WEBVTT 00:00:00.000 --> 00:00:03.160 Imagine building the brain for a powerful AI 00:00:03.160 --> 00:00:08.000 model. Not over months, but maybe in just a couple 00:00:08.000 --> 00:00:09.919 of hours. Right. And doing it with something 00:00:09.919 --> 00:00:12.519 like, what, 200 ,000 graphics cards? That sounds 00:00:12.519 --> 00:00:14.519 like sci -fi, but it's actually happening now. 00:00:14.939 --> 00:00:17.300 Welcome to the Deep Dive. Today, we're going 00:00:17.300 --> 00:00:19.539 to unpack a really fascinating set of sources. 00:00:19.739 --> 00:00:22.879 They focus on this rapidly accelerating world 00:00:22.879 --> 00:00:26.539 of AI, specifically the intense battle that's 00:00:26.539 --> 00:00:28.859 going on, sometimes behind the scenes, for computing 00:00:28.859 --> 00:00:31.160 power. Yeah, and our mission today is to explore 00:00:31.160 --> 00:00:33.859 what some people are calling the GPU arms race. 00:00:34.039 --> 00:00:35.759 We want to understand what it means for big tech, 00:00:35.820 --> 00:00:38.179 for everyone else. And then we'll dive into some 00:00:38.179 --> 00:00:40.439 of the latest kind of surprising AI developments 00:00:40.439 --> 00:00:42.240 that are popping up and shaping things. We'll 00:00:42.240 --> 00:00:43.679 try to highlight what's really important here, 00:00:43.719 --> 00:00:45.780 maybe how these shifts could impact you as you, 00:00:45.799 --> 00:00:47.939 you know. navigate this incredibly fast -moving 00:00:47.939 --> 00:00:51.140 landscape. So let's dig in. Okay, let's unpack 00:00:51.140 --> 00:00:54.020 this first part. The core idea seems pretty straightforward, 00:00:54.179 --> 00:00:56.539 actually. Training the biggest, most advanced 00:00:56.539 --> 00:01:00.119 AI models. It now depends almost entirely on 00:01:00.119 --> 00:01:02.700 who controls the most computing power. Simple 00:01:02.700 --> 00:01:04.920 as that. It really is a new kind of arms race, 00:01:05.019 --> 00:01:07.200 isn't it? But instead of weapons, it's silicon 00:01:07.200 --> 00:01:10.040 and, well, massive amounts of energy. Exactly. 00:01:10.079 --> 00:01:12.379 And the biggest player kind of stepping out into 00:01:12.379 --> 00:01:16.519 the light right now, that seems to be XAI's Colossus 00:01:16.519 --> 00:01:19.480 supercomputer. It's being called the Undisputed 00:01:19.480 --> 00:01:23.200 Heavyweight, packing, get this, 200 ,000 H100 00:01:23.200 --> 00:01:27.200 equivalent GPUs, graphics processing units. 100 00:01:27.200 --> 00:01:29.340 ,000. It's hard to even picture that. It's truly 00:01:29.340 --> 00:01:31.640 massive. And what's really fascinating, maybe 00:01:31.640 --> 00:01:33.980 even mind -bending, is this machine can apparently 00:01:33.980 --> 00:01:37.620 train a model like GPT -3 in under two hours. 00:01:37.719 --> 00:01:40.439 Under two hours, not weeks. Nope, not weeks or 00:01:40.439 --> 00:01:44.799 days, hours. Beat, whoa. Yeah. I mean, just imagine 00:01:44.799 --> 00:01:47.459 scaling that kind of training capacity. Think 00:01:47.459 --> 00:01:50.239 about applying it to like a billion search queries 00:01:50.239 --> 00:01:52.879 or way more complex tasks. It's kind of staggering. 00:01:53.099 --> 00:01:54.879 That speed changes the whole development cycle. 00:01:55.120 --> 00:01:57.439 Completely. It allows for super rapid iteration 00:01:57.439 --> 00:02:00.140 experimentation, just fundamentally changes the 00:02:00.140 --> 00:02:04.359 pace of AI progress. And it's not just XAI, right? 00:02:04.680 --> 00:02:07.079 Other big tech players are scaling up incredibly 00:02:07.079 --> 00:02:09.840 fast, too. We're seeing Meta, Microsoft, and 00:02:09.840 --> 00:02:12.900 OpenAI, Oracle. They're all running these huge 00:02:12.900 --> 00:02:17.379 clusters, 65 ,000 to maybe 100 ,000 GPUs. And 00:02:17.379 --> 00:02:19.479 Tesla's definitely in the game, too, 50 ,000 00:02:19.479 --> 00:02:23.319 GPUs for their Cortex Phase 1 project. And it's 00:02:23.319 --> 00:02:25.840 interesting. Most of these top 10 players, the 00:02:25.840 --> 00:02:27.860 ones we know about anyway, are based in the U 00:02:27.860 --> 00:02:30.729 .S. Right. But, you know, what's really noticeable 00:02:30.729 --> 00:02:33.729 is who's not on that public list. Google and 00:02:33.729 --> 00:02:35.669 Amazon. Yeah, that's a good point. They're obviously 00:02:35.669 --> 00:02:38.469 dominant forces in AI model development. Totally. 00:02:38.650 --> 00:02:41.389 But they're suspiciously quiet about their own 00:02:41.389 --> 00:02:43.830 supercomputer specs, given their history, their 00:02:43.830 --> 00:02:46.750 investments. It feels highly likely they're building 00:02:46.750 --> 00:02:49.330 huge clusters, too. Just maybe. In stealth mode. 00:02:49.469 --> 00:02:52.349 Or guarding trade secrets very closely. It suggests 00:02:52.349 --> 00:02:54.729 this foundational AI infrastructure is seen as 00:02:54.729 --> 00:02:57.909 deeply strategic, which brings up a really important 00:02:57.909 --> 00:02:59.550 question, especially when you look at the money 00:02:59.550 --> 00:03:01.990 involved. The sources say the cost of building 00:03:01.990 --> 00:03:04.870 and running these massive GPU clusters, it's 00:03:04.870 --> 00:03:08.509 doubling every 13 months. Doubling every 13 months. 00:03:08.689 --> 00:03:11.610 Wow. That's not just a big expense. That's an 00:03:11.610 --> 00:03:14.509 accelerating financial strain, even for the biggest 00:03:14.509 --> 00:03:17.439 companies out there. Even for them. And the sources 00:03:17.439 --> 00:03:20.060 really emphasize this point. Whoever controls 00:03:20.060 --> 00:03:23.740 these chips, this compute, they essentially control 00:03:23.740 --> 00:03:26.120 the next wave of AI. Yeah, it's like the new 00:03:26.120 --> 00:03:29.479 strategic resource, like oil refineries or maybe 00:03:29.479 --> 00:03:32.460 fiber optic cables. These GPU clusters are that 00:03:32.460 --> 00:03:34.740 critical infrastructure now. And that's probably 00:03:34.740 --> 00:03:36.860 why you see national labs like Lawrence Livermore 00:03:36.860 --> 00:03:39.159 also scaling up their own compute power pretty 00:03:39.159 --> 00:03:41.780 rapidly. They see the strategic importance, too. 00:03:42.169 --> 00:03:44.569 So thinking about these crazy costs and this 00:03:44.569 --> 00:03:47.849 intense competition, what does this whole escalating 00:03:47.849 --> 00:03:52.169 GPU arms race really mean for, say, smaller AI 00:03:52.169 --> 00:03:54.990 innovators or startups just trying to get going? 00:03:55.110 --> 00:03:56.870 Yeah, that's the big worry, isn't it? Is it creating 00:03:56.870 --> 00:03:58.990 this kind of AI oligarchy? That's exactly the 00:03:58.990 --> 00:04:01.050 concern. I mean, this rising barrier to entry. 00:04:01.819 --> 00:04:03.919 It kind of implies that only the players with 00:04:03.919 --> 00:04:06.219 the deepest pockets can really afford to train 00:04:06.219 --> 00:04:08.139 and deploy these cutting edge frontier models. 00:04:08.340 --> 00:04:10.280 It definitely pushes innovation towards them, 00:04:10.340 --> 00:04:13.300 centralizing power. OK, so while that hardware 00:04:13.300 --> 00:04:19.800 battle is clearly intense, AI's impact goes way 00:04:19.800 --> 00:04:23.569 beyond just the raw compute, right? Let's maybe 00:04:23.569 --> 00:04:25.850 pivot a bit now. Look at some broader ways AI 00:04:25.850 --> 00:04:29.370 is reshaping things like policy or new creative 00:04:29.370 --> 00:04:31.829 tools, even how we interact with tech itself. 00:04:32.129 --> 00:04:34.569 Yeah, sounds good. So policy wise, there was 00:04:34.569 --> 00:04:36.589 actually a pretty big move recently. The Trump 00:04:36.589 --> 00:04:39.829 administration released this really extensive 00:04:39.829 --> 00:04:42.930 AI action plan. Action plan. Yeah. And it's not 00:04:42.930 --> 00:04:44.870 just some quick memo. It's apparently a huge 00:04:44.870 --> 00:04:47.829 checklist over 90 different points, all aimed 00:04:47.829 --> 00:04:50.529 at making sure the U .S. stays, you know, an 00:04:50.529 --> 00:04:52.839 AI superpower. 90 points. That's comprehensive. 00:04:53.100 --> 00:04:54.839 It is. And what's interesting is they apparently 00:04:54.839 --> 00:04:57.779 shaped it using feedback from over 10 ,000 public 00:04:57.779 --> 00:04:59.839 comments. So lots of different voices went into 00:04:59.839 --> 00:05:01.660 it. That's quite a bit of public input. Yeah. 00:05:01.699 --> 00:05:03.259 And alongside policy, there's also this push 00:05:03.259 --> 00:05:06.620 to maybe democratize the building of AI. Exactly. 00:05:06.660 --> 00:05:09.339 Like, look at GitHub Spark. It now lets developers 00:05:09.339 --> 00:05:11.620 build and deploy AI apps just using prompts, 00:05:11.740 --> 00:05:13.879 like plain English instructions. Just prompts, 00:05:14.000 --> 00:05:16.959 not complex code. Right. Which is obviously super 00:05:16.959 --> 00:05:20.160 appealing to the, what, 150 million programmers 00:05:20.160 --> 00:05:22.480 already on GitHub. It basically turns coding 00:05:22.480 --> 00:05:24.779 into more of a conversation. Lowers the barrier 00:05:24.779 --> 00:05:26.439 significantly. Totally. And then you see these 00:05:26.439 --> 00:05:28.560 incredible creative uses popping up. There's 00:05:28.560 --> 00:05:32.000 an AI filmmaker, Lu Huang, who showed 10 really 00:05:32.000 --> 00:05:35.740 top tier examples of using JSON. JSON. OK, that's 00:05:35.740 --> 00:05:38.240 a data format, right? Makes things easy for chatbots 00:05:38.240 --> 00:05:40.800 to read. Yeah, exactly. Using JSON to generate 00:05:40.800 --> 00:05:43.800 these stunningly realistic, really polished commercials. 00:05:44.839 --> 00:05:48.720 quality is wow it shows ai can handle seriously 00:05:48.720 --> 00:05:52.319 complex creative work now that's impressive and 00:05:52.319 --> 00:05:54.480 if we connect that to how we interact with technology 00:05:54.480 --> 00:05:58.100 meta seems to be taking a pretty bold step Kind 00:05:58.100 --> 00:05:59.779 of like Neuralink, but different. They've got 00:05:59.779 --> 00:06:02.399 this AI wristband. Yeah, I saw that. It's designed 00:06:02.399 --> 00:06:04.980 to decode signals in your wrist nerve signals, 00:06:05.100 --> 00:06:06.959 basically to let you control devices with just 00:06:06.959 --> 00:06:09.540 tiny little hand gestures, almost invisible ones. 00:06:09.680 --> 00:06:12.040 So like minority report, but subtle. Kind of. 00:06:12.060 --> 00:06:14.139 It's definitely a big step towards making human 00:06:14.139 --> 00:06:16.500 computer interaction much more seamless, more 00:06:16.500 --> 00:06:19.319 intuitive, like tech becoming an extension of 00:06:19.319 --> 00:06:21.240 your thoughts. Imagine the possibilities there 00:06:21.240 --> 00:06:24.180 for accessibility, for just efficiency. For sure. 00:06:24.579 --> 00:06:28.060 And speaking of big moves. OpenAI looks like 00:06:28.060 --> 00:06:33.000 the next big thing is GPT -5. Expected maybe 00:06:33.000 --> 00:06:36.259 early August. GPT -5, okay. What's the word on 00:06:36.259 --> 00:06:38.680 that? Is it just a bigger GPT -4? Doesn't sound 00:06:38.680 --> 00:06:41.959 like it. It's described more as combining the 00:06:41.959 --> 00:06:44.879 powerful tech stuff from the GPT series that 00:06:44.879 --> 00:06:47.959 we know with their O -series models, like O3, 00:06:48.139 --> 00:06:51.000 which seem more focused on real -time, maybe 00:06:51.000 --> 00:06:55.740 multimodal interaction, seeing and hearing. Perhaps. 00:06:55.740 --> 00:07:00.660 So not just text, more integrated, more perceptive. 00:07:00.660 --> 00:07:02.480 That seems to be the idea. So you won't just 00:07:02.480 --> 00:07:04.779 see it as a text model like GPT -4 was initially. 00:07:04.879 --> 00:07:08.139 It's expected to be a more. evolved system. Could 00:07:08.139 --> 00:07:10.300 change how we interact with AI quite a bit. Okay. 00:07:10.360 --> 00:07:12.180 Interesting. And one more on the business side. 00:07:12.319 --> 00:07:14.860 Yeah. Legalon. It's an AI legal tech company. 00:07:14.959 --> 00:07:17.220 They just raised another $50 million for their 00:07:17.220 --> 00:07:19.680 tools. $50 million. What do their tools do? They 00:07:19.680 --> 00:07:22.120 help with legal work, specifically contract review. 00:07:22.339 --> 00:07:24.839 Apparently they're already used by 7 ,000 organizations. 00:07:25.060 --> 00:07:27.819 And get this, they speed up contract review by 00:07:27.819 --> 00:07:31.569 85%. 85%. That's... Massive efficiency gain, 00:07:31.689 --> 00:07:34.250 right? They already serve like a quarter of Japan's 00:07:34.250 --> 00:07:35.949 public companies and they're expanding fast. 00:07:36.250 --> 00:07:39.430 So clear, tangible impact in a field that's usually 00:07:39.430 --> 00:07:41.209 kind of slow to change. Okay, so with all these 00:07:41.209 --> 00:07:43.889 things happening, the policy, the building tools, 00:07:44.089 --> 00:07:46.310 the interfaces, the business applications, how 00:07:46.310 --> 00:07:49.810 quickly do you really think these new AI tools 00:07:49.810 --> 00:07:51.709 could fundamentally change our day -to -day work? 00:07:52.139 --> 00:07:54.779 our professional lives? I think very fast, honestly. 00:07:54.959 --> 00:07:57.540 Many of these seem designed for immediate practical 00:07:57.540 --> 00:08:00.120 use. They slot right into existing workflows, 00:08:00.420 --> 00:08:02.860 so the impact could be felt quite quickly. Okay, 00:08:02.939 --> 00:08:05.100 let's dive into maybe a few quick hits now. Things 00:08:05.100 --> 00:08:06.759 from the AI world that just kind of caught our 00:08:06.759 --> 00:08:09.120 eye, sparked some curiosity. And then let's end 00:08:09.120 --> 00:08:11.860 with something truly remarkable, connecting AI 00:08:11.860 --> 00:08:15.500 to human history. Sounds great. Yeah, a few things 00:08:15.500 --> 00:08:17.699 popped out. There was a funny piece. Someone 00:08:17.699 --> 00:08:21.319 asked an AI if their job writing humor was at 00:08:21.319 --> 00:08:24.220 risk. Kind of meta, right? Chuckles softly. Yeah. 00:08:24.980 --> 00:08:28.379 And I'll admit it, I still wrestle with prompt 00:08:28.379 --> 00:08:30.420 drift myself sometimes, you know, where the AI 00:08:30.420 --> 00:08:32.559 kind of slowly forgets what you originally asked 00:08:32.559 --> 00:08:35.720 it to do. So, yeah, I get the anxiety. It's a 00:08:35.720 --> 00:08:37.899 real thing. We also saw this alert from scientists 00:08:37.899 --> 00:08:40.799 sounding a pretty light alarm, actually, saying 00:08:40.799 --> 00:08:43.419 AI might soon get smart enough to outsmart our 00:08:43.419 --> 00:08:46.080 current safety checks. That's sobering. It definitely 00:08:46.080 --> 00:08:49.059 raises some immediate concerns. Yeah. On a lighter, 00:08:49.059 --> 00:08:51.440 more practical note, Google's got that new AI 00:08:51.440 --> 00:08:54.129 feature, right? Let's you... Virtually try on 00:08:54.129 --> 00:08:56.330 clothes you see in search results. Oh, yeah. 00:08:56.389 --> 00:08:59.049 The virtual try on. That could be a game changer 00:08:59.049 --> 00:09:01.649 for shopping online. Make it way more interactive. 00:09:02.009 --> 00:09:04.070 And Google's also rethinking search itself, aren't 00:09:04.070 --> 00:09:07.350 they? With this new AI curated web guide. Yeah. 00:09:07.389 --> 00:09:09.509 Trying to get more synthesized answers, not just 00:09:09.509 --> 00:09:12.909 links. Right. Moving beyond just keywords. We 00:09:12.909 --> 00:09:14.529 also saw this really interesting comparison. 00:09:15.190 --> 00:09:18.190 Someone gave the exact same complex financial 00:09:18.190 --> 00:09:21.330 task to an AI and a human financial planner. 00:09:21.490 --> 00:09:23.700 Oh, how'd that turn out? out well it just really 00:09:23.700 --> 00:09:25.399 highlighted their different strengths you know 00:09:25.399 --> 00:09:27.779 the ai was amazing at crunching the data finding 00:09:27.779 --> 00:09:30.559 patterns the human brought in like nuance emotional 00:09:30.559 --> 00:09:33.279 understanding strategic thinking about life goals 00:09:33.279 --> 00:09:36.299 different approaches fascinating comparison okay 00:09:36.299 --> 00:09:38.799 but here's the one that really got me connecting 00:09:38.799 --> 00:09:43.570 ai to our past google deep mind quietly released 00:09:43.570 --> 00:09:46.330 something called Aeneas. It's fully open source. 00:09:46.470 --> 00:09:49.490 And it's pretty groundbreaking. What it does 00:09:49.490 --> 00:09:53.029 is it reads, analyzes, and actually reconstructs 00:09:53.029 --> 00:09:56.710 ancient Roman texts using images of fragments, 00:09:56.830 --> 00:09:59.789 bits of text. Wow. Wait, so you feed it like 00:09:59.789 --> 00:10:03.169 a picture of a broken stone tablet with faded 00:10:03.169 --> 00:10:05.169 letters. Exactly. Maybe just a blurry photo. 00:10:05.269 --> 00:10:07.149 Incomplete inscription. And Aeneas pulls from 00:10:07.149 --> 00:10:10.850 this huge database like 176 ,000 ancient writings. 00:10:11.320 --> 00:10:14.220 And then it intelligently guesses where it might 00:10:14.220 --> 00:10:16.860 be from, when it was likely written, and crucially, 00:10:16.899 --> 00:10:19.279 what it used to say. That's incredible. How accurate 00:10:19.279 --> 00:10:21.279 is it? The numbers are pretty remarkable. It 00:10:21.279 --> 00:10:23.980 gets 72 % accuracy assigning inscriptions to 00:10:23.980 --> 00:10:26.259 the correct Roman province. 72%, okay. And it 00:10:26.259 --> 00:10:28.779 restores the broken or missing text with 73 % 00:10:28.779 --> 00:10:32.039 accuracy. Wow. And maybe just as important, 90 00:10:32.039 --> 00:10:34.360 % of the historians who tried it said it significantly 00:10:34.360 --> 00:10:36.679 boosted their confidence in their own research 00:10:36.679 --> 00:10:40.700 by like 44%. That's a huge endorsement from the 00:10:40.700 --> 00:10:43.659 experts themselves. And here's the kicker. It's 00:10:43.659 --> 00:10:46.299 totally free and open source. Open source. So 00:10:46.299 --> 00:10:49.860 anyone can use it. Adapt it. Yep. Researchers 00:10:49.860 --> 00:10:52.659 can download the model, fine tune it for their 00:10:52.659 --> 00:10:55.480 specific projects, maybe even adapt it to other 00:10:55.480 --> 00:10:57.539 ancient languages. That's amazing. It really 00:10:57.539 --> 00:10:59.840 raises this important question, doesn't it? How 00:10:59.840 --> 00:11:02.799 AI? through tools like Aeneas could just unlock 00:11:02.799 --> 00:11:05.820 vast amounts of human history, stuff that was 00:11:05.820 --> 00:11:08.779 basically inaccessible or would take lifetimes 00:11:08.779 --> 00:11:11.500 to piece together before. Yeah, definitely. So 00:11:11.500 --> 00:11:13.779 thinking about that open source aspect, could 00:11:13.779 --> 00:11:16.679 Aeneas or tools built like it potentially unlock 00:11:16.679 --> 00:11:19.320 other forgotten historical texts, maybe even 00:11:19.320 --> 00:11:22.019 decipher completely lost ancient languages? Oh, 00:11:22.039 --> 00:11:24.399 absolutely. I mean, it's open source nature is 00:11:24.399 --> 00:11:26.299 practically an invitation for that, applying 00:11:26.299 --> 00:11:27.940 it to different scripts, different cultures. 00:11:28.279 --> 00:11:30.899 Yeah, it could potentially reveal countless historical 00:11:30.899 --> 00:11:33.240 insights we just don't have access to right now. 00:11:33.840 --> 00:11:35.980 Sponsor. OK, so let's try to pull this all together. 00:11:36.159 --> 00:11:38.100 What does this all mean for you listening right 00:11:38.100 --> 00:11:41.840 now? We've seen today that this AI boom It's 00:11:41.840 --> 00:11:43.919 not just about clever software. It's deeply tied 00:11:43.919 --> 00:11:46.899 into this really high state's global arms race 00:11:46.899 --> 00:11:50.419 for raw computing power. And this battle, it's 00:11:50.419 --> 00:11:52.940 driving innovation like crazy, but it's also 00:11:52.940 --> 00:11:55.799 creating these immense financial pressures, these 00:11:55.799 --> 00:11:57.799 strategic pressures. Right. And if we connect 00:11:57.799 --> 00:12:00.279 that to the bigger picture, AI isn't just about 00:12:00.279 --> 00:12:02.940 the hardware, the compute power. It's rapidly 00:12:02.940 --> 00:12:05.860 transforming almost everything from how governments 00:12:05.860 --> 00:12:08.440 make policy to how commercials get made, even 00:12:08.440 --> 00:12:11.179 down to our fundamental ability to like. preserve 00:12:11.179 --> 00:12:14.200 and understand ancient history. It really feels 00:12:14.200 --> 00:12:16.240 like a foundational shift happening across so 00:12:16.240 --> 00:12:18.399 many different areas. It's redefining what's 00:12:18.399 --> 00:12:22.200 even possible. So as you think about all this, 00:12:22.360 --> 00:12:25.919 there's maybe a question to mull over, a provocative 00:12:25.919 --> 00:12:29.159 thought. As these AI models keep getting bigger 00:12:29.159 --> 00:12:30.940 and bigger, demanding more and more resources, 00:12:31.399 --> 00:12:33.480 where do you think the physical infrastructure 00:12:33.480 --> 00:12:36.639 is actually going to end up? The chips, the massive 00:12:36.639 --> 00:12:38.700 amounts of energy needed, these enormous data 00:12:38.700 --> 00:12:41.220 centers. where will they physically be built 00:12:41.220 --> 00:12:43.559 in the future? Yeah, that's a big question. And 00:12:43.559 --> 00:12:45.500 how might that geographical concentration, wherever 00:12:45.500 --> 00:12:48.080 it ends up being, how might that shift global 00:12:48.080 --> 00:12:51.179 power dynamics in the years ahead? That's definitely 00:12:51.179 --> 00:12:53.679 something worth pondering as you, you know, engage 00:12:53.679 --> 00:12:56.639 with this AI world unfolding around us. Hopefully 00:12:56.639 --> 00:12:58.879 this deep dive gave you some valuable nuggets 00:12:58.879 --> 00:13:01.159 to think about. Thank you for joining us on this 00:13:01.159 --> 00:13:03.899 deep dive into the latest in AI OTRO music.