WEBVTT 00:00:00.000 --> 00:00:02.879 I watched this video recently and it really stuck 00:00:02.879 --> 00:00:05.759 with me. You see these six small agile robots. 00:00:05.980 --> 00:00:07.580 Yeah. And they're not just, you know, standing 00:00:07.580 --> 00:00:10.619 there. They're executing this perfectly synchronized 00:00:10.619 --> 00:00:13.880 front flip, like a dance. The physical control 00:00:13.880 --> 00:00:16.859 was just... Mesmerizing. It really was. That 00:00:16.859 --> 00:00:19.039 physical precision, that mastery of gravity, 00:00:19.179 --> 00:00:21.260 it's incredible. Now, contrast that with the 00:00:21.260 --> 00:00:23.320 digital side, the frustration we all feel, right? 00:00:23.399 --> 00:00:26.800 You spend 20 minutes writing this perfect, complex 00:00:26.800 --> 00:00:29.820 pomp for an AI video tool, and the model basically 00:00:29.820 --> 00:00:33.380 ignores 90 % of it. Welcome to the Deep Dive. 00:00:33.840 --> 00:00:35.880 Today, we're tackling a stack of sources that 00:00:35.880 --> 00:00:38.159 I think perfectly encapsulate this dual revolution 00:00:38.159 --> 00:00:40.619 happening in tech right now. We're looking at 00:00:40.619 --> 00:00:43.659 hardware that seems to defy physics through hyperefficiency. 00:00:44.179 --> 00:00:46.380 And at the same time, software that demands we 00:00:46.380 --> 00:00:48.640 become hyperefficient just to talk to it. And 00:00:48.640 --> 00:00:51.399 our mission in this deep dive is to quickly distill 00:00:51.399 --> 00:00:53.039 what you need to know about this new landscape. 00:00:53.500 --> 00:00:55.520 We're going to cover the specific physics that 00:00:55.520 --> 00:00:59.039 forces humanoids to stay short and fast. Then, 00:00:59.140 --> 00:01:01.399 the essential seven rules for video prompting 00:01:01.399 --> 00:01:04.269 that actually work. The staggering reality of 00:01:04.269 --> 00:01:06.730 AI infrastructure costs. That's the real shocker. 00:01:06.769 --> 00:01:09.530 And yeah, Disney's ingenious, temperature -aware 00:01:09.530 --> 00:01:12.609 Olaf robot. It's all connected by a single theme. 00:01:13.010 --> 00:01:15.590 The ruthless need for efficient resource management. 00:01:15.810 --> 00:01:17.769 Okay, let's dive straight into the physical world 00:01:17.769 --> 00:01:21.129 first. Those synchronized unitary G1 flips? This 00:01:21.129 --> 00:01:23.430 wasn't one robot doing a slow move in a lab. 00:01:23.590 --> 00:01:26.409 This was six of them. performing a complicated 00:01:26.409 --> 00:01:29.450 dynamic dance routine all in perfect sync. It 00:01:29.450 --> 00:01:32.370 even caught Elon Musk's eye. What's fascinating 00:01:32.370 --> 00:01:34.209 here is understanding why they succeed where, 00:01:34.269 --> 00:01:36.650 you know, these larger platforms tend to struggle. 00:01:36.790 --> 00:01:38.930 Most flashy humanoid demos are single units. 00:01:38.969 --> 00:01:40.930 They're heavily scripted, and they often struggle 00:01:40.930 --> 00:01:43.890 with heat and endurance. The Unitree G1 was engineered 00:01:43.890 --> 00:01:46.569 to be super athletic, but it got there through 00:01:46.569 --> 00:01:49.269 strategic subtraction, not addition. They are 00:01:49.269 --> 00:01:51.370 noticeably short, right? Only around, what, 1 00:01:51.370 --> 00:01:55.930 .27 to 1 .32 meters tall? Weighing about 35 kilograms. 00:01:56.310 --> 00:01:58.650 That short stature isn't a design limitation. 00:01:58.870 --> 00:02:01.790 It's a fundamental engineering choice. And it's 00:02:01.790 --> 00:02:04.530 driven entirely by physics and economics. Exactly. 00:02:04.549 --> 00:02:07.010 The sources make a clear case for this minimum 00:02:07.010 --> 00:02:09.729 viable robot approach. Shorter robots, I mean, 00:02:09.729 --> 00:02:11.949 it's common sense. They need less material, so 00:02:11.949 --> 00:02:14.629 that reduces cost. And they can use smaller, 00:02:14.710 --> 00:02:16.870 less powerful actuators, which are basically 00:02:16.870 --> 00:02:19.129 the motors that drive movement. And it goes beyond 00:02:19.129 --> 00:02:21.490 just the hardware cost. There's logistics. The 00:02:21.490 --> 00:02:25.069 G1 folds down. dramatically to just 690 millimeters. 00:02:25.229 --> 00:02:27.710 That makes shipping, storage, all of that so 00:02:27.710 --> 00:02:29.990 much cheaper. But I think the real secret weapon, 00:02:30.090 --> 00:02:32.590 the thing that makes these flips possible is 00:02:32.590 --> 00:02:34.509 torque management. That's the critical detail. 00:02:34.750 --> 00:02:36.969 Shorter limbs means significantly lower torque 00:02:36.969 --> 00:02:39.210 demands on the motors. Torque is just force applied 00:02:39.210 --> 00:02:40.990 over a distance. So if you have a longer arm, 00:02:41.050 --> 00:02:42.650 the motor has to fight way harder. It demands 00:02:42.650 --> 00:02:45.169 more power. And when a motor demands less torque, 00:02:45.289 --> 00:02:47.310 it operates faster. And crucially, it generates 00:02:47.310 --> 00:02:50.629 way less waste heat. So it's a heat problem masquerading 00:02:50.629 --> 00:02:54.250 as a physics problem. That low torque lets the 00:02:54.250 --> 00:02:56.949 robot accelerate quicker, execute those sudden 00:02:56.949 --> 00:02:59.710 powerful movements, even with pretty modest motors. 00:02:59.870 --> 00:03:02.009 They can hit peak performance without crossing 00:03:02.009 --> 00:03:04.110 that thermal red line. And you can't overlook 00:03:04.110 --> 00:03:06.110 the practical benefits, especially as they move 00:03:06.110 --> 00:03:09.229 into the real world. 130 centimeter body is just 00:03:09.229 --> 00:03:11.330 inherently safer and less threatening around 00:03:11.330 --> 00:03:14.150 people. It fits into existing workspaces and 00:03:14.150 --> 00:03:17.289 homes way better than some hulking six foot metal 00:03:17.289 --> 00:03:19.750 frame. And the affordability point is a big one. 00:03:19.830 --> 00:03:21.389 The hardware is coming out of China at around 00:03:21.389 --> 00:03:24.949 $16 ,000, which, OK, it's still a significant 00:03:24.949 --> 00:03:27.409 investment, but it is magnitudes cheaper than 00:03:27.409 --> 00:03:30.569 other humanoids. But I have to ask, is $16 ,000 00:03:30.569 --> 00:03:33.550 truly accessible? Or is it just affordable for 00:03:33.550 --> 00:03:36.430 a robot? It's a great question. The sources suggest 00:03:36.430 --> 00:03:38.870 it pushes these robot platforms closer to the 00:03:38.870 --> 00:03:41.629 price point of, say, specialized industrial equipment, 00:03:41.810 --> 00:03:44.430 which means they could become pretty common in 00:03:44.430 --> 00:03:46.590 technical classrooms or research labs very soon. 00:03:46.710 --> 00:03:48.889 It democratizes the platform, you know, even 00:03:48.889 --> 00:03:51.250 if it's not a household item tomorrow. The focus 00:03:51.250 --> 00:03:54.169 on efficiency driven by physics directly translates 00:03:54.169 --> 00:03:56.669 into a cost that just opens the door for so many 00:03:56.669 --> 00:03:59.729 more developers. So if the G1's efficiency is 00:03:59.729 --> 00:04:02.400 all about minimizing torque, to manage heat and 00:04:02.400 --> 00:04:05.740 speed. What does that tell us about the future 00:04:05.740 --> 00:04:08.639 of physical robot design? It tells us the primary 00:04:08.639 --> 00:04:11.520 goal is rapid, safe movement. It's driven by 00:04:11.520 --> 00:04:14.000 low -demand actuators, prioritizing efficiency 00:04:14.000 --> 00:04:16.920 over just trying to imitate human scale. That's 00:04:16.920 --> 00:04:18.939 a perfect pivot. If managing physical torque 00:04:18.939 --> 00:04:21.100 makes hardware efficient, we need that same kind 00:04:21.100 --> 00:04:22.779 of resource efficiency on the software side, 00:04:22.899 --> 00:04:25.540 because unstructured, wasted input is just as 00:04:25.540 --> 00:04:27.699 costly. Let's shift to that digital frustration 00:04:27.699 --> 00:04:30.439 of massive AI models and prompt mastery. Oh, 00:04:30.540 --> 00:04:32.899 this segment hits home for everyone who uses 00:04:32.899 --> 00:04:35.879 generative AI. It's that punch your screen moment. 00:04:36.060 --> 00:04:39.439 You spend all this time crafting a long, detailed 00:04:39.439 --> 00:04:42.220 prompt for a video tool only to get something 00:04:42.220 --> 00:04:45.819 blurry, broken, or it just completely ignores 00:04:45.819 --> 00:04:48.079 half of your story. I still wrestle with prompt 00:04:48.079 --> 00:04:50.100 drift myself. You know, you start with this amazing 00:04:50.100 --> 00:04:52.720 idea, but as the prompt gets longer and more 00:04:52.720 --> 00:04:55.220 complex, the model just seems to drift off. It 00:04:55.220 --> 00:04:57.500 loses the plot. It's a universal vulnerability 00:04:57.500 --> 00:04:59.649 when we're interacting with these systems. Well, 00:04:59.709 --> 00:05:02.410 the sources reveal the truth here. Most AI video 00:05:02.410 --> 00:05:05.529 tools are highly sensitive to unstructured input. 00:05:05.730 --> 00:05:09.029 They often ignore 90 % of a long rambling prompt. 00:05:09.329 --> 00:05:11.810 This happens because the model's attention mechanism, 00:05:12.129 --> 00:05:13.889 that's how it weighs the importance of different 00:05:13.889 --> 00:05:16.310 words, it just gets overwhelmed. It loses the 00:05:16.310 --> 00:05:18.250 signal and all the noise. So it's not that the 00:05:18.250 --> 00:05:21.009 AI can't read all the words. It just can't effectively 00:05:21.009 --> 00:05:23.889 process them when they're in, like, an unstructured 00:05:23.889 --> 00:05:27.310 word dump. Exactly. The solution isn't some magic 00:05:27.310 --> 00:05:29.850 words. It's structure. you have to teach the 00:05:29.850 --> 00:05:32.250 model how to parse your command. Think of it 00:05:32.250 --> 00:05:34.949 like a recipe versus a random grocery list. A 00:05:34.949 --> 00:05:37.230 grocery list has all the parts, but a recipe 00:05:37.230 --> 00:05:39.709 gives the AI the order, the measurements, the 00:05:39.709 --> 00:05:42.439 method. The newsletter highlighted seven dead 00:05:42.439 --> 00:05:45.240 simple prompt styles that work across multiple 00:05:45.240 --> 00:05:49.860 platforms, like VO, Sora, Pika, Runway. Can you 00:05:49.860 --> 00:05:51.779 give us a concrete example of what a structured 00:05:51.779 --> 00:05:54.120 style looks like? Absolutely. One of the most 00:05:54.120 --> 00:05:55.939 effective methods is what they call the reporter 00:05:55.939 --> 00:05:58.699 style. So instead of writing, a dog runs through 00:05:58.699 --> 00:06:00.939 a park, you structure it. You use that classic 00:06:00.939 --> 00:06:03.519 who, what, where, when, and why framework. You 00:06:03.519 --> 00:06:06.480 define the setting where, a foggy autumnal park. 00:06:06.720 --> 00:06:09.540 The action what, a golden retriever sprints. 00:06:09.959 --> 00:06:13.879 The style Y, shot in 4K cinematic lighting. And 00:06:13.879 --> 00:06:17.180 the focus who? The specific dog. That makes so 00:06:17.180 --> 00:06:18.819 much sense. You're giving the AI these explicit 00:06:18.819 --> 00:06:20.560 categories instead of just hoping it figures 00:06:20.560 --> 00:06:22.680 out the hierarchy on its own. Another style they 00:06:22.680 --> 00:06:24.379 mentioned was chain of thought prompting, which 00:06:24.379 --> 00:06:26.620 is basically telling the model, think step by 00:06:26.620 --> 00:06:29.370 step. And that's critical for complex tasks. 00:06:29.629 --> 00:06:32.250 Chain of thought prompting forces the large language 00:06:32.250 --> 00:06:34.750 model. That's just an AI trained on vast data 00:06:34.750 --> 00:06:37.810 to generate human -like text to pause. It makes 00:06:37.810 --> 00:06:40.129 it lay out its logic internally before it gives 00:06:40.129 --> 00:06:43.050 you the final output. It massively improves accuracy. 00:06:43.350 --> 00:06:46.300 It's like asking the AI to show its work. And 00:06:46.300 --> 00:06:48.000 this isn't just about getting a better video. 00:06:48.199 --> 00:06:50.639 It's about making better use of the limited compute 00:06:50.639 --> 00:06:54.379 time that AI is dedicating to your request. Wasted 00:06:54.379 --> 00:06:58.120 input is wasted resources. Precisely. If you're 00:06:58.120 --> 00:07:00.160 a beginner, the fundamental mistake to correct 00:07:00.160 --> 00:07:03.240 is thinking that length equals quality. Focus 00:07:03.240 --> 00:07:05.800 on those structured styles. So the core lesson 00:07:05.800 --> 00:07:08.279 is to stop throwing unstructured words at the 00:07:08.279 --> 00:07:11.379 model. and instead focus on using defined structured 00:07:11.379 --> 00:07:13.860 styles to manage that information load. That's 00:07:13.860 --> 00:07:16.079 the entire game. Now, connecting this efficiency 00:07:16.079 --> 00:07:19.259 challenge to the bigger picture, the whole AI 00:07:19.259 --> 00:07:22.680 ecosystem is shifting so incredibly fast, especially 00:07:22.680 --> 00:07:24.819 with infrastructure and scale. And this brings 00:07:24.819 --> 00:07:27.779 us to the staggering financial reality of the 00:07:27.779 --> 00:07:29.920 compute crisis. Yeah, the sources documented 00:07:29.920 --> 00:07:32.279 a dizzying amount of activity from the major 00:07:32.279 --> 00:07:35.800 players. I mean, Google had a massive 2025 recap, 00:07:36.060 --> 00:07:39.560 60 announcements, including updates to their 00:07:39.560 --> 00:07:42.519 foundational models like Gemini 3, new niche 00:07:42.519 --> 00:07:45.060 tools like Nano Banana and their research assistant 00:07:45.060 --> 00:07:48.670 Notebook LM. It's just... Relentless. And Andrej 00:07:48.670 --> 00:07:51.350 Karpathy, who's a major voice in the field, he 00:07:51.350 --> 00:07:54.110 laid out six paradigm shifts just in his LLM 00:07:54.110 --> 00:07:56.470 year in review. Whether it's the move from pure 00:07:56.470 --> 00:07:59.449 text to multimodal AI or the focus on agents 00:07:59.449 --> 00:08:01.850 that can act on their own, the ground is constantly 00:08:01.850 --> 00:08:04.550 moving. We're even seeing a new focus on quality 00:08:04.550 --> 00:08:06.850 control. There are now sites popping up that 00:08:06.850 --> 00:08:09.269 just continuously monitor if major models are 00:08:09.269 --> 00:08:11.709 getting dumber. They run the same tasks over 00:08:11.709 --> 00:08:13.470 and over to generate these failure rate charts. 00:08:13.769 --> 00:08:15.670 Developers are struggling just to keep the existing 00:08:15.670 --> 00:08:17.689 model stable while they're constantly updating 00:08:17.689 --> 00:08:20.170 them. Before we hit the cost shocker, we should 00:08:20.170 --> 00:08:22.129 acknowledge how far the capabilities have come. 00:08:22.449 --> 00:08:24.949 Quinn just dropped Image Layered, a tool that 00:08:24.949 --> 00:08:27.790 breaks images into editable layers. It's basically 00:08:27.790 --> 00:08:30.389 bringing Photoshop -style functionality right 00:08:30.389 --> 00:08:32.309 into the generative space. That's a huge deal 00:08:32.309 --> 00:08:35.509 for creatives. And Claude in Chrome is an enormous 00:08:35.509 --> 00:08:39.110 productivity hack. It lets the LLM, that large 00:08:39.110 --> 00:08:42.090 language model, actually see, click, type, and 00:08:42.090 --> 00:08:44.690 navigate directly in your browser. So AI can 00:08:44.690 --> 00:08:47.409 now truly take on complex multi -step digital 00:08:47.409 --> 00:08:50.529 tasks without a human in the middle. But this 00:08:50.529 --> 00:08:52.269 brings us right back to that central theme of 00:08:52.269 --> 00:08:54.490 resource management, because these capabilities 00:08:54.490 --> 00:08:57.450 are not free. The sources flag the staggering 00:08:57.450 --> 00:09:00.250 resource consumption of OpenAI's new Atlas browser, 00:09:00.590 --> 00:09:03.389 which handles a lot of these agentic tasks. And 00:09:03.389 --> 00:09:05.309 here's the shocking detail, the one that really 00:09:05.309 --> 00:09:08.029 grounds the reality of scaling this tech. The 00:09:08.029 --> 00:09:11.970 Atlas browser ate 72 gigabytes of RAM with just 00:09:11.970 --> 00:09:15.399 four text documents open. Wait, 72 gigabytes 00:09:15.399 --> 00:09:17.700 for four text documents. A typical browser session 00:09:17.700 --> 00:09:20.100 might use, what, two or three gigs? Stops. That 00:09:20.100 --> 00:09:22.720 number, 72 GB, it feels like the real physical 00:09:22.720 --> 00:09:25.649 consequence of chasing AGI. It's not just code. 00:09:25.809 --> 00:09:28.389 It's a tangible multi -billion dollar heat and 00:09:28.389 --> 00:09:30.769 power problem. It speaks to the huge overhead 00:09:30.769 --> 00:09:33.669 required to run these sophisticated models. It's 00:09:33.669 --> 00:09:36.330 not just standard browser memory. That RAM is 00:09:36.330 --> 00:09:38.250 likely loaded with the model's context window, 00:09:38.509 --> 00:09:40.629 various inference streams, all the parameters 00:09:40.629 --> 00:09:43.210 it needs for agentic action. That memory has 00:09:43.210 --> 00:09:45.870 to be ready to process new data instantly. Whoa. 00:09:46.399 --> 00:09:50.480 Just imagine scaling that 72 GB RAM load to a 00:09:50.480 --> 00:09:53.720 billion daily queries globally. That is an exponential 00:09:53.720 --> 00:09:55.879 capital expenditure crisis waiting to happen. 00:09:56.039 --> 00:09:58.259 It fundamentally changes the economic calculation 00:09:58.259 --> 00:10:01.139 for every startup trying to build on these platforms. 00:10:01.480 --> 00:10:04.399 It absolutely explains the frantic, massive fundraising 00:10:04.399 --> 00:10:07.539 we're seeing. OpenAI plans to raise $100 billion 00:10:07.539 --> 00:10:11.539 at an $830 billion valuation. They need those 00:10:11.539 --> 00:10:13.940 funds specifically to cover those skyrocketing 00:10:13.940 --> 00:10:15.659 compute costs and the global infrastructure. 00:10:15.720 --> 00:10:17.679 structure build out. The hunger for resources 00:10:17.679 --> 00:10:19.860 is exponential. It's driven by the sheer size 00:10:19.860 --> 00:10:21.639 of these models. So the Atlas RAM consumption 00:10:21.639 --> 00:10:24.059 fundamentally shows us that running these massive 00:10:24.059 --> 00:10:26.539 models requires enormous compute resources and 00:10:26.539 --> 00:10:29.039 the capital required to scale is rising exponentially, 00:10:29.220 --> 00:10:31.419 threatening the business model itself. Precisely. 00:10:31.419 --> 00:10:34.120 The cost curve is almost vertical. Now, let's 00:10:34.120 --> 00:10:36.419 wrap up with the most delightfully weird and 00:10:36.419 --> 00:10:38.820 clever breakthrough of the year. It brings us 00:10:38.820 --> 00:10:41.559 back to physical efficiency, but this time in 00:10:41.559 --> 00:10:44.240 the form of a Disney character. The fully autonomous 00:10:44.240 --> 00:10:47.620 Olaf robot from Frozen. This is great because 00:10:47.620 --> 00:10:50.500 Olaf, I mean, he violates every single rule of 00:10:50.500 --> 00:10:52.700 stable robotics. He has a massive top heavy head, 00:10:52.740 --> 00:10:56.179 a tiny body, and those famously unstable snowball 00:10:56.179 --> 00:10:58.960 feet. By all accounts, this thing should be constantly 00:10:58.960 --> 00:11:01.840 falling over. But he walks, he emotes, he balances 00:11:01.840 --> 00:11:05.600 flawlessly. How did Disney manage to cheat physics 00:11:05.600 --> 00:11:08.620 on such an inherently unstable design? The real 00:11:08.620 --> 00:11:11.340 breakthrough, just like the Unitree G1, is efficiency 00:11:11.340 --> 00:11:14.379 and resource management. But applied to heat... 00:11:14.779 --> 00:11:17.220 Disney trained an AI to take real -time temperature 00:11:17.220 --> 00:11:19.539 input from the robot's motors, especially the 00:11:19.539 --> 00:11:21.960 ones in the neck and joints. This lets the AI 00:11:21.960 --> 00:11:24.779 adjust behavior on the fly. So if Olaf is doing 00:11:24.779 --> 00:11:27.379 too many enthusiastic head wobbles and the neck 00:11:27.379 --> 00:11:29.179 actuator starts heating up toward the danger 00:11:29.179 --> 00:11:32.500 zone... The AI eases off the torque. It slightly 00:11:32.500 --> 00:11:35.440 repositions the head in a less strenuous posture, 00:11:35.620 --> 00:11:38.200 and it avoids crossing that 80 degree Celsius 00:11:38.200 --> 00:11:41.809 threshold where the motor might fail. The robot 00:11:41.809 --> 00:11:44.409 is using its brain for internal thermal self 00:11:44.409 --> 00:11:46.470 -management. It's basically saying, I'm too hot, 00:11:46.549 --> 00:11:48.970 let me chill for a sec, but still look cute for 00:11:48.970 --> 00:11:51.490 the kids. That's incredible. It's such an intelligent 00:11:51.490 --> 00:11:53.990 solution to an impossible mechanical design. 00:11:54.470 --> 00:11:56.789 But the clever engineering goes even deeper than 00:11:56.789 --> 00:11:59.179 the AI, right? They had to work hard to hide 00:11:59.179 --> 00:12:02.240 the mechanics. Oh, they did. To prevent the limbs 00:12:02.240 --> 00:12:04.899 from colliding inside the foam costume, they 00:12:04.899 --> 00:12:07.539 had to use asymmetric hidden legs with inverted 00:12:07.539 --> 00:12:10.240 joint setups. The left leg works differently 00:12:10.240 --> 00:12:12.519 than the right. It's a brilliant hidden solution. 00:12:12.860 --> 00:12:15.039 And the illusion of those floating feet is maintained 00:12:15.039 --> 00:12:18.419 by the soft foam skirts that hide the real complex 00:12:18.419 --> 00:12:20.980 walking legs underneath. It perfectly maintains 00:12:20.980 --> 00:12:23.320 the look while the robot is doing serious work. 00:12:23.659 --> 00:12:26.000 And they focus so heavily on the user experience. 00:12:26.620 --> 00:12:29.059 Footstep optimization reduced the walking noise 00:12:29.059 --> 00:12:32.960 by 13 .5 decibels, making Olaf almost silent 00:12:32.960 --> 00:12:35.340 when he shuffles around. And for safety in a 00:12:35.340 --> 00:12:38.440 theme park, Olaf's arms, nose, and hair all have 00:12:38.440 --> 00:12:41.120 magnetic snap -offs. They just detach safely 00:12:41.120 --> 00:12:43.700 if a child pulls too hard. It's just a beautifully 00:12:43.700 --> 00:12:46.970 holistic piece of design. So Disney had to use 00:12:46.970 --> 00:12:49.750 this complex thermal monitoring and AI adjustment 00:12:49.750 --> 00:12:52.230 instead of just better mechanical cooling. What 00:12:52.230 --> 00:12:54.669 was the driving factor there? The unstable top 00:12:54.669 --> 00:12:57.889 -heavy design requires the AI to constantly manage 00:12:57.889 --> 00:13:00.669 heat by adjusting physical strain. It had to 00:13:00.669 --> 00:13:03.110 be a totally integrated system. This deep dive 00:13:03.110 --> 00:13:06.269 has shown us the dual revolution of 2025. On 00:13:06.269 --> 00:13:08.649 the hardware side, humanoids like the G1 are 00:13:08.649 --> 00:13:11.610 getting small, efficient, accessible. It's all 00:13:11.610 --> 00:13:13.669 driven by physics and the relentless economics 00:13:13.669 --> 00:13:15.970 of low torque. They succeed through resource 00:13:15.970 --> 00:13:18.190 optimization. And on the software and infrastructure 00:13:18.190 --> 00:13:21.490 side, AI models are getting vast and incredibly 00:13:21.490 --> 00:13:23.909 resource intensive. They're demanding exponentially 00:13:23.909 --> 00:13:27.149 rising capital and, crucially, requiring highly 00:13:27.149 --> 00:13:29.090 structured human input through these precise 00:13:29.090 --> 00:13:31.230 prompting styles. So what does this all mean 00:13:31.230 --> 00:13:33.649 for the big picture? I mean, whether we're talking 00:13:33.649 --> 00:13:37.049 about a 1 .3 meter robot intelligently managing 00:13:37.049 --> 00:13:39.789 its motor heat or a massive cloud server dealing 00:13:39.789 --> 00:13:43.190 with 72 gigs of RAM per session. Efficient resource 00:13:43.190 --> 00:13:46.330 management. be it power, torque, data structure, 00:13:46.450 --> 00:13:49.950 or thermal load, is the central unifying challenge. 00:13:50.070 --> 00:13:52.570 It's the defining battle of this era. We saw 00:13:52.570 --> 00:13:56.090 how cost forces the Unitree G1 robot to optimize 00:13:56.090 --> 00:13:58.929 its size for efficiency. So here's the thing 00:13:58.929 --> 00:14:01.590 to think about. If these massive AI models keep 00:14:01.590 --> 00:14:03.830 growing at their current rate, where does the 00:14:03.830 --> 00:14:06.090 practical limit of a truly general purpose model 00:14:06.090 --> 00:14:09.389 scale actually lie before the sheer compute cost 00:14:09.389 --> 00:14:11.990 completely breaks the business model? That is 00:14:11.990 --> 00:14:13.929 the fascinating high stakes question we leave 00:14:13.929 --> 00:14:16.269 you to mull over. So consider how these shifts 00:14:16.269 --> 00:14:18.769 will impact your own workflow. Maybe it's adopting 00:14:18.769 --> 00:14:21.289 that reporter style prompting today or thinking 00:14:21.289 --> 00:14:23.490 about where application specific hyper efficient 00:14:23.490 --> 00:14:26.049 robotics could solve a new problem in your industry 00:14:26.049 --> 00:14:28.769 tomorrow. That's the deep dive for today. Until 00:14:28.769 --> 00:14:29.330 next time.