AI is smarter than you think with Shane Legg of Google DeepMind

0:00Wondery Plus subscribers can listen to How I Built This early and add free right now. Join Wondery Plus in the Wondery app or on Apple Podcast. Is your AI built to work with the tools your business relies on? IBM's AI agents can easily integrate with the tools you're already using, so they can work across your business, not just some parts of it. Get started with AI agents at IBM.com. The AI built for business, IBM. Listening on Audible helps your imagination soar. And no matter what you like, Audible's romance collection has something to make you swoon. Here's your invitation to have it all. Find a book boyfriend in the city and another on the hockey field, Or if nothing on this earth touches your heart, you can always find love in another realm.

0:43Hear modern rom-coms from authors like Lily Chu and Allie Hazelwood, the latest romantic-y series from Sarah J. Maas and Rebecca Yaros, and Regency favorites like Bridgerton and Outlander. And that's only the beginning. Audible has an incredible selection with over one million audiobooks, podcasts, and Audible originals, all in one easy app. And you can enjoy Audible anytime while doing other things. Household chores, exercising, on the road, commuting, you name it. Audible makes it easy to drop into your fantasies during your everyday routine without needing to set aside extra time. There's more to imagine when you listen.

1:20Your first great love story is free when you sign up for a free 30-day Audible trial. Visit audible.com slash built. I love traveling with my family. We did an awesome trip this summer. And one of the things that made the trip so special were the Airbnb experiences we did. Immersive tours, cooking classes, a chance to get coffee with a world-class barista. I had so much fun on those experiences that I decided to host my own Airbnb original experience in San Francisco. Designed to help you think about how to unlock your next big move in your career or even in your life. To learn more about my Airbnb original experience, head to Airbnb.com slash Guy.

2:09Hey, it's Guy here. And before we start the show, I want to tell you about a super exciting thing. We are launching on How I Built This. So if you own your own business or trying to get one off the ground, we might put you on the show. Yes, on the show. And when you come on, you won't just be joining me, but you'll be speaking with some of our favorite former guests who also happen to be some of the greatest entrepreneurs on earth. And together, we'll answer your most pressing questions about launching and growing your business. Imagine getting real-time branding advice from Sunbum's Tom Rinks or marketing tips from Vaughn Weaver of Uncle Nearest Whiskey.

2:50If you'd like to be considered, send us a one-minute message that tells us about your business and the issues or questions that you'd like help with. And make sure to tell us how to reach you. Each week, we'll pick a few callers to join us on this show. You can send us a voice memo at hibt at id.wondery.com, or you can call 1-800-433-1298 and leave a message there. That's 1-800-433-1298. And that's it. Hope to hear from you soon. And we are so excited to have you come on the show. And now, on to the show.

3:40Hello and welcome to How I Built This Lab. I'm Guy Raz. So artificial intelligence is already changing the way we live and work, believe it or not, even if you don't realize it. And so far, most of the AI applications we've seen have been focused on specific tasks, like generating images or video from a text prompt, or identifying chemical compounds for new drug treatments, or using satellite imagery to spot forest fires before they spread, or even producing a podcast like this one. But what if there was a single AI program that could do just about anything a human could do, but even better? Well, my guest today, Shane Legg, has devoted his entire career to creating a general-purpose AI that can be effective for, well, just about anything.

4:30In 2010, Shane co-founded DeepMind with the mission of bringing this artificial general intelligence to life. Today, the company is part of Google's AI division, and Shane is on the front lines of a technological revolution. But his journey started back in the 1980s when he was just a kid in New Zealand with a new birthday present. So for my 10th birthday, my parents bought a small computer. It was called a VZ200. And it had an 8-bit microprocessor and had whole 8 kilobytes of memory. and had built in basic programming language. And as a 10-year-old boy, I'd heard about computers in movies or read about them in books or things like that.

5:18But I had a real computer of my own now. And this was prior to the internet and all these sorts of things. So there was really only one thing to do on the computer, and that was the program. It was a place where you could make things. You could write your own programs. You could start learning the language and so on. And you could just dream. You could invent things. You can invent little spaces with little pixels or little graphics that would chase each other around the screen or do all sorts of things like this. And it was just sort of a playground for my imagination. I could bring these little worlds to life.

5:51I could bring these little characters to life. And that just absolutely captivated me when I was 10, 11, and so on. You would go on to study mathematics and statistics in New Zealand. This is in the late 90s when you kind of start your career. The notion of artificial intelligence had been around. It had been written about in science fiction novels and it had already been depicted in films. Was that on your radar as a young man when you were starting out your career? When I first started, it wasn't. I started doing mathematics. I like the subject because it was very challenging. I did some computer science as well.

6:29I found computer science very easy because I'd spent my adolescence programming. But it was only at the end of my second year when in my spare time, I wrote some software that did calculus. And I showed that to some of the professors. And then they said, well, actually, we've got a summer job. And on a machine learning project, would you be interested in coming to work for us for the summer? What did machine learning mean in 1983? It's like the equivalent of like a, I don't know, like a mainframe computer to a laptop today, right? Yeah, it was early days then. It was a relatively new field then that had sort of split off from artificial intelligence not that many years earlier.

7:15And the difference in flavor was the emphasis on, instead of engineering like reasoning systems or something like that, the emphasis was on using data and learning a model of something. And so that's where the machine learning comes from. And presumably at that time, by the way, you had to feed the data into the machine because it's really – I mean there really wasn't – I mean the internet was around, but it wasn't populated with endless reams of information. Yeah, it wasn't – and also even if you had endless reams of information, the algorithms and the computers couldn't handle it. Yeah. So what you would typically do is you would have specific small data sets.

7:54I mean by small, you could like also print them out on a few sheets of paper kind of size. and you would have algorithms, particularly classification algorithms. So what they would do is you'd have a list of, I don't know, let's say people and you'd have different measurements about their blood results and then whether or not they have a disease. And you would learn a function, which is given this person's blood results, do they or do they not have this disease? And then in a new instance, you'd get the blood results and you wouldn't know if they have the disease, but you'd learn to predict it, to classify that individual as say having that disease or not.

8:25So that's a classification problem. So that was your first contact with what sort of a kind of AI, machine learning. Yeah. But in early 2000, right, most people who were going into the field of computing were really focused on the Internet. But you went right into artificial intelligence at a time when it was like really nascent. It wasn't really much of an industry yet. Yeah, there wasn't much of an industry. It was certainly an active area in academia. And so, yeah, after doing my undergraduate and then master's degree, I ended up working to do things like document classification for newspapers.

9:05Like determining this document here that uses the word bank, is it really about rivers or is it about finance, right? Yeah. So what you're talking about is basically creating a specific task for a machine that can get, you know, more or less can get a bit more sophisticated doing that single task. And this was sort of a kind of a crude – compared to what these machines can do today, relatively crude. But this is what you pursued. I mean you pursued a PhD in artificial intelligence already starting in 2003 and really began to think about machine superintelligence. This was the subject of your thesis, machine superintelligence.

9:52was the, I mean, when you talk about machine superintelligence, at that time, were you already using the term artificial general intelligence? Yeah. So maybe it's worth going back a little bit to about 1999. So I ended up working for a company called Intelligenesis, and they were doing various things, including this sort of document classification and so on. But one of the things they were interested in is building artificial intelligence systems that were very general and capable, building a thinking machine or something, if you like. And they didn't, I'd say they didn't really get very far in that, but it did get me thinking about that subject.

10:30And then when that company collapsed, I spent a lot more time thinking about the subject. And I actually became convinced, particularly after reading The Age of Spiritual Machines by Ray Kurzweil, that artificial, very powerful and general AI systems were going to come, but some a few decades into the future. So I actually came up with an estimate then that it was about a 50 % chance of what I call AGI now by about 2028. And that's also about the time that I proposed the term artificial general intelligence. And to be clear, the way I define artificial general intelligence is it's an artificial agent that can do all the kinds of cognitive tasks that people can typically do and possibly more.

11:14And I like that definition because I think it's very intuitive. We have a good understanding of the sorts of cognitive things that people can typically do. and we can look around at existing systems and see that they can do some of the sorts of cognitive tasks people can do, but they can't do others. And so we have some sort of intuitive sense of the kind of breadth and capability that you would need to be classified as an artificial general intelligence. Okay, so we're about four years away from seeing whether your prediction is right. We're going to come back to that later in this conversation, but I want to talk about what you would go on to do, because while you were in graduate school, you met Demis Hassab, who's still at Google today, and Mustafa Suleiman in the news now because he was just brought on by Microsoft to head up their AI program.

12:00And together you founded a company called DeepMind in the UK. Yeah. What was the goal? Presumably it was to develop an AGI? Yes. Our business plan from September 2010 had our logo on the front and DeepMind, and it had one sentence, which is build the world's first artificial general intelligence. And eventually you would develop a program, you'd develop a system that could defeat the greatest Go players in the world. And for people who don't know this game, and I don't know it very well, this is a much more complex game than chess. This was a huge challenge to try and figure out how to actually build that.

12:43Can you explain why you guys were focused on Go? Yeah. To be clear, we did work on many projects. We had a broad range of things. But AlphaGo ended up being, of course, one of the big famous ones. Yeah, it was a longstanding problem in artificial intelligence. In a game like chess, the number of possible moves is not so great that you can actually use brute force computation, more or less, to actually search through all the space of possibilities of which moves and which counter moves and so on and find very good combinations. In the game of Go, that's much more difficult because the board is much, much larger and there's a much larger number of possibilities at each point.

13:26So when you start looking through the trees of different possible moves, it exponentially grows at a much, much faster rate. And so this basically tripped up all the more brute force approaches to search and planning that would try to play this game well. So we had to come up with something new. And what we did is we actually blended together search techniques, something called MCTS, with deep learning. And the deep learning would actually learn two things. It would learn which moves were likely to be good moves just by looking at the patterns on the board using a deep neural network. And it would also learn, given a certain state of the game, who is more likely to win or lose.

14:14And then what we would do is we'd get our AI system to play against itself. And as it would win or lose games, it would take that signal, and then it would apply these learning algorithms to improve these deep learning networks. So they'd become better and better at anticipating which are the likely good moves and at any given point in time, which of the two players was most likely to win the game. Now, if you combine that with a search, then you start having a very, very powerful algorithm that both has this sort of classical search, but this sort of slightly more intuitive, deep learning aspect where it's kind of picking up subtle patterns in the game and trying to figure out, you know, which way it's going based on sort of these more subtle structures and so on.

14:57And it was that beautiful combination of the two that led through to the breakthrough in performance. We're going to take a quick break, but when we come back, how DeepMind went from mastering Go to working on some of the biggest challenges of our time. Stay with us. I'm Guy Raz, and you're listening to How I Built This Lab.

15:25Airbnb original experience that I'm hosting. And I'm so excited to have a chance to connect with listeners in a completely new way. It's called the Reinvention Lab, and it's designed to help you think about how to unlock your next big move in your career or even in your life. I'll help you discover your own story in ways I do on this show with my guests and attendees will get a chance to take a deeper dive with me on so many lessons I've learned from this show. lessons that have transformed how I work and think about the future. All proceeds from the event will go to support the Ronald McDonald houses that helps families stay near their children who are being treated at nearby hospitals.

16:06It's going to be really fun, and I can't wait to meet you. So come join me in San Francisco and take your idea to the next level. To grab your spot, visit airbnb.com slash guy. It's going to be great. Shopping for eggs should be simple. A happy hen makes a happy egg. And that's why eggs from Happy Egg are so delicious. Happy Egg partners with family farms across the Midwest to raise happy hens outdoors. The proof is inside the shell. A tasty orange yolk. It's the difference you can see and taste. I just made an incredible omelet with eggs from Happy Egg. It was delicious and so fresh. Once you crack open a happy egg, you can see and taste the difference right away.

16:57It's obvious. Visit www.happyegg.com slash built to find happy egg near you. Ever had one of those afternoons where your brain just quits on you? You're sluggish, hangry, maybe even a little foggy. What if it's your glucose? See, glucose is an energy currency for your mind and body. When it's stable, you're on point. When it crashes, so can you. That's why Lingo is so interesting. Lingo is a glucose wearable designed to help you connect the dots between your glucose and what you eat, how you move, and how you feel. It shows your glucose data in real time. Instead of guessing, you see the impact of your choices.

17:40Maybe that healthy snack is actually sending your glucose on a roller coaster. Or that afternoon walk is the perfect stabilizer. It's about unlocking your consistent best all day long by truly understanding your body's unique responses. Get to know your glucose and learn about how to build healthy habits that work for you with Lingo, designed for you by Abbott. Through November 30th, use code GUY10 on HelloLingo.com to get 10 % off a Lingo plan purchase, one use per customer. This offer cannot be combined with other offers, U.S., Puerto Rico, and U.K. only. The lingo-glucose system is for users 18 years and older, not on insulin.

18:20It is not intended for diagnosis of diseases, including diabetes. Individual responses may vary.

18:34Welcome back to How I Built This Lab. I'm Guy Raz. So in 2014, only a few years after its founding, DeepMind was acquired by Google. This is, from a business perspective, an amazing story, but you had only around 75 employees, I think, at DeepMind. And Google acquired the company for between reportedly$5 and$650 million. And then you became, for about nine years, sort of an independent part of Google. And we'll talk about why that changed recently. But essentially working on different projects, including projects around drug development and working with trying to model proteins that could potentially offer life-saving cures.

19:22Yeah, yeah. So that was another big success, which is protein folding. So your body is largely built out of proteins. These are the molecules that are the building blocks of most of biology. and it's not very difficult to know what the atoms are that make up the molecule. It's a certain chain of atoms. But what happens is the different atoms in that chain attract or repel each other in different ways. And the result of that is that chain actually folds up into a three-dimensional shape. And that three-dimensional shape can be all sorts of things. It can be an axle. It can be a spiral. It can be sort of a sheet.

20:01It can be sort of a sphere that contains something. you can build all sorts of interesting structures out of these shapes, sort of like 3D Lego blocks, if you like. And these shapes are very important if you want to understand what that protein does. Now, the problem is that while it's easy to find out the molecules that make up the protein, it's very difficult to find out the shape. And the techniques that people used to use would often require several years of research to find out what the shape was and cost maybe $200 ,000. It's maybe something that somebody would do as a PhD thesis is find out the shape of one protein.

20:40And there are hundreds of millions of proteins. So there was this computational challenge that had been around for many decades, which is, is it possible to take the molecule, just the knowledge of the atoms, and compute what the shape is directly, rather than go through this laborious process of several years of experiments and all these sorts of things. And so that was called the protein folding problem. How does a protein fold into the three-dimensional shape, and can you predict this computationally? And so people have been trying to do that for a long time, and we thought using advanced machine learning, deep learning techniques, and so on, that we had a shot at solving this problem.

21:22And yeah, long story short, we spent a few years working on it, and we basically solved the problem, yeah. And just for some clarification around it, I mean, what does it mean in practical terms? I mean, is it ready to go in terms of enabling treatments and drug development? Yeah, it's not quite that direct. What we did is we folded all the proteins known to science and we released them all to the public for free. So you can find them all online on the internet and you can look at the shapes and so on. And we've had about 1.7 million researchers use that resource. It doesn't mean you suddenly know how to build a drug or something rather.

22:01but it does mean that you can now see what the different shapes of the proteins are. The proteins that you might have in the drugs, the proteins that might be, I don't know, part of your liver or some other part of your biology. And you can see how they maybe interact with each other and all sorts of things like this. So it's not like you suddenly can solve the problem. But before, when you're operating in the dark and you didn't even have any idea what any of these things looked like, let alone how they might connect together and other things like that, now you can see a whole lot of this information.

22:29And so that's incredibly useful when you want to go and then develop drugs and so on. So you might see, I don't know, there's a particular problem taking place and it's to do with a particular protein. You might be able to then go, okay, what proteins are going to connect into this other protein and act on it in certain ways? And you can then target specific things that look like they're going to be very interesting and so on. So it's not a solution to drug discovery and everything, but it's a great enabler, an accelerator of this kind of process. Can you explain what happened around 2017 in the field of AI research?

23:04My understanding is that large language models were essentially introduced. And I guess that was like a turning point in the acceleration of the development towards artificial general intelligence? I think it was. We at Google had invented an extremely powerful algorithm called Transformer. And we'd been experimenting with it for things like translation between different multiple language targets in and out and all sorts of things like that. And we'd found it to be very, very scalable. And then what happened was that another company, OpenAI, latched onto the idea that this is, in fact, extremely scalable, more scalable than anybody had appreciated.

23:47And they just basically scaled it up and scaled it up and scaled it up. And, you know, it just kept on scaling, basically. And so it was a bit of a surprise in just seeing how far these language models could go when you made them extremely big and you started feeding in a significant chunk of text from the internet. And so that came as a surprise to many people. They didn't think it would go quite so far. So last year in 2023, DeepMind, which had been essentially an independent part of Google for a decade, was merged with Google's AI division. And I think that was in response to what was happening with OpenAI and maybe even some other competitors in the space to really kind of ramp up what Google could do around artificial intelligence.

24:43did part of that feel like you were joining an arms race that that you know did it feel like okay all hands on deck we've got to compete against these other companies yeah so so what happened basically was that you know it was becoming clear that extremely scaled up models were going to become a really important thing in the future and you know for google it was important that we had the biggest and best models and it doesn't make sense to have you know two different groups, both developing big models. We needed to come together. We both had a lot of expertise in this sort of thing, come together and use all the resources in terms of all the people together, and then all the compute and everything to make the best models we possibly could.

25:30All right. Around that time, Shane, you signed a letter that was signed by many other people in in the AI industry, warning of extinction risk. This is one of the quotes says, mitigating the risk of extinction from AI should be a global priority. Everyone signed this letter. I mean, Sam Altman signed this letter, Jeffrey Hinton, the sort of the godfather of AI, the heads of AI at Microsoft and Anthropic, and everyone signed it. And again, this technology is going to happen, right? But part of me is like, okay, All these people who are creating this technology and warning that there's a risk of extinction from the technology they're creating are signing this letter to sort of say we should be figuring this out.

26:18But at the same time, like we're just going to keep moving forward and marching forward. And so to me, there's sort of a – again, I'm not criticizing or attacking you. I'm just trying to figure, understand your thinking around this. Because on the one hand, you're signing a letter that's sort of ringing the alarm bells. And then a couple hours later, you're going back to your desk and doing your job. Yeah. And doing my job is often working on safety, to be clear. Okay. I mean, why am I doing what I do? Well, one, I think that something deeply transformational is about to happen in the coming years.

26:52Yeah. The world that we live in is being shaped by human intelligence. The clothes I wear, the headphones I have on, the internet we're talking to each other, the words we're using, the concepts we're using. Even the atmosphere we're breathing at the moment has been affected by human intelligence and combustion engines that we've invented and all these sorts of things, right? So human intelligence is a profoundly powerful thing. It's affected the world very, very deeply in many, many ways. Now, what's about to happen is that machine intelligence is going to arrive. and it is going to be potentially very, very powerful.