[0:00:00] The best guest of our exploration life. Harvard University professor and bestseller writer. This is the first time a Nobel Prize winner has come out. This is the Conrad Hotel. To commemorate the 10th anniversary of Isedol and AlphaGo, the founder of DeepMind, who created AlphaGo, and AlphaFold, who applied AlphaGo, we made a program called AlphaFold,
[0:00:28] It's been 10 years since AlphaGo. Yeah, it's amazing to think in many ways, feels very short amount of time. So it feels like just yesterday I was here with the match with Lee Sedol. But on the other hand, so much has happened in AI since that match too.
[0:00:55] Have you been able to sit down with Isador yet? I am seeing him on Wednesday, but we've had a good chat on Zoom call and it was nice to reconnect. We get on very well. On my taxi ride here, I got on my Gemini and I said, I'm meeting your dad.
[0:01:13] What did it say? It said, that sounds like a very interesting meeting. Since I'm on AI, I don't actually have a dad in the traditional sense. And then it said, if you happen to be meeting one of them, or maybe somebody like Sundar Pichai, tell them I said hello, that I'm hard at work. Very good. And I said, I'm meeting Demis. And he says, meeting Demis Hassabis is about as close to meeting my dad as it gets. Well, that's great. That's great that it answers that. And it's great that it's working hard. So this is good to hear. Good answers.
[0:01:42] So you're in Korea, which is the traditional land of Go. But not every young Korean plays Go or understands what the game is like. So after AlphaGo, you went on to make a lot of algorithms that went on to solve some of the biggest problems in science. And everybody says that that fateful day when AlphaGo beat Isador was when it all started. For somebody who's not scientifically minded or somebody who's not a Go player, they might say...
[0:02:09] What's the big deal about Go? Why is beating this board game the watershed moment for AI to go on solving protein folding problems? What's the relationship? No, it's a great question. I mean, I think AlphaGo, the match with Lisa Doe was the watershed moment for the modern AI industry. You can think of modern AI really starting with that match. And why was it such a big deal? It's because Go...
[0:02:35] It's such a complex game, very beautiful game, very intuitive game. Unlike chess, which is more about calculation and was cracked in the 1990s, we knew that if you wanted to play at the world champion level with Go, it must mean you built a very interesting algorithm. So it was always the pinnacle of kind of AI theory.
[0:02:57] Can a system eventually be world champion level at Go? And obviously when we created AlphaGo, it learned how to play Go itself. That was the important thing about the system. Instead of being programmed with the solution, like all previous AI systems, like the chess computers, it learned to play Go by playing itself.
[0:03:15] And learn to understand the important elements of the game. So that was the critical part. And because it was a learning system, it's able to generalize to many other areas of the world. Real world problems, chatbots, but also science systems like AlphaFold. So the idea was to use Go as a training ground for these ideas. Many young Korean people don't play Go, but they definitely do play StarCraft. Yes. And a lot of Korean people.
[0:03:44] They know about AlphaGo, but they don't know about your next big algorithm, which is AlphaStar. And in Korea, you usually don't think the brain that you required for playing StarCraft will translate to brain that you require for solving protein folding problems. It's more like if you're playing too much StarCraft, your mom's going to slap you on the back and say, go study for your SATs, right? So what gave you that idea that after Go, StarCraft was a good testing ground for...
[0:04:13] What AI could potentially do in the future.
[0:04:16] Well, taking a step back, what we did from the beginning of DeepMind in 2010 was to use games, starting with very simple ones, Atari games, really simple like Space Invaders, this kind of classic game, and then getting increasingly harder to test our algorithms and see if they were ready to use in the real world. So after Go, the next thing was StarCraft, as many of your viewers will know, is one of the most complex real-time strategy games. So it has different challenges from Go. Go is the most complicated, perfect.
[0:04:45] information game you obviously you see everything on the board it's and it's incredibly complex.
[0:04:50] But StarCraft is partial information. You don't know what your opponent has. You have to build your units. So it has many other kind of like resource optimization challenges, which is also quite interesting for AI to see if they can do those kinds of tasks. So we wanted to build a fully general AI system, Artificial General Intelligence, AGI. That was the mission from the start of DeepMind. And in order to do that, you should be able to play any type of game.
[0:05:17] From perfect information game like Go to partial information like Poker or Starcraft, where you're orchestrating many, many different units and agents into some bigger strategy. And so these are kind of general principles that can be applied to many other areas. Of course, especially areas like science, which is my passion, is why in the first place I started DeepMiner working on AI because I wanted to use it as the ultimate tool for advancing science.
[0:05:46] So what you're known for is as the father of AlphaGo, but also as the man who solved the protein folding problems for which you received the Nobel Prize.
[0:06:14] But our viewers are more humanities arts and history-oriented viewers, and they might not understand what protein folding problem exactly is and why it is so difficult for human beings to do it. And what in the way that AI works would make it so easy to do it by AI while it's almost impossible to do it with the human mind? Yeah. Well, proteins are everything in your body relies on for to, you know, do all the functions in your body from your neurons firing to your muscles moving. It's all relies on proteins.
[0:06:43] And proteins are described by like a genetic sequence. But what the protein folding problem that we solved is, is you need to understand in the body, how does it become a 3D structure? How does it roll up into a 3D structure? And the 3D structure, the structure of the protein is what tells you what the function protein does in the body.
[0:07:03] So it's very important for understanding biology, but also very important for things like drug discovery and understanding disease. So it's the kind of key for unlocking many things in biology. And the problem until we did AlphaFold is that usually it takes the rule of thumb, one PhD student, one, their whole PhD to solve one protein if they do it experimentally with very complicated, expensive equipment, electron microscopes and many other techniques. So if you, in the history,
[0:07:33] History of the last 50 years of experimental biology, we've only folded about 100,000 proteins, okay? Because it takes so long. But with AlphaFold, we were able to fold all 200 million proteins known to science. So if you think about that, that would have taken five years each times 200 million. So one...
[0:07:53] billion PhD years to have folded all the proteins if you were to do it in the experimental way by hand. And we were able to do it in one year and then released it for free out into the world open source for everybody to use. So now 3 million researchers around the world, pretty much every biology and medical researcher makes use of AlphaFold now in their very important work. So it's a big impact on the world.
[0:08:19] I remember that I watched your documentary, The Thinking Game, and I remember the scene where you realized that you could fold every protein onto a pen and you're telling your teammate, oh, let's just fold all of them. Yes, exactly. Which would be unheard of because usually even with the previous protein systems.
[0:08:39] You would get the experimental scientists would send you the one protein they were interested in. And then maybe it takes one or two weeks to make it even on the computer systems. But AlphaFold was so accurate and also so fast, we could fold one protein in a few seconds. So we worked out we could do everything in one year. So if your lab level AI was to sit down with AlphaGo and if there were people.
[0:09:05] And they were having a conversation. What would be the gap in their intelligence? Would sort of Gemini 3.1 Pro look at AlphaGo and say, oh, it's a two-year-old or three-year-old or it's a retriever? Yes. Well, I think it's actually, it's more like a kind of prototype system is the way I look at it. So, you know, AlphaGo was the first.
[0:09:27] System of its kind. So it had deep learning to learn the game of Go and learn the principles and then the patterns in Go. And then it had reinforcement learning to kind of make the decisions, you know? So it was an agent based system, which is now all of the fashion today with agents, but it was a very early one and it had all the ingredients that we now use today. But of course, today's systems like Gemini, they're much, much more powerful. And the main thing is they're more general.
[0:09:57] So AlphaGo was built to play Go, and then we built a second system, AlphaZero, which could play any two-player game, but it's still just limited to games. So the question always was, could it generalize these techniques to language and the real world and understanding things in the real world? And that's what we've done and others have done with the new modern foundation models like Gemini. Wow.
[0:10:21] You've been working with some of the top scientists around the world. And a lot of parents that are raising kids right now, they want to raise somebody that can thrive in an AI environment. And I'm sure you've met a lot of traditional scientists, some of whom thrive when you give them an AI tool. They're like, this is what I've been waiting for. Now I can do great things. And you might have met some people who are more reticent. They're like, oh, maybe it's not ready yet. Or how can I trust this new tool? So if we want to raise somebody, raise the next generation who's going to thrive.
[0:10:49] In the age of AI, who's going to do things like fold proteins with it? What are some of the traits that we need to inculcate in order to have the next generation who can use this to really, really make the world a better place?
[0:11:01] Well, I think first of all, studying traditional STEM subjects, science and math subjects, is going to be important because even though these systems are very capable of doing those sorts of things, it's still better if you are the user, you understand the technical system, why it's able to do that. You'll be able to use these systems better. So the traditional studying of mathematics and science is still important.
[0:11:24] Also, along with that, I would recommend the young kids of today to also immerse themselves in these new tools because they're like kind of magical, incredible capabilities they have. And most people only are scratching the surface of how to use them. So to me, it reminds me when I was young, growing up in the 80s and 90s of the first computer wave, home computer wave. For me, it was Sinclair Spectrum and Amiga. And kids like me were...
[0:11:53] In their spare time outside of school, playing around with programming games and other things on these machines, most of the adults didn't understand what these computers were for. But I knew and my friends knew that this was going to become a very important thing in the world. And it was very interesting. And I have the feeling the same today with these AI systems. They can do a lot more than most people imagine. So I think... More than asking AI. Yeah, exactly. Exactly. So that's what we would do.
[0:12:22] I think the youth of today would be using it to build incredible things, games, websites, applications, helping with their science. So I think if the youth of today...
[0:12:33] I realized just the last few days that after AlphaGo, AlphaZero, AlphaFour, there are at least 60, 70 projects, alpha something, solving some of the biggest problems in the world.
[0:12:59] Keeps you up at night. Which two or three of those projects, you look at them and you say, this is big. This is as big as Alpha 4 or bigger. Well, we're lucky we get to work on so many interesting things because the tool we're building is so general. You can apply it, I think, to nearly every area of science. The number one thing I always wanted to apply to was improving human health.
[0:13:20] So of course, alpha fold is part of that. But knowing the structure of a protein is only one small part of a drug discovery process. You need to understand the chemistry and the side effects, these kinds of things. So we spun out a new company, Isomorphic Labs, last two, three years. We've been working very hard to build more alpha folds to connect with the protein structure. So we're making very good progress. So I would really like to cure cancer.
[0:13:46] Really, not metaphorically with these systems. And I think in the next five, 10 years, we'll be able to solve many, many diseases. So this is one exciting thing. Another area is material science, designing new types of materials that have really cool properties. So better batteries, superconductors, maybe better solar materials. Are we going to finally solve cold fusion? Maybe. If it's possible. I think with the help of AI, we'll work out the limits of the physics understanding we have of the world.
[0:14:15] Last question, yes. I've been using Gemini a lot in my ordinary life. I use it for things like learning languages and asking historical questions. And what I feel like is that the next generation is going to grow up with basically 100 employees in their pocket, a group of assistants that can help them with research or writing or whatever they want to do. So they need to be able to think like a CEO from day one. You're an actual CEO. What are some of the skills that you acquired as a CEO?
[0:14:43] You think it's necessary to survive in the AI world. And I'm glad you're finding Gemini useful. And, you know, I'm really happy to see how popular it is in Korea, especially we worked hard on things like the Korean language, multilingual capabilities. So it's useful for that.
[0:14:57] You're right. I think in the future, we're all going to have dozens of agents each that are doing different things, research analysts, you know, helping with brainstorming. So we need to work out how to break down a big problem into smaller problems and then delegate the smaller problems to the right kind of agent with the right sort of instructions. And then a lot of it will come down to kind of optimizing that coordination and also asking the right question, the valuable question.
[0:15:27] One of the hardest things in science and business is to articulate the right question to answer and the right North Star to go for. So these are the kind of skills actually, you know, that we now need to think about teaching in schools and universities, not just the technical skills, but more like almost like business, operational, organizational skills. I hope more people, more of our viewers after watching this interview will...
[0:15:50] Watch the thinking game to realize that AI gives you this enormous potential to really, really solve big problems for the humankind. Thank you very much for being with us today. It's the first time I'm sitting down with a Nobel laureate. I'm super nervous. Thank you so much. It was really fun. Thanks for having me. Thank you. Great. I watched your interview with Cleo Abram. Oh, yes. And then Lex Friedman. And I sort of got into that algorithmic loophole where now all I feel is about...
[0:16:18] Oh, it's about this? Because you've been watching too many? Yes, it's really insightful. Thank you very much. Oh, thank you.
[0:16:43] In other words, AI can solve some of the problems that the smartest scientists in the world have never solved in the lab. So what Demis said in another interview was, this is like the invention of fire.
[0:17:01] Or like the industrial revolution, it will be an opportunity to completely change the way humans live. For example, if you go to a farmer in the 16th century and say that obesity will be the greatest problem of mankind in the 20th century, it will be as if the farmer in the army did not understand. Now that we have a lot of knowledge, in the 20th century, the scientists were worried because they could not solve environmental problems. Why is that a problem? Wasn't this just something that could be solved?
[0:17:27] That's how big the evolution of evolution is. That's what happened in 2016. He told me that it was the country of AlphaGo and Lee Sedol. Personally, in a way, AlphaGo's... I'm using Gemini a lot in my daily life. I'm not a scientist. As Demis said earlier, 100% prediction of future climate. Developing new ingredients that kill human cancer cells.
[0:17:57] I don't think it's my job. I like reading and studying, so I realized that JAEMIN is a great help for studying. For example, I'm studying Arabic reading recently. I've tried Arabic a few times and I've given up. Because we usually give up when we learn European languages because of verb changes. Unlike English, most of the European words are 1st, 2nd, 3rd.
[0:18:26] The terms are all different, and it changes into the past tense, the past perfect tense, the future tense, the future perfect tense, and so on. And there are a lot of things that look similar. Oh, it's so confusing. And I give up. But that's enough for my head. So I did French and Italian. Then I tried Arabic. This language, known as the language with the most difficult verb structure in the world. I thought I could do Italian and French, but I thought I couldn't.
[0:18:53] I did it a few times, but I gave up because I thought I wouldn't be able to do it for the rest of my life. Recently, I uploaded the content that I gave up on it. So, what do you have to do in your 20s? I told you that you can't do it in your 40s, so I told you to study Arabic. But what do you have to study Arabic for? I don't think I can study Arabic without a teacher. In France, it was easy to get a tutor because there were a lot of Arabic-speaking people, but in Korea, it was not the case. So, it's a very complicated subject, but there are no teachers around.
[0:19:21] Let's use Artificial Intelligence. And I told Jaemin, You're going to read Koran five times a day from now on. Give me five Koran lines. All the words in one Koran line. And what kind of grammar the word is in. And if the word was written differently than the original meaning, I made a report on what kind of collection or what kind of text it was written differently. I ordered five pages. It's amazing. For example, Alif here. The one that's been deleted is...
[0:19:50] The spelling we're using right now is the spelling of the caliphate called Uthman. How is it fixed? Tell me all of this. So it wasn't confusing. I was going to try this a few times. I was going to give up on three lines, but now the longest one is the second chapter. Korah is very unique because the chapters are in the order of length. So the first chapter is just the opening, but the second chapter is the longest and the third chapter is shorter. But I finished the second chapter. There's a saying that if you finish the second chapter, you're done.
[0:20:16] He also gives me motivation. If I do chapter 2, the rest is easy. Chapter 2 is the most difficult. But chapter 3 is difficult. Oh, he's coaching. When I read a history book, what was he doing here in this era? Or, as I read in the past, a Portuguese adventurer was here. If a Japanese merchant went to the Philippines at that time, would they have met? If you put that kind of fun in it, of course, this is a fact check later.
[0:20:46] According to the record in the source you know, there is no possibility that you two met because this person arrived in a few months. It's also very useful to let you know this. Oh, and what else is there? When you make content for a long time, you read a book and after about three years, the fact keeps changing and you feel like you're doing it. It's really good to look for that. For example, I had a watch that I had a long time ago. This was a watch I bought from Swatch.
[0:21:10] I suddenly thought of that watch. I was going to make a watch content, and I was going to talk about my first watch, but I couldn't find the exact model name of that watch. But anyway, if you block the parameter, you search for the number of cases in it. So I came out of Swatch, and it was a watch that was sold between 1994 and 1999, and the line was metal, and I could see the inside, but the middle was made of rubber, and then there was no number plate in front of it, and the machine behind it was visible.
[0:21:39] So I googled one of them and found out that it was a watch called Swatch SAG-001 Irony Heart and Body. So in the next watch content, I can talk about my first watch. For example, it's a movie I saw a long time ago. I think it was in the Middle Ages, and a noble woman was kidnapped by thieves while riding a carriage. If you ask me, what is the movie that goes to the 9th episode with a strange machine? I give you a few movies with these lines. Then there's one of them.
[0:22:08] So if you ask me, is it true that I read this book? If you ask me, is it this person? If you ask me, I can find out that I remember it wrong. If you just believe the answer, there's still a little bit of a mistake, so check it out. It even says, I can make mistakes. There's nothing to believe in my head. I read the book a long time ago, and I remember the content and the characters. But I don't remember who the writer is.
[0:22:35] JAEMIN is really good at that. Also, what's good is that JAEMIN has great strengths as a philosophical and strategic partner. Sometimes, when I make content, there's a slump. One of the reasons for the slump is that if you do a job called YouTube, you have no choice but to follow the algorithm. We originally started as a channel that only reviewed historical movies, but as the trend changes, other channels say, it's good to do this, so let's try it. If you get an amazing interview request like today,
[0:23:04] I was doing a Nobel Prize-winning interview. I said, I have to do it. And then after three years, I was confused about my identity. What channel are we on? What is the value we give to our subscribers? There are times when I'm confused. I made this content for Jaemin. What is the consistent commonality of the message I want to deliver? If you talk about it like this, it's very accurate. If you look at the world from the pyramid,
[0:23:32] I'm telling you that you can melt your heart. I'm telling you that you can melt your heart. I'm telling you that you can melt your heart. I'm telling you that you can melt your heart. I'm telling you that you can melt your heart. I'm telling you that you can melt your heart. I'm telling you that you can melt your heart.
[0:24:02] There's a summary. I ask them to make it into a table, so I put it on the table. The three things I use the most to make it fun. One is that when you study a language, it's too complicated to study it in a book or a book that doesn't come out in a book. I'm going to read the actual sentence or word and listen to the interpretation of it. I think this is very powerful. And the second one is that I'm going to put a label on what I remember. And the third one is that it's surprisingly deep and philosophical.
[0:24:31] They organized it. But I think this was very surprising. When we think about the relationship between humans and computers, humans are good at deep philosophical and humanities accidents, and computers will play the role of executing the humanities accidents I did. In the pre-AI era, we thought. When I make a robot, this robot does what I asked it to do repeatedly. That's the standard of humans and machines. Surprisingly, he's good at deep thinking. Rather, I think he's good at thinking about life.
[0:24:59] I thought it was necessary to redefine the relationship between humans and machines. When I was young, I read Lee Won-bok's Do you remember what country your neighbor is from? I thought I should learn Latin. I want to learn Latin, but what my mother said at first was, Where do you get a Latin teacher? But now, if I were a second-grade student in Yeouido Middle School, it's very simple. Like Koran, bring Kaisar's Galia War or Bigelis's Ainaid. The first line, Arma virum cucano.
[0:25:28] What does it mean? Where did all these words come from? What kind of form is it in grammar? Teach me what the word order is. Then you just have to do that one line today. Of course, I can't be a priest who preaches in Latin fluently. Basically, when I meet a teacher, I think I can get to the point where I can start from the middle school. I think it's really good for people who don't want to study at school.
[0:25:54] And unfortunately, Korea doesn't have any Korean books about that field. Since the language users are small countries, for example, if I say I want to find a book about the history of Laos food, it's actually very difficult to find a book that's well organized about the history of Laos food in Korean. It's all in English. Because all the scholars around the world use English. But if you explain the Laos food culture, they'll scratch the data from all over the world and explain it in Korean.
[0:26:23] There's no such consideration. So I think it would be great for those who have a lot of curiosity. But for those who don't have any curiosity, it could be a bit like that. They say they use it all, but the first question is that. What should I help you with? But if I don't have anything I want to do, I don't have anything to add to the prompt. People like that might say, where do you use this? Wow, if I had this in middle school, I think I'd be able to speak about 20 languages now. Today, I talked about Demesne, and I personally talked about what I felt using Gemini.
[0:26:52] In the end, I think the size of the dream is proportional to how much work you can do when such a new tool appears. Demis is a person who dreams of getting rid of cancer. But in the end, Cho Seung-hyun's dream size is sitting in the pelvis. So what I think is that we've been telling us to dream big since a long time ago. But I don't even know if it's time to dream really big.
[0:27:20] People who thought they couldn't make their dreams come true with the help of AI, I think people will think that AI is the new generation. Since I'm a dad, how can I make a big dream come true? Dad told me to look into it, but I'm doing this, should I make a big dream come true? Don't you have any questions related to parenting? I asked about that once. I was so stressed out when my child cried, and I asked if I was a bad dad. There are a lot of people who ask this question unexpectedly. That's right. You're all doing that.
[0:27:49] We are a channel that deals with history, history, and human relationships, but when tools appear, human relationships change. I think there will be a huge change in consciousness in the AI era, as the human relationship between the times when we lived in apartments and the times when we lived in supermarkets and our consciousness does not match.
[0:28:09] The father of AlphaGo and Gemini's father, David Hassabis, who leads it ahead. I interviewed him and made a content about how AI changed my daily life. If you're using AI in a unique way in your life, please leave a comment. I think we'll learn a lot from each other. That's the end of Cho Seung-hyun's exploration life today. I'll be back with a more interesting history and culture story next time. Thank you. Goodbye.