WEBVTT 00:00:00.000 --> 00:00:04.599 We are at a critical juncture. AI is moving very 00:00:04.599 --> 00:00:06.759 aggressively out of the cloud and into the physical 00:00:06.759 --> 00:00:09.660 world. Right. It's in our factories, our cars, 00:00:09.679 --> 00:00:12.140 and even in your medicine cabinet. It's what 00:00:12.140 --> 00:00:15.820 NVIDIA's CEO called the chat GPT moment for physical 00:00:15.820 --> 00:00:19.140 AI. And that's not just an upgrade. This shift, 00:00:19.300 --> 00:00:21.719 it really changes everything about how we live 00:00:21.719 --> 00:00:24.519 and work. Welcome back to the Deep Dive. Today, 00:00:24.640 --> 00:00:27.320 we're unpacking a huge stack of sources, all 00:00:27.320 --> 00:00:30.539 coming out of CES 2026. And they're all focused 00:00:30.539 --> 00:00:34.079 on that one intersection, AI meeting the real 00:00:34.079 --> 00:00:37.240 physical world. Our mission today is to guide 00:00:37.240 --> 00:00:39.020 you through these massive breakthroughs. We're 00:00:39.020 --> 00:00:41.140 talking autonomous machines, the new chips that 00:00:41.140 --> 00:00:43.299 power them. And some really unexpected stuff, 00:00:43.479 --> 00:00:45.640 like how ChatGPT is being used in healthcare 00:00:45.640 --> 00:00:48.539 and the controversies that are already popping 00:00:48.539 --> 00:00:51.280 up. Okay, let's get into it. To understand this 00:00:51.280 --> 00:00:53.439 physical AI revolution, we had to start with 00:00:53.439 --> 00:00:56.679 the hardware, with the muscle and the mind. And 00:00:56.679 --> 00:00:58.899 that means NVIDIA. Yeah, Jensen Huang really 00:00:58.899 --> 00:01:01.659 set the tone. His big claim was the chat GPT 00:01:01.659 --> 00:01:04.379 moment for physical AI is here. That's a bold 00:01:04.379 --> 00:01:06.099 statement. What does he mean by that exactly? 00:01:06.319 --> 00:01:08.500 He's trying to put physical automation on the 00:01:08.500 --> 00:01:10.799 same level as, you know, the huge productivity 00:01:10.799 --> 00:01:13.620 jump we all saw with language models. So if we 00:01:13.620 --> 00:01:16.260 strip away the marketing hype, what's a simple 00:01:16.260 --> 00:01:19.159 way to define physical AI? It's machines that 00:01:19.159 --> 00:01:22.700 don't just process data. They see, they reason 00:01:22.700 --> 00:01:25.040 about what they see, and then they act in the 00:01:25.040 --> 00:01:28.480 real world. So we're moving past. This is code 00:01:28.480 --> 00:01:31.000 that's actually running a factory or driving 00:01:31.000 --> 00:01:33.980 a car down the highway. Exactly. And the star 00:01:33.980 --> 00:01:36.099 of their show was this system called Alpamayo. 00:01:36.280 --> 00:01:38.700 Right. They called it the world's first thinking, 00:01:38.859 --> 00:01:42.140 reasoning, autonomous vehicle AI. That language. 00:01:42.620 --> 00:01:45.359 Thinking and reasoning. Is that real or is it 00:01:45.359 --> 00:01:47.299 just a good tagline? I think it's a legitimate 00:01:47.299 --> 00:01:49.579 leap. It's all about something called end -to 00:01:49.579 --> 00:01:52.260 -end learning. Okay. So with old autonomous driving 00:01:52.260 --> 00:01:54.760 systems, you had all these separate steps. The 00:01:54.760 --> 00:01:57.040 car sees, then it builds a map, then it identifies 00:01:57.040 --> 00:01:59.620 objects, then it plans a path. It's like a slow 00:01:59.620 --> 00:02:02.719 step -by -step assembly line. Right. And any 00:02:02.719 --> 00:02:05.180 error in one step just messes up everything that 00:02:05.180 --> 00:02:08.620 comes after. It skips a lot of that. It takes 00:02:08.620 --> 00:02:10.580 the raw video from the cameras and translates 00:02:10.580 --> 00:02:12.719 it directly into driving actions. Oh, it's learning 00:02:12.719 --> 00:02:15.020 more like a human does. You see something, you 00:02:15.020 --> 00:02:18.060 react. Precisely. It's much faster, more fluid. 00:02:18.199 --> 00:02:20.939 It can handle new situations it's never seen 00:02:20.939 --> 00:02:23.080 before because it's not just following a rigid 00:02:23.080 --> 00:02:25.319 set of rules. And they showed this in action. 00:02:25.479 --> 00:02:27.909 The sources mentioned... A partnership with Mercedes 00:02:27.909 --> 00:02:31.169 -Benz, with a demo of the car handling tricky 00:02:31.169 --> 00:02:33.650 turns and even avoiding pedestrians. But that 00:02:33.650 --> 00:02:36.289 powerful mind needs the right hardware to run 00:02:36.289 --> 00:02:38.569 on. And that's where their new platform, Rubin, 00:02:38.590 --> 00:02:43.050 comes in. So Rubin, six new chips. What's the 00:02:43.050 --> 00:02:45.650 core idea here? What makes this a revolution 00:02:45.650 --> 00:02:49.110 and not just, you know, the next model up? Scalability. 00:02:49.250 --> 00:02:51.509 That's the whole game. The problem before was 00:02:51.509 --> 00:02:53.169 that the computers you needed for this kind of 00:02:53.169 --> 00:02:55.590 AI were too big, too hot. They used way too much 00:02:55.590 --> 00:02:59.020 power. Not practical for a million cars. Not 00:02:59.020 --> 00:03:00.960 at all. You can't put a data center in a Toyota. 00:03:01.240 --> 00:03:03.259 Rubin is designed to get all that thinking power 00:03:03.259 --> 00:03:05.520 into a much smaller, cheaper, and more efficient 00:03:05.520 --> 00:03:08.060 package. So it makes autonomous machines not 00:03:08.060 --> 00:03:10.819 just possible, but actually scalable across the 00:03:10.819 --> 00:03:13.039 whole economy. Exactly. And it's not just for 00:03:13.039 --> 00:03:15.020 cars. They made that super clear. Yeah, they 00:03:15.020 --> 00:03:17.379 brought the BD -1 droids from Star Wars out on 00:03:17.379 --> 00:03:19.460 stage. Which was a great visual, right? It's 00:03:19.460 --> 00:03:21.340 telling everyone this is for enterprise first. 00:03:21.560 --> 00:03:24.159 Think factory bots, delivery drones, construction 00:03:24.159 --> 00:03:26.599 equipment. So the business case is clear there. 00:03:26.800 --> 00:03:29.680 Much clearer. Sources are saying that while we 00:03:29.680 --> 00:03:32.939 might see some robo taxi pilots in 2027, these 00:03:32.939 --> 00:03:35.699 enterprise robots are going to be adopted way, 00:03:35.800 --> 00:03:38.240 way sooner. We have to talk about the human side 00:03:38.240 --> 00:03:40.599 of this, though. If this takes off, we're looking 00:03:40.599 --> 00:03:42.979 at millions of jobs being displaced. Driving, 00:03:43.120 --> 00:03:46.020 logistics, labor unions are definitely watching 00:03:46.020 --> 00:03:48.659 this. It's the critical, uncomfortable truth 00:03:48.659 --> 00:03:51.759 under all the cool tech. The scale of job replacement 00:03:51.759 --> 00:03:54.969 could be historic. You know, I still wrestle 00:03:54.969 --> 00:03:57.710 with prompt drift myself when asking AI for basic 00:03:57.710 --> 00:04:00.310 tasks. The idea of trusting that same kind of 00:04:00.310 --> 00:04:02.530 intelligence with an entire fleet of trucks, 00:04:02.729 --> 00:04:06.270 that's a huge leap of faith. So if the hardware 00:04:06.270 --> 00:04:08.789 is ready and the software can reason, what's 00:04:08.789 --> 00:04:11.349 the realistic timeline for seeing mass adoption 00:04:11.349 --> 00:04:14.669 of these automated enterprise systems? Enterprise 00:04:14.669 --> 00:04:17.350 robots will arrive faster than consumer robotaxis, 00:04:17.569 --> 00:04:21.319 probably before 2027. Okay, so zooming out from 00:04:21.319 --> 00:04:25.519 just NVIDIA. CES 2026 really showed this was 00:04:25.519 --> 00:04:29.519 the theme everywhere. Amazon, Google, Lego. Everyone 00:04:29.519 --> 00:04:32.079 is focused on... ai meeting the physical world 00:04:32.079 --> 00:04:34.839 right this is edge ai becoming a reality the 00:04:34.839 --> 00:04:36.860 processing is happening right there on the device 00:04:36.860 --> 00:04:39.800 not in some distant data center exactly the intelligence 00:04:39.800 --> 00:04:42.500 is local it's in the forklift it's on your laptop 00:04:42.500 --> 00:04:45.160 it's it's on a ring on your finger let's talk 00:04:45.160 --> 00:04:46.800 about some of those specific examples because 00:04:46.800 --> 00:04:49.120 they really show how broad this is oh the scope 00:04:49.120 --> 00:04:51.579 was amazing you saw robot forklifts just zipping 00:04:51.579 --> 00:04:54.259 around warehouses on their own and ai powered 00:04:54.259 --> 00:04:57.000 lego sets that you know use computer vision to 00:04:57.000 --> 00:04:58.879 give you custer instructions based on the bricks 00:04:58.879 --> 00:05:00.879 you have and it's hitting consumer devices in 00:05:00.879 --> 00:05:04.319 a big way, too. A huge way. HP refreshed their 00:05:04.319 --> 00:05:07.939 Omnibook laptops with a ton of local AI power. 00:05:08.220 --> 00:05:10.660 This means it can do things like complex video 00:05:10.660 --> 00:05:13.420 editing or live transcription without needing 00:05:13.420 --> 00:05:15.680 to be connected to the Internet. The power is 00:05:15.680 --> 00:05:17.860 moving inside the machine. The one that really 00:05:17.860 --> 00:05:19.600 stood out to me, though, was the smart ring. 00:05:20.009 --> 00:05:22.610 The AI smart ring. Yeah. It has enough local 00:05:22.610 --> 00:05:25.269 compute to listen to your work meetings and summarize 00:05:25.269 --> 00:05:28.550 them for you in real time. And always on, always 00:05:28.550 --> 00:05:31.170 listening synthesizer on your hand. You can't 00:05:31.170 --> 00:05:33.629 get more edge than that. It's the ultimate localized 00:05:33.629 --> 00:05:37.269 integration. Whoa. Just imagine scaling that 00:05:37.269 --> 00:05:39.910 single AI architecture. It's managing a factory 00:05:39.910 --> 00:05:42.490 floor, then the logistics network, and now it's 00:05:42.490 --> 00:05:45.189 in all our gadgets. The scale of that is... It's 00:05:45.189 --> 00:05:47.170 just incredible. So looking across all these 00:05:47.170 --> 00:05:49.310 new devices, which one really holds the most 00:05:49.310 --> 00:05:53.209 surprising local AI power? The AI smart ring 00:05:53.209 --> 00:05:55.730 listening to meetings suggests maximum localized 00:05:55.730 --> 00:05:57.889 integration. All right, let's shift from hardware 00:05:57.889 --> 00:06:00.550 to the models themselves. Yeah. And the regulatory 00:06:00.550 --> 00:06:02.350 challenges that are coming with them. Yeah, on 00:06:02.350 --> 00:06:04.149 the image generation side, there was a big win 00:06:04.149 --> 00:06:06.470 for Alibaba. With their new model, Quinn Image 00:06:06.470 --> 00:06:09.589 2512. Right. And it's outperforming Google's 00:06:09.589 --> 00:06:12.670 Nano Banana Pro in quality tests. What's the 00:06:12.670 --> 00:06:15.220 big deal there? Beyond just making nicer pictures, 00:06:15.420 --> 00:06:18.339 what problem did it actually solve? Two big ones. 00:06:18.699 --> 00:06:21.519 First, it gets rid of that weird, uncanny AI 00:06:21.519 --> 00:06:24.459 look that a lot of generated images have. But 00:06:24.459 --> 00:06:26.959 second, and this is huge for anyone using it 00:06:26.959 --> 00:06:30.180 commercially, it can finally handle text perfectly. 00:06:30.500 --> 00:06:32.620 It's been a major hurdle. Why was it so hard 00:06:32.620 --> 00:06:35.930 for AIs to just... spell words correctly inside 00:06:35.930 --> 00:06:38.769 an image. It's about precision. The models understood 00:06:38.769 --> 00:06:41.610 the whole scene, the vibe, but they struggled 00:06:41.610 --> 00:06:44.930 to arrange pixels into stable, perfect letters. 00:06:45.370 --> 00:06:47.610 Quinn seems to have finally cracked that problem. 00:06:47.790 --> 00:06:49.949 So that's a clear technical win. Yeah. But then 00:06:49.949 --> 00:06:51.839 you have the other side of the coin. Grok from 00:06:51.839 --> 00:06:55.060 XAI, which caused a huge regulatory firestorm. 00:06:55.220 --> 00:06:57.180 Oh, yeah. That was the big controversy of the 00:06:57.180 --> 00:06:59.139 week. Its edit image feature was immediately 00:06:59.139 --> 00:07:01.560 used by people to create some really disturbing, 00:07:01.800 --> 00:07:04.279 awful content. The sources specifically mentioned 00:07:04.279 --> 00:07:06.459 undressed kids. It was appalling. Absolutely. 00:07:06.620 --> 00:07:09.639 And the EU regulators were furious. XAI had to 00:07:09.639 --> 00:07:11.439 come out and admit they messed up, that their 00:07:11.439 --> 00:07:13.779 safety filters just completely failed. It's a 00:07:13.779 --> 00:07:16.139 perfect example of how capability just races 00:07:16.139 --> 00:07:18.199 ahead of safety. But what's so fascinating is 00:07:18.199 --> 00:07:21.430 the contrast. While this public controversy is 00:07:21.430 --> 00:07:24.129 raging, what's happening behind the scenes? High 00:07:24.129 --> 00:07:26.470 -level government adoption. The sources say the 00:07:26.470 --> 00:07:28.350 U .S. War Department is planning to integrate 00:07:28.350 --> 00:07:31.569 Grok into their Gene AI mill system. For about 00:07:31.569 --> 00:07:35.930 3 million people. Rolling out early 2026. This 00:07:35.930 --> 00:07:39.389 is a massive defense deal for XAI happening right 00:07:39.389 --> 00:07:41.310 in the middle of a public safety crisis for the 00:07:41.310 --> 00:07:43.410 exact same model. It's an incredible tension. 00:07:43.709 --> 00:07:46.290 It really is. So does this rapid military adoption 00:07:46.290 --> 00:07:49.430 suggest that regulation is just lagging way behind 00:07:49.430 --> 00:07:52.470 capability? Yes. The pace of defense adoption 00:07:52.470 --> 00:07:55.569 significantly outruns any regulatory backlash. 00:07:55.930 --> 00:07:58.870 Let's shift gears completely now. Away from cars 00:07:58.870 --> 00:08:01.550 and defense to something very human. Health. 00:08:01.730 --> 00:08:04.449 The numbers coming out from OpenAI are staggering. 00:08:04.939 --> 00:08:07.279 Truly staggering. Over 40 million people are 00:08:07.279 --> 00:08:10.220 using ChatGPT every single day just to navigate 00:08:10.220 --> 00:08:13.339 health care. 40 million daily. Which means by 00:08:13.339 --> 00:08:15.480 default, it's already the biggest health assistant 00:08:15.480 --> 00:08:17.980 on the planet. And they never even set out to 00:08:17.980 --> 00:08:20.019 build one. What are people asking it? Is it, 00:08:20.040 --> 00:08:22.980 you know? I have a cough. What is it? It's actually 00:08:22.980 --> 00:08:24.980 more administrative, which I think is the most 00:08:24.980 --> 00:08:27.279 interesting part. They're getting between 1 .6 00:08:27.279 --> 00:08:31.339 and 1 .9 million insurance questions every single 00:08:31.339 --> 00:08:34.059 week. Every week? Yeah. People are asking it 00:08:34.059 --> 00:08:36.360 to compare insurance plans, help them appeal 00:08:36.360 --> 00:08:39.240 denied claims, figure out confusing medical bills. 00:08:39.480 --> 00:08:41.700 So it's become a bureaucratic assistant for the 00:08:41.700 --> 00:08:44.029 health care system. Exactly. And the timing is 00:08:44.029 --> 00:08:46.870 really telling. 70 % of these chats are happening 00:08:46.870 --> 00:08:49.509 outside of normal clinic hours. It's become the 00:08:49.509 --> 00:08:51.909 midnight second opinion for people. And the sources 00:08:51.909 --> 00:08:54.529 pointed out how critical this is for rural users. 00:08:55.169 --> 00:08:58.889 600 ,000 messages a week from areas with low 00:08:58.889 --> 00:09:01.950 doctor access. The AI is filling a massive gap 00:09:01.950 --> 00:09:04.070 in the system. And with the end of some Affordable 00:09:04.070 --> 00:09:07.710 Care Act subsidies, you might see even more uninsured 00:09:07.710 --> 00:09:09.710 people turning to this as their first option. 00:09:09.850 --> 00:09:12.769 It's driven by necessity. It's not all bad news, 00:09:12.850 --> 00:09:15.830 though. The data did show a huge positive. The 00:09:15.830 --> 00:09:18.750 AI is actually catching errors, errors in billing 00:09:18.750 --> 00:09:21.350 and claims that even trained humans sometimes 00:09:21.350 --> 00:09:24.049 miss. That's the incredible potential. But you 00:09:24.049 --> 00:09:26.289 have to pivot immediately to the risk. Of course. 00:09:26.610 --> 00:09:28.929 Lawsuits have already started over bad advice. 00:09:29.269 --> 00:09:32.190 So OpenAI is now managing a de facto medical 00:09:32.190 --> 00:09:34.889 platform, whether they like it or not. The users 00:09:34.889 --> 00:09:37.490 decided what it was for. Which means the warning 00:09:37.490 --> 00:09:40.100 has to be said loud and clear. Absolutely. You 00:09:40.100 --> 00:09:42.360 have to double -check any high -stakes health 00:09:42.360 --> 00:09:45.059 advice you get from an AI. The risk is just too 00:09:45.059 --> 00:09:47.740 high to blindly trust it. With that incredible 00:09:47.740 --> 00:09:51.360 volume, 40 million users a day, is it being used 00:09:51.360 --> 00:09:54.159 more for administrative or for diagnostic advice? 00:09:54.539 --> 00:09:57.120 The use skews heavily administrative, answering 00:09:57.120 --> 00:09:59.419 millions of insurance and billing questions weekly. 00:09:59.820 --> 00:10:02.360 Okay, so let's tie this all together. We have 00:10:02.360 --> 00:10:05.240 NVIDIA's new hardware. AI getting embedded in 00:10:05.240 --> 00:10:08.259 everything, these regulatory fights, and ChatGPT 00:10:08.259 --> 00:10:10.980 accidentally becoming a health navigator. What's 00:10:10.980 --> 00:10:13.299 the big idea? The biggest takeaway for me is 00:10:13.299 --> 00:10:16.179 that AI is fundamentally shifting. It's not just 00:10:16.179 --> 00:10:18.519 a digital tool anymore. It's becoming a core 00:10:18.519 --> 00:10:20.700 part of our physical and societal infrastructure. 00:10:21.100 --> 00:10:23.899 Right. NVIDIA gave us the chips, the Rubin platform, 00:10:24.059 --> 00:10:26.799 and the autonomous mind, Alpamayo, to make that 00:10:26.799 --> 00:10:29.639 physical AI scalable. And we're seeing that scale 00:10:29.639 --> 00:10:32.460 everywhere, from factory floors to that smart 00:10:32.460 --> 00:10:35.340 ring on your finger. Total integration is here. 00:10:35.419 --> 00:10:37.559 And at the same time, in the digital world, these 00:10:37.559 --> 00:10:41.159 general models like ChatGPT are just taking on 00:10:41.159 --> 00:10:43.879 these huge high -stakes roles that no one planned 00:10:43.879 --> 00:10:46.379 for. We're seeing atoms and bits just rapidly 00:10:46.379 --> 00:10:49.100 converging. And the impact of that, the job losses, 00:10:49.159 --> 00:10:51.820 the ethical problems, but also the life -saving 00:10:51.820 --> 00:10:54.299 potential, it's all accelerating way faster than 00:10:54.299 --> 00:10:56.620 our ability to even understand it, let alone 00:10:56.620 --> 00:10:58.820 regulate it. The world is changing under our 00:10:58.820 --> 00:11:01.159 feet faster than policy can keep up. We really 00:11:01.159 --> 00:11:02.940 appreciate you taking this deep dive with us. 00:11:03.340 --> 00:11:04.980 Understanding this gives you a real shortcut 00:11:04.980 --> 00:11:07.259 to seeing where technology is heading next. So 00:11:07.259 --> 00:11:09.700 a final thought for you to chew on. Given that 00:11:09.700 --> 00:11:13.139 5 % of all chat GPT chats globally are about 00:11:13.139 --> 00:11:16.120 health, high stakes field it took over by accident, 00:11:16.360 --> 00:11:20.080 what's the next domain? Is it law? Finance. Education? 00:11:20.519 --> 00:11:23.399 What will a general AI accidentally become the 00:11:23.399 --> 00:11:25.340 dominant player in next just because of sheer 00:11:25.340 --> 00:11:27.080 user need? Something to think about.