WEBVTT 00:00:00.000 --> 00:00:02.480 Imagine needing the power of a nuclear plant, 00:00:02.660 --> 00:00:05.000 a whole gigawatt. Right. And imagine needing 00:00:05.000 --> 00:00:08.980 that much compute power every week just for AI. 00:00:09.259 --> 00:00:11.460 Yeah, it's kind of staggering, isn't it? That's 00:00:11.460 --> 00:00:13.300 the new reality Sam Altman is talking about. 00:00:13.380 --> 00:00:16.059 He calls it abundant intelligence. Exactly. And 00:00:16.059 --> 00:00:19.320 the response from industry, it was immediate 00:00:19.320 --> 00:00:22.440 and huge. Welcome to the Deep Dive. Today we're 00:00:22.440 --> 00:00:26.940 taking a calm. But I think really curious look 00:00:26.940 --> 00:00:30.579 at the sources that are defining this huge AI 00:00:30.579 --> 00:00:32.659 infrastructure race. Yeah, we're drawing from 00:00:32.659 --> 00:00:34.659 a newsletter that really digs into, you know, 00:00:34.679 --> 00:00:36.640 the hardware side, the software breakthroughs. 00:00:36.640 --> 00:00:39.020 And maybe most interestingly, these new ways 00:00:39.020 --> 00:00:41.000 people are trying to measure AI performance. 00:00:41.399 --> 00:00:44.039 Who's really best? So our mission today is really 00:00:44.039 --> 00:00:46.359 to unpack three main things. First, this literal 00:00:46.359 --> 00:00:50.820 race, building the grid for AI. Second, the just 00:00:50.820 --> 00:00:53.020 absurd speed of breakthroughs and what these 00:00:53.020 --> 00:00:55.880 models can do. AI's capability. Yeah, the pace 00:00:55.880 --> 00:00:58.179 is nuts. And third, a pretty revolutionary new 00:00:58.179 --> 00:01:00.140 way to figure out which AI works best because, 00:01:00.219 --> 00:01:02.640 spoiler, it's not the same for everyone. Let's 00:01:02.640 --> 00:01:05.719 start with the Giga Ones then. Altman's big idea. 00:01:05.939 --> 00:01:08.519 Right, this vision of abundant intelligence. 00:01:08.640 --> 00:01:12.180 He basically argues that access to AI or intelligence 00:01:12.180 --> 00:01:16.280 should be like a fundamental human right. That's 00:01:16.280 --> 00:01:18.599 a pretty profound statement. It is, but making 00:01:18.599 --> 00:01:22.340 that real, that takes scale. Almost unimaginable 00:01:22.340 --> 00:01:23.840 scale. And that's where the numbers come in. 00:01:24.219 --> 00:01:27.260 OpenAI's goal. Yeah. One gigawatt of compute 00:01:27.260 --> 00:01:30.700 capacity per week. Exactly. And that demand just 00:01:30.700 --> 00:01:32.920 kicked off this massive infrastructure build 00:01:32.920 --> 00:01:35.879 out. Hyperscalers, VCs, everyone jumped in. The 00:01:35.879 --> 00:01:38.640 investment scale is. Yeah. It's genuinely tough 00:01:38.640 --> 00:01:41.790 to visualize. We hear big numbers, but. 5 .5 00:01:41.790 --> 00:01:43.650 gigawatts. Yeah, that's Oracle. They're building 00:01:43.650 --> 00:01:46.510 these huge data centers. Texas, New Mexico, the 00:01:46.510 --> 00:01:49.430 Midwest. 5 .5 gigawatts. What does that even 00:01:49.430 --> 00:01:51.769 compare to? Well, think about a major city like 00:01:51.769 --> 00:01:53.950 Dallas, maybe. That's more than its base power 00:01:53.950 --> 00:01:55.849 needs. So, yeah, it's massive. And they're talking, 00:01:55.950 --> 00:01:58.590 what, 25 ,000 jobs just from that push? Wow. 00:01:59.170 --> 00:02:01.310 Construction tech. Right. They're not just building 00:02:01.310 --> 00:02:02.810 server farms. It feels like they're building, 00:02:02.890 --> 00:02:04.930 you know, the next century's economic engine. 00:02:05.030 --> 00:02:06.730 And soft banks in the mix, too, right? Yeah, 00:02:06.730 --> 00:02:09.110 correct. Aggressively. Oh, yeah. They committed 00:02:09.110 --> 00:02:12.469 1 .5 gigawatts and they want it done in 18 months. 00:02:12.590 --> 00:02:15.030 18 months. That's incredibly fast. New sites 00:02:15.030 --> 00:02:17.729 in Ohio, Texas. These aren't small steps. These 00:02:17.729 --> 00:02:21.169 are like moonshots, big bets. So if you add it 00:02:21.169 --> 00:02:23.810 all up, where are we heading? Well, the early 00:02:23.810 --> 00:02:25.710 commitments, if you tally them, point towards 00:02:25.710 --> 00:02:30.479 something like $500 billion. And 10 gigawatts 00:02:30.479 --> 00:02:33.599 total by the end of 2025. Half a trillion dollars. 00:02:33.719 --> 00:02:35.680 Yeah. 10 gigawatts. And they're already well 00:02:35.680 --> 00:02:38.560 on their way. Like 400 billion and 7 gigawatts 00:02:38.560 --> 00:02:40.680 are basically planned and funded already. It's 00:02:40.680 --> 00:02:43.180 like stacking these incredibly advanced Lego 00:02:43.180 --> 00:02:46.099 blocks, you know, but at warp speed. And then 00:02:46.099 --> 00:02:47.939 you have players like Alibaba taking a slightly 00:02:47.939 --> 00:02:50.139 different angle. Right. They talk about being 00:02:50.139 --> 00:02:52.500 the electric company for AI, but they're also 00:02:52.500 --> 00:02:55.360 going vertical. Super fast. Like dropping six 00:02:55.360 --> 00:02:58.719 major product launches in one day. Exactly. It 00:02:58.719 --> 00:03:00.960 highlights that speed, but also owning the whole 00:03:00.960 --> 00:03:03.039 process from the chip all the way up to the app. 00:03:03.120 --> 00:03:05.620 That seems key to their strategy. So why does 00:03:05.620 --> 00:03:09.139 this infrastructure layer matter so much to someone 00:03:09.139 --> 00:03:11.659 listening right now? Well, because this isn't 00:03:11.659 --> 00:03:13.979 just about tech companies anymore. It's becoming 00:03:13.979 --> 00:03:18.439 a race between like governments, big money VCs, 00:03:18.479 --> 00:03:22.360 the cloud giants to own the fundamental infrastructure, 00:03:22.500 --> 00:03:25.340 the plumbing for intelligence itself. It goes 00:03:25.340 --> 00:03:28.479 way beyond just making chat bots a bit better. 00:03:28.539 --> 00:03:30.719 It's foundational. OK, so the scale is huge. 00:03:30.900 --> 00:03:34.000 The cost is astronomical. Does all this frantic 00:03:34.000 --> 00:03:39.099 building actually mean that AI access gets cheaper? 00:03:39.800 --> 00:03:43.060 and sooner for regular people. Yeah, I think 00:03:43.060 --> 00:03:45.300 the signs point that way. This level of aggressive 00:03:45.300 --> 00:03:47.719 competition, the sheer amount of money pouring 00:03:47.719 --> 00:03:50.500 in, it suggests better, cheaper access is coming. 00:03:50.659 --> 00:03:52.580 And probably fast. Right, because of that competition 00:03:52.580 --> 00:03:54.860 and just the pace of the tech improving. Exactly. 00:03:55.060 --> 00:03:57.300 Okay, so all this hardware is being built because 00:03:57.300 --> 00:03:59.080 the software, the capabilities are demanding 00:03:59.080 --> 00:04:02.020 it. Let's shift gears then from building speed 00:04:02.020 --> 00:04:04.599 to intelligence speed. yeah the physical stuff 00:04:04.599 --> 00:04:06.520 is really just trying to keep up with the software 00:04:06.520 --> 00:04:08.840 breakthroughs the pace of improvement and models 00:04:08.840 --> 00:04:11.979 and what they can do it's uh it's kind of absurd 00:04:11.979 --> 00:04:14.039 right now we saw that play out recently didn't 00:04:14.039 --> 00:04:16.560 we there's that live marketing showdown oh yeah 00:04:16.560 --> 00:04:20.860 with the big llms chat gpt gemini perplexity 00:04:20.860 --> 00:04:23.420 claude right and in a real business test one 00:04:23.420 --> 00:04:25.550 of them just walked away with five out of the 00:04:25.550 --> 00:04:28.129 six wins. That shows you like the immediate practical 00:04:28.129 --> 00:04:30.269 value is already there. It's not theoretical. 00:04:30.649 --> 00:04:32.750 And then there's the sheer size increase. You 00:04:32.750 --> 00:04:35.050 mentioned Quinn 3 Max. Yeah. Alibaba's new model. 00:04:35.209 --> 00:04:39.029 One trillion parameters. Whoa. One trillion. 00:04:39.660 --> 00:04:42.019 Just try to imagine scaling that. A trillion 00:04:42.019 --> 00:04:44.439 variables for the model to learn from. That's 00:04:44.439 --> 00:04:47.120 a huge jump in complexity. It's a massive number. 00:04:47.199 --> 00:04:49.339 And it's immediately showing results, beating 00:04:49.339 --> 00:04:52.620 older models, even early GPT -5 versions on some 00:04:52.620 --> 00:04:55.819 tasks. But the really stunning thing. What's 00:04:55.819 --> 00:04:58.259 that? It's math performance. It scored 100 % 00:04:58.259 --> 00:05:01.959 on the AME25 math benchmark. 100%. Wow. Okay, 00:05:02.100 --> 00:05:03.939 that's the moment of wonder right there. That's 00:05:03.939 --> 00:05:07.500 not just good. Perfect mastery on a really tough 00:05:07.500 --> 00:05:10.639 academic test. Right. Models hitting expert levels 00:05:10.639 --> 00:05:12.439 almost right out of the gate. And it's translating 00:05:12.439 --> 00:05:14.060 to the real world, too, isn't it? The finance 00:05:14.060 --> 00:05:17.180 exam. Oh, yeah. The CFA exam. Claude Opus. Gemini 00:05:17.180 --> 00:05:20.300 2 .5 Pro. They both passed level three. Which 00:05:20.300 --> 00:05:23.139 is notoriously the hardest level. Takes humans, 00:05:23.220 --> 00:05:25.560 what, over a thousand hours of study? Usually, 00:05:25.720 --> 00:05:29.060 yeah. Intense study. And these models did it. 00:05:29.500 --> 00:05:32.579 In minutes. In minutes. The implication there 00:05:32.579 --> 00:05:35.420 for, like, knowledge work, professional training, 00:05:35.500 --> 00:05:38.939 it's just huge and immediate. You have to wonder, 00:05:39.019 --> 00:05:41.959 what's the value of that certification if a machine 00:05:41.959 --> 00:05:45.500 can ace it instantly? Exactly. And look how fast 00:05:45.500 --> 00:05:48.220 it gets integrated. Microsoft already plugged 00:05:48.220 --> 00:05:51.639 Cloud into 365 Copilot. That's the first external 00:05:51.639 --> 00:05:54.019 AI model inside the main Office tools, right? 00:05:54.079 --> 00:05:56.139 Yep. The speed from breakthrough to application 00:05:56.139 --> 00:05:59.019 is relentless. It really is. Okay, here's maybe 00:05:59.019 --> 00:06:02.060 a vulnerable admission. Uh -oh. Even trying to 00:06:02.060 --> 00:06:03.699 follow this closely, I still kind of wrestle 00:06:03.699 --> 00:06:06.379 with prompt drift. Oh, absolutely. You're definitely 00:06:06.379 --> 00:06:08.399 not alone there, where you ask the same thing 00:06:08.399 --> 00:06:10.759 or use the same image prompts like two weeks 00:06:10.759 --> 00:06:12.540 apart. And the results are completely different 00:06:12.540 --> 00:06:14.959 because the model changed underneath without 00:06:14.959 --> 00:06:17.720 you knowing. Yeah. It's a real challenge, especially, 00:06:17.800 --> 00:06:20.540 like you said, with visual tools. That Redditor's 00:06:20.540 --> 00:06:25.220 game, real versus AI images. It's getting almost 00:06:25.220 --> 00:06:28.120 impossible to tell consistently now. For anyone. 00:06:28.639 --> 00:06:30.839 We're struggling to keep up with the tools, really. 00:06:31.079 --> 00:06:33.300 But the tools just keep coming. They do. Quick 00:06:33.300 --> 00:06:36.259 hits here. Cling 2 .5 Turbo. Getting really good 00:06:36.259 --> 00:06:39.379 at realistic images, video. Pushing towards that 00:06:39.379 --> 00:06:41.980 believable synthetic media. And that AI bandage 00:06:41.980 --> 00:06:45.079 thing. That sounds wild. Isn't it? An AI -powered 00:06:45.079 --> 00:06:48.300 smart bandage monitors the wound, adjusts things. 00:06:48.480 --> 00:06:52.420 They claim it heals 25 % faster. And maybe less 00:06:52.420 --> 00:06:54.639 exciting, but relevant. Ads are probably coming 00:06:54.639 --> 00:06:58.639 to free chat GPT by 2026. Ah, ties back to the 00:06:58.639 --> 00:07:00.579 compute costs, I guess. Pretty much has to, yeah. 00:07:00.779 --> 00:07:03.139 Okay, so with these models acing things like 00:07:03.139 --> 00:07:06.139 the CFA exam in minutes, what does that tell 00:07:06.139 --> 00:07:08.079 us about where the ceiling is? Is there even 00:07:08.079 --> 00:07:11.120 a ceiling for these current models? That immediate 00:07:11.120 --> 00:07:13.420 mastery, it really shows we have to constantly 00:07:13.420 --> 00:07:15.839 reevaluate. Even models that were top tier just 00:07:15.839 --> 00:07:19.740 months ago, the ceiling just keeps rising perpetually. 00:07:19.860 --> 00:07:21.339 Okay, so if the capabilities are moving that 00:07:21.339 --> 00:07:23.279 fast, the way we measure them needs to change 00:07:23.279 --> 00:07:27.079 too, right? Exactly. Which brings us to the metrics, 00:07:27.279 --> 00:07:30.300 how we decide what's good. Traditionally, we've 00:07:30.300 --> 00:07:33.360 had leaderboards. El Marina, places like that, 00:07:33.420 --> 00:07:35.019 they've been the standard. Yeah, but they had 00:07:35.019 --> 00:07:37.500 this huge blind spot, a really unavoidable one. 00:07:37.620 --> 00:07:40.240 They basically treat every user the same. Doesn't 00:07:40.240 --> 00:07:42.259 matter who you are, where you are, why you're 00:07:42.259 --> 00:07:45.740 using it. An engineer in Tokyo, a lawyer in London. 00:07:45.879 --> 00:07:48.740 Or a student in Buenos Aires, yeah. The leaderboards 00:07:48.740 --> 00:07:52.180 just give you a raw performance score, undifferentiated. 00:07:52.379 --> 00:07:54.420 That doesn't really tell you if it's useful for 00:07:54.420 --> 00:07:57.329 your specific need, though. Not at all. So the 00:07:57.329 --> 00:08:01.430 new idea is from Scale AI. They launched Seal 00:08:01.430 --> 00:08:03.889 Showdown. Seal Showdown. Okay, what's different? 00:08:04.089 --> 00:08:06.829 It completely flips the ranking method. Instead 00:08:06.829 --> 00:08:09.149 of just raw speed or whatever, it ranks models 00:08:09.149 --> 00:08:12.389 based on real user preferences, and it connects 00:08:12.389 --> 00:08:16.069 those preferences to demographic info. Oh, okay. 00:08:16.149 --> 00:08:18.850 So it's subjective based on who's using it. Exactly. 00:08:18.990 --> 00:08:21.490 Think about, say you're a bank choosing an LLM 00:08:21.490 --> 00:08:23.509 for your Spanish -speaking customers in Miami. 00:08:24.170 --> 00:08:27.589 El Marina, useless for that decision. But SEAL 00:08:27.589 --> 00:08:29.810 Showdown could actually give you useful data 00:08:29.810 --> 00:08:33.070 because it breaks down results by age, education, 00:08:33.450 --> 00:08:36.759 language, country. So you can pick the model 00:08:36.759 --> 00:08:40.379 that that specific group actually prefers. Precisely. 00:08:40.379 --> 00:08:42.700 That's actionable utility. How are they getting 00:08:42.700 --> 00:08:45.139 the data? Globally. They say preferences from 00:08:45.139 --> 00:08:48.860 100 countries, 70 languages. Try and get a real 00:08:48.860 --> 00:08:51.340 global snapshot. And how do they stop people 00:08:51.340 --> 00:08:54.200 from gaming the system? Like model creators trying 00:08:54.200 --> 00:08:56.360 to boost their scores. Seems like they've thought 00:08:56.360 --> 00:09:00.289 about that. Voting is voluntary, anonymous, and 00:09:00.289 --> 00:09:02.250 they hold back the results from public view for 00:09:02.250 --> 00:09:04.950 60 days to prevent that kind of manipulation. 00:09:05.330 --> 00:09:07.730 Okay, so it's a big shift. El Marina gives you 00:09:07.730 --> 00:09:10.409 raw speed, like engine horsepower. SEAL Showdown 00:09:10.409 --> 00:09:12.730 tells you which car people actually like driving 00:09:12.730 --> 00:09:14.490 in their specific neighborhoods. That's a great 00:09:14.490 --> 00:09:17.230 analogy, yeah. Which tool is perceived as best 00:09:17.230 --> 00:09:19.929 by your users? That creates real market pressure. 00:09:20.330 --> 00:09:23.210 So you'd expect the big players, OpenAI, Anthropic, 00:09:23.490 --> 00:09:27.120 Google. They'll have to react somehow. Oh, definitely. 00:09:27.360 --> 00:09:29.059 They'll likely try to build their own versions 00:09:29.059 --> 00:09:31.179 or maybe just plug into scale system. You can't 00:09:31.179 --> 00:09:33.100 really ignore that kind of specific customer 00:09:33.100 --> 00:09:35.519 feedback. So what's the core assumption about 00:09:35.519 --> 00:09:38.779 AI performance that this new benchmark really 00:09:38.779 --> 00:09:41.440 challenges? It fundamentally challenges that 00:09:41.440 --> 00:09:43.980 old idea that there's one single best model, 00:09:44.100 --> 00:09:47.399 one size fits all, that it can serve every single 00:09:47.399 --> 00:09:50.460 user everywhere equally well. SEAL says, no, 00:09:50.500 --> 00:09:53.440 it's more complicated. Okay, let's try to wrap 00:09:53.440 --> 00:09:55.740 this deep dive up. Quick recap of the big layers 00:09:55.740 --> 00:09:57.559 we talked about. Sure. First, the foundation, 00:09:57.840 --> 00:10:00.960 the infrastructure, this gigawatt race heading 00:10:00.960 --> 00:10:04.100 towards $500 billion, all chasing that abundant 00:10:04.100 --> 00:10:06.779 intelligence vision. It's a battle for the means 00:10:06.779 --> 00:10:09.059 of production, really. Second, the capabilities, 00:10:09.340 --> 00:10:12.259 just stunning growth, models mastering complex 00:10:12.259 --> 00:10:15.519 skills like that CFA exam almost instantly. It's 00:10:15.519 --> 00:10:17.460 constantly pushing the performance ceiling up, 00:10:17.539 --> 00:10:20.289 impacting work and education. Yeah. And third, 00:10:20.409 --> 00:10:22.889 how we measure it all. Shifting towards personalized, 00:10:23.250 --> 00:10:25.809 user -focused benchmarks like SEAL Showdown. 00:10:26.009 --> 00:10:28.970 Moving past Rossby to ask, who is this really 00:10:28.970 --> 00:10:31.269 best for? It brings us right back to Altman's 00:10:31.269 --> 00:10:33.470 original claim, doesn't it? Yeah. If AI access 00:10:33.470 --> 00:10:35.970 becomes a fundamental right, how does that change 00:10:35.970 --> 00:10:39.070 global power dynamics? When only a handful of 00:10:39.070 --> 00:10:42.250 massive companies, Oracle, SoftBank, Alibaba, 00:10:42.269 --> 00:10:46.330 can actually afford to build the 10 gigawatts 00:10:46.330 --> 00:10:48.490 needed to deliver that right. That's the billion. 00:10:49.120 --> 00:10:51.019 Maybe trillion dollar question, isn't it? Who 00:10:51.019 --> 00:10:53.419 owns the intelligence grid and what responsibility 00:10:53.419 --> 00:10:56.360 comes with that ownership? That's the core geopolitical 00:10:56.360 --> 00:10:59.039 tension underneath this whole race. Something 00:10:59.039 --> 00:11:01.539 to think about. We'd encourage you listening 00:11:01.539 --> 00:11:04.440 to consider the sources we unpacked. Maybe think 00:11:04.440 --> 00:11:06.940 about how your own use of AI tools or how you 00:11:06.940 --> 00:11:09.120 judge them might be skewed by those old metrics. 00:11:09.299 --> 00:11:11.720 Are you using the fastest model or the one that 00:11:11.720 --> 00:11:14.059 actually works best for you? Good question to 00:11:14.059 --> 00:11:16.019 ponder. Thank you for joining us on the Deep 00:11:16.019 --> 00:11:16.379 Dive.