WEBVTT 00:00:00.000 --> 00:00:03.700 Imagine this. There's a robot, Neo, and it has 00:00:03.700 --> 00:00:07.280 one simple task. Pull a tissue from a box. Okay. 00:00:07.440 --> 00:00:11.019 But before it moves, it just stops. And inside 00:00:11.019 --> 00:00:14.119 its own process, it generates eight short five 00:00:14.119 --> 00:00:16.620 -second video clips. Some of those videos show 00:00:16.620 --> 00:00:19.719 it working perfectly. In others, it fails. The 00:00:19.719 --> 00:00:22.719 tissue rips. Maybe the box falls over. It sees 00:00:22.719 --> 00:00:24.699 it all first. It's incredible. It's not just 00:00:24.699 --> 00:00:27.219 running a program. That robot is using what you 00:00:27.219 --> 00:00:30.140 could call synthetic imagination. Synthetic imagination. 00:00:30.199 --> 00:00:33.119 Yeah, it's mapping out possible futures to guide 00:00:33.119 --> 00:00:35.880 what it does in the real world, which is messy 00:00:35.880 --> 00:00:38.899 and unpredictable. It's a huge shift. Welcome 00:00:38.899 --> 00:00:42.000 back to the Deep Dive. That idea, that hook, 00:00:42.140 --> 00:00:44.380 it really captures the two big themes we're looking 00:00:44.380 --> 00:00:46.500 at today. It really does. Okay, so let's unpack 00:00:46.500 --> 00:00:48.719 this. In this Deep Dive, we're focusing on two 00:00:48.719 --> 00:00:52.060 major things. how physical AI is learning to 00:00:52.060 --> 00:00:55.740 imagine, and then the surprising realities of 00:00:55.740 --> 00:00:57.979 how AI is being adopted around the world. That's 00:00:57.979 --> 00:01:00.039 the mission. We've synthesized the key insights 00:01:00.039 --> 00:01:01.979 from the sources you shared with us. We'll start 00:01:01.979 --> 00:01:04.400 with that breakthrough from 1X Robotics. Then 00:01:04.400 --> 00:01:06.280 we're going to do a rapid -fire look at the current 00:01:06.280 --> 00:01:10.359 market. Some big deals, some friction and a few 00:01:10.359 --> 00:01:12.819 clever research tricks. And then the global report. 00:01:13.019 --> 00:01:14.939 And then, yeah, global report that really challenges 00:01:14.939 --> 00:01:17.400 how we think about where the U .S. actually stands 00:01:17.400 --> 00:01:19.480 in all of this. There's a lot of crucial ground 00:01:19.480 --> 00:01:21.719 to cover. Let's jump right in with how these 00:01:21.719 --> 00:01:24.400 robots are starting to teach themselves. So the 00:01:24.400 --> 00:01:27.280 one next world model for these neo robots is 00:01:27.280 --> 00:01:30.280 it's a really critical inflection point. This 00:01:30.280 --> 00:01:33.359 is the step that allows them to learn new physical 00:01:33.359 --> 00:01:36.140 tasks. Without a person having to code every 00:01:36.140 --> 00:01:39.219 single. movement exactly without that intense 00:01:39.219 --> 00:01:42.420 you know bespoke human coding for every single 00:01:42.420 --> 00:01:45.400 action it's a move away from the old way of doing 00:01:45.400 --> 00:01:47.819 things and the how is what's so important here 00:01:47.819 --> 00:01:49.780 it's not the traditional approach where you calculate 00:01:49.780 --> 00:01:52.379 exact joint angles and all that no it's much 00:01:52.379 --> 00:01:55.200 more visual much more predictive the system takes 00:01:55.200 --> 00:01:57.680 a simple text prompt like your pull a tissue 00:01:57.680 --> 00:02:02.219 example and it takes the robots current camera 00:02:02.219 --> 00:02:04.959 view its context it feeds both of those into 00:02:04.959 --> 00:02:08.210 this world model which basically acts like an 00:02:08.210 --> 00:02:11.330 internal simulator. And that's what generates 00:02:11.330 --> 00:02:14.229 those little five -second imagined videos of 00:02:14.229 --> 00:02:16.330 what might happen? Yep. They use a concept for 00:02:16.330 --> 00:02:18.590 this, right? Yeah. Something like video diffusion 00:02:18.590 --> 00:02:21.199 for motion planning. That's the technical term, 00:02:21.259 --> 00:02:23.159 and we should probably define that. It just means 00:02:23.159 --> 00:02:26.080 the robot plans its moves by trying to generate 00:02:26.080 --> 00:02:29.520 a successful visual outcome first. It's all based 00:02:29.520 --> 00:02:32.520 on pixels, not just lines of code. Which is so 00:02:32.520 --> 00:02:34.300 much better for dealing with new situations. 00:02:34.680 --> 00:02:37.360 Fundamentally better. Because in the old way, 00:02:37.520 --> 00:02:39.780 if the box was tilted or the light was a bit 00:02:39.780 --> 00:02:42.039 dim, the whole thing could fail if you didn't 00:02:42.039 --> 00:02:45.039 code for it. Right. With this, the model just 00:02:45.039 --> 00:02:47.259 imagines what success looks like, the successful 00:02:47.259 --> 00:02:50.090 pixels, and works backward from there. So it 00:02:50.090 --> 00:02:52.229 runs, what, eight different imagined futures? 00:02:52.449 --> 00:02:54.710 Yeah. And then it picks the best one. Correct. 00:02:55.009 --> 00:02:57.370 An internal critic looks at those eight rollouts, 00:02:57.449 --> 00:02:59.550 and it selects the one that seems most likely 00:02:59.550 --> 00:03:02.550 to succeed. Only then does a second model translate 00:03:02.550 --> 00:03:06.090 that chosen video into the actual joint commands. 00:03:06.449 --> 00:03:08.849 And the results really back this up. The sources 00:03:08.849 --> 00:03:11.949 were clear that for that tissue -pulling task, 00:03:12.310 --> 00:03:15.840 the success rate jumped from 30 % to 45%. Just 00:03:15.840 --> 00:03:18.080 by sampling. Just by sampling and choosing the 00:03:18.080 --> 00:03:20.599 best imagined future before it even moves. Yeah. 00:03:20.659 --> 00:03:22.819 That's a huge improvement in reliability. It 00:03:22.819 --> 00:03:25.280 is. And what's really exciting is that It works 00:03:25.280 --> 00:03:27.379 for tasks it's seen before, what they call in 00:03:27.379 --> 00:03:30.319 distribution, but also for totally new tasks. 00:03:30.580 --> 00:03:32.819 It can generalize. OK, but here's the reality 00:03:32.819 --> 00:03:36.479 check. Speed. Ah, yes. The latency. The sources 00:03:36.479 --> 00:03:39.439 highlight a big limitation. It takes 11 seconds 00:03:39.439 --> 00:03:41.659 for that whole imagination and planning phase. 00:03:41.840 --> 00:03:44.400 Plus one second for the action itself. So 12 00:03:44.400 --> 00:03:46.800 seconds total. Which is a long time. 11 seconds 00:03:46.800 --> 00:03:48.879 is definitely the defining challenge for this 00:03:48.879 --> 00:03:51.860 to be practical, you know, in real time. It shows 00:03:51.860 --> 00:03:54.319 the AI race has moved beyond just... virtual 00:03:54.319 --> 00:03:57.520 agents like ChatGPT. Now it's about physical 00:03:57.520 --> 00:04:00.979 agents and 1x is shipping soon. This tech is 00:04:00.979 --> 00:04:03.719 the key, but that latency has to come down. So 00:04:03.719 --> 00:04:06.219 that raises the question, how does this 11 second 00:04:06.219 --> 00:04:09.300 planning latency affect its real world usefulness 00:04:09.300 --> 00:04:11.979 right now? It's too slow for instantaneous work, 00:04:12.060 --> 00:04:14.879 but it definitively proves the concept. Using 00:04:14.879 --> 00:04:17.060 visual imagination for physical planning works. 00:04:17.279 --> 00:04:20.319 It's a proof of concept. Okay, so while 1x is 00:04:20.319 --> 00:04:23.379 figuring that out, let's pivot to the... absolute 00:04:23.379 --> 00:04:25.839 chaos of the immediate market. This is where 00:04:25.839 --> 00:04:28.339 the pace is just relentless. It's a flurry of 00:04:28.339 --> 00:04:30.860 news for sure. Let's start with a kind of wild 00:04:30.860 --> 00:04:34.860 story. The Shopify CEO went viral for using Claude 00:04:34.860 --> 00:04:37.720 to build a custom tool to analyze his own x -ray. 00:04:38.009 --> 00:04:40.810 Yeah, I saw that. The ability to use an LLM, 00:04:40.970 --> 00:04:43.370 not just for text, but to code a specific tool 00:04:43.370 --> 00:04:46.230 for yourself in an afternoon. Right. That speed 00:04:46.230 --> 00:04:49.069 is remarkable. It really blurs the line between 00:04:49.069 --> 00:04:52.310 a developer tool and just a consumer app. I mean, 00:04:52.370 --> 00:04:54.689 it still feels kind of wild that people are doing 00:04:54.689 --> 00:04:56.610 their own medical reads, even if they're technically 00:04:56.610 --> 00:04:59.329 capable. And moving over to the media side, Google 00:04:59.329 --> 00:05:02.949 just upgraded VO 3 .1, their video model. The 00:05:02.949 --> 00:05:05.589 quality enhancements are huge. We're talking. 00:05:05.980 --> 00:05:08.860 Full support for vertical, you know, 9 .16 videos, 00:05:09.139 --> 00:05:13.000 4K upscaling, and the clips just feel more dynamic, 00:05:13.139 --> 00:05:15.279 more natural, even with really short prompts. 00:05:15.500 --> 00:05:17.639 The gap between just typing something and getting 00:05:17.639 --> 00:05:19.740 a high -quality video back is shrinking almost 00:05:19.740 --> 00:05:21.199 every week. It's going to have a huge impact. 00:05:21.519 --> 00:05:23.519 And speaking of unexpected things, this was a 00:05:23.519 --> 00:05:26.480 real aha moment from the sources. Google researchers 00:05:26.480 --> 00:05:30.180 found what the source called a dumb trick. A 00:05:30.180 --> 00:05:33.040 dumb trick, as in simple but incredibly effective. 00:05:33.439 --> 00:05:37.480 Exactly. And this trick boosted accuracy by 76 00:05:37.480 --> 00:05:40.959 % on certain tasks. And it worked across the 00:05:40.959 --> 00:05:45.339 board. Gemini, GPT, Claude, DeepSeek, all of 00:05:45.339 --> 00:05:48.060 them. So what was the trick? It wasn't some complex 00:05:48.060 --> 00:05:52.180 new architecture. It was just specific phrasing 00:05:52.180 --> 00:05:54.879 in the prompt itself. To stop the model from 00:05:54.879 --> 00:05:57.250 forgetting things. Precisely. They found that 00:05:57.250 --> 00:05:59.829 phrasing it like, let's generate four detailed 00:05:59.829 --> 00:06:02.370 options first and then evaluate them one by one 00:06:02.370 --> 00:06:05.009 before you give me a final answer, made a massive 00:06:05.009 --> 00:06:06.930 difference. You're forcing it to show its work. 00:06:07.089 --> 00:06:09.410 It's like metacognition for an LLM. It is. And 00:06:09.410 --> 00:06:11.170 that just shows why prompt engineering is still 00:06:11.170 --> 00:06:13.430 so hard. You know, I still wrestle with prompt 00:06:13.430 --> 00:06:15.430 drift myself. What do you mean by that? It's 00:06:15.430 --> 00:06:17.290 when you make a tiny change to the wording and 00:06:17.290 --> 00:06:19.490 suddenly the quality of the output just plummets. 00:06:19.790 --> 00:06:22.370 So to find that one simple instruction can give 00:06:22.370 --> 00:06:25.569 you such a huge cross -model return, it's... 00:06:25.740 --> 00:06:28.300 It's humbling. Right. Now, shifting to market 00:06:28.300 --> 00:06:31.180 power, the huge deal between Apple and Google. 00:06:31.339 --> 00:06:33.720 The Gemini deal. Yeah. The reporting says it's 00:06:33.720 --> 00:06:35.899 around a billion dollars a year for Gemini to 00:06:35.899 --> 00:06:39.019 power Siri and other Apple AI features. That 00:06:39.019 --> 00:06:42.259 is just massive. We're seeing this profound centralization 00:06:42.259 --> 00:06:45.240 of capability. You have two of the biggest tech 00:06:45.240 --> 00:06:47.600 companies in the world combining their AI reach. 00:06:47.939 --> 00:06:50.560 Which, of course, drew immediate criticism. Elon 00:06:50.560 --> 00:06:52.939 Musk was very vocal, saying it represents an 00:06:52.939 --> 00:06:55.670 unreasonable concentration of power. Well, and 00:06:55.670 --> 00:06:57.589 that concentration of power is creating friction 00:06:57.589 --> 00:06:59.370 on the ground. Literally, we're seeing physical 00:06:59.370 --> 00:07:02.050 protests against AI infrastructure. Microsoft. 00:07:02.110 --> 00:07:05.230 Yeah, Microsoft unveiled these new community 00:07:05.230 --> 00:07:08.069 first data center plans. They're promising things 00:07:08.069 --> 00:07:11.209 like no local electricity bill hikes, more jobs, 00:07:11.310 --> 00:07:13.290 all of that. But the sources say that locals 00:07:13.290 --> 00:07:16.370 in 24 different states are still actively protesting 00:07:16.370 --> 00:07:19.009 the rollouts. The energy and land use are just 00:07:19.009 --> 00:07:21.670 too visible. And the promise of jobs tomorrow 00:07:21.670 --> 00:07:24.029 doesn't always outweigh the noise and environmental 00:07:24.029 --> 00:07:27.439 impact. of a huge data center today. Yeah, that's 00:07:27.439 --> 00:07:29.620 a key conflict. And finally, on hardware, while 00:07:29.620 --> 00:07:32.019 NVIDIA is king, challengers are getting serious 00:07:32.019 --> 00:07:35.339 funding. A startup called Etched just raised 00:07:35.339 --> 00:07:37.899 half a billion dollars. 500 million. To take 00:07:37.899 --> 00:07:40.600 on NVIDIA directly. They're betting on specialized 00:07:40.600 --> 00:07:43.180 chips optimized just for large language models. 00:07:43.420 --> 00:07:46.120 Which tells you the market sees the GPU bottleneck 00:07:46.120 --> 00:07:48.759 as a real vulnerability. That specialized hardware 00:07:48.759 --> 00:07:50.839 might not beat NVIDIA on everything, but for 00:07:50.839 --> 00:07:53.819 certain tasks, it could be way more efficient. 00:07:54.139 --> 00:07:57.699 So does that massive Apple and Gemini deal pretty 00:07:57.699 --> 00:08:00.100 much confirm the market is moving toward... A 00:08:00.100 --> 00:08:02.100 centralized, maybe two or three player landscape. 00:08:02.339 --> 00:08:05.079 Yes. The sheer value of that deal suggests strong 00:08:05.079 --> 00:08:07.699 centralization, despite understandable criticism 00:08:07.699 --> 00:08:10.680 about concentrated power. Mid -roll sponsor read 00:08:10.680 --> 00:08:13.560 insert here. OK, we are back. And now we're going 00:08:13.560 --> 00:08:15.680 to transition from that fast paced, high value 00:08:15.680 --> 00:08:18.759 U .S. market to the broader world stage. We're 00:08:18.759 --> 00:08:21.220 focusing on this Microsoft AI Economy Institute 00:08:21.220 --> 00:08:25.860 report. And the data on global adoption is genuinely 00:08:25.860 --> 00:08:28.399 surprising. It really is. This is where we see 00:08:28.399 --> 00:08:30.180 the difference between who's developing the models 00:08:30.180 --> 00:08:32.720 versus who is actually adopting them. The sources 00:08:32.720 --> 00:08:35.539 show that despite all the innovation here, the 00:08:35.539 --> 00:08:38.799 U .S. ranks only 24th in actual AI adoption. 00:08:39.039 --> 00:08:41.580 24th. 24th. The adoption rate here is reported 00:08:41.580 --> 00:08:45.779 at 24%. Now, compare that to, say, the UAE. They're 00:08:45.779 --> 00:08:48.820 at 64%. That is an enormous gap. It suggests 00:08:48.820 --> 00:08:51.399 that, you know, a national strategy and top -down 00:08:51.399 --> 00:08:54.480 prioritization really matter, especially in smaller 00:08:54.480 --> 00:08:56.960 economies. And globally, the average is just 00:08:56.960 --> 00:09:00.120 over 16%. But developed countries are adopting 00:09:00.120 --> 00:09:02.139 at nearly double the rate of developing ones. 00:09:02.379 --> 00:09:04.580 But the report highlights an even more critical 00:09:04.580 --> 00:09:07.980 story, a counter -narrative to the big tech dominance. 00:09:08.399 --> 00:09:11.600 Because while everyone is focused on GPT, Gemini, 00:09:11.600 --> 00:09:13.960 and Claude, The model that's really succeeding 00:09:13.960 --> 00:09:17.320 in underserved markets is DeepSeek. Right, especially 00:09:17.320 --> 00:09:20.159 across Africa and Southeast Asia. Why them? They 00:09:20.159 --> 00:09:22.500 are running the classic Android playbook. It's 00:09:22.500 --> 00:09:25.179 a strategy focused on being everywhere. Be free, 00:09:25.259 --> 00:09:28.159 be open source. Be free, be open source, be low 00:09:28.159 --> 00:09:30.799 cost, and be good enough. That's the formula 00:09:30.799 --> 00:09:32.919 for high volume adoption. And it is working. 00:09:33.539 --> 00:09:36.019 DeepSeek is seeing two to four times higher usage 00:09:36.019 --> 00:09:38.480 than the big proprietary models in some African 00:09:38.480 --> 00:09:41.149 markets. And it's not just about the model. It's 00:09:41.149 --> 00:09:43.909 an infrastructure play. They're using key partnerships, 00:09:44.169 --> 00:09:47.330 especially with Huawei, to roll out the necessary 00:09:47.330 --> 00:09:50.830 hardware and cloud support. So open source democratizes 00:09:50.830 --> 00:09:53.490 access. Exactly. The world doesn't always need 00:09:53.490 --> 00:09:56.590 the most expensive, leading -edge model if 90 00:09:56.590 --> 00:09:59.429 % of people can't afford it. It needs tools that 00:09:59.429 --> 00:10:02.509 are functional and accessible right now. And 00:10:02.509 --> 00:10:04.590 this confirms a major point from the sources. 00:10:05.049 --> 00:10:08.490 AI adoption in 2026 isn't just a tech stat anymore. 00:10:08.970 --> 00:10:11.490 It's a development stat. It is. If you ignore 00:10:11.490 --> 00:10:14.029 the next billion users in emerging economies, 00:10:14.250 --> 00:10:16.370 someone else is going to capture that market. 00:10:16.570 --> 00:10:19.629 Whoa. I mean, imagine scaling that kind of open 00:10:19.629 --> 00:10:22.090 source service to a billion queries. That's a 00:10:22.090 --> 00:10:24.149 colossal opportunity, and it's being realized 00:10:24.149 --> 00:10:26.570 through open access. It's a totally different 00:10:26.570 --> 00:10:28.490 way of thinking about the market. So this raises 00:10:28.490 --> 00:10:30.649 one last important question for this segment. 00:10:31.100 --> 00:10:34.120 What's the long term implication of these open 00:10:34.120 --> 00:10:36.559 source models leading adoption in emerging markets? 00:10:36.840 --> 00:10:39.679 Open source accessibility is closing the adoption 00:10:39.679 --> 00:10:43.220 gap, proving that AI tools must work for everyone 00:10:43.220 --> 00:10:47.600 to achieve true global reach. So to synthesize 00:10:47.600 --> 00:10:50.259 the big ideas for you, the learner, let's connect 00:10:50.259 --> 00:10:52.360 these two threads. On one hand, you have robotics, 00:10:52.559 --> 00:10:54.740 which is making this fundamental shift. Right. 00:10:55.149 --> 00:10:58.110 Physical agents moving from rigid code to what 00:10:58.110 --> 00:11:00.129 we're calling synthetic imagination. And that 00:11:00.129 --> 00:11:02.909 visual planning sets the stage for robots that 00:11:02.909 --> 00:11:05.049 can actually teach themselves. It's a move from 00:11:05.049 --> 00:11:08.929 pure reaction to to real internal deliberation, 00:11:08.990 --> 00:11:11.549 even if it's a slow 11 second deliberation right 00:11:11.549 --> 00:11:14.669 now. And at the exact same time, the global landscape 00:11:14.669 --> 00:11:16.970 is full of friction. You have protests against 00:11:16.970 --> 00:11:19.070 infrastructure. You have this intense market 00:11:19.070 --> 00:11:21.389 centralization with the Apple and Gemini deal. 00:11:21.590 --> 00:11:24.309 And also incredible opportunity. Right. And that. 00:11:24.429 --> 00:11:27.230 opportunity is being seized by accessible open 00:11:27.230 --> 00:11:29.850 source models like DeepSeek. They're the ones 00:11:29.850 --> 00:11:31.929 driving true adoption in emerging economies. 00:11:32.129 --> 00:11:34.169 It's a fascinating race, absolute performance 00:11:34.169 --> 00:11:36.669 at the very top versus sheer accessibility for 00:11:36.669 --> 00:11:39.509 everyone else. And that 1x Neo technology, that 00:11:39.509 --> 00:11:41.850 really feels like the fundamental marker of progress 00:11:41.850 --> 00:11:45.259 here. It is. The shift from an AI predicting 00:11:45.259 --> 00:11:48.279 what should happen based on code to an AI that 00:11:48.279 --> 00:11:50.879 actively imagines and then selects the best path 00:11:50.879 --> 00:11:53.980 for many possible futures, that changes everything 00:11:53.980 --> 00:11:56.639 about how a robot can learn. They're modeling 00:11:56.639 --> 00:11:58.919 reality before they touch it. It's the difference 00:11:58.919 --> 00:12:01.419 between simulating a single line and generating 00:12:01.419 --> 00:12:03.899 a whole field of possibilities. That's intelligence. 00:12:04.159 --> 00:12:06.399 That really is the next frontier. So we have 00:12:06.399 --> 00:12:07.960 to leave you with a final thought to mull over. 00:12:08.100 --> 00:12:10.860 What new autonomous capabilities will unlock 00:12:10.860 --> 00:12:14.480 when that AI imagination latency, that critical 00:12:14.480 --> 00:12:17.679 11 second planning time, drops down to just one 00:12:17.679 --> 00:12:20.379 second? When a robot's imagination becomes instantaneous, 00:12:20.860 --> 00:12:23.100 physical AI changes everything. Until the next 00:12:23.100 --> 00:12:25.759 deep dive, keep learning. Out to you, music.