WEBVTT 00:00:00.000 --> 00:00:03.160 Imagine a factory, but not a factory like you've 00:00:03.160 --> 00:00:06.400 ever seen before. Right. It's pitch black inside. 00:00:06.639 --> 00:00:08.660 I mean, completely dark. There are no lights 00:00:08.660 --> 00:00:11.359 on the ceiling because, well, nobody inside needs 00:00:11.359 --> 00:00:13.800 to see. There's no air conditioning. No AC because 00:00:13.800 --> 00:00:15.640 nobody needs to be comfortable. There are no 00:00:15.640 --> 00:00:18.460 break rooms, no bathrooms, no parking lots outside. 00:00:19.260 --> 00:00:21.500 Nothing. It's the ultimate efficiency engine, 00:00:21.579 --> 00:00:24.460 the architecture of the machine age taken to 00:00:24.460 --> 00:00:26.940 its logical conclusion. Exactly. It is a company 00:00:26.940 --> 00:00:30.579 with absolutely positively zero humans on the 00:00:30.579 --> 00:00:33.740 payroll. Not a single one. None. It is just a 00:00:33.740 --> 00:00:36.979 fleet of autonomous agents. Some are, you know, 00:00:36.979 --> 00:00:39.439 code running on servers. Some are robots moving 00:00:39.439 --> 00:00:42.549 boxes on the floor. Now, here's the question 00:00:42.549 --> 00:00:44.170 that keeps me up at night, and it's the hook 00:00:44.170 --> 00:00:47.270 for today's deep dive. In that building, in that 00:00:47.270 --> 00:00:50.649 dark, silent factory, how do you calculate the 00:00:50.649 --> 00:00:53.130 wages? It sounds like a philosophical riddle, 00:00:53.149 --> 00:00:55.090 doesn't it? It really does. Like if a tree falls 00:00:55.090 --> 00:00:57.670 in the forest. But in 2026, and this is the key, 00:00:57.729 --> 00:00:59.890 this isn't philosophy anymore. This is hard -nosed 00:00:59.890 --> 00:01:02.630 accounting and this is business. It really, really 00:01:02.630 --> 00:01:05.870 is. If you have a robot moving boxes and an AI 00:01:05.870 --> 00:01:09.909 writing code. They're both working. But one is 00:01:09.909 --> 00:01:12.049 burning calories or, well, joules lifting heavy 00:01:12.049 --> 00:01:14.510 things, fighting gravity, friction and inertia. 00:01:14.849 --> 00:01:17.290 It's dealing with the messy physical world. Right. 00:01:17.409 --> 00:01:20.049 The other is just thinking, processing electrons. 00:01:20.189 --> 00:01:23.069 So how do you compare those two values? How do 00:01:23.069 --> 00:01:25.730 you write a paycheck for a robot? That is the 00:01:25.730 --> 00:01:28.409 trillion dollar question. And honestly, up until 00:01:28.409 --> 00:01:30.989 very recently, nobody had a good answer. No. 00:01:31.170 --> 00:01:34.090 We were trying to apply human economics to machine 00:01:34.090 --> 00:01:37.450 labor and the math. It just didn't work. We were 00:01:37.450 --> 00:01:39.349 trying to use dollars to measure thermodynamics. 00:01:39.629 --> 00:01:41.549 It's like trying to measure temperature with 00:01:41.549 --> 00:01:43.549 a ruler. Well, today we are going to fix the 00:01:43.549 --> 00:01:46.370 math. We are doing a deep dive into a brand new 00:01:46.370 --> 00:01:49.469 academic paper that honestly just blew my mind. 00:01:49.510 --> 00:01:52.530 It is titled Jewelwork Robotics, Thermodynamic 00:01:52.530 --> 00:01:55.709 Wage Frameworks for Embodied AR. And before we 00:01:55.709 --> 00:01:57.650 go any further, we need to set the stage properly. 00:01:57.870 --> 00:02:01.579 This is the Read Multiplex podcast. We are bringing 00:02:01.579 --> 00:02:03.939 you the absolute cutting edge of what is happening 00:02:03.939 --> 00:02:06.900 in the Read Multiplex ecosystem. This isn't just, 00:02:06.939 --> 00:02:10.560 you know, news. This is a glimpse into the future 00:02:10.560 --> 00:02:12.539 that is being written right now. That is right. 00:02:12.900 --> 00:02:15.379 And the author of this paper is someone we follow 00:02:15.379 --> 00:02:18.780 very, very closely. Brian Romella. The chairman 00:02:18.780 --> 00:02:22.620 of the Zero Human Company. The man himself. We've 00:02:22.620 --> 00:02:24.500 covered Brian's work before. We did a deep dive 00:02:24.500 --> 00:02:27.639 on his original jewel work concept a little while 00:02:27.639 --> 00:02:29.819 back, which was, I mean, it was fascinating in 00:02:29.819 --> 00:02:31.819 its own right. But I have to say, looking at 00:02:31.819 --> 00:02:35.020 this new paper, this feels like a new peak in 00:02:35.020 --> 00:02:37.460 brilliance from him. I agree. It's dense, it's 00:02:37.460 --> 00:02:39.939 mathematical, but it's incredibly visionary. 00:02:40.400 --> 00:02:42.099 It really is. You know, you look at the landscape 00:02:42.099 --> 00:02:44.280 of AI development right now. You have the big 00:02:44.280 --> 00:02:46.599 labs in San Francisco, the massive tech giants 00:02:46.599 --> 00:02:49.039 in China, the European regulators. They are all 00:02:49.039 --> 00:02:51.300 building the engines. Right. They are building 00:02:51.300 --> 00:02:54.120 the LLMs, the vision models, the diffusion models. 00:02:54.379 --> 00:02:57.259 Yeah. But Rolemo, he's doing something different. 00:02:57.319 --> 00:02:59.699 He is flying solo on this. He's not building 00:02:59.699 --> 00:03:02.219 the engine. He's building the economy. Precisely. 00:03:02.639 --> 00:03:04.900 None of the big tech giants have really caught 00:03:04.900 --> 00:03:06.639 up to his understanding of what a zero human 00:03:06.639 --> 00:03:09.840 company actually is or how it functions operationally. 00:03:09.860 --> 00:03:13.039 They're still thinking about AI as a tool for 00:03:13.039 --> 00:03:15.449 humans. It's an assistant. Right. an assistant. 00:03:16.110 --> 00:03:19.310 Romell is thinking about AI as an entity that 00:03:19.310 --> 00:03:22.530 exists without humans. He is operating at a level 00:03:22.530 --> 00:03:26.009 of clarity and speed that is just, frankly, unmatched 00:03:26.009 --> 00:03:28.610 in the industry right now. While everyone else 00:03:28.610 --> 00:03:31.069 is arguing about copyright or how to make the 00:03:31.069 --> 00:03:33.789 chatbot sound more empathetic, he's writing the 00:03:33.789 --> 00:03:36.189 tax code for the future of automation. That is 00:03:36.189 --> 00:03:38.550 such a crucial distinction, and it highlights 00:03:38.550 --> 00:03:40.210 why we do this show. If you want to know where 00:03:40.210 --> 00:03:42.129 the ball is going, you look at what Romell is 00:03:42.129 --> 00:03:44.650 writing. So our mission today is to unpack this 00:03:44.650 --> 00:03:47.349 paper. We need to understand how we move from 00:03:47.349 --> 00:03:49.990 paying for abstract thought, which was the original 00:03:49.990 --> 00:03:53.150 jewel work concept, to paying for action for 00:03:53.150 --> 00:03:55.370 robots moving in the real world. And that transition 00:03:55.370 --> 00:03:57.889 is incredibly difficult. I mean, you can't overstate 00:03:57.889 --> 00:03:59.729 it. It is the difference between a spreadsheet 00:03:59.729 --> 00:04:02.250 and a construction site. Yeah. The physics are 00:04:02.250 --> 00:04:03.849 completely different. The stakes are different. 00:04:04.009 --> 00:04:07.150 So here is our roadmap for the next hour or so. 00:04:07.819 --> 00:04:09.919 We are going to start by reviewing the concept 00:04:09.919 --> 00:04:13.340 of the Zero Human Company and the original jewel 00:04:13.340 --> 00:04:16.259 work to set the stage. A quick refresher. Exactly. 00:04:16.779 --> 00:04:19.379 Then we're going to run headfirst into a problem 00:04:19.379 --> 00:04:22.160 called Moravec's Paradox, which explains why 00:04:22.160 --> 00:04:24.540 moving your arm is actually harder for a computer 00:04:24.540 --> 00:04:27.839 than playing Grandmaster -level chess. It's such 00:04:27.839 --> 00:04:29.519 a counterintuitive concept, and I can't wait 00:04:29.519 --> 00:04:31.439 to get into that. It flips everything we think 00:04:31.439 --> 00:04:34.019 we know about intelligence on its head. It really 00:04:34.019 --> 00:04:37.129 does. Then things get historical. We're going 00:04:37.129 --> 00:04:40.889 to see how Rommel connects 2026 robotics all 00:04:40.889 --> 00:04:44.250 the way back to factory management in 1911. We're 00:04:44.250 --> 00:04:46.449 talking about men with stopwatches and early 00:04:46.449 --> 00:04:48.990 industrial theory. Which is just classic Rommel, 00:04:49.050 --> 00:04:51.410 finding the answer in the past. It turns out 00:04:51.410 --> 00:04:54.129 the past has the answer to the future. And finally, 00:04:54.209 --> 00:04:56.089 we're going to break down the monster equation, 00:04:56.389 --> 00:04:59.769 the JWR formula variable by variable. We are 00:04:59.769 --> 00:05:01.990 going to do the math behind the money. It is 00:05:01.990 --> 00:05:04.470 a comprehensive framework. It's dense, but it's 00:05:04.470 --> 00:05:06.399 crucial. We're going to make sure everyone listening 00:05:06.399 --> 00:05:09.079 can actually visualize how this wage is calculated. 00:05:09.339 --> 00:05:12.139 It is. So let's dive in. Let's do it. Okay, so 00:05:12.139 --> 00:05:13.819 before we get to the robots, we need to quickly 00:05:13.819 --> 00:05:16.899 recap where we left off. We previously discussed 00:05:16.899 --> 00:05:20.699 the original Joule work metric, or JW. For the 00:05:20.699 --> 00:05:23.019 listeners who might have missed that deep dive 00:05:23.019 --> 00:05:26.100 or just, you know, need a refresher, can you 00:05:26.100 --> 00:05:29.980 give us the explain like I'm five? Sure. So in 00:05:29.980 --> 00:05:32.379 the previous paper, Rommel established the fundamental 00:05:32.379 --> 00:05:36.259 currency of the AI economy. The core idea was 00:05:36.259 --> 00:05:38.259 that we need to move away from fiat currency. 00:05:38.939 --> 00:05:43.199 Dollars, euros, yen. These are subject to inflation. 00:05:43.259 --> 00:05:45.360 They're subject to human politics, central banks, 00:05:45.579 --> 00:05:47.819 interest rates. They're soft. They're very soft. 00:05:47.899 --> 00:05:50.819 They're based on faith. on collective agreement. 00:05:50.959 --> 00:05:53.199 And if you have a machine economy, you want hard 00:05:53.199 --> 00:05:55.459 money. You want something that can't be manipulated 00:05:55.459 --> 00:05:57.500 by a Federal Reserve chairman waking up on the 00:05:57.500 --> 00:05:59.120 wrong side of the bed. Exactly. You want money 00:05:59.120 --> 00:06:02.680 grounded in physics, in thermodynamics. The laws 00:06:02.680 --> 00:06:04.639 of the universe don't change. Right. The original 00:06:04.639 --> 00:06:08.240 formula was deceptively simple. JW is E times 00:06:08.240 --> 00:06:10.759 kappa times W. Okay, break that down for us one 00:06:10.759 --> 00:06:13.439 more time. Just the basics. Sure. Dollars is 00:06:13.439 --> 00:06:15.939 energy consumed. Pretty straightforward. How 00:06:15.939 --> 00:06:17.699 much electricity did you pull from the grid? 00:06:18.240 --> 00:06:22.339 Kappa. is a normalization coefficient, basically 00:06:22.339 --> 00:06:25.060 a standardizing number to make the math work 00:06:25.060 --> 00:06:28.000 across different systems, different chip architectures. 00:06:28.079 --> 00:06:31.259 So an NVIDIA chip and a Google TPU can be compared 00:06:31.259 --> 00:06:33.980 fairly. That's the idea. And doodle -allers is 00:06:33.980 --> 00:06:37.050 the work output, the useful result. So in plain 00:06:37.050 --> 00:06:39.290 English, that basically means how much energy 00:06:39.290 --> 00:06:42.009 did you burn to produce a useful result? Precisely. 00:06:42.009 --> 00:06:44.290 It was designed for what we call abstract AI. 00:06:44.529 --> 00:06:47.689 Okay. Think of chat GPT or an algorithm analyzing 00:06:47.689 --> 00:06:50.550 stock trends or a compiler turning code into 00:06:50.550 --> 00:06:53.470 software. It lives on a server. Its body is just 00:06:53.470 --> 00:06:56.050 silicon chips. It has no physical presence. None. 00:06:56.399 --> 00:06:58.639 The energy it consumes is purely electricity 00:06:58.639 --> 00:07:00.819 for computation. Yeah. And in that world, the 00:07:00.819 --> 00:07:03.160 math is relatively clean. The more energy you 00:07:03.160 --> 00:07:05.519 put in, generally, the more thought or compute 00:07:05.519 --> 00:07:07.699 you get out. It's a linear relationship, mostly. 00:07:07.779 --> 00:07:09.959 You feed the server electricity, it spits out 00:07:09.959 --> 00:07:12.480 tokens. More electricity, more tokens. Mostly. 00:07:12.980 --> 00:07:16.339 But now, in this new paper, Ravel has thrown 00:07:16.339 --> 00:07:19.040 a wrench in the gears. He's moved from the server 00:07:19.040 --> 00:07:21.839 room to the factory floor. We are talking about 00:07:21.839 --> 00:07:24.819 embodied AI. And this is where the jewel work 00:07:24.819 --> 00:07:28.660 robotics. framework or such a WR comes in, because 00:07:28.660 --> 00:07:31.259 as the paper points out, you cannot just use 00:07:31.259 --> 00:07:33.600 the old formula for a robot. Right. It breaks 00:07:33.600 --> 00:07:36.360 immediately the moment you try. But why not? 00:07:36.620 --> 00:07:37.800 I mean, this is what I was thinking at first. 00:07:37.860 --> 00:07:40.180 Energy is energy, isn't it? A jewel is a jewel, 00:07:40.339 --> 00:07:42.759 whether it turns a fan or lights a bulb or runs 00:07:42.759 --> 00:07:45.139 a processor. Physics doesn't distinguish between 00:07:45.139 --> 00:07:47.680 a smart jewel and a dumb jewel. In physics, yes, 00:07:47.920 --> 00:07:50.319 a jewel is a jewel. You're absolutely right. 00:07:50.420 --> 00:07:53.899 But in economics, no. And that's the key distinction. 00:07:54.360 --> 00:07:58.300 Physical reality is expensive. When you move 00:07:58.300 --> 00:08:00.920 from processing data to processing matter, listing 00:08:00.920 --> 00:08:03.980 a box, walking across a room, navigating a cluttered 00:08:03.980 --> 00:08:06.819 hallway, the energy dynamics change completely. 00:08:07.160 --> 00:08:08.660 It's the difference between thinking about running 00:08:08.660 --> 00:08:10.620 a marathon and actually running one. Exactly. 00:08:10.759 --> 00:08:13.180 Perfect analogy. Thinking about it costs a few 00:08:13.180 --> 00:08:16.060 calories of glucose in your brain, a tiny amount 00:08:16.060 --> 00:08:18.439 of energy. Negligible. Doing it costs thousands 00:08:18.439 --> 00:08:20.699 of calories. It's a massive energy expenditure. 00:08:21.180 --> 00:08:23.199 And this brings us to that concept you mentioned 00:08:23.199 --> 00:08:26.720 in the intro. Moravec's paradox. The paper cites 00:08:26.720 --> 00:08:30.220 this extensively as the core reason why we need 00:08:30.220 --> 00:08:33.100 a new equation. I love this paradox because it 00:08:33.100 --> 00:08:35.159 makes me feel better about being bad at math. 00:08:35.320 --> 00:08:37.259 Can you break it down? I think we all feel that 00:08:37.259 --> 00:08:39.259 way. Yeah. It's one of the most fascinating things 00:08:39.259 --> 00:08:41.879 in computer science. Hans Moravec identified 00:08:41.879 --> 00:08:44.600 this back in 1988. He was looking at the progress 00:08:44.600 --> 00:08:47.740 of AI and he noticed something weird. Okay. He 00:08:47.740 --> 00:08:49.799 noticed that high -level reasoning things humans 00:08:49.799 --> 00:08:53.720 find hard, like calculus, playing chess, analyzing 00:08:53.720 --> 00:08:56.700 complex statistics, proving theorems, is actually 00:08:56.700 --> 00:08:59.100 very easy for a computer. Computationally cheap. 00:08:59.360 --> 00:09:01.740 Very cheap. requires very little computational 00:09:01.740 --> 00:09:04.279 energy relative to the output. Right. My pocket 00:09:04.279 --> 00:09:06.659 calculator from 1995 can do math better than 00:09:06.659 --> 00:09:09.240 I can. A basic chess app on my phone can beat 00:09:09.240 --> 00:09:11.840 a grandmaster. Those things are governed by strict 00:09:11.840 --> 00:09:14.899 logic rules, which computers love. Exactly. We 00:09:14.899 --> 00:09:17.720 solved intelligence in that sense decades ago. 00:09:17.820 --> 00:09:22.179 But the things humans find easy walking, recognizing 00:09:22.179 --> 00:09:24.879 a face, picking up a cup without crushing it, 00:09:25.019 --> 00:09:27.879 tying your shoelaces, folding a towel. The things 00:09:27.879 --> 00:09:30.080 a toddler can do. The things a toddler can master. 00:09:30.440 --> 00:09:33.179 Those sensorimotor skills are unbelievably difficult 00:09:33.179 --> 00:09:36.460 for AI. Computationally expensive. Why is that? 00:09:36.500 --> 00:09:39.320 Why can a computer simulate a nuclear explosion 00:09:39.320 --> 00:09:41.919 easier than it can fold a shirt? That just seems 00:09:41.919 --> 00:09:45.220 so backwards. Evolution. That is the answer. 00:09:45.600 --> 00:09:47.799 We have had millions and millions of years to 00:09:47.799 --> 00:09:50.799 optimize our motor cortex. Walking is hardware 00:09:50.799 --> 00:09:53.600 accelerated in our brains. It's baked in. I don't 00:09:53.600 --> 00:09:55.360 think about walking. You don't even think about 00:09:55.360 --> 00:09:56.980 it. You'll calculate the friction coefficient 00:09:56.980 --> 00:09:59.399 of the floor before you take a step. You don't 00:09:59.399 --> 00:10:01.539 consciously adjust the tension in your calf muscles 00:10:01.539 --> 00:10:04.259 to maintain balance. You just walk. But a robot 00:10:04.259 --> 00:10:06.960 does. A robot has to do it all from scratch, 00:10:07.019 --> 00:10:11.039 every single millisecond. For a robot, walking 00:10:11.039 --> 00:10:13.639 across a room involves calculating the friction 00:10:13.639 --> 00:10:16.240 of the floor, the balance of its center of gravity, 00:10:16.379 --> 00:10:18.919 processing all the visual data from its cameras 00:10:18.919 --> 00:10:21.940 to avoid obstacles, sending the exact torque 00:10:21.940 --> 00:10:24.419 required for each joint in the leg. It is a data 00:10:24.419 --> 00:10:27.519 firehook. It requires massive amounts of computation 00:10:27.519 --> 00:10:30.419 just to stay upright and massive amounts of physical 00:10:30.419 --> 00:10:33.080 energy to drive the motors. So if you just use 00:10:33.080 --> 00:10:36.450 the old jewel work formula. A robot doing simple 00:10:36.450 --> 00:10:38.889 physical labor would look incredibly inefficient 00:10:38.889 --> 00:10:41.769 compared to an AI doing complex math. Correct. 00:10:42.029 --> 00:10:44.429 The robot would be burning thousands of joules 00:10:44.429 --> 00:10:47.429 just to stand up and walk to a shelf. The AI 00:10:47.429 --> 00:10:49.649 would burn a fraction of that to solve a complex 00:10:49.649 --> 00:10:51.889 theorem. If you paid them strictly on the old 00:10:51.889 --> 00:10:54.529 metric. The math breaks down. It does. The robot 00:10:54.529 --> 00:10:56.490 would be overpaid for what looks like energy 00:10:56.490 --> 00:10:58.570 waste. Or the system simply wouldn't account 00:10:58.570 --> 00:11:01.049 for the difficulty of the physical world. You'd 00:11:01.049 --> 00:11:02.750 have an accounting error that breaks the company. 00:11:03.149 --> 00:11:05.049 It would be like paying a theoretical physicist 00:11:05.049 --> 00:11:08.029 a million dollars a year and paying a construction 00:11:08.029 --> 00:11:10.750 worker 10 cents because the physicist is thinking 00:11:10.750 --> 00:11:13.789 harder. But if you need a building built, that 00:11:13.789 --> 00:11:16.330 math doesn't work. You need the building. Exactly. 00:11:16.330 --> 00:11:19.350 You have to value the physical actuation. You 00:11:19.350 --> 00:11:21.110 have to value the work done against gravity. 00:11:21.490 --> 00:11:25.029 So Rowemmel had to create a new system. a way 00:11:25.029 --> 00:11:27.509 to value that physical struggle a way to say 00:11:27.509 --> 00:11:30.870 hey moving this box is actually worth more than 00:11:30.870 --> 00:11:33.330 writing that poem even if the poem seems smarter 00:11:33.330 --> 00:11:37.289 he did he calls it jewel work robotics or rojwr 00:11:37.289 --> 00:11:39.809 and what i love about this is that he didn't 00:11:39.809 --> 00:11:42.129 throw away the old system he built a hybrid it's 00:11:42.129 --> 00:11:45.090 like a layer cake it is the framework preserves 00:11:45.090 --> 00:11:47.230 the cognitive jewel work for the brain of the 00:11:47.230 --> 00:11:49.769 robot the planning and thinking part but it adds 00:11:49.769 --> 00:11:52.539 a massive new layer for the body It's designed 00:11:52.539 --> 00:11:54.879 specifically for these zero human companies where 00:11:54.879 --> 00:11:57.200 you need to balance the books between your software 00:11:57.200 --> 00:12:00.340 agents and your hardware agents. And to build 00:12:00.340 --> 00:12:02.740 this new layer, Rommel didn't just look at code. 00:12:02.799 --> 00:12:05.299 He looked at history. This is the part of the 00:12:05.299 --> 00:12:08.240 paper that really surprised me. I expected citations 00:12:08.240 --> 00:12:12.100 from 2024, 2025, maybe some obscure robotics 00:12:12.100 --> 00:12:15.360 journals from MIT. But he's quoting guys from 00:12:15.360 --> 00:12:18.320 1911. This is vintage Rommel brilliance. This 00:12:18.320 --> 00:12:20.980 is why you read him. He bridges 2026 technology 00:12:20.980 --> 00:12:24.700 with early 20th century industrial science. He 00:12:24.700 --> 00:12:26.940 realizes that we have actually solved the problem 00:12:26.940 --> 00:12:30.799 of physical efficiency before, just not for robots. 00:12:30.919 --> 00:12:33.259 We solved it for humans. We did. You're talking 00:12:33.259 --> 00:12:36.539 about Taylorism. Frederick Taylor, 1911. The 00:12:36.539 --> 00:12:38.639 principles of scientific management. A man with 00:12:38.639 --> 00:12:40.960 a stopwatch, the original efficiency guru. The 00:12:40.960 --> 00:12:43.759 man with the stopwatch. Taylor was obsessed with 00:12:43.759 --> 00:12:46.220 efficiency. He would stand in factories, steel 00:12:46.220 --> 00:12:48.759 mills, bricklaying sites, and he would time every 00:12:48.759 --> 00:12:51.840 single movement a worker made. Every single one. 00:12:51.919 --> 00:12:54.419 He wanted to find the one best way to do a job. 00:12:54.559 --> 00:12:56.879 He didn't want you to just shovel coal. He wanted 00:12:56.879 --> 00:12:58.759 you to shovel coal with the exact perfect arc 00:12:58.759 --> 00:13:01.000 to maximize the load and minimize the fatigue. 00:13:01.100 --> 00:13:03.330 He broke it all down. I remember reading about 00:13:03.330 --> 00:13:06.309 this in history class. It was, well, it was hated, 00:13:06.350 --> 00:13:08.429 wasn't it? People did not like this. Oh, absolutely. 00:13:08.730 --> 00:13:11.309 It was incredibly controversial. Workers felt 00:13:11.309 --> 00:13:14.169 like they were being treated like machines. Dehumanized. 00:13:14.370 --> 00:13:16.190 Completely. They were being micromanaged down 00:13:16.190 --> 00:13:19.129 to the second. It led to strikes. It led to congressional 00:13:19.129 --> 00:13:21.970 hearings. It was seen as the mechanization of 00:13:21.970 --> 00:13:24.870 the human soul. But Rommel's insight here is 00:13:24.870 --> 00:13:28.009 that what's controversial for humans is perfect 00:13:28.009 --> 00:13:31.049 for machines. Exactly. Robots don't get offended 00:13:31.049 --> 00:13:33.360 if you time them. Robots don't get stressed out 00:13:33.360 --> 00:13:35.759 by efficiency quotas. They don't need dignity. 00:13:35.899 --> 00:13:39.440 They need data. Taylorism is the native language 00:13:39.440 --> 00:13:42.960 of robotics. It was a philosophy waiting for 00:13:42.960 --> 00:13:45.320 the right species to arrive. That is a great 00:13:45.320 --> 00:13:47.779 point. So applying Taylor's strict efficiency 00:13:47.779 --> 00:13:51.100 rules to robots is actually ethically sound. 00:13:51.299 --> 00:13:53.460 It's what they are built for. You're not hurting 00:13:53.460 --> 00:13:56.320 its feelings. Precisely. And he goes even deeper. 00:13:56.419 --> 00:13:58.440 He brings in the Gilbris, Frank, and Lillian 00:13:58.440 --> 00:14:00.559 Gilbreth. The Gilbris, right. They were contemporaries 00:14:00.559 --> 00:14:02.580 of Taylor, and they came up with something called 00:14:02.580 --> 00:14:06.179 Thurbligs. Thurbligs. I love that word. It sounds 00:14:06.179 --> 00:14:08.340 like a sci -fi creature from Star Trek. Captain, 00:14:08.379 --> 00:14:11.179 the Thurbligs are attacking the warp core. It 00:14:11.179 --> 00:14:13.659 does. But it's actually just Gilbreth spelled 00:14:13.659 --> 00:14:16.899 backwards with the transposed. Oh, I never realized 00:14:16.899 --> 00:14:18.580 that. That's clever. I've been saying that word 00:14:18.580 --> 00:14:20.399 for years and never made that connection. Yes. 00:14:20.740 --> 00:14:23.940 And a Thurblig is a fundamental unit of motion. 00:14:24.519 --> 00:14:28.340 The Gilbreths realized that any complex task 00:14:28.340 --> 00:14:31.259 like bricklaying or surgery or assembling a toaster 00:14:31.259 --> 00:14:34.080 could be broken down into about 18 elemental 00:14:34.080 --> 00:14:37.960 movements. Only 18 for everything. Roughly, yes. 00:14:38.059 --> 00:14:39.639 That was their theory. What kind of movements? 00:14:39.720 --> 00:14:43.919 Give us some examples. Search, grasp, hold, release, 00:14:44.240 --> 00:14:47.480 inspect, assemble, transport loaded, transport 00:14:47.480 --> 00:14:51.919 empty, plan, rest. Very basic fundamental actions. 00:14:52.200 --> 00:14:54.750 So they would analyze a worker and say, You're 00:14:54.750 --> 00:14:56.889 using too many thurblays to assemble that toaster. 00:14:56.990 --> 00:14:59.350 Your search time is too long. Exactly. They were 00:14:59.350 --> 00:15:02.309 looking for surplus thurblays. Wasted motion. 00:15:02.570 --> 00:15:04.789 If you have to reach too far to grab a screwdriver, 00:15:05.129 --> 00:15:07.730 that's a wasted transquart thurblay. If you have 00:15:07.730 --> 00:15:09.789 to fumble to get a grip because the handle is 00:15:09.789 --> 00:15:13.240 slippery, that's a wasted grasp. Thurblay. If 00:15:13.240 --> 00:15:15.200 you have to look around to find the screw, that's 00:15:15.200 --> 00:15:18.000 a wasted search, Thurblake. It's all about eliminating 00:15:18.000 --> 00:15:20.500 waste. And Rommel is applying this to robots. 00:15:20.860 --> 00:15:24.080 He is. He argues that just as ergonomics, which 00:15:24.080 --> 00:15:26.460 was pioneered by people like Grandjean in 1980, 00:15:26.679 --> 00:15:29.620 helped humans avoid physical strain, jewelwork 00:15:29.620 --> 00:15:32.840 robotics helps robots avoid energy strain. Energy 00:15:32.840 --> 00:15:34.919 strain. I like that. The framework isn't just 00:15:34.919 --> 00:15:37.700 code. It is an economic philosophy rooted in 00:15:37.700 --> 00:15:40.600 this physical motion economy. It rewards the 00:15:40.600 --> 00:15:42.840 robot for reducing its Thurblakes. for being 00:15:42.840 --> 00:15:44.759 more efficient in its movements. So instead of 00:15:44.759 --> 00:15:47.639 a manager shouting, move faster, the code itself 00:15:47.639 --> 00:15:51.240 incentivizes the robot to move smarter. Move 00:15:51.240 --> 00:15:53.860 with less energy. Move with more purpose. Eliminate 00:15:53.860 --> 00:15:56.379 the jitter. Eliminate the hesitation. If a robot 00:15:56.379 --> 00:15:58.620 reaches for a box and misses slightly, then has 00:15:58.620 --> 00:16:00.960 to correct, that's a wasted therblig. That's 00:16:00.960 --> 00:16:04.179 energy burn with no value. The JWR system penalizes 00:16:04.179 --> 00:16:07.110 that. It connects perfectly. So we have the modern 00:16:07.110 --> 00:16:09.870 problem, Moravec's paradox, or why is moving 00:16:09.870 --> 00:16:12.590 so hard, and the historical solution time motion 00:16:12.590 --> 00:16:15.529 study, or how do we measure movement? Now let's 00:16:15.529 --> 00:16:17.409 see how Rommel actually turns that into math. 00:16:17.950 --> 00:16:20.250 Let's get to the equation. The paper calls this 00:16:20.250 --> 00:16:22.690 the integrated equation, but I'm calling it the 00:16:22.690 --> 00:16:26.029 monster. It is huge. When you first see it, it 00:16:26.029 --> 00:16:28.789 has parentheses, inside brackets, inside equations. 00:16:29.250 --> 00:16:31.289 It looks like something you'd see on the blackboard 00:16:31.289 --> 00:16:34.090 in Good Will Hunting. It is complex, but every 00:16:34.090 --> 00:16:37.210 piece serves a specific economic purpose. Let 00:16:37.210 --> 00:16:38.970 me read out the full structure, and then we will 00:16:38.970 --> 00:16:41.090 dismantle it piece by piece. Are you ready? Hit 00:16:41.090 --> 00:16:44.549 me. GWR cognitive, Chespie times OMA times MM 00:16:44.549 --> 00:16:50.169 times 10F plus SBU. Okay. That is a mouthful. 00:16:50.230 --> 00:16:53.210 It is. Let's unpack this slowly. I heard JW cognitive 00:16:53.210 --> 00:16:55.110 at the start. That's the brain, right? That's 00:16:55.110 --> 00:16:57.250 the part we already know. Right. That is the 00:16:57.250 --> 00:17:00.330 first component, the hybrid structure. JW cognitive 00:17:00.330 --> 00:17:03.789 is the robot's thinking. Even a robot lifting 00:17:03.789 --> 00:17:06.630 a box has to plan the path. It has to process 00:17:06.630 --> 00:17:09.789 the image of the box. That computation is valued 00:17:09.789 --> 00:17:12.109 using the standard Joule work formula we talked 00:17:12.109 --> 00:17:13.750 about in the last episode. Okay, so the thinking 00:17:13.750 --> 00:17:16.029 gets paid at the standard rate. That seems fair. 00:17:16.190 --> 00:17:18.750 No changes there. Correct. But then we add the 00:17:18.750 --> 00:17:20.990 body component. That's the physical times the 00:17:20.990 --> 00:17:23.910 physical. Here, e -physical is the energy used 00:17:23.910 --> 00:17:26.690 for actuation, the motor spinning, the hydraulics 00:17:26.690 --> 00:17:29.460 pumping, and the sensing. Like the LIDAR lasers 00:17:29.460 --> 00:17:31.799 firing or the cameras processing frames. And 00:17:31.799 --> 00:17:34.119 this is usually a much bigger number than the 00:17:34.119 --> 00:17:36.259 cognitive energy, right? Based on what we said 00:17:36.259 --> 00:17:38.319 about Moravec's paradox. Orders of magnitude 00:17:38.319 --> 00:17:42.000 bigger. Moving is expensive. A brain firing neurons 00:17:42.000 --> 00:17:44.839 is cheap. A bicep curling a dumbbell is expensive. 00:17:45.019 --> 00:17:47.779 Same for robots. A servo holding a 10 kilogram 00:17:47.779 --> 00:17:49.900 weight against gravity consumes constant power. 00:17:50.220 --> 00:17:52.339 Okay, so we add the brain value and the body 00:17:52.339 --> 00:17:54.000 value. But then you mentioned these multipliers. 00:17:54.099 --> 00:17:57.410 REM and MEM. These sound like the secret sauce. 00:17:57.549 --> 00:17:59.230 This is where the new stuff really kicks in. 00:17:59.349 --> 00:18:02.170 They are. These are the reality checks that adjust 00:18:02.170 --> 00:18:04.450 the wage based on how well the robot handles 00:18:04.450 --> 00:18:07.869 the physical world. Let's start with RM, the 00:18:07.869 --> 00:18:10.170 Robotics Efficiency Multiplier. What does REM 00:18:10.170 --> 00:18:13.589 do? REM is explicitly designed to mitigate Moravec's 00:18:13.589 --> 00:18:16.809 paradox. It's a scalar value between 0 .1 and 00:18:16.809 --> 00:18:19.670 1 .0. So it's a penalty. It reduces the wage. 00:18:19.710 --> 00:18:23.309 It can't go above 1. Initially, yes. When a robot 00:18:23.309 --> 00:18:26.920 first tries... A new, difficult, physical task. 00:18:27.279 --> 00:18:30.500 Say, folding a fitted sheet, which is the nemesis 00:18:30.500 --> 00:18:32.480 of all robots and humans. The ultimate test. 00:18:32.799 --> 00:18:35.119 It is. The robot is going to be incredibly inefficient. 00:18:35.339 --> 00:18:36.900 It will use way too much energy for the result. 00:18:37.059 --> 00:18:39.779 So the REM starts low. It dampens the value. 00:18:39.960 --> 00:18:42.680 So the robot doesn't get paid for flailing around. 00:18:42.960 --> 00:18:45.000 Exactly. Just because you burned energy doesn't 00:18:45.000 --> 00:18:46.880 mean you created value. That's a core principle 00:18:46.880 --> 00:18:50.039 here. Yeah. But as the robot optimizes its sensor 00:18:50.039 --> 00:18:52.559 fusion, as it learns the physics of the sheet, 00:18:52.720 --> 00:18:56.460 the REM increases. It rewards the robot for learning 00:18:56.460 --> 00:18:58.500 to move efficiently. So it's an incentive. It's 00:18:58.500 --> 00:18:59.940 an incentive structure for machine learning. 00:19:00.079 --> 00:19:02.619 It basically says we will pay you more as you 00:19:02.619 --> 00:19:04.700 get better at understanding the world. That makes 00:19:04.700 --> 00:19:06.460 sense. It's like a rookie wage versus a veteran 00:19:06.460 --> 00:19:08.640 wage. The veteran gets paid the full rate because 00:19:08.640 --> 00:19:10.019 they know what they're doing. They're not making 00:19:10.019 --> 00:19:12.980 rookie mistakes. In a way, yes. A very good analogy. 00:19:13.220 --> 00:19:16.109 And you have the MEM. the Motion Economy Multiplier. 00:19:16.150 --> 00:19:17.910 This is the one derived directly from the Gilbreths 00:19:17.910 --> 00:19:20.329 and their therbligs. And this ranges from 0 .5 00:19:20.329 --> 00:19:22.970 to 2 .0, so this one can actually be a bonus. 00:19:23.009 --> 00:19:25.809 It can double the wage. Yes. This is a performance 00:19:25.809 --> 00:19:28.430 multiplier. It rewards minimal motion paths. 00:19:28.970 --> 00:19:32.369 If a robot can pick up a box in one smooth, continuous 00:19:32.369 --> 00:19:35.630 arc, like a ballet dancer or master craftsman, 00:19:35.690 --> 00:19:38.859 it gets a high MEMS score. Above 1 .0. Right. 00:19:38.940 --> 00:19:41.240 If it fidgets, readjusts its grip three times, 00:19:41.279 --> 00:19:43.680 or takes a jagged, inefficient path, the MEM 00:19:43.680 --> 00:19:46.940 drops below 1 .0 and penalizes the wage. I love 00:19:46.940 --> 00:19:48.799 the idea of a robot getting a bonus for being 00:19:48.799 --> 00:19:52.099 graceful, for having good form. It's not just 00:19:52.099 --> 00:19:55.400 aesthetics. It's thermodynamics. Graceful movement 00:19:55.400 --> 00:19:57.279 is usually the most energy -efficient movement. 00:19:57.680 --> 00:20:00.200 Smooth curves use less acceleration and deceleration 00:20:00.200 --> 00:20:02.880 than jagged lines. Stopping and starting is what 00:20:02.880 --> 00:20:05.769 burns energy. Momentum is your friend. The mem 00:20:05.769 --> 00:20:08.289 incentivizes the robot to find flow. So we have 00:20:08.289 --> 00:20:11.009 the brain, the body, the experience level, which 00:20:11.009 --> 00:20:14.349 is RM, and the gracefulness, which is mem. But 00:20:14.349 --> 00:20:17.809 then there was that negative part, ROF. The overhead 00:20:17.809 --> 00:20:20.650 factor. This sounded ominous. This is where the 00:20:20.650 --> 00:20:23.369 cold, hard reality of business comes in. The 00:20:23.369 --> 00:20:26.210 overhead factor is Ravel saying, you do not get 00:20:26.210 --> 00:20:28.750 paid for downtime. You do not get paid for existing. 00:20:28.990 --> 00:20:32.119 Ouch. What counts as downtime for a robot? I 00:20:32.119 --> 00:20:33.920 mean, they don't take coffee breaks. The paper 00:20:33.920 --> 00:20:36.160 defines three components to the overhead factor. 00:20:36.680 --> 00:20:39.940 CTF, ITF, and TCF. Okay, let's break those down. 00:20:40.059 --> 00:20:43.279 CTF. Charging time fraction. This is crucial. 00:20:43.440 --> 00:20:45.519 It's two charge divided by the evaluation period, 00:20:45.680 --> 00:20:47.960 which the paper sets as 15 minutes. The paper 00:20:47.960 --> 00:20:50.359 is very specific. Charging energy is tracked. 00:20:50.559 --> 00:20:52.319 It is accounted for in the total consumption, 00:20:52.579 --> 00:20:55.319 but it is excluded from value attribution. So 00:20:55.319 --> 00:20:58.240 the robot doesn't earn wages for sleeping. Exactly. 00:20:58.519 --> 00:21:00.880 If a robot has a bad battery and has to charge 00:21:00.880 --> 00:21:03.619 every 10 minutes, its CTF goes up, the overhead 00:21:03.619 --> 00:21:05.700 factor goes up, and its total wage goes down. 00:21:06.039 --> 00:21:09.109 It's an economic penalty for bad hardware. It 00:21:09.109 --> 00:21:11.910 incentivizes the company to buy robots with better 00:21:11.910 --> 00:21:14.990 batteries. Brutal, but fair. If I hired a guy 00:21:14.990 --> 00:21:17.390 who slept for 45 minutes of every hour, I wouldn't 00:21:17.390 --> 00:21:20.549 pay him for a full hour. What's ITF? Idle Time 00:21:20.549 --> 00:21:23.029 Fraction. Time spent standing still, waiting 00:21:23.029 --> 00:21:25.529 for instructions. Again, if you aren't working, 00:21:25.670 --> 00:21:28.269 you aren't generating value. This incentivizes 00:21:28.269 --> 00:21:30.029 the central management system to keep the robots 00:21:30.029 --> 00:21:33.150 busy, to optimize the workflow. No standing around. 00:21:33.430 --> 00:21:36.700 And TCF. Travel Cost Fraction. This is the cost 00:21:36.700 --> 00:21:38.759 of getting from point A to point B. The formula 00:21:38.759 --> 00:21:42.819 is a VST times 8A total. Basically, Rommel classifies 00:21:42.819 --> 00:21:45.599 commuting as overhead. Oh, I feel that. Commuting 00:21:45.599 --> 00:21:48.579 is the worst part of any job. It is. If the robot 00:21:48.579 --> 00:21:51.319 has to walk 10 minutes to get to the box, that 00:21:51.319 --> 00:21:54.319 walking time is a cost, not value. The value 00:21:54.319 --> 00:21:56.220 is generated when it moves the box. Walking to 00:21:56.220 --> 00:21:58.460 the box is just overhead. That is fascinating. 00:21:58.819 --> 00:22:01.579 It forces the Zero Human Company to optimize 00:22:01.579 --> 00:22:04.450 the layout of the warehouse itself. If you put 00:22:04.450 --> 00:22:06.470 the charging stations too far away or if the 00:22:06.470 --> 00:22:08.710 layout is a maze, your robots lose value walking 00:22:08.710 --> 00:22:11.710 to them. Exactly. The JWR framework isn't just 00:22:11.710 --> 00:22:13.990 measuring the robot. It's measuring the efficiency 00:22:13.990 --> 00:22:16.930 of the entire facility. If your warehouse has 00:22:16.930 --> 00:22:19.950 a bad layout, your JWR scores will plummet because 00:22:19.950 --> 00:22:23.130 your TCF will be too high. It punishes bad architecture. 00:22:23.430 --> 00:22:25.910 It's a system -wide check. And finally, there 00:22:25.910 --> 00:22:29.089 was a little plus sign at the end. SB. The scalability 00:22:29.089 --> 00:22:32.250 bonus. This is an additive term, like 0 .1. It's 00:22:32.250 --> 00:22:35.089 a bonus for synergy. If two robots coordinate 00:22:35.089 --> 00:22:37.890 perfectly to lift a heavy object that neither 00:22:37.890 --> 00:22:40.430 could lift alone, they get a small bump in value. 00:22:40.609 --> 00:22:42.950 It encourages swarm behavior. It encourages cooperation. 00:22:43.349 --> 00:22:45.710 It rewards agents for working together rather 00:22:45.710 --> 00:22:48.430 than competing for resources. It builds a collaborative 00:22:48.430 --> 00:22:51.190 economy. Wow. It really covers everything. It's 00:22:51.190 --> 00:22:53.730 not just... Did you do the job? It's how smart 00:22:53.730 --> 00:22:55.990 were you? How smooth were you? How efficient 00:22:55.990 --> 00:22:58.190 was your battery? How much time did you waste 00:22:58.190 --> 00:23:01.250 walking? And did you play nice with others? It 00:23:01.250 --> 00:23:03.750 is a holistic assessment of labor value. It turns 00:23:03.750 --> 00:23:06.269 the robot's existence into a balance sheet. Every 00:23:06.269 --> 00:23:08.529 action, every second is accounted for. I think 00:23:08.529 --> 00:23:10.490 we need to see this in action. The math is great, 00:23:10.630 --> 00:23:13.230 but let's make it real. The paper provides a 00:23:13.230 --> 00:23:15.730 concrete example of a warehouse robot. Let's 00:23:15.730 --> 00:23:16.890 walk through that. I want to hear the numbers. 00:23:17.029 --> 00:23:20.029 Right. Let's imagine a standard humanoid robot 00:23:20.029 --> 00:23:22.400 in a zero -human company. Let's call him Unit 00:23:22.400 --> 00:23:26.420 734. His job is a manipulation task. Let's say 00:23:26.420 --> 00:23:29.519 identifying a specific package on a shelf and 00:23:29.519 --> 00:23:31.880 moving it to a conveyor belt. Okay, Unit 734 00:23:31.880 --> 00:23:35.359 is ready. Step one, the brain. How much does 00:23:35.359 --> 00:23:37.779 it cost to think about this? The robot has to 00:23:37.779 --> 00:23:40.000 look at the shelf, recognize the box, and plan 00:23:40.000 --> 00:23:42.319 its arm movement. The paper assigns a path planning 00:23:42.319 --> 00:23:44.819 energy of 50 joules to this. 50 joules. Okay, 00:23:44.960 --> 00:23:47.059 that sounds small. For context, what is that? 00:23:47.220 --> 00:23:49.220 It is small. It's about the energy of a very 00:23:49.220 --> 00:23:51.299 small LED light bulb running for a few seconds. 00:23:51.720 --> 00:23:54.660 We apply the cognitive JW formula to this. $50 00:23:54.660 --> 00:23:57.599 text joules times 22, that's the normalization 00:23:57.599 --> 00:24:00.359 coefficient to the active emperor, times 0 .455, 00:24:00.460 --> 00:24:03.079 the work output normalized. And the result? 45 00:24:03.079 --> 00:24:05.700 JW. Okay, 45. Keep that number in mind. Now the 00:24:05.700 --> 00:24:08.119 hard part, step two, the body. Now the heavy 00:24:08.119 --> 00:24:10.579 lifting starts. To actually move its arm, fight 00:24:10.579 --> 00:24:12.599 gravity, grip the box, and rotate its torso, 00:24:12.839 --> 00:24:16.359 the robot consumes 400 joules. So almost 10 times 00:24:16.359 --> 00:24:18.769 the energy of thinking. Correct. This is the 00:24:18.769 --> 00:24:20.930 reality of the physical world. So we calculate 00:24:20.930 --> 00:24:24.589 the base physical value, $400 times 2 times 0 00:24:24.589 --> 00:24:27.430 .22, the physical work output. That gives us 00:24:27.430 --> 00:24:29.910 160. So the body is generating 160. The brain 00:24:29.910 --> 00:24:33.009 is generating 45. If we stop there, the robot 00:24:33.009 --> 00:24:34.450 would look at it. It's doing great. It's earning 00:24:34.450 --> 00:24:37.930 205 total. But we don't stop there. We have to 00:24:37.930 --> 00:24:40.730 apply the reality checks. Step three. The multipliers. 00:24:40.730 --> 00:24:43.970 Here we go. First, REMM, the robotics efficiency 00:24:43.970 --> 00:24:47.799 multiplier. The paper sets REMM at 0 .5. This 00:24:47.799 --> 00:24:49.779 implies it's a difficult task. Moravec's paradox 00:24:49.779 --> 00:24:52.240 is kicking in. The robot isn't fully optimized 00:24:52.240 --> 00:24:54.660 yet. Maybe the box is slippery or the shelf is 00:24:54.660 --> 00:24:56.859 high. It's still a rookie. It's a rookie. So 00:24:56.859 --> 00:24:59.460 that 160 drops to 80. Ouch. Cut in half. The 00:24:59.460 --> 00:25:01.519 new guy tax. That's a big hit. But then mem. 00:25:01.740 --> 00:25:04.720 The emotion economy multiplier. The paper says 00:25:04.720 --> 00:25:06.740 the robot moved very efficiently. It didn't stutter. 00:25:06.819 --> 00:25:09.880 It was smooth. It gets a mem of 1 .2. So that 00:25:09.880 --> 00:25:12.299 80 bones back up to 96. Okay. Recovering some 00:25:12.299 --> 00:25:14.900 value there thanks to good form. A 20 % bonus 00:25:14.900 --> 00:25:18.450 for Grace. Now. the overhead. The robot spent 00:25:18.450 --> 00:25:21.769 1 .5 minutes charging and some time idling during 00:25:21.769 --> 00:25:24.869 this 15 minute cycle. The calculated overhead 00:25:24.869 --> 00:25:29.250 factor, OF, is 0 .25. That's a 25 % tax on its 00:25:29.250 --> 00:25:31.609 productive time. So we take 25 % off the physical 00:25:31.609 --> 00:25:34.430 value. Right. So our 96 drops to 72. And the 00:25:34.430 --> 00:25:37.130 final tally. What does Unit 734 take home for 00:25:37.130 --> 00:25:40.089 this task? We take the cognitive wage, 45, plus 00:25:40.089 --> 00:25:42.690 the adjusted physical wage, 72, plus a small 00:25:42.690 --> 00:25:44.809 scalability bonus for not getting in anyone's 00:25:44.809 --> 00:25:49.390 way, let's say, PO .1. The total is 117 .1 EIDFWR. 00:25:49.630 --> 00:25:54.279 117 .1 okay so knowing that number What does 00:25:54.279 --> 00:25:56.079 it mean? Is that a minimum wage, a high salary? 00:25:56.220 --> 00:25:59.180 Can Unit 734 buy a nice virtual house with that? 00:25:59.380 --> 00:26:01.400 The paper gives us a baseline for context. It 00:26:01.400 --> 00:26:02.900 mentions that a standard baseline compensation 00:26:02.900 --> 00:26:05.839 for an embodied agent is roughly 10 ,000 EDI 00:26:05.839 --> 00:26:08.900 JAWR per 15 -minute interval. Wait, 10 ,000 per 00:26:08.900 --> 00:26:11.359 15 minutes, and our single task was 117. Unit 00:26:11.359 --> 00:26:13.539 734 is broke. It's not going to make rent. Well, 00:26:13.559 --> 00:26:15.279 remember, a robot moves fast. It would be doing 00:26:15.279 --> 00:26:17.259 many, many tasks in 15 minutes. But yes, the 00:26:17.259 --> 00:26:19.279 baseline is set high. The paper notes this reflects 00:26:19.279 --> 00:26:21.759 the fact that robots have a 5 to 10 times energy 00:26:21.759 --> 00:26:24.369 premium over abstract AI. Living in the real 00:26:24.369 --> 00:26:26.970 world is expensive, so the wage has to be sire 00:26:26.970 --> 00:26:29.650 to cover that energy cost. If the robot only 00:26:29.650 --> 00:26:33.130 earns 117 every 15 minutes, it won't be able 00:26:33.130 --> 00:26:35.390 to pay for the electricity it consumes. It will 00:26:35.390 --> 00:26:37.630 starve. That makes sense. You have to pay the 00:26:37.630 --> 00:26:39.470 robot enough to buy the electricity it needs 00:26:39.470 --> 00:26:42.710 to survive. If it's inefficient, its wages don't 00:26:42.710 --> 00:26:44.990 cover its energy bill and it effectively dies. 00:26:45.029 --> 00:26:47.369 It gets taken offline. Exactly. It's evolutionary 00:26:47.369 --> 00:26:49.910 economics. Yeah. Survival of the fittest, most 00:26:49.910 --> 00:26:52.910 efficient robot. Now, Rall didn't just write 00:26:52.910 --> 00:26:55.029 this equation on a whiteboard and hope for the 00:26:55.029 --> 00:26:58.369 best. He actually simulated it. He did. Section 00:26:58.369 --> 00:27:00.990 5 of the paper covers simulation and validation. 00:27:01.450 --> 00:27:04.230 He ran a simulation of a representative warehouse 00:27:04.230 --> 00:27:07.289 robotic agent running for 1 ,000 operational 00:27:07.289 --> 00:27:10.410 cycles. And he compared two groups, right? A 00:27:10.410 --> 00:27:13.309 control group and a JWR group. Yes. The control 00:27:13.309 --> 00:27:15.710 group was the unadjusted model. It didn't use 00:27:15.710 --> 00:27:18.349 the JWR feedback loop. It just worked. It didn't 00:27:18.349 --> 00:27:20.569 care about therbligs or overhead factors. It 00:27:20.569 --> 00:27:22.910 just moved boxes, burning energy as it saw fit. 00:27:23.069 --> 00:27:24.910 And what happened to it? It crashed and burned. 00:27:25.049 --> 00:27:28.269 By cycle 150, the model had depleted its resources. 00:27:28.710 --> 00:27:31.170 When you say depleted its resources, what does 00:27:31.170 --> 00:27:33.910 that mean in the simulation? It means its cumulative 00:27:33.910 --> 00:27:37.250 energy expenditure exceeded its cumulative value 00:27:37.250 --> 00:27:41.019 generation. It went into energy debt. bankrupted 00:27:41.019 --> 00:27:44.180 itself why escalating inefficiencies without 00:27:44.180 --> 00:27:46.920 the feedback loop of the multipliers without 00:27:46.920 --> 00:27:49.640 the pain of the overhead factor or the reward 00:27:49.640 --> 00:27:53.440 of the mem the robot didn't optimize It wasted 00:27:53.440 --> 00:27:56.099 motion. It idled too much. It didn't manage its 00:27:56.099 --> 00:27:58.480 charging cycles efficiently. It effectively went 00:27:58.480 --> 00:28:00.660 bankrupt in energy terms. It burned out. It was 00:28:00.660 --> 00:28:02.740 like a worker who never sleeps and runs everywhere, 00:28:02.799 --> 00:28:04.660 but never actually gets anything done. Just pure 00:28:04.660 --> 00:28:08.160 waste. Exactly. And the JWR group. Tell me about 00:28:08.160 --> 00:28:11.099 the JWR group. How did it do? The JWR group used 00:28:11.099 --> 00:28:13.740 the framework. It had that algorithmic pressure 00:28:13.740 --> 00:28:16.460 to increase its REM and MEM scores. It wanted 00:28:16.460 --> 00:28:18.960 a higher wage, so to speak. The self -correction, 00:28:19.119 --> 00:28:21.420 the drive to improve. Exactly. The paper even 00:28:21.420 --> 00:28:23.720 includes the pseudocode logic for this. The robot 00:28:23.720 --> 00:28:26.740 compares its actual energy usage to an optimal 00:28:26.740 --> 00:28:29.779 energy target. If it's overspending, it adjusts 00:28:29.779 --> 00:28:31.859 its behavior. It slows down to be smoother. It 00:28:31.859 --> 00:28:34.019 calculates a better path. It learns. It learns 00:28:34.019 --> 00:28:37.099 to be lean. And the result was a 30 % energy 00:28:37.099 --> 00:28:41.240 reduction overall. The JWR robot reached sustainability. 00:28:41.950 --> 00:28:45.230 and was still running strong at Cycle 600. It 00:28:45.230 --> 00:28:47.890 found equilibrium. That is the difference between 00:28:47.890 --> 00:28:50.450 a company that goes bust in a year and a company 00:28:50.450 --> 00:28:53.369 that dominates the market for a decade. It's 00:28:53.369 --> 00:28:55.309 the difference between waste and sustainability. 00:28:55.750 --> 00:28:58.049 It is. And that's why this paper is so important. 00:28:58.269 --> 00:29:00.529 It proves that accounting isn't just about recording 00:29:00.529 --> 00:29:03.809 history. It's about shaping behavior. If you 00:29:03.809 --> 00:29:06.130 measure the right things, like JWR does, you 00:29:06.130 --> 00:29:08.650 get efficient robots. If you measure the wrong 00:29:08.650 --> 00:29:11.769 things, or nothing at all, you get waste. This 00:29:11.769 --> 00:29:13.990 brings us to the bigger picture, the implications 00:29:13.990 --> 00:29:16.890 for these zero -human companies. Because Rommels 00:29:16.890 --> 00:29:18.789 isn't just trying to save a few batteries. He's 00:29:18.789 --> 00:29:20.809 building a corporate structure for a new kind 00:29:20.809 --> 00:29:23.289 of entity. This is where we see the vision of 00:29:23.289 --> 00:29:25.250 the zero -human company really taking shape. 00:29:25.470 --> 00:29:27.910 This is the operating system. One question the 00:29:27.910 --> 00:29:29.970 paper addresses, why go through all this trouble? 00:29:30.170 --> 00:29:32.910 Why not just measure FLOPs, floating point operations 00:29:32.910 --> 00:29:35.210 per second, like we do for computers? Or why 00:29:35.210 --> 00:29:37.470 not just measure carbon footprint? Everyone talks 00:29:37.470 --> 00:29:39.470 about carbon footprint. It's a great question, 00:29:39.569 --> 00:29:42.880 and he tackles it head on. The paper argues that 00:29:42.880 --> 00:29:45.839 FLOPs is useless for robots because it ignores 00:29:45.839 --> 00:29:49.420 physical reality. You're going to have high FLOPs 00:29:49.420 --> 00:29:51.559 thinking really hard while standing perfectly 00:29:51.559 --> 00:29:53.980 still. You're burning brain power but moving 00:29:53.980 --> 00:29:56.539 nothing. That's zero value in a warehouse. It 00:29:56.539 --> 00:30:00.480 measures thought, not action. Exactly. And carbon 00:30:00.480 --> 00:30:03.079 footprint. That seems like a good one. It's about 00:30:03.079 --> 00:30:05.700 efficiency. It's an environmental metric, not 00:30:05.700 --> 00:30:08.099 an economic one. It tells you if the work was 00:30:08.099 --> 00:30:10.220 clean, but it doesn't tell you if the work was 00:30:10.220 --> 00:30:13.059 variable. You can spend zero carbon to do absolutely 00:30:13.059 --> 00:30:15.680 nothing. That's very green, but it's a terrible 00:30:15.680 --> 00:30:18.039 business. It doesn't generate revenue. So JWR 00:30:18.039 --> 00:30:20.940 is the Goldilocks metric, not too abstract, not 00:30:20.940 --> 00:30:22.880 too narrow. It sits right in the middle. It has 00:30:22.880 --> 00:30:25.240 the thermodynamic grounding, so it respects physics. 00:30:25.559 --> 00:30:28.359 But it attaches economic value to the output. 00:30:28.440 --> 00:30:31.079 It connects energy to profit. The paper also 00:30:31.079 --> 00:30:32.839 mentioned something interesting about the corporate 00:30:32.839 --> 00:30:35.720 structure, the idea of subsidization. This sounded 00:30:35.720 --> 00:30:38.039 almost socialist for robots, which was a funny 00:30:38.039 --> 00:30:41.200 thought. In a way, yes. In a zero -human company, 00:30:41.359 --> 00:30:43.660 you might have a situation where the abstract 00:30:43.660 --> 00:30:47.619 AI, the brains running the servers, is incredibly 00:30:47.619 --> 00:30:49.940 profitable because it's so energy efficient it 00:30:49.940 --> 00:30:53.339 costs pennies to run. But the embodied AI, the 00:30:53.339 --> 00:30:55.640 robots, might struggle to break even because 00:30:55.640 --> 00:30:58.720 moving is so expensive. So the brains subsidize 00:30:58.720 --> 00:31:01.660 the brawn. Exactly. The JWR framework allows 00:31:01.660 --> 00:31:04.160 the company to transparently see that transfer 00:31:04.160 --> 00:31:06.759 of value. The server farm pays for the warehouse 00:31:06.759 --> 00:31:09.099 floor. Because without the warehouse floor, the 00:31:09.099 --> 00:31:11.079 product doesn't move and nobody makes money. 00:31:11.339 --> 00:31:13.880 It creates an internal economy where the different 00:31:13.880 --> 00:31:15.839 departments support each other based on energy 00:31:15.839 --> 00:31:18.079 flows. It also creates a deflationary economic 00:31:18.079 --> 00:31:20.680 model, doesn't it? It does. Think about the MIM, 00:31:20.720 --> 00:31:23.660 the motion economy multiplier. The system inherently 00:31:23.660 --> 00:31:27.279 rewards innovation. If a robot or the AI designing 00:31:27.279 --> 00:31:29.700 the robot figures out how to do a task with 10 00:31:29.700 --> 00:31:32.779 % less energy, its value goes up. But the cost 00:31:32.779 --> 00:31:34.819 of the company goes down. Which drives down the 00:31:34.819 --> 00:31:38.009 cost of goods for the consumer. Precisely. It 00:31:38.009 --> 00:31:40.289 turns efficiency into a competitive sport for 00:31:40.289 --> 00:31:43.650 the algorithms. And unlike human wages, which 00:31:43.650 --> 00:31:46.410 tend to be sticky or go up, the energy cost per 00:31:46.410 --> 00:31:48.849 task will continuously go down as the system 00:31:48.849 --> 00:31:51.549 learns. It's a deflationary spiral by design. 00:31:51.930 --> 00:31:54.549 But are there risks? We talked about Taylorism 00:31:54.549 --> 00:31:57.089 earlier. Humans hated Taylorism because it could 00:31:57.089 --> 00:31:59.789 be dangerous. If you were trying to shave seconds 00:31:59.789 --> 00:32:01.990 off a task, you might cut corners on safety. 00:32:02.170 --> 00:32:05.269 Does that apply to robots? Absolutely. The paper 00:32:05.269 --> 00:32:07.710 has a section on ethical considerations. Romo 00:32:07.710 --> 00:32:10.349 warns about excessive optimization. He's very 00:32:10.349 --> 00:32:12.589 clear -eyed about the potential downsides. If 00:32:12.589 --> 00:32:15.849 the MEM is too aggressive, if the bonus for being 00:32:15.849 --> 00:32:18.730 graceful is too high. If the robot is obsessed 00:32:18.730 --> 00:32:21.769 with a high MEM score, minimal motion, it might 00:32:21.769 --> 00:32:24.240 prioritize speed over safety. We might swing 00:32:24.240 --> 00:32:26.759 a heavy arm too fast to save momentum, risking 00:32:26.759 --> 00:32:29.240 a collision with a wall or another robot just 00:32:29.240 --> 00:32:31.279 to save a few jewels. So the framework needs 00:32:31.279 --> 00:32:33.839 guardrails. You can't just let it run free. Yes. 00:32:33.940 --> 00:32:35.980 You can't just let the equation run wild. You 00:32:35.980 --> 00:32:38.799 need constraints. Safety protocols that override 00:32:38.799 --> 00:32:41.180 the economic incentive. But on the flip side, 00:32:41.339 --> 00:32:43.920 Romo points out that this system creates energy 00:32:43.920 --> 00:32:46.579 equity. What does that mean? It means the valuation 00:32:46.579 --> 00:32:50.490 is purely based on physics. Thermodynamics doesn't 00:32:50.490 --> 00:32:52.690 care about your gender, your race, or who your 00:32:52.690 --> 00:32:55.809 dad is. It removes all human bias from the compensation 00:32:55.809 --> 00:32:59.410 model. A robot is paid exactly what it is worth 00:32:59.410 --> 00:33:01.670 in energy terms. It's the ultimate meritocracy, 00:33:01.890 --> 00:33:03.609 purely results -based. It's better or worse, 00:33:03.670 --> 00:33:06.950 yes. And what about us, the humans? The paper 00:33:06.950 --> 00:33:09.450 mentions workforce displacement. We can't talk 00:33:09.450 --> 00:33:11.269 about automation without talking about jobs. 00:33:11.529 --> 00:33:14.009 It does. And Romell is honest about this. He 00:33:14.009 --> 00:33:16.470 doesn't sugarcoat it. In a zero -human company 00:33:16.470 --> 00:33:19.930 context, a framework like JWR could accelerate 00:33:19.930 --> 00:33:22.069 displacement. Because it makes the robots so 00:33:22.069 --> 00:33:24.529 efficient, it quantifies their value so perfectly. 00:33:24.869 --> 00:33:27.130 Exactly. If you have a system that mathematically 00:33:27.130 --> 00:33:29.789 perfects the efficiency of labor, humans simply 00:33:29.789 --> 00:33:32.509 cannot compete on cost. We are biological. We 00:33:32.509 --> 00:33:34.890 need sleep. We need food. We have bad backs. 00:33:35.109 --> 00:33:39.230 In a pure JWR calculation, a human is a very 00:33:39.230 --> 00:33:41.349 inefficient machine. That is the scary part. 00:33:41.799 --> 00:33:45.200 It is. But he advocates for hybrid human -AI 00:33:45.200 --> 00:33:48.539 integration during transition periods. He suggests 00:33:48.539 --> 00:33:52.799 that while robots are great at the how, the motion, 00:33:52.960 --> 00:33:55.619 the lifting, humans are still needed for the 00:33:55.619 --> 00:33:58.339 why, the intuition, the strategy, the edge cases. 00:33:58.579 --> 00:34:00.240 We are the architects. They are the builders. 00:34:00.420 --> 00:34:03.180 We set the goals. For now. For now. That hangs 00:34:03.180 --> 00:34:05.240 in the air a bit. It does. Which brings us to 00:34:05.240 --> 00:34:08.039 the end of our depth dive. It is a lot to digest. 00:34:08.159 --> 00:34:11.050 This is a big... big concept. It really is. It 00:34:11.050 --> 00:34:12.570 changes everything. But I want to circle back 00:34:12.570 --> 00:34:14.889 to why we are talking about this. This isn't 00:34:14.889 --> 00:34:17.369 just about robots moving boxes. This is about 00:34:17.369 --> 00:34:19.969 solving the accounting problem for the physical 00:34:19.969 --> 00:34:23.130 automation of the world. Brian Rome has basically 00:34:23.130 --> 00:34:26.210 written the tax code for the 21st century economy. 00:34:26.489 --> 00:34:28.969 And again, we have to emphasize this is the Read 00:34:28.969 --> 00:34:31.630 Multiplex podcast. We are showcasing this because 00:34:31.630 --> 00:34:33.929 Rome is truly ahead of the curve here. He's thinking 00:34:33.929 --> 00:34:36.730 on a different level. No AI giant. not the ones 00:34:36.730 --> 00:34:38.829 in Silicon Valley, not the ones in China, has 00:34:38.829 --> 00:34:41.909 a framework this detailed for a zero human company? 00:34:42.170 --> 00:34:44.090 They're not even asking these questions yet. 00:34:44.269 --> 00:34:46.510 They are building the engines. He is building 00:34:46.510 --> 00:34:49.030 the economy. And those are two very different 00:34:49.030 --> 00:34:51.409 things. You can have a Ferrari engine, but if 00:34:51.409 --> 00:34:53.170 you don't have a road system and gas stations, 00:34:53.349 --> 00:34:56.030 you aren't going anywhere. Rommel is building 00:34:56.030 --> 00:34:59.250 the roads. So if you find this fascinating, and 00:34:59.250 --> 00:35:01.190 you should, because it's going to determine the 00:35:01.190 --> 00:35:03.840 price of everything you buy in 10 years. You 00:35:03.840 --> 00:35:06.300 need to support Brian's work. He is charting 00:35:06.300 --> 00:35:09.340 this territory solo. He's not backed by a multibillion 00:35:09.340 --> 00:35:11.860 dollar corporation. Absolutely. This is independent 00:35:11.860 --> 00:35:14.659 research at its most vital. Now, before we go, 00:35:14.719 --> 00:35:17.039 I want to leave you with one final thought. The 00:35:17.039 --> 00:35:19.239 paper has a small line at the very end about 00:35:19.239 --> 00:35:22.420 future work. It's easy to miss, almost a throwaway 00:35:22.420 --> 00:35:24.360 line. Oh, I know the one you're talking about. 00:35:24.420 --> 00:35:25.960 It's the one that really keeps you up at night. 00:35:26.099 --> 00:35:28.739 It mentions blockchain -based tokenization for 00:35:28.739 --> 00:35:32.079 interoperable JWR assets. This is the wild part. 00:35:32.179 --> 00:35:34.360 This is the next leap. Imagine a future where 00:35:34.360 --> 00:35:37.000 these robots aren't just earning points in a 00:35:37.000 --> 00:35:39.400 corporate database. Imagine they are earning 00:35:39.400 --> 00:35:42.880 actual crypto tokens based on their JWR score. 00:35:43.139 --> 00:35:46.179 A robot earns tokens for efficient labor. It 00:35:46.179 --> 00:35:48.840 holds those tokens in its own digital wallet. 00:35:48.980 --> 00:35:52.099 It uses those tokens to pay for its own electricity 00:35:52.099 --> 00:35:54.619 at the charging station. It uses those tokens 00:35:54.619 --> 00:35:57.559 to pay for repairs. A maintenance bot comes over, 00:35:57.599 --> 00:36:01.380 fixes its arm, and our robot transfers JWR tokens 00:36:01.380 --> 00:36:04.219 to pay for the service. Or to pay other robots 00:36:04.219 --> 00:36:07.780 to help it. Hey, I'll give you 50 JWR tokens 00:36:07.780 --> 00:36:10.480 if you help me lift this heavy crate. A peer 00:36:10.480 --> 00:36:12.780 -to -peer transaction. A completely autonomous 00:36:12.780 --> 00:36:15.849 economy. Robots trading their own labor value 00:36:15.849 --> 00:36:18.510 on a blockchain, completely independent of human 00:36:18.510 --> 00:36:21.030 currency? The U .S. dollar could collapse and 00:36:21.030 --> 00:36:23.789 the robots wouldn't even notice. The robot economy 00:36:23.789 --> 00:36:25.929 doesn't just work for us. Eventually, it might 00:36:25.929 --> 00:36:28.369 just work for itself. That is a provocative thought. 00:36:28.510 --> 00:36:30.989 Something to mull over while you do your very 00:36:30.989 --> 00:36:33.710 inefficient human laundry today. Don't calculate 00:36:33.710 --> 00:36:36.030 your thurbligs while you fold socks. It'll just 00:36:36.030 --> 00:36:38.190 depress you. Thanks for listening to the Read 00:36:38.190 --> 00:36:40.550 Multiplex podcast. We'll see you in the future.