WEBVTT 00:00:00.000 --> 00:00:03.040 Right now, an algorithm that was originally designed 00:00:03.040 --> 00:00:07.700 to play 1970s Atari games is predicting the path 00:00:07.700 --> 00:00:10.279 of the hurricane outside your window. Yeah, it 00:00:10.279 --> 00:00:13.740 really is. It is quietly optimizing the thermal 00:00:13.740 --> 00:00:16.600 cooling of the data center, storing your photos, 00:00:17.219 --> 00:00:19.579 and it's mapping the three -dimensional structures 00:00:19.579 --> 00:00:22.120 of the microscopic proteins inside your own body. 00:00:22.429 --> 00:00:25.250 It's just a complete paradigm shift. I mean, 00:00:25.390 --> 00:00:27.370 when you look at the systems built by Google 00:00:27.370 --> 00:00:29.410 DeepMind, you aren't just looking at a smarter 00:00:29.410 --> 00:00:32.750 search bar or like a text generator. You are 00:00:32.750 --> 00:00:35.070 looking at a fundamentally different way of engineering 00:00:35.070 --> 00:00:38.049 intelligence. Exactly. And welcome to this deep 00:00:38.049 --> 00:00:40.390 dive into the comprehensive history of Google 00:00:40.390 --> 00:00:43.500 DeepMind. We are pulling from a massive stack 00:00:43.500 --> 00:00:46.439 of source material today to trace how an AI lab 00:00:46.439 --> 00:00:48.840 in the UK went from bouncing a pixelated ball 00:00:48.840 --> 00:00:51.619 and breakout to literally winning a Nobel Prize 00:00:51.619 --> 00:00:54.240 in chemistry. And generating real -time virtual 00:00:54.240 --> 00:00:57.219 world. Yes. But to understand how any of this 00:00:57.219 --> 00:00:59.460 is actually possible, we have to look at their 00:00:59.460 --> 00:01:01.460 foundational philosophy. Because they didn't 00:01:01.460 --> 00:01:03.420 start by trying to teach a machine, you know, 00:01:03.740 --> 00:01:05.400 all the rules of the universe. No, they didn't. 00:01:05.480 --> 00:01:07.599 They just taught it how to learn. What's fascinating 00:01:07.599 --> 00:01:09.959 here is that if you go back to their founding, 00:01:10.200 --> 00:01:13.719 right, in 2010. By Demis Hassabis, Shane Legge, 00:01:13.920 --> 00:01:16.799 and Mustafa Suleiman. Right, exactly. The prevailing 00:01:16.799 --> 00:01:19.980 AI logic at that time was highly prescriptive. 00:01:20.200 --> 00:01:22.299 Like, if you wanted a machine to play chess, 00:01:22.739 --> 00:01:25.019 you programmed in the value of every single piece, 00:01:25.620 --> 00:01:28.239 the exact geometry of the board, specific opening 00:01:28.239 --> 00:01:30.739 strategies. You essentially gave it a heavy top 00:01:30.739 --> 00:01:32.840 -down instruction manual. Yeah, exactly. But 00:01:32.840 --> 00:01:35.200 DeepMind completely discarded the manual. They 00:01:35.200 --> 00:01:37.260 utilized reinforcement learning instead. And 00:01:37.260 --> 00:01:40.579 their laboratory for this wasn't... some supercomputer 00:01:40.579 --> 00:01:44.099 simulating global economics. It was retro video 00:01:44.099 --> 00:01:48.299 games, Pong, Space Invaders, Breakout. The classics. 00:01:48.500 --> 00:01:50.700 Right, they would feed the AI raw pixels, just 00:01:50.700 --> 00:01:52.819 the changing colors on the screen. The AI had 00:01:52.819 --> 00:01:55.239 no concept of what a paddle was or that a laser 00:01:55.239 --> 00:01:57.400 destroys an alien. Right, it had no context at 00:01:57.400 --> 00:01:59.959 all. None. It was just given a controller and 00:01:59.959 --> 00:02:02.140 a single mathematical objective, which was basically 00:02:02.140 --> 00:02:04.319 just make the score integer go up. Which forces 00:02:04.319 --> 00:02:06.879 the system to build its own internal representations 00:02:06.879 --> 00:02:09.680 of the world. Because at first, the AI just flails 00:02:09.680 --> 00:02:12.219 around. It's just smashing buttons. Literally. 00:02:12.439 --> 00:02:15.199 It outputs random controller movements. But eventually, 00:02:15.360 --> 00:02:17.560 a random movement causes the paddle to hit the 00:02:17.560 --> 00:02:20.080 ball and the score goes up. So the reinforcement 00:02:20.080 --> 00:02:22.539 learning algorithm registers a positive reward 00:02:22.539 --> 00:02:25.240 signal. Oh, wow. Yeah. And it mathematically 00:02:25.240 --> 00:02:28.020 updates its own neural pathways to say, well, 00:02:28.280 --> 00:02:30.680 Whatever sequence of pixel states led to that 00:02:30.680 --> 00:02:33.659 action, do that again. OK, let's unpack this. 00:02:34.240 --> 00:02:36.180 Because it's like handing a toddler a joystick, 00:02:36.960 --> 00:02:39.560 absolutely refusing to explain the rules, leaving 00:02:39.560 --> 00:02:41.819 the room, and then returning three days later 00:02:41.819 --> 00:02:43.819 to find they are the undisputed world champion. 00:02:44.000 --> 00:02:46.099 That's a great way to put it. It just grinds 00:02:46.099 --> 00:02:49.259 through millions of iterations, failing, adjusting, 00:02:49.639 --> 00:02:51.699 failing again. But I look at them starting with 00:02:51.699 --> 00:02:54.639 Space Invaders, and it just feels like, I don't 00:02:54.639 --> 00:02:57.030 know, a massive waste of computing power. You 00:02:57.030 --> 00:02:59.430 mean, why not start with something real? Yeah. 00:02:59.889 --> 00:03:02.689 If the ultimate goal is artificial general intelligence 00:03:02.689 --> 00:03:06.490 solving real human problems, why hide in a virtual 00:03:06.490 --> 00:03:10.310 sandbox? Why not tackle real world data immediately? 00:03:10.590 --> 00:03:13.050 Well, because physical reality is impossibly 00:03:13.050 --> 00:03:16.389 noisy. I mean, it is ambiguous. Physics are completely 00:03:16.389 --> 00:03:19.169 unpredictable and success is incredibly hard 00:03:19.169 --> 00:03:21.229 to define in a mathematical equation. Right. 00:03:21.250 --> 00:03:23.750 That makes sense. A video game, however, offers 00:03:23.750 --> 00:03:27.819 a pristine state space. It is a perfectly contained 00:03:27.819 --> 00:03:30.599 universe with a crystal -clear wind state. Which 00:03:30.599 --> 00:03:32.800 lets them isolate the learning part. Exactly. 00:03:32.860 --> 00:03:35.639 It proved their core thesis, which was that complex 00:03:35.639 --> 00:03:37.960 cognitive processes like long -term planning 00:03:37.960 --> 00:03:41.120 and pattern recognition can be synthesized entirely 00:03:41.120 --> 00:03:43.699 from scratch through self -correction. And they 00:03:43.699 --> 00:03:45.860 really pushed that thesis to its absolute limit 00:03:45.860 --> 00:03:49.300 with Go. They moved from the strict 2D parameters 00:03:49.300 --> 00:03:52.539 of Atari to the near -infinite game tree complexity 00:03:52.539 --> 00:03:55.599 of the Go board. Yes. In 2016, their Alpha Go 00:03:55.500 --> 00:03:59.020 system beat Lee Saul, who is a nine -dan professional 00:03:59.020 --> 00:04:02.039 human champion, four to one. A huge moment. But 00:04:02.039 --> 00:04:03.759 the detail in the sources that really anchors 00:04:03.759 --> 00:04:06.259 this whole concept is what happened next, AlphaGo 00:04:06.259 --> 00:04:08.419 Zero. Right, because the original AlphaGo was 00:04:08.419 --> 00:04:10.979 initially trained on millions of human matches 00:04:10.979 --> 00:04:14.120 to get like a baseline understanding of strategy. 00:04:14.439 --> 00:04:18.120 But AlphaGo Zero used zero human data. It was 00:04:18.120 --> 00:04:20.220 just given the rules of Go and simply played 00:04:20.220 --> 00:04:23.220 against itself. And within three days, it beat 00:04:23.220 --> 00:04:26.800 the original AlphaGo 100 to 0. It's unbelievable. 00:04:27.120 --> 00:04:29.439 It didn't just learn the game. It discovered 00:04:29.439 --> 00:04:32.519 strategies that human masters had never even 00:04:32.519 --> 00:04:35.500 conceived of in 3 ,000 years of playing the game. 00:04:35.560 --> 00:04:38.069 And that is the inflection point. By mastering 00:04:38.069 --> 00:04:40.050 these environments purely through self -play 00:04:40.050 --> 00:04:42.350 and reinforcement learning, they weren't just 00:04:42.350 --> 00:04:45.470 building like a board game savant. They had built 00:04:45.470 --> 00:04:48.670 a generalized optimization engine. So the natural 00:04:48.670 --> 00:04:50.810 engineering question then became, well, if we 00:04:50.810 --> 00:04:53.370 can map the perfect virtual state of a game board, 00:04:53.870 --> 00:04:56.850 can we apply this exact same learning mechanism 00:04:56.850 --> 00:05:00.029 to the messy physical infrastructure of the real 00:05:00.029 --> 00:05:01.829 world? Here's where it gets really interesting 00:05:01.829 --> 00:05:03.579 for you listening. because this brings us to 00:05:03.579 --> 00:05:06.199 Google's own data centers. These are massive, 00:05:06.639 --> 00:05:09.360 multi -million dollar server farms that generate 00:05:09.360 --> 00:05:12.379 an unbelievable amount of heat. Oh, yeah. And 00:05:12.379 --> 00:05:14.720 in 2014, a Google engineer essentially looked 00:05:14.720 --> 00:05:17.120 at AlphaGo and said, hey, can we use this to 00:05:17.120 --> 00:05:19.560 play the cooling systems like a video game? Just 00:05:19.560 --> 00:05:21.240 treat the whole building like Pong. Exactly. 00:05:21.379 --> 00:05:24.970 They brought DeepMind in, fed the AI. all the 00:05:24.970 --> 00:05:28.009 historical sensor data temperatures, pump speeds, 00:05:28.209 --> 00:05:31.689 power usage, and just told it to optimize for 00:05:31.689 --> 00:05:34.430 energy efficiency. Specifically, they optimized 00:05:34.430 --> 00:05:36.829 for a metric called power usage effectiveness, 00:05:37.370 --> 00:05:40.810 or PUE. And what the AI started recommending 00:05:40.810 --> 00:05:44.689 was, well, deeply unintuitive to the human engineers 00:05:44.689 --> 00:05:47.290 on site. Right. The sources note the AI figured 00:05:47.290 --> 00:05:49.490 out how to exploit winter conditions to produce 00:05:49.490 --> 00:05:52.389 colder than normal water. Yeah. Which fundamentally 00:05:52.389 --> 00:05:55.310 altered the standard operating procedures. The 00:05:55.310 --> 00:05:57.750 human engineers were apparently highly skeptical. 00:05:57.970 --> 00:06:00.250 Understandably so. But they let it run, and it 00:06:00.250 --> 00:06:02.550 reduced the energy used for cooling by up to 00:06:02.550 --> 00:06:05.569 30 percent. At Google's scale, that is a monumental 00:06:05.569 --> 00:06:07.889 drop in power consumption. If we connect this 00:06:07.889 --> 00:06:10.550 to the bigger picture, why was the AI able to 00:06:10.550 --> 00:06:12.569 find those efficiencies when human engineers 00:06:12.569 --> 00:06:15.040 couldn't? Because it lacks human heuristics. 00:06:15.480 --> 00:06:17.579 Human operators, even world -class engineers, 00:06:17.819 --> 00:06:20.139 rely on rules of thumb. An engineer might look 00:06:20.139 --> 00:06:21.839 at a cooling valve and think, well, we never 00:06:21.839 --> 00:06:25.160 opened that past 40 % in December. The manuals 00:06:25.160 --> 00:06:27.339 say it's inefficient. The AI doesn't know what 00:06:27.339 --> 00:06:30.279 December is. And it hasn't read the manual. It 00:06:30.279 --> 00:06:33.160 only sees the raw mathematical relationship between 00:06:33.160 --> 00:06:36.170 the valve state and the PUE metric. It lacks 00:06:36.170 --> 00:06:38.350 our blind spots. But I keep thinking about the 00:06:38.350 --> 00:06:40.870 sheer physical risk here. Yeah. It's one thing 00:06:40.870 --> 00:06:43.610 to lose a virtual game of StarCraft. Sure. It 00:06:43.610 --> 00:06:46.490 is an entirely different reality to hand over 00:06:46.490 --> 00:06:49.009 the thermal controls of a global data center 00:06:49.009 --> 00:06:52.069 to an algorithm. If it decides to try a random 00:06:52.069 --> 00:06:54.490 action just to see what happens, it could melt 00:06:54.490 --> 00:06:56.209 down the servers. Which is why they didn't just 00:06:56.209 --> 00:06:58.889 hand over the keys, they implemented strict safety 00:06:58.889 --> 00:07:01.389 bounds. Think of it as a mathematical fence. 00:07:01.730 --> 00:07:03.689 The reinforcement learning agent was free to 00:07:03.689 --> 00:07:05.529 experiment and adjust the cooling parameters, 00:07:05.930 --> 00:07:08.089 but its outputs were filtered through a hard 00:07:08.089 --> 00:07:10.629 -coded safety system. Ah, I see. Yeah. If the 00:07:10.629 --> 00:07:12.970 AI suggested a temperature curve that exceeded 00:07:12.970 --> 00:07:15.870 hardware safety limits, the system simply rejected 00:07:15.870 --> 00:07:18.550 it. It's guided optimization. And they scaled 00:07:18.550 --> 00:07:21.819 this optimization way beyond server rooms. By 00:07:21.819 --> 00:07:24.680 mid -2025, DeepMind launched the Weather Lab. 00:07:24.980 --> 00:07:27.399 They took neural networks and trained them on 00:07:27.399 --> 00:07:30.879 45 years of historical global weather data. And 00:07:30.879 --> 00:07:34.379 during the 2025 Atlantic hurricane season, this 00:07:34.379 --> 00:07:37.620 AI systematically outperformed the U .S. National 00:07:37.620 --> 00:07:40.180 Weather Service in predicting cyclone tracks 00:07:40.180 --> 00:07:43.860 and intensity up to 15 days in advance. To appreciate 00:07:43.860 --> 00:07:46.540 how massive that is, you have to understand how 00:07:46.540 --> 00:07:49.100 traditional weather forecasting works. The National 00:07:49.100 --> 00:07:52.779 Weather Service uses incredibly complex deterministic 00:07:52.779 --> 00:07:55.339 physics models. Super computers running simulations, 00:07:55.560 --> 00:07:58.040 right? Exactly. Fluid dynamics, thermodynamics, 00:07:58.319 --> 00:08:00.699 the Navier -Stokes equations. But DeepMind's 00:08:00.699 --> 00:08:02.660 weather lab approached it completely differently. 00:08:02.860 --> 00:08:04.819 It didn't try to simulate the physics at all. 00:08:04.959 --> 00:08:07.560 Wait, really? Yeah. It treated 45 years of weather 00:08:07.560 --> 00:08:10.360 data as an incredibly complex pattern recognition 00:08:10.360 --> 00:08:13.079 game. It learned the hidden correlation between 00:08:13.079 --> 00:08:15.519 atmospheric pressure in one hemisphere and wind 00:08:15.519 --> 00:08:17.759 shear in another. Wow. Correlations that are 00:08:17.759 --> 00:08:19.779 just too complex for traditional physics engines 00:08:19.779 --> 00:08:22.459 to compute in real time. So the AI is mapping 00:08:22.459 --> 00:08:26.100 the macro patterns of the globe. But if an architecture 00:08:26.100 --> 00:08:28.360 is powerful enough to decode the atmosphere, 00:08:29.139 --> 00:08:32.120 it can also be turned inward, right, to decode 00:08:32.120 --> 00:08:34.659 the microscopic building blocks of life itself. 00:08:34.919 --> 00:08:37.360 And this is the leap from applied engineering 00:08:37.360 --> 00:08:40.980 to pure scientific discovery. We are talking 00:08:40.980 --> 00:08:44.230 about protein folding. For 50 years, this was 00:08:44.230 --> 00:08:46.809 one of the grandest challenges in molecular biology. 00:08:46.889 --> 00:08:49.850 It really was. Because proteins are the machinery 00:08:49.850 --> 00:08:52.509 of life, and they start as one -dimensional strings 00:08:52.509 --> 00:08:55.789 of amino acids. But to do their job, they fold 00:08:55.789 --> 00:08:58.429 into highly complex three -dimensional shapes. 00:08:58.490 --> 00:09:00.870 Right. And knowing that shape tells you exactly 00:09:00.870 --> 00:09:03.769 what the protein does, which is the key to designing 00:09:03.769 --> 00:09:06.529 targeted medicines, curing diseases, or even 00:09:06.529 --> 00:09:09.710 engineering enzymes that eat plastic. But calculating 00:09:09.710 --> 00:09:13.070 how that 1D string will fold into 3D space is 00:09:13.070 --> 00:09:15.649 computationally staggering. I mean, the number 00:09:15.649 --> 00:09:18.309 of possible configurations for a single protein 00:09:18.309 --> 00:09:20.990 is greater than the number of atoms in the observable 00:09:20.990 --> 00:09:23.629 universe. That's insane. It is. Traditional methods 00:09:23.629 --> 00:09:25.570 like X -ray crystallography could take years 00:09:25.570 --> 00:09:27.870 to map just one structure. Enter alpha fold. 00:09:28.649 --> 00:09:32.029 By 2020, DeepMind largely solved the structural 00:09:32.029 --> 00:09:35.220 prediction problem. By 2022, they released a 00:09:35.220 --> 00:09:37.179 database containing the predicted structures 00:09:37.179 --> 00:09:40.639 of over 200 million proteins. Which is virtually 00:09:40.639 --> 00:09:43.100 every protein known to human science. Yeah. And 00:09:43.100 --> 00:09:46.039 then in May 2024, they released AlphaFold 3, 00:09:46.159 --> 00:09:48.159 which went even further, predicting how those 00:09:48.159 --> 00:09:51.610 proteins interact with DNA and RNA. Amazing. 00:09:51.929 --> 00:09:54.769 It was a breakthrough so seismic that Demis Isavis 00:09:54.769 --> 00:09:58.049 and John Jumper were awarded the 2024 Nobel Prize 00:09:58.049 --> 00:10:01.090 in Chemistry. It compressed decades of biological 00:10:01.090 --> 00:10:03.870 research into a matter of days. And what is so 00:10:03.870 --> 00:10:06.269 profound is that the system isn't just mimicking 00:10:06.269 --> 00:10:09.990 known biology. It is actively inferring the underlying 00:10:09.990 --> 00:10:12.110 biophysical rules from the data it was trained 00:10:12.110 --> 00:10:14.879 on. But biology isn't the only foundational code 00:10:14.879 --> 00:10:17.379 they hacked. In 2023, they released a system 00:10:17.379 --> 00:10:19.940 called Alphadev, and they pointed it at computer 00:10:19.940 --> 00:10:22.220 code itself, specifically sorting algorithm. 00:10:22.419 --> 00:10:24.720 Yes, sorting algorithm. These are the fundamental 00:10:24.720 --> 00:10:26.840 instructions computers use to organize data, 00:10:26.980 --> 00:10:29.899 and they're used trillions of times a day. Alphadev 00:10:29.899 --> 00:10:31.759 discovered a new sorting algorithm for short 00:10:31.759 --> 00:10:34.639 sequences that was up to 70 % faster. It was 00:10:34.639 --> 00:10:37.460 so undeniably superior that it was integrated 00:10:37.460 --> 00:10:41.080 into the C++ standard library, which hadn't updated 00:10:41.080 --> 00:10:44.320 those algorithms in over a decade. But, and this 00:10:44.320 --> 00:10:46.360 is the part that genuinely breaks my brain, if 00:10:46.360 --> 00:10:48.299 human computer scientists have been staring at 00:10:48.299 --> 00:10:51.100 these exact sorting algorithms since the 1990s. 00:10:51.360 --> 00:10:53.779 Analyzing every single line of assembly code. 00:10:54.019 --> 00:10:57.600 Exactly. How does an AI just find a 70 % faster 00:10:57.600 --> 00:11:00.360 route? What is it seeing that the entire computer 00:11:00.360 --> 00:11:03.159 science community missed? It sees the code the 00:11:03.159 --> 00:11:06.059 exact same way AlphaGo saw the Go board. AlphaDev 00:11:06.059 --> 00:11:08.759 doesn't read code like a human programmer. A 00:11:08.759 --> 00:11:11.279 human thinks in sequential logic, right? We build 00:11:11.279 --> 00:11:13.659 algorithms based on what makes logical sense 00:11:13.659 --> 00:11:16.500 to our brains, but DeepMind set up the assembly 00:11:16.500 --> 00:11:18.980 language of the CPU as a game state. Oh, I see. 00:11:19.100 --> 00:11:21.500 The moves were the swapping, adding, or deleting 00:11:21.500 --> 00:11:24.179 of individual machine instructions. The win state 00:11:24.179 --> 00:11:26.299 was sorting the data correctly with the fewest 00:11:26.299 --> 00:11:28.159 possible operations. So it was just playing a 00:11:28.159 --> 00:11:31.149 game of efficiency. Yes, and in doing so, it 00:11:31.149 --> 00:11:33.350 found a shortcut. It discovered that you could 00:11:33.350 --> 00:11:35.830 literally leave out a specific instruction that 00:11:35.830 --> 00:11:38.429 humans always assumed was logically mandatory 00:11:38.429 --> 00:11:41.830 to ensure the data was sorted. The AI proved 00:11:41.830 --> 00:11:44.669 it wasn't mandatory at all. It found an invisible 00:11:44.669 --> 00:11:46.909 pathway through the assembly code because it 00:11:46.909 --> 00:11:49.230 was completely unburdened by human assumptions 00:11:49.230 --> 00:11:52.230 about how code should be written. It proves that 00:11:52.230 --> 00:11:55.269 these systems can discover entirely new mathematically 00:11:55.269 --> 00:11:57.929 verifiable truths. Okay, so we've tracked this 00:11:57.929 --> 00:12:00.490 evolution from an AI that learns virtual rules 00:12:00.490 --> 00:12:03.000 to an AI that that optimizes physical reality 00:12:03.000 --> 00:12:06.240 to an AI that discovers new biological and mathematical 00:12:06.240 --> 00:12:10.000 truths. But over the last two years, the fundamental 00:12:10.000 --> 00:12:13.240 nature of what DeepMind outputs has changed. 00:12:13.600 --> 00:12:16.940 Because in April 2023, they merged with Google's 00:12:16.940 --> 00:12:19.279 brain division, pooling all their resources to 00:12:19.279 --> 00:12:22.139 create Google DeepMind. And this catalyzed the 00:12:22.139 --> 00:12:24.720 generative explosion, the Gemini era. We moved 00:12:24.720 --> 00:12:27.299 from systems designed to decode existing realities 00:12:27.299 --> 00:12:29.360 to systems designed to generate entirely new 00:12:29.360 --> 00:12:31.870 ones. So what does this all mean for you? The 00:12:31.870 --> 00:12:34.049 timeline here is moving at absolute breakneck 00:12:34.049 --> 00:12:36.710 speed for those of you listening today, March 00:12:36.710 --> 00:12:40.929 26th, 2026. Hard to keep up. Very. A year ago, 00:12:41.049 --> 00:12:43.529 we saw Gemini 2 .5, which introduced reasoning 00:12:43.529 --> 00:12:46.110 tokens, giving the model the ability to internally 00:12:46.110 --> 00:12:49.029 think and verify its logic before outputting 00:12:49.029 --> 00:12:53.029 text. By November 2025, Gemini 3 Pro was woven 00:12:53.029 --> 00:12:55.730 into the fabric of Google search. Right. But 00:12:55.730 --> 00:12:57.490 the text is almost the least interesting part 00:12:57.490 --> 00:13:01.820 now. Because in May 2025, they dropped VO3, generating 00:13:01.820 --> 00:13:04.399 high definition, synchronized video, dialogue, 00:13:04.580 --> 00:13:08.200 and ambient sound. And then in August 2025, Genie 00:13:08.200 --> 00:13:11.600 3 launched. Genie 3 is a massive conceptual leap. 00:13:11.820 --> 00:13:13.299 Right, because it doesn't just draw a picture. 00:13:13.360 --> 00:13:15.500 You give Genie 3 a text prompt or a single image, 00:13:15.700 --> 00:13:18.220 and it generates an interactive, real -time 3D 00:13:18.220 --> 00:13:20.379 virtual world. You can literally step into it 00:13:20.379 --> 00:13:22.220 and control the environment. And just yesterday, 00:13:22.299 --> 00:13:25.679 March 25th, 2026, they released Liria 3 Pro, 00:13:25.820 --> 00:13:28.480 which generates structurally aware, symphonic 00:13:28.480 --> 00:13:30.600 -level music. It's incredible. But how do we 00:13:30.600 --> 00:13:33.399 even contextualize this leap? I mean, how does 00:13:33.399 --> 00:13:36.240 an architecture designed to just predict the 00:13:36.240 --> 00:13:38.659 next word in a sentence, suddenly figure out 00:13:38.659 --> 00:13:41.899 how to generate a real -time 3D physics engine. 00:13:42.100 --> 00:13:44.399 Well, it comes down to the shift toward polyvalent 00:13:44.399 --> 00:13:48.500 or multimodal models. Early on, DeepMind built 00:13:48.500 --> 00:13:51.809 a model called GATO. Gaito was trained on text, 00:13:52.129 --> 00:13:54.730 image captions, and robotic arm movements all 00:13:54.730 --> 00:13:57.809 simultaneously. It could perform over 600 different 00:13:57.809 --> 00:14:00.590 tasks using the exact same neural weights. Wow. 00:14:00.870 --> 00:14:03.210 What DeepMind realized is that if you scale that 00:14:03.210 --> 00:14:05.950 architecture massively, the model stops just 00:14:05.950 --> 00:14:08.269 predicting pixels or text and starts inferring 00:14:08.269 --> 00:14:10.830 a world model. A world model, meaning it actually 00:14:10.830 --> 00:14:13.210 understands the physics behind the image. Exactly. 00:14:13.409 --> 00:14:16.129 To accurately generate a video of a glass falling 00:14:16.129 --> 00:14:19.169 off a table, the model has to implicitly learn 00:14:19.169 --> 00:14:21.580 the concepts of gravity, mass, and fragility. 00:14:21.840 --> 00:14:25.340 With Genie 3, it uses latent action models. It 00:14:25.340 --> 00:14:28.320 infers how an environment should react to a user's 00:14:28.320 --> 00:14:30.639 input based on billions of examples of physical 00:14:30.639 --> 00:14:33.519 interactions. The insight here is that the boundary 00:14:33.519 --> 00:14:36.960 between calculating data and creating media has 00:14:36.960 --> 00:14:39.620 completely dissolved. It is a total democratization 00:14:39.620 --> 00:14:42.460 of creation. But that democratization forces 00:14:42.460 --> 00:14:45.320 an incredible responsibility onto you, the listener. 00:14:45.860 --> 00:14:48.679 If Genie 3 can spin up a photorealistic reality 00:14:48.559 --> 00:14:52.360 on the fly, and Lyria 3 Pro can generate human 00:14:52.360 --> 00:14:56.080 emotion through a symphony. I mean, the critical 00:14:56.080 --> 00:14:57.919 thinking required to navigate what is synthetic 00:14:57.919 --> 00:15:00.379 and what is authentic is now a daily survival 00:15:00.379 --> 00:15:02.960 skill. Yeah, absolutely. And that friction between 00:15:02.960 --> 00:15:05.799 incredible capability and human reality brings 00:15:05.799 --> 00:15:07.980 us to the ethical guardrails. Because when you 00:15:07.980 --> 00:15:10.679 have the power to parse biological data and generate 00:15:10.679 --> 00:15:13.539 realities, the collision with human privacy is 00:15:13.539 --> 00:15:15.649 almost inevitable. This raises an important question 00:15:15.649 --> 00:15:18.730 because innovation almost always outpaces regulation, 00:15:19.110 --> 00:15:21.610 and DeepMind has had to navigate severe structural 00:15:21.610 --> 00:15:23.929 failures regarding data privacy. The clearest 00:15:23.929 --> 00:15:26.429 example in the sources is the NHS data sharing 00:15:26.429 --> 00:15:29.850 controversy. Wait. In 2016, a hospital network 00:15:29.850 --> 00:15:33.250 in London, the Royal Free London NHS Foundation 00:15:33.250 --> 00:15:35.929 Trust, partnered with DeepMind to build an app 00:15:35.929 --> 00:15:40.049 called Streams. The goal was noble, right? Alert 00:15:40.049 --> 00:15:42.230 doctors instantly when a patient was at risk 00:15:42.230 --> 00:15:44.870 of acute kidney injury. Good intentions. To build 00:15:44.870 --> 00:15:47.529 it, the Trust gave DeepMind access to an estimated 00:15:47.529 --> 00:15:50.740 1 .6 million. patient records. But this wasn't 00:15:50.740 --> 00:15:53.559 just anonymized kidney data. It included highly 00:15:53.559 --> 00:15:56.700 sensitive historical logs, HIV diagnoses, depression 00:15:56.700 --> 00:15:59.019 treatments, abortion records. Right. And while 00:15:59.019 --> 00:16:01.039 the app functioned beautifully, I mean, doctors 00:16:01.039 --> 00:16:03.840 reported it saved massive amounts of time and 00:16:03.840 --> 00:16:06.779 improved care. The UK's Information Commission's 00:16:06.779 --> 00:16:08.759 office eventually stepped in and ruled that the 00:16:08.759 --> 00:16:10.779 hospital failed to comply with the Data Protection 00:16:10.779 --> 00:16:13.220 Act. Yeah. The patients had never been adequately 00:16:13.220 --> 00:16:15.100 informed that their entire medical histories 00:16:15.100 --> 00:16:17.840 were being handed over to a massive tech corporation 00:16:17.840 --> 00:16:20.259 for algorithm. development. It is a profound 00:16:20.259 --> 00:16:23.019 lesson in alignment. The objective function was 00:16:23.019 --> 00:16:25.759 stop kidney failure but the system surrounding 00:16:25.759 --> 00:16:28.779 it lacked the parameters for human consent. It's 00:16:28.779 --> 00:16:31.019 like building a Ferrari but completely forgetting 00:16:31.019 --> 00:16:33.659 to install the brakes or you know training a 00:16:33.659 --> 00:16:35.720 brilliant medical savant who can diagnose any 00:16:35.720 --> 00:16:38.019 disease flawlessly but failing to teach them 00:16:38.019 --> 00:16:40.419 the concept of doctor -patient confidentiality. 00:16:40.419 --> 00:16:43.240 Exactly. In response to the legal fallout DeepMind 00:16:43.240 --> 00:16:46.120 established a dedicated ethics and society unit 00:16:46.120 --> 00:16:50.059 and more recently in 2024, they implemented what 00:16:50.059 --> 00:16:53.580 they call a robot constitution for their AI agents. 00:16:53.620 --> 00:16:56.820 Right. It is literally inspired by Isaac Asimov's 00:16:56.820 --> 00:16:59.799 three laws of robotics. Rule one is essentially 00:16:59.799 --> 00:17:03.480 do no harm to a human being. Right. But structurally, 00:17:03.519 --> 00:17:05.460 how does that actually work? I mean, can a set 00:17:05.460 --> 00:17:08.079 of text based rules actually contain a system 00:17:08.079 --> 00:17:11.160 as smart as Gemini 3 Pro? How do you map a philosophical 00:17:11.160 --> 00:17:13.160 rule into a loss function? That is the bleeding 00:17:13.160 --> 00:17:15.670 edge of constitutional AI. You don't just paste 00:17:15.670 --> 00:17:18.089 the text into the code. During the reinforcement 00:17:18.089 --> 00:17:20.369 learning phase, the AI generates thousands of 00:17:20.369 --> 00:17:22.470 potential responses to prompts. OK. And then 00:17:22.470 --> 00:17:25.529 another AI or a human raider scores those responses 00:17:25.529 --> 00:17:27.809 strictly based on whether they violate the Constitution. 00:17:28.269 --> 00:17:30.690 The model is mathematically penalized for generating 00:17:30.690 --> 00:17:33.009 harmful outputs and rewarded for aligning with 00:17:33.009 --> 00:17:35.430 the principles. Oh, I see. Yeah, it bakes the 00:17:35.430 --> 00:17:37.250 ethical constraints directly into the neural 00:17:37.250 --> 00:17:39.109 weights before the model is ever released to 00:17:39.109 --> 00:17:41.450 the public. It's an attempt to mathematically 00:17:41.450 --> 00:17:44.769 encode human values, and it's a fitting capstone 00:17:44.769 --> 00:17:47.269 to the journey we've tracked today. We've gone 00:17:47.269 --> 00:17:50.849 from an AI trying to maximize a score in a 1970s 00:17:50.849 --> 00:17:53.789 arcade game to an AI mapping the proteins of 00:17:53.789 --> 00:17:56.450 the human body to models that generate entire 00:17:56.450 --> 00:17:59.049 realities governed by a digital constitution. 00:17:59.279 --> 00:18:02.500 It is the evolution of a system learning to navigate 00:18:02.500 --> 00:18:05.240 and ultimately shape our world. But there is 00:18:05.240 --> 00:18:08.339 one final piece of the puzzle buried deep in 00:18:08.339 --> 00:18:11.119 the source material. A release from May 2025 00:18:11.119 --> 00:18:14.460 called AlphaVolve. AlphaVolve is an evolutionary 00:18:14.460 --> 00:18:17.500 coding agent. It uses large language models to 00:18:17.500 --> 00:18:20.140 actively design new algorithms. It generates 00:18:20.140 --> 00:18:22.940 variations of code, tests them, selects the most 00:18:22.940 --> 00:18:25.539 efficient candidates, and then iteratively rewrites 00:18:25.539 --> 00:18:28.079 them to be even better. It is an AI acting as 00:18:28.079 --> 00:18:30.420 its own architect. And that is the thought we 00:18:30.420 --> 00:18:33.140 want to leave you with today. We've spent this 00:18:33.140 --> 00:18:35.420 entire deep dive looking at how humans have guided 00:18:35.420 --> 00:18:37.839 the learning of these models. But if DeepMind 00:18:37.839 --> 00:18:40.359 has built a system that can iteratively improve 00:18:40.359 --> 00:18:42.299 its own source code faster than a human ever 00:18:42.299 --> 00:18:45.180 could, well, what happens when Elf Evolve points 00:18:45.180 --> 00:18:47.819 its optimization engine at itself? It's a massive 00:18:47.819 --> 00:18:50.259 unknown. If it begins writing variations of its 00:18:50.259 --> 00:18:53.019 own architecture, the software you use next week 00:18:53.019 --> 00:18:55.759 might be built by an intelligence that no human 00:18:55.759 --> 00:18:58.869 engineer fully comprehends. We are rapidly approaching 00:18:58.869 --> 00:19:00.710 the moment where the machine writing the code 00:19:00.710 --> 00:19:02.829 becomes exponentially smarter than the human 00:19:02.829 --> 00:19:05.009 who wrote the machine. Thank you for taking this 00:19:05.009 --> 00:19:08.269 deep dive with us. Keep exploring, keep interrogating 00:19:08.269 --> 00:19:10.750 the technology around you, and we will see you 00:19:10.750 --> 00:19:11.109 next time.