop. I had to do customer services. I ran that shop for seven years and three of them were remote from Pasadena, Caltech. So a lot of customer service was remote. It was good. It builds character. But then you're a professor at Princeton and then you get to Stanford also pretty rapidly from there, right? Yeah. So I actually, from Caltech, I went to a bad at Champagne for one year. I was very grateful. They gave me a job. But then Princeton called me and it's my alma mater.

My parents at that time, I was already taking care of them. So they were with me, but they wanted to go back to New Jersey. So I took them back to Princeton. And I was going to live there happily ever after. But your prestige colleagues at Stanford called me, especially Daphne Cullors, Sebastian Brown, Andrew, and Bill Dali, and it took me two years of agony. It's hard to resist Stanford and Silicon Valley when AI was taking off. Vijay, my field is visual intelligence, making AI be visually intelligent and seeing things. What is seeing? Seeing is not just sensing light and colors and shape.

Seeing is really making sense out of things. You see, and eventually, because of that, you can do things in the world. And a huge part of making sense in human visual intelligence is understanding the semantic world. You actually know you're looking at a desk, a puppy, a phone, or your own phone, and so on. So that's called object recognition. And it's actually a really hard problem. Evolution took hundreds of millions of years to crack this problem. And humans are by far the most capable visual animal when it comes to object recognition.

So I was a faculty at Princeton 2007, and at that time, I had a lot of conviction that this is my North Star problem to work on, because as physics has taught us, that you have to learn to identify important North Star problem to go after as a scientist. And to me, that's a North Star problem. But if I look at my field at that time, a very small number of people are working on this problem, but they're working with very small-scale data. We're still stuck in the Bayesian modeling world that we are tweaking parameters. It's really hard to get it to work. And I guess the insight come from me and my students is that mathematically we know AI is about generalization. The flip set of generalization is overfitting. There are two ingredients when it comes to overfitting.

One is your model is not good enough. It's too small. The other one is your data is not good enough. So the whole field was working on modeling per se. I was working on modeling. We were writing pages of equations. But I guess an insight that we're missing data. That's driving algorithms.

That's driving AI. Nature has taught us that there is a top. ton of data out there. Well, conveniently, that was year 2007. What has happened is internet has happened. It's still early internet, but it was already big. So there is data. If you look, you want to get it. So I decided with my students, let's just go absolutely crazy. Our field was working with a data set of thousands of images. Let's go tens of millions if we can. So image that eventually was a North Star visual intelligence data set that has 15 million images organized of the whole world's nouns that are visual,

which turned out to be around 22,000 of them, and it became the data set that drives visual intelligence. How do you gather that? Because that's no small feat. Yeah, well, it starts with the delusional optimism, which in hindsight, if I wasn't delusional, I would have. have started it. Well, but was it delusional because it worked, right?

Well, I'll tell you, there was some delusional part because I wanted to go big. And this is where advisors can be more delusional than graduate students. The graduate students say, okay, I downloaded a few million. I'm like, that's not enough. By the time we downloaded the entire internet, based on WordNet vocabularies, we had a billion images, right? So what do we do? But did they have labels? I forget.

So here's the thing. We use labels and some trick of label expansion to download, but they're very noisy, right? Especially early days of Internet. But even today, you type in the word German Shepherd and what you're going to get is not necessarily... It could be a German human being that mine's sheep. Exactly. It could be a T-shirt. It could be a book.

It could be a cartoon, whatever, right? So they were guided by labels, but you have to do serious curation and cleaning and organization, and we have to do it by humans. And I thought, the first time I thought, well, how about hiring undergrads, right? Princeton undergrads are somewhat smart, but turned out they're too smart to want to work on this problem. And also, back of envelope computation says it's going to take, I don't know, 19 years to finish. And my PhD student, Jadden, who was my first author for ImageNet, was like, I'm not going to take a PhD. for 19 years.

So I thought about another way is to use machines to clean because we're thinking, oh, can we get machines to clean? But that's a circular reasoning because machines are not good enough. If we use machines to clean, we'll never create that real data set

to drive new machine learning algorithms. So eventually, again, serendipitously, we discover Amazon Mechanical Turk as an online market. that was just launched less than a year. And we actually at some point was probably the biggest academic users. And that online market gave us tens of thousands of online workers across more than 100 countries for three years.

And I think it really was that vision of scale that was the difference, right? I mean, that's, I think, the big aha moment, not just for images, but I think for all of us, is that if you could get to enough scale, AI could work. Yes. No. ImageNet was such a foundational pillar for AI, but even that was like, what, 2010? 2012, yeah. So now 10 years ago, what do you think about sort of how AI has emerged in the last decade?

Are you surprised it's going this fast? Are you surprised it took so long? I'm going to say, Vijay, roughly speaking, I'm impressed by the speed. What about you? I mean, you're in the field, too. I'm kind of shocked, and I think part of it is that. I think having Google Brain and meta and these collaborations also between these companies and academics, there's so many smart people working on it with so many resources, so much

compute, so much data that just wasn't there before. It was different than if it was purely an academic pursuit. And that you go to archive, it feels every week there's like an interesting archive paper that maybe in previous times would have been every quarter or every year or something like that. Yeah, no, I cannot catch up, to be honest. Yeah, so by and large, I'm very impressed that you're right. What has converged is that because we're seeing glimpse of hope from the early days, that resources start to come in, especially the large tech companies.

Then, you know, war's law continue to carry us, right? It's a genuine inflection point. But I'm also concerned. I'm concerned not from this typical rhetoric of human extinction. I think that's a very far concern. I think it's just powerful tools can be used good and bad in powerful ways. Yeah. So I agree with you both that the extension is overblown.

And yeah, like any human tool, it could be nuclear energy, it could be a nuclear bomb. A car can be transportation. You could kill somebody with a car. So how do you see us moving forward responsibly? First of all, I guess coming out of my dry cleaner shop, I am a pragmatist. Being responsible also involves being pragmatic. The world has so many problems that AI can help solving. And one part of the pragmatism is to embrace good tools, right?

Especially in the area you work in bio and I work in healthcare. I cannot imagine we don't embrace these tools. But in the meantime, recognize responsibility and recognize collective responsibility. We all have a responsibility, whether it's big tech, individual researchers, policy makers, civil society, school teachers, artists. And that is important so that we move from pointing fingers or hyperbolic rhetoric to let's get together and actually do something that's positive. So you're thinking something like collective. come up with guidelines that we think allow innovation in a responsible way? Yeah, I think there's guidelines, there's norms, there is guardrail in terms of regulation,

there is also good policy to incentivize good work. I think it's really multidimensional. There's also partnership and dialogues. I'm hard to see that tech and policy world are starting to talk. A few years ago, they don't talk to each other. The flight is too long, but now... Well, it's funny because I grew up in a suburb of Washington, D.C. And these places are just so far apart, I guess, metaphorically and literally, at times. And I think bringing us together is a very important first step. You mentioned regulation. What do you think regulation would look like? This is a problem, actually, Vijay, we think a lot about, because as I am the founding co-director of Stanford Institute for Human Center AI. In addition to tech, we actually work a lot in social

sciences, humanities, and policy. So I think I want to say two things. One thing is that policy is not just guard rails. Policy is good incentive structure. America has always win when we have good public investment, especially public sector research, think tank, academia. So we have a lot of have been leading the public sector to encourage the U.S. government to put in much more resources in our universities in terms of compute GPUs and CPUs, as well as data. You know this, but I still want to say this. Not a single university today in America can train a chat GPT model, not even probably GPT 3. And all universities combined, I doubt we can train a chat chat And to be clear, you're talking somewhat about just the dollars, right? Just the cloud cost. But also to some

degree, it's very challenging to hire programmers, professional programmers, into universities. And so it's just hard to have the staffing, right? Right. We don't have the machines. We don't have enough talents and our talents are exiting and going to industry because of this. So I think there should be a lot of public sector investment as part of policy. Now, that's just one part of the, incentivization, but there is also regulation, right? That's important. And healthcare and medicine is actually a perfect example. I really think we should be careful how we think about regulation. There is applications that rubber meets the road, and we have regulatory framework. Let's not reinvent the wheels. Exactly. Especially in healthcare and life sciences, there's a very robust regulatory

framework. Exactly. We need to update it. We need to. inform, educate, and communicate with our regulators so that they know how this new thing might change some of the details. But this is the pragmatic way of delivering the right kind of regulation to businesses and consumers. Of course, there's also a catastrophic layer where it's newer to our world, for example, disinformation, especially AI mixed with social media. I think there we need, let's start with well-informed policymakers and deep dialogue of multi-stakeholders to figure this out. How do you think things look like in five or ten years? Let's give the optimistic version.

If we can build this sort of consensus to be thoughtful on how we go forward, where do you think we get to? I've been dreaming about this and I could be a little too crazy. VJ, one day I want us the entire society to move into what I would call dignity economy rather than labor economy, meaning that productivity would be increased so much by technology that each of us go to work because we feel this gives us agency and dignity and reward rather than purely this kind of day-to-day grind that many people still do and it's still not great wages. So I hope that the technology would boost productivity, but also help us to create an economy that preserves human dignity or even

enhances human dignity. That is the optimist version. We do have to work hard, though. I don't want to be totally delusional about this, because I think if we don't work hard, there are a lot of potential pitfalls that are dire, sometimes to communities that are underserved and underrepresented it will be first hit. Yes. So to end with one last question, what do you do for your own health? Great question.

Let me see. One is happiness. What do you mean by that? Tell us more. Well, I'm grateful for the family I have. We enjoy each other. Yeah. We love our family life.

We have one amazing cook in our family, which is not me. So credit to my husband. And we eat good food. And I think that's, so important. With that, maybe that's a good place to end it. Fifei, thank you so much for joining us.

Thank you, BJ. Always a joy to talk to you. If you liked this episode, if you made it this far, help us grow the show. Share with a friend, or if you're feeling really ambitious, you can leave us a review at rate thispodcast.com slash A6D. You know, candidly, producing a podcast can sometimes feel like

you're just talking into a void. And so if you did like this episode, if you liked any of our episodes, please let us know. I'll see you next time.