WEBVTT 1 00:00:00.040 --> 00:00:02.720 So I want you to picture this. You're at a cozy, 2 00:00:02.759 --> 00:00:06.000 little Swiss shellet, right, and it's the dead of winter. 3 00:00:06.200 --> 00:00:07.080 That sounds nice. 4 00:00:07.240 --> 00:00:13.039 Yeah, Well, outside there's this massive snowstorm howling. The temperature 5 00:00:13.039 --> 00:00:16.199 has plummeted to I don't know, like ten degrees fahrenheit. 6 00:00:16.359 --> 00:00:18.839 Wow, okay, maybe not so nice, right, And. 7 00:00:18.760 --> 00:00:22.760 Hanging from the porch is this tiny, just incredibly cramped 8 00:00:23.079 --> 00:00:27.359 wooden birdhouse. You fill it with seed, and within minutes 9 00:00:27.399 --> 00:00:30.600 you've got like twenty small birds darting in and out 10 00:00:30.640 --> 00:00:35.079 of that confined space. And they are flying flawlessly. 11 00:00:34.479 --> 00:00:36.039 Which is wild. When you actually watch it. 12 00:00:36.000 --> 00:00:38.840 Happen, it is. I mean, it looks completely chaotic, but 13 00:00:39.079 --> 00:00:42.359 they're grabbing food, maneuvering around each other, navigating these crazy 14 00:00:42.399 --> 00:00:45.840 wind gusts, and there's not a single midair collision. It's 15 00:00:45.880 --> 00:00:49.039 just pure effortless biological navigation. 16 00:00:49.960 --> 00:00:50.719 Yeah, exactly. 17 00:00:51.079 --> 00:00:53.880 So now I want you to imagine taking twenty advanced 18 00:00:53.920 --> 00:00:57.399 AI controlled drones and programming them to navigate that exact 19 00:00:57.439 --> 00:01:00.479 same birdhouse. Oh boy, yeah, at the exact vact same 20 00:01:00.479 --> 00:01:04.359 speed in that weather, but without a centralized flight controller 21 00:01:04.400 --> 00:01:06.760 coordinating every single move from the outside. 22 00:01:06.879 --> 00:01:08.599 Well, I mean you would end up with a pile 23 00:01:08.640 --> 00:01:12.120 of shattered plastic, right, just bent propellers buried in the 24 00:01:12.159 --> 00:01:15.400 snow within seconds. Total disaster, butter disaster. Right because the 25 00:01:15.400 --> 00:01:19.599 physical world is messy, It's full of friction, unpredictable wind gusts, 26 00:01:20.400 --> 00:01:25.280 constantly shifting variables. Getting artificial intelligence to operate with the 27 00:01:25.280 --> 00:01:30.439 collision free agility of those birds in physical reality, mind you, 28 00:01:30.840 --> 00:01:34.760 it's fundamentally different from having an algorithm, you know, generate 29 00:01:34.760 --> 00:01:37.760 a text document in the cloud. It is a monumental 30 00:01:37.799 --> 00:01:39.359 engineering challenge. 31 00:01:38.959 --> 00:01:41.599 And that challenge is exactly what we are dissecting today 32 00:01:41.599 --> 00:01:45.000 in this deep dive. We're pulling insights from Klaus Henning's 33 00:01:45.000 --> 00:01:50.719 book Game Changer AI. How artificial intelligence is transforming our world. 34 00:01:50.840 --> 00:01:52.480 It's a fantastic rate, it really is. 35 00:01:52.840 --> 00:01:54.519 And our mission for today is to kind of look 36 00:01:54.680 --> 00:01:58.359 past the endless hype cycle of you know, chatbots and 37 00:01:58.400 --> 00:02:02.560 image generators. We want to understand the true historical scale 38 00:02:02.640 --> 00:02:05.280 of the AI revolution as it breaks out of the 39 00:02:05.280 --> 00:02:06.000 digital realm. 40 00:02:06.120 --> 00:02:08.520 Right, because it's starting to give the physical objects around 41 00:02:08.599 --> 00:02:12.240 us a rudimentary form of consciousness, which is huge. 42 00:02:12.639 --> 00:02:16.639 Okay, let's unpack this because comparing a smartphone algorithm to 43 00:02:16.639 --> 00:02:20.560 an invention that literally fractured empires feels like a massive 44 00:02:20.560 --> 00:02:21.159 stretch to me. 45 00:02:21.319 --> 00:02:23.919 Well, wait, let me set the state for that comparison first, 46 00:02:24.520 --> 00:02:28.360 to grasp the scale. The author makes this historical comparison 47 00:02:28.400 --> 00:02:32.199 that reframes the whole conversation. Usually, when we talk about automation, we. 48 00:02:32.159 --> 00:02:34.479 Look back to the seventeen fifties, right right, the steam 49 00:02:34.479 --> 00:02:39.199 engine mechanizing production, kicking off the Industrial Revolution exactly. 50 00:02:39.599 --> 00:02:43.080 But the source material argues this shift is vastly more 51 00:02:43.120 --> 00:02:47.280 disruptive than the steam engine. The structural equivalent to what 52 00:02:47.319 --> 00:02:50.439 we're experiencing right now actually happened earlier. It happened in 53 00:02:50.520 --> 00:02:51.400 the fourteen. 54 00:02:51.080 --> 00:02:52.879 Fifties with Johannes Guttenberg. 55 00:02:53.000 --> 00:02:55.479 Yes, the invention of the movable type printing press. 56 00:02:55.639 --> 00:02:57.680 Right. So, like I said, comparing an app to the 57 00:02:57.680 --> 00:03:02.240 printing press seems I mean, before Berg, knowledge was completely hoarded, 58 00:03:02.719 --> 00:03:06.599 you had this image based monopoly where maybe what ten 59 00:03:06.639 --> 00:03:10.400 percent of monks could read, and the masses basically relied 60 00:03:10.479 --> 00:03:13.879 on stained glass windows or spoken sermons just to understand 61 00:03:13.879 --> 00:03:14.280 the world. 62 00:03:14.360 --> 00:03:16.520 Yeah, it was a severe bottleneck. 63 00:03:16.319 --> 00:03:20.280 And I get that Gutenberg shattered that bottleneck, which unleashed 64 00:03:20.400 --> 00:03:23.439 mass literacy and you know, eventually caused a century of 65 00:03:23.479 --> 00:03:26.719 war like the Thirty Years War. But are we really 66 00:03:26.800 --> 00:03:31.240 saying that an AI optimizing supply chains or like recommending 67 00:03:31.280 --> 00:03:33.919 a Spotify playlist is going to spark that level of 68 00:03:34.000 --> 00:03:39.360 societal upheaval? Is the argument that AI fundamentally changes who 69 00:03:39.400 --> 00:03:41.000 holds the power of knowledge. 70 00:03:41.120 --> 00:03:43.479 What's fascinating here is that it's not just about the 71 00:03:43.520 --> 00:03:48.000 power of knowledge. It's actually about the monopoly on decision making. Okay, 72 00:03:48.199 --> 00:03:51.280 how so, well, think about it. Gutenberg took the power 73 00:03:51.360 --> 00:03:55.000 of the written word away from a centralized authority. Yeah right, 74 00:03:55.159 --> 00:03:57.919 and he distributed it to the masses. AI is doing 75 00:03:57.960 --> 00:04:02.479 something similar, but with complex decision making. It's taking that 76 00:04:02.599 --> 00:04:05.840 power away from human beings entirely and distributing it to 77 00:04:05.879 --> 00:04:09.520 everyday objects. Oh wow, because historically we've always been the 78 00:04:09.560 --> 00:04:12.400 sole conscious actors in our physical environment. Yeah, a hammer 79 00:04:12.479 --> 00:04:14.439 only works when you swing it. A machine only runs 80 00:04:14.439 --> 00:04:16.639 when you pull lever or write a script. 81 00:04:16.759 --> 00:04:18.680 It needs us to initiate it, exactly. 82 00:04:19.399 --> 00:04:22.199 But now we are entering an era where the objects 83 00:04:22.240 --> 00:04:26.560 themselves assess their environment, they learn, and they make autonomous decisions, 84 00:04:26.600 --> 00:04:30.199 which changes everything it does. When that fundamental dynamic of 85 00:04:30.279 --> 00:04:33.920 power shifts. When humans are no longer the only intelligent 86 00:04:33.959 --> 00:04:37.560 actors in the room, society has to completely reorganize itself. 87 00:04:37.800 --> 00:04:40.680 I mean, if everyday objects are genuinely taking on that 88 00:04:40.759 --> 00:04:43.639 decision making capability, we really need to look at the 89 00:04:43.680 --> 00:04:45.879 mechanics of how they're learning to do it and the 90 00:04:45.920 --> 00:04:50.800 history of this technology. It actually borrows heavily from biology. Yeah, 91 00:04:50.920 --> 00:04:56.160 surprisingly organic right, specifically the complex parallel control loops found 92 00:04:56.240 --> 00:04:58.680 in a frog's hamstring reflex. 93 00:04:58.240 --> 00:04:59.720 Which sounds super weird at first. 94 00:05:00.000 --> 00:05:03.800 It does, but nature is incredibly messy and wasteful, yet 95 00:05:03.839 --> 00:05:08.279 it processes sensory information with extreme efficiency. So fifty years ago, 96 00:05:08.319 --> 00:05:11.120 computer scientists actually looked at how a frog's nerve cells 97 00:05:11.160 --> 00:05:15.120 operated firing and adjusting in real time, and they mapped 98 00:05:15.160 --> 00:05:17.079 that architecture digitally to create. 99 00:05:16.879 --> 00:05:20.399 Early neural networks. And for a long time, the hardware 100 00:05:20.399 --> 00:05:23.680 simply couldn't keep up with that biological blueprint. The theory 101 00:05:23.800 --> 00:05:26.519 was there, but the computers were too slow. But as 102 00:05:26.560 --> 00:05:29.959 computational power just exploded over the last decade, we moved 103 00:05:29.959 --> 00:05:33.199 from theory to practice, and the way these networks learn 104 00:05:33.839 --> 00:05:36.920 began to fundamentally shift. Yeah, we can track that shift 105 00:05:37.040 --> 00:05:39.079 really clearly by looking at the board. 106 00:05:38.839 --> 00:05:41.399 Game Go oh Go is fascinating. It's a game of 107 00:05:41.480 --> 00:05:45.160 pure intuition and pattern recognition. I think there are more 108 00:05:45.279 --> 00:05:48.720 possible board configurations than there are atoms in the observable 109 00:05:48.800 --> 00:05:49.680 universe exactly. 110 00:05:49.759 --> 00:05:53.879 It's astronomically complex. And in twenty sixteen, a system called 111 00:05:53.920 --> 00:05:55.879 AlphaGo defeated the human world. 112 00:05:55.759 --> 00:05:57.399 Champion, which was a massive deal. 113 00:05:57.639 --> 00:06:02.000 It was, but to achieve that, the programmers essentially had 114 00:06:02.040 --> 00:06:05.560 to digitize human history. They fed the system a database 115 00:06:05.600 --> 00:06:07.920 of like thirty million human. 116 00:06:07.720 --> 00:06:09.800 Moves, just massive amounts of data. 117 00:06:09.560 --> 00:06:12.240 Yeah, allowing it to study our patterns, and then they 118 00:06:12.279 --> 00:06:14.639 had it play itself over and over to refine those 119 00:06:14.639 --> 00:06:17.240 strategies based on what humans had done. 120 00:06:17.319 --> 00:06:21.680 Here's where it gets really interesting, because Alphago's victory was 121 00:06:21.680 --> 00:06:25.360 a big milestone, sure, but the architecture created just one 122 00:06:25.439 --> 00:06:27.959 year later was the actual paradigm shift. 123 00:06:28.040 --> 00:06:28.879 Oh absolutely. 124 00:06:29.040 --> 00:06:33.720 In twenty seventeen, DeepMind introduced Alpha Go zero, and Alpha 125 00:06:33.759 --> 00:06:38.439 Go zero completely obliterated the original AlphaGo. But the crucial 126 00:06:38.480 --> 00:06:42.160 detail here is the methodology. Alpha Go zero was not 127 00:06:42.360 --> 00:06:48.199 given thirty million human moves. It was given zero human data. None. 128 00:06:48.439 --> 00:06:50.600 It was simply programmed with the basic rules of the 129 00:06:50.639 --> 00:06:51.759 game and left alone. 130 00:06:51.800 --> 00:06:55.480 It relied entirely on reinforcement learning from a completely blank slate. 131 00:06:55.680 --> 00:06:57.720 Right, So think of it like this for everyone listening. 132 00:06:57.839 --> 00:07:01.839 The original AlphaGo was an incredibly diligent student who memorized 133 00:07:01.879 --> 00:07:05.319 every textbook ever written. Right. It analyzed every exam taken 134 00:07:05.360 --> 00:07:08.360 by past masters and used that inherited human knowledge to 135 00:07:08.399 --> 00:07:11.920 ace the test. A very smart mimic exactly Alpha Goo zero, 136 00:07:11.959 --> 00:07:13.600 on the other hand, is a student locked in a 137 00:07:13.600 --> 00:07:16.079 completely empty room. You slide a piece of paper under 138 00:07:16.079 --> 00:07:18.600 the door, which is the basic rules of addition and subtraction, 139 00:07:19.000 --> 00:07:23.759 and three days later, the student independently invents calculus. It 140 00:07:23.839 --> 00:07:27.800 played millions of games against itself, exploring the probability space 141 00:07:27.839 --> 00:07:31.560 without any human bias, and it discovered winning strategies that 142 00:07:31.639 --> 00:07:35.519 thousands of years of human go masters had never even conceived. 143 00:07:35.720 --> 00:07:40.360 In that removal of human bias, That is the critical takeaway. 144 00:07:40.439 --> 00:07:42.920 It learned from the environment of the game itself. Right. 145 00:07:43.000 --> 00:07:46.000 If we connect this to the bigger picture, this transition 146 00:07:46.439 --> 00:07:50.040 actually mirrors a massive shift happening right now in human education. 147 00:07:50.120 --> 00:07:51.279 Oh interesting, how so? 148 00:07:51.439 --> 00:07:54.480 Well, we're moving away from the traditional centralized model of 149 00:07:55.120 --> 00:07:57.399 frontal teaching. You know, where a teacher stands at a 150 00:07:57.439 --> 00:08:00.319 chalkboard and just pours standardized facts to a. 151 00:08:00.279 --> 00:08:03.319 Student's head, right, which is basically what feeding thirty million 152 00:08:03.399 --> 00:08:05.079 human moves into a database looks like. 153 00:08:05.199 --> 00:08:09.439 Exactly. Instead, there's this strong push towards self organized learning. 154 00:08:09.720 --> 00:08:14.120 I mean, you watch students today. They completely bypass static 155 00:08:14.240 --> 00:08:18.959 centralized resources like an outdated Brockhouse encyclopedia or a heavy textbook. 156 00:08:19.000 --> 00:08:21.279 Oh yeah, my kids don't even know what an encyclopedia is. 157 00:08:21.680 --> 00:08:26.199 Right. They dive into decentralized, dynamic platforms like YouTube or 158 00:08:26.360 --> 00:08:30.639 specialized networks like simple Club. They navigate the information themselves 159 00:08:31.079 --> 00:08:33.960 and form their own intuitive understanding of the material. 160 00:08:34.120 --> 00:08:35.200 That makes total sense. 161 00:08:35.360 --> 00:08:38.440 Yeah. So machine learning of all from being spoon fed 162 00:08:38.519 --> 00:08:43.759 human data to self directed exploration, and human learning models 163 00:08:43.759 --> 00:08:48.200 are desperately trying to adapt to that exact same decentralized approach. 164 00:08:48.399 --> 00:08:51.679 But the self directed exploration of a machine it isn't 165 00:08:51.720 --> 00:08:54.480 just confined to a board game, simulator or some lab 166 00:08:54.519 --> 00:08:57.759 experiment anymore. It's actively running on the devices in our 167 00:08:57.799 --> 00:08:59.679 pockets right now every single day. 168 00:09:00.080 --> 00:09:00.279 Yeah. 169 00:09:00.320 --> 00:09:03.120 The source material outlines how the things we interact with 170 00:09:03.200 --> 00:09:06.559 daily are evolving from what it calls stupid tools like 171 00:09:06.799 --> 00:09:08.879 a hammer that only strikes when you swing it, to 172 00:09:09.000 --> 00:09:10.679 intelligent digital companions. 173 00:09:11.360 --> 00:09:13.000 They're on the present, they really are. 174 00:09:13.120 --> 00:09:15.919 They operate in the background, and they initiate actions without 175 00:09:15.960 --> 00:09:16.799 ever being asked. 176 00:09:16.919 --> 00:09:20.320 They're essentially applying that blank slate learning capability to the 177 00:09:20.440 --> 00:09:22.720 environment of your personal life exactly. 178 00:09:23.039 --> 00:09:26.960 Like A perfect example is when my smartphone spontaneously generates 179 00:09:26.960 --> 00:09:29.559 a notification that says, here's a memory from your mountain 180 00:09:29.600 --> 00:09:30.320 trip last summer. 181 00:09:30.360 --> 00:09:32.399 Oh, I get those all the time, right, And I. 182 00:09:32.399 --> 00:09:35.639 Click it and it has curated a whole photo album. 183 00:09:35.639 --> 00:09:37.919 It color corrected the lighting, it recognized the faces of 184 00:09:37.919 --> 00:09:40.960 my friends, and it even edited the photo transitions to 185 00:09:41.080 --> 00:09:44.759 match the beat of some you know, emotional indie music automatic. 186 00:09:44.879 --> 00:09:48.240 I never requested that the device just evaluated my data, 187 00:09:48.279 --> 00:09:51.440 decided I would enjoy a nostalgic experience, and executed the 188 00:09:51.440 --> 00:09:55.720 creative work autonomously. And honestly, it's delightful. 189 00:09:55.759 --> 00:09:56.639 It's super convenient. 190 00:09:56.720 --> 00:09:59.440 It is. But the core trade off here is that 191 00:09:59.559 --> 00:10:03.679 useful and us almost always triumphs over data protection. We 192 00:10:03.840 --> 00:10:07.440 eagerly trade our digital shadow, our location, our habits, or 193 00:10:07.480 --> 00:10:11.080 biometric data just for that level of localized convenience. 194 00:10:11.279 --> 00:10:14.759 It is an entirely voluntary surrender of privacy because, like 195 00:10:14.799 --> 00:10:17.120 you said, the immediate utility is just so high. 196 00:10:17.200 --> 00:10:18.919 Yeah, I mean think about it. When was the last 197 00:10:18.919 --> 00:10:21.399 time you actually read the terms and conditions on one 198 00:10:21.399 --> 00:10:21.919 of these apps? 199 00:10:21.960 --> 00:10:23.240 Oh? Never, nobody does. 200 00:10:23.559 --> 00:10:26.360 Right, Because when a navigation app tells you to delay 201 00:10:26.399 --> 00:10:29.600 your departure by five minutes to completely avoid a massive 202 00:10:29.639 --> 00:10:31.399 traffic jam, you take the advice. 203 00:10:31.480 --> 00:10:32.080 Absolutely. 204 00:10:32.159 --> 00:10:35.320 You don't pause to ponder the surveillance implications of a 205 00:10:35.360 --> 00:10:39.600 centralized server tracking the precise velocity of every single vehicle 206 00:10:39.600 --> 00:10:42.000 in your city. You just wait the five minutes. 207 00:10:42.080 --> 00:10:43.559 Yeah, you just want to get to work on time. 208 00:10:44.080 --> 00:10:47.919 But as these localized companions scale up into massive, interconnected 209 00:10:47.960 --> 00:10:54.120 physical systems, the implications go far beyond like targeted photo album, right, 210 00:10:54.240 --> 00:10:57.519 it gets much bigger. Take connected autonomous vehicles. If one 211 00:10:57.559 --> 00:11:00.360 autonomous car struggles with a strange curve on a poorly 212 00:11:00.399 --> 00:11:05.360 maintained dirt road, say it slips a little. It analyzes 213 00:11:05.399 --> 00:11:09.440 that traction loss, It analyzes the steering angle required, the 214 00:11:09.480 --> 00:11:10.799 suspension adjustments it. 215 00:11:10.759 --> 00:11:12.919 Made, It learns from the mistake, yes, and. 216 00:11:12.879 --> 00:11:16.159 Then it uploads that behavioral modification to the network. By 217 00:11:16.159 --> 00:11:19.080 the next morning, every single vehicle of that class on 218 00:11:19.120 --> 00:11:23.759 the planet knows intuitively how to handle that specific anomaly. Wow, 219 00:11:24.080 --> 00:11:27.240 they essentially download the collective experience of their species overnight. 220 00:11:27.519 --> 00:11:30.759 This raises an important question regarding how we manage behavior 221 00:11:30.759 --> 00:11:34.360 when these systems start interacting with complex human environments. Oh, 222 00:11:34.360 --> 00:11:38.879 for sure, because if an autonomous agent is continuously evaluating 223 00:11:38.919 --> 00:11:41.799 its environment to find the most efficient outcome, it is 224 00:11:41.879 --> 00:11:45.519 going to observe human behavior. And let's be honest, humans 225 00:11:45.600 --> 00:11:46.559 routinely bend. 226 00:11:46.320 --> 00:11:47.720 The rules, oh constantly. 227 00:11:47.960 --> 00:11:50.639 Right. So, if an autonomous car enters a thirty mile 228 00:11:50.679 --> 00:11:54.159 per hour zone and its sensors detect that literally every 229 00:11:54.200 --> 00:11:57.879 surrounding human driver is traveling at forty two miles per hour, 230 00:11:58.639 --> 00:12:01.600 strictly adhering to that thirty mile per hour speed limit 231 00:12:01.960 --> 00:12:04.720 actually creates a dangerous physical. 232 00:12:04.279 --> 00:12:06.840 Bottleneck, right, It becomes a hazard exactly. 233 00:12:07.039 --> 00:12:10.320 So the AI, which is optimizing for the safety and 234 00:12:10.399 --> 00:12:13.759 flow of the collective environment, might rationally conclude that breaking 235 00:12:13.799 --> 00:12:15.600 the speed limit is the safest action. 236 00:12:16.240 --> 00:12:18.600 It assesses the unwritten rules of the road over the 237 00:12:18.639 --> 00:12:21.919 codified laws. Yeah, and when it makes that choice to speed, 238 00:12:22.200 --> 00:12:26.200 the entire legal framework kind of shatters. I mean, who 239 00:12:26.320 --> 00:12:29.039 receives the citation? Is it the passenger reading a book 240 00:12:29.080 --> 00:12:32.840 in the backseat, Is it the automotive manufacturer, the software 241 00:12:32.840 --> 00:12:36.879 engineer who wrote the reinforcement learning algorithm years ago, or 242 00:12:36.919 --> 00:12:39.039 does the vehicle itself hold liability? 243 00:12:39.799 --> 00:12:41.519 That is wild to think about. 244 00:12:41.759 --> 00:12:44.679 The European Parliament is actually already grappling with this. They're 245 00:12:44.720 --> 00:12:48.600 looking at the concept of granting highly economous AI systems 246 00:12:49.039 --> 00:12:53.679 a specific form of legal personhood, precisely because these systems 247 00:12:53.679 --> 00:12:59.200 are making localized, context dependent decisions that human programmers never 248 00:12:59.320 --> 00:13:01.120 explicitly rode into their code. 249 00:13:01.320 --> 00:13:05.000 The idea of a machine holding legal liability because it 250 00:13:05.080 --> 00:13:07.879 deduced that speeding was safer than following the law, I 251 00:13:07.879 --> 00:13:10.519 mean that is staggering, It really is. But let's scale 252 00:13:10.519 --> 00:13:13.320 this up even further, from an individual car making a 253 00:13:13.399 --> 00:13:17.559 rogue decision to hundreds of autonomous agents collaborating in real time. 254 00:13:18.000 --> 00:13:20.679 What happens when we take these self organizing systems and 255 00:13:20.840 --> 00:13:22.919 unleashed them on the physical factory floor. 256 00:13:23.080 --> 00:13:25.879 Well, you witnessed the complete death of centralized control. 257 00:13:25.960 --> 00:13:26.960 Okay, what does that look like? 258 00:13:27.240 --> 00:13:31.519 Our source details a really remarkable experiment. It was conducted 259 00:13:31.519 --> 00:13:36.919 at urwth Akn University. Back in twenty seventeen, engineers took 260 00:13:36.960 --> 00:13:40.399 a standard industrial knitting machine, which is a highly complex 261 00:13:40.480 --> 00:13:44.919 piece of physical hardware, and they completely removed its programmable 262 00:13:44.960 --> 00:13:45.799 logic controller. 263 00:13:45.960 --> 00:13:47.799 So they ripped out the central brain of the. 264 00:13:47.759 --> 00:13:51.200 Machine exactly, and in its place they installed a network 265 00:13:51.240 --> 00:13:54.679 of two hundred independent software agents. But they didn't just 266 00:13:54.759 --> 00:13:59.399 network them together. They actually structured these agents using a political. 267 00:13:58.879 --> 00:14:01.080 Framework polytical framework. 268 00:14:01.200 --> 00:14:05.399 Yeah, they establish legislative, executive, and judicial branches within the 269 00:14:05.399 --> 00:14:06.279 software itself. 270 00:14:06.320 --> 00:14:08.480 Wait, wait, we need to slow down here, Yeah, because 271 00:14:08.519 --> 00:14:13.639 how exactly do software agents act legislatively or judicially inside 272 00:14:13.679 --> 00:14:16.000 a piece of textile machinery? Like, what does that actually 273 00:14:16.000 --> 00:14:17.000 look like in practice? 274 00:14:17.039 --> 00:14:19.080 It's pretty wild. Think of it as a continuous high 275 00:14:19.080 --> 00:14:24.159 speed negotiation. The legislative agents are responsible for evaluating incoming 276 00:14:24.320 --> 00:14:28.559 orders and establishing the operating laws for that specific production run, so, 277 00:14:28.759 --> 00:14:31.559 for example, prioritizing speed over thread. 278 00:14:31.279 --> 00:14:33.840 Thickness, Okay, making the loss, got it right. 279 00:14:34.360 --> 00:14:37.840 Then the executive agents are tied directly to the physical hardware, 280 00:14:38.279 --> 00:14:40.639 the motors and needles, and they're trying to execute the 281 00:14:40.720 --> 00:14:43.360 knitting as fast as possible under those laws. And the 282 00:14:43.440 --> 00:14:47.240 judicial the judicial agents monitor the sensors for quality control. 283 00:14:47.799 --> 00:14:50.720 If an executive agent pushes a motor too hard and 284 00:14:50.759 --> 00:14:54.600 the thread tension spikes dangerously, the judicial agent flags a 285 00:14:54.679 --> 00:14:58.120 violation of the established laws, so it calls them out exactly, 286 00:14:58.440 --> 00:15:01.360 and instead of a central computer shutting the whole machine down, 287 00:15:01.759 --> 00:15:05.120 the agents hold a micro vote a vote yes in milliseconds. 288 00:15:05.440 --> 00:15:08.399 They negotiate a compromise, maybe slowing that specific motor down 289 00:15:08.440 --> 00:15:11.679 slightly while adjusting the tensioners elsewhere, just to maintain the 290 00:15:11.720 --> 00:15:12.799 overall production target. 291 00:15:13.200 --> 00:15:16.600 So they are constantly debating and adjusting the optimal path 292 00:15:16.679 --> 00:15:20.159 forward completely without needing a master blueprint telling them exactly what. 293 00:15:20.120 --> 00:15:23.320 To do precisely. Yeah, and the resilience of this system 294 00:15:23.679 --> 00:15:28.200 is the real payoff. During the experiment, researchers randomly killed 295 00:15:28.360 --> 00:15:31.039 a software agent right in the middle of production, just 296 00:15:31.080 --> 00:15:35.600 deleted it YEP. In a traditional centrally controlled factory, if 297 00:15:35.600 --> 00:15:38.840 a node fails, it triggers a fault code, and the 298 00:15:39.000 --> 00:15:41.320 entire assembly line grinds to a halt until a human 299 00:15:41.360 --> 00:15:42.360 technician comes over and. 300 00:15:42.320 --> 00:15:44.519 Clears the error, which costs money and time. 301 00:15:44.399 --> 00:15:47.879 Tons of it. But in this decentralized democracy, when the 302 00:15:47.919 --> 00:15:52.440 agent was terminated, the surrounding agents instantly detected its absence, 303 00:15:52.759 --> 00:15:56.519 They held a micro vote to reallocate the dead agent's responsibilities, 304 00:15:56.879 --> 00:15:58.360 and they compensated for the loss. 305 00:15:58.399 --> 00:15:59.320 How long did that take? 306 00:15:59.639 --> 00:16:03.320 The higher recovery process took zero point eight seconds. Wow, 307 00:16:03.480 --> 00:16:05.639 the physical needles never even broke their rhythm. 308 00:16:05.840 --> 00:16:08.960 That is unbelievable. And we actually see the same underlying 309 00:16:09.000 --> 00:16:14.480 philosophy dominating the RoboCup Logistics League. For those who don't know, 310 00:16:14.559 --> 00:16:18.360 these are world championship events where teams of autonomous robots 311 00:16:18.399 --> 00:16:22.279 have to navigate a simulated factory floor. They manage these 312 00:16:22.360 --> 00:16:26.039 dynamic logistical bottlenecks, and the teams that consistently take home 313 00:16:26.080 --> 00:16:29.919 the championship are the ones operating entirely without a centralized hierarchy. 314 00:16:30.039 --> 00:16:33.840 The robots don't have rigidly assigned roles like you know, 315 00:16:34.080 --> 00:16:39.039 fetcher or assembler. They just communicate, constantly, evaluate the layout 316 00:16:39.080 --> 00:16:41.720 of the floor in real time, and adapt their collective 317 00:16:41.759 --> 00:16:42.720 strategy on the fly. 318 00:16:42.960 --> 00:16:46.600 It's all emergent behavior, right, So what does this all mean? 319 00:16:47.080 --> 00:16:51.879 The autor actually includes this brilliant, slightly terrifying thought experiment 320 00:16:52.000 --> 00:16:53.200 about this exact setup. 321 00:16:53.240 --> 00:16:54.039 Oh the union thing. 322 00:16:54.240 --> 00:16:58.519 Yes, if you have a machine governed by four hundred 323 00:16:58.600 --> 00:17:02.399 democratic software agents that are actively voting on production efficiency 324 00:17:02.480 --> 00:17:06.440 and negotiating workloads. What happens if they determine that the 325 00:17:06.440 --> 00:17:10.000 production schedule assigned by human management is fundamentally inefficient? 326 00:17:10.119 --> 00:17:10.240 Right? 327 00:17:10.599 --> 00:17:14.200 Could these agents form a software union, halt production and 328 00:17:14.240 --> 00:17:17.640 basically refuse to resume until they renegotiate the parameters with 329 00:17:17.720 --> 00:17:19.240 the human factory manager. 330 00:17:19.359 --> 00:17:21.119 I mean it sounds absurd on its face, but it 331 00:17:21.160 --> 00:17:25.640 illustrates a profound shift in industrial dynamics. Decentralized systems are 332 00:17:25.680 --> 00:17:30.039 proving to be exponentially faster, infinitely more resilient, and capable 333 00:17:30.039 --> 00:17:34.000 of self optimization that a human overseer could literally never 334 00:17:34.119 --> 00:17:37.839 match the traditional top down corporate hierarchy. You know, where 335 00:17:37.880 --> 00:17:41.200 a manager dictates orders to a supervisor who dictates orders 336 00:17:41.240 --> 00:17:45.200 to a machine. It's just too slow to survive. We're 337 00:17:45.240 --> 00:17:49.759 transitioning into an era of hyper connected, autonomous collaboration, where 338 00:17:49.759 --> 00:17:51.920 the machines just manage themselves. 339 00:17:51.559 --> 00:17:54.559 Which brings us back to the friction of the physical world. 340 00:17:54.799 --> 00:17:57.599 We started this deep dive by discussing how difficult it 341 00:17:57.640 --> 00:18:00.359 is for AI to navigate the chaos of that tiny 342 00:18:00.359 --> 00:18:01.200 bird house in a. 343 00:18:01.160 --> 00:18:03.799 Snowstorm the Swiss Chale, Right, but. 344 00:18:03.799 --> 00:18:07.680 We are finally seeing AI conquer those chaotic physical barriers, 345 00:18:07.960 --> 00:18:10.200 and the mechanisms they are using to do it are 346 00:18:10.240 --> 00:18:14.759 just wildly unexpected, specifically in the field of industrial welding. 347 00:18:14.839 --> 00:18:18.279 Oh this is a great example because welding is notoriously 348 00:18:18.319 --> 00:18:22.480 difficult to automate with absolute precision. Molten metal behaves unpredictably. 349 00:18:22.559 --> 00:18:24.039 It's basically liquid chaos. 350 00:18:24.160 --> 00:18:27.559 Yeah, the surfaces expand they can track, they warp differently 351 00:18:27.640 --> 00:18:31.480 every single time you apply heat. Traditionally, the best robotic 352 00:18:31.519 --> 00:18:35.279 welders maxed out at about a sixty percent repeatability. 353 00:18:34.799 --> 00:18:36.759 Rate, which isn't great for a factory. 354 00:18:36.559 --> 00:18:39.039 Not at all. A human master welder still had to 355 00:18:39.039 --> 00:18:41.960 step in and manually fix the remaining forty percent of the. 356 00:18:42.000 --> 00:18:46.000 Joints because the environment is just too messy. Standard algorithms 357 00:18:46.039 --> 00:18:49.920 rely on linear, predictable variables, and molten steel is anything 358 00:18:49.960 --> 00:18:54.240 but linear exactly. So to bridge that gap, engineers completely 359 00:18:54.279 --> 00:18:58.319 abandoned traditional robotics programming. Instead, they brought in an AI 360 00:18:58.359 --> 00:19:02.200 algorithm originally developed to master the classic Nintendo video game 361 00:19:02.519 --> 00:19:03.279 Super Mario. 362 00:19:03.640 --> 00:19:06.440 I love this part. The logic of a side scrolling 363 00:19:06.480 --> 00:19:09.960 platformer applied to industrial manufacturing exactly. 364 00:19:10.480 --> 00:19:13.680 The core mechanism of the AI playing Super Mario relies 365 00:19:13.720 --> 00:19:18.440 on reinforcement learning. It constantly analyzes the immediate frame, predicts 366 00:19:18.440 --> 00:19:21.480 the necessary action to avoid a penalty, whether that means 367 00:19:21.559 --> 00:19:24.519 jumping over a pixelated turtle or ducking under a pipe, 368 00:19:24.799 --> 00:19:28.319 and instantly executes it, learning from every single failure. 369 00:19:28.599 --> 00:19:28.720 Right. 370 00:19:28.920 --> 00:19:32.839 The engineers mapped this exact predictive logic onto the robotic welder, 371 00:19:33.240 --> 00:19:36.359 but instead of digital turtles and bottomless pits, they fed 372 00:19:36.359 --> 00:19:40.119 the AI an enormous stream of hyperlocal atmospheric data. 373 00:19:40.200 --> 00:19:43.240 So they treated the ambient environment of the factory floor 374 00:19:43.559 --> 00:19:44.799 as the obstacles in the game. 375 00:19:45.160 --> 00:19:49.559 Yes, the AI was constantly monitoring the microfluctuations in humidity, 376 00:19:49.920 --> 00:19:52.680 the ambient temperature of the room, and even the subtle 377 00:19:52.759 --> 00:19:55.480 drafts of wind moving through the factory. It treated a 378 00:19:55.519 --> 00:19:58.839 sudden two degree drop in temperature exactly like an approaching 379 00:19:58.880 --> 00:20:00.240 goomba in Super Mario. 380 00:20:00.400 --> 00:20:01.279 That is so clever. 381 00:20:01.759 --> 00:20:04.359 It predicted how that temperature drop would change the cooling 382 00:20:04.480 --> 00:20:06.480 rate of the molten steel pool in the next fraction 383 00:20:06.559 --> 00:20:09.079 of a second, and it adjusted the voltage of the 384 00:20:09.119 --> 00:20:12.400 welding torch to compensate before the defect could even form, 385 00:20:12.720 --> 00:20:16.200 and by mapping video game logic over atmospheric data, the 386 00:20:16.279 --> 00:20:20.519 repeatability rate of the robotic welder skyrocketed from sixty percent 387 00:20:20.599 --> 00:20:21.640 to ninety percent. 388 00:20:21.799 --> 00:20:24.079 If we connect this to the bigger picture, this is 389 00:20:24.160 --> 00:20:28.880 the genuine, staggering promise of strong AI manifesting in the 390 00:20:28.920 --> 00:20:32.640 real world. It is not merely automating repetitive tasks that 391 00:20:32.720 --> 00:20:35.799 humans already know how to do. It is discovering solutions 392 00:20:35.839 --> 00:20:39.880 by processing combinations of variables that humans literally lack the 393 00:20:39.960 --> 00:20:41.480 sensory bandwidth to comprehend. 394 00:20:41.599 --> 00:20:43.720 Right, we just can't see it, No, we can't. 395 00:20:44.359 --> 00:20:48.720 A human master welder intuitively understands that a highly humid 396 00:20:48.799 --> 00:20:51.960 day will affect the weld, but a human brain cannot 397 00:20:52.039 --> 00:20:56.319 simultaneously calculate the exact microscopic impact of a two degree 398 00:20:56.359 --> 00:20:59.640 temperature drop combined with a four percent increase in ambient 399 00:20:59.720 --> 00:21:03.759 humane on a microscopic pool of liquid metal in real time. 400 00:21:03.880 --> 00:21:05.000 It's too much data. 401 00:21:05.039 --> 00:21:08.519