WEBVTT 00:00:00.000 --> 00:00:02.120 Welcome back to The Deep Dive. This is where 00:00:02.120 --> 00:00:04.719 we take your stack of sources, the academic journals, 00:00:04.940 --> 00:00:07.019 the historical records, the career milestones, 00:00:07.360 --> 00:00:09.820 and turn them into a clear, compelling narrative, 00:00:10.060 --> 00:00:13.259 giving you the deep context fast. Today, we are 00:00:13.259 --> 00:00:16.120 unpacking a figure whose career trajectory is, 00:00:16.199 --> 00:00:19.320 it's less a series of steps and more a kind of 00:00:19.320 --> 00:00:22.679 grand, unified theory of problem solving. We 00:00:22.679 --> 00:00:25.219 are diving into a life and work of Sir Demis 00:00:25.219 --> 00:00:28.039 Hassabis. If you know the name, it's likely linked 00:00:28.039 --> 00:00:30.239 to DeepMind, the artificial intelligence company 00:00:30.239 --> 00:00:32.320 he co -founded, which has, I mean, it's fundamentally 00:00:32.320 --> 00:00:35.079 redefined the boundaries of computation. Absolutely. 00:00:35.100 --> 00:00:38.579 But to view him solely as an AI mogul, you'd 00:00:38.579 --> 00:00:41.640 be missing the astonishing sequence of foundational 00:00:41.640 --> 00:00:44.179 disciplines he mastered first. We're talking 00:00:44.179 --> 00:00:47.020 about a child chess prodigy, a pioneering video 00:00:47.020 --> 00:00:49.340 game designer, and then crucially, a cognitive 00:00:49.340 --> 00:00:51.700 neuroscientist. And when you lay out the sources, 00:00:51.979 --> 00:00:55.140 you know, his FIE chess rating history, his records 00:00:55.140 --> 00:00:57.679 in the world. board games championship, papers 00:00:57.679 --> 00:01:00.240 in nature and science, and then, of course, the 00:01:00.240 --> 00:01:02.219 record of his recent monumental achievement, 00:01:02.500 --> 00:01:04.819 you see a very, very deliberate path. Right. 00:01:04.879 --> 00:01:06.819 It's not random. There's a thread. There's a 00:01:06.819 --> 00:01:08.840 definite thread. And that central achievement, 00:01:09.040 --> 00:01:10.859 the capstone that connects all these seemingly 00:01:10.859 --> 00:01:14.359 disparate worlds, has to be his joint award of 00:01:14.359 --> 00:01:17.700 the 2024 Nobel Prize in chemistry. Yeah, in chemistry 00:01:17.700 --> 00:01:20.480 for an AI company. He won it alongside John M. 00:01:20.519 --> 00:01:23.189 Jumper. For their work on AI research for protein 00:01:23.189 --> 00:01:25.849 structure prediction, the system we now all know 00:01:25.849 --> 00:01:28.989 as AlphaFold. Exactly. And that really is the 00:01:28.989 --> 00:01:31.090 mission of this deep dive for you, the listener. 00:01:31.530 --> 00:01:33.390 We're going to trace that evolution. We want 00:01:33.390 --> 00:01:35.590 to show how Hassabis' background in strategy, 00:01:35.829 --> 00:01:39.530 simulation, and the study of human memory didn't 00:01:39.530 --> 00:01:41.870 just accidentally intersect with molecular biology. 00:01:41.909 --> 00:01:44.489 No, it was the opposite. Right. How those fields 00:01:44.489 --> 00:01:46.849 acted as essential intellectual building blocks, 00:01:46.989 --> 00:01:49.590 really, that enabled the solution to one of biology's 00:01:49.590 --> 00:01:52.469 grandest challenges. We're tracing the evolution 00:01:52.469 --> 00:01:55.430 of a generalist genius who decided to specialize 00:01:55.430 --> 00:01:58.670 in, well, the impossible. Okay, let's unpack 00:01:58.670 --> 00:02:00.310 this journey. Let's start right at the beginning. 00:02:01.000 --> 00:02:04.000 Section one, the foundations of a prodigy. This 00:02:04.000 --> 00:02:06.379 covers his childhood through his early 20s. So 00:02:06.379 --> 00:02:09.699 he was born in London in 1976. He has a mixed 00:02:09.699 --> 00:02:12.250 background. His father is Greek Cypriot. His 00:02:12.250 --> 00:02:15.689 mother is Chinese Singaporean. But what the sources 00:02:15.689 --> 00:02:19.050 just immediately scream at you is the level of 00:02:19.050 --> 00:02:21.729 mental acuity he displayed from an almost unbelievably 00:02:21.729 --> 00:02:23.530 young age. We're talking about chess, right? 00:02:23.590 --> 00:02:26.129 Chess mastery. He was a child prodigy, started 00:02:26.129 --> 00:02:28.650 playing at age four. And I have to stress, this 00:02:28.650 --> 00:02:31.729 wasn't just, you know, a fun hobby. By age 13, 00:02:31.889 --> 00:02:34.669 he had reached master standard. Thirteen. Thirteen. 00:02:34.930 --> 00:02:38.789 He hit an ELO rating of 2300 in January 1990. 00:02:39.469 --> 00:02:42.169 Okay, so for listeners who might not be deep 00:02:42.169 --> 00:02:43.689 into competitive chess, what does that number 00:02:43.689 --> 00:02:46.370 actually mean? An ELO of 2 ,300 at that age. 00:02:46.729 --> 00:02:48.930 That's got to be rare territory. It's exceptionally 00:02:48.930 --> 00:02:51.050 rare. I mean, it's off the charts. A master rating 00:02:51.050 --> 00:02:52.909 means you're performing in the top fraction of 00:02:52.909 --> 00:02:55.770 1 % of all competitive chess players globally. 00:02:55.969 --> 00:02:59.490 Wow. To be there at 13, it means your brain is 00:02:59.490 --> 00:03:03.009 already executing multi -step planning and deep 00:03:03.009 --> 00:03:05.349 strategic look ahead at a professional adult 00:03:05.349 --> 00:03:08.340 level. He captained many England junior chess 00:03:08.340 --> 00:03:10.439 teams, representing Cambridge in the big Oxford 00:03:10.439 --> 00:03:13.219 -Cambridge matches. And chess, at its core, it's 00:03:13.219 --> 00:03:15.580 a perfect information game, right? It's solved 00:03:15.580 --> 00:03:18.500 by something called a mini -max search. That's 00:03:18.500 --> 00:03:20.680 it. It's the process of recursively minimizing 00:03:20.680 --> 00:03:23.000 your opponent's maximum possible gain. You're 00:03:23.000 --> 00:03:26.110 playing out the future in your head. This becomes 00:03:26.110 --> 00:03:27.969 foundational for how he would later approach 00:03:27.969 --> 00:03:30.710 AI. He was literally training the ultimate search 00:03:30.710 --> 00:03:33.050 algorithm in his own mind before he was even 00:03:33.050 --> 00:03:35.530 a teenager. So he's building this mental engine 00:03:35.530 --> 00:03:38.900 for strategy. But his strategic output wasn't 00:03:38.900 --> 00:03:41.599 isolated to just the 64 squares of a chessboard. 00:03:41.900 --> 00:03:44.800 The sources show a competitive streak that spanned, 00:03:44.819 --> 00:03:46.780 well, a lot of disciplines. This is where the 00:03:46.780 --> 00:03:48.919 polymathic nature at least starts to show. He 00:03:48.919 --> 00:03:51.199 wasn't satisfied just being a chess master. He 00:03:51.199 --> 00:03:53.939 was a master of competitive multi -games. He 00:03:53.939 --> 00:03:56.340 went on to dominate the world of all -round board 00:03:56.340 --> 00:03:59.020 games. The pentamind. The pentamind. He won the 00:03:59.020 --> 00:04:01.819 World Board Games Championship five times. 1998, 00:04:02.120 --> 00:04:06.379 1999, 2000, 2001, and then again in 2003. also 00:04:06.379 --> 00:04:08.819 the world decametathlon champion in 03 and 04. 00:04:09.180 --> 00:04:11.280 Let's just pause on that for a second. Winning 00:04:11.280 --> 00:04:13.879 one world championship is incredible. Winning 00:04:13.879 --> 00:04:16.660 five in a competition that tests multiple games, 00:04:16.779 --> 00:04:19.939 that suggests a really unique cognitive flexibility. 00:04:20.540 --> 00:04:23.910 What does the pentamind actually test? Well, 00:04:23.930 --> 00:04:25.829 that's the thing. It's a competition that demands 00:04:25.829 --> 00:04:28.230 peak performance across five different intellectual 00:04:28.230 --> 00:04:31.649 disciplines. And these can be anything from abstract 00:04:31.649 --> 00:04:34.930 strategy, positional games, real -time tactical 00:04:34.930 --> 00:04:38.560 games, memory challenges. You name it. So the 00:04:38.560 --> 00:04:41.180 skill isn't just excellence in one game like 00:04:41.180 --> 00:04:44.259 chess. Not at all. It's the rapid transfer of 00:04:44.259 --> 00:04:47.819 complex strategic frameworks across radically 00:04:47.819 --> 00:04:50.420 different rule sets. It requires a high degree 00:04:50.420 --> 00:04:52.519 of what we now call general intelligence. Which 00:04:52.519 --> 00:04:54.819 is precisely the thing he later sought to engineer 00:04:54.819 --> 00:04:56.720 in a machine. Exactly. And if you look at the 00:04:56.720 --> 00:04:59.259 specific games he mastered, they map almost perfectly 00:04:59.259 --> 00:05:01.680 to the different modalities an AGI would need 00:05:01.680 --> 00:05:03.600 to master. Right. Because the sources don't just 00:05:03.600 --> 00:05:05.339 list chess. They list things like diplomacy. 00:05:05.399 --> 00:05:08.050 Yes. He was the world team champion in diplomacy 00:05:08.050 --> 00:05:11.629 in 2004. Yeah. And competitor to poker. He cashed 00:05:11.629 --> 00:05:13.730 six times at the World Series of Poker. Okay, 00:05:13.769 --> 00:05:16.310 I have to ask about diplomacy. Because diplomacy 00:05:16.310 --> 00:05:19.350 isn't a search problem like chess. It's a human 00:05:19.350 --> 00:05:22.389 problem. It's about negotiation, backstabbing, 00:05:22.389 --> 00:05:25.910 alliances. Predicting the intent of seven other 00:05:25.910 --> 00:05:29.490 intelligent and often deceitful human players. 00:05:30.009 --> 00:05:33.339 Yes. So how does that connect to AGI? Well, it's 00:05:33.339 --> 00:05:35.040 a crucial distinction. In a game like chess, 00:05:35.279 --> 00:05:37.180 the environment is static, the information is 00:05:37.180 --> 00:05:39.480 perfect, you see the whole board. Diplomacy is 00:05:39.480 --> 00:05:42.319 an exercise in multi -agent cooperation and, 00:05:42.480 --> 00:05:45.279 more importantly, theory of mind. Modeling what 00:05:45.279 --> 00:05:47.319 the other guy is thinking. Exactly. A successful 00:05:47.319 --> 00:05:50.560 player, human or AI, has to model the beliefs, 00:05:50.680 --> 00:05:52.819 the desires, the intentions of the other players. 00:05:52.959 --> 00:05:55.540 That modeling of human belief systems is vital 00:05:55.540 --> 00:05:58.079 for developing complex, ethical and interactive 00:05:58.079 --> 00:06:01.220 AI systems today. He was mastering theory of 00:06:01.220 --> 00:06:03.620 mind in a competitive setting two decades ago. 00:06:03.860 --> 00:06:06.259 And poker then adds another layer on top of that. 00:06:06.639 --> 00:06:08.879 Probabilistic reasoning under conditions of incomplete 00:06:08.879 --> 00:06:11.439 information. You don't know their hand. Correct. 00:06:11.920 --> 00:06:14.720 Poker is a test of probability, risk assessment, 00:06:14.819 --> 00:06:19.079 and behavioral bluffing. So you see, these games, 00:06:19.180 --> 00:06:21.759 they weren't just trophies on a shelf for him. 00:06:21.819 --> 00:06:24.560 They were early successful experiments in applied 00:06:24.560 --> 00:06:28.079 human intelligence in complex, real -world mimicking 00:06:28.079 --> 00:06:30.500 environments. Environments full of noise, hidden 00:06:30.500 --> 00:06:33.639 variables, unreliable agents. All of it. And 00:06:33.639 --> 00:06:35.600 while this is all happening, his technological 00:06:35.600 --> 00:06:38.420 journey is starting in parallel, rooted in that 00:06:38.420 --> 00:06:40.750 same pattern of self -directed learning. That 00:06:40.750 --> 00:06:44.069 spark ignited in 1984. He used his chess winnings 00:06:44.069 --> 00:06:47.209 to buy his first computer, the iconic ZX Spectrum 00:06:47.209 --> 00:06:49.569 48K. And he didn't take classes. He just taught 00:06:49.569 --> 00:06:51.329 himself programming from books. Just from books. 00:06:51.370 --> 00:06:53.790 And immediately, what does he do? He applies 00:06:53.790 --> 00:06:55.470 that programming knowledge right back to his 00:06:55.470 --> 00:06:58.209 passion strategy. He wrote his very first AI 00:06:58.209 --> 00:07:00.629 program, an algorithm designed to play the board 00:07:00.629 --> 00:07:03.550 game Reversi on a Commodore Amiga. This establishes 00:07:03.550 --> 00:07:06.389 a pattern, a relentless focus. Learn the tool, 00:07:06.490 --> 00:07:08.389 which is programming, then apply it immediately 00:07:08.389 --> 00:07:10.949 to the goal, which is creating systems that exhibit 00:07:10.949 --> 00:07:13.930 intelligent behavior. That ability to rapidly 00:07:13.930 --> 00:07:16.689 acquire technical skills and deploy them against 00:07:16.689 --> 00:07:19.990 ambitious intellectual targets, it defines his 00:07:19.990 --> 00:07:22.569 entire career. And this rapid intellectual trajectory 00:07:22.569 --> 00:07:25.529 meant his formal education also moved at, well, 00:07:25.629 --> 00:07:28.290 light speed. He completed his A -level exams 00:07:28.290 --> 00:07:32.139 a year early, at just 16. And by 1997, he had 00:07:32.139 --> 00:07:34.199 graduated from Queens College, University of 00:07:34.199 --> 00:07:36.639 Cambridge, with a double first in computer science 00:07:36.639 --> 00:07:39.750 tripos. So by the time he's 21, he is a world 00:07:39.750 --> 00:07:42.589 -class strategic thinker, a proven master of 00:07:42.589 --> 00:07:44.889 multiple learning environments, and a highly 00:07:44.889 --> 00:07:48.149 skilled foundational computer scientist. He understands 00:07:48.149 --> 00:07:50.370 the theory of complex systems, the mathematics 00:07:50.370 --> 00:07:52.490 of probability, and the engineering required 00:07:52.490 --> 00:07:55.069 to simulate those systems. That combination is 00:07:55.069 --> 00:07:57.250 what propelled him into his next phase, which 00:07:57.250 --> 00:07:59.629 was, maybe surprisingly, the world of commercial 00:07:59.629 --> 00:08:02.089 video games. And that leads us to Section 2, 00:08:02.250 --> 00:08:05.050 the next crucial chapter. The Architect of Virtual 00:08:05.050 --> 00:08:09.850 Worlds. This covers roughly 1994 to 2005. And 00:08:09.850 --> 00:08:11.790 it's fascinating because he wasn't just in the 00:08:11.790 --> 00:08:14.829 video game industry. He was a revolutionary force 00:08:14.829 --> 00:08:17.430 in simulation design. The transition was almost 00:08:17.430 --> 00:08:22.110 accidental, or at least mandated by his own precociousness. 00:08:22.550 --> 00:08:24.769 Cambridge University actually requested he take 00:08:24.769 --> 00:08:26.930 a gap year because he was so young when he finished 00:08:26.930 --> 00:08:28.930 his A -levels. And that gap year wasn't spent 00:08:28.930 --> 00:08:31.189 backpacking. Not at all. It was spent building. 00:08:31.720 --> 00:08:34.600 He landed at Bullfrog Productions, an iconic 00:08:34.600 --> 00:08:37.500 game studio of the era, after winning an Amiga 00:08:37.500 --> 00:08:39.740 Power Win a Job at Bullfrog competition. The 00:08:39.740 --> 00:08:42.379 competition. Yes. And this is where he got his 00:08:42.379 --> 00:08:45.500 first chance to build a truly large -scale complex 00:08:45.500 --> 00:08:49.059 simulation. At just 17 years old, he co -designed 00:08:49.059 --> 00:08:51.559 and was the lead programmer on the 1994 simulation 00:08:51.559 --> 00:08:54.120 game Theme Park. Alongside the industry legend 00:08:54.120 --> 00:08:55.960 Peter Molyneux, yeah. The commercial success 00:08:55.960 --> 00:08:58.820 was immediate and massive. Theme Park wasn't 00:08:58.820 --> 00:09:01.179 just a hit. The sources note it sold millions 00:09:01.179 --> 00:09:03.240 of copies and essentially created the simulation 00:09:03.240 --> 00:09:06.340 sandbox genre that we all know today. Absolutely. 00:09:06.559 --> 00:09:08.860 The kind of game where players manage complex 00:09:08.860 --> 00:09:11.000 systems with all these interconnected variables. 00:09:11.580 --> 00:09:14.820 Finances, staff happiness, customer needs, park 00:09:14.820 --> 00:09:18.179 layout, even traffic flow. It was a complex systems 00:09:18.179 --> 00:09:20.820 model dressed up as entertainment. And here's 00:09:20.820 --> 00:09:22.980 a fantastic financial anecdote from the sources. 00:09:23.580 --> 00:09:26.100 He earned enough from his lead programming role 00:09:26.100 --> 00:09:28.679 on Theme Park during that single gap year to 00:09:28.679 --> 00:09:31.139 entirely fund his subsequent university education 00:09:31.139 --> 00:09:33.379 at Cambridge. That's incredible. It's a complete 00:09:33.379 --> 00:09:35.500 professional validation before he even started 00:09:35.500 --> 00:09:38.360 his degree. Right. So after completing his double 00:09:38.360 --> 00:09:41.639 first at Cambridge in 1997, He rejoins Peter 00:09:41.639 --> 00:09:43.919 Molyneux at his new company, Lionhead Studios. 00:09:44.320 --> 00:09:46.379 And there he continues his focus on sophisticated 00:09:46.379 --> 00:09:49.500 AI. He worked as a lead AI programmer on the 00:09:49.500 --> 00:09:52.659 2001 God game, Black and White. The complexity 00:09:52.659 --> 00:09:55.960 here really ramps up. In a God game, the AI has 00:09:55.960 --> 00:09:58.360 to manage not just physics or resources, but 00:09:58.360 --> 00:10:01.220 belief systems, evolution, moral choices for 00:10:01.220 --> 00:10:03.620 the players, creatures, and worshippers. It's 00:10:03.620 --> 00:10:06.620 an early form of adaptive, evolving AI inside 00:10:06.620 --> 00:10:08.830 a simulation. But the real expression of his 00:10:08.830 --> 00:10:11.009 ambition to build a massive simulation engine 00:10:11.009 --> 00:10:13.769 came when he decided to run his own lab. His 00:10:13.769 --> 00:10:16.970 own studio, yeah. He left Lionhead in 1998 to 00:10:16.970 --> 00:10:19.789 found Elixir Studios, an independent developer 00:10:19.789 --> 00:10:22.309 in London. He was the executive designer on their 00:10:22.309 --> 00:10:25.149 major titles. The sources highlight two big releases. 00:10:25.669 --> 00:10:29.070 Evil Genius, which was a successful parody game 00:10:29.070 --> 00:10:31.389 about building a villain's lair. It did pretty 00:10:31.389 --> 00:10:33.269 well, right? It did well, yeah. Metacritic score 00:10:33.269 --> 00:10:36.389 of 75 out of 100. It was well received. But the 00:10:36.389 --> 00:10:39.279 other project. That was the ambitious one. Republic. 00:10:39.639 --> 00:10:42.799 The revolution. And Republic is really the key 00:10:42.799 --> 00:10:45.620 to understanding his subsequent pivot. Why was 00:10:45.620 --> 00:10:47.600 that game so much more than just entertainment? 00:10:47.960 --> 00:10:50.440 Because Republic was basically Hassabis trying 00:10:50.440 --> 00:10:54.110 to build AGI in commercial gaming clothing. It 00:10:54.110 --> 00:10:56.269 was described as a hugely ambitious political 00:10:56.269 --> 00:10:59.269 simulation that attempted to model the political 00:10:59.269 --> 00:11:01.950 workings, the social unrest, the economic levers 00:11:01.950 --> 00:11:04.470 of an entire fictional country. That sounds less 00:11:04.470 --> 00:11:06.710 like a game and more like a computational sociology 00:11:06.710 --> 00:11:09.370 project. It was. I mean, that sounds incredibly 00:11:09.370 --> 00:11:11.169 complex, especially for the hardware limitations 00:11:11.169 --> 00:11:14.789 of the late 90s and early 2000s. And that's precisely 00:11:14.789 --> 00:11:16.850 the learning point we have to emphasize here. 00:11:17.070 --> 00:11:20.919 His ambition just completely outstripped. the 00:11:20.919 --> 00:11:22.919 reality of the technology, and the constraints 00:11:22.919 --> 00:11:25.120 of the commercial market. The scope was too big. 00:11:25.379 --> 00:11:28.919 The scope was so vast, simulating an entire functioning 00:11:28.919 --> 00:11:32.080 country's political machine, that it led to massive 00:11:32.080 --> 00:11:34.879 development delays. The final version had to 00:11:34.879 --> 00:11:37.679 be significantly scaled back from its original 00:11:37.679 --> 00:11:41.000 theoretical vision. And the result was a pretty 00:11:41.000 --> 00:11:44.179 lukewarm critical reception. Lukewarm is putting 00:11:44.179 --> 00:11:46.879 it kindly. It scored 62 out of 100 on Metacritic. 00:11:47.240 --> 00:11:49.320 The reviews basically said it was an amazing 00:11:49.320 --> 00:11:52.230 idea that... didn't quite work so his relentless 00:11:52.230 --> 00:11:54.789 drive for complexity the same drive honed through 00:11:54.789 --> 00:11:57.190 five world championships in all -round strategy 00:11:57.190 --> 00:12:00.210 finally hit a technological wall in the commercial 00:12:00.210 --> 00:12:02.509 world yes He essentially failed commercially 00:12:02.509 --> 00:12:04.710 because he was trying to build something that 00:12:04.710 --> 00:12:08.370 required AGI before AGI existed. And that failure, 00:12:08.370 --> 00:12:10.429 or at least that necessary reduction of scope, 00:12:10.570 --> 00:12:12.730 was the ultimate catalyst for what came next. 00:12:13.049 --> 00:12:15.929 Exactly. He realized that the tools he needed, 00:12:16.009 --> 00:12:17.889 the fundamental principles of intelligence that 00:12:17.889 --> 00:12:20.190 could handle the complexity of simulating a nation. 00:12:20.840 --> 00:12:23.100 They just didn't exist yet. And you certainly 00:12:23.100 --> 00:12:25.159 couldn't shoehorn them into a commercially viable 00:12:25.159 --> 00:12:28.220 game engine. He had spent a decade building incredibly 00:12:28.220 --> 00:12:31.500 complex external worlds, but he needed to understand 00:12:31.500 --> 00:12:34.059 the ultimate internal simulation engine. The 00:12:34.059 --> 00:12:36.620 human brain. The human brain. Which necessitated 00:12:36.620 --> 00:12:40.000 a really radical move. He sold the IP and technology 00:12:40.000 --> 00:12:44.340 rights for Elixir Studios in 2005. He just stepped 00:12:44.340 --> 00:12:46.580 away from a high profile, high earning career 00:12:46.580 --> 00:12:49.429 to go back to school. That pivot speaks volumes 00:12:49.429 --> 00:12:51.629 about his underlying mission. And that pivot 00:12:51.629 --> 00:12:54.710 brings us to Section 3, the academic pivot. Right. 00:12:54.769 --> 00:12:56.830 This is where he intentionally sought to connect 00:12:56.830 --> 00:12:59.230 brain function with computation. And you could 00:12:59.230 --> 00:13:01.090 argue this is the most crucial transition of 00:13:01.090 --> 00:13:03.610 his entire career. He returned to academia, not 00:13:03.610 --> 00:13:05.950 for a career change, but with this explicit mission. 00:13:06.460 --> 00:13:09.159 to find inspiration in the human brain for new 00:13:09.159 --> 00:13:12.399 AI algorithms. He needed to figure out how the 00:13:12.399 --> 00:13:15.059 biological machines solved the complex real -world 00:13:15.059 --> 00:13:17.360 problems that his video game simulations just 00:13:17.360 --> 00:13:20.080 couldn't handle. So he goes and gets a PhD in 00:13:20.080 --> 00:13:22.399 cognitive neuroscience from University College 00:13:22.399 --> 00:13:26.419 London. His thesis, completed in 2009, was on 00:13:26.419 --> 00:13:28.799 neural processes underpinning episodic memory. 00:13:29.610 --> 00:13:31.710 But he didn't just stop there. No, he made sure 00:13:31.710 --> 00:13:34.190 he was steeped in the very best computational 00:13:34.190 --> 00:13:37.789 neuroscience environments. A postdoctoral fellowship 00:13:37.789 --> 00:13:40.370 at the Gatsby Computational Neuroscience Unit 00:13:40.370 --> 00:13:43.049 at UCL. And visiting scientist roles at MIT, 00:13:43.429 --> 00:13:46.409 specifically in Tommaso Poggio's lab and at Harvard. 00:13:46.549 --> 00:13:49.009 He was systematically retraining his brain and 00:13:49.009 --> 00:13:51.110 absorbing all the latest research on how the 00:13:51.110 --> 00:13:53.250 brain works, especially how it handles memory 00:13:53.250 --> 00:13:55.820 and planning. And this period of research yielded 00:13:55.820 --> 00:13:58.539 truly groundbreaking results. He co -authored 00:13:58.539 --> 00:14:00.940 several highly influential papers in Nature, 00:14:01.120 --> 00:14:04.360 Science, PNAS, which established a core theory 00:14:04.360 --> 00:14:06.820 about how the brain functions, a theory that 00:14:06.820 --> 00:14:09.000 later became the blueprint for DeepMind's architecture. 00:14:09.379 --> 00:14:11.340 It really did. Let's focus on that core finding 00:14:11.340 --> 00:14:13.960 from his very first academic paper. It fundamentally 00:14:13.960 --> 00:14:17.279 redefined how we view amnesia. That landmark 00:14:17.279 --> 00:14:20.600 paper demonstrated, through really systematic 00:14:20.600 --> 00:14:24.100 testing, that Patients who had damage to the 00:14:24.100 --> 00:14:26.539 hippocampus, you know, the brain structure we 00:14:26.539 --> 00:14:28.679 all associate with amnesia, the inability to 00:14:28.679 --> 00:14:31.320 form new episodic memories. Right. They also 00:14:31.320 --> 00:14:33.580 suffered from another specific critical deficit. 00:14:34.159 --> 00:14:37.360 They were unable to imagine themselves in new 00:14:37.360 --> 00:14:40.440 future experiences. So it wasn't just a failure 00:14:40.440 --> 00:14:43.100 to recall the past. It was a failure to construct 00:14:43.100 --> 00:14:45.740 the future. That's a massive insight. It's huge. 00:14:45.879 --> 00:14:48.460 It connects memory, which we traditionally view 00:14:48.460 --> 00:14:51.399 as a reconstructive process, a playback, with 00:14:51.399 --> 00:14:53.549 imagination. which is a constructive process. 00:14:54.190 --> 00:14:56.669 The same underlying neural machinery that pieces 00:14:56.669 --> 00:14:58.649 together the details of what happened last week 00:14:58.649 --> 00:15:00.970 is needed to piece together the details of what 00:15:00.970 --> 00:15:03.309 might happen tomorrow. So how did he and his 00:15:03.309 --> 00:15:05.169 colleagues formalize that connection? What was 00:15:05.169 --> 00:15:07.220 the theory? They developed a new theoretical 00:15:07.220 --> 00:15:09.679 account based on what they term scene construction. 00:15:10.080 --> 00:15:12.600 The mechanism is this. The brain has to generate 00:15:12.600 --> 00:15:15.559 and maintain a complex, coherent, spatial, and 00:15:15.559 --> 00:15:17.600 temporal scene in the mind. Like a movie set. 00:15:17.740 --> 00:15:19.460 A bit like a movie set, yeah. If they have a 00:15:19.460 --> 00:15:21.759 campus that's damaged, you lose the ability to 00:15:21.759 --> 00:15:24.919 fluently link and generate the specific contextual 00:15:24.919 --> 00:15:27.879 details, the who, what, where, and when that 00:15:27.879 --> 00:15:30.000 are necessary to build that scene. Without the 00:15:30.000 --> 00:15:31.860 ability to construct the scene of the past, you 00:15:31.860 --> 00:15:34.279 cannot construct the scene of a novel future 00:15:34.279 --> 00:15:37.070 experience. That takes the concept of memory 00:15:37.070 --> 00:15:39.789 way beyond simple data retrieval. It turns it 00:15:39.789 --> 00:15:42.629 into an active system simulation. Precisely. 00:15:42.669 --> 00:15:45.779 And this is the intellectual leap. Hassabis generalized 00:15:45.779 --> 00:15:48.759 the specific fighting into this grander notion 00:15:48.759 --> 00:15:51.139 of the simulation engine of the mind. The simulation 00:15:51.139 --> 00:15:53.740 engine of the mind. He proposed that the brain's 00:15:53.740 --> 00:15:56.639 primary evolutionary advantage and its main computational 00:15:56.639 --> 00:15:59.220 role isn't just reacting to the present moment, 00:15:59.340 --> 00:16:01.899 but constantly acting as a simulation engine. 00:16:02.039 --> 00:16:05.500 It uses past data stored as episodic memories 00:16:05.500 --> 00:16:08.720 to internally imagine or simulate novel events 00:16:08.720 --> 00:16:10.820 and scenarios. And the whole purpose of this 00:16:10.820 --> 00:16:13.299 simulation is to get better at planning. Better 00:16:13.299 --> 00:16:15.740 planning. Better prediction, more adaptive decision 00:16:15.740 --> 00:16:18.100 -making for the future. That's the goal. Which 00:16:18.100 --> 00:16:20.179 is a direct parallel to the competitive games 00:16:20.179 --> 00:16:23.220 he mastered, isn't it? Chess, poker, diplomacy, 00:16:23.480 --> 00:16:26.279 they all rely on mentally simulating future board 00:16:26.279 --> 00:16:29.419 states or future social outcomes. And now he 00:16:29.419 --> 00:16:31.639 had found the biological mechanism behind that 00:16:31.639 --> 00:16:34.799 ability. The synergy is undeniable. When he built 00:16:34.799 --> 00:16:36.639 Republic, he was trying to build an external 00:16:36.639 --> 00:16:39.740 simulation engine for political planning. When 00:16:39.740 --> 00:16:41.919 he studied the hippocampus, he identified the 00:16:41.919 --> 00:16:44.500 ultimate internal simulation engine for planning. 00:16:44.960 --> 00:16:47.399 His neuroscience research provided the intellectual 00:16:47.399 --> 00:16:50.220 architecture for the AGI he would go on to build. 00:16:50.340 --> 00:16:52.580 And the scientific community recognized the power 00:16:52.580 --> 00:16:55.740 of this idea immediately. They did. This research 00:16:55.740 --> 00:16:57.720 on memory and imagination was listed in the top 00:16:57.720 --> 00:17:00.080 10 scientific breakthroughs of the year by the 00:17:00.080 --> 00:17:03.279 journal Science in 2007. That's just an astonishing 00:17:03.279 --> 00:17:05.380 recognition for someone who had just pivoted 00:17:05.380 --> 00:17:07.700 from running a video game studio. It underscores 00:17:07.700 --> 00:17:09.920 that the academic pivot wasn't just a sabbatical. 00:17:10.039 --> 00:17:12.759 It was a mission -critical research phase. He 00:17:12.759 --> 00:17:15.279 went into the deepest realm of biology, the human 00:17:15.279 --> 00:17:18.519 mind, to find the missing blueprint for generalized 00:17:18.519 --> 00:17:21.619 intelligence. And he found it. He found it in 00:17:21.619 --> 00:17:25.019 our ability to simulate and predict. Once he 00:17:25.019 --> 00:17:27.140 had that blueprint, that theoretical framework, 00:17:27.480 --> 00:17:30.500 it was time to build. This brings us to Section 00:17:30.500 --> 00:17:34.079 4, DeepMind and the Quest for AGI, covering the 00:17:34.079 --> 00:17:37.880 years 2010 to 2017. He founded DeepMind in London 00:17:37.880 --> 00:17:40.539 in 2010 with Shane Legg and Mustafa Suleiman, 00:17:40.700 --> 00:17:44.079 and crucially, he recruited David Silver, a collaborator 00:17:44.079 --> 00:17:46.799 from his Elixir Studios gaming days. The founding 00:17:46.799 --> 00:17:49.779 mission was... Well, it was perhaps the most 00:17:49.779 --> 00:17:52.559 ambitious in modern corporate history to solve 00:17:52.559 --> 00:17:54.759 intelligence and then use that intelligence to 00:17:54.759 --> 00:17:56.440 solve everything else. That's a modest goal. 00:17:56.619 --> 00:17:58.799 Just a small thing, yeah. They were aiming for 00:17:58.799 --> 00:18:02.420 artificial general intelligence, or AGI, by integrating 00:18:02.420 --> 00:18:04.720 the insights from systems neuroscience to Savas' 00:18:04.819 --> 00:18:07.369 latest specialty with machine learning. And where 00:18:07.369 --> 00:18:09.109 do you start when you want to create a general 00:18:09.109 --> 00:18:11.009 purpose simulation engine? You go back to the 00:18:11.009 --> 00:18:12.849 world of structured, contained environments, 00:18:12.990 --> 00:18:15.849 games. But this time, they weren't designing 00:18:15.849 --> 00:18:18.150 the simulation. They were designing the AI to 00:18:18.150 --> 00:18:20.789 master the simulation autonomously. That's where 00:18:20.789 --> 00:18:24.630 the DeepQ network, or DQN, comes in. This was 00:18:24.630 --> 00:18:26.789 their pioneering breakthrough in deep reinforcement 00:18:26.789 --> 00:18:30.359 learning, announced in 2013. The goal was to 00:18:30.359 --> 00:18:33.299 train an algorithm to master dozens of classic 00:18:33.299 --> 00:18:36.059 Atari games like Breakout or Space Invaders at 00:18:36.059 --> 00:18:39.299 a superhuman level. What made DQN such a breakthrough 00:18:39.299 --> 00:18:42.660 compared to previous AI attempts at gaming? The 00:18:42.660 --> 00:18:46.240 novelty lay in the generalizability and the input. 00:18:46.859 --> 00:18:49.359 Previous game -playing AI was often specifically 00:18:49.359 --> 00:18:52.180 coded with the rules of the game. An expert had 00:18:52.180 --> 00:18:54.480 to sit down and tell it how to play. But DQN 00:18:54.480 --> 00:18:56.859 was different. Completely different. It only 00:18:56.859 --> 00:18:59.339 received the raw, unstructured pixels on the 00:18:59.339 --> 00:19:02.359 screen as input, just like a human seeing a video 00:19:02.359 --> 00:19:04.720 game for the first time. It used deep learning 00:19:04.720 --> 00:19:06.680 to interpret that visual input and reinforcement 00:19:06.680 --> 00:19:09.019 learning to figure out the rules of the environment 00:19:09.019 --> 00:19:11.480 and maximize its reward, which was the score. 00:19:11.660 --> 00:19:14.200 So it was building its own internal model of 00:19:14.200 --> 00:19:16.420 the environment, its own simulation, simply by 00:19:16.420 --> 00:19:19.220 trial and error and seeking reward. This is the 00:19:19.220 --> 00:19:21.880 machine version of the brain's simulation engine 00:19:21.880 --> 00:19:25.329 concept, isn't it? Precisely. It learns the dynamics 00:19:25.329 --> 00:19:27.329 of the environment entirely from experience, 00:19:27.589 --> 00:19:30.430 creates a map of possible future states, and 00:19:30.430 --> 00:19:32.930 then chooses the action that maximizes its predicted 00:19:32.930 --> 00:19:36.210 future reward. That process of planning through 00:19:36.210 --> 00:19:38.569 predicting reward is essentially what Hassabis 00:19:38.569 --> 00:19:41.849 discovered the hippocampus does. The DQN breakthrough 00:19:41.849 --> 00:19:44.349 proved this approach could work for general tasks 00:19:44.349 --> 00:19:47.509 across multiple diverse environments without 00:19:47.509 --> 00:19:50.900 any manual intervention. If DQN was the proof 00:19:50.900 --> 00:19:53.460 of concept, then AlphaGo was the proof of scale 00:19:53.460 --> 00:19:56.960 and complexity. It took on the game of Go, which, 00:19:57.039 --> 00:19:59.799 as we know, was the holy grail of AI. Go was 00:19:59.799 --> 00:20:02.259 considered a much harder problem than chess because 00:20:02.259 --> 00:20:05.220 of its sheer scope. Chess has complexity, but 00:20:05.220 --> 00:20:07.359 Go's branching factor, the number of moves possible 00:20:07.359 --> 00:20:10.019 from any single position, is exponentially larger. 00:20:10.240 --> 00:20:12.259 The sources confirm that the number of possible 00:20:12.259 --> 00:20:14.759 Go board positions exceeds the number of atoms 00:20:14.759 --> 00:20:17.440 in the observable universe. It does. which meant 00:20:17.440 --> 00:20:19.400 that brute force search was just completely impossible. 00:20:19.819 --> 00:20:22.079 So how did AlphaGo solve a problem that decades 00:20:22.079 --> 00:20:24.519 of AI research couldn't even touch? It didn't 00:20:24.519 --> 00:20:27.660 rely on the old minimax search alone. AlphaGo 00:20:27.660 --> 00:20:30.400 combined the deep learning of the DQN architecture, 00:20:30.859 --> 00:20:33.319 analyzing the visual input of the board state, 00:20:33.539 --> 00:20:37.069 with advanced Monte Carlo tree search. It essentially 00:20:37.069 --> 00:20:40.089 learned to evaluate board positions like a grandmaster. 00:20:40.250 --> 00:20:42.269 It learned intuition. It learned an artificial 00:20:42.269 --> 00:20:45.069 form of intuition, yes. It could determine which 00:20:45.069 --> 00:20:47.869 moves were strategically promising rather than 00:20:47.869 --> 00:20:50.190 simply exploring every single possible path. 00:20:50.680 --> 00:20:53.779 It had developed an intuition informed by vast 00:20:53.779 --> 00:20:56.279 experience, much like a human expert. And the 00:20:56.279 --> 00:20:58.559 trajectory of its public dominance was astounding. 00:20:58.660 --> 00:21:00.720 It's all meticulously documented in the sources. 00:21:01.019 --> 00:21:03.220 It started by beating the European champion Fan 00:21:03.220 --> 00:21:07.339 Huan 5 -0 in late 2015. But the real moment of 00:21:07.339 --> 00:21:10.799 global shock came in March 2016 when it defeated 00:21:10.799 --> 00:21:13.200 the legendary former world champion Lee Sedol 00:21:13.200 --> 00:21:15.740 4 -1. And Lee Sedol wasn't just a master. He 00:21:15.740 --> 00:21:17.519 was considered one of the most innovative and 00:21:17.519 --> 00:21:19.980 strategic players in history. A true genius of 00:21:19.980 --> 00:21:22.569 the game. To see him defeated so decisively was, 00:21:22.730 --> 00:21:25.250 it was a watershed moment. And that chapter was 00:21:25.250 --> 00:21:28.630 definitively closed in 2017 when AlphaGo beat 00:21:28.630 --> 00:21:44.140 the then world's top -ranked player, KG. The 00:21:44.140 --> 00:21:47.579 scale of these breakthroughs made DeepMind too 00:21:47.579 --> 00:21:50.160 valuable to remain independent for long. Google 00:21:50.160 --> 00:21:53.109 purchased them in 2014 for $400 million. And 00:21:53.109 --> 00:21:54.990 the subsequent real -world applications show 00:21:54.990 --> 00:21:56.829 that this general -purpose learning wasn't just 00:21:56.829 --> 00:21:59.109 for simulated strategy. No, they immediately 00:21:59.109 --> 00:22:01.250 started applying it to real -world complexity. 00:22:01.490 --> 00:22:04.009 The first major one listed in the sources is 00:22:04.009 --> 00:22:06.609 optimizing Google's massive data centers. Right. 00:22:06.690 --> 00:22:09.190 These facilities consume staggering amounts of 00:22:09.190 --> 00:22:11.710 energy, particularly for cooling. DeepMind's 00:22:11.710 --> 00:22:13.829 AI was deployed to manage the climate controls. 00:22:14.029 --> 00:22:16.170 It made predictions about future thermal load 00:22:16.170 --> 00:22:18.009 and adjusted the cooling systems proactively. 00:22:18.230 --> 00:22:20.930 The result? A 40 % reduction in the energy required 00:22:20.930 --> 00:22:23.400 for cooling. A 40 %? percent reduction. That's 00:22:23.400 --> 00:22:25.900 a staggering efficiency gain derived from a system 00:22:25.900 --> 00:22:28.099 that was born from reinforcement learning research 00:22:28.099 --> 00:22:30.920 for video games. And the second major area of 00:22:30.920 --> 00:22:33.720 application was health moving into high -stakes 00:22:33.720 --> 00:22:36.240 environments where pattern recognition could 00:22:36.240 --> 00:22:39.839 save time and possibly sight. Yes. They partnered 00:22:39.839 --> 00:22:42.359 with the UK's National Health Service, specifically 00:22:42.359 --> 00:22:45.539 Moorfield's Eye Hospital. They used deep learning 00:22:45.539 --> 00:22:48.319 to analyze scans of the eye, looking for complex 00:22:48.319 --> 00:22:50.759 patterns that are indicative of degenerative 00:22:50.759 --> 00:22:53.440 eye conditions. And the AI could identify these 00:22:53.440 --> 00:22:55.900 conditions with the same accuracy as a human 00:22:55.900 --> 00:22:58.599 specialist. The same accuracy, but instantly. 00:22:59.200 --> 00:23:02.079 dramatically speeding up diagnosis and treatment 00:23:02.079 --> 00:23:05.039 planning. This whole period just cemented DeepMind's 00:23:05.039 --> 00:23:07.880 status. The development of other advanced systems, 00:23:07.960 --> 00:23:10.740 like the neural Turing machine, and AI that combined 00:23:10.740 --> 00:23:13.259 deep learning with an external memory bank. Another 00:23:13.259 --> 00:23:15.500 architectural nod to human cognitive function. 00:23:15.700 --> 00:23:17.940 It just showed they were building general scalable 00:23:17.940 --> 00:23:20.259 intelligence, an intelligence that was ready 00:23:20.259 --> 00:23:22.799 to tackle the ultimate challenge, fundamental 00:23:22.799 --> 00:23:25.480 science. And that leads us directly into Section 00:23:25.480 --> 00:23:28.920 5. The Nobel breakthrough. This is the culmination 00:23:28.920 --> 00:23:30.960 of everything Hassabis had learned about games, 00:23:31.180 --> 00:23:33.740 simulation, and the architecture of the planning 00:23:33.740 --> 00:23:39.000 mind. DeepMind pivoted its focus to, well, arguably 00:23:39.000 --> 00:23:41.279 the hardest scientific problem of the last half 00:23:41.279 --> 00:23:43.720 century. They focused on protein structure prediction, 00:23:44.059 --> 00:23:46.660 often called the protein folding problem. For 00:23:46.660 --> 00:23:48.720 the listener, can you break down why this specific 00:23:48.720 --> 00:23:51.839 problem is so monumental? Of course. So proteins 00:23:51.839 --> 00:23:53.799 are the workhorses of life. They build, they 00:23:53.799 --> 00:23:56.140 repair, they catalyze reactions. They do everything. 00:23:56.400 --> 00:23:58.779 They start as a simple linear chain of amino 00:23:58.779 --> 00:24:01.500 acids, which is called a primary structure. But 00:24:01.500 --> 00:24:04.220 to function, they must fold rapidly into a complex, 00:24:04.319 --> 00:24:07.309 unique 3D shape. the tertiary structure. And 00:24:07.309 --> 00:24:09.190 the function is completely dependent on that 00:24:09.190 --> 00:24:11.650 final 3D shape. Entirely. So we know the recipe, 00:24:11.769 --> 00:24:14.470 the amino acid sequence, but predicting the final 00:24:14.470 --> 00:24:16.769 3D cooked dish from that recipe has been the 00:24:16.769 --> 00:24:18.910 monumental challenge for 50 years. And why is 00:24:18.910 --> 00:24:20.509 it so difficult for traditional computation? 00:24:21.009 --> 00:24:23.430 The number of possible ways a protein could fold 00:24:23.430 --> 00:24:26.390 is astronomically large. It's even more complex 00:24:26.390 --> 00:24:29.849 than Go. A small chain of just 100 amino acids 00:24:29.849 --> 00:24:32.769 has a theoretical number of configurations that 00:24:32.769 --> 00:24:35.230 just dwarfs the search space of any board game. 00:24:35.690 --> 00:24:37.990 Traditional computational methods simply couldn't 00:24:37.990 --> 00:24:40.029 search that space quickly or accurately enough 00:24:40.029 --> 00:24:43.190 to find the single correct lowest energy folded 00:24:43.190 --> 00:24:46.170 state. This was a massive bottleneck for drug 00:24:46.170 --> 00:24:49.069 discovery, for biological research. A huge bottleneck. 00:24:49.069 --> 00:24:51.420 The entire field was stuck. This challenge is 00:24:51.420 --> 00:24:54.160 formalized in a competition called CASP, the 00:24:54.160 --> 00:24:56.119 Critical Assessment of Techniques for Protein 00:24:56.119 --> 00:24:58.900 Structure Prediction. And DeepMind entered this 00:24:58.900 --> 00:25:02.119 highly specialized arena and introduced AlphaFold1 00:25:02.119 --> 00:25:05.480 in 2018. And AlphaFold1 immediately exceeded 00:25:05.480 --> 00:25:08.859 all expectations. It won the 13th CASP competition 00:25:08.859 --> 00:25:12.099 by accurately predicting 25 out of 43 target 00:25:12.099 --> 00:25:14.940 protein structures. This was a significant advance 00:25:14.940 --> 00:25:17.319 over all existing computational methods. And 00:25:17.319 --> 00:25:19.500 Hassabis himself described this as DeepMind's 00:25:19.500 --> 00:25:22.700 first major investment into a fundamental, very 00:25:22.700 --> 00:25:25.259 important real -world scientific problem. It 00:25:25.259 --> 00:25:27.240 was a signal of their new direction. It was a 00:25:27.240 --> 00:25:30.299 huge signal, a shift from games to solving core 00:25:30.299 --> 00:25:33.140 human challenges. But the real seismic shift 00:25:33.140 --> 00:25:35.839 came with AlphaFold 2, which they released in 00:25:35.839 --> 00:25:39.599 2020. This didn't just win CASP14, it essentially 00:25:39.599 --> 00:25:42.180 ended the competition. The jump in performance 00:25:42.180 --> 00:25:44.619 was revolutionary. It's like the breakthrough 00:25:44.619 --> 00:25:46.460 we saw between the first computer chess programs 00:25:46.460 --> 00:25:49.779 in AlphaGo, but for biology. AlphaFold2 achieved 00:25:49.779 --> 00:25:52.079 world -beating results by shifting the entire 00:25:52.079 --> 00:25:54.079 paradigm of the problem. Let's talk about the 00:25:54.079 --> 00:25:56.519 metric they use, the Global Distance Test, or 00:25:56.519 --> 00:25:59.460 GDT score. This is how CSP measures accuracy. 00:25:59.859 --> 00:26:02.220 Right. The GDT score measures how closely the 00:26:02.220 --> 00:26:04.819 predicted 3D structure matches the actual experimental 00:26:04.819 --> 00:26:07.240 structure. Yeah. A score of 100. means a perfect 00:26:07.240 --> 00:26:10.279 match. In 2018, AlphaFold1 was achieving median 00:26:10.279 --> 00:26:12.660 GDT scores of under 60. Which was already state 00:26:12.660 --> 00:26:15.599 -of -the -art. It was. But in 2020, AlphaFold2 00:26:15.599 --> 00:26:19.619 hit a median GDT score of 87 .0 in the most challenging 00:26:19.619 --> 00:26:21.759 free modeling category. And what does that 87 00:26:21.759 --> 00:26:24.480 .0 score mean in physical scientific terms? Why 00:26:24.480 --> 00:26:27.000 was it considered solved? It means the system's 00:26:27.000 --> 00:26:30.039 overall prediction error was less than the width 00:26:30.039 --> 00:26:32.880 of a single carbon atom, less than one angstrom. 00:26:33.319 --> 00:26:35.880 This level of precision is comparable to the 00:26:35.880 --> 00:26:38.339 resolution you get from time -consuming and expensive 00:26:38.339 --> 00:26:42.059 experimental lab methods, like X -ray crystallography 00:26:42.059 --> 00:26:45.660 or cryo -electron microscopy. So the prediction 00:26:45.660 --> 00:26:47.500 was as good as the measurement? Essentially, 00:26:47.599 --> 00:26:51.160 yes. The CASP organizers, upon seeing these results, 00:26:51.380 --> 00:26:54.099 declared the 50 -year problem essentially solved. 00:26:54.539 --> 00:26:57.599 The predictive power of the AI had reached parity 00:26:57.599 --> 00:27:00.029 with physical measurement. That is the moment 00:27:00.029 --> 00:27:02.170 where the strategy learned from chess and Go, 00:27:02.369 --> 00:27:04.869 and the simulation insight learned from the human 00:27:04.869 --> 00:27:08.190 hippocampus, all converge to unlock molecular 00:27:08.190 --> 00:27:11.230 biology. It's a stunning synthesis. And it was 00:27:11.230 --> 00:27:14.029 recognized as such. AlphaFold2 was named the 00:27:14.029 --> 00:27:16.430 winner of Science's 2021 Breakthrough of the 00:27:16.430 --> 00:27:18.950 Year. This marked the fourth time Hassabis' work 00:27:18.950 --> 00:27:21.470 had made that prestigious list. It just cemented 00:27:21.470 --> 00:27:24.329 his place not just as a technology pioneer, but 00:27:24.329 --> 00:27:26.170 as one of the most impactful scientists of his 00:27:26.170 --> 00:27:28.279 generation. And all of that foundational work 00:27:28.279 --> 00:27:30.839 led to the ultimate scientific validation, the 00:27:30.839 --> 00:27:34.259 Nobel Prize. Exactly. Hassabis and John M. Jumper, 00:27:34.359 --> 00:27:36.880 the lead scientist on AlphaFold, were jointly 00:27:36.880 --> 00:27:40.299 awarded the 2024 Nobel Prize in Chemistry for 00:27:40.299 --> 00:27:43.140 this achievement. The Swedish Academy recognized 00:27:43.140 --> 00:27:45.460 that this AI methodology fundamentally changed 00:27:45.460 --> 00:27:47.799 how we approach chemistry and biology. It's a 00:27:47.799 --> 00:27:50.099 Nobel awarded for the invention of a powerful 00:27:50.099 --> 00:27:53.099 new tool that unlocks knowledge previously inaccessible 00:27:53.099 --> 00:27:55.660 to humanity. But the story doesn't end with the 00:27:55.660 --> 00:27:58.720 competition win or the Nobel. DeepMind took this 00:27:58.720 --> 00:28:01.220 technology and scaled its contribution immensely. 00:28:01.500 --> 00:28:04.539 They moved from solving a few dozen CASP targets 00:28:04.539 --> 00:28:07.420 to solving the entire known universe of proteins. 00:28:08.119 --> 00:28:10.960 Over the next year, DeepMind used AlphaFold2 00:28:10.960 --> 00:28:14.200 to fold all 200 million proteins known to science. 00:28:14.440 --> 00:28:16.829 All of them? All of them. And crucially, they 00:28:16.829 --> 00:28:18.690 didn't hoard this information. They made the 00:28:18.690 --> 00:28:20.849 entire database of predicted structures and the 00:28:20.849 --> 00:28:23.549 system itself openly and freely available via 00:28:23.549 --> 00:28:26.309 the AlphaFold protein structure database in collaboration 00:28:26.309 --> 00:28:29.170 with EMBL -EB. It was a gift to the world. It 00:28:29.170 --> 00:28:31.109 was a gift of scientific knowledge to the world. 00:28:31.210 --> 00:28:33.289 It immediately accelerated research in everything 00:28:33.289 --> 00:28:36.009 from malaria to cancer. And Hassabis is now taking 00:28:36.009 --> 00:28:38.210 this capability directly to the most practical 00:28:38.210 --> 00:28:40.799 application possible. drug development. That's 00:28:40.799 --> 00:28:43.640 the purpose of Isomorphic Labs, which he co -founded 00:28:43.640 --> 00:28:47.140 and serves as CEO. It was founded in 2021 as 00:28:47.140 --> 00:28:49.799 a separate alphabet company dedicated specifically 00:28:49.799 --> 00:28:52.839 to leveraging these advanced AI capabilities 00:28:52.839 --> 00:28:56.339 for AI driven drug discovery. This is the ultimate 00:28:56.730 --> 00:28:59.430 high -stakes simulation environment. Right. No 00:28:59.430 --> 00:29:01.670 longer planning a move and go or managing an 00:29:01.670 --> 00:29:04.369 amusement park. But simulating molecular interactions 00:29:04.369 --> 00:29:07.109 to design life -saving medicines. As impressive 00:29:07.109 --> 00:29:08.910 as these technical and scientific achievements 00:29:08.910 --> 00:29:12.430 are, we have to address Section 6. Context, recognition, 00:29:12.750 --> 00:29:15.029 and the ethical imperatives that come with developing 00:29:15.029 --> 00:29:17.990 artificial general intelligence. Hassabis has 00:29:17.990 --> 00:29:19.829 helped unleash a technology that the sources 00:29:19.829 --> 00:29:22.329 suggest carries, well, massive potential, but 00:29:22.329 --> 00:29:25.250 also massive risks. His stance on AI safety is 00:29:25.250 --> 00:29:27.529 a central theme. in his public commentary. And 00:29:27.529 --> 00:29:29.630 it's complex. It reflects a deep awareness of 00:29:29.630 --> 00:29:32.369 both utopian possibility and potential catastrophe. 00:29:32.769 --> 00:29:35.670 On the one hand, he's intensely optimistic. He's 00:29:35.670 --> 00:29:37.670 predicted AI will be one of the most beneficial 00:29:37.670 --> 00:29:40.450 technologies of mankind ever. He champions the 00:29:40.450 --> 00:29:43.349 idea that the potential upsides, solving climate 00:29:43.349 --> 00:29:46.109 change, curing diseases, are too significant 00:29:46.109 --> 00:29:49.000 to just ignore. Right. But on the other hand, 00:29:49.079 --> 00:29:51.819 he has been a consistent voice calling for caution 00:29:51.819 --> 00:29:53.880 and for formal safety research. He's not just 00:29:53.880 --> 00:29:56.019 passively cautious either. He's engaged in some 00:29:56.019 --> 00:29:59.160 high -stakes advocacy. He has. He was a signatory 00:29:59.160 --> 00:30:02.160 in that widely publicized 2023 statement, which 00:30:02.160 --> 00:30:04.180 declared that mitigating the risk of extinction 00:30:04.180 --> 00:30:07.119 from AI should be treated as a global priority, 00:30:07.299 --> 00:30:10.099 right alongside other existential societal risks 00:30:10.099 --> 00:30:12.799 like pandemics and nuclear war. That's a very 00:30:12.799 --> 00:30:15.339 clear articulation of the seriousness with which 00:30:15.339 --> 00:30:17.640 he views the power of the systems he is creating. 00:30:17.799 --> 00:30:20.940 It is. However, he has resisted calls for a complete 00:30:20.940 --> 00:30:24.559 global pause on AI development. He makes a distinction 00:30:24.559 --> 00:30:27.339 between safe, regulated progress and just halting 00:30:27.339 --> 00:30:30.440 research entirely. He argues that a global pause 00:30:30.440 --> 00:30:32.940 is politically and practically impossible to 00:30:32.940 --> 00:30:35.299 enforce, especially internationally. And more 00:30:35.299 --> 00:30:36.920 importantly, he believes the potential benefits, 00:30:37.099 --> 00:30:39.000 particularly in fundamental science and medicine, 00:30:39.160 --> 00:30:42.339 are too great to delay. His alternative solution 00:30:42.339 --> 00:30:44.559 is to focus on urgent research into evaluation 00:30:44.559 --> 00:30:47.859 tests. Evaluation tests. What does that specifically 00:30:47.859 --> 00:30:51.609 entail in the context of AGI safety? It means 00:30:51.609 --> 00:30:55.029 developing scientific, reliable ways to measure 00:30:55.029 --> 00:30:58.849 exactly how capable new AI models are long before 00:30:58.849 --> 00:31:01.890 they are deployed. And even more crucially, developing 00:31:01.890 --> 00:31:04.670 tests to measure their controllability. So we 00:31:04.670 --> 00:31:06.349 need to know that we can keep them in check. 00:31:06.589 --> 00:31:08.829 We need to know that as these systems become 00:31:08.829 --> 00:31:11.569 vastly more intelligent, we can reliably align 00:31:11.569 --> 00:31:13.769 their goals with human values and ensure they 00:31:13.769 --> 00:31:16.829 remain docile and controllable. This is the safety 00:31:16.829 --> 00:31:19.529 field he champions, understanding and managing 00:31:19.529 --> 00:31:22.650 advanced AI, not just stopping it. And the recognition 00:31:22.650 --> 00:31:24.690 he's received, which is reflected across all 00:31:24.690 --> 00:31:27.089 the sources, it mirrors his massive influence 00:31:27.089 --> 00:31:29.529 across science, technology and government policy. 00:31:29.769 --> 00:31:31.970 It's an unprecedented array of honors. On the 00:31:31.970 --> 00:31:34.789 policy side, he was knighted in 2024 for services 00:31:34.789 --> 00:31:37.089 to artificial intelligence. He was appointed 00:31:37.089 --> 00:31:38.549 a commander of the Order of the British Empire, 00:31:38.789 --> 00:31:41.910 a CBE, back in 2017. He currently serves as a 00:31:41.910 --> 00:31:44.839 UK government AI advisor. and is an advisor to 00:31:44.839 --> 00:31:47.160 the Advanced Research and Invention Agency, or 00:31:47.160 --> 00:31:50.440 ARIA. He is actively shaping the policy landscape 00:31:50.440 --> 00:31:53.119 for the technology he pioneered. And the academic 00:31:53.119 --> 00:31:56.019 world has embraced him fully, despite his decade 00:31:56.019 --> 00:31:58.900 -long detour into video game design. Oh, completely. 00:31:59.099 --> 00:32:01.140 He was elected a Fellow of the Royal Society 00:32:01.140 --> 00:32:04.259 in 2018, a Fellow of the Royal Academy of Engineering 00:32:04.259 --> 00:32:07.299 in 2017. But the most telling measure of his 00:32:07.299 --> 00:32:09.579 impact is the repeated scientific validation. 00:32:10.319 --> 00:32:12.859 As we've noted, His work has been singled out 00:32:12.859 --> 00:32:15.640 by science as a top 10 scientific breakthrough 00:32:15.640 --> 00:32:18.400 four separate times. The neuroscience work in 00:32:18.400 --> 00:32:22.519 2007. AlphaGo in 2016, AlphaFold1 in 2020, and 00:32:22.519 --> 00:32:25.940 then AlphaFold the winner in 2021. This level 00:32:25.940 --> 00:32:27.900 of consistent cross -disciplinary breakthrough 00:32:27.900 --> 00:32:31.240 is virtually unparalleled. He's also transcended 00:32:31.240 --> 00:32:33.319 the academic bubble and become a major public 00:32:33.319 --> 00:32:35.779 figure in the technology sphere. He was recognized 00:32:35.779 --> 00:32:37.960 in the Time 100 Most Influential People in the 00:32:37.960 --> 00:32:41.140 World in 2017 and again in 2025. He was also 00:32:41.140 --> 00:32:42.859 part of the group recognized as the architects 00:32:42.859 --> 00:32:46.099 of AI for Times 2025 Person of the Year. And 00:32:46.099 --> 00:32:48.359 the fact that his life story is the subject of 00:32:48.359 --> 00:32:51.160 a major 2024 documentary, The Thinking Game, 00:32:51.299 --> 00:32:53.519 by the same filmmaker who did the famous AlphaGo 00:32:53.519 --> 00:32:56.140 documentary, it just confirms his role as a true 00:32:56.140 --> 00:32:58.200 cultural touchstone and intellectual pioneer. 00:32:58.579 --> 00:33:01.240 His career path is just a masterclass in how 00:33:01.240 --> 00:33:03.559 different disciplines, when applied with a singular 00:33:03.559 --> 00:33:06.559 goal, can feed into each other to create a synthesis 00:33:06.559 --> 00:33:09.619 of knowledge that no single field could produce 00:33:09.619 --> 00:33:12.039 on its own. own. The entire narrative is about 00:33:12.039 --> 00:33:14.299 the search for the ultimate simulation engine. 00:33:14.579 --> 00:33:17.039 Absolutely. He started with the structured simulation 00:33:17.039 --> 00:33:19.640 of games, realized the limitations, went back 00:33:19.640 --> 00:33:21.220 to study the architecture of the human brain 00:33:21.220 --> 00:33:23.640 simulation engine, and then used that architectural 00:33:23.640 --> 00:33:25.920 blueprint to build the ultimate computational 00:33:25.920 --> 00:33:28.740 simulation engine, DeepMind. And that engine 00:33:28.740 --> 00:33:31.559 has now successfully navigated the incredible 00:33:31.559 --> 00:33:34.400 complexity of molecular biology and protein folding. 00:33:34.990 --> 00:33:37.329 So what does this extraordinary journey mean 00:33:37.329 --> 00:33:39.849 for you, the listener, as you look at the world 00:33:39.849 --> 00:33:42.930 of technology and science? We've traced the path 00:33:42.930 --> 00:33:46.170 of Demis Asabis, from child chess master with 00:33:46.170 --> 00:33:49.690 a master level ELO by 13, the five -time world 00:33:49.690 --> 00:33:52.490 champion board game player, mastering the complexities 00:33:52.490 --> 00:33:54.789 of theory of mind and incomplete information, 00:33:55.089 --> 00:33:57.650 all the way to a pioneering video game simulation 00:33:57.650 --> 00:34:00.670 designer, pushing systems way beyond their limits. 00:34:01.180 --> 00:34:03.539 We saw him make that radical pivot to cognitive 00:34:03.539 --> 00:34:06.359 neuroscience, to prove the brain's power relies 00:34:06.359 --> 00:34:09.300 on active imagination and the internal simulation 00:34:09.300 --> 00:34:11.719 of future scenarios for planning, a finding that 00:34:11.719 --> 00:34:13.880 earned him a spot on science's breakthrough list. 00:34:14.139 --> 00:34:16.360 And finally, we saw how that same architecture 00:34:16.360 --> 00:34:19.420 was used to build AGI systems like DeepMind, 00:34:19.519 --> 00:34:21.739 culminating in the Nobel -winning achievement 00:34:21.739 --> 00:34:24.280 of AlphaFold, solving the protein folding problem. 00:34:24.679 --> 00:34:27.119 The consistent thread that binds this entire 00:34:27.119 --> 00:34:29.800 incredibly diverse trajectory is the pursuit 00:34:29.800 --> 00:34:32.300 of mastering or building systems that can execute 00:34:32.300 --> 00:34:35.719 deep general learning and highly structured simulation 00:34:35.719 --> 00:34:39.199 to solve previously intractable problems. He 00:34:39.199 --> 00:34:41.400 moves from games of perfect information to games 00:34:41.400 --> 00:34:44.019 of incomplete information to simulated human 00:34:44.019 --> 00:34:46.300 political reality to the simulated molecular 00:34:46.300 --> 00:34:48.900 reality of protein structure. The major takeaway 00:34:48.900 --> 00:34:51.199 here has to be the transcendent power of interdisciplinary 00:34:51.199 --> 00:34:54.579 thinking. His explicit insight from studying 00:34:54.579 --> 00:34:56.760 human memory, that the simulation engine of the 00:34:56.760 --> 00:34:59.539 mind is key for successful prediction and planning, 00:34:59.780 --> 00:35:02.739 was the critical intellectual framework that 00:35:02.739 --> 00:35:05.139 informed the deep reinforcement learning architecture 00:35:05.139 --> 00:35:09.329 of deep mind. The human and the machine understanding 00:35:09.329 --> 00:35:12.989 of general intelligence converged at this single 00:35:12.989 --> 00:35:15.730 idea of simulation. You see that link so clearly 00:35:15.730 --> 00:35:18.630 when you compare the process. The human hippocampus 00:35:18.630 --> 00:35:21.389 uses stored memory past data to construct and 00:35:21.389 --> 00:35:24.329 simulate a novel scenario, a plan for the future. 00:35:24.809 --> 00:35:27.329 And AlphaGo, or AlphaFold, used vast training 00:35:27.329 --> 00:35:30.050 data to construct and simulate the optimal future 00:35:30.050 --> 00:35:33.130 board state or the correct folded molecular structure. 00:35:33.369 --> 00:35:35.769 He built an external engine that perfectly mirrored 00:35:35.769 --> 00:35:37.849 his findings about our internal one. And this 00:35:37.849 --> 00:35:40.329 leaves us with a final, provocative thought that 00:35:40.329 --> 00:35:42.530 builds directly on the theoretical work Hassabis 00:35:42.530 --> 00:35:44.989 himself conducted on the brain. We know his research 00:35:44.989 --> 00:35:47.369 formalized the brain's simulation engine of the 00:35:47.369 --> 00:35:50.070 mind, which enables us to imagine events for 00:35:50.070 --> 00:35:52.619 better planning and collective survival. But 00:35:52.619 --> 00:35:56.239 now he has helped build a vastly superior external 00:35:56.239 --> 00:36:00.119 computational simulation engine AGI and powerfully 00:36:00.119 --> 00:36:02.840 demonstrated by AlphaFold success in mapping 00:36:02.840 --> 00:36:06.559 all 200 million known proteins. This engine operates 00:36:06.559 --> 00:36:09.599 at a scale and speed that no human mind or even 00:36:09.599 --> 00:36:11.940 a collective of human minds can possibly match. 00:36:12.179 --> 00:36:14.039 So the ultimate question isn't just whether AGI 00:36:14.039 --> 00:36:16.539 can solve scientific problems. It's what are 00:36:16.539 --> 00:36:18.420 the full implications for humanity's ability 00:36:18.420 --> 00:36:21.139 to plan, imagine and solve societal problems. 00:36:21.800 --> 00:36:23.619 we now possess an external simulation engine 00:36:23.619 --> 00:36:25.960 capable of accurately modeling extremely complex 00:36:25.960 --> 00:36:28.519 systems, climate dynamics, global pandemics, 00:36:28.559 --> 00:36:30.980 economic policies, the core challenge and the 00:36:30.980 --> 00:36:33.119 massive benefit of the technologies he has created 00:36:33.119 --> 00:36:35.179 will be leveraging that computational simulation 00:36:35.179 --> 00:36:37.699 engine to overcome our own human limitations 00:36:37.699 --> 00:36:40.519 in collective planning, execution, and long -term 00:36:40.519 --> 00:36:42.599 foresight. That is the grand challenge Hassabis 00:36:42.599 --> 00:36:43.579 has laid before the world.