WEBVTT 00:00:00.000 --> 00:00:03.060 The scale of modern A .I. is it's almost impossible 00:00:03.060 --> 00:00:05.740 to really grasp. I mean, right now, the cost 00:00:05.740 --> 00:00:07.580 of building the current A .I. infrastructure 00:00:07.580 --> 00:00:10.759 is already more than the Manhattan Project and 00:00:10.759 --> 00:00:13.640 the Apollo space program combined. That is a 00:00:13.640 --> 00:00:16.739 staggering comparison. It truly is. We're talking 00:00:16.739 --> 00:00:19.019 about physical structures that need the same 00:00:19.019 --> 00:00:23.179 amount of land as, what, 450 soccer fields? And 00:00:23.179 --> 00:00:25.320 they require enough electricity to light up a 00:00:25.320 --> 00:00:27.699 million homes. This isn't just, you know, scaling 00:00:27.699 --> 00:00:30.079 up tech. This is a whole new standard for global 00:00:30.079 --> 00:00:32.869 infrastructure. Welcome to the Deep Dive. We've 00:00:32.869 --> 00:00:34.549 gone through the latest research to bring you 00:00:34.549 --> 00:00:37.250 the most crucial insights into this, this massive 00:00:37.250 --> 00:00:39.490 shift. Our mission today is pretty clear. We're 00:00:39.490 --> 00:00:42.509 exploring two huge frontiers in AI that are happening 00:00:42.509 --> 00:00:44.670 at the same time. First, there's the conceptual 00:00:44.670 --> 00:00:47.509 one, why LLMs have sort of hit a wall, and what 00:00:47.509 --> 00:00:49.570 this idea of spatial intelligence actually means 00:00:49.570 --> 00:00:52.090 for what's next. And second, we're going to dive 00:00:52.090 --> 00:00:55.429 into the physical reality of it all, the just 00:00:55.429 --> 00:00:58.890 eye -watering cost and the unbelievable scale 00:00:58.890 --> 00:01:01.969 of these new ultra -mega data centers that have 00:01:01.969 --> 00:01:05.209 to run these models. That includes the, frankly, 00:01:05.310 --> 00:01:09.590 astonishing $32 billion Stargate project. So 00:01:09.590 --> 00:01:11.349 let's unpack this. We have to start with Fei 00:01:11.349 --> 00:01:13.359 -Fei Li. For anyone who doesn't know, she's the 00:01:13.359 --> 00:01:16.040 Stanford professor behind ImageNet, which is, 00:01:16.060 --> 00:01:17.700 I mean, it's essentially the foundation for the 00:01:17.700 --> 00:01:19.379 whole deep learning revolution we're living through. 00:01:19.519 --> 00:01:21.379 Right. And she just put out what is basically 00:01:21.379 --> 00:01:23.939 a manifesto. It's a really strong declaration 00:01:23.939 --> 00:01:27.260 that LLMs, large language models, have pretty 00:01:27.260 --> 00:01:29.200 much reached their ceiling. She argues that the 00:01:29.200 --> 00:01:31.280 future of AI isn't just about better language. 00:01:31.459 --> 00:01:33.980 It requires something she calls spatial intelligence. 00:01:34.340 --> 00:01:36.620 So what does that actually mean for an AI, I 00:01:36.620 --> 00:01:38.700 mean? What is spatial intelligence when you take 00:01:38.700 --> 00:01:41.170 it out of the human context? It's the kind of 00:01:41.170 --> 00:01:43.290 thing we do every single day without even thinking 00:01:43.290 --> 00:01:45.269 about it. Yeah. It's the ability to understand 00:01:45.269 --> 00:01:48.450 and navigate and interact with three -dimensional 00:01:48.450 --> 00:01:51.189 space. Yeah. Like think about catching a set 00:01:51.189 --> 00:01:54.010 of keys, someone tosses you in a dark room. Or 00:01:54.010 --> 00:01:56.810 a firefighter who has to instantly read a chaotic 00:01:56.810 --> 00:01:59.510 smoke -filled space to find the safest way out. 00:01:59.819 --> 00:02:02.120 It's about applying the basic rules of physics 00:02:02.120 --> 00:02:03.959 and prediction to the world you can actually 00:02:03.959 --> 00:02:06.739 touch. Exactly. Yeah. Light's big critique is 00:02:06.739 --> 00:02:09.419 that LLMs, for all their amazing chat abilities, 00:02:09.680 --> 00:02:12.960 are completely blind to reality. They don't know 00:02:12.960 --> 00:02:15.919 that if you drop something, it falls down. Always. 00:02:16.060 --> 00:02:18.639 They have zero concept of gravity or mass unless 00:02:18.639 --> 00:02:20.960 we spell it out for them in text. So to fix that, 00:02:21.020 --> 00:02:22.860 she's proposing that AI needs what she calls 00:02:22.860 --> 00:02:25.539 world models. And she says these models have 00:02:25.539 --> 00:02:27.780 to have three core capabilities to bridge that 00:02:27.780 --> 00:02:30.240 gap between just language and actual physics. 00:02:30.500 --> 00:02:32.219 Okay, so the first one is that they have to be 00:02:32.219 --> 00:02:34.199 generative. The models have to be able to create 00:02:34.199 --> 00:02:37.199 these really complex 3D environments that, and 00:02:37.199 --> 00:02:39.919 this is key, strictly obey the rules of real 00:02:39.919 --> 00:02:43.159 world physics. No weird floating teacups. Second, 00:02:43.360 --> 00:02:46.000 they have to be completely multimodal. So they 00:02:46.000 --> 00:02:48.979 need to process everything at once. Text. images, 00:02:49.159 --> 00:02:52.379 video, depth maps, even data from real world 00:02:52.379 --> 00:02:54.900 sensors. It's a full sensory understanding. Right. 00:02:55.060 --> 00:02:57.699 And finally, they have to be interactive. They 00:02:57.699 --> 00:03:00.659 need to predict with really high accuracy what's 00:03:00.659 --> 00:03:02.620 going to happen when a user does something inside 00:03:02.620 --> 00:03:05.400 that simulation. That's how you get real cause 00:03:05.400 --> 00:03:07.479 and effect learning. And this isn't just a theory, 00:03:07.580 --> 00:03:10.379 right? Her team at World Labs is already shipping 00:03:10.379 --> 00:03:12.979 a tool called Marble, which takes a simple text 00:03:12.979 --> 00:03:15.419 prompt and turns it into a 3D scene you can actually 00:03:15.419 --> 00:03:17.580 walk around in. Yeah, and what's so fascinating 00:03:17.580 --> 00:03:19.699 is how they're doing it. They're probably using 00:03:19.699 --> 00:03:22.699 things like neural lamberis, neural radiance 00:03:22.699 --> 00:03:24.400 fields, which are these models that can build 00:03:24.400 --> 00:03:27.479 a 3D scene from 2D images. It's moving from AI 00:03:27.479 --> 00:03:30.449 describing the world. to AI building it. I still 00:03:30.449 --> 00:03:32.530 kind of wrestle with trying to figure out how 00:03:32.530 --> 00:03:35.449 to bridge what I type into an LLM and what the 00:03:35.449 --> 00:03:37.569 real world does. It just seems like such a massive 00:03:37.569 --> 00:03:39.610 leap. That's a totally fair struggle. It really 00:03:39.610 --> 00:03:42.129 gets to the heart of the challenge. But the need 00:03:42.129 --> 00:03:44.669 for agents that can operate in the real world, 00:03:44.810 --> 00:03:47.810 you know, robots, self -driving cars, it means 00:03:47.810 --> 00:03:50.370 the industry is being forced to push for these 00:03:50.370 --> 00:03:52.949 hybrid spatial models right now. Since language 00:03:52.949 --> 00:03:56.930 is so, so foundational to how we think. How quickly 00:03:56.930 --> 00:03:59.889 is this shift from pure language to these spatial 00:03:59.889 --> 00:04:03.189 models actually going to impact the AI tools 00:04:03.189 --> 00:04:05.409 that we're all using every day? The need for 00:04:05.409 --> 00:04:07.870 real -world interaction is already pushing the 00:04:07.870 --> 00:04:10.819 immediate deployment of these new... hybrid models. 00:04:11.020 --> 00:04:13.379 So moving from the theoretical, let's look at 00:04:13.379 --> 00:04:15.259 what's happening on the ground right now. Even 00:04:15.259 --> 00:04:17.180 as we're looking towards spatial AI, we're still 00:04:17.180 --> 00:04:19.240 grappling with some really basic challenges and 00:04:19.240 --> 00:04:22.240 seeing some fascinating new applications. Yeah, 00:04:22.279 --> 00:04:24.079 on the challenge side, there's still that fundamental 00:04:24.079 --> 00:04:27.160 problem of why AI struggles so much to tell the 00:04:27.160 --> 00:04:28.959 difference between a fact and a subjective belief. 00:04:29.199 --> 00:04:31.740 One report called it the missing piece, which 00:04:31.740 --> 00:04:34.360 is really about causal modeling. The AI often 00:04:34.360 --> 00:04:36.899 gets the what, but it totally misses the why 00:04:36.899 --> 00:04:39.680 behind the data. And at the same time, agents 00:04:39.680 --> 00:04:42.680 are getting so much smarter. I love this example 00:04:42.680 --> 00:04:45.639 of the Minecraft AI agent named Steve. Oh, it's 00:04:45.639 --> 00:04:47.839 brilliant. You can give Steve just one high -level 00:04:47.839 --> 00:04:50.680 command. Something like, mine some iron or build 00:04:50.680 --> 00:04:52.920 me a castle. And Steve doesn't just do it. It 00:04:52.920 --> 00:04:55.079 actually spawns a bunch of other little agents 00:04:55.079 --> 00:04:58.040 that coordinate and work together like a real 00:04:58.040 --> 00:05:00.720 team to get it done. OK, let's talk about friction, 00:05:00.879 --> 00:05:03.420 because that is definitely heating up, especially 00:05:03.420 --> 00:05:06.019 around data. We saw a huge example of this with 00:05:06.019 --> 00:05:09.000 Wikipedia recently. Right. Wikipedia, which is, 00:05:09.019 --> 00:05:12.139 let's be honest, the bedrock for so much AI training 00:05:12.139 --> 00:05:15.220 data, is basically telling AI companies to please 00:05:15.220 --> 00:05:18.519 stop scraping its entire site. They want everyone 00:05:18.519 --> 00:05:21.600 to use their paid API instead. Why is that such 00:05:21.600 --> 00:05:23.779 a big deal? Well, it's about fairness. Yeah. 00:05:23.800 --> 00:05:26.439 You know. Giving credit for all that human labor 00:05:26.439 --> 00:05:28.639 and making sure the data quality stays high. 00:05:28.819 --> 00:05:31.579 They even made a pointed reference to Grokopedia, 00:05:31.860 --> 00:05:34.759 which was a pretty clear jab at Grok and XAI 00:05:34.759 --> 00:05:37.899 for allegedly relying so heavily on scraped content. 00:05:38.199 --> 00:05:40.240 So it sounds like quality training data, the 00:05:40.240 --> 00:05:42.199 fuel for all of this, is going to be heavily 00:05:42.199 --> 00:05:44.740 monetized from now on. And that ties directly 00:05:44.740 --> 00:05:48.040 into the money, into financing. Crusoe, which 00:05:48.040 --> 00:05:50.899 is an AI energy and infrastructure company, just 00:05:50.899 --> 00:05:54.790 secured a huge investment. 1 .3. eight billion 00:05:54.790 --> 00:05:57.250 dollars it puts their valuation at 10 billion 00:05:57.250 --> 00:05:59.870 that is a massive vote of confidence what makes 00:05:59.870 --> 00:06:02.689 them so special crusoe's all about tackling the 00:06:02.689 --> 00:06:05.449 energy problem They capture wasted energy, like 00:06:05.449 --> 00:06:07.529 natural gas that would just be flared off at 00:06:07.529 --> 00:06:09.689 a site, and they use it to power computing centers 00:06:09.689 --> 00:06:12.170 right there. The fact that NVIDIA is a major 00:06:12.170 --> 00:06:14.649 investor just shows how critical that link between 00:06:14.649 --> 00:06:17.069 energy and compute has become. We're also seeing 00:06:17.069 --> 00:06:20.370 friction in creative ethics. The showrunner for 00:06:20.370 --> 00:06:23.209 Amazon's House of David called using over 350 00:06:23.209 --> 00:06:27.689 AI -generated shots magical filmmaking. But a 00:06:27.689 --> 00:06:29.509 lot of critics just immediately called it cheap, 00:06:29.629 --> 00:06:32.490 a way to replace human artists. And that tension 00:06:32.490 --> 00:06:34.420 is not gone. going away. It's a real philosophical 00:06:34.420 --> 00:06:37.480 split between efficiency and human artistry. 00:06:37.519 --> 00:06:39.519 Let's end this section on a positive note though. 00:06:39.939 --> 00:06:42.420 Privacy. Google launched something called Private 00:06:42.420 --> 00:06:45.160 AI Compute. It lets you use the full cloud power 00:06:45.160 --> 00:06:47.779 of Gemini, but it ensures that no one, not even 00:06:47.779 --> 00:06:50.120 Google, can see the data you're processing. Yeah, 00:06:50.160 --> 00:06:52.920 they do it using secure enclaves, which are like 00:06:52.920 --> 00:06:55.360 these hardware -level black boxes. It allows 00:06:55.360 --> 00:06:57.980 for really sensitive data to be computed privately, 00:06:58.240 --> 00:07:00.319 which is a huge deal for a lot of applications. 00:07:00.779 --> 00:07:03.759 So what does Wikipedia demanding payment mean 00:07:03.759 --> 00:07:06.259 for the future long -term availability of truly 00:07:06.259 --> 00:07:08.839 open training data? It suggests quality training 00:07:08.839 --> 00:07:11.779 data, the fuel of AI, will be heavily monetized 00:07:11.779 --> 00:07:14.300 moving forward. Okay, let's shift now to the 00:07:14.300 --> 00:07:16.810 physical frontier. the sheer infrastructure you 00:07:16.810 --> 00:07:20.029 need for all this. An Epoch AI report said that 00:07:20.029 --> 00:07:22.569 spending on these specialized AI data centers 00:07:22.569 --> 00:07:26.529 is on track to pass $300 billion by the end of 00:07:26.529 --> 00:07:29.170 2025. You have to put that number in perspective. 00:07:29.470 --> 00:07:32.589 $300 billion is almost 1 % of the entire U .S. 00:07:32.589 --> 00:07:35.470 GDP. It's more than the Apollo program and the 00:07:35.470 --> 00:07:39.029 Manhattan Project combined. This isn't just an 00:07:39.029 --> 00:07:41.589 investment. It's a state -level commitment to 00:07:41.589 --> 00:07:43.879 a single piece of technology. And the headline 00:07:43.879 --> 00:07:45.819 example, the one everyone's talking about, is 00:07:45.819 --> 00:07:49.000 OpenAI's proposed Stargate Abilene project. The 00:07:49.000 --> 00:07:51.500 numbers are just. They're hard to believe. We're 00:07:51.500 --> 00:07:55.000 talking a $32 billion price tag. It needs those 00:07:55.000 --> 00:07:58.660 450 soccer fields of land. And the critical part, 00:07:58.800 --> 00:08:02.139 it will draw one gigawatt of power. That is enough 00:08:02.139 --> 00:08:04.600 electricity for about a million homes, all for 00:08:04.600 --> 00:08:07.480 one site. And it will have 250 times the compute 00:08:07.480 --> 00:08:10.620 capacity of GPT -4. But here's the thing that 00:08:10.620 --> 00:08:12.800 really changes the paradigm, the engineering 00:08:12.800 --> 00:08:15.459 implication of this. It's the paradox of latency. 00:08:15.699 --> 00:08:17.360 It used to be that you had to build data centers 00:08:17.360 --> 00:08:20.839 near users to be fast. Now, latency just doesn't 00:08:20.839 --> 00:08:22.899 matter as much. The reason is that the time it 00:08:22.899 --> 00:08:24.519 takes the model to actually think and generate 00:08:24.519 --> 00:08:27.459 an answer, the inference time, is about 100 times 00:08:27.459 --> 00:08:29.480 longer than it takes to send data all the way 00:08:29.480 --> 00:08:31.670 around the globe. Whoa. So imagine having so 00:08:31.670 --> 00:08:33.149 much computing power that you could literally 00:08:33.149 --> 00:08:35.190 bounce data off the moon and you would still 00:08:35.190 --> 00:08:37.929 be bottlenecked by the model itself, not by how 00:08:37.929 --> 00:08:40.129 fast you could send the signal. That changes 00:08:40.129 --> 00:08:42.549 everything. It really does. The era of needing 00:08:42.549 --> 00:08:44.950 to be geographically close for speed is just 00:08:44.950 --> 00:08:47.230 over. Yeah. The entire physical constraint has 00:08:47.230 --> 00:08:49.690 shifted. Now, data centers get built wherever 00:08:49.690 --> 00:08:52.610 power is cheapest and most available, not where 00:08:52.610 --> 00:08:55.139 the users are. Which brings us right back to 00:08:55.139 --> 00:08:58.080 that power challenge. Only a few countries can 00:08:58.080 --> 00:09:00.519 realistically handle multiple sites that each 00:09:00.519 --> 00:09:03.639 need over a gigawatt of power constantly. The 00:09:03.639 --> 00:09:05.860 solutions seem to be starting with natural gas, 00:09:06.100 --> 00:09:09.100 then layering on solar and wind through big grid 00:09:09.100 --> 00:09:11.379 interconnects. The race isn't for faster chips 00:09:11.379 --> 00:09:14.320 anymore. It's a race for massive, reliable energy 00:09:14.320 --> 00:09:16.860 sources. So if the competition has completely 00:09:16.860 --> 00:09:20.539 shifted from speed to just sheer scale and access 00:09:20.539 --> 00:09:23.190 to power, does that fundamentally change? who's 00:09:23.190 --> 00:09:26.370 going to lead the next phase of AI. Absolutely. 00:09:26.570 --> 00:09:29.210 The future is being shaped by the very few entities 00:09:29.210 --> 00:09:32.029 that can build, finance, and secure these colossal 00:09:32.029 --> 00:09:34.330 power sources faster than anyone else on the 00:09:34.330 --> 00:09:36.649 planet. So if we synthesize everything from our 00:09:36.649 --> 00:09:38.509 deep dive today, we really covered three major 00:09:38.509 --> 00:09:41.450 themes. First, we're redefining the core objective 00:09:41.450 --> 00:09:44.029 of AI. We're moving away from just pure language 00:09:44.029 --> 00:09:46.370 and towards these complex spatial world models. 00:09:46.549 --> 00:09:49.250 Second, at the exact same time, we're battling 00:09:49.250 --> 00:09:51.789 this intense data and ethical friction in the 00:09:51.789 --> 00:09:53.490 applications we have now. You know, everything 00:09:53.490 --> 00:09:56.490 from creative tension in Hollywood to who owns 00:09:56.490 --> 00:09:59.149 and gets to use Wikipedia's data. And finally, 00:09:59.250 --> 00:10:02.269 all of this innovation, both the big ideas and 00:10:02.269 --> 00:10:04.470 the stuff happening today, is being built on 00:10:04.470 --> 00:10:07.769 an infrastructure of just completely unprecedented 00:10:07.769 --> 00:10:11.710 scale. It's far bigger than any historical megaproject 00:10:11.710 --> 00:10:14.610 in its cost and its power demand. So you now 00:10:14.610 --> 00:10:16.490 have a pretty clear structure for understanding 00:10:16.490 --> 00:10:19.029 how AI is going to evolve over the next decade. 00:10:19.190 --> 00:10:21.409 You can see the conceptual limits, the immediate 00:10:21.409 --> 00:10:24.289 frictions, and the staggering physical cost of 00:10:24.289 --> 00:10:26.090 making it all happen. And here's something to 00:10:26.090 --> 00:10:28.210 think about. If one gigawatt of power becomes 00:10:28.210 --> 00:10:31.309 the minimum price to play in advanced AI, what 00:10:31.309 --> 00:10:33.409 happens to the innovative coder in a garage? 00:10:33.970 --> 00:10:36.570 What happens to that decentralized future we 00:10:36.570 --> 00:10:39.759 all used to imagine? Something to mull on. Thank 00:10:39.759 --> 00:10:41.659 you for joining us for this deep dive into the 00:10:41.659 --> 00:10:43.960 dual frontiers of artificial intelligence. We 00:10:43.960 --> 00:10:45.779 really encourage you to keep exploring these 00:10:45.779 --> 00:10:46.100 topics.