WEBVTT 00:00:00.000 --> 00:00:01.740 So imagine this. You're sitting in the back of 00:00:01.740 --> 00:00:05.559 a robotaxi. It's a Tuesday, maybe 2 p .m. You're 00:00:05.559 --> 00:00:07.900 just, you know, answering emails, watching the 00:00:07.900 --> 00:00:09.960 suburbs roll by. It's totally mundane. Right. 00:00:10.060 --> 00:00:12.960 And then suddenly the sky turns this bruised 00:00:12.960 --> 00:00:17.059 purple. A tornado touches down like three blocks 00:00:17.059 --> 00:00:19.719 ahead. Oh, wow. Or maybe you turn a corner in 00:00:19.719 --> 00:00:22.160 Phoenix and there's a Texas Longhorn steer just 00:00:22.160 --> 00:00:26.019 standing dead center in the lane. Or the neighborhood 00:00:26.019 --> 00:00:29.859 is quite literally on fire. That is a terrifying 00:00:29.859 --> 00:00:32.439 Tuesday. It is. But here's the question that 00:00:32.439 --> 00:00:35.039 I think keeps engineers up at night. How do you 00:00:35.039 --> 00:00:37.299 train a computer for that? Yeah. You can't exactly 00:00:37.299 --> 00:00:39.640 drive a million miles just waiting for a steer 00:00:39.640 --> 00:00:41.539 to wander onto the highway. You don't wait for 00:00:41.539 --> 00:00:43.840 the disaster. Hallucinate it. Welcome to the 00:00:43.840 --> 00:00:46.520 era where AI dreams up nightmares just to keep 00:00:46.520 --> 00:00:49.920 us safe. It's poetic, sure. But, you know. Practically, 00:00:49.920 --> 00:00:51.780 it's just data efficiency. You can't wait 100 00:00:51.780 --> 00:00:54.280 years for it steered across the road. You have 00:00:54.280 --> 00:00:56.899 to force the error. Force the error. I like it. 00:00:57.179 --> 00:00:59.780 Welcome to the Deep Dive. It is Monday, February 00:00:59.780 --> 00:01:03.560 9th, 2026. We're sitting right at this intersection 00:01:03.560 --> 00:01:06.299 of simulation and physical reality today. I've 00:01:06.299 --> 00:01:08.420 got a stack of notes here that paint a pretty 00:01:08.420 --> 00:01:11.379 wild picture of where we are. And looking at 00:01:11.379 --> 00:01:13.379 the sources we have today, it feels like we've 00:01:13.379 --> 00:01:15.519 hit a tipping point. We're seeing what people 00:01:15.519 --> 00:01:19.140 are calling the super cycle of 2026. Yeah. Really 00:01:19.140 --> 00:01:21.359 taking shape. And numbers are staggering. We'll 00:01:21.359 --> 00:01:23.900 get to that 800 million users, which is just 00:01:23.900 --> 00:01:26.920 it's mind blowing. But I want to stick with that 00:01:26.920 --> 00:01:29.579 image of the Robotaxi and the Texas Longhorn 00:01:29.579 --> 00:01:31.879 for a second, because this comes directly from 00:01:31.879 --> 00:01:34.689 Waymo. This is the new Waymo world model. And 00:01:34.689 --> 00:01:36.750 the engine underneath it is what really matters. 00:01:36.930 --> 00:01:39.609 They're using Google's Genie 3. Genie 3. Okay, 00:01:39.650 --> 00:01:41.329 let's unpack this. Because usually when we talk 00:01:41.329 --> 00:01:43.750 about simulations, my mind goes to, like, video 00:01:43.750 --> 00:01:46.790 games. Grand Theft Auto, but for robots. Is that 00:01:46.790 --> 00:01:48.730 what this is? No, and that's a crucial distinction. 00:01:48.950 --> 00:01:51.629 In a video game engine, like Unreal or Unity, 00:01:51.870 --> 00:01:54.650 a human programmer has defined all the physics. 00:01:54.849 --> 00:01:57.090 Right. They wrote the code that says, if car 00:01:57.090 --> 00:02:00.489 hits wall, then crumple metal. It's rigid. It's 00:02:00.489 --> 00:02:03.129 just a render. And Genie 3. Genie 3 is generative. 00:02:03.599 --> 00:02:05.620 It's not rendering the world based on a set of 00:02:05.620 --> 00:02:08.620 rules. It's dreaming it up based on memory. It 00:02:08.620 --> 00:02:10.740 has watched millions of hours of driving footage. 00:02:11.060 --> 00:02:14.259 So when it creates a scenario, it's predicting 00:02:14.259 --> 00:02:16.960 the next pixel, the same way ChatGPT predicts 00:02:16.960 --> 00:02:19.319 the next word. So it's hallucinating the physics. 00:02:19.719 --> 00:02:22.639 In a way, yes. But here's the kicker. It's not 00:02:22.639 --> 00:02:25.120 just generating video. It's generating the sensor 00:02:25.120 --> 00:02:27.680 data. What do you mean? It simulates the LiDAR 00:02:27.680 --> 00:02:29.659 returns, the radar waves, the camera inputs, 00:02:29.879 --> 00:02:33.400 all of it. Wait, so... Does the car's computer 00:02:33.400 --> 00:02:36.599 even know it's in a simulation? It has no idea. 00:02:36.879 --> 00:02:39.060 To the perception stack, the brain of the car, 00:02:39.159 --> 00:02:42.139 a simulated photon and a real photon are mathematically 00:02:42.139 --> 00:02:45.379 identical. That is slightly unsettling. It solves 00:02:45.379 --> 00:02:47.439 the biggest problem in autonomous driving, the 00:02:47.439 --> 00:02:50.479 long tail. You can drive a billion miles in sunny 00:02:50.479 --> 00:02:52.879 California and never see a snowstorm or a flood. 00:02:53.000 --> 00:02:55.199 Yeah, you just won't encounter it. But with Genie 00:02:55.199 --> 00:02:58.039 3, engineers can just type in a command. They 00:02:58.039 --> 00:03:00.620 use natural language to say add a flood or make 00:03:00.620 --> 00:03:03.500 it nighttime or insert a literal elephant. A 00:03:03.500 --> 00:03:05.259 literal elephant. That is actually in the notes. 00:03:05.400 --> 00:03:07.800 It is. And it matters because of this safe danger 00:03:07.800 --> 00:03:11.620 concept. You can test what if decisions. What 00:03:11.620 --> 00:03:14.199 if the car swerves left? What if it breaks hard? 00:03:14.340 --> 00:03:16.979 You can run that scenario 10 ,000 times in the 00:03:16.979 --> 00:03:19.840 cloud without ever risking a passenger or, you 00:03:19.840 --> 00:03:21.419 know, a pedestrian. It's interesting because 00:03:21.419 --> 00:03:24.340 they call it language control. So an engineer 00:03:24.340 --> 00:03:27.300 is basically being a god of this little virtual 00:03:27.300 --> 00:03:29.939 world, just speaking disasters into existence. 00:03:30.240 --> 00:03:32.199 Let there be a tornado. Exactly. And controlling 00:03:32.199 --> 00:03:34.039 the scene layout, the traffic flow, everything. 00:03:34.500 --> 00:03:37.340 Waymo's bet here is really specific. If you can 00:03:37.340 --> 00:03:40.080 simulate everything. you can make real -world 00:03:40.080 --> 00:03:43.199 failure impossible, or at least, you know, statistically 00:03:43.199 --> 00:03:45.840 negligible. It creates this real moment of wonder 00:03:45.840 --> 00:03:48.400 for me. Just thinking about the computational 00:03:48.400 --> 00:03:51.360 power required to simulate how a tornado affects 00:03:51.360 --> 00:03:53.639 a LIDAR sensor, that's not just pixels, that's 00:03:53.639 --> 00:03:55.800 light physics being predicted by a neural net. 00:03:55.960 --> 00:03:58.479 It's massive. And they're even building a Gemini 00:03:58.479 --> 00:04:00.599 -based voice assistant for inside the car, so 00:04:00.599 --> 00:04:03.340 the interaction inside is evolving too. So here's 00:04:03.340 --> 00:04:05.340 where it gets really interesting for me. If the 00:04:05.340 --> 00:04:08.439 simulation is indistinguishable from reality 00:04:08.439 --> 00:04:12.300 for the car, does the real world even matter 00:04:12.300 --> 00:04:16.259 for training anymore? Not really. If the data 00:04:16.259 --> 00:04:18.839 is perfect, the training is valid. The car doesn't 00:04:18.839 --> 00:04:20.879 care if the photon came from the sun or from 00:04:20.879 --> 00:04:22.839 a server. Okay, so that's simulating the world. 00:04:22.879 --> 00:04:25.000 We're building the matrix for cars. But we also 00:04:25.000 --> 00:04:27.620 have this breakthrough from Harvard and Stanford 00:04:27.620 --> 00:04:30.000 about moving through the world. And this one 00:04:30.000 --> 00:04:32.420 seems to bridge the gap between chatbots and 00:04:32.420 --> 00:04:35.959 robots. This is the OAT system. OAT. The name 00:04:35.959 --> 00:04:37.540 isn't as important as the mechanism. This is 00:04:37.540 --> 00:04:40.160 about predicting robot actions like their text. 00:04:40.339 --> 00:04:42.560 Right, because we know transformers, the tech 00:04:42.560 --> 00:04:45.000 behind GPT and clot, are really good at predicting 00:04:45.000 --> 00:04:47.339 the next word in the sentence. The cat sat on 00:04:47.339 --> 00:04:50.000 the... and the AI knows math. Exactly. But robots 00:04:50.000 --> 00:04:52.639 don't move in words. They move in continuous 00:04:52.639 --> 00:04:56.259 kind of messy physical arcs. A robot arm reaching 00:04:56.259 --> 00:04:59.240 for a cup isn't a discrete word. It's a flow 00:04:59.240 --> 00:05:02.759 of analog data. Voltage, torque, velocity. All 00:05:02.759 --> 00:05:05.240 of that. So how do you turn a backflip into a 00:05:05.240 --> 00:05:07.730 sentence? I have no idea. You have to digitize 00:05:07.730 --> 00:05:10.089 it. Think of it like music. Sound is a continuous 00:05:10.089 --> 00:05:13.089 wave, right? But to put it on a CD or an MP3, 00:05:13.269 --> 00:05:15.250 you have to chop it up into digital bits. You 00:05:15.250 --> 00:05:17.870 take samples. Okay, I follow that. OAT does that 00:05:17.870 --> 00:05:20.829 for movement. The researchers built an encoder 00:05:20.829 --> 00:05:23.209 that takes that continuous motion, the robot 00:05:23.209 --> 00:05:27.000 arm swinging, and splits it into chunks. Then 00:05:27.000 --> 00:05:29.839 it uses a process called finite scalar quantization. 00:05:30.060 --> 00:05:32.439 That is a mouthful. It is. But just think of 00:05:32.439 --> 00:05:35.300 it as creating a vocabulary. It forces the infinite 00:05:35.300 --> 00:05:38.139 complexity of a robot arm's arc into a fixed 00:05:38.139 --> 00:05:41.939 menu of specific movement words or tokens. So 00:05:41.939 --> 00:05:43.860 instead of a smooth wave, it becomes a series 00:05:43.860 --> 00:05:46.759 of steps, like token A, then token B, then token 00:05:46.759 --> 00:05:49.459 C. Exactly. And because these tokens flow left 00:05:49.459 --> 00:05:51.720 to right, just like a sentence in English, standard 00:05:51.720 --> 00:05:54.000 large language models can process them. So you 00:05:54.000 --> 00:05:56.860 could feed a movement sequence into GPT -5 or 00:05:56.860 --> 00:06:00.639 CLAWD and it can autocomplete the movement. Yes. 00:06:01.000 --> 00:06:03.740 The breakthrough is that it turns physical dexterity 00:06:03.740 --> 00:06:06.579 into a language problem. The first few tokens 00:06:06.579 --> 00:06:08.899 might describe the general direction, move arm 00:06:08.899 --> 00:06:11.819 up, and the later tokens fill in the fine motor 00:06:11.819 --> 00:06:15.300 details, rotate wrist 10 degrees. That feels 00:06:15.300 --> 00:06:18.040 huge. Yeah. It means we aren't reinventing the 00:06:18.040 --> 00:06:20.360 wheel for robotics. We're just piggybacking on 00:06:20.360 --> 00:06:22.259 the massive intelligence we already built for 00:06:22.259 --> 00:06:25.160 chatbots. And it is beating previous methods 00:06:25.160 --> 00:06:28.079 across 20 different tasks. It's vastly more efficient. 00:06:28.279 --> 00:06:31.660 The implication is wild. If a robot's brain is 00:06:31.660 --> 00:06:34.879 just a large language model, then the robot effectively 00:06:34.879 --> 00:06:38.319 knows everything the Internet knows. So does 00:06:38.319 --> 00:06:40.220 this mean we can eventually just talk a robot 00:06:40.220 --> 00:06:43.480 into learning a backflip? Essentially, yes. By 00:06:43.480 --> 00:06:45.540 treating the backflip as a sentence of movement 00:06:45.540 --> 00:06:47.699 tokens, you're just prompting it to complete 00:06:47.699 --> 00:06:49.980 the thought, but physically. Complete the thought 00:06:49.980 --> 00:06:52.240 physically. I like that. It connects perfectly 00:06:52.240 --> 00:06:53.939 to the sheer scale of what is happening right 00:06:53.939 --> 00:06:57.420 now. We mentioned the date February 2026, and 00:06:57.420 --> 00:06:59.759 the industry is calling this the super cycle. 00:06:59.939 --> 00:07:01.920 The numbers coming out of OpenAI this week are 00:07:01.920 --> 00:07:06.050 frankly absurd. 800 million weekly users. 800 00:07:06.050 --> 00:07:08.949 million. That is nearly the population of the 00:07:08.949 --> 00:07:11.769 entire generic Western Hemisphere interacting 00:07:11.769 --> 00:07:13.930 with these models every single week. I remember 00:07:13.930 --> 00:07:15.930 when getting to 100 million was the fastest in 00:07:15.930 --> 00:07:18.329 history. Now they're reporting growth is back 00:07:18.329 --> 00:07:21.430 to over 10 % monthly. And it's not just chat. 00:07:21.990 --> 00:07:26.389 Codex usage. The coding AI surged. 50 % just 00:07:26.389 --> 00:07:29.589 after the GPT 5 .3 launch. This tells me that 00:07:29.589 --> 00:07:31.569 this isn't a novelty anymore. This is infrastructure. 00:07:31.870 --> 00:07:34.730 Speaking of infrastructure, A16Z, that's Andreas 00:07:34.730 --> 00:07:39.089 and Horowitz, just dropped $1 .7 billion into 00:07:39.089 --> 00:07:41.589 AI infrastructure. Wow. And that's out of a $15 00:07:41.589 --> 00:07:44.209 billion fund. They are literally rebuilding the 00:07:44.209 --> 00:07:46.050 platforms from the ground up. They're seeing 00:07:46.050 --> 00:07:49.519 2026 as the year. The experimental phase ends 00:07:49.519 --> 00:07:52.180 and the utility phase begins. This is about chips, 00:07:52.300 --> 00:07:54.759 data centers, you know, the pipes that run the 00:07:54.759 --> 00:07:57.000 Internet. But with utility comes commercialization. 00:07:57.139 --> 00:07:59.839 And I have to admit, I had a bit of a sigh moment 00:07:59.839 --> 00:08:02.100 reading the news this morning. The ads. The ads. 00:08:02.120 --> 00:08:05.000 Ads are officially live in ChatGPT. Adobe is 00:08:05.000 --> 00:08:06.980 one of the launch partners. They're testing ads 00:08:06.980 --> 00:08:09.439 for Photoshop and Acrobat right there in the 00:08:09.439 --> 00:08:11.870 chat interface. It was inevitable, right? I know, 00:08:11.949 --> 00:08:13.850 I know. But I have to admit, I just hated seeing 00:08:13.850 --> 00:08:17.189 it. The pristine white box. It felt like a sanctuary, 00:08:17.470 --> 00:08:21.689 you know? Just pure intelligence. No noise. Now, 00:08:21.810 --> 00:08:23.769 if I ask for a summary or a photo editing tip, 00:08:24.009 --> 00:08:27.269 I might get a nudge to buy Firefly. It feels 00:08:27.269 --> 00:08:29.170 like the magic just got a corporate logo slapped 00:08:29.170 --> 00:08:32.169 on it. I get the nostalgia for the research preview 00:08:32.169 --> 00:08:35.070 era. Yeah. But look at the scale. Right. You 00:08:35.070 --> 00:08:39.190 can't serve 800 million users on compute -heavy 00:08:39.190 --> 00:08:42.049 models for free. Forever. It's the tradeoff. 00:08:42.210 --> 00:08:44.149 We're moving from the wild west of discovery 00:08:44.149 --> 00:08:47.210 to the utility grade of electricity. The electric 00:08:47.210 --> 00:08:49.070 company sends you a bill. That's a fair point. 00:08:49.230 --> 00:08:51.289 It is now utility grade infrastructure. It's 00:08:51.289 --> 00:08:53.029 like electricity or the telephone network. It's 00:08:53.029 --> 00:08:55.809 just there. It's just there. Right. And because 00:08:55.809 --> 00:08:58.730 it's just there, the way we use it is changing. 00:08:58.870 --> 00:09:01.070 So we have the 800 million users and we have 00:09:01.070 --> 00:09:02.970 the massive silicon build out. The question is, 00:09:02.990 --> 00:09:05.409 what are they actually doing with it? Because 00:09:05.409 --> 00:09:08.049 this week the tools shifted. We move from look 00:09:08.049 --> 00:09:10.120 at this funny video to this is. It's how you 00:09:10.120 --> 00:09:13.000 run a Fortune 500 company. The empowered AI tool 00:09:13.000 --> 00:09:16.299 sector is really maturing. But the video generation 00:09:16.299 --> 00:09:19.840 space is what catches your eye first. ByteDance 00:09:19.840 --> 00:09:23.039 released Seedens 2 .0. The demos for that are 00:09:23.039 --> 00:09:26.379 wild. I saw one with 2 .6 million views. It's 00:09:26.379 --> 00:09:29.080 a massive leap. And then you have Kling 3 .0. 00:09:29.139 --> 00:09:31.840 Kling 3 .0. This is the one claiming scene level 00:09:31.840 --> 00:09:34.759 control. What does that actually mean? It means 00:09:34.759 --> 00:09:37.860 we're moving past the slot machine era of AI 00:09:37.860 --> 00:09:41.490 video. type of prompt pull the lever and just 00:09:41.490 --> 00:09:43.590 hope the video isn't a nightmare right scene 00:09:43.590 --> 00:09:46.450 level means multi -shot control longer takes 00:09:46.450 --> 00:09:49.049 yeah it understands continuity you are directing 00:09:49.049 --> 00:09:51.389 not just prompting but the biggest shift for 00:09:51.389 --> 00:09:53.529 me and this ties back to that utility idea is 00:09:53.529 --> 00:09:56.629 open ai frontier ah this is the enterprise play 00:09:56.629 --> 00:09:59.210 and this is where the money is explain the difference 00:09:59.210 --> 00:10:01.840 here Because we have team accounts, we have enterprise 00:10:01.840 --> 00:10:04.679 accounts, what is Frontier? So Frontier isn't 00:10:04.679 --> 00:10:06.639 just about giving your employees access to a 00:10:06.639 --> 00:10:10.059 chatbot. It's a platform for managing AI agents. 00:10:10.320 --> 00:10:13.480 Right, not chatbots that answer questions, but 00:10:13.480 --> 00:10:16.519 agents that perform workflows. Frontier provides 00:10:16.519 --> 00:10:19.360 shared context, onboarding for these digital 00:10:19.360 --> 00:10:23.039 workers, and oversight. It treats the AI as a 00:10:23.039 --> 00:10:26.460 workforce, not a software tool. See, that's the 00:10:26.460 --> 00:10:28.720 shift. We're moving from generating content, 00:10:28.960 --> 00:10:32.100 write me a poem, make me a video, to orchestrating 00:10:32.100 --> 00:10:34.600 labor. Exactly. The human role is shifting. You 00:10:34.600 --> 00:10:36.559 aren't playing the violin anymore. You're waving 00:10:36.559 --> 00:10:39.779 the baton. You are the conductor. The conductor. 00:10:39.940 --> 00:10:43.200 That is both exciting and terrifying. It feels 00:10:43.200 --> 00:10:45.200 like the friction is disappearing, whether it's 00:10:45.200 --> 00:10:47.639 creating a video scene or automating a business 00:10:47.639 --> 00:10:50.279 process. The barrier to entry is just melting 00:10:50.279 --> 00:10:53.190 away. And that brings its own chaos. But it's 00:10:53.190 --> 00:10:55.830 the reality of 2026. I feel like we need to take 00:10:55.830 --> 00:10:57.990 a breath and look at the big picture here. We've 00:10:57.990 --> 00:11:00.169 covered a lot of ground. We have the God mode 00:11:00.169 --> 00:11:02.789 simulation at Waymo. We have the robot language 00:11:02.789 --> 00:11:05.750 at Harvard. And we have the super cycle of adoption. 00:11:06.049 --> 00:11:08.409 That's a lot. If we zoom out, what's the thread 00:11:08.409 --> 00:11:11.070 connecting these for you? I think it's the synthetic 00:11:11.070 --> 00:11:13.830 tipping point. Look at the three pillars. First, 00:11:13.990 --> 00:11:16.690 simulation. We're replacing real -world testing 00:11:16.690 --> 00:11:19.809 with synthetic experiences. That's Waymo. Okay. 00:11:20.090 --> 00:11:23.000 Second. Translation, we're converting physical 00:11:23.000 --> 00:11:25.360 movement into synthetic language tokens, which 00:11:25.360 --> 00:11:29.740 is OAT. And third, scale. We have 800 million 00:11:29.740 --> 00:11:32.340 people living and working inside this synthetic 00:11:32.340 --> 00:11:34.360 infrastructure. It really does feel like we're 00:11:34.360 --> 00:11:36.059 waking up in a different world every Monday. 00:11:36.179 --> 00:11:38.460 The lines between the physical and the digital 00:11:38.460 --> 00:11:41.480 are just gone. They're blurring, certainly. I 00:11:41.480 --> 00:11:42.799 want to leave everyone with a thought to mull 00:11:42.799 --> 00:11:45.840 over. We talked about Waymo simulating tornadoes. 00:11:45.840 --> 00:11:48.820 We talked about OAT turning backflips into words. 00:11:49.000 --> 00:11:52.139 Right. If an AI can simulate a tornado perfectly 00:11:52.139 --> 00:11:54.240 to the point where the sensors can't tell the 00:11:54.240 --> 00:11:57.320 difference and another AI can predict robot movements 00:11:57.320 --> 00:12:01.330 like words. How long until the simulation is 00:12:01.330 --> 00:12:03.549 the training ground for all physical labor? Like, 00:12:03.549 --> 00:12:05.509 do we ever need to practice anything in the real 00:12:05.509 --> 00:12:07.470 world again? That's the billion -dollar question. 00:12:07.730 --> 00:12:10.049 If the simulation is perfect, the real world 00:12:10.049 --> 00:12:12.409 is just the final exam. Something to think about 00:12:12.409 --> 00:12:14.509 while you're looking out the window of your robo 00:12:14.509 --> 00:12:17.190 -taxi. Thanks for diving in with us. Keep exploring. 00:12:17.309 --> 00:12:18.669 We'll catch you in the next deep dive.