WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:21.920 examine the ideas shaping the future of artificial minds. 5 00:00:23.920 --> 00:00:27.399 Welcome back to the Deep Dive. We are doing something 6 00:00:27.440 --> 00:00:29.320 a little bit different today. Usually we look at a 7 00:00:29.359 --> 00:00:32.280 product launch or a new piece of software, something that 8 00:00:32.359 --> 00:00:34.240 might change your workflow next month. 9 00:00:34.079 --> 00:00:36.159 Right, the usual tech updates, Yeah, exactly. 10 00:00:37.439 --> 00:00:40.119 But today we are looking at something that changes the 11 00:00:40.200 --> 00:00:44.920 fundamental math of civilization. We're digging into a stack of reports, 12 00:00:45.079 --> 00:00:49.880 industry analysis, and some really wild operational data from and 13 00:00:50.039 --> 00:00:53.759 this is the important part. February twenty twenty six. 14 00:00:54.640 --> 00:00:56.079 Right now, It is happening right now. 15 00:00:56.119 --> 00:00:59.159 And the headline here isn't that robots will build robots 16 00:00:59.159 --> 00:01:01.119 someday that they already are. 17 00:01:01.399 --> 00:01:03.399 It sounds like science fiction, doesn't it. I mean, when 18 00:01:03.399 --> 00:01:06.760 you say that phrase out loud, robots building robots, it 19 00:01:06.760 --> 00:01:09.319 feels like we should be breaking down a movie script. Yeah, 20 00:01:09.319 --> 00:01:11.519 but we aren't. Now we're on We're talking about actual 21 00:01:11.599 --> 00:01:15.239 factory floors in California, Texas and Shenzen that are hum 22 00:01:15.239 --> 00:01:15.840 and along. 23 00:01:15.640 --> 00:01:19.159 Today exactly the core topic for you all listening today 24 00:01:19.200 --> 00:01:22.760 is recursive manufacturing. And to be super clear right up front, 25 00:01:22.840 --> 00:01:25.920 we aren't talking about like a robotic arm welding a 26 00:01:26.000 --> 00:01:28.159 car door. We've had that for forty years. 27 00:01:28.200 --> 00:01:31.280 That's old news, right, That is just standard industrial automation. 28 00:01:32.359 --> 00:01:35.760 What we are analyzing today is a massive transition. We 29 00:01:35.840 --> 00:01:40.319 are talking about humanoid robots, machines that have our form factor, 30 00:01:40.359 --> 00:01:41.400 that look and move. 31 00:01:41.400 --> 00:01:43.879 Like us, standing right there on the line. 32 00:01:43.799 --> 00:01:47.239 Yes, standing on the assembly line, using highly dexterous hands 33 00:01:47.239 --> 00:01:51.480 to wire, assemble, and package the very next generation of themselves. 34 00:01:51.760 --> 00:01:53.640 It is a closed loop production system. 35 00:01:54.040 --> 00:01:56.719 Closed loop. That's a phrase that keeps popping up in 36 00:01:56.760 --> 00:01:59.519 the research we're going through. It feels like the holy 37 00:01:59.599 --> 00:02:01.319 grail of industrial engineering. 38 00:02:01.400 --> 00:02:04.359 It absolutely is, yeah, because once you manage to close 39 00:02:04.400 --> 00:02:08.360 that loop, you fundamentally detach production growth from the limitations 40 00:02:08.360 --> 00:02:09.319 of the human workforce. 41 00:02:09.680 --> 00:02:12.360 Before we get into the heavy hitters like Tesla and 42 00:02:12.400 --> 00:02:16.520 Figure AI and the economic implications which are honestly mind bending. 43 00:02:16.560 --> 00:02:18.560 I want to clear up a definition. Sure, when we 44 00:02:18.599 --> 00:02:21.840 say robots building robots, my brain and probably your brain 45 00:02:21.919 --> 00:02:24.439 listening to this goes straight to the matrix or those 46 00:02:24.439 --> 00:02:28.080 self replicating space probes. Yeah, the Von Neumann probes. 47 00:02:27.759 --> 00:02:30.199 Right right, the classic Von Neumann probe. It's a great 48 00:02:30.199 --> 00:02:33.159 Sci Fi trope. A machine lands on an asteroid, It 49 00:02:33.240 --> 00:02:37.039 mines the raw iron, smelts, it builds a factory and 50 00:02:37.120 --> 00:02:40.159 replicates itself with literally zero human input. 51 00:02:40.439 --> 00:02:40.680 Right. 52 00:02:41.039 --> 00:02:43.439 I want to be very precise here. That is not 53 00:02:43.479 --> 00:02:45.240 what is occurring in twenty twenty six. 54 00:02:45.319 --> 00:02:47.960 Okay, so we aren't quite there yet. We don't have 55 00:02:48.120 --> 00:02:51.319 robots wandering out into the wilderness to build entirely new 56 00:02:51.360 --> 00:02:52.319 cities from scratch. 57 00:02:52.520 --> 00:02:56.639 No, that is autonomous self replication in an unstructured environment 58 00:02:57.039 --> 00:03:02.680 that involves mining, refining, complex chemical processing. That's decades away. 59 00:03:03.039 --> 00:03:04.280 So what are we seeing today? 60 00:03:04.439 --> 00:03:07.439 What we are seeing today is recursive manufacturing and strictly 61 00:03:07.439 --> 00:03:11.719 controlled settings. Imagine a highly structured factory environment. The lights 62 00:03:11.759 --> 00:03:14.080 are on, the floors are flat, the parts are delivered 63 00:03:14.080 --> 00:03:16.960 to specific bins by automated logistics systems. 64 00:03:17.000 --> 00:03:18.400 But the actual labor. 65 00:03:18.159 --> 00:03:20.919 Part, right, the labor, the actual act of snapping the 66 00:03:20.919 --> 00:03:24.280 plastic housing together, threading a delicate wire through a joint, 67 00:03:24.520 --> 00:03:27.120 testing the motor. That physical labor is shifting from human 68 00:03:27.159 --> 00:03:28.439 hands to robotic Hans. 69 00:03:28.800 --> 00:03:31.680 So the distinction is really the environment and the scope. 70 00:03:32.039 --> 00:03:34.639 It's not a robot mining or in a cave. It's 71 00:03:34.680 --> 00:03:37.120 a robot standing at a workstation doing what a human 72 00:03:37.159 --> 00:03:38.879 worker used to do just a couple of years. 73 00:03:38.719 --> 00:03:43.080 Ago, exactly. But the economic impact of that is almost 74 00:03:43.120 --> 00:03:45.560 exactly the same as the Sci Fi version. How so, 75 00:03:45.960 --> 00:03:49.039 because once you replace the human labor in the production 76 00:03:49.159 --> 00:03:53.439 of the labor force itself, you fundamentally change the entire 77 00:03:53.520 --> 00:03:56.840 economic equation. You stop being limited by how many people 78 00:03:56.919 --> 00:03:59.719 you can hire, or housing shortages in the area, or 79 00:03:59.759 --> 00:04:00.719 wage inflation. 80 00:04:00.879 --> 00:04:01.919 You just need materials. 81 00:04:02.000 --> 00:04:04.759 You start being limited only by how much silicon and 82 00:04:04.800 --> 00:04:06.840 steel and electricity you can buy. 83 00:04:07.199 --> 00:04:11.719 That's such a heavy concept, detaching production from the human population. Usually, 84 00:04:11.800 --> 00:04:13.599 I mean, if you want to double your factory output, 85 00:04:13.680 --> 00:04:16.360 you need to hire double the workers, and humans are scarce. 86 00:04:16.519 --> 00:04:19.680 Humans take eighteen years to reach working age, They need sleep, 87 00:04:19.839 --> 00:04:22.399 they get injured, they eventually need to retire, They. 88 00:04:22.279 --> 00:04:23.839 Have lives outside the factory floor. 89 00:04:23.959 --> 00:04:27.319 Precisely, but if the product, the robot can build the product. 90 00:04:27.560 --> 00:04:30.720 The only constraints left are raw materials and energy. 91 00:04:30.680 --> 00:04:33.360 Which brings us to the speed of all this. The 92 00:04:33.399 --> 00:04:36.879 sources we're looking at. Describe this theoretical framework called the 93 00:04:36.920 --> 00:04:40.480 triple exponential. This is something Elon Musk and a few 94 00:04:40.480 --> 00:04:43.800 other industry leaders have been harping on lately. Break this 95 00:04:43.920 --> 00:04:45.800 down for us. Why triple So. 96 00:04:45.839 --> 00:04:50.160 Most technology grows on a single exponential curve, you know 97 00:04:50.399 --> 00:04:53.439 More's law. For example, computer chips get faster and cheaper 98 00:04:53.480 --> 00:04:57.439 every two years. Right, But recursive robotics is writing three 99 00:04:57.839 --> 00:05:00.720 simultaneous waves that interlock with the each other. If you 100 00:05:00.720 --> 00:05:03.120 miss one, the whole system stalls out. But if you 101 00:05:03.199 --> 00:05:05.240 hit all three at the same time, you get this 102 00:05:05.319 --> 00:05:06.120 vertical takeoff. 103 00:05:06.360 --> 00:05:08.720 Okay, walk us through these three factors. What is the 104 00:05:08.720 --> 00:05:09.199 first one? 105 00:05:09.240 --> 00:05:12.920 The first one is AI software intelligence. In twenty twenty six. 106 00:05:12.959 --> 00:05:15.639 We aren't just running you know, standard hard coded logic. 107 00:05:15.959 --> 00:05:19.079 We are running massive foundation models. We're seeing these scaling 108 00:05:19.160 --> 00:05:21.560 laws play out in real time scaling laws. 109 00:05:21.600 --> 00:05:23.920 We hear that in the AI space constantly. That basically 110 00:05:24.000 --> 00:05:26.519 means if you dump more data and more compute into 111 00:05:26.519 --> 00:05:29.040 the system, it gets smarter at a very predictable rate. 112 00:05:29.160 --> 00:05:34.720 Right, yes, exactly, But specifically for robotics, it's about multimodal training. 113 00:05:35.319 --> 00:05:38.079 The AI isn't just reading text from the Internet anymore. 114 00:05:38.399 --> 00:05:43.160 It's processing live video force feedback from its fingertips. Proprioception. 115 00:05:43.560 --> 00:05:47.519 Wait, proprioception, what is that? Exactly? In this context, that's. 116 00:05:47.360 --> 00:05:50.000 The robot sense of where its limbs are in physical space. 117 00:05:50.600 --> 00:05:52.360 Just like you can close your eyes and still touch 118 00:05:52.360 --> 00:05:55.399 your nose. The robot knows exactly the angle and tension 119 00:05:55.480 --> 00:05:58.879 of every joint. Add spatial awareness to that, and the 120 00:05:58.920 --> 00:06:01.839 software is evolving to understand the physical world, not just 121 00:06:01.879 --> 00:06:02.680 the digital one. 122 00:06:02.800 --> 00:06:05.439 So factor one is the brain. The brain is getting smarter, 123 00:06:05.600 --> 00:06:08.000 way faster than we expected. What's factor two? 124 00:06:08.399 --> 00:06:11.399 Compute power. The brain needs a body, sure, but it 125 00:06:11.399 --> 00:06:14.399 also needs calories or in this case, watts and flops. 126 00:06:15.040 --> 00:06:19.040 We are seeing these companies, specifically Tesla and Figure, building 127 00:06:19.120 --> 00:06:21.680 massive in house supercomputer clusters. 128 00:06:21.839 --> 00:06:24.759 This is the massive infrastructure piece we keep hearing about. 129 00:06:24.600 --> 00:06:27.319 Right, because you can have the smartest AI architecture in 130 00:06:27.360 --> 00:06:29.360 the world, but if you don't have the next generation 131 00:06:29.439 --> 00:06:31.600 chips to run the inference in real time. 132 00:06:31.639 --> 00:06:33.000 So it can react fast enough. 133 00:06:33.120 --> 00:06:36.360 Exactly, so the robot can react to a falling screw 134 00:06:36.839 --> 00:06:41.920 in milliseconds. Without that compute it's useless. The compute density 135 00:06:41.959 --> 00:06:46.399 in these dedicated clusters is skyrocketing. They are literally building 136 00:06:46.439 --> 00:06:50.959 the physical capacity to think for millions of robots simultaneously. 137 00:06:51.199 --> 00:06:55.319 And the third factor in this triple exponential hardware dexterity. 138 00:06:55.759 --> 00:06:59.079 This is the physical machine itself, the actuators, the motors, 139 00:06:59.319 --> 00:07:01.079 the tactiles, answers on the hands. 140 00:07:00.800 --> 00:07:02.519 The muscles in the nerves exactly. 141 00:07:03.040 --> 00:07:06.360 Think back just a few years ago. Robots were incredibly clumsy. 142 00:07:06.879 --> 00:07:09.360 They could lift a heavy car chassis with ease, but 143 00:07:09.399 --> 00:07:12.160 they couldn't button a shirt. They couldn't handle a flexible 144 00:07:12.199 --> 00:07:13.000 floppy wire. 145 00:07:13.120 --> 00:07:15.120 They were strong, but dumb and stiff. 146 00:07:16.000 --> 00:07:19.279 Now we have custom actuators that offer human level precision. 147 00:07:19.480 --> 00:07:22.240 When you combine those three things, a smarter brain, a 148 00:07:22.279 --> 00:07:25.160 faster processor, and a highly dexterous body, you get the 149 00:07:25.199 --> 00:07:26.199 recursive exponential. 150 00:07:26.279 --> 00:07:28.279 It's a multiplicative effect. They don't just add up, they 151 00:07:28.319 --> 00:07:29.160 multiply each other. 152 00:07:29.480 --> 00:07:32.519 It creates a compounding loop. Better robots are able to 153 00:07:32.519 --> 00:07:35.639 build better factories, which in turn build even better robots. 154 00:07:36.199 --> 00:07:39.079 Must calls it pushing the margin. The goal is to 155 00:07:39.120 --> 00:07:43.759 push the robot into superhuman performance for economically valuable tasks. 156 00:07:44.439 --> 00:07:46.839 The timeline in our notes here is fascinating. It says 157 00:07:46.879 --> 00:07:49.600 we are currently right now, in early twenty twenty six, 158 00:07:49.759 --> 00:07:51.879 in a phase of partial integration. 159 00:07:52.240 --> 00:07:55.079 I think that's a fair assessment. We aren't one hundred 160 00:07:55.079 --> 00:07:58.720 percent automation yet. If you walk into these manufacturing facilities today, 161 00:07:58.759 --> 00:08:00.319 you still see people walking around. 162 00:08:00.519 --> 00:08:03.959 But the projection for a full loop where robots are 163 00:08:03.959 --> 00:08:06.399 the primary force building the robots is just a three 164 00:08:06.399 --> 00:08:09.480 to five year window. That feels incredibly fast to me. 165 00:08:09.639 --> 00:08:13.480 It's lightning fast in industrial terms. Usually retooling a major 166 00:08:13.519 --> 00:08:17.040 factory takes a decade. We're seeing physical literations happen in 167 00:08:17.199 --> 00:08:20.480 months now. The recursion is already active on the lines. 168 00:08:21.040 --> 00:08:24.079 Now it's just a game of percentages. How So, today 169 00:08:24.120 --> 00:08:26.639 maybe the robots do twenty percent of the assembly. Next 170 00:08:26.680 --> 00:08:29.240 year it'll be forty percent, then eighty percent. 171 00:08:29.600 --> 00:08:33.120 Let's get into some specific examples. Theories are great, but 172 00:08:33.200 --> 00:08:35.080 I want to know what's actually happening on the ground. 173 00:08:35.440 --> 00:08:38.440 We have a case study here on Figure AI. They 174 00:08:38.440 --> 00:08:42.039 seem to be the ones really pushing this recursive strategy explicitly. 175 00:08:42.320 --> 00:08:45.279 Figure is absolutely the poster child for this right now. 176 00:08:45.320 --> 00:08:48.240 They aren't being subtle about it at all. Their entire 177 00:08:48.320 --> 00:08:50.720 business model is built around closing this loop. 178 00:08:50.799 --> 00:08:54.720 They have a facility called BOTQ in California. What exactly 179 00:08:54.799 --> 00:08:56.720 goes on inside botq Q is. 180 00:08:56.720 --> 00:08:59.440 Designed for one thing, and that is scale. It's not 181 00:08:59.519 --> 00:09:01.879 a standard assembly plant where you have people standing at 182 00:09:01.879 --> 00:09:04.840 stations with screwdrivers and torque wrenches. It was built from 183 00:09:04.879 --> 00:09:08.120 the ground up to transition from making thousands of units 184 00:09:08.120 --> 00:09:12.000 a year to a highly aggressive, high volume output. 185 00:09:11.799 --> 00:09:14.200 Using the robots themselves as the primary workers. 186 00:09:14.399 --> 00:09:15.120 Exactly so. 187 00:09:15.240 --> 00:09:18.279 Brett Adcock, the CEO over at Figure has this roadmap 188 00:09:18.320 --> 00:09:19.759 involving the Figure O three. 189 00:09:19.639 --> 00:09:23.080 Model right, the Figure O three and specifically the Helix 190 00:09:23.120 --> 00:09:24.240 AI stack that runs it. 191 00:09:24.440 --> 00:09:26.840 Explain this Helix stack to us? Is that just their 192 00:09:26.919 --> 00:09:28.559 version of an operating system. 193 00:09:28.720 --> 00:09:31.919 Think of Helix as a collective intelligence. This is where 194 00:09:31.960 --> 00:09:35.080 the math gets really interesting. Every single time a figure 195 00:09:35.159 --> 00:09:40.159 robot moves, it's not just executing code, it's learning. This 196 00:09:40.279 --> 00:09:42.600 leads to what they call the data flywheel. This is 197 00:09:43.039 --> 00:09:47.960 easily the most critical concept for understanding figures massive valuation right. 198 00:09:47.840 --> 00:09:50.799 Now, the flywheel. We hear that term in business school 199 00:09:50.840 --> 00:09:52.759 all the time, but how does it apply to a 200 00:09:52.759 --> 00:09:55.360 physical robot assembling say a battery pack. 201 00:09:55.440 --> 00:09:57.840 Okay, let's play it out. Imagine a Figure O three 202 00:09:58.039 --> 00:10:01.600 robot on the line at BOTQ. It's tasked with wiring 203 00:10:01.639 --> 00:10:04.279 a subassembly. Let's say it struggles a little bit on 204 00:10:04.320 --> 00:10:07.240 its first try. Maybe it fumbles a tiny connector or 205 00:10:07.279 --> 00:10:10.360 the wire is bent in an unpredictable way. It tries, 206 00:10:10.480 --> 00:10:12.759 it adjusts its grip, it tries again, and eventually it 207 00:10:12.799 --> 00:10:13.360 gets it right. 208 00:10:13.519 --> 00:10:14.559 It figures the problem out. 209 00:10:14.639 --> 00:10:16.600 It figures it out. But here is the magic of 210 00:10:16.639 --> 00:10:20.279 the flywheel. That data the struggle, the micro corrections, the 211 00:10:20.399 --> 00:10:23.799 video feed, the exact force it felt on its fingertips. 212 00:10:24.159 --> 00:10:27.120 All of that is immediately uploaded to the central helix model. 213 00:10:27.200 --> 00:10:29.440 It's essentially recording the entire experience. 214 00:10:29.799 --> 00:10:33.200 Yes, and then that experience is used to train the 215 00:10:33.240 --> 00:10:36.639 next version of the model. That update is then pushed 216 00:10:36.679 --> 00:10:40.960 out to the entire fleet. Wow. So suddenly every single 217 00:10:41.039 --> 00:10:43.960 robot in the factory and every robot working out in 218 00:10:44.000 --> 00:10:48.080 the field knows exactly how to handle that specific connector 219 00:10:48.440 --> 00:10:49.639 perfectly on the first try. 220 00:10:49.799 --> 00:10:52.919 So if one robot learns something hard, they all learn 221 00:10:52.960 --> 00:10:54.799 it instantly, instantly. 222 00:10:54.600 --> 00:10:57.519 And here is where the recursive part kicks into high gear. 223 00:10:58.320 --> 00:11:01.320 The bitter they get it manufacturing the faster and cheaper 224 00:11:01.399 --> 00:11:03.519 figure can build more robots, which means. 225 00:11:03.320 --> 00:11:05.840 They can put more robots out in the world, collecting 226 00:11:05.879 --> 00:11:06.919 even more real. 227 00:11:06.720 --> 00:11:10.159 World data, which makes the central model smarter, which makes 228 00:11:10.159 --> 00:11:13.519 the manufacturing even better. Round and round, the flywheel goes 229 00:11:13.559 --> 00:11:14.200 gaining speed. 230 00:11:14.279 --> 00:11:16.720 That really explains the valuation numbers we are seeing. The 231 00:11:16.720 --> 00:11:19.440 notes here say Figure AI is valued at over thirty 232 00:11:19.519 --> 00:11:22.639 nine billion dollars. That's a massive number for a company 233 00:11:22.639 --> 00:11:24.519 that just a few years ago didn't even exist. 234 00:11:24.799 --> 00:11:27.840 Investors aren't really betting on the robot's figure is selling today, 235 00:11:28.080 --> 00:11:31.159 They are betting entirely on the flywheel. They see the 236 00:11:31.240 --> 00:11:37.360 direct correlation between this recursive strategy and exponential scaling. If 237 00:11:37.399 --> 00:11:39.879 you can make the robot that builds the robot, you 238 00:11:40.000 --> 00:11:42.559 essentially win the future manufacturing economy. 239 00:11:42.679 --> 00:11:45.600 And the vision here is scaling to billions of units, 240 00:11:45.840 --> 00:11:47.399 not millions billions. 241 00:11:47.840 --> 00:11:49.879 Ed Gock has been very clear on this point. He 242 00:11:49.960 --> 00:11:53.200 wants these things in factories and households and eventually helping 243 00:11:53.200 --> 00:11:54.159 with space colonization. 244 00:11:54.360 --> 00:11:56.279 Okay, let's pivot to the other giant in the room. 245 00:11:56.480 --> 00:11:59.440 You can't talk about manufacturing scale without talking about Tesla. 246 00:11:59.519 --> 00:12:03.240 Tesla's optimist program. It's a very different beast from Figure 247 00:12:03.440 --> 00:12:06.399 mostly because of their intense vertical integration. 248 00:12:06.159 --> 00:12:09.360 Right because figure is laser focused on just the robot. 249 00:12:09.639 --> 00:12:12.399 Tesla is already a car company, an energy company, and 250 00:12:12.480 --> 00:12:16.320 a massive AI company. What is the actual operational status 251 00:12:16.360 --> 00:12:19.360 of Optimists right now in February or twenty twenty six. 252 00:12:19.679 --> 00:12:22.519 We are looking closely at their Fremont pilot production lines. 253 00:12:23.240 --> 00:12:26.399 Right now they are assembling generation two and some early 254 00:12:26.480 --> 00:12:30.519 generation three prototypes. But if you look at the ground reports, 255 00:12:31.360 --> 00:12:34.360 the vibe there is quite different from a fully automated 256 00:12:34.480 --> 00:12:37.360 lights outline. It's very much human in the loop right now. 257 00:12:37.519 --> 00:12:39.000 So it's definitely not a ghost. 258 00:12:38.759 --> 00:12:41.720 Factory yet, No, not yet. You walk that line and 259 00:12:41.759 --> 00:12:44.960 you see teams of human engineers working right alongside the 260 00:12:45.000 --> 00:12:49.480 automated robotic processes. They're tweaking the actuators, adjusting tension in 261 00:12:49.519 --> 00:12:52.919 the wiring harnesses, calibrating the sensors by hand. It's a 262 00:12:53.039 --> 00:12:54.080 very close collaboration. 263 00:12:54.519 --> 00:12:57.159 Musk has said they are currently useful for simple tasks. 264 00:12:57.519 --> 00:13:00.600 That's the official line today, simple tasks now, but with 265 00:13:00.639 --> 00:13:04.240 a very aggressive trajectory toward complex autonomy by the end 266 00:13:04.279 --> 00:13:07.399 of twenty twenty six. But the hardware design itself is 267 00:13:07.440 --> 00:13:11.000 where Tesla really flexes its manufacturing muscles biomimicry. 268 00:13:11.159 --> 00:13:11.960 Right, We've seen a. 269 00:13:11.879 --> 00:13:14.399 Lot about that. Yes, Optimis is designed to mimic the 270 00:13:14.480 --> 00:13:17.879 human form biologically more than almost any other robot on 271 00:13:17.919 --> 00:13:20.879 the market. The specific motors they design are meant to 272 00:13:20.919 --> 00:13:24.360 act like biological muscles. The electronics are incredibly compact to 273 00:13:24.399 --> 00:13:25.399 fit inside the frame. 274 00:13:25.720 --> 00:13:28.600 Why is that so important? I mean, why not just 275 00:13:28.679 --> 00:13:31.320 build a robot that looks like a forklift or a 276 00:13:31.440 --> 00:13:33.559 robotic arm on wheels because. 277 00:13:33.320 --> 00:13:36.720 The world is already built for humans. The modern factory 278 00:13:36.759 --> 00:13:39.879 is built for humans. If you want a truly general 279 00:13:39.919 --> 00:13:42.440 purpose robot, it needs to be able to fit through 280 00:13:42.440 --> 00:13:45.639 a standard door frame. It needs to use standard hand tools, 281 00:13:45.639 --> 00:13:48.000 It needs to walk up stairs that were built for 282 00:13:48.120 --> 00:13:49.240 size ten shoes. 283 00:13:49.399 --> 00:13:50.279 Ah. I see. 284 00:13:50.360 --> 00:13:53.080 The whole design intent is to create a general purpose 285 00:13:53.159 --> 00:13:56.559 humanoid that can seamlessly slot into a world that was 286 00:13:56.720 --> 00:13:57.559 entirely built. 287 00:13:57.399 --> 00:14:00.000 For us, including the factory that builds the robot itself. 288 00:14:00.360 --> 00:14:03.720 Exactly. If the Tesla factory is already designed for people 289 00:14:03.759 --> 00:14:06.200 to walk around in, the robot needs to move and 290 00:14:06.279 --> 00:14:07.279 operate like a person. 291 00:14:07.399 --> 00:14:08.879 I want to ask about this thing in the notes 292 00:14:08.919 --> 00:14:11.759 called the Optimus Academy It sounds like a prep school 293 00:14:11.840 --> 00:14:14.039 from a sci fi novel, but the research says it's 294 00:14:14.120 --> 00:14:16.559 absolutely crucial for their data strategy. 295 00:14:16.840 --> 00:14:19.000 It sort of is a prep school. Honestly. This is 296 00:14:19.039 --> 00:14:21.080 their massive simulation engine. 297 00:14:21.480 --> 00:14:22.000 You see. 298 00:14:22.120 --> 00:14:24.879 You can't train a robot entirely in the real world. 299 00:14:25.039 --> 00:14:26.360 It's way too slow. 300 00:14:26.200 --> 00:14:28.879 And dangerous, I imagine, and very dangerous. 301 00:14:29.600 --> 00:14:32.120 If a robot loses its balance and falls over in 302 00:14:32.159 --> 00:14:35.279 the real world, you just broke a ten thousand dollars 303 00:14:35.519 --> 00:14:38.559 custom actuator. If it falls over in a simulation, you 304 00:14:38.720 --> 00:14:40.240 just reset the code and try again. 305 00:14:40.399 --> 00:14:42.960 So they essentially put them in the matrix to learn. 306 00:14:43.279 --> 00:14:47.679 Essentially, Yes, they have thousands upon thousands of virtual Optimist 307 00:14:47.720 --> 00:14:51.120 units running in these hyperreal simulations twenty four hours a day. 308 00:14:51.720 --> 00:14:56.200 They are practicing fine dexterity balance and complex object manipulation 309 00:14:56.600 --> 00:14:59.080 in a totally physics compliant virtual world. 310 00:14:59.200 --> 00:15:01.399 They call it self. Play in the documents, right. 311 00:15:01.799 --> 00:15:04.720 They play out these scenarios millions of times. Pick up 312 00:15:04.759 --> 00:15:07.840 the box, plug in the high voltage cable, walk over 313 00:15:07.840 --> 00:15:11.559 the uneven concrete floor. Then they take that brain the 314 00:15:11.559 --> 00:15:13.919 neural net weights they developed in the simulation, and they 315 00:15:13.960 --> 00:15:17.240 download it straight into the physical robot. This accelerates the 316 00:15:17.320 --> 00:15:20.919 learning curve massively without the risk of breaking expensive hardware. 317 00:15:21.120 --> 00:15:23.080 So you have the real world data coming from the 318 00:15:23.080 --> 00:15:26.559 Fremont pilot line combining with the massive simulation data from 319 00:15:26.600 --> 00:15:28.000 the Optimus academy. 320 00:15:27.799 --> 00:15:31.320 And that combination is what triggers the recursion tipping point. 321 00:15:31.440 --> 00:15:34.320 Once the internal fleets are capable enough to bootstrap their 322 00:15:34.320 --> 00:15:38.080 own output, the data looks tighten up significantly, becomes a 323 00:15:38.120 --> 00:15:39.159 blur of improvement. 324 00:15:39.440 --> 00:15:42.080 What's the actual production roadmap look like? When can I 325 00:15:42.080 --> 00:15:43.879 actually buy one to fold my laundry? 326 00:15:44.600 --> 00:15:46.840 Well, I wouldn't get your checkbook out just yet. Mid 327 00:15:46.879 --> 00:15:50.120 twenty twenty six is their target for repurposing lines for 328 00:15:50.240 --> 00:15:53.600 low volume internal use. Tesla wants to be its own, 329 00:15:53.639 --> 00:15:54.919 first and biggest customer. 330 00:15:55.080 --> 00:15:57.320 That makes total sense, eat your own dog food, as 331 00:15:57.360 --> 00:15:57.879 they say in. 332 00:15:57.879 --> 00:16:01.799 Software, or use your own robot. Then they are projecting 333 00:16:02.039 --> 00:16:05.919 ramping up to actual customer deliveries in late twenty twenty 334 00:16:05.960 --> 00:16:09.559 six or early twenty twenty seven. The volume targets they've 335 00:16:09.559 --> 00:16:12.360 set are staggering. Fremont is looking at fifty to one 336 00:16:12.440 --> 00:16:13.320 hundred thousand. 337 00:16:13.120 --> 00:16:15.759 Units initially, and the Giga Texas facility. 338 00:16:15.480 --> 00:16:18.440 Millions long term they are eyeing millions of units out 339 00:16:18.480 --> 00:16:20.960 of Texas. And this brings us to the cost factor. 340 00:16:21.120 --> 00:16:23.000 This is the part that blows my mind every single 341 00:16:23.000 --> 00:16:24.559 time I look at the economic church, the. 342 00:16:24.559 --> 00:16:26.720 Cost collapse phenomenon. Let's dig into that. 343 00:16:27.039 --> 00:16:30.080 As recursion scales up, the cost to produce each unit 344 00:16:30.159 --> 00:16:33.240 is projected to absolutely plummet. We are talking about a 345 00:16:33.279 --> 00:16:37.159 target range of twenty to thirty thousand dollars per robot. 346 00:16:37.360 --> 00:16:40.360 That is, I mean, that's comparable to a pretty cheap 347 00:16:40.440 --> 00:16:41.159 compact car. 348 00:16:41.320 --> 00:16:44.159 Actually it's cheaper than most new cars today. And remember 349 00:16:44.200 --> 00:16:46.840 a car sits idle in your driveway ninety percent of 350 00:16:46.879 --> 00:16:50.559 the time. This robot works twenty four to seven. The 351 00:16:50.639 --> 00:16:54.279 value proposition is just off the charts. The cost collapse 352 00:16:54.320 --> 00:16:57.200 happens because you are permanently removing human labor from the 353 00:16:57.240 --> 00:16:58.039 production equation. 354 00:16:58.279 --> 00:17:01.320 If the robot that is building the row doesn't draw salary, 355 00:17:01.360 --> 00:17:04.400 doesn't need health insurance, the cost of the new robot 356 00:17:04.480 --> 00:17:07.559 basically approaches the raw cost of the materials. 357 00:17:07.039 --> 00:17:09.359 And the energy to run the machines. That's it, that's 358 00:17:09.400 --> 00:17:11.880 the bottom line. If you take human wages completely out 359 00:17:11.880 --> 00:17:14.039 of the cost of goods sold, you are left with 360 00:17:14.160 --> 00:17:15.720 baseline commodity prices. 361 00:17:16.240 --> 00:17:18.680 We've talked a lot about California and Texas, but we 362 00:17:18.799 --> 00:17:21.559 really need to look east. The ecosystem in China is 363 00:17:21.599 --> 00:17:23.599 moving at an absolute breakneck pace. 364 00:17:23.960 --> 00:17:28.279 Relentless. That's the only word for the Chinese industrial sector 365 00:17:28.400 --> 00:17:31.720 right now. If the US companies are innovating heavily on 366 00:17:31.759 --> 00:17:35.000 the software models, China is innovating on the sheer scale 367 00:17:35.039 --> 00:17:35.599 of deployment. 368 00:17:35.759 --> 00:17:38.160 Let's talk about Unitree. They're a massive player here. 369 00:17:38.519 --> 00:17:42.200 Unitree has already deployed their G one humanoids on actual 370 00:17:42.240 --> 00:17:46.559 factory production lines in Shenzen. It isn't a controlled lab environment. 371 00:17:46.599 --> 00:17:49.519 This is Shenzen. It's the manufacturing capital of the world. 372 00:17:49.720 --> 00:17:52.240 And they are using this system called the Uniform LM 373 00:17:52.440 --> 00:17:53.480 X one zero. 374 00:17:53.640 --> 00:17:57.799 That's their proprietary embodied AI architecture. And the visual evidence 375 00:17:57.839 --> 00:18:00.680 we have from these factory floors is really strike. We 376 00:18:00.720 --> 00:18:04.000 see complex by manual manipulation, meaning the robot is using 377 00:18:04.000 --> 00:18:08.519 two hands together cooperatively to screw in tiny components, assemble subassemblies, 378 00:18:08.640 --> 00:18:09.920 and wire up actuators. 379 00:18:10.039 --> 00:18:13.079 Wiring is notoriously hard for robots, isn't it. It's floppy, 380 00:18:13.119 --> 00:18:14.880 it bends, it's unpredictable, it. 381 00:18:14.960 --> 00:18:17.599 Is incredibly difficult. A rigid metal part is easy for 382 00:18:17.599 --> 00:18:19.720 a robot to grab. A wire moves out of the way. 383 00:18:20.000 --> 00:18:22.960 Seeing a robot handle flexible cables with that level of 384 00:18:23.000 --> 00:18:26.599 precision is a major industry milestone. It signals a hard 385 00:18:26.599 --> 00:18:29.680 transition from cool tech demos to actual operational production. 386 00:18:29.880 --> 00:18:31.039 And what about their pricing. 387 00:18:31.519 --> 00:18:35.279 China always competes fiercely on cost. Currently, that level of 388 00:18:35.319 --> 00:18:38.559 capability is priced around ninety to one hundred thousand dollars, 389 00:18:39.000 --> 00:18:42.279 but they're scaling is pushing that number down extremely fast. 390 00:18:42.640 --> 00:18:46.039 They are actively projecting shipments of ten to twenty thousand 391 00:18:46.160 --> 00:18:47.640 units in twenty twenty six. 392 00:18:47.720 --> 00:18:50.400 Low I have to mention the Spring Festival showcases. We 393 00:18:50.440 --> 00:18:53.359 all saw online the videos of the robots doing backflips 394 00:18:53.359 --> 00:18:55.720 and kung fu. It felt a bit like a circus 395 00:18:55.720 --> 00:18:56.160 act to me. 396 00:18:56.480 --> 00:18:59.079 It's fantastic marketing, isn't it. But it serves a very 397 00:18:59.119 --> 00:19:02.319 real dual per pose. Yes, it looks incredibly cool on 398 00:19:02.359 --> 00:19:05.039 social media and gets the public excited about the tech, 399 00:19:05.480 --> 00:19:07.359 but it's also an extreme stress test. 400 00:19:07.720 --> 00:19:09.720 How so a backflip isn't building a car. 401 00:19:09.880 --> 00:19:11.960