WEBVTT 00:00:00.000 --> 00:00:04.179 Imagine looking at a timeline of modern technology, 00:00:04.320 --> 00:00:07.639 right? And on the far left side, you have a computer 00:00:07.639 --> 00:00:10.919 slowly figuring out how to bounce this little 00:00:10.919 --> 00:00:14.720 square digital ball across a black screen. Ah, 00:00:14.919 --> 00:00:18.179 yeah, the 1970s arcade game, POM. Exactly, POM. 00:00:18.280 --> 00:00:20.879 Now, if you trace that line all the way to the 00:00:20.879 --> 00:00:23.460 right, like to the very other end, you have a 00:00:23.460 --> 00:00:25.920 literal Nobel Prize in chemistry. Right. And 00:00:25.920 --> 00:00:29.120 the artificial intelligence engine powering basically 00:00:29.120 --> 00:00:31.440 the most advanced world -altering systems on 00:00:31.440 --> 00:00:33.939 the planet. I mean, it is a staggering trajectory. 00:00:34.039 --> 00:00:36.539 You are looking at a direct through line from 00:00:36.539 --> 00:00:40.240 a retro video game to a system that fundamentally 00:00:40.240 --> 00:00:42.299 decodes the building blocks of human biology. 00:00:42.520 --> 00:00:45.000 Welcome to the deep dive. Today, we are mapping 00:00:45.000 --> 00:00:47.240 out that wild evolution, looking through this 00:00:47.240 --> 00:00:50.299 massive, constantly updated dossier we have on 00:00:50.299 --> 00:00:52.539 Google DeepMind. And this covers everything up 00:00:52.539 --> 00:00:54.560 through March 2026. By the way, you realize this 00:00:54.560 --> 00:00:56.299 isn't just a standard corporate history. No, 00:00:56.299 --> 00:00:58.939 not at all. The mission for this deep dive is 00:00:58.939 --> 00:01:02.500 to figure out how a relatively niche gaming lab 00:01:02.500 --> 00:01:05.719 in London became the absolute epicenter of the 00:01:05.719 --> 00:01:10.159 global AI arms race. Yeah. More importantly, 00:01:10.480 --> 00:01:12.459 how the systems they build are already quietly 00:01:12.459 --> 00:01:14.579 running the background of your daily life. Yeah, 00:01:14.579 --> 00:01:16.840 because what we are really tracking today is 00:01:16.840 --> 00:01:19.519 the evolution of a very specific concept, which 00:01:19.519 --> 00:01:21.659 is reinforcement learning. Right. That is the 00:01:21.659 --> 00:01:23.219 mathematical engine connecting, like you said, 00:01:23.299 --> 00:01:26.439 a vintage arcade game to a literal breakthrough 00:01:26.439 --> 00:01:28.439 in cancer research. OK, let's unpack this. We 00:01:28.439 --> 00:01:30.000 have to start at the beginning. It's November 00:01:30.000 --> 00:01:35.319 2010, London. Three founders, Demis Asabas, Shane 00:01:35.319 --> 00:01:37.640 Legge, and Mostafa Suleiman. Yeah. They start 00:01:37.640 --> 00:01:40.299 this company called DeepMind. and their goal 00:01:40.299 --> 00:01:42.739 right out of the gate is incredibly audacious. 00:01:42.859 --> 00:01:44.819 They didn't want to build an AI to just do one 00:01:44.819 --> 00:01:46.359 specific thing. Right, because, I mean, we all 00:01:46.359 --> 00:01:48.500 remember IBM's Deep Blue in the 90s, right? Oh, 00:01:48.560 --> 00:01:50.260 yeah, the computer that beat Garry Kasparov at 00:01:50.260 --> 00:01:52.980 chess. Exactly. Deep Blue was a marvel, but it 00:01:52.980 --> 00:01:55.299 was hard -coded by humans specifically to play 00:01:55.299 --> 00:01:57.879 chess. If you asked Deep Blue to play checkers 00:01:57.879 --> 00:02:00.180 or, I don't know, boil an egg... It'd just crash. 00:02:00.519 --> 00:02:02.579 It would completely crash. It didn't actually 00:02:02.579 --> 00:02:05.540 know how to learn. But DeepMind wanted to build 00:02:05.540 --> 00:02:08.750 a general -purpose AI, right? Like an intelligence 00:02:08.750 --> 00:02:11.189 that could learn anything totally from scratch. 00:02:11.229 --> 00:02:13.430 Yeah, and to build a brain that can learn anything 00:02:13.430 --> 00:02:15.550 You need a safe environment where it can fail 00:02:15.550 --> 00:02:17.849 millions of times without causing real -world 00:02:17.849 --> 00:02:21.590 damage So their testing ground was vintage video 00:02:21.590 --> 00:02:24.009 games, which makes sense But the way they did 00:02:24.009 --> 00:02:27.250 it is what absolutely blows my mind They didn't 00:02:27.250 --> 00:02:31.050 feed the AI the actual code of the game. No, 00:02:31.069 --> 00:02:33.189 they did They just fed it the raw pixels on the 00:02:33.189 --> 00:02:35.050 screen. Think about a game like space invaders 00:02:35.050 --> 00:02:39.259 or breakout. Yeah the AI was given zero prior 00:02:39.259 --> 00:02:41.439 knowledge. I mean, no rule book at all. Nothing. 00:02:41.580 --> 00:02:43.680 It didn't know what a spaceship was or what a 00:02:43.680 --> 00:02:47.900 paddle was, or even that the concept of a score 00:02:47.900 --> 00:02:50.580 even existed. Or just match buttons randomly. 00:02:50.699 --> 00:02:53.000 Pretty much. But here is where that reinforcement 00:02:53.000 --> 00:02:56.080 learning kicks in. The AI takes an action. sees 00:02:56.080 --> 00:02:58.159 the result on the screen and gets a mathematical 00:02:58.159 --> 00:03:00.800 reward if its score goes up. It then updates 00:03:00.800 --> 00:03:03.599 its internal neural network to favor that specific 00:03:03.599 --> 00:03:07.020 action. It's essentially mimicking human cognitive 00:03:07.020 --> 00:03:10.080 processes, you know, just trial and error. It's 00:03:10.080 --> 00:03:12.860 like dropping a kid in a room with a hyper complex 00:03:12.860 --> 00:03:15.479 board game and absolutely no rule book. They 00:03:15.479 --> 00:03:18.039 just start moving pieces around. But by the end 00:03:18.039 --> 00:03:20.759 of the week, they are the undisputed world champion. 00:03:20.800 --> 00:03:22.900 That's a great way to put it. And Google saw 00:03:22.900 --> 00:03:26.180 this happening. realized the huge potential and 00:03:26.180 --> 00:03:30.300 swooped in by 2014 acquiring DeepMind for somewhere 00:03:30.300 --> 00:03:33.780 between 400 and 650 million dollars. Which is 00:03:33.780 --> 00:03:36.039 what gave them the resources to tackle the holy 00:03:36.039 --> 00:03:38.379 grail of artificial intelligence at the time, 00:03:38.699 --> 00:03:41.259 which was the ancient board game Go. Right, because 00:03:41.259 --> 00:03:42.979 Go is a whole different beast than chess. You 00:03:42.979 --> 00:03:45.379 can't just brute force calculate your way to 00:03:45.379 --> 00:03:47.900 a win. No, absolutely not. The game has more 00:03:47.900 --> 00:03:50.120 possible board configurations than there are 00:03:50.120 --> 00:03:52.740 atoms in the observable universe. Which is just 00:03:52.740 --> 00:03:55.580 a crazy statistic. It is. And that's exactly 00:03:55.580 --> 00:03:58.060 why experts thought an AI beating a human at 00:03:58.060 --> 00:04:00.699 Go was easily a decade away. You need something 00:04:00.699 --> 00:04:03.740 akin to human intuition to evaluate a board state 00:04:03.740 --> 00:04:07.580 that complex. But then, in 2015, their program, 00:04:07.719 --> 00:04:11.449 AlphaGo, beat the European champion. And the 00:04:11.449 --> 00:04:14.310 following year, it beat the world champion, Lise 00:04:14.310 --> 00:04:18.230 et al. 4 to 1. It sent a massive shockwave through 00:04:18.230 --> 00:04:20.550 the global tech community. It really did. But 00:04:20.550 --> 00:04:22.870 the real paradigm shift wasn't beating Lise et 00:04:22.870 --> 00:04:25.350 al. It was what happened the following year with 00:04:25.350 --> 00:04:27.870 a program called AlphaGo Zero. Oh, this part 00:04:27.870 --> 00:04:30.389 is wild. Yeah. So the original AlphaGo learned 00:04:30.389 --> 00:04:33.410 by studying 30 million human moves. It learned 00:04:33.410 --> 00:04:36.930 from us. But AlphaGo Zero was given zero human 00:04:36.930 --> 00:04:39.680 data. Nothing. Nothing at all. It just played 00:04:39.680 --> 00:04:42.019 against itself millions of times. And within 00:04:42.019 --> 00:04:44.319 three days, it played the original AlphaGo, the 00:04:44.319 --> 00:04:46.420 one that beat the world champion, mind you, and 00:04:46.420 --> 00:04:50.180 defeated it 100 to 0. Wow. 100 to 0. That is 00:04:50.180 --> 00:04:52.160 the moment everything changed. Because when an 00:04:52.160 --> 00:04:55.379 AI learns from human data, it inherits our limitations, 00:04:55.639 --> 00:04:58.339 our blind spots basically. By generating its 00:04:58.339 --> 00:05:00.959 own data through self -play, AlphaGo Zero discovered 00:05:00.959 --> 00:05:03.120 entirely new strategies. It proved that human 00:05:03.120 --> 00:05:05.160 knowledge wasn't the baseline for artificial 00:05:05.160 --> 00:05:08.220 intelligence. Human knowledge was the bottleneck. 00:05:08.459 --> 00:05:11.240 And they took that exact same self -play concept 00:05:11.240 --> 00:05:13.939 and applied it to real -time strategy games too. 00:05:14.189 --> 00:05:17.790 Like, by 2019, a program called AlphaStar reached 00:05:17.790 --> 00:05:20.189 grand master level in the video game StarCraft 00:05:20.189 --> 00:05:23.250 2. Which is incredibly hard. Yeah, in StarCraft, 00:05:23.250 --> 00:05:25.110 you don't even see the whole board. You have 00:05:25.110 --> 00:05:27.610 hidden information, you have to manage economic 00:05:27.610 --> 00:05:30.470 resources, and you have to plan long -term strategies 00:05:30.470 --> 00:05:33.069 in real time while your opponent is actively 00:05:33.069 --> 00:05:35.250 attacking you. It requires immense strategic 00:05:35.250 --> 00:05:37.550 depth. But, you know, there is a crucial limitation 00:05:37.550 --> 00:05:39.850 here. What's that? Games like StarCraft or Go, 00:05:39.949 --> 00:05:42.689 no matter how incredibly complex they are, are 00:05:42.689 --> 00:05:46.110 still closed systems. They're digital sandboxes 00:05:46.110 --> 00:05:49.449 with perfect, unbreakable rules. If DeepMind 00:05:49.449 --> 00:05:51.970 wanted to build a truly general intelligence, 00:05:52.269 --> 00:05:54.430 they had to take this digital brain and apply 00:05:54.430 --> 00:05:57.689 it to the messy, unpredictable, open -loop system 00:05:57.689 --> 00:06:00.170 of the physical world. And that pivot from games 00:06:00.170 --> 00:06:02.209 to reality starts in a very practical place, 00:06:02.509 --> 00:06:05.050 Google's own data centers. These massive server 00:06:05.050 --> 00:06:07.790 farms generate an insane amount of heat, and 00:06:07.790 --> 00:06:10.009 keeping them cool takes a ridiculous amount of 00:06:10.009 --> 00:06:14.310 energy. So in 2016, DeepMind brings their reinforcement 00:06:14.310 --> 00:06:17.389 learning AI in to manage the cooling systems. 00:06:17.949 --> 00:06:19.889 Right. So the AI was given control over things 00:06:19.889 --> 00:06:23.470 like fans, cooling towers, and chillers. It started 00:06:23.470 --> 00:06:25.709 reading all the sensor data, like temperatures, 00:06:26.129 --> 00:06:29.490 pump speeds, power usage. OK. And because it 00:06:29.490 --> 00:06:31.769 wasn't burdened by human assumptions about how 00:06:31.769 --> 00:06:34.009 a building should be run, it started recommending 00:06:34.009 --> 00:06:37.629 actions that longtime human operators found completely 00:06:37.629 --> 00:06:40.889 unintuitive. Like, give me an example. So for 00:06:40.889 --> 00:06:43.670 example, it figured out how to aggressively exploit 00:06:43.670 --> 00:06:46.089 winter conditions. Oh, interesting. Yeah. It 00:06:46.089 --> 00:06:48.110 realized that if it dynamically adjusted the 00:06:48.110 --> 00:06:50.509 system to draw in specific amounts of outside 00:06:50.509 --> 00:06:53.449 cold air at exact times, it could produce colder 00:06:53.449 --> 00:06:55.889 than normal water for cooling. And that allowed 00:06:55.889 --> 00:06:58.490 it to shut down other power -hungry systems entirely. 00:06:58.730 --> 00:07:01.889 Wow. It recognized complex thermodynamic patterns 00:07:01.889 --> 00:07:04.689 that human engineers simply couldn't see. The 00:07:04.689 --> 00:07:06.350 human engineers were probably looking at the 00:07:06.350 --> 00:07:09.110 screens terrified, but they followed the AI's 00:07:09.110 --> 00:07:11.350 recommendations and it ultimately saved Google 00:07:11.350 --> 00:07:14.930 30 % on energy used for cooling. Which is massive 00:07:14.930 --> 00:07:17.290 at that scale. Yeah. Every time you run a Google 00:07:17.290 --> 00:07:20.089 search today, it requires less electricity because 00:07:20.089 --> 00:07:22.230 an AI treated a building like a giant puzzle. 00:07:23.149 --> 00:07:25.930 But taking over a thermostat is one thing. Solving 00:07:25.930 --> 00:07:28.490 a 50 -year -old biological mystery is another. 00:07:29.389 --> 00:07:32.329 We need to talk about Alpha Fold. Yes, alphafold. 00:07:32.569 --> 00:07:35.350 This is really their crown jewel. For decades, 00:07:35.610 --> 00:07:38.649 biology faced this grand challenge, which was... 00:07:38.170 --> 00:07:41.410 protein folding. Proteins are the building blocks 00:07:41.410 --> 00:07:43.610 of life. They start as a long string of amino 00:07:43.610 --> 00:07:46.589 acids, but then they crumple and fold into highly 00:07:46.589 --> 00:07:49.610 complex 3D shapes. And the specific shape of 00:07:49.610 --> 00:07:51.870 a protein dictates exactly what it does in the 00:07:51.870 --> 00:07:54.569 human body, whether it causes a disease or cures 00:07:54.569 --> 00:07:56.790 one. OK, but here's where I struggle, though. 00:07:56.810 --> 00:07:59.790 I understand how an AI can predict a chess move 00:07:59.790 --> 00:08:02.730 or figure out space invaders, but how does playing 00:08:02.730 --> 00:08:05.329 a game against yourself translate to predicting 00:08:05.329 --> 00:08:08.670 how a microscopic protein physically fold inside 00:08:08.670 --> 00:08:11.129 a biological cell. If we connect this to the 00:08:11.129 --> 00:08:13.870 bigger picture, think of it like a microscopic 00:08:13.870 --> 00:08:17.069 hyper complex piece of origami. Okay. If you 00:08:17.069 --> 00:08:19.550 fold the paper wrong, the cell gets a disease. 00:08:20.189 --> 00:08:23.149 If you fold it right, you create a cure. The 00:08:23.149 --> 00:08:26.170 problem is, a single protein can fold in an almost 00:08:26.170 --> 00:08:29.170 infinite number of ways. It would take longer 00:08:29.170 --> 00:08:31.329 than the age of the universe to test every possible 00:08:31.329 --> 00:08:34.190 shape manually. Wow. But at a foundational level, 00:08:34.529 --> 00:08:37.210 nature operates on rules, the laws of physics 00:08:37.210 --> 00:08:40.590 and chemistry. So DeepMind basically treated 00:08:40.590 --> 00:08:43.409 protein folding as a spatial optimization game. 00:08:43.789 --> 00:08:45.889 So instead of trying to maximize a high score 00:08:45.889 --> 00:08:48.590 in a video game, the AI is trying to find the 00:08:48.590 --> 00:08:51.870 most stable, energy efficient 3D structure. based 00:08:51.870 --> 00:08:54.250 on the laws of physics. Precisely. It looks at 00:08:54.250 --> 00:08:56.750 the sequence of amino acids and predicts the 00:08:56.750 --> 00:08:59.009 distance and angles between every single pair 00:08:59.009 --> 00:09:01.169 of them. And it actually works. It absolutely 00:09:01.169 --> 00:09:04.789 worked. By July 2022, AlphaFold had predicted 00:09:04.789 --> 00:09:07.289 the structures of over 200 million proteins. 00:09:07.470 --> 00:09:09.070 That's practically all of them, right? That is 00:09:09.070 --> 00:09:11.309 virtually every protein known to science, yes. 00:09:11.649 --> 00:09:14.049 Which fundamentally changed biological research 00:09:14.049 --> 00:09:16.970 overnight. I mean, Dimas Hasabas and John Jumper 00:09:16.970 --> 00:09:19.929 literally won the 2024 Nobel Prize in Chemistry 00:09:19.929 --> 00:09:22.289 for this. Yeah. And according to the source material, 00:09:22.629 --> 00:09:25.450 they just kept pushing. In 2024, they released 00:09:25.450 --> 00:09:28.409 AlphaFold 3, which expanded the AI to predict 00:09:28.409 --> 00:09:31.980 how proteins interact with DNA and RNA. It bumped 00:09:31.980 --> 00:09:34.799 the accuracy on DNA interaction tests from a 00:09:34.799 --> 00:09:38.559 baseline of 28 % all the way up to 65%. And then 00:09:38.559 --> 00:09:41.039 they took that same underlying philosophy -finding 00:09:41.039 --> 00:09:43.679 patterns and massive chaotic data sets and applied 00:09:43.679 --> 00:09:46.740 it to the atmosphere. Right. In mid -2025, DeepMind 00:09:46.740 --> 00:09:49.220 launched WeatherLab. Now, weather forecasting 00:09:49.220 --> 00:09:52.600 is notoriously difficult. How does an AI approach 00:09:52.600 --> 00:09:55.440 a hurricane differently than, say, the meteorologists 00:09:55.440 --> 00:09:58.009 we see on the local news? Well, traditional forecasting 00:09:58.009 --> 00:10:01.149 relies on rigid, physics -based models, supercomputers 00:10:01.149 --> 00:10:03.409 crunching massive fluid dynamics equations. Right, 00:10:03.429 --> 00:10:05.710 lots of math. Right, but DeepMind took a different 00:10:05.710 --> 00:10:07.769 route. They train what are called stochastic 00:10:07.769 --> 00:10:10.470 neural networks on 45 years of global weather 00:10:10.470 --> 00:10:13.669 data. Stochastic just means it embraces randomness 00:10:13.669 --> 00:10:16.549 and probability. Instead of relying purely on 00:10:16.549 --> 00:10:19.590 rigid equations, the AI looks at decades of chaotic 00:10:19.590 --> 00:10:22.929 historical weather patterns and learns the probabilistic 00:10:22.929 --> 00:10:25.710 rules of the atmosphere. So it's pattern recognition 00:10:25.710 --> 00:10:29.029 on a planetary scale. And the dossier notes that 00:10:29.029 --> 00:10:32.110 during the 2025 Atlantic hurricane season, this 00:10:32.110 --> 00:10:35.129 weather lab AI actually outperformed the U .S. 00:10:35.350 --> 00:10:37.409 National Weather Service's traditional models. 00:10:37.490 --> 00:10:39.950 It did. It was predicting the formation and tracks 00:10:39.950 --> 00:10:43.210 of hurricanes up to 15 days in advance. Giving 00:10:43.210 --> 00:10:45.710 cities an extra week to prepare for a disaster 00:10:45.710 --> 00:10:48.429 is just a monumental leap. It really is. But, 00:10:48.450 --> 00:10:50.370 you know, predicting the physical world, whether 00:10:50.370 --> 00:10:52.850 it's weather or proteins, is fundamentally an 00:10:52.850 --> 00:10:56.019 analytical task. By 2023, the entire tech industry 00:10:56.019 --> 00:10:58.120 was obsessing over a completely different kind 00:10:58.120 --> 00:11:00.740 of AI, generative AI. The ability to create things 00:11:00.740 --> 00:11:04.639 from scratch. Right. Open AI drops chat GPT and 00:11:04.639 --> 00:11:07.299 the world goes crazy. Google obviously had to 00:11:07.299 --> 00:11:11.320 respond. So in April 2023, DeepMind merges with 00:11:11.320 --> 00:11:14.419 Google Brain to form Google DeepMind. Their mandate 00:11:14.419 --> 00:11:17.299 was clear tackle human language, reasoning, and 00:11:17.299 --> 00:11:20.120 creativity. And this merger marks their entry 00:11:20.120 --> 00:11:23.100 into the generative arms race. They became the 00:11:23.100 --> 00:11:27.500 engine behind Google's consumer -facing AI, specifically 00:11:27.500 --> 00:11:30.259 the Gemini models. And the evolution here happens 00:11:30.259 --> 00:11:33.960 at a blistering pace. Like, by March 2025, they 00:11:33.960 --> 00:11:36.899 released Gemini 2 .5. And the key innovation 00:11:36.899 --> 00:11:39.080 here wasn't just generating text, right? No. 00:11:39.159 --> 00:11:41.500 It was the introduction of a feature where the 00:11:41.500 --> 00:11:45.070 AI actually pauses to think. before responding. 00:11:45.370 --> 00:11:47.929 Right. It simulates human reasoning. Instead 00:11:47.929 --> 00:11:50.309 of just spitting out the most statistically likely 00:11:50.309 --> 00:11:53.389 next word, it takes a beat, explores different 00:11:53.389 --> 00:11:56.769 logical paths, verifies its own work, and then 00:11:56.769 --> 00:11:58.509 gives you the answer. Exactly. And they followed 00:11:58.509 --> 00:12:01.669 that up in November 2025 with Gemini 3 Pro, which 00:12:01.669 --> 00:12:04.490 is a fully multimodal reasoning model deeply 00:12:04.490 --> 00:12:06.490 integrated into Google search. But they didn't 00:12:06.490 --> 00:12:08.730 just hoard this technology for themselves. They 00:12:08.730 --> 00:12:11.470 also released the Gemma series. Yes. The Gemma 00:12:11.470 --> 00:12:13.700 models are what we call open weight models. So 00:12:13.700 --> 00:12:15.919 a foundation model is the massive underlying 00:12:15.919 --> 00:12:18.659 brain trained on vast amounts of data. Usually 00:12:18.659 --> 00:12:21.220 companies lock these behind APIs. Like, you can 00:12:21.220 --> 00:12:23.179 talk to it, but you can't see the engine. Right. 00:12:23.820 --> 00:12:26.700 But Openweight means DeepMind actually released 00:12:26.700 --> 00:12:29.320 the core architecture and parameters to the public, 00:12:29.820 --> 00:12:32.639 allowing developers to run and modify the AI 00:12:32.639 --> 00:12:35.480 on their own hardware. Which leads to easily 00:12:35.480 --> 00:12:38.720 the most fascinating project in the entire dossier, 00:12:39.000 --> 00:12:41.419 in my opinion, Dolphin Gemma, released in April 00:12:41.419 --> 00:12:44.679 2025. Oh, yeah. This was a highly specialized 00:12:44.679 --> 00:12:47.240 attempt to decode dolphin communication. It's 00:12:47.240 --> 00:12:50.220 just wild. We spend decades aiming satellite 00:12:50.220 --> 00:12:52.799 dishes at outer space looking for alien intelligence. 00:12:53.519 --> 00:12:56.240 And DeepMind is using AI to talk to the highly 00:12:56.240 --> 00:12:58.799 intelligent aliens swimming in our own oceans. 00:12:59.120 --> 00:13:01.019 That's a great analogy. But how do you even begin 00:13:01.019 --> 00:13:03.259 to map a language when you have no Rosetta Stone? 00:13:03.519 --> 00:13:06.480 Well, you treat the audio like an unknown linguistic 00:13:06.480 --> 00:13:09.299 structure. The AI analyzes thousands of hours 00:13:09.299 --> 00:13:11.250 of dolphin clicks and whistles. I don't know 00:13:11.250 --> 00:13:13.389 what the words mean, obviously, but it looks 00:13:13.389 --> 00:13:16.029 for statistical patterns in the noise. It learns 00:13:16.029 --> 00:13:18.169 the syntax. It figures out that when dolphin 00:13:18.169 --> 00:13:21.009 A makes this sequence of clicks, dolphin B responds 00:13:21.009 --> 00:13:23.570 with that sequence. By mapping these structural 00:13:23.570 --> 00:13:25.929 relationships, the foundation model can potentially 00:13:25.929 --> 00:13:29.070 generate novel, contextually accurate dolphin 00:13:29.070 --> 00:13:31.809 -like sound sequences. It's structural linguistics 00:13:31.809 --> 00:13:35.549 driven by raw computing power. Exactly. And that 00:13:35.549 --> 00:13:37.950 computing power extends to human media, too. 00:13:38.039 --> 00:13:41.279 The source details an explosion in audio and 00:13:41.279 --> 00:13:44.399 video generation. In May 2025, they launched 00:13:44.399 --> 00:13:47.559 VO3. This isn't just generating a silent video 00:13:47.559 --> 00:13:50.139 from a text prompt. No, it generates video with 00:13:50.139 --> 00:13:52.799 perfectly synchronized audio simultaneously. 00:13:52.940 --> 00:13:55.139 Yeah, you type in a prompt and it hallucinates 00:13:55.139 --> 00:13:57.240 the visuals, the dialogue, the sound effects 00:13:57.240 --> 00:14:00.179 and the ambient noise all at once, mathematically 00:14:00.179 --> 00:14:02.980 perfectly synced. And they did the same for Complex 00:14:02.980 --> 00:14:06.419 Music Generation with Liria 3 Pro in March 2026. 00:14:06.639 --> 00:14:08.740 But the absolute peak of this generative era 00:14:08.740 --> 00:14:11.559 has to be Project Genie, which became available 00:14:11.559 --> 00:14:14.639 to premium subscribers in January 2026. Yes. 00:14:14.779 --> 00:14:16.720 Project Genie sounds like science fiction. You 00:14:16.720 --> 00:14:19.340 give it a single 2D image or just a text prompt, 00:14:19.679 --> 00:14:22.419 and it generates an entire playable interactive 00:14:22.419 --> 00:14:25.340 3D virtual environment. How is that computationally 00:14:25.340 --> 00:14:27.860 possible in real time? It's essentially hallucinating 00:14:27.860 --> 00:14:30.429 a physics engine on the fly. When you interact 00:14:30.429 --> 00:14:32.750 with the environment, say, you command your character 00:14:32.750 --> 00:14:35.909 to jump, the AI predicts what the next frame 00:14:35.909 --> 00:14:37.809 of that world should look like based on its deep 00:14:37.809 --> 00:14:40.789 understanding of spatial dynamics. It is generating 00:14:40.789 --> 00:14:43.389 the rules of a world in real time as you move 00:14:43.389 --> 00:14:45.450 through it. We're talking about an AI creating 00:14:45.450 --> 00:14:50.269 entire realities from scratch. But what happens 00:14:50.269 --> 00:14:53.490 when you use this technology to mediate the reality 00:14:53.490 --> 00:14:56.129 we actually live in? Ah, right. There's a 2024 00:14:56.129 --> 00:14:58.049 experiment mentioned in the source called the 00:14:58.049 --> 00:15:01.600 Habermas machine. DeepMind brought together groups 00:15:01.600 --> 00:15:04.500 of people with highly polarized, differing views 00:15:04.500 --> 00:15:07.559 to debate a topic. Yeah. And they used AI to 00:15:07.559 --> 00:15:09.480 mediate the discussion and try to find common 00:15:09.480 --> 00:15:11.639 ground. And the results were really striking. 00:15:12.000 --> 00:15:13.980 Participants actually rated the AI summaries 00:15:13.980 --> 00:15:16.179 and proposed compromises higher than a human 00:15:16.179 --> 00:15:19.299 moderator's 56 % of the time. Wait, really? It 00:15:19.299 --> 00:15:22.019 was better than a human? Yes. It's taking the 00:15:22.019 --> 00:15:24.080 temperature of the room and finding the mathematical 00:15:24.080 --> 00:15:28.019 center of gravity in a human debate. An AI, completely 00:15:28.019 --> 00:15:31.080 free from human ego or emotional bias, was just 00:15:31.080 --> 00:15:33.159 better at identifying our shared values than 00:15:33.159 --> 00:15:36.840 we were. It is a profound proof of concept. But 00:15:36.840 --> 00:15:38.759 pulling this off in the real world hasn't been 00:15:38.759 --> 00:15:42.039 without serious friction. Moving fast and breaking 00:15:42.039 --> 00:15:44.240 things works perfectly when you're playing Space 00:15:44.240 --> 00:15:47.019 Invaders. Right. When you enter healthcare, academia, 00:15:47.340 --> 00:15:50.039 and critical infrastructure, the guardrails are 00:15:50.039 --> 00:15:52.320 very different. This raises an important question 00:15:52.320 --> 00:15:55.940 about how society audits AI. Because the source 00:15:55.940 --> 00:15:59.039 material highlights several major controversies 00:15:59.039 --> 00:16:02.220 where DeepMind's ambitions slammed right into 00:16:02.220 --> 00:16:03.779 those guardrails. Right, we have to look at the 00:16:03.779 --> 00:16:06.440 facts here. Take the NHS data controversy back 00:16:06.440 --> 00:16:09.460 in 2016. Yes. DeepMind Health developed an app 00:16:09.460 --> 00:16:13.039 called Streams. The goal was fantastic, right? 00:16:13.360 --> 00:16:15.720 Yeah. Alert doctors to acute kidney injury early 00:16:15.720 --> 00:16:18.639 to save patients' lives. But to train the system, 00:16:18.779 --> 00:16:21.200 they gained access to the healthcare data of 00:16:21.200 --> 00:16:24.779 1 .6 million UK patients. The central issue there 00:16:24.779 --> 00:16:27.720 was consent. The UK's Information Commissioner's 00:16:27.720 --> 00:16:30.000 Office investigated the arrangement and ruled 00:16:30.000 --> 00:16:32.139 that the hospital trust involved had failed to 00:16:32.139 --> 00:16:34.120 comply with the Data Protection Act. Because 00:16:34.120 --> 00:16:36.559 the patients simply were not adequately informed 00:16:36.559 --> 00:16:38.480 that their medical data was being handed over 00:16:38.480 --> 00:16:40.940 to a tech company for app development. Exactly. 00:16:41.440 --> 00:16:43.299 Furthermore, DeepMind had initially promised 00:16:43.299 --> 00:16:45.519 that this patient data would be kept strictly 00:16:45.519 --> 00:16:48.789 separate from Google accounts. But later on, 00:16:49.149 --> 00:16:51.490 Google just absorbed the health division anyway. 00:16:52.309 --> 00:16:53.929 Privacy advocates pointed out that this looked 00:16:53.929 --> 00:16:56.330 like a blatant betrayal of the initial promises 00:16:56.330 --> 00:16:58.690 made to the public. Yeah. And you see a similar 00:16:58.690 --> 00:17:00.850 friction within the scientific community itself, 00:17:01.029 --> 00:17:03.750 where DeepMind's corporate PR often clashes with 00:17:03.750 --> 00:17:07.089 the demands of academic rigor. Consider the alpha 00:17:07.089 --> 00:17:09.789 chip debate. Oh, this one is fascinating. Yeah. 00:17:09.950 --> 00:17:12.369 DeepMind claimed they were using reinforcement 00:17:12.369 --> 00:17:15.210 learning to design computer chips. Imagine a 00:17:15.210 --> 00:17:17.690 city planner trying to fit millions of buildings 00:17:17.690 --> 00:17:21.009 into a microscopic grid. Right. The AI treats 00:17:21.009 --> 00:17:23.589 that silicon layout like a game board. placing 00:17:23.589 --> 00:17:25.890 components to minimize wire length and energy 00:17:25.890 --> 00:17:29.130 use. DeepMind claimed this reduced chip design 00:17:29.130 --> 00:17:32.309 time from weeks to mere hours, and that these 00:17:32.309 --> 00:17:35.049 AI designs were actively being used in Google's 00:17:35.049 --> 00:17:37.390 own hardware. Which sounds like a revolution, 00:17:37.730 --> 00:17:39.730 but independent experts and publications like 00:17:39.730 --> 00:17:42.130 Communications of the ACM pushed back heavily 00:17:42.130 --> 00:17:44.730 on this. The criticism was that DeepMind failed 00:17:44.730 --> 00:17:47.509 to provide transparent, independent benchmarks. 00:17:48.299 --> 00:17:51.039 In the scientific community, if you claim a superhuman 00:17:51.039 --> 00:17:53.339 breakthrough, you have to share the underlying 00:17:53.339 --> 00:17:56.559 comparative data so other researchers can independently 00:17:56.559 --> 00:17:59.160 verify it. Right. You can't just say, trust us, 00:17:59.160 --> 00:18:02.240 it's faster. Exactly. And we saw the exact same 00:18:02.240 --> 00:18:05.819 skepticism with Genome, their materials science 00:18:05.819 --> 00:18:08.960 AI. DeepMind announced this tool had discovered 00:18:08.960 --> 00:18:12.119 millions of new crystalline structures. They 00:18:12.119 --> 00:18:14.640 essentially claimed to have revolutionized material 00:18:14.640 --> 00:18:17.400 science overnight. But then, researchers like 00:18:17.400 --> 00:18:19.400 Ant... Anthony Cheatham published reviews stating 00:18:19.400 --> 00:18:22.319 the tool failed to make a useful, practical contribution. 00:18:22.920 --> 00:18:25.279 Right. When human scientists actually dug into 00:18:25.279 --> 00:18:27.279 the data, they found that the vast majority of 00:18:27.279 --> 00:18:30.200 these new materials were just minor, highly predictable 00:18:30.200 --> 00:18:32.200 variants of structures we already knew about. 00:18:32.299 --> 00:18:35.259 They weren't practical breakthroughs. It highlights 00:18:35.259 --> 00:18:38.039 the inherent tension of this era, a corporation 00:18:38.039 --> 00:18:41.119 eager to announce world -changing abilities versus 00:18:41.119 --> 00:18:43.740 a scientific community demanding rigorous, peer 00:18:43.740 --> 00:18:46.420 -reviewed proof. To their credit, the dossier 00:18:46.420 --> 00:18:48.700 notes DeepMind is attempting to self -regulate 00:18:48.700 --> 00:18:51.940 as these systems get more powerful. Like, they 00:18:51.940 --> 00:18:54.819 formed a DeepMind ethics and society unit. And 00:18:54.819 --> 00:18:58.299 in 2024, they introduced a robot constitution 00:18:58.299 --> 00:19:01.200 for their AI products. Which is heavily inspired 00:19:01.200 --> 00:19:03.680 by Isaac Asimov's science fiction laws of robotics. 00:19:04.299 --> 00:19:07.740 Rule number one, a robot may not injure a human 00:19:07.740 --> 00:19:10.059 being. So what does this all mean for the listener? 00:19:10.640 --> 00:19:13.140 We have an AI that is smart enough to win a Nobel 00:19:13.140 --> 00:19:16.880 Prize, but still needs a literal robot constitution 00:19:16.880 --> 00:19:19.309 so it behaves. It sounds a bit theatrical, I 00:19:19.309 --> 00:19:21.589 know, but you realize why it's necessary when 00:19:21.589 --> 00:19:23.529 you see that models like Gemini Robotics are 00:19:23.529 --> 00:19:26.430 now being deployed to control actual physical 00:19:26.430 --> 00:19:29.190 robotic arms in the real world. Oh, wow. You 00:19:29.190 --> 00:19:31.349 need hard -coded behavioral guardrails when the 00:19:31.349 --> 00:19:33.089 AI can physically interact with its environment. 00:19:33.509 --> 00:19:35.670 It means intelligence is scaling faster than 00:19:35.670 --> 00:19:38.029 our traditional societal frameworks can comfortably 00:19:38.029 --> 00:19:40.190 accommodate. Exactly. Which brings us full circle. 00:19:40.460 --> 00:19:42.759 DeepMind is no longer just a bunch of researchers 00:19:42.759 --> 00:19:45.240 in a London lab trying to teach a computer to 00:19:45.240 --> 00:19:47.920 play a vintage arcade game. It has become the 00:19:47.920 --> 00:19:50.240 invisible infrastructure of your daily life. 00:19:50.380 --> 00:19:53.160 It really has. Think about it. When you pull 00:19:53.160 --> 00:19:56.960 out your phone to check a 15 -day hurricane forecast. 00:19:57.180 --> 00:19:59.059 Wait, no, let me rephrase that. When you pull 00:19:59.059 --> 00:20:02.079 out your phone to check that forecast, that is 00:20:02.079 --> 00:20:04.259 DeepMind's weather lab running the probabilities. 00:20:04.380 --> 00:20:06.180 And when you watch a video on YouTube and it 00:20:06.180 --> 00:20:09.720 loads instantly on a cellular connection, That's 00:20:09.720 --> 00:20:13.279 DeepMind's MuZero algorithm, which quietly optimize 00:20:13.279 --> 00:20:16.140 the compression to reduce video bitrates across 00:20:16.140 --> 00:20:19.720 the entire platform by 6 .28%. Even the battery 00:20:19.720 --> 00:20:22.099 inside your Android phone? Yes. If you have adaptive 00:20:22.099 --> 00:20:24.539 battery turned on, it is using reinforcement 00:20:24.539 --> 00:20:27.339 learning to study your habits, predicting which 00:20:27.339 --> 00:20:29.599 apps you'll open, and routing power accordingly. 00:20:30.220 --> 00:20:32.440 DeepMind is actively optimizing your reality 00:20:32.440 --> 00:20:34.750 right now. literally in your pocket. Which leads 00:20:34.750 --> 00:20:36.849 to one final concept from the source material 00:20:36.849 --> 00:20:38.769 that is just, well, it's worth lingering on. 00:20:38.809 --> 00:20:41.630 Oh, yeah. In May 2025, DeepMind unveiled something 00:20:41.630 --> 00:20:43.710 called AlphaVolve. This is the one that really 00:20:43.710 --> 00:20:46.309 gets me. So AlphaVolve is an evolutionary coding 00:20:46.309 --> 00:20:49.269 agent. It uses advanced language models to design, 00:20:49.390 --> 00:20:52.109 test, and optimize its own algorithm. Wait, its 00:20:52.109 --> 00:20:55.609 own? Yes. It writes a piece of code, tests its 00:20:55.609 --> 00:20:58.089 efficiency, mutates the code to see if it improves, 00:20:58.430 --> 00:21:01.029 and selects the best version to iterate on. It 00:21:01.029 --> 00:21:03.509 is an AI that has already matched or beaten state 00:21:03.509 --> 00:21:06.309 -of -the -art human algorithms in dozens of complex 00:21:06.309 --> 00:21:08.630 mathematical problems. Look at the progression 00:21:08.630 --> 00:21:11.289 here. DeepMind started by teaching AI to play 00:21:11.289 --> 00:21:14.549 video games. Then they taught AI to solve biology 00:21:14.549 --> 00:21:17.170 and physics. And now, with AlphaVolve, they're 00:21:17.170 --> 00:21:20.680 teaching AI to write better AI. The loop is closing. 00:21:21.019 --> 00:21:23.220 It leaves us with a profound question for you 00:21:23.220 --> 00:21:26.200 to mull over. If human developers are no longer 00:21:26.200 --> 00:21:28.720 the primary bottleneck for creating and refining 00:21:28.720 --> 00:21:31.460 intelligence, what do the next five years look 00:21:31.460 --> 00:21:34.279 like when the AI is the one driving its own evolution? 00:21:34.559 --> 00:21:37.180 That is the ultimate question. From the jagged 00:21:37.180 --> 00:21:40.640 pixels of a 1970s TV screen to an intelligence 00:21:40.640 --> 00:21:43.460 that is rewriting its own source code. Keep questioning, 00:21:43.599 --> 00:21:46.059 keep learning, and keep an eye on how these invisible 00:21:46.059 --> 00:21:48.299 systems are shaping your world. We'll catch you 00:21:48.299 --> 00:21:49.259 on the next deep dive.