WEBVTT 00:00:00.000 --> 00:00:02.740 We often talk about A .I. in terms of, you know, 00:00:02.740 --> 00:00:05.360 the algorithms, the models themselves. But maybe 00:00:05.360 --> 00:00:07.299 the real story, the one that's kind of mind blowing 00:00:07.299 --> 00:00:11.119 right now is the physical scale. Yeah. The infrastructure. 00:00:11.419 --> 00:00:13.480 Oh, absolutely. It's staggering. And there's 00:00:13.480 --> 00:00:15.980 this one stat about open A .I. that just puts 00:00:15.980 --> 00:00:17.699 everything into perspective. They're apparently 00:00:17.699 --> 00:00:20.120 aiming to build something they call an infrastructure 00:00:20.120 --> 00:00:23.559 factory. An infrastructure factory. Yeah. Designed 00:00:23.559 --> 00:00:27.079 to produce one gigawatt of A .I. compute every 00:00:27.079 --> 00:00:32.950 single. One gigawatt weekly. I mean, it's hard 00:00:32.950 --> 00:00:35.130 to even picture that kind of pace. It really 00:00:35.130 --> 00:00:36.770 is. It means they're operating on a completely 00:00:36.770 --> 00:00:39.030 different level. If you try to contextualize 00:00:39.030 --> 00:00:41.429 that, you're basically looking at combining like... 00:00:41.679 --> 00:00:44.200 All the big NVIDIA deployments, plus the Tesla 00:00:44.200 --> 00:00:47.359 Gigafactory, plus major AWS data centers, all 00:00:47.359 --> 00:00:49.420 running at once. And they plan to just keep doing 00:00:49.420 --> 00:00:51.539 that. Yeah. Week after week. That seems to be 00:00:51.539 --> 00:00:53.740 the plan. Yeah. Just to handle the training and 00:00:53.740 --> 00:00:56.140 deployment needs for whatever comes next. It 00:00:56.140 --> 00:00:58.520 forces a total rethink of, well, compute economics. 00:00:59.200 --> 00:01:01.359 Welcome to the Deep Dive. Today, we're going 00:01:01.359 --> 00:01:04.609 to dive into the latest industry sources. cut 00:01:04.609 --> 00:01:07.109 through some of the noise and pull out the key 00:01:07.109 --> 00:01:11.189 things you probably need to know. We'll be looking 00:01:11.189 --> 00:01:14.230 at some big corporate shifts, some very real 00:01:14.230 --> 00:01:16.170 practical threats that are emerging, and also 00:01:16.170 --> 00:01:18.450 this rising global competition, especially around 00:01:18.450 --> 00:01:21.189 the cost of performance. Right. So roadmap wise, 00:01:21.430 --> 00:01:24.409 first up, we'll unpack the new structure at OpenAI, 00:01:24.609 --> 00:01:27.040 what that means for their speed. Then we'll get 00:01:27.040 --> 00:01:28.920 into some immediate disruptions, things like 00:01:28.920 --> 00:01:32.420 fake AI receipts, workforce impacts, new tools. 00:01:32.739 --> 00:01:34.760 And finally, we absolutely have to talk about 00:01:34.760 --> 00:01:37.680 these super competitive Chinese models. They're 00:01:37.680 --> 00:01:40.079 beating Western models pretty badly on cost. 00:01:40.340 --> 00:01:43.659 Let's jump in. Segment one, OpenAI's new reality. 00:01:44.079 --> 00:01:47.700 OK, so focusing first on the, let's say, the 00:01:47.700 --> 00:01:50.579 strategic and financial side. The big news is 00:01:50.579 --> 00:01:53.159 OpenAI has finally finished this really complex 00:01:53.159 --> 00:01:56.010 legal reshuffle. They are now officially a for 00:01:56.010 --> 00:01:58.150 -profit company. That's right. The pivot is complete. 00:01:58.609 --> 00:02:01.250 But that ownership is still intentionally messy. 00:02:01.370 --> 00:02:02.750 You know, it's not like a standard corporate 00:02:02.750 --> 00:02:05.069 setup, which keeps people guessing. Yeah. We 00:02:05.069 --> 00:02:06.790 learned the original nonprofit foundation, it 00:02:06.790 --> 00:02:09.069 still owns 26 % of the new thing, the OpenAI 00:02:09.069 --> 00:02:11.789 group. Right. And meanwhile, Microsoft, they 00:02:11.789 --> 00:02:14.789 are dug in deep. They've got about 27 % of this 00:02:14.789 --> 00:02:17.569 new group. That stake is valued at, what, around 00:02:17.569 --> 00:02:22.259 $135 billion now? Wow. But the real power might 00:02:22.259 --> 00:02:24.960 be the exclusive model rights they hold. That 00:02:24.960 --> 00:02:28.599 runs until 2032. 2032. That's a long time. That 00:02:28.599 --> 00:02:30.599 gives Microsoft huge control over where this 00:02:30.599 --> 00:02:32.439 tech goes commercially, doesn't it? For the next 00:02:32.439 --> 00:02:34.860 decade, pretty much. Yeah. And speaking of valuation 00:02:34.860 --> 00:02:37.400 frenzy, here's a kind of interesting side note. 00:02:37.460 --> 00:02:39.860 Elon Musk apparently tried to buy them not long 00:02:39.860 --> 00:02:45.139 ago for $97 billion. $97 billion. Obviously, 00:02:45.159 --> 00:02:46.560 that didn't happen. Maybe it wasn't even enough. 00:02:46.599 --> 00:02:49.439 Who knows? But the why behind this whole for 00:02:49.439 --> 00:02:52.580 profit shift is key. They basically have to remove 00:02:52.580 --> 00:02:55.719 all the legal friction. The whole structure made 00:02:55.719 --> 00:02:58.120 it hard to raise the truly astronomical amounts 00:02:58.120 --> 00:02:59.919 of money they need for that physical infrastructure 00:02:59.919 --> 00:03:02.139 we talked about. So open the floodgates for investment. 00:03:02.560 --> 00:03:06.199 Scale up faster. Exactly. Which brings us back 00:03:06.199 --> 00:03:08.740 to that infrastructure goal. The infrastructure 00:03:08.740 --> 00:03:13.300 factory. One gigawatt per week. Whoa. I mean, 00:03:13.319 --> 00:03:17.520 just imagine trying to scale to handle say a 00:03:17.520 --> 00:03:20.740 billion queries consistently. That kind of investment, 00:03:20.900 --> 00:03:23.039 that output. It's removed all the hardware bottlenecks 00:03:23.039 --> 00:03:24.939 they've ever had, right? It unlocks speed in 00:03:24.939 --> 00:03:26.900 a way we haven't seen. And you see that urgency 00:03:26.900 --> 00:03:29.659 reflected internally, too. Oh, so? Well, their 00:03:29.659 --> 00:03:32.280 chief scientist, Jakub Pachocki, recently suggested 00:03:32.280 --> 00:03:34.520 superintelligence might be less than 10 years 00:03:34.520 --> 00:03:36.699 away. Less than 10 years. Yeah, and that doesn't 00:03:36.699 --> 00:03:38.539 sound like a casual prediction. It sounds like 00:03:38.539 --> 00:03:40.300 it's based on the speed they're seeing right 00:03:40.300 --> 00:03:43.300 now internally. So the implication there is we 00:03:43.300 --> 00:03:46.379 could see big jumps in model performance, maybe 00:03:46.379 --> 00:03:49.189 even in the next six months. Before we even get 00:03:49.189 --> 00:03:52.030 to GPT -6. That seems to be the hint, yes. The 00:03:52.030 --> 00:03:53.990 training runs are moving faster than maybe even 00:03:53.990 --> 00:03:55.949 they predicted. Okay, so here's a question then. 00:03:56.669 --> 00:03:59.110 With all this money pouring in, this infrastructure, 00:03:59.330 --> 00:04:02.349 this speed, does that guarantee open AI market 00:04:02.349 --> 00:04:04.889 dominance long term? Well, the capital flow is 00:04:04.889 --> 00:04:08.150 huge, definitely. But those proprietary rights, 00:04:08.270 --> 00:04:10.629 the control aspects, they really complicate the 00:04:10.629 --> 00:04:12.689 market picture. Right. So the money's there, 00:04:12.830 --> 00:04:15.229 but control rights make it complicated. Exactly. 00:04:15.710 --> 00:04:18.629 OK, let's shift gears from the big money story 00:04:18.629 --> 00:04:22.810 to more immediate practical things hitting people 00:04:22.810 --> 00:04:25.810 segment two. And first up is something that's 00:04:25.810 --> 00:04:29.670 honestly a bit scary. The realism of these models 00:04:29.670 --> 00:04:32.980 is creating a well. a crisis in digital trust. 00:04:33.399 --> 00:04:35.839 We need to talk about fake AI receipts. Yeah, 00:04:35.920 --> 00:04:37.819 this is becoming a million dollar problem. Literally. 00:04:38.060 --> 00:04:40.160 We're not talking about clumsy Photoshop anymore. 00:04:40.300 --> 00:04:42.540 These are like professional level fakes. Good 00:04:42.540 --> 00:04:44.740 enough to fool corporate expense systems. So 00:04:44.740 --> 00:04:47.040 the takeaway is you just can't trust your eyes 00:04:47.040 --> 00:04:49.800 anymore. Not with digital documents. Pretty much. 00:04:49.819 --> 00:04:52.519 Visual evidence alone isn't enough. Assume verification 00:04:52.519 --> 00:04:55.800 is needed. Digital trust is eroding fast. Okay. 00:04:56.079 --> 00:04:58.279 At the same time, though, powerful creative tools 00:04:58.279 --> 00:05:00.889 are getting easier to access, right? Adobe just 00:05:00.889 --> 00:05:02.910 put AI assistance into Express and Photoshop. 00:05:03.110 --> 00:05:05.850 That's huge, yeah. Tools that used to need serious 00:05:05.850 --> 00:05:08.490 design skills, now usable with simple text prompts. 00:05:08.689 --> 00:05:11.129 It makes creativity more accessible, sure. But 00:05:11.129 --> 00:05:13.250 it also probably fuels that forgery problem we 00:05:13.250 --> 00:05:15.149 just mentioned. Exactly. Double -edged sword. 00:05:15.930 --> 00:05:20.110 And look at commerce integration. OpenAI and 00:05:20.110 --> 00:05:22.689 PayPal are partnering up. For instant checkout 00:05:22.689 --> 00:05:25.269 within ChatGPT. Right. So you could be chatting, 00:05:25.389 --> 00:05:27.350 the AI suggests something, and boom, you buy 00:05:27.350 --> 00:05:29.730 it right there. Conversation at checkout. Seamless. 00:05:29.870 --> 00:05:33.470 And speaking of access, look at India. They're 00:05:33.470 --> 00:05:37.250 getting 12 months of free ChatGPT Go access for 00:05:37.250 --> 00:05:39.970 a limited time. That's a big strategic play for 00:05:39.970 --> 00:05:43.870 market share. And this tech leap is causing weird... 00:05:44.110 --> 00:05:46.670 contrasts in the corporate world, too, like Amazon 00:05:46.670 --> 00:05:48.889 laying off thousands. Right. Thousands of people 00:05:48.889 --> 00:05:50.930 gone. But it's not because Amazon's struggling 00:05:50.930 --> 00:05:54.290 financially. No, they're doing it while massively 00:05:54.290 --> 00:05:56.850 increasing their spending on AI development. 00:05:57.009 --> 00:06:00.129 It's a really stark restructuring of the workforce 00:06:00.129 --> 00:06:02.470 happening right now. It really shows the tradeoff 00:06:02.470 --> 00:06:05.290 companies are making, swapping human capital 00:06:05.290 --> 00:06:08.389 for compute capital. And sometimes, you know, 00:06:08.410 --> 00:06:11.350 that compute capital doesn't quite behave. Yeah. 00:06:11.370 --> 00:06:14.360 Yeah. Slight juggle. The community reactions 00:06:14.360 --> 00:06:17.439 are always quick when models stumble. I saw that 00:06:17.439 --> 00:06:20.699 post about ChatGPT Atlas being dragged into the 00:06:20.699 --> 00:06:23.360 bin after everyone initially called it top one. 00:06:23.579 --> 00:06:25.740 Oh yeah, that went viral. You know, that kind 00:06:25.740 --> 00:06:27.860 of hit home for me. I mean, I still wrestle with 00:06:27.860 --> 00:06:29.740 prompt drift myself sometimes. You figure something 00:06:29.740 --> 00:06:31.560 out one day, the perfect prompt, and the next 00:06:31.560 --> 00:06:35.439 day, poof. It just doesn't work the same. It's 00:06:35.439 --> 00:06:37.399 frustrating. It really is. It's just the nature 00:06:37.399 --> 00:06:40.620 of working on the bleeding edge, I guess. Unstable. 00:06:41.689 --> 00:06:43.589 Meanwhile, governments are trying to react to 00:06:43.589 --> 00:06:46.449 all this. The EU just launched a $5 billion fund. 00:06:46.970 --> 00:06:50.069 Right, for AI, quantum computing, semiconductors. 00:06:51.269 --> 00:06:53.550 Trying to catch up. Explicitly, yeah. Trying 00:06:53.550 --> 00:06:55.490 to match the scale of investment coming out of 00:06:55.490 --> 00:06:57.949 the U .S. and Asia. Okay, so thinking about those 00:06:57.949 --> 00:07:01.370 fake receipts again. Yeah. Given that risk, how 00:07:01.370 --> 00:07:03.870 much should we actually rely on digital confirmation 00:07:03.870 --> 00:07:07.560 for, say, financial stuff? I think we have to 00:07:07.560 --> 00:07:09.920 assume verification is needed outside the document 00:07:09.920 --> 00:07:11.680 itself. You can't just trust the digital file 00:07:11.680 --> 00:07:15.259 anymore. Right. Digital trust is eroding. Always 00:07:15.259 --> 00:07:18.980 cross verify the source. OK, let's move to our 00:07:18.980 --> 00:07:22.519 final segment, the cost frontier. This is where 00:07:22.519 --> 00:07:25.240 models from China are really shaking things up. 00:07:25.300 --> 00:07:27.560 Yeah, this is a potentially huge market shift 00:07:27.560 --> 00:07:30.019 coming from the east. We need to talk about a 00:07:30.019 --> 00:07:33.800 model called Minimax M2. OK, Minimax M2. What's 00:07:33.800 --> 00:07:36.319 the goal there? The goal is basically GPT -4 00:07:36.319 --> 00:07:38.540 level performance, but without the crazy high 00:07:38.540 --> 00:07:40.680 price tag you get with the big U .S. models. 00:07:40.920 --> 00:07:43.800 And the cost difference is big. Staggering. M2 00:07:43.800 --> 00:07:48.180 apparently runs at just 8 % of the cost of running 00:07:48.180 --> 00:07:50.740 something comparable like Cloud Sonnet. 8 %? 00:07:50.800 --> 00:07:53.639 Wow. That changes everything, doesn't it? How 00:07:53.639 --> 00:07:55.879 do they do that? Well, the key is the tech underneath. 00:07:56.060 --> 00:07:58.660 M2 is what's called a sparse mixture of experts 00:07:58.660 --> 00:08:02.300 model and Moe. Okay, jargon alert. Mixture of 00:08:02.300 --> 00:08:04.100 experts. Can you break that down simply? Sure. 00:08:04.139 --> 00:08:06.759 Think of it like this. Instead of using its entire 00:08:06.759 --> 00:08:09.459 massive brain for every single question, a MoE 00:08:09.459 --> 00:08:12.480 model only activates a small, specialized part, 00:08:12.600 --> 00:08:15.019 like an expert team, that's best suited for that 00:08:15.019 --> 00:08:17.680 specific task. Ah, okay, so it's not firing up 00:08:17.680 --> 00:08:20.019 the whole engine every time. Exactly. Maybe it 00:08:20.019 --> 00:08:22.579 only uses, say, 10 billion parameters out of 00:08:22.579 --> 00:08:25.180 a much larger total for one query. That makes 00:08:25.180 --> 00:08:27.980 it way lighter, much faster, and radically more 00:08:27.980 --> 00:08:31.339 efficient than running a huge, dense model constantly. 00:08:31.639 --> 00:08:33.220 And the practical result of that efficiency? 00:08:33.639 --> 00:08:37.049 It's fast. uses less memory, and crucially, you 00:08:37.049 --> 00:08:40.750 don't need those giant, super expensive GPU clusters 00:08:40.750 --> 00:08:43.429 to run it well. That just lowers the barrier 00:08:43.429 --> 00:08:46.190 to entry massively for innovators everywhere. 00:08:46.590 --> 00:08:48.190 Okay, and how does it actually perform? You said 00:08:48.190 --> 00:08:51.210 GPT -4 level. The benchmarks was really strong. 00:08:51.350 --> 00:08:54.049 It's reportedly outperforming older Cloud 4 models 00:08:54.049 --> 00:08:56.909 in some areas, and it's nearly matching the newer 00:08:56.909 --> 00:08:59.909 Cloud 4 .5. So the big picture here is Chinese 00:08:59.909 --> 00:09:04.139 labs, Minimax, DeepSeq, Moonshot. They're putting 00:09:04.139 --> 00:09:06.440 out high quality open models, models that compete 00:09:06.440 --> 00:09:08.759 performance wise with the West. But crush them 00:09:08.759 --> 00:09:11.059 on operational costs. Yes. These are good for 00:09:11.059 --> 00:09:13.159 what kinds of applications? They're ideal if 00:09:13.159 --> 00:09:15.539 you're building, say, agent workflows or custom 00:09:15.539 --> 00:09:17.620 code co -pilots, retrieval bots, specialized 00:09:17.620 --> 00:09:20.480 tool chains, anything where cost efficiency matters, 00:09:20.639 --> 00:09:22.620 especially if you want to self -host outside 00:09:22.620 --> 00:09:26.659 the big cloud platforms. Right. So probing question 00:09:26.659 --> 00:09:31.159 then. If the running cost drops by like 90 percent. 00:09:32.090 --> 00:09:34.490 What does that cost advantage really mean for 00:09:34.490 --> 00:09:37.149 companies trying to build new agents, maybe outside 00:09:37.149 --> 00:09:40.370 the orbit of AWS or Google Cloud? Well, fundamentally, 00:09:40.490 --> 00:09:42.490 it democratizes access, right? It shifts innovation 00:09:42.490 --> 00:09:45.870 power. Yeah. Away from just the giant labs, potentially 00:09:45.870 --> 00:09:48.809 towards smaller, more agile builders anywhere 00:09:48.809 --> 00:09:51.450 in the world. So cheaper models mean faster innovation 00:09:51.450 --> 00:09:54.529 and wider access for startups. Exactly. Mid -roll 00:09:54.529 --> 00:09:57.309 sponsor, Reed Placeholder. Okay, so wrapping 00:09:57.309 --> 00:09:59.549 up this deep dive, the main threads seem to be 00:09:59.549 --> 00:10:01.769 this incredible race for physical infrastructure, 00:10:02.009 --> 00:10:04.820 this hyper -financed build -out. You know, OpenAI's 00:10:04.820 --> 00:10:07.419 one gigawatt goal really symbolizes that. And 00:10:07.419 --> 00:10:09.440 that massive spending really sets the stage for 00:10:09.440 --> 00:10:11.620 the competitive shift we're seeing. These cost 00:10:11.620 --> 00:10:13.600 -effective open models, particularly from the 00:10:13.600 --> 00:10:16.700 East, like MiniMax M2, they're completely changing 00:10:16.700 --> 00:10:18.860 the definition of high performance for developers 00:10:18.860 --> 00:10:21.240 because now it's actually affordable. It feels 00:10:21.240 --> 00:10:23.220 like AI is becoming more distributed, moving 00:10:23.220 --> 00:10:25.750 beyond just a few huge labs. I think that's right. 00:10:25.929 --> 00:10:28.230 And connecting back to the physical buildout, 00:10:28.269 --> 00:10:30.230 the infrastructure, we talked about the scale, 00:10:30.450 --> 00:10:33.950 the 1GW factory, the Amazon layoffs linked to 00:10:33.950 --> 00:10:36.830 AI spending. There's one more piece here related 00:10:36.830 --> 00:10:39.730 to secrecy. Yeah, there are reports suggesting 00:10:39.730 --> 00:10:42.730 that nondisclosure agreements, NDAs, are being 00:10:42.730 --> 00:10:45.509 used quite aggressively, specifically to keep 00:10:45.509 --> 00:10:48.490 details about these new AI data centers, the 00:10:48.490 --> 00:10:51.110 tech, the operations hidden from the public, 00:10:51.169 --> 00:10:55.299 particularly in the U .S. Given the sheer scale 00:10:55.299 --> 00:10:58.179 and speed of all this building the equivalent 00:10:58.179 --> 00:11:00.080 of a massive data center complex every single 00:11:00.080 --> 00:11:02.639 week, what exactly are we not being told about 00:11:02.639 --> 00:11:04.860 the physical foundation supporting arguably the 00:11:04.860 --> 00:11:06.820 most powerful technology we've ever built? That's 00:11:06.820 --> 00:11:08.700 the question to maybe keep mulling over. What's 00:11:08.700 --> 00:11:10.419 behind the curtain of all this infrastructure? 00:11:10.840 --> 00:11:13.059 Indeed. Something to think about. Thanks for 00:11:13.059 --> 00:11:14.899 joining us for this deep dive, El Trio Music.