Andrew Huberman interviews AI pioneer Dr Fei Fei Li on intelligence, vision science, and the human future of AI
Andrew Huberman speaks with Dr Fei Fei Li, computer scientist and Stanford professor, about artificial intelligence, human cognition, and the future of technology.
Summary
Dr Fei Fei Li, professor of computer science at Stanford and co-director of the Stanford Institute for Human-Centered AI, joins Andrew Huberman to discuss artificial intelligence, neuroscience, and the future of human-machine collaboration. Dr Fei Fei Li traces the origins of modern AI through the convergence of neural network algorithms, large-scale visual datasets — particularly her own ImageNet project — and GPU computing, arguing that this 2012 inflection point defined the modern AI era. She draws sharp distinctions between what AI can do — pattern recognition at massive scale — and what remains uniquely human, including inaccessible internal states, embodied intuition, and creativity rooted in personal memory that has never been digitized. She argues forcefully that AI should augment human agency rather than replace it, and that teachers, parents, and students are the most neglected population in the current public discourse around AI.
Key Takeaways
FULL TRANSCRIPT
Vision as the Foundation of Intelligence
Andrew Huberman: Dr Fei Fei Li, welcome.
Dr Fei Fei Li: Thank you. I'm excited to be here, Andrew.
Andrew Huberman: This is a long time coming. You are a luminary in the AI field, but I also consider you a neuroscientist and computer scientist, and we share a common path through vision science.
Dr Fei Fei Li: And fellow colleagues.
Andrew Huberman: And fellow colleagues at Stanford. So I'd like to start in vision. What is so special about vision and seeing and light as it pertains to AI and where it's all going? Because I think for most people those probably sound like very divorced themes, but actually that's where it all starts.
Dr Fei Fei Li: I see vision as a cornerstone of intelligence in almost two parallel ways. One is what evolution has taught us — the evolution of vision and animal intelligence and human intelligence. The other is computer vision and AI, and what that relationship is.
On the evolution side, I always say that 540 million years ago animals saw the first light. These are simple sea creatures — trilobites and their cousins. Before that there was very little sensing. Around that same time, tactile and haptic sensing was starting to emerge in animal bodies, but there was no hearing, no smelling, and absolutely no nervous system. But the first photoreceptive cells created an evolutionary force that propelled animals to evolve, because sensing the external world changes your self-perception and changes your relationship with the external world. To put it simply, if you can see food, it changes your life from an evolutionary point of view. You become someone else's food, and you're also actively seeking food, actively seeking mates, and all of that. So because of sensing and perception, evolution took an incredibly accelerated pace in terms of animal speciation. Fossil studies have told us that ten million years after the first light for animals was what we call the big bang of evolution — the Cambrian explosion of animal speciation.
Fast forward, and vision has always played a huge role not only in the early evolution of animals but also in advanced intelligence and how that emerged. You and I are both vision students and scientists. It is estimated that half of the cortical activity in the human brain is involved in visual function. Children are first visual before they are verbal in development. So vision really, to this day, plays a central role in both the evolution of animal intelligence and in the daily life of human beings.
Now in parallel, vision as a discipline or area of artificial intelligence played a pivotal role in what we see as this modern AI moment in a couple of ways. First, the neural network algorithms. Computer scientists started dabbling with those in the early 1950s. And you might remember what was happening on the neuroscience side in the early 1950s — neuroscientists like Hubel and Wiesel were starting to record visual cells in the mammalian brain and starting to realize there is a hierarchical structure of nerve cells that stack against each other and pass neural information across these hierarchies. It goes from collecting light from the retina all the way to recognizing there is a shape in front of you. That very neural architecture we see in the mammalian brain is also part of the inspiration for neural network algorithms. Today's neural network algorithms run on hundreds of billions and even trillions of parameters — a complexity that departs from what we recorded in the mammalian visual pathway — but the origin is very close to each other, about half a century ago, a little more than half a century ago.
That's one aspect of vision's contribution to AI. There is another aspect that is also pivotal, which comes closer to my own work, and that is big data. AI around the turn of the century was a field of machine learning where a lot of different labs and research scientists were trying out different algorithms — not just neural networks, but other methods, jargon words like Bayesian methods, support vector machine methods. It doesn't matter what these methods are; it was an exploratory phase where we were trying to get these algorithms to work so that we could empower the machine to read or to see.
A group of us computer vision scientists were struggling with these algorithms, and I was a very young faculty member — first-year faculty in 2006 at Princeton — and my students and I were looking at these algorithms and how little data were fed into them to learn. So I turned to the cognitive neuroscience literature, namely the vision literature, and started to study how much humans learn, how much humans can see. The numbers were incredible. Humans by age six can learn tens of thousands of different object categories, and the exposure to the visual world is also massive. Babies can see most of the time from the moment they're born, so they are inundated with this big data. We conjectured that the lack of data was a huge part of the reason for the lack of progress in AI. So we took a departure from everybody else who were really focusing only on algorithms and said that we need data to drive these algorithms to learn.
Long story short, we led this ImageNet project that collected the first ever internet-scale large dataset for the field of artificial intelligence, really through the field of vision, because ImageNet is a collection of 15 million images. The goal of ImageNet was to drive machines to recognize everyday objects — microphones, cups, chairs. That work converged with the advances in neural network algorithms as well as in GPU computing, and by 2012 that convergence of the three elements became the defining moment of what modern AI is.
Andrew Huberman: I recall somewhere around 2012 there was this debate at a vision course at Cold Spring Harbor — held every other summer — about whether a computer could learn to recognize specific faces as well as humans. Now I think most people would say computers are actually much better at it than humans, even though you have these super-recognizer people who are exceptional at this. Could you tell us how this technology went from a state where it would confuse you and maybe a cousin, or even someone who looks somewhat like you, to the point where it is now exquisitely precise? How did we get here?
Dr Fei Fei Li: I want to double and triple click on the convergence of this technology. Around the second decade of the 21st century — like you said, around 2012 — the huge convergence was the capability of GPU computing, which basically accelerated or parallelized computing so that you can have more processing power going through algorithms. You need that speed. Then you also had, after many decades of research, neural network algorithms getting more mature. Starting, as we said, in the 1950s, people began creating these very simple algorithms that behave similarly to neurons but much simpler. Neurons as you know are very complex, but here the idea is that you have one unit or node that takes some input and outputs another, and within it there's just a very simple function. You stack them together — that's what a neural network is. By around 2010 or so, the maturity of these algorithms had gotten to a level where they were becoming really good. But last but not least was the recognition of big data. The internet definitely fueled that. It made data more available. But the reckoning moment — the realization that big data needs to be part of the equation — was also critical. So this convergence of these three things really set off the revolution of AI.
The specific moment is also worth mentioning. After we collected this enormous dataset, we put out a public challenge for the research community starting in 2010 — the ImageNet challenge — and invited people to solve this major computer vision problem called object recognition. The task was straightforward. We have a dataset of a thousand different categories of objects, and this dataset is more than a million images large. The task for the algorithm is: I'll show you a picture, you have to name the main object inside, and if you guess right you get a point. We later benchmarked human performance on that challenge, done by a very smart graduate student at Stanford, and the human error rate was roughly 4%. Random chance would be one in a thousand, so 4% for humans is not that bad. The first few years, machines were not as good as humans. The turning point was 2012 — the convergence of neural network algorithms, the ImageNet dataset, and GPU computing. Even that year, even though the error rate was cut down significantly, it still wasn't at the human performance level. But 2012 was so momentous because the error rate from previous algorithms dropped so dramatically with this neural network approach. We knew in the research community that when something this drastic happens, it means an inflection point. It still took another three or four years — by around 2015 or 2016 — for the algorithm to beat humans in naming a thousand objects.
Andrew Huberman: Could I ask where this 4% error is coming from in this very smart graduate student? Is it that they don't recognize the objects, or is it recognition against time pressure — like they're being fed images fast enough that occasionally they make an incorrect assignment?
Dr Fei Fei Li: I don't think time pressure was the main issue, even though for a graduate student to do this, I don't think they want to do it forever. I think the human brain, as you know, has limited memory — whether it's long-term or short-term — so retaining the patterns of a thousand object classes, even if some classes you're familiar with, is not that easy. There is confusion, and also, for example, different species of dogs get really close.
Andrew Huberman: Mhm.
Dr Fei Fei Li: And that's a challenge.
From Object Recognition to Language and Video
Andrew Huberman: I can see the rationale for doing this in the vision domain. But has a similar thing been explored with hearing, with sounds? As humans, we are amazing at recognizing speech inflection, emotional tone, things like that. But if I had to discriminate even 15 different sound frequencies, I can tell you as a non-musician it would be very difficult for me.
Dr Fei Fei Li: Absolutely. What you see is the floodgate got opened and every sub-area of AI — whether it's speech recognition, sound recognition, natural language processing, vision — all areas got a real boost in terms of the technology. We have colleagues at Stanford who are studying whale sounds, whale songs, using machine learning and AI now. Speech recognition is another area that did so well in the early days of this AI revolution.
Of course, the technology continues to advance. By the time the transformer paper was published around 2016 to 2017, it quickly showed that it is even more powerful than the early ImageNet AlexNet algorithm. It was not the field of computer vision that made the next big progress — it was the field of natural language processing. The recipe hadn't changed: we now had an even more powerful neural network algorithm called the transformer, but we also had even more data on the internet, more readily available in the form of text, and now more powerful GPUs. So companies like OpenAI and Google quickly rallied around this very important technology. It still took about five years from 2017 to 2022 to get to the ChatGPT moment in natural language. But that was yet another step forward.
How Machines Learn Context and Probability
Andrew Huberman: For people who are not computer scientists nor neuroscientists, the natural human experience will perhaps resonate with them. When a child learns that there's something called a kitty cat, they say "cat." If they have enough interactions with a cat, they'll realize what a cat is — even if they see it from the side, from the back, and eventually if they see a tail that looks a little bit like a cat and it's behind some books, they're very likely to say "cat." Even if they've also seen foxes and other animals with tails, just based on their experience, they're making a probability judgment. And that's essentially what AI can do, what machine learning can do. But it seems to me there's a key moment that had to happen in the progression from calculators to the AI we have now — to be able to see an image of a tail and make the reasonable assumption that it's most likely a cat if it's indoors, because foxes generally aren't indoors. At what point did machine learning and AI gain the ability to do this kind of contextual learning and come up with the most likely assignment of what something is? This object constancy piece — where if something is moving and you're only getting a partial image — this isn't what most people think of in terms of intelligence, but it's part of what makes our brains so remarkable. When did AI achieve this, and how was that scripted into these computers?
Dr Fei Fei Li: Let's take the problem you've described so well — seeing a glimpse of a cat's tail and being able to recognize cat. The interesting thing is that generations of machine learning computer scientists tried this problem. Before today's machines could reliably do it, there were different algorithms. A common-sense way of thinking about it: maybe we should recognize all the furniture to know it's indoors, so it's unlikely to be a fox. There are rules like that which were built into previous generations of algorithms. There were also rules like, instead of guessing it's a cat, let's only guess from among ten potential animals, with cat being one of them — that limits the search space and would help. Many ideas were tried.
The moment it became much more reliable is this current era, when the huge amount of data that these algorithms have learned — let's take Gemini or GPT — really created the capability in the machine's learned space. So much knowledge, so much pattern, that when presented with a new photo of a cat's tail sticking out from behind a bookshelf, that pattern activated the learned weights or learned parameters that put the machine's assessment closer to what it has seen, which is likely to be a cat's tail. There's just so much data.
This is where, as neuroscientists, I think we depart from the human brain. That child who learns about the kitty cat would not have the chance to download the internet of images of cats. They've likely seen three cats, ten cats at most, and yet they're able to identify that tail as a cat's tail instead of a fox's tail through a different kind of learning pathway. These are mysteries we haven't fully solved. But I do want to point out that departure between today's AI algorithm, which has learned from a humongous amount of data, versus how humans have evolved.
Andrew Huberman: If we continue to ascend the kind of hierarchy from simple object recognition to what you and I would call higher-order brain functions — moving more towards what most people, when they hear the word intelligence, think of as creativity and imagination — let's go to a middle step and then a much further step out. Staying with the cat example, if a computer or a child learns to recognize a cat through the tail or the whole thing, and they've seen a cat move, it's a very new world at that point for that brain, because now they know that the cat generally moves in the direction of its head, not its tail. It might go after mice but might run from dogs. The next layer up in terms of intelligence is to assign likelihoods of direction to move, directions not to move, other objects that object is likely to interact with. This all sounds very basic to people, but this is how brains learn and this is how machines learn. So when was the next big inflection in terms of giving a computer AI a picture of a cat and saying, animate this cat for me, make it move like a cat, without giving it any specific instructions about how to move its limbs?
Dr Fei Fei Li: It's really funny you asked this, and you put it beautifully. That moment came when video became part of the training data. I'm going back to the training data again. Around 2023, very shortly after the ChatGPT moment, multiple research teams started to put video into the training data. Remember January 2024, Sora was released, and that's where people saw that a video could be generated — literally what you just said. People could type "a cat running towards a mouse" and a few-second clip would be generated with a cat moving its legs in a plausible way running towards the mouse. At that time there were still mistakes, and even today it's not perfect, but things have gotten a lot better. That opened the floodgate of video generation.
What happened there is actually not as revolutionary as you might think, because the bottom line is it's still data. As a scientist, I can tell you there are all kinds of algorithm tweaks and changes and improvements. But overall, if you zoom out, it's still part of this great neural network era. What happened is that we're now able to process video data in a way — some clever engineering, tokenizing it, whatever you call it — and now we can generate these short clips of video, which are frames put together, that look like plausible cat movement.
Now you might ask: does the algorithm know the muscle structure of a cat's legs so that when it shows the cat moving in a plausible way with the paws in sequence, does it understand that? I would say the algorithm doesn't. But what it does have is so many videos of cats on the internet that it learned what it should look like. In a way, humans do that too — most of us without education would not know how muscles move in cats. We've just gotten so used to seeing cats moving that we have a plausible idea of how cats move. That's how similar AI is. It's the statistics, the large amount of data that showed you what the plausible generation of cat movements looks like.
Andrew Huberman: So when people have heard that the brain is a prediction machine, a learning machine, this is exactly what you're referring to.
Dr Fei Fei Li: Yes.
The Limits of AI: What Cannot Be Uploaded
Andrew Huberman: Let's go to a really far-out aspect of brain function that we know exists in humans. These are genuine questions, not loaded questions. If you use an apple as an example, we could go from low-level seeing an apple, to mid-level seeing that an apple always drops and doesn't fly off, to the highest level — what is the equation that governs the apple's movement. That's ascending to a higher-order, more reductionist analysis. What do you think about the idea that while AI is indeed intelligent and can do things that brains can do — maybe even things that individual human brains can't do, as we know by virtue of beating humans at chess — the idea right now, as I understand it, is that AI is trained on the internet: images, discussions, videos, songs. But that's not all of human cognition. Are there aspects of AI — whether it's ChatGPT, Claude, or even the most powerful not-yet-released tools — that don't have access to features of human brain function yet, because they've never been uploaded to the internet in a way that AI can pull out?
You could put a symphony there and it follows certain rules of music and mathematics and sound, but you have thoughts all day long and I have thoughts all day long that don't quite mesh with language in a way that I can just type them out on the internet. In the world of art, we have abstraction — occasionally somebody will come up with a painting or a drawing that doesn't look like anything specific. This happens in music too, where you just feel something — there's a fundamental rule or an emotion associated with it. They've tapped into some aspect of brain function, but you can't say what it is. I feel like this is the sort of thing that is complicated for AI. You can put that piece of art into AI and say, what fundamental feature of human experience does this reveal, and it only has access to what's on the internet. How can you capture a complex constellation of feelings and experience with AI? That seems to be the gap for me. It's hard to get to these higher-order abstract representations that can't be spoken or written down or drawn. If I say, give me your experience of nostalgia for your childhood home, you could write about it, but those are just words. I can't understand your experience at a first-person level.
Dr Fei Fei Li: I know you put a lot of thought into this question and I think it's a very important one. Let's peel it one step at a time. The short answer is I agree with you that we have to be very careful recognizing what AI can do and is likely to do. I recognize that what you just described are these extremely nuanced, personalized, hard-to-characterize human cognitive behaviors, and because they were not captured, they were not uploaded to the internet, and today's AI doesn't have a way to access them.
When you call the internet the source of AI's data, let's be very clear about what the internet is. The internet is not some random thing. The internet is the biggest collection of human behavior in multimodal forms. Let's break it down further. The internet has the world's population typing on it for multiple decades now. That typing is a sensing mechanism that captured everything from teenage chitchat all the way to deep scientific articles that got digitized and uploaded. Capturing human language is what the internet is super good at. Then the internet captures images — because we now have digital cameras so prevalent in smartphones, and humans love taking photos, from the cat in your house to selfies to beautiful BBC-captured photos. Those also got uploaded into our digital sphere. On top of that, there are videos, which have sound and movement. On top of that, there's music — speeches, singing, orchestras — that also got uploaded. So now we have created this humongous library of human knowledge in words, human behavior in videos, human expressions in sound, and now AI gets trained on that. This is why it's so powerful, especially on the language front — AI can recognize patterns and synthesize patterns because so much of this is already there.
But the thing you just talked about — when Picasso had that incredibly profound thought about that particular way of expressing that portrait of the young woman — that thought has never been captured. As neuroscientists, if I ask you which brain area that thought came from, you don't know. Is it Broca's? Is it V1? Is it motor? Is it prefrontal? We don't know. Maybe it's diffused everywhere because that thought is so personalized, so special. You can call it creativity, you can call it emotion, you can call it whatever you want — but that thought is not captured, therefore it's not on the internet, therefore AI has not seen it. That is where humans still remain so unique.
But we also need to give credit to AI, because AI has learned so many things it can combine information in highly creative ways. Do you remember Move 37?
Andrew Huberman: This is AlphaGo.
Dr Fei Fei Li: Right. Move 37 has symbolized AI's creativity. I think it's both true but can be taken out of context, because that was a game when AlphaGo was playing Lee Sedol, and in the third game out of five, AlphaGo made a move that the human masters of Go had never thought about. That is an incredible move — these are the masters, and collectively they never thought about it. But if you really go deep into what AI did there, it was because Go is a highly mathematical game with very clear mathematical objectives and very clear mathematical rules in terms of moves. When AI, having bigger compute and ways to retain how many moves it can remember, was able to do things that human brains don't typically do. Is that called creativity? I think it is, but we do have to recognize that's a special kind of creativity.
I was talking to an incredible mathematician of our time and I was asking him about the unsolved problems of mathematics and how AI can contribute. He was very positive. He said there are many problems in today's mathematics that, as hard as they are, even as a Fields Medalist, I've probably forgotten there are known methods in math that can solve these problems, because I have a human brain and I don't remember all of math's solutions from the past hundreds of years. So AI can help us solve these problems. But as a mathematician, he was also telling me he didn't know if AI can solve all math problems, because some of these problems require solutions that have not yet been invented — that will push creativity to a whole different level. My current conjecture is that it's hybrid — that humans working alongside AI would help us solve these problems whose solutions have yet to be invented.
And what you touched on regarding emotion is even more personalized. This is not necessarily logic or deductive reasoning. Maybe you look at this cup and it evokes an emotion in me — a childhood moment that a gray cup might mean something that only me and my best friend share. That is a completely inaccessible piece of information in my brain that is never uploaded to the internet, and no matter how mighty AI is today, it cannot access that. So my reaction to this cup, and potentially what I would do with it because of that piece of memory, can be completely different. You can call it creativity, expression, storytelling — but that's where AI cannot access.
Andrew Huberman: I feel like at some point in the not too distant future, computers will have access to our brain activity in non-invasive ways. You know, like I might imagine in five to ten years I'm wearing something on my head — some electrodes that are just there on the outside of my skull, not bothering me, sensing my activity inside the brain, maybe also sensing my heart rate, autonomic activity, how alert I am, and comparing that to what I'm saying and what I'm doing. This is all totally within reach and it's going to happen. There's this world where a computer that I own — and I'm not worried about data getting out — is sensing all these aspects of me and picking up on the fact that yes, what I say might be important, but there are aspects of my internal state and brain activity that I'm not even aware of. And I can decide to collaborate with this and say, let's come up with a really interesting picture that I've never seen before, but that comes from some experience of mine that's important, based on whatever — and it could reveal that to me because it has access to unconscious features of my brain activity.
I think most people have an inherent interest in what's going on for them. I would love to know why I trip up in certain ways and don't have the best day, or why some days I have the best day, or where ideas come from in me. If you think about how primitively we go about this right now — one cup of coffee good, one and a half a little better, two is too much — it's kind of crazy. Everyone has a different method and we all try to get this right, and then you've aged enough by the time you get it right that you have to update it. We're probably not getting the most out of our biology and our brains at all right now.
Dr Fei Fei Li: No, we're not. And this is why it bothers me when people talk about AI as if it's replacing humanity. What you describe is about enhancing and augmenting humanity. This doesn't even have to go as sci-fi as a smart hairnet accessing your brain waves. Just AI learning your patterns of writing can already help you be a better communicator, a more effective communicator, a more efficient communicator — and that is an empowering capability we could unleash in today's AI.
I think one of the most important things — as a neuroscientist and also as faculty, we know — is that agency is so important for humanity. That boils down to motivation, agency, and dignity at every individual level. We need to recognize that we need to think about AI as a tool that helps us in our agency. It should not take away our agency, and people who lead in today's AI should not talk as if this work will take away agency from people.
Andrew Huberman: I think people who are very familiar with the technology — whether it's computers, biology, or any technology, cars for that matter — they become such nerds of that thing that we forget that it can be scary to people, and that the language around it is essential. I remember a time in the early 90s when genetic testing was viewed as this thing — would you want to do a blood test? Because you might see something that could really scare you. That discussion is happening now around self-elected MRIs and things like that. I come from the stance that more information is better. But I've come to understand that not everyone feels that way. Some people don't want to know.
Dr Fei Fei Li: But they should have the choice. In the meantime, we should have enough public education and communication to let people know the pros and cons, but not to deny them the choice, and also not to say, "Well, since you don't understand this, let me decide for you what's good." That is not good. The rhetoric around AI right now is getting really skewed because people who know what this is tend to talk down at the public — whether the motivation is positive or negative, there's a rhetoric of "you guys don't know what this is and I will tell you and I will make you happy or safe or whatever it is and I will decide for you." These are not healthy and not helpful.
Andrew Huberman: I agree. And I think one of the reasons for starting this podcast was to showcase the scientists and physicians who really have a benevolence about them and have no interest in dumbing things down, but do have an interest in people understanding things. Many people would feel that health information is among the more important things to understand.
Dr Fei Fei Li: Absolutely. And I think we need to hear voices of different backgrounds, because there are plenty of scholars, technologists, builders, and thinkers out there who have been dealing with AI, using AI, thinking hard about how to use AI to empower people, and these voices are so important.
AI in Medicine and Scientific Discovery
Andrew Huberman: Since you're here, I'm going to go next to something that I think most everybody would agree would be a wonderful thing if it existed, and it's already starting to happen — the use of AI to augment health discovery, treatment of disease, and so on. Using the AlphaGo example from before, and the cat example, those follow certain rules. AlphaGo is a very complicated set of rules, but if you learn them, there's a constrained set. With the cat, it seems unconstrained — infinite possibilities — but it's constrained enough that machines and humans can learn it really well.
When you start getting into medicine, there are rules of medicine. There are rules of science. You have a question, you pose a hypothesis, you test it, you try and rule out your hypothesis — the scientific method. And in medicine, every field has its methods. We observe disease, we observe who recovers, we have a case report, we do a randomized controlled trial. So there are rules, and the internet knows these rules. LLMs can be used to mine health information very well because there are constrained rules. But the rules of biology are still revealing themselves to us. Every month it seems a discovery comes out that violates the rule. I learned that action potentials are unitary — they always look the same, you either fire or not. But there was a paper not but twelve years ago that showed the shape of an action potential can vary quite a lot. It was published in Nature. Everyone saw it and then no one wanted to deal with it. It's just too much — it changes the rule. Neurons are supposed to be either graded or all-or-none, and it's in every single textbook. So now if you take a bunch of neural activity and give it the rule that action potentials can be a hundred different shapes, it completely confuses everything we understand about neuroscience.
But if you gave AI the rule that it could be a hundred different shapes of this signal, AI could probably do a lot more than even the very best graduate student at Stanford or MIT or Caltech. So I'd like to get your thoughts on how humans in healthcare, the general public, and AI can collaborate to help solve disease and ideally come up with new rules for discovery so that we can finally understand our biology at a level that can really change the course of humanity for the better.
Dr Fei Fei Li: Andrew, you touch on one of the most exciting uses of AI, which is scientific discovery. In the case of biomedicine, scientific discovery directly connects to human health and diseases. I think we're ready for a complete rewriting of how scientific discovery can be done, because for ages it has relied on smart humans retaining what they have learned from other smart humans and doing things at the speed of our own muscles. By and large, the scientist's brain is the only central character in this process. Now we have a new tool whose brain can retain a humongous amount of information, can help us synthesize knowledge, and can go across disciplines in ways that you and I cannot. We happen to be both in the vision neuroscience and AI domain. I know nothing about, say, olfactory neuroscience — I don't even know how to spell most of the words our colleagues in that field know. It's so hard for our brains. But now we have a tool that can break open those boundaries.
I was just thinking — 150 years ago or so, electricity changed everything in our life. I'm sure that was a moment of thinking about the changes, the opportunities, the scary moments. I think we have to come to reckon that scientific discovery is one of the most exciting opportunities for AI and for health — how information can be synthesized, how information can be presented not only to clinicians but also to patients, and how patients can participate in that process from diagnosis to treatment. There is just so much we can do now.
Andrew Huberman: I mean, AI was able to disambiguate vertigo from low blood pressure for me a few months back, and one of the people who got it wrong was an ENT who works on the vestibular system.
Dr Fei Fei Li: What information did you provide?
Andrew Huberman: Just my subjective experience over a day or two. It turns out it was a medication that a doctor had prescribed me that I had a mild but adverse event to. It's a weird thing to step and feel like the whole world's dropping down and then kind of spinning. I thought, this feels like vertigo, but I remembered that dizzy and lightheaded are different. So I started looking into that, and sure enough it was a blood pressure issue — it brought my blood pressure down too low. I consulted some smart doctors, none of these were at Stanford, I will say that. But it's remarkable. It was zero cost. It took a morning to know that if I drank some electrolytes at what I would have thought was an excessive level, two hours later I would be fine. And two hours later, I was fine. It's also very consoling to the patient to have this. It's not to say don't go to a doctor, but it's incredible. This exists now.
Dr Fei Fei Li: Doctors can use this tooling too. By the way, I have a very interesting example. You know we had to reschedule this conversation because my father was going through a surgery right at Stanford with an incredible surgeon. But the surgery was done by a robot — the Da Vinci robot system — because it was a liver surgery, and the surgeon was driving the robot. So it was a deep human-machine collaboration. After the surgery, I asked the surgeon: "Do you imagine if you collected all of human surgeons' data for this type of liver surgery — can we possibly train an automatic AI to do this?" The answer was not clear. We went a little bit down the rabbit hole because the liver is a very complicated organ. It's extremely vascular, it has a lot of vessels, and everybody's liver is very different. Given the reality of how many patients undergo liver surgery per year, even if you aggregate the world's liver surgeries, you might not have enough data to train these algorithms. This speaks to a very important fact: AI learns from patterns. When the patterns are not abundant, we have to be careful about how to use AI or how not to use AI. In this case, having a human collaborating with the robot is way better than an undertrained robot doing the surgery by itself. But the same issue might be true for surgeons — how many surgeries can a surgeon get trained on? These are opportunities where humans and AI can totally collaborate and might reveal the best result.
Right now the future remains to be seen. Can we create an artificial simulation of a liver that we can now train on infinite possibilities? These are all incredibly open scientific possibilities waiting ahead of us. But then there are situations like your situation, where vertigo versus low blood pressure has probably been reported so many times that in the database there's enough of that for AI to have learned it. We can then take advantage of that for people who don't have immediate access to doctors.
Andrew Huberman: Is your father's surgery okay?
Dr Fei Fei Li: It did go well. He actually lost ten times less blood than a typical surgery, thanks to the laparoscopic capability of robotic surgery.
Intuition, Creativity, and What Machines Cannot Feel
Andrew Huberman: I'd like to talk a little bit about some features that we think are uniquely human — and then I'd also like to get educated on how AI is structured to allow these things to happen. For instance, intuition. We all like to think of intuition as this mystical, very powerful thing that we own, that no one can take from us, that can't be mimicked. But I could also break intuition down to be: it's my experience over time, it's a dataset coupled to some bodily and brain sensations and some prediction cues — the last time I felt this, this happened. You could assign these rules to a computer. But there are other aspects of our deeper self — things like creativity, intuition, premonition, the idea that you really sense something is coming but it hasn't happened yet — that don't really have a location or even a network to point to.
Within AI systems and GPUs, can we actually allocate more energetic flow through particular learning rules? Could we tell a machine, based on everything you know about my sister whom I love, what is your intuition about how our relationship will evolve over time? Can it actually become sort of mind-like and come up with a sense of what might actually be worthwhile, or does it just need more and more prompts — so I'm actually doing the work?
Dr Fei Fei Li: Such an interesting question. I do want to separate intuition from creativity for the sake of argument here. Let's talk about this intuition of, given my sibling love, what's going to happen. Is it really intuition? Today when you go to an AI chatbot, you're going to prompt it — you know, "I'm a Stanford professor and a neuroscientist, give me this information." That is already called context. Because you gave that piece of information, the AI's answer for you is already going to be different than if I type that I'm a 14-year-old teenager who loves race cars. Even if we ask the same question, it'll have a customized answer. That is a mathematical fact of how these algorithms take context and tailor the outputs. It's called context. That's one type of intuition that is fairly shallow, because you're already able to use language to describe it.
The deeper intuition you described — where you don't even know where it comes from, is it because I smell something, is it hormones, is it the mixture of mood, is it my breakfast — that intuition is what I would say is inaccessible. There's no sensory apparatus yet that can glean that data and feed it to AI. Sometimes as a couple you might have a moment where you're just rubbing each other the wrong way. If you're really familiar with each other, you can kind of sense it but you can't quite tell. Maybe you just leave quietly, leave that person alone. Whatever that intuition is, they could not even express it in words or a gesture to give it to another person to use as a piece of information. So when you cannot even access that, neither a different human nor a machine can do anything about it, because there's no access to that highly individualized intuition.
What I'm trying to say is: it's not about whether it's deep or not — it's about whether the data is accessible, either through language, picture, imaging, or brain waves. If it's accessible and we've collected enough of it, you can train machines with it. What I'm trying to do here is not to make it sound mystical, but to give it a scientific process — to describe how, if it were to happen, it would happen.
Andrew Huberman: Pattern recognition based on big datasets and rules gets us a long way, is what I'm hearing. And we were talking earlier about where doctors fail and robots and machines perhaps do better, or they collaborate to do better than either one alone. As a neuroscientist, you spend a lot of time looking at cells at some point in your career. The electrophysiologists for decades developed an intuition — I learned to recognize cells based on things that were not written up in any papers. Like if there was a straighter edge along this thing and it had a certain shape and roundness, I can tell you right now that's a transient off-alpha cell in the retina. Eventually we developed genetic labels to reveal that that was true in every case. But then you also saw some that didn't fit the rule. Machines can learn that. With all that information from all those papers, now we have a pretty good parts list of the retina.
What I was trying to get to with intuition is: what are some internal states of humans that are really hard to imagine machines could recapitulate? Like motivation. Do machines, do robots get motivated? When I'm really motivated to do something, we call that urgency. I might move faster to do it. Less activation energy. Can you say to a machine, "I want you to seek this out, but with a heightened level of urgency," or are they just constrained by the mathematical rules they can work with?
Dr Fei Fei Li: You could build this into the mathematics. Certain things — whether you call it motivation or, in machine learning, objective functions — you can build certain things into math. For example, now you go to ChatGPT and it has different modes, like a deeper thinking mode or a quick answer mode. If you don't know how this works, you might think, "Oh, one has more urgency, that gives me a quicker answer, the other one has to go deeper into the search and takes longer." As a human, if you anthropomorphize it too much, you might call it urgency or motivation. But the truth is this is just a different kind of objective for the algorithm. The one that thinks quicker has a time limit or token limit; the one that thinks slower can activate a different part of the model that takes longer. It becomes actually mathematically very dry and not that deep. But for a human you can call that motivation or urgency.
But let's go deeper, because you're asking something deeper than that. There are cognitive states that humans truly have — whether it's motivation, urgency, fear, or love — that are very hard to access and express. Do machines have that today? No. Let's make that very clear. We tend to imagine that machines feel, but they don't have that data, they don't have that mathematical objective function. When the machine says "I'm sorry you're so sick today," it's very different from how your friend says it to you. The machine said that because it has learned through pattern that when someone tells it "I'm sick," you should say "I'm sorry you're sick" instead of "I'm so glad you're sick," because that data exists. Whereas your friend who hears that genuinely wants your well-being. They love you. They don't want to see you suffer. They have that empathetic feeling — if you're in pain, they've experienced pain. The machine doesn't have any of that.
So we do need to make sure we differentiate that. A lot of what drives humans, what ticks humans, what triggers humans, doesn't exist in today's machines. We operate fundamentally differently from today's AI, and we have to recognize that, respect that, and this is where public communication is so important. We cannot confuse the public about this.
Andrew Huberman: I feel like people assume there's an emotion, a person, or whatever inside of the AI chatbot because we're so language-oriented. It's talking to us, it's writing things to me. And we've gotten very accustomed to receiving communications in fairly deprived language. Texts are not like extensive prose. Language has changed. Modes of communication have changed — more deprived as opposed to more enriched.
But at some point soon, I'm guessing faces are going to start to enter the picture. How far off are we from, if you or I were to text the other person, that text actually being a photo or video image of you just talking to me telling me that? The technology is there.
Dr Fei Fei Li: But we have to zoom out a little and think about the social parameters, the legal implications. Humans are capable of doing a lot of things with our tools, but we don't do all of them. For example, today any car manufacturer could make it so the brakes don't work every Friday — there's a clock in the car's computer and it just turns off the brakes. But we don't do that because it has deeply bad implications for human society. That's where rules come in, laws come in, social norms come in, morality comes in. And I think this is where we exit the pure technical discussion of AI and need to enter the social discussion of AI.
Governance, Ethics, and the Multi-Stakeholder Problem
Andrew Huberman: Let's do that, because one thing I know about biologists or technologists is they like to go fast because it's exciting — it's the next edge. I remember long ago I had a friend studying viral vectors for getting genes expressed as experimental tools in animals. There came the opportunity to put a modified rabies virus into Drosophila, into fruit flies. Now, that's fine in my opinion if you are absolutely certain — 100% certainty — that it's a nonfunctional version of the rabies virus, because you can put other cargo in there and do all sorts of important experiments on disease. But if there's just one fruit fly that somehow escapes and you get the actual rabies virus, now you have fruit flies with rabies and those things move really fast. So there's a reason why you don't do that experiment. But it was exciting for them to think about, and then they got denied, for good reason.
The point is that technologists love to go fast. They love sensing that next edge of things. So how is it that between government, the general public, and technologists, that conversation can occur in a way that's going to satisfy each of those groups enough, not hold us back? Because we're also supposedly in an AI race right now, which warrants going faster, not slower. How do you think about this?
Dr Fei Fei Li: Andrew, this is why I returned from Google eight years ago back to Stanford and started the Human-Centered AI Institute. These are profound societal questions we had to face. Back in 2018, there was no ChatGPT. But as an AI scientist, I knew that this was only going to accelerate. This is why I went to my colleagues and university leadership and said, let's put a framework together. But it's not just my framework or Stanford's framework — the entire society in every way needs to wake up to the social implications, as we have done in human history whether it was cars, airplanes, or biotech. It's multi-dimensional with multiple stakeholders.
There is the professional norm — you as biologists don't sneak into the lab and try to put rabies into fruit flies because that's a professional norm and your ethical training. There are industry rules — for example, IRBs: every human subject experiment today on university campuses is subject to the IRB regulatory framework. And then there are laws and regulatory frameworks depending on whether it's applied to humans versus crops or whatever. AI has to go through the same process. We need to have our professional norms. We need to have education — computer scientists are not educated in ethics and societal studies. They're starting to be; this is why a number of universities including Stanford are feverishly putting that part of the curriculum into our education now. But we also should work with governments, and different kinds of governments and societies have different norms, traditions, and heritage. We should look at where regulatory measures should apply to AI — for example, crossing into biology and the FDA. I think that's a very important area to look at how AI should be used to help but also to guard against harm.
What I would not like to see is one person or a few people coming from industry and telling everybody what to do. I think that would be dangerous, because market forces are different from societal norms and culture and heritage, which are different from education and ethics. These are multi-stakeholder problems to solve together.
AI, Young Brains, and the Future of Learning
Andrew Huberman: I'd like to get your thoughts on how the human brain is being shaped by machines and how machines are being shaped by our understanding of the human brain. Many people, parents and kids, are thinking, "Oh, my kid is never going to learn anything now — they're just going to look everything up on a chatbot." But if you look back in the history of learning, similar arguments were made about calculators, computers, the typewriter, and on and on.
This hardware that we have in our heads evolved to process physical things in the world — light, sound, smells — and then it got this really cool piece up front, the prefrontal cortex, that can learn learning rules and can update those learning rules. So if anything, we were gifted with a learn-to-learn machine. I grew up in Palo Alto when the personal computer showed up — here's Pong, there's the Apple IIe — and I think the brain can mature around technology and collaborate with technology in a way that I think my life has been greatly enriched by it. But I think the smartphone, and perhaps the camera-smartphone combination, as people like Jonathan Haidt have pointed out, have created a situation where most people love these technologies for the ease and convenience, but we're all a little bit more aware now that we're giving up something too. There are traps that people, in particular young people, can fall down.
So what is the very optimistic, the middle, and the very pessimistic view in your mind of how young brains can be enriched, unaffected, or harmed by AI as it exists now?
Dr Fei Fei Li: Great question, Andrew. The answer almost falls out of our previous conversations, because you used the word motivation and I was using the word agency. The absolute bad outcome is that our young generation's agency and human-level motivation for learning and living is taken away by tools. Doom-scrolling, passive watching of short videos — all of this is not helping human agency. Learning fundamentally — respecting the hardware you're talking about — takes time, takes effort, sometimes takes some pain. That is just how our brain is. It doesn't matter how transistors move; our neurons move in certain ways, our chemistry and our hormones move in certain ways. So for the young generation, no matter how different society will be, no matter how different jobs will be, our human body needs to go through a deeply developmental phase where learning needs to happen. That agency of learning, that motivation of learning, cannot and should not be taken away by humans nor by machines. That would be my concern — if AI is not used right, the agency and motivation are taken away, and we are left with generations who have not properly developed.
The other kind of danger is that in the name of agency and motivation, the tools are denied to our students because we're worried they'll cheat or only get their answer from ChatGPT. That is very bad as well. Because with the proper agency, proper motivation, and proper ways of using this tool, we can go a lot deeper with AI than we have ever learned.
I was just thinking about being a premed student for a while. Organic chemistry was hard. I remembered trying to learn the molecules, their orientations, but the TA hours were too short or they overlapped with my other class, and my professors only had certain office hours. It was just a struggle to learn. If today I were to have an AI companion, I would ask so many questions about organic chemistry because I know where I'm stuck. I have the motivation to learn — I just need guidance. That would be such a powerful tool for me to learn. So we should not deny students that.
Both things worry me: either denying the tool or taking away agency and motivation. The flip side is great — let's find a way to keep our children and students' motivation and agency. Let's find a way to give them access and the right way of using these tools. Then this coming generation and many generations to come will be way smarter than us because they are superpowered.
Andrew Huberman: I love that answer. I have great faith in neuroplasticity and the younger generations. Even our own neuroplasticity — I find AI a great tool for my learning. For me it's been a remarkable discovery of what it can do. I tend to approach it from the position of consumer if I know nothing about something, and from the position of creator if I have some knowledge set inside of whatever it is I'm asking.
Dr Fei Fei Li: I actually have another thing, because a Stanford undergrad taught me something last year. I realized that before ChatGPT, sometimes I got lazy — if I had a question, I'd ask the person I thought was smart next to me. Now I realize I should not ask lazy questions, because it's so much easier to get information before you spend somebody else's time on something that's too lazy. AI is forcing me not to be too lazy.
Andrew Huberman: How essential is the specificity of the prompt to getting the best information out of AI?
Dr Fei Fei Li: Prompting is very important. And that's a skill. This is why public education is so important. I would love to see our K-12 schools teaching prompting. Here's a quiz: who is humanity's best prompter?
Andrew Huberman: I'm going to flunk this quiz.
Dr Fei Fei Li: Socrates, if he were alive — because that is the method of prompting. Think about it. What is the Socratic method? It's prompting and seeking truth by asking questions. We should go back to teaching kids that.
Andrew Huberman: And taking a walk while you have those discussions.
Dr Fei Fei Li: Yes.
Embodied AI, Robotics, and the Physical World
Andrew Huberman: Which is actually a good transition perhaps to this notion of embodied AI. My good childhood friend Dr. Eddie Chang, chair of neurosurgery, bioengineer, who studies speech and language — he and others have figured out the transformation of neural activity to control of the larynx and pharynx. He's brought people essentially out of locked-in syndrome so they can speak. For the first time in ten years, he has this patient who was sadly paralyzed and could speak through a computer. He has many examples of these. But the incredible thing is when he started putting an iPad next to this person — one woman in particular who is wheelchair-bound — they had a video of her at her wedding. So they knew her voice, they knew her emotive patterns, they knew a bit about how she moved her body. And she now speaks through an iPad next to her frozen real face. But she can interact with the world and it can interact with her in a completely different level of depth than if it were just a microphone — the sort of Stephen Hawking thing — and it's constantly being updated through machine learning, paying attention to the people she's speaking to and their responses. This is embodiment.
Dr Fei Fei Li: It's not just embodiment of people — embodied AI also goes into robotics. The next frontier of AI, as I have been saying, is beyond language, because humans develop first preverbally. Evolution took 500 million years without verbal communication. And the world, in the right version, would be a lot better place with robots helping humans.
Andrew Huberman: Could you give some examples? I love this idea. Robots — we've got self-driving cars. The Waymo always stops for me and my puppy. A lot of people won't stop; they'll almost run us over in the morning. The Waymo is very respectful. Where do you think this is going to show up first? And what's it going to look like if we zoom out — two years, three years?
Dr Fei Fei Li: I would say if we zoom out thirty years. I'm not saying that's the first time robots hit the street — we already have robotic cars. I'm just saying it takes longer for especially a hardware-involved technology to manifest. But I would say hopefully in you and my lifetime, I would love to see robots being part of our society, helping us.
For example, I'm a single grown-up child taking care of two very advanced-aged and very sick parents who also don't speak English. The amount of work I do is incredible. I would love to have help. It doesn't take away family's responsibility. It doesn't take away love. It doesn't take away the necessary communication. But the physical labor — certain parts of it — I would love to get help with.
We live in the state of California. What is the one thing we all experience? Wildfires. Who is fighting these wildfires? Putting humans in danger of rescue in natural disasters is not a great idea. And I was just thinking about an elderly person living alone — how do they go get groceries? How do they go get medicine? What if they want to go for a walk or go to a park? And by the way, we don't have an excess of caretakers. We have a shortage of caretakers. Our nurses are deeply fatigued and overworked. I was literally in the hospital with my dad for the past month, just watching the amount of work nurses do. On a given shift, nurses walk miles to fetch things, get medicine. Can you imagine robots helping? There are just so many ways our society can benefit from that kind of help.
Andrew Huberman: So many spring to mind. You know, crossing guards. You imagine with video that somebody who's homebound because of age or illness could navigate to the store and pick things off the shelf. It doesn't have to be so disconnected that they just program it and it comes back — that could be an option too.
I think we have to revise our notions of what this picture looks like. One thing about robots and computers that scares people is their physical hardness — the way we share space with them is very different than the way we share space with other things. I'm not thinking you cuddle with a robot, although some people might think that. But I am thinking, okay, if I had a robot that could fold clothes, vacuum, water the plants, and feed my fish — although I like to feed my fish myself. I really enjoy it. I love seeing them eat. I love being tactile with them. They'll eat from my hand.
Dr Fei Fei Li: Does your puppy like your fish?
Andrew Huberman: He does. He has his own fish tank. I just got him some tropical fish, right in front of his little — he looks at them. He's not equipped to take care of them yet, I don't think. Unfortunately, there's not enough prefrontal cortex in him. He's a bulldog mutt. They're not the smartest breed.
Dr Fei Fei Li: They only have a few learning rules, but they're very kind.
Andrew Huberman: But the idea here is, if one robot is doing one thing and another robot is doing another, it feels like a lot of hardware in my life. And I think that's kind of how people feel. But you could imagine a multimorphic robot. Do you know Baymax?
Dr Fei Fei Li: I don't.
Andrew Huberman: Disney's robot, probably ten or fifteen years ago. Google the image. This is the white medicine robot — a healthcare robot that is very spongy, like a big balloon.
Dr Fei Fei Li: More contours. Yeah. And more multitasking from the same robot. That feels like a world I could adjust to more quickly than the idea of my world filled with robots.
Andrew Huberman: Yes. And I think as we imagine the future and talk about how we imagine it, I keep coming back to the word agency. Humanity should have the agency to decide how we imagine this. It cannot just be a company or an investor deciding that the world should be filled with metal robots. Our society should be collectively and proactively imagining this, and one thing I worry about in this AI rhetoric is that the public is put in a position of being reactive when it feels like some people are just deciding, and the multiple stakeholders are not participating in designing the future together.
With your example of your father's surgery — crossing the problem where there's a vulnerability with a robot that clearly makes things better — the picture changes in the right direction. I'm thinking of a few examples: most people would agree that if their kids could walk themselves to school and home, it would be great, but you worry about safety. But if a robot was a good guardian of your kid — to the point where it could alert the authorities or even physically protect your child — that would be awesome. Give them more agency in the world.
You think about some of the darker but unfortunately real predatory behavior online. Parents can only oversee their kids' behavior so much. But you could imagine a kind of avatar in there with you that's really advocating for you, that can spot things and keep predators at bay.
Dr Fei Fei Li: Here you go. That's a great startup idea.
Andrew Huberman: But here's what's missing from the picture for me. I remember seeing this incredible guy walking around downtown Palo Alto when I was a postdoc and when I was a kid growing up working at the Palo Alto Sport World — and that was Steve Jobs. No shoes, kind of looked like a hippie. He shouted at people at work, and HR wouldn't look too kindly upon him nowadays. But he understood that these things we call computers needed to have rounded edges. They needed to fit seamlessly in our pocket. They needed to have Bob Dylan on the landing page or whatever, so that it softened the relationship to technology. Some people would say it went too far, that it was a Trojan horse. But I don't think so.
Somebody who really understands human nature needs to allow these clearly benevolent collaborations between robots and humans to happen, because as you've pointed out, there's a hardness to either the way they're being presented or what they're capable of sharing that is a real separator.
Dr Fei Fei Li: Yes. And I'm not a therapist, but if I could wrap my arms around them, I'd be like, "Listen, you're the smartest people in the room, guys and gals. But people don't like you because they don't understand you, and maybe you need a collaborator to help you share your vision in a way that isn't going to allow the press to build this chasm." Because the media is guilty of building this chasm — these technologists, they're coming for us. I think that's a total trick of media, just to put money in their pocket. So who's the Steve Jobs of this? Someone who really understands human beings.
Andrew Huberman: Many of them — there are many of us. Stanford started the Human-Centered AI Institute. Well, there's you.
Dr Fei Fei Li: Okay. But there are many. There are plenty of entrepreneurs who are doing incredible startups on AI for drug discovery, AI for healthcare, AI for aging, AI for mental health. These people care about AI. There are many designers and product managers who are trying to — I do think the megaphone is too much focused on people pumping their chests and talking about tech in a certain particular way. Even this podcast is making a positive difference, I hope — putting that human angle, the rounded human perspective, the human future into these conversations.
I don't feel despair, Andrew. I'm an educator. I'm a builder. I'm a technologist. I see many people around me, including my entire startup. These brilliant young technologists could join any startup or company they want, but they come to World Labs because they want to empower people. I see many people, but I don't think there's enough. The public discourse is not balanced right now. There is too much extreme rhetoric — either extreme doomerism and lack of safety that's just freaking people out, or extreme utopianism as if technology can do no wrong. That's disingenuous. People would say, "Well, okay, you're the haves, of course you say that." We should come to the middle and talk about what this technology is, how to use it, and how we can collectively have that agency to guide the future.
Andrew Huberman: One thing that was pointed out to me by one of my podcast colleagues — that should have been obvious but wasn't — is that people don't really want to hear stories about machines. But people love hearing that some person cured their dog's cancer, or their child that was experiencing crazy symptoms, the doctors had no clue, and at their fingertips AI solved the problem. These are the stories that really need amplification, because we can relate to them and they're beautiful stories. They're incredible stories, but they're not getting nearly as much attention as the other stuff.
Dr Fei Fei Li: That's a challenge. Traditional media doesn't really care about the long arc of things — they're on a twelve to twenty-four hour cycle.
World Labs and Spatial Intelligence
Andrew Huberman: I would love to learn more about your startup, because you don't pick projects haphazardly. What is the project? What's the goal?
Dr Fei Fei Li: My startup, co-founded with a couple of other co-founders, is called World Labs. We co-founded it at the beginning of 2024. It really is, for me, a kind of my life's work. We both come from vision, and the recognition that there's more beyond language intelligence is what really motivated me to think hard about what's the next chapter of the AI frontier. We recognize that unlocking spatial and physical intelligence is really the next chapter — not excluding language, of course, the language technology is incredible — but where we can devote more time to build models and eventually products that can help unlock capabilities in spatial intelligence, like generating 3D and 4D worlds that are deeply useful for creators, for robot training, for architecture design, to enable those interactive environments. Whether you're talking about healthcare usage, education usage, robotics usage, or industry usage, these capabilities go beyond language per se.
World Labs was founded based on that premise. We are still a young company, very much a model-focused company where we're building this foundation model. We started with a lot of PhDs, but now we're starting to build products. It's still the beginning. It's very exciting, and as a technologist, I feel deep in my heart I'm a builder. Maybe it's also because I'm an immigrant — that rolling your sleeves up and getting in with the young generation that's so incredibly smart and just building something from scratch is just so exciting.
Andrew Huberman: I recall a time not but fifteen or twenty years ago when there were cars driving around taking images. There are certainly aerial views as well. You could imagine little tiny drones picking up information about every nook and cranny of the fjords in Norway. Has that been done to sort of map the three-dimensional world?
Dr Fei Fei Li: First of all, let's not make it sound scary that drones are getting into people's homes and properties. The ability to capture imagery of the world is really rapidly advancing. Our cell phones are incredible sensors — they're not drones, but people take a lot of photos. And of course our camera technology has improved. What World Labs is doing is not just taking real-world images. We allow people to imagine what's in their mind's eye. As long as you could type a sentence, show a picture, or sketch what you imagine, we try to turn that into worlds and environments. Why is it useful? Because the entertainment industry would use it, the design industry would use it, the robotics industry very much would use it for training environments. So the combination of capturing what's in the real world as well as capturing what's in your imagined world is the new frontier.
AI and the Future of Storytelling
Andrew Huberman: If you don't mind, I'd like to take a couple more minutes and talk about moving from imagination to something. Because this is Los Angeles, it occurred to me that a lot of people write scripts and then try to get their movie made. But with AI, in theory, you could take a script and give it to AI and it could make the movie — going from words to pictures to video. Has a successful movie been made start to finish using AI?
Dr Fei Fei Li: This is a very nuanced topic. This is where we also get into people's wariness of AI and creativity, where if not careful it might sound like we're taking away storytellers' and creators' jobs. Let's separate the job conversation from the technology conversation a little, even though they're entangled.
Technology has advanced enough that taking scripts and generating video shots is getting really good. We have seen short movies, even almost feature-length films, being assembled by AI tools. There are many companies — US companies, Asian companies — creating this technology. But what remains deeply human and important is every part of storytelling and story creation. There are humans behind it with their unique emotion, story technique, how they see the world, how they move the cameras, how they characterize characters. A lot of that is what Hollywood and novel writers are about. So how do we meet the human need and human desire for storytelling with modern tools is actually a challenge, because there is a fear very much coming from Hollywood that AI is taking over and that storytellers, actors, and screenwriters' jobs are being impacted.
I think they are being impacted, but how is it being impacted, and what are we doing about it? Who is working in a constructive way? This is not my industry per se, but I would love to see much more nuanced work and nuanced public discussion about that. I do think, just like healthcare, AI is absolutely changing the way we're doing storytelling. Speaking of which, I have a co-founder whose name is Ben, and Ben and I met with Ben Affleck, who is also thinking very avant-garde about using AI tools in filmmaking. Having conversations between technologists and storytellers or movie makers at this moment is critical.
Andrew Huberman: I feel like in every example of technology, there's some crossover point where somebody who's truly an insider embraces a technology and then it just kind of takes off. You need an insider and an outsider to do it right, because you have to understand both cultures and how to include the industry and the people. So I really hope — because World Labs works with the VFX industry as well — that our customers and users feel empowered. Technology should not be taking their jobs away. Technology should be making their jobs better, superpowering their creativity. That's how I see this technology and how I would like to work with users and customers.
Dr Fei Fei Li: It's wild to think that when I was a kid on California Avenue in Palo Alto, there was this store — and it was just a photograph store and camera store. You'd go in there, get your film developed, and there were all these guys behind the counter telling you about long-distance lenses. None of that exists anymore. Everything went digital. But there are still camera stores. Industries can morph. They don't always get obliterated.
It morphs. People also get reskilled and upskilled. We are working with a lot of creators who are using AI tools because they see where technology is going and they want to reskill and upskill themselves. Moments of change are moments of both opportunity and loss. We need to be really thoughtful about that.
The Young Generation and the Role of Teachers
Andrew Huberman: My last question is about the young generation. How do they feel about AI? I'm talking about kids between the age of seven and twenty.
Dr Fei Fei Li: That's literally my kids.
Andrew Huberman: Yeah. So I might have asked that question for a reason. How do they feel about it? Are they excited by it? There is this phenomenon where computers come along and your handwriting teacher gets nervous that people aren't going to know how to write. These stories have been around for a long time about how we're just going to dissolve into a puddle of our own neurons if we don't embrace the past as much as the future. How do the kids feel? What do they think?
Dr Fei Fei Li: This is actually my pet project as an educator and technologist. Everywhere I go, I try to talk to students, parents, and teachers, because I think that is the most forgotten population. Our policymakers, technologists, and investors — they don't talk about teachers, parents, and students. They all have opinions and they all have kids, but they don't talk about it.
I always have hope for kids. Maybe because I'm an educator. I think the biggest thing humanity never learns is the older generation lamenting about the future generation as if the future generation doesn't know anything — they're rude, they're forgetting the past. But if you look at the arc of history of humanity, by and large, we advance for the better. Now, I'm not denying the atrocities. I'm not denying the setbacks. But fundamentally I'm an optimist in humanity. I look at kids — they're curious. Of course, they get massively entertained by this technology, but they also are starting to use it.
What I worry about are teachers and some parents, because I think our society today and especially Silicon Valley are not doing them a service. We're forgetting about them. We are lecturing them. We are berating them. We are looking down at them. They are the most important people in our society. We should be talking to them. We should be uplifting them. We should be supporting them. We should be providing resources to them. K-12 teachers and K-16 teachers share the most important critical burden of our society.
I'll tell you a real story. November 2022, ChatGPT came out. Obviously I'm an insider in terms of technology, but the first thing I did was email the principal of the elementary school my kid was in and say, "I would like to come and guest lecture for your students and teachers." It's not because I'm so special. It's because I want them in real time to know what's happening. Because nobody — nobody in Silicon Valley, no investors, no multi-trillion dollar companies — when ChatGPT came out, the first thing they thought was, "What about our teachers in the neighborhood?" Nobody thinks like that. But we need to be talking to teachers. We need to show teachers. Of course, they're going to ask the question about what if kids cheat. It's okay. They ask those questions. Let's just show them. Let's work with them and empower them to come up with ways to deal with that. They are smart too. They are eager to change. They're just forgotten.
So I have hope for kids, but in order not to have a blind hope, I think we should all remember our teachers and help our teachers and parents so that we can help our kids.
Andrew Huberman: I absolutely love that answer, and I know that sentiment is shared by many many people listening. God bless the teachers, and they need help, support, and information, because now they turn on most podcasts and they're just scared. They hear the doomerism, or they hear "don't worry, it's utopian." Neither of these messages can help our teachers, and if they're not helped, our kids are not helped.
Dr Fei Fei Li: Couldn't agree more.
Andrew Huberman: Fei-Fei, thank you so much for taking the time out of your incredibly busy schedule. I'm so glad to hear your father's okay. And that is also part of your schedule — taking care of your parents, kids, and all the rest — to come educate us on this thing that's not just important, it's a major wedge of where we're at and where we're headed. I share great optimism with caution, even more so on the basis of what you shared today. And also thank you for teaching us more neuroscience as we went along, because these machines are informed by the brain and the brain is informed by these machines. This is the world we're living in, and I have great optimism in no small part thanks to the fact that you exist in this world. Thank you for taking the time to come here and share.
Dr Fei Fei Li: Thank you, Andrew. I really appreciated this conversation. It's a civilizational moment.