Using AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei Li

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. I'm not denying this. 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.

Thank you. I'm excited to be here, Andrew.

Yeah. I see vision as a cornerstone of intelligence in almost two parallel ways. One is what evolution has taught us. You know, what's the evolution of vision and animal intelligence and human intelligence. The other one is computer vision and AI, what that relationship is. So I'll go into each. Evolution. I always say that 540 million years ago animals saw the first light. These are simple sea ocean animals, trilobites and their cousins. And before that there was very little sensing. Around that same time tactile and haptics was starting also to emerge in animal bodies but there was no hearing. There's no, you know, smelling, there's no, but there's 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, changes the way your relationship with the external world. To put it simply, if you seek, you can see food, it changes your life, right? From an evolution point of view, and you become someone else's food, and also you're actively seeking food. You're actively seeking mates and all that. So really because of sensing and perception, evolution took an incredibly accelerated pace in terms of animal speciation. Fossil studies have told us that 10 million years after the first light for animals was what we call the big bang of evolution or Cambrian explosion of animal speciation. And fast forward, I think vision has always played a huge role in not only in the early evolution of animals but as well as advanced intelligence and how that emerged. You and I are both vision students and scientists. It is estimated half of the cortical activities in human brain is involved in visual function. Children were first visual before they were verbal in development. So vision really to this day plays a central role in both the evolution of animal intelligence as well as in the daily life of human life. Now in parallel, vision as a discipline or as an area of artificial intelligence really played a pivotal role in what we see as this modern AI moment in a couple of ways. First of all is the algorithms, the neural network algorithms. Neural network algorithms were first computer scientists started dabbling in that in the early 1950s. And Andrew, you might remember what's happening on the neuroscience side in the early 1950s is that neuroscientists like Hubel and Wiesel were starting to record visual cells in mammalian brain and starting to realize there is a hierarchical structure of nervous cells that stack against each other and pass neuroinformation across these hierarchy. And it goes from, you know, collecting light from retina all the way to recognizing there is a shape in front of you. And that very neuroarchitecture that we see in mammalian brain is also part of the inspiration of neural network algorithm. Now today's neural network algorithm runs on hundreds of billions and even trillions of parameters. It has the complexity that departs from what we recorded in the mammalian brain or the 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 of vision's contribution to AI that is also pivotal, which is through big data, that comes closer to my own work, is that AI around the century was a field of machine learning, a lot of different labs, different research scientists were trying out different algorithms and it's not just neural network, there are other methods, jargon words like Bayesian methods, support vector machine methods. It doesn't matter what these methods are, but it's an explorative phase that we're trying to get these algorithms to work so that we can 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, first year faculty 2006 at Princeton and my students and I are looking at these algorithms and how little data were fed into these algorithms to learn. So I turned to cognitive neuroscience literature, namely vision literature, and started to study how much humans learn, how much humans can see and the numbers were incredible. Humans by age six can learn tens of thousands of different object categories and the exposure to visual world is also massive. Right? Babies can see most of the time the moment they're born. So they're inundated with this big data. So 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 are really focusing only on algorithm and said that we need data. We need data to drive these algorithms. So long story short, we led this ImageNet project that collected the first ever internet scale large dataset for the field of artificial intelligence, but really through the field of vision because ImageNet is a collection of 15 million images. And the goal of ImageNet was to drive machines to recognize everyday objects, you know, microphones, cups, chairs. And that work converged with the advances in neural network algorithm as well as in GPU computing. And by 2012 that work, that the convergence of the three elements of modern AI became the defining moment of what modern AI is.

I want to definitely double, triple click on the convergence of this technology. I think around the second decade of the 21st century. So 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 flops going through algorithms, right? You need that speed. Then you also have, after many decades of research, neural network algorithms getting more mature. You know, starting as we said, 1950s people start to create 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 of node that takes some input and outputs another input and within it it's just a function, a very simple function. So you stack them together. That's what neural network is. But by the time it's in the, after, you know, around 2010ish, the maturity of these algorithms have gotten to a level that it's becoming really good. But also last but not the least, the recognition of big data. Internet definitely fueled that. It made data more available. But the reckoning moment of wow, big data needs to be part of that equation. We need to use big data to drive these algorithms to learn these patterns. So this convergence of these three things really set off the revolution of AI. The specific moment is also worth mentioning because you mentioned face recognition, is this ImageNet challenge. My lab put forward that starting 2010 after we collected this humongous dataset, we at that point GPU was not yet mature and we put out a public challenge for the research community for multiple years in a row and invited people to solve this major computer vision problem called object recognition. The task was very easy. We have a dataset of a thousand different categories of objects and this dataset is more than a million images large. It's what we call the testing dataset and the task for the algorithm is I'll show you a picture. You have to name the main objects inside and if you guess right you get a point. If you guess wrong you don't get a point. So that ImageNet challenge, we later a couple of years later benchmarked human performance by a very smart graduate student at Stanford and that was roughly 4%. So random chance will be one over a thousand, right? 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, ImageNet dataset and GPU. Even that year, even though the error rate was cut to, oh by the way the human performance error rate was 4%. Sorry, I need to correct that, the error rate was cut down to the teens. It wasn't where human performance was. So this is looking at images and assigning a one out of a thousand labels. Yeah. But 2012 was so momentous that year because the error rate from previous algorithm dropped a lot by this neural network algorithm. And we know in the research community when something this drastic happens it means an inflection point. But it still took another three years, I remember by 2012, 2016, for the algorithm to beat humans in naming a thousand objects.

I don't think the time pressure was the main issue even though for a graduate student to do this I don't think they want to do this forever. But I think, you know, the human brain as you know has limited memory whether it's long-term or short-term, right? So retaining the patterns of a thousand object classes even if some classes you're familiar with is not that easy. So I think there is the confusion and also for example different species of dogs get really close. And that's a challenge.

Absolutely. I think that what you see is the floodgate got open and every sub area of AI whether it's speech recognition, sound recognition, natural language processing which is more than recognition, vision, all areas got really a boost in terms of the technology. We have colleagues at Stanford who are studying whale sound, right, whale songs using machine learning and AI now. And speech recognition is another area that did so well in the early days of this AI revolution and of course the technology continues to advance. By the time the transformer paper was published around 2016, 2017 it quickly showed that it is even more powerful than the early ImageNet AlexNet algorithm. There it was not the field of computer vision that made the next big progress. It's the field of natural language processing. So because the recipe hasn't changed, now we have an even more powerful neural network algorithm called transformer, but we have even more data on the internet, or at least more readily available data on the internet in the form of texts, and now we have more powerful GPUs. So companies like OpenAI and Google quickly rallied beyond this very important technology and it still took about 5 years from 2017 to 2022 to get to the ChatGPT moment in natural language. But that's yet another step forward.

So let's just take the problem very, you have described it so well, this problem of seeing a glimpse of a cat tail and being able to recognize cat, right, or a sign of high likelihood there is a cat. The interesting thing is Andrew, generations of machine learning computer scientists have tried this problem. So before today that machines can reliably do it, there were different algorithms. You know, you can imagine a common sense way of thinking about this is, oh maybe we should recognize all the furniture to know it's indoor so it's unlikely to be a fox. So though there are rules like that that were built into previous generations of algorithms, there are also rules like, well let's only, instead of guess it's a cat, let's only guess one out of the 10 potential animals, you know, cat being one of them. That limits the search or guess space and that would help. So many ideas were tried. So when was the moment it became much more reliable? It's this current era when the huge data that these algorithms have learned, let's take Gemini or GPT, have learned really created the capability in the machines' learned space, so much knowledge, so much pattern that when presented with this more or less maybe a newish photo of a cat's tail sticking outside of a bookshelf, that pattern activated the learned, what we call learned weights or learned parameters that put the machine's assessment or guess of this object closer to what it has seen which is likely to be a cat tail or just tail because there's just so much data. This is where Andrew, as neuroscientists, I think we depart from the human brain because that child who learns about what you say kitty cat would not have the chance to download the internet of images of cat. They likely have seen three cats, 10 cats at most, but yet they're able to identify that tail as a cat tail instead of a fox tail through a different kind of learning pathway. These are the mysteries we haven't fully solved. But I do want to point out that departure between today's AI algorithm that is learned with a humongous amount of data versus how humans have evolved.

Yeah. So it's really funny you asked this and you put it beautifully. I never thought to put it in this way for a public audience but that moment came when video became part of the training data. So see again I'm going back to the training data. So around 2023, very shortly after the ChatGPT moment, multiple research teams start to put video into the training data. Of course, I'm not going to get into the nuance stuff, the algorithm. There's a little bit of changes and variations. So, remember January 2024, Sora was released and that's where people see a video can be generated literally what you just said. People can then type and say a cat running towards a mouse and then a few second clip would be generated and there would be a cat moving its leg in a plausible way running towards the mouse. At that time there were still mistakes, still to even today it's not perfect but things have gotten a lot better but that opened the floodgate of video generation as you described it. So what happened there? 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 and all that. But overall, if you zoom out, it's still part of this great neural network era, right? But what happened is that we're now able to process video data in a way, again some clever engineering, tokenize it whatever you call it. And now we can generate these short clips of videos which is 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 the algorithm shows that the cat is moving in a plausible way with the paws, you know, in a sequence, I would say the algorithm doesn't but what it does have is so many videos, especially cat on the internet, so many videos of cat, so it learned what it should look like. So in a way humans do that. Most of us without education would not know how muscles move in cats. I still don't know. You know, our colleagues in medical school might know but we have just got so used to seeing cats moving this way that we have a plausible idea of how cats move. So that is similar, that's how similar AI is. It's the statistics. It's the large amount of data that showed you what is the plausible generation of cat movements.

Totally. Andrew, I know you put a lot of thoughts into this question and I think it's a very important question and let's peel this one step at a time. First of all, TLDR short answer is I agree with you that we do have to be very careful recognizing what AI can do, is likely to do, not conjecturing over 100 years or whatever. I recognize what you just said are these extremely nuanced personalized hard to characterize or not even captured human cognitive behaviors and because they were not captured, they then were not uploaded on the internet and we don't have, today's AI doesn't have a way to do that. So when you call internet, which is the source of AI's data, let's be very clear what is internet. Internet is not some random thing. Internet is the biggest collection of human behavior in multimodal forms. Let's break it down further. Internet has the world's population typing on it for many, many, at this point multiple decades. That typing is a sensing mechanism that captured everything from teenager chit-chat all the way to deep scientific articles who got digitized and get uploaded. Right? So that capturing human language is what internet is super good at. Then internet captures images. How? Because we now have digital cameras. That's so prevalent in smartphones and digital cameras. So that humans love taking photos from, you know, the cat in your house to selfies to beautiful, you know, BBC captured photos. Those also got uploaded in our digital sphere. On top of that, there's videos. Videos now has sound, has movements that also got uploaded to our digital sphere. On top of that, there's music. We're not even getting into the legal discussion of copyrights, but let's just table that aside. I'm just talking about the forms of data. The speeches and singing and music and orchestra that also got uploaded into the digital sphere. So now we have created this humongous library of human knowledge in words, human behavior in videos, human expressions or even nature's whatever in sound and now AI gets trained on that. That is why it's so powerful. This is why especially in the words front that AI can recognize patterns, can synthesize patterns because so much of this is already there. But the thing that you just talked about that when let's say Picasso had that incredibly profound thought about that particular way of expressing that portrait of the young woman, that thought has never been captured. In fact, as neuroscientists, if I ask you which brain area did that thought come from, you don't know, right? Is it Broca? 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. You can call it cat 231, whatever name you can give it, that thought is not captured therefore it's not on the internet therefore AI has not seen it. So 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 way. Did you remember Move 37?

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 I think it's the third game out of the five games that AlphaGo as a computer algorithm made a move that the human masters of Go never thought about. And that is an incredible move, right? Because it really, humans collectively, these are the masters, never thought about it. But if you really go deep into what AI did there, it was because first of all Go is a highly mathematical game. It has very clear mathematical objective, very clear mathematical rules in terms of move. So when AI having a bigger compute and ways to retain how many moves it can remember, it was able to do things that human brains don't typically do. So 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 problem of mathematics and how AI can contribute to that and he was very positive. He said there are many problems in today's mathematics. As hard as they are, even as say a Fields medalist, I probably have forgotten there are known methods in math that can solve these problems because I have a human brain. I don't remember, I don't know all of math's, you know, solutions in the past hundreds of years, even if I were a Fields medalist. So AI can help us to solve these problems. But as a mathematician, he was also telling me, he said, I don't know if AI can solve all of math problems because some of these math problems require solutions that have not been invented that will push creativity to a whole different level. And this is where, you know, we should be curious, is it going to be a human creativity or AI would go through its iterations of improvement and get to a point of creativity that humans don't have or is it a combined creativity. My current conjecture is hybrid, is that humans working alongside AI would help us to solve these problems whose solutions have yet to be invented. And then what you said especially you touched on emotion is even more personalized. This is not necessarily logic. This is not necessarily deductive reasoning. This is maybe Andrew you look at this cup and say it's a great cup. What if it evoked 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 on the internet and no matter how mighty AI is today cannot access that. So that 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. You can call it expression. You can call it storytelling. You can call it in many ways. But that's where AI cannot access.

No, we're not. And this is why I keep saying, this is why it bothers me when people talk about AI. Some people make it sound like it's replacing humanity. But what we really — what you describe — is about enhancing and augmenting humanity. Right. This is where it doesn't even have to go as sci-fi as a smart headset accessing your brain waves. Just AI learning your patterns of writing can already help you to be, you know, a better communicator, a more effective communicator, a more efficient communicator, and that is an empowering capability that we could unleash in today's AI.

I think one of the most important things, Andrew, that as a neuroscientist and also faculty we know is agency is so important for humanity. You know, that boils down to motivation, agency, and dignity at every individual level. And I think 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 try to talk like this work will take away agency from people.

Yeah. 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 take away, you know, and say, well, since you don't understand this, let me decide for you what's good. That is not good, you know. And 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 a positive one or a negative one, there's a rhetoric of, you guys don't know what this is and I will tell you, I will make you happy, safe, whatever it is, and I will decide for you. These are not healthy and not helpful.

Absolutely. Well, thankfully you're breaking the mold of the, you know, the phenotype you just described. And there are a few others, but you've really been doing this at the highest levels, really encouraging people to think about the collaboration that is AI, the agency that exists, and whether to use it or not to use it, and so forth.

One of the agency I do think it's important for individual humans, whether you're a student, a teacher, a doctor, a policy maker, is learn about this. Not necessarily learn about how to code. I don't think that's necessary, depending on your job. So for example, if you're an artist or if you're a teacher or a doctor, you don't necessarily need to code, but learn about what this technology is, learn about how you can use it yourself to empower yourself, your learning or your work or your expression. By learning, one feels more in control. By learning, you're less scared of trying. And by learning, you retain that agency and that dignity, because at the end of the day, no matter how advanced technology is or medicine is, as humans, we want that benevolence that helps us to live better, keep our dignity, and make our community better.

Yeah. There's a power in not sharing. There's also a power to say, just trust me, I will tell you. And neither as educators — that is, we don't go to our lectures and say, just trust me, you know, 2 plus 2 equals 4. We actually say, here's how you break it down and learn about it so next time you can do it yourself. Right. I also think that especially your podcast is so important as part of public communication, education of knowledge. I also think that we need to hear voices of different backgrounds, right? So because there are plenty of scholars, technologists, builders, 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.

Yeah. No, Andrew, this is probably — perhaps you touch one of the most exciting usages of AI, which is scientific discovery. And in the case of biomedicine, you know, 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, I don't even know how long, it relies on smart humans retaining what they have learned from other smart humans and doing things at the speed of our own muscles, I guess, you know. Most likely, of course, there's like supercolliders and all that, but by and large, the ways of doing scientific discovery — human brain or scientist's brain — are 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, can go across disciplines in ways that you and I cannot go. So for example, we happen to be both in the vision neuroscience AI domain. I know nothing about, you know, olfactory — zero. Like I don't even know how to spell most of probably these words that our colleagues know, right? So it's so hard for our brain, but now we have a tool that can break open. So I think that we need to change, we need to use this tool. We absolutely — I was just thinking 150 or I don't know exactly when years ago, electricity changed everything in our life, right? I'm sure that's a moment we were thinking about the changes, the opportunities, the scary moment. I think we have to come to reckon that scientific discovery is one of the most exciting opportunities for AI and for health, right? 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.

What information did you provide? Just your subjective —

Okay, good.

Doctor can use this tooling. By the way, I have a very interesting example. You know that we had to reschedule our 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, incredible surgeon, was driving the robot. So it was a deep human-machine collaboration. After the surgery, I asked the surgeon, I said, do you imagine if say you've done a million — which is impossible for a human surgeon — but let's collect all of human surgeons for this liver, this type of liver surgery data. Can we possibly train an automatic AI to do this? The answer was not clear. So we went a little bit down the rabbit hole because liver is a very complicated organ. It's extremely vascular. It has a lot of vessels, and everybody's liver is very different. So given the reality of how many patients undergo liver surgery per year, even if you aggregate the world's liver patient surgeries, you might not have enough data to train these algorithms.

So this speaks of a very important fact that AI learns from patterns. When the patterns are not abundant, then we have to be careful. We have to know how to use AI or how not to use AI. You know, in this case, having a human collaborating with the robot is way better than an underlearned robot doing the surgery by itself. But the same issue might be true for surgeons, because how many surgeries a surgeon can get trained on. So these are opportunities that humans and AI can totally collaborate with 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 infinite possibility? These are all incredibly open scientific possibilities that are waiting ahead of us. But then there are situations like your situation where the vertigo versus low blood pressure probably have been reported so many times that in the database there's enough of that that AI has learned that. So we can then now take advantage of that for people who don't have immediate access to doctors.

It did. It actually lost 10x less blood than a typical surgery, thanks to the laparoscopic capability of a robot surgery.

Such an interesting question, Andrew. So I do want to separate intuition from creativity for the sake of argument here, and maybe we'll come back to merging. So let's talk about this intuition of, given my sibling love, what's going to happen, right? Is it really intuition? So today when you go to an AI chatbot, you're going to prompt, you know, I'm a Stanford professor and a neuroscientist, give me this information. That is already called context. I don't know if you call it intuition, but because you gave that piece of information, the AI's answer for you is already going to be different. If I type that I'm a 14-year-old teenager, you know, loving race cars, even if we ask the same question, it'll have a customized answer. That is a mathematical — I wouldn't call it energy. I want to be — that is just a mathematical fact of how these algorithms take these contexts and tailor the outputs, and it's called context. It's not that deep in computer science.

That's one type of intuition that is fairly shallow because you already are able to use language to describe it, or you can say I'll upload an image, that also is already expressible, and then AI gets it. The deeper intuition you just said is like you don't even know where they come from, right? Like is it because I smell something? Is it hormones? Is it, you know, the mixture of mood? Is it my breakfast? That intuition — what would AI do with it? That is what I would say is inaccessible. There's no sensory apparatus yet that can glean that data and feed it to — not only AI, cannot even feed it to, you know, for example, sometimes as a couple you might have a moment that you're just rubbing each other in the wrong way.

If you're really familiar with each other, you kind of can sense it, but you can't quite tell. Maybe you just leave quietly, leave that person alone. So that means whatever that intuition that person has, 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 human, a different human, nor a machine can do anything about it because there's no access to that highly individualized intuition. There's no technology that can do that till you say we put brainwave collectors or, you know, skin conductance sensors. I mean, by the time we do those, maybe they become accessible. So we have to recognize. So what I'm trying to say here is, what's not very deep is, is the data accessible, you know, either through language or through picture or through imaging or through brain waves, whatever it is, it needs to be an accessible piece of information. If it's accessible, then if we have collected enough of that, you can train machines with, or if a machine is well trained, it can, like you said, in a private way — forget about privacy breach — but in a private way, the machine can probably use it. What I'm trying to do, Andrew, here is not to make it sound mystical, but try to give it a scientific process to describe if it were to happen, how would that happen?

So you could build this in the mathematics. So certain things, whether you call it motivation or in the machine learning world we call them objective functions, you can build certain things into math. For example, now you go to say GPT, it has different modes, like think deeper mode or like give me a quick answer mode. If you don't know how this works, you're like, oh, this is interesting, one has more urgency that gives me a quicker answer, the other one has to go deeper into the search, right? And take longer to give me the answer. So 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. You can say, well, 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 would take longer. So 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, right? Is that there are cognitive states that humans truly just — whether it's motivation or urgency or fear or love — that is very hard to access and express. And do machines have it today? No. Let's make it very clear. We tend to imagine that the machines feel, or they don't have that data, they don't have that mathematical objective function. So when the machine says, I'm sorry you're so sick today, it's very different from how your friend says it to you, because the machine said that because it has learned through pattern, 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, they genuinely want your well-being. They love you. They don't want to see you suffer. They have that empathetic feel of, oh, wow, if you're in pain, I've experienced pain. So that's — it's not mirror neuron, but it's at least a memory of what pain means. The machine doesn't have any of that. So we do need to make sure we differentiate that. So a lot of what drives human, what ticks human, what triggers human doesn't exist in today's machine. 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.

Yeah. Certainly, 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.

The technology is there. But we have to now zoom out a little bit and think about the social parameters, the legal implications. I mean, 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 can say every Friday the brake doesn't work. This is a trivial technology. There's a clock in the car's computer and it just turns off the brake every Friday. But we don't do that because it has deeply bad implications to our 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.

Oh my god.

But if there's just one fruit fly that somehow an escaper and you get the actual rabies virus.

I mean, 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. And back in 2018, there was no ChatGPT. But as an AI scientist, I knew that this is only going to accelerate. This is why I went to my colleagues and university leadership and said, let's put a framework. But it's not just my framework or Stanford's framework. The entire society in every way needs to wake up to the social implication, as we have done this in human history, whether it was cars or airplanes or biotech, is that it's multi-dimensional with multi-stakeholders, right? There is the professional norm, for example, you guys as biologists don't sneak into the lab and try to put rabies into Drosophila or fruit flies because that's a professional norm and your ethical training. There is industry rules, for example, IRBs — every human subject experiment today on university campuses is subject to the IRB regulatory framework so that we can look at this. And then there are laws and regulatory laws depending on if it's applied to humans versus crops or, you know. So AI has to go through the same, right? We need to have our professional norms, we need to have education. Computer scientists are not educated in ethics and societal studies, you know, they're starting to. I mean, this is why a number of universities, including Stanford, are feverishly putting that part of curriculum into our education now. Those are the norms and education, but we also should work with the government, and different kinds of governments and societies have different kinds of norms and traditions and heritage, and look at where the regulatory measures should apply AI, for example, crossing biology, FDA. I think that's a very important area to look at how AI should be used to help but also guardrail to harm so that we can avoid 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 are different from —

Education and ethics, and these are multistakeholder problems to solve together.

And that they're traps that people, in particular young people, can fall down.

Great question, Andrew. And the answers almost fall out of our previous conversations because you use the word motivation and I was using the word agency. The absolute bad outcome is that our young generation, their agency and human-level motivation of learning and living, is taken away by tools. So doom scrolling, passive watching of shorts—all this is not helping agency, 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, our hormones move in a certain way. So for the young generation, no matter how the society will be different, jobs will be different, our human body needs to go through a deeply developmental phase where learning needs to happen. And that agency of learning, that motivation of learning, cannot be taken away by anybody—should not be taken away by humans, nor should it be taken away by machines. That would be my concern, which is that if AI is not used right, the agency and motivation is taken away, then we are left with generations or generations to come who have not properly developed the brain. The other kind of danger is, in the name of agency and motivation, the tools are denied to our students because we're worried you cheat or worry you only got your answer from ChatGPT. That is very bad as well, because with the proper agency, proper motivation, proper ways of using this tool, we can go a lot deeper with AI than we have ever learned. I was just thinking about—I was a premed student for a while. Man, organic chemistry was hard, you know. I remember trying to learn the molecules, their orientations, but the TA hours are too short or it overlaps with my other class, and my professors only have a certain number of office hours. It was just a struggle to learn that, right? 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, right? I have the motivation to learn. I just need guidance. That would be such a powerful tool for me to learn. So that we should not deny students from. So both things worry me: either denying the tool or taking away agency and motivation. Of course, the flip side is great: let's find a way to keep our children's and students' motivation and agency. Let's find a way to give them the access and the right way of using these tools. Then this generation, this coming generation, and many generations to come will be way smarter than us because they are superpowered.

Not so. Let's give ourselves some credit. Plasticity does exist throughout the lifespan.

I mean, for me, it's been a remarkable discovery of what it can do.

Well, I actually have another thing because Stanford undergrad taught me something last year, and I realized before ChatGPT, sometimes I got lazy. If I have a question, I ask the person I think is 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 to ask something that's too lazy. And AI is forcing me not to be too lazy.

Prompting is very important.

That is a skill. This is why public education is so important. This is why education is so important. I would love to see our schools K-12 teaching prompting. Here's a quiz: Who is humanity's best prompter?

Socrates, if he were alive.

And taking a walk while you have those discussions.