WEBVTT - The Coming Generative AIPocalypse 0:00:02.800 --> 0:00:08.319 Zone Media. Hello, and welcome to Better Offline. I'm your 0:00:08.320 --> 0:00:22.759 host ed Zitron. What In the last episode, I walked 0:00:22.760 --> 0:00:25.040 you through how GENERALIFAI has hit its ceiling and how 0:00:25.040 --> 0:00:27.800 big tech has burned billions of dollars chasing a technology 0:00:28.000 --> 0:00:30.000 that really isn't going to give them the returns they 0:00:30.040 --> 0:00:33.560 got from the smartphone and cloud computing revolutions, not that 0:00:33.560 --> 0:00:34.080 that's going. 0:00:34.040 --> 0:00:35.440 To stop them. 0:00:35.680 --> 0:00:37.800 In this episode, I'm going to walk you through the 0:00:37.840 --> 0:00:41.159 potentially massive problems they have ahead, the ones they've created 0:00:41.280 --> 0:00:44.320 by building these massive supercomputers, the ones necessary to capture 0:00:44.320 --> 0:00:47.440 this imaginary demand of this kind of bullshit AI revolution, 0:00:47.920 --> 0:00:50.080 and I'm going to try and guess what might happen next. 0:00:50.720 --> 0:00:53.640 All right, So, the AI boom helped the S and 0:00:53.680 --> 0:00:56.360 P five hundred hit record levels in twenty twenty four, 0:00:56.600 --> 0:00:59.240 largely thanks to chip giant Nvidia, a company that makes 0:00:59.280 --> 0:01:01.800 both the GPUs necessary to train and run generative AI 0:01:02.080 --> 0:01:04.520 and the software architecture behind them, as well as some 0:01:04.600 --> 0:01:08.440 other server parts too. Part of Nvidia's remarkable growth has 0:01:08.440 --> 0:01:11.640 been its ability to capitalize on Kuda architecture, the software 0:01:11.680 --> 0:01:15.160 layer that lets you do complex computing on GPUs rather 0:01:15.200 --> 0:01:17.679 than simply using them to render video games or complex 0:01:17.720 --> 0:01:20.080 three D images. Someone's going to email and say that 0:01:20.080 --> 0:01:20.880 that's partially wrong. 0:01:21.000 --> 0:01:22.120 Just bear with me. 0:01:23.360 --> 0:01:25.840 And of course, one of Nvidia's other things they do 0:01:25.959 --> 0:01:28.520 is continually create new GPUs to sell for tens of 0:01:28.560 --> 0:01:30.679 thousands of dollars to tech companies that want to burn 0:01:30.760 --> 0:01:34.080 billions of dollars on GENERATIVII, which has led Nvidia's stock 0:01:34.120 --> 0:01:36.000 to pop more than one hundred and seventy nine percent 0:01:36.040 --> 0:01:39.080 over the last year. Back in May, in Vidia's CEO 0:01:39.120 --> 0:01:42.120 and professional Carnival Barker Jensen Wang said that the company 0:01:42.160 --> 0:01:44.120 was now and I quote on a one year rhythm 0:01:44.120 --> 0:01:48.120 in AIGPU production with its latest Blackwell GPUs, specifically the 0:01:48.120 --> 0:01:50.760 B one hundred, B two hundred and GBT two hundred 0:01:50.960 --> 0:01:53.360 used of Regenerative AI, supposedly due at the end of 0:01:53.400 --> 0:01:56.360 twenty twenty four, except they're now delayed until at least 0:01:56.400 --> 0:02:00.000 March twenty twenty five. Now, before we go any firm, 0:02:00.200 --> 0:02:02.400 it's worth noting that when I say GPU, I don't 0:02:02.440 --> 0:02:04.360 mean the one you'd find in the gaming PC, but 0:02:04.400 --> 0:02:06.760 a much much larger chip put on a specialized server 0:02:07.000 --> 0:02:10.239 and actually a specialized board as well. With multiple other GPUs, 0:02:10.320 --> 0:02:13.560 all integrated with special case in cooling and networking infrastructure, 0:02:13.880 --> 0:02:15.760 all of which is sold by Nvidia. By the way, 0:02:15.840 --> 0:02:18.280 you can use other things, but as well buy it 0:02:18.280 --> 0:02:21.840 in one place, right, That's their product strategy. In simple terms, 0:02:21.880 --> 0:02:23.640 these are the things necessary to make sure all these 0:02:23.720 --> 0:02:26.520 chips work together efficiently and also stop them from overheating 0:02:26.800 --> 0:02:29.880 because they get extremely hot, and generative AI makes them 0:02:29.960 --> 0:02:31.680 run at full speed all the time. 0:02:32.560 --> 0:02:33.519 Now, the initial. 0:02:33.240 --> 0:02:35.200 Delay of the new Blackwell chips is caused by a 0:02:35.240 --> 0:02:38.720 now fixed design, floor and production. But as I've previously suggested, 0:02:38.720 --> 0:02:41.760 the problem isn't just creating the chips, it's making sure 0:02:41.800 --> 0:02:44.560 they actually work at scale for the jobs they've. 0:02:44.400 --> 0:02:45.000 Been bought for. 0:02:46.840 --> 0:02:52.560 Now what if that wasn't possible a few weeks ago? 0:02:52.639 --> 0:02:55.200 The Information Report and Video is currently grappling with the 0:02:55.200 --> 0:02:58.560 oldest problem in computing, how to cool the fucking things. 0:02:59.080 --> 0:03:01.640 According to their report, and Video has been asking supplies 0:03:01.680 --> 0:03:04.800 to change the design of its three thousand pound seventy 0:03:04.840 --> 0:03:08.440 five GPU server racks several times to overcome these heating problems, 0:03:08.600 --> 0:03:11.880 which the Information calls quote the most complicated design that 0:03:11.960 --> 0:03:14.639 Nvideo had ever come up with. According to the report. 0:03:14.680 --> 0:03:17.200 A few months after revealing the racks, engineers found that 0:03:17.240 --> 0:03:21.120 they didn't work properly even when they use the smaller 0:03:21.200 --> 0:03:24.079 thirty six chip racks, and they've been kind of scrambling 0:03:24.080 --> 0:03:25.000 to fix it ever since. 0:03:25.760 --> 0:03:27.560 While one can dazzle investors with. 0:03:27.480 --> 0:03:31.640 Buzzwords, charts and the term superintelligence, the laws of physics 0:03:31.680 --> 0:03:34.240 are a little bit more of a harsh mistress. And 0:03:34.280 --> 0:03:36.560 if in Video is struggling mere months before the first 0:03:36.560 --> 0:03:39.200 installations to begin, though I've heard some are going out now, 0:03:39.680 --> 0:03:43.000 it's kind of unclear how they practically plan to launch 0:03:43.040 --> 0:03:46.320 this next generation of chips, let alone doing another one 0:03:46.320 --> 0:03:49.520 in a year. The Information reports that these changes have 0:03:49.560 --> 0:03:52.520 been made late in production, which is scaring customers that 0:03:52.560 --> 0:03:55.200 desperately need them now so that their models can continue 0:03:55.240 --> 0:03:57.280 to do something that they will work out later. 0:03:57.800 --> 0:03:59.320 To quote the Information again. 0:04:00.000 --> 0:04:02.440 Secutives at large cloud providers that have ordered the new 0:04:02.520 --> 0:04:05.760 chips said they're concerned that such last minute difficulties might 0:04:05.760 --> 0:04:07.480 push back the timeline for when they can get their 0:04:07.480 --> 0:04:11.000 GPU clusters up and running next year. The fact that 0:04:11.000 --> 0:04:14.520 in video is having such significant difficulties with thermal performance. 0:04:14.160 --> 0:04:15.320 Is very, very bad. 0:04:15.760 --> 0:04:18.800 These chips are very expensive thirty to seventy grand I 0:04:18.880 --> 0:04:21.400 hear and will be running, as I've mentioned, at full speed, 0:04:21.600 --> 0:04:24.360 generating an incredible amount of heat that must be dissipated 0:04:24.640 --> 0:04:27.200 while sat next to anywhere from thirty five to seventy 0:04:27.240 --> 0:04:30.040 one other chips, which will in turn be densely packed 0:04:30.120 --> 0:04:32.440 so that you can cram more servers into a data center. 0:04:33.000 --> 0:04:36.280 New more powerful chips require entirely new methods to rack, 0:04:36.320 --> 0:04:38.839 mount them, operate them, and cool them, and all of 0:04:38.880 --> 0:04:44.520 these parts must operate and sync as overheating GPUs will die. 0:04:44.600 --> 0:04:46.599 And while these units are big, some of their internal 0:04:46.600 --> 0:04:49.600 components and nanometers in size, and unless properly cool their 0:04:49.600 --> 0:04:51.680 circus will start to crumble when roasted by a guy 0:04:51.720 --> 0:04:54.440 typing Garfield with a guarn and a bra into chat GPT, 0:04:54.839 --> 0:04:58.360 which I have never done. Of course, Remember Blackwell is 0:04:58.360 --> 0:05:01.200 supposed to be in videos big thing. It's meant to 0:05:01.240 --> 0:05:04.000 represent a major leap forward in performance. If in Vidia 0:05:04.040 --> 0:05:06.720 doesn't solve its cooling problem, and solve it well, its 0:05:06.760 --> 0:05:10.400 customers will undoubtedly encounter thermal throttling, where the chip reduces 0:05:10.440 --> 0:05:13.240 its speed in order to avoid causing any permanent damage. 0:05:13.520 --> 0:05:17.080 It could eliminate some more all of the performance gains 0:05:17.200 --> 0:05:20.560 obtained from the new architecture a new manufacturing process, despite 0:05:20.600 --> 0:05:23.440 costing much much more, both for the chips itself and 0:05:23.480 --> 0:05:26.799 the housing for them. In Vidio's problem isn't just bringing 0:05:26.880 --> 0:05:29.680 these thermal performance issues under control, but both keeping them 0:05:29.720 --> 0:05:32.360 under control and being able to educate their customers on 0:05:32.400 --> 0:05:35.400 how to do so. In Vidia has, according to the information, 0:05:35.720 --> 0:05:39.000 repeatedly tried to influence its customer server integrations to follow 0:05:39.040 --> 0:05:42.240 its designs because it thinks it will lead to better performance. 0:05:42.279 --> 0:05:43.839 But in this case one has to worry. If in 0:05:43.920 --> 0:05:47.760 Vidia's black Well chips can be reliably cooled, maybe they 0:05:47.800 --> 0:05:50.360 can in videos worked out crazy things in the past, 0:05:50.720 --> 0:05:52.479 and while in video might be able to fix this 0:05:52.520 --> 0:05:54.919 problem in isolation within its racks, it remains to be 0:05:54.920 --> 0:05:56.599 seen how this works at scale as they ship and 0:05:56.600 --> 0:05:59.760 integrate hundreds of thousands of Blackwell GPUs starting in the 0:05:59.760 --> 0:06:03.719 front half of twenty twenty five. Things also get a 0:06:03.720 --> 0:06:06.000 little worse when you realize how these chips are being 0:06:06.040 --> 0:06:09.760 installed in these giant supercomputer data centers, where tens of 0:06:09.800 --> 0:06:12.160 thousands are as many as one hundred thousand of them. 0:06:12.160 --> 0:06:14.960 In the case of Elon Musk's Colossus Data center, which 0:06:15.000 --> 0:06:18.200 is just very funny that that's what it's for. 0:06:18.279 --> 0:06:20.120 It's for GROC on a dying social network. 0:06:20.600 --> 0:06:23.160 But yeah, these GPUs, they're running concert to power generative 0:06:23.200 --> 0:06:23.800 AI models. 0:06:23.839 --> 0:06:25.039 That's all they're really for. 0:06:25.760 --> 0:06:27.760 And The Wall Street Journal reported a few weeks ago 0:06:27.800 --> 0:06:31.440 that building these vast data centers creates entirely new engineering challenges, 0:06:31.560 --> 0:06:33.720 with one expert saying that big tech companies could be 0:06:33.800 --> 0:06:36.800 spending as much of half of their capital expenditures on 0:06:36.880 --> 0:06:39.520 replacing parts that are broken down, in large part because 0:06:39.520 --> 0:06:41.880 these clusters are running on GPUs that are running at 0:06:41.880 --> 0:06:44.240 full speed that need to all be cooled and need 0:06:44.279 --> 0:06:44.880 to be cooled. 0:06:44.880 --> 0:06:45.159 Well. 0:06:46.040 --> 0:06:49.159 Remember the capital expenditures on generative AI and the associated 0:06:49.200 --> 0:06:52.120 infrastructure have gone over two hundred billion dollars in the 0:06:52.160 --> 0:06:55.520 last year. If half of that's dedicated to replacing broken shit. 0:06:55.560 --> 0:06:58.719 What happens when there's no path of profitability in any 0:06:58.760 --> 0:07:02.000 case is fine, they'll work it out well enough to 0:07:02.040 --> 0:07:04.320 keep flogging these things. They've already made billions of dollars 0:07:04.400 --> 0:07:06.919 selling Blackwells. They're sold out for a year, in fact, 0:07:07.000 --> 0:07:09.960 and they'll continue to do so for now. But any 0:07:10.000 --> 0:07:12.679 manufacturing and cooling issues are going to be very costly, 0:07:13.320 --> 0:07:15.480 and even then, at some point somebody has to ask 0:07:15.480 --> 0:07:18.720 the question, why do we need all these GPUs if 0:07:18.760 --> 0:07:22.560 we've reached PKI. Despite the remarkable power of these chips. 0:07:22.640 --> 0:07:25.600 In Vidia's entire enterprise GPU business model centers around the 0:07:25.640 --> 0:07:28.200 idea that throwing more power at these problems will finally 0:07:28.200 --> 0:07:29.280 create some solutions. 0:07:30.400 --> 0:07:31.880 What if that isn't the case? 0:07:32.000 --> 0:07:37.720 Though? So as ever, I'm kind of pissed off. I'm 0:07:37.760 --> 0:07:40.080 pissed off reading this stuff. And it's not just because 0:07:40.080 --> 0:07:42.600 I was saying this months ago. It's because all of 0:07:42.600 --> 0:07:44.560 this money could have gone literally anywhere else. They could 0:07:44.560 --> 0:07:47.000 have done something else, they could have I don't know, 0:07:47.480 --> 0:07:49.520 I'm not running a big tech company, but I would 0:07:49.560 --> 0:07:52.160 have put it into climate stuff or new battery technology. 0:07:52.680 --> 0:07:55.040 Everyone's saying, oh, we'll make our own chips now, Yes, 0:07:55.080 --> 0:07:57.440 see you in a few years, dickhead, Well, you go 0:07:57.520 --> 0:08:00.000 to TSMC and ask if they've got any spare slot. 0:08:00.120 --> 0:08:02.760 I'm sure they've got plenty. I'm sure that you can 0:08:02.840 --> 0:08:05.200 just go and build a chip tomorrow. Sam Moultman talking 0:08:05.200 --> 0:08:09.880 about building his own goddamn chips. What an absolute fast? 0:08:09.920 --> 0:08:10.280 This is? 0:08:10.520 --> 0:08:13.960 This fucking fast? All of this money burned in search 0:08:14.000 --> 0:08:16.360 of a technology that only kind of works on something. 0:08:16.520 --> 0:08:20.200 Sometimes it's just kind of a joke. The tech industry 0:08:20.240 --> 0:08:23.040 is over leverage. They've doubled, they've tripled the kadrupled down 0:08:23.040 --> 0:08:25.680 on generative AI now and this technology does not do 0:08:25.880 --> 0:08:27.720 much more than it did a few months ago, And 0:08:27.760 --> 0:08:29.400 I don't think it's going to do very much more 0:08:29.440 --> 0:08:32.760 in a few months. Every single big tech companies pulled 0:08:32.760 --> 0:08:35.520 tens of billions of dollars into building out these massive 0:08:35.559 --> 0:08:39.719 superclusters with the intent of capturing AI demand. Yet they 0:08:39.760 --> 0:08:42.280 never seem to think whether they were actually building things 0:08:42.280 --> 0:08:45.040 that people wanted or would pay for, or that would 0:08:45.040 --> 0:08:47.960 make the money, or that would help you manity in 0:08:48.000 --> 0:08:50.760 any way, or really anything to do with an outcome 0:08:50.800 --> 0:08:53.960 other than line go up. While some have claimed that 0:08:54.000 --> 0:08:58.120 agents are the next frontier AI, Agents frontier Jesus, Nope, 0:08:58.240 --> 0:09:00.439 keep him, The reality is that agents may be the 0:09:00.559 --> 0:09:04.199 last generative AI product. Multiple large language models and integrations 0:09:04.240 --> 0:09:06.520 bouncing off of each other in attempt to simulate what 0:09:06.559 --> 0:09:08.480 a human might do at a cost that won't be 0:09:08.480 --> 0:09:12.040 sustainable for the majority of businesses. While Anthropics demo of 0:09:12.080 --> 0:09:15.160 its alleged model that can control a few browser windows 0:09:15.160 --> 0:09:17.840 with a prompt might have seemed impressive to credulous people, 0:09:18.520 --> 0:09:22.160 specifically mean casing you. And here these were controlled demos 0:09:22.160 --> 0:09:25.440 which Anthropic added with slow and make lots of mistakes. Hey, 0:09:25.559 --> 0:09:28.079 almost like it's hallucinating. I sure hope they fixed that 0:09:28.200 --> 0:09:30.360 unfixable problem I'd be talking about for months. 0:09:30.440 --> 0:09:31.080 Jesus Christ. 0:09:32.360 --> 0:09:36.120 Even if it does, Anthropic has now successfully replaced an 0:09:36.240 --> 0:09:40.079 entry full data worker position at an indeterminate cost, and 0:09:40.280 --> 0:09:43.320 likely an unprofitable one too. And in many organizations those 0:09:43.400 --> 0:09:45.840 jobs have already been outsourced or automated or start with 0:09:45.920 --> 0:09:48.320 cheap contractors. So the people who are going to lose 0:09:48.360 --> 0:09:50.520 out here are people who are already getting shot on 0:09:50.559 --> 0:09:51.120 by life. 0:09:51.320 --> 0:09:52.160 It really sucks. 0:09:52.679 --> 0:09:55.880 And the obscenity of this mass delusion, it's just nauseating. 0:09:56.080 --> 0:09:58.160 It's a monolith to bad decision making and the herd 0:09:58.160 --> 0:10:01.440 mentality of tech's most powerful people, as well as an 0:10:01.480 --> 0:10:04.160 outright attempt to manipulate the media into believing something was 0:10:04.200 --> 0:10:07.280 possible that never was, and the media bought it, hook 0:10:07.360 --> 0:10:08.240 line and sinker. 0:10:09.120 --> 0:10:09.920 You all bought it. 0:10:10.920 --> 0:10:13.480 Hundreds of billions of dollars have been wasted building giant 0:10:13.559 --> 0:10:16.240 data centers to crunch numbers for software that has no 0:10:16.320 --> 0:10:18.640 real product market fit, or while trying to hammer it 0:10:18.640 --> 0:10:21.760 into various shapes to make it pretend that it's alive, conscious, 0:10:21.840 --> 0:10:25.360 or even fucking useful. There is no path, from what 0:10:25.440 --> 0:10:27.880 I can see, to turning generative AI from what it 0:10:27.920 --> 0:10:31.520 is today into anything resembling a sustainable business. And the 0:10:31.559 --> 0:10:34.760 only path that big tech appeared to have seen or 0:10:34.840 --> 0:10:37.240 thought about was to throw as much money, power and 0:10:37.320 --> 0:10:40.320 data at the problem. An avenue that appeared to be 0:10:40.440 --> 0:10:42.880 and definitely appears to be now a complete dead end, 0:10:43.640 --> 0:10:47.120 and we're still nothing's really come out of this movement. 0:10:47.960 --> 0:10:50.640 I've used a handful of AI products that I found useful, 0:10:50.720 --> 0:10:54.320 and aipower journal, for example, but these are not products 0:10:54.320 --> 0:10:57.560 that one associates with revolutions. They're useful tools that would 0:10:57.600 --> 0:11:00.000 have been a welcome surprised if it didn't require burning 0:11:00.120 --> 0:11:03.080 billions of dollars, blowing past emissions targets, and stealing the 0:11:03.080 --> 0:11:06.640 creative works of millions of people to train them. It's 0:11:06.679 --> 0:11:10.360 just sickening. It's sickening because this did not have to happen. 0:11:10.520 --> 0:11:13.320 Had suned up Ashay seen what Microsoft did with open 0:11:13.360 --> 0:11:16.520 AI and said not for me, mate, That's definitely not 0:11:16.559 --> 0:11:19.440 how he talks. But nevertheless, none of this would have continued. 0:11:19.559 --> 0:11:21.360 You think Microsoft was going to hold this shit up 0:11:21.360 --> 0:11:24.880 on their own. God, No, they have no juice such 0:11:24.920 --> 0:11:27.640 an Adela. He's just a he's a cult leader, something 0:11:27.679 --> 0:11:30.400 I'm going to get into in a later episode. And 0:11:30.440 --> 0:11:33.160 it's just it's all very sad. But if you want 0:11:33.160 --> 0:11:35.520 to cheer yourself up, maybe you could listen to one 0:11:35.520 --> 0:11:38.200 of the following advertisements. And this is a short episode, 0:11:38.240 --> 0:11:39.920 so if they don't put one after this, you might 0:11:39.960 --> 0:11:43.560 just have a short gap between me talking. That's up 0:11:43.600 --> 0:11:50.800 to Matt. MAT's a genius. That's my producer, by the way, Mattasowski. Anyway, ads, 0:11:59.160 --> 0:12:04.079 and we're back. Look, I truly don't know what happens next, 0:12:04.160 --> 0:12:05.920 but I'm going to walk you through what I'm thinking. 0:12:06.640 --> 0:12:08.480 If we're truly at the diminishing. 0:12:08.040 --> 0:12:10.720 Return stage of transformer based models, it's going to be 0:12:10.760 --> 0:12:14.800 extremely difficult to justify buying further iterations of Nvidia GPUs 0:12:14.840 --> 0:12:19.199 past Blackwell, and also I have doubts they're going to 0:12:19.240 --> 0:12:21.719 be able to make a new one every year. The 0:12:21.920 --> 0:12:24.679 entire generative AI movement lives and dies by the idea 0:12:24.720 --> 0:12:27.079 that more compute power and more training data makes these 0:12:27.120 --> 0:12:29.720 things better. And if that's no longer the case, there's 0:12:29.960 --> 0:12:32.120 not really a reason to keep buying bigger and better. 0:12:32.640 --> 0:12:36.760 What's the point even now? What exactly happens when Microsoft 0:12:36.880 --> 0:12:39.600 or Google has racks worth of Blackwell GPUs? What does 0:12:39.600 --> 0:12:43.640 aren't getting better faster? I guess what does faster even mean? 0:12:43.760 --> 0:12:46.480 What does better even mean? Better has yet to me 0:12:46.520 --> 0:12:49.800 in a new product? Better has not seem to be 0:12:50.200 --> 0:12:54.680 measurable by humans just by bullshit benchmarks. This also makes 0:12:54.720 --> 0:12:58.120 the lives of open Ai and Anthropic a little more difficult. 0:12:58.440 --> 0:13:00.760 Sam Altman has grown rich in power for lying about 0:13:00.800 --> 0:13:04.480 how GPT will somehow lead to AGI some conscious computer. 0:13:04.920 --> 0:13:09.080 But at this point, what exactly is open ai meant 0:13:09.080 --> 0:13:11.720 to do? What are they doing? The only way it's 0:13:12.160 --> 0:13:14.480 this company has ever been able to develop new models 0:13:14.559 --> 0:13:17.079 is by throwing masses of compute and training data at them. 0:13:17.400 --> 0:13:19.839 And there are only other choices to start stapling its 0:13:20.040 --> 0:13:23.120 their reasoning model onto the side of their large language model, 0:13:23.240 --> 0:13:29.200 at which point something happens something everyone It's happening something 0:13:29.400 --> 0:13:31.880 that is so good that literally nobody working for open 0:13:31.920 --> 0:13:33.719 aire in the media appears to be able to tell 0:13:33.720 --> 0:13:34.200 you what it is. 0:13:35.160 --> 0:13:36.679 Yeah. 0:13:36.760 --> 0:13:39.680 Putting that aside, open ai is also a terrible business 0:13:40.160 --> 0:13:42.520 that has to burn five billion dollars to make three 0:13:42.520 --> 0:13:45.040 point seven billion dollars, and there's no proof that they're 0:13:45.080 --> 0:13:47.840 able to bring down their costs. The constant thing I 0:13:47.840 --> 0:13:50.720 hear from VC's and AI fantasist is that the chips 0:13:50.720 --> 0:13:53.600 will bring down the cost of inference. Yet I don't 0:13:53.640 --> 0:13:57.400 see that happening. I've yet to see that happening. There's 0:13:57.440 --> 0:14:00.040 cerebrus dinner plate sized chips, but you know what, I 0:14:00.160 --> 0:14:02.520 just don't think that shit's going to scale. And at 0:14:02.559 --> 0:14:05.400 that point you have to wonder do they have like 0:14:05.640 --> 0:14:08.880 where is this going? I know you're listening to an 0:14:08.920 --> 0:14:12.160 episode where I tell you, but still even reading this 0:14:12.240 --> 0:14:14.800 stuff all day, you just look at it too hard 0:14:15.040 --> 0:14:18.920 and you start feeling a little crazy. I've read pretty 0:14:18.960 --> 0:14:21.680 much every article on open ai at this point. I've 0:14:21.680 --> 0:14:24.880 read Microsoft's earnings multiple quarters straight. Same with the rest 0:14:24.880 --> 0:14:27.000 of them. I've read all their blog posts, I've read 0:14:27.000 --> 0:14:30.359 all their shit. They don't have any idea. I'm confident 0:14:30.440 --> 0:14:33.480 that they don't have any idea. It's terrifying and this 0:14:33.600 --> 0:14:36.880 is just this dismal situation where the only other there's 0:14:36.920 --> 0:14:40.000 really only a few options. If I'm honest, you stop now. 0:14:40.720 --> 0:14:45.160 That is the biggest one. You just stop. Someone has 0:14:45.200 --> 0:14:48.200 to admit, Microsoft, Google, They'll never do this. They need 0:14:48.240 --> 0:14:50.640 to go this is not going to scale. We need 0:14:50.680 --> 0:14:52.560 to pair this back or we need to stop it. 0:14:53.240 --> 0:14:55.800 And it will hurt their stock, it will hurt their media. 0:14:56.040 --> 0:14:58.480 But it's the right thing to do for the environment alone, 0:14:58.760 --> 0:15:03.360 but as a sustainable busines. But assuming they don't do that, 0:15:03.880 --> 0:15:08.040 we have another problem, which is the only other option 0:15:08.160 --> 0:15:10.880 they can have is to keep flooring the gas. It 0:15:10.920 --> 0:15:13.800 costs one hundred million dollars to train GPT four to zero, 0:15:14.000 --> 0:15:17.000 and anthropic CEO Dario ami Day estimated a few months 0:15:17.040 --> 0:15:20.000 ago that training future models would cost one billion, if 0:15:20.040 --> 0:15:23.080 not ten billion dollars, with one researcher claiming that training 0:15:23.160 --> 0:15:26.760 open AI's next model, GPT five will cost around a 0:15:26.760 --> 0:15:29.320 billion dollars outside of a miracle. 0:15:29.320 --> 0:15:32.200 We're about to enter an era of desperation in generative AI. 0:15:32.640 --> 0:15:34.760 We're two years in and we have no killer apps, 0:15:34.760 --> 0:15:38.080 no industry defining products other than chat GPT, a product 0:15:38.120 --> 0:15:40.520 that burns billions of doddlers, and nobody. 0:15:40.200 --> 0:15:41.920 Can really describe every episode. 0:15:41.960 --> 0:15:44.600 I'm like, email me, tell me what chat GPT is, 0:15:44.800 --> 0:15:48.880 and I'll get some smart us who sends me three paragraphs. No, no, no, no, no, 0:15:49.200 --> 0:15:51.680 that's not a description, that's an essay. May you actually 0:15:51.680 --> 0:15:53.320 need to tell me what the hell this is? And 0:15:53.400 --> 0:15:55.720 if you can't, you've proven my point in my extremely 0:15:55.760 --> 0:16:00.280 thin and rigged game. Anyway, neither Microsoft, nor Meta, nor Google, 0:16:00.320 --> 0:16:02.000 no Amazon seemed to be able to come up with 0:16:02.080 --> 0:16:05.280 any profitable use cases, let alone one their users actually like, 0:16:05.560 --> 0:16:07.560 Nor have any of the people that have raised billions 0:16:07.560 --> 0:16:09.720 of dollars in venture capital for anything with AI take 0:16:09.800 --> 0:16:13.240 to the side. An investor interest in AI is already cooling. 0:16:13.280 --> 0:16:17.280 According to the information, It's kind of unclear how far 0:16:17.400 --> 0:16:20.040 this fast goes from here, if only because it isn't 0:16:20.080 --> 0:16:22.480 obvious what it is that anybody gets by investing in 0:16:22.520 --> 0:16:25.560 future rounds of open Aianthropic or any of these other 0:16:25.600 --> 0:16:29.000 generative AI companies. At some point, they must make money, 0:16:29.080 --> 0:16:31.440 and the entire dream has been built around the idea 0:16:31.480 --> 0:16:33.200 that all of these GPUs and all of this money 0:16:33.400 --> 0:16:38.960 would eventually spit out something, something revolutionary, something world changing. 0:16:40.200 --> 0:16:43.280 What we have is this clunky, ugly, messy, lasceness, environmentally 0:16:43.280 --> 0:16:45.080 destructive and mediocre. 0:16:44.560 --> 0:16:45.280 Piece of shit. 0:16:46.240 --> 0:16:48.760 Generative AI was a reckless pursuit, one that shows a 0:16:48.760 --> 0:16:50.960 total lack of creativity and sense in the minds of 0:16:51.000 --> 0:16:53.200 big tech in venture capital, one where there was never 0:16:53.320 --> 0:16:55.920 anything really impressive other than the amount of money it 0:16:55.920 --> 0:16:57.800 could burn on the amount of time samal And could 0:16:57.800 --> 0:17:00.800 say something stupid and get quoted for it. I'll be 0:17:00.800 --> 0:17:03.920 honest with you, I don't really know what happens here. 0:17:04.400 --> 0:17:06.840 The future was always one that demanded big tech spend 0:17:06.840 --> 0:17:08.600 more to make even bigger models that would at some 0:17:08.600 --> 0:17:11.840 point become useful, and it isn't happening. In pursuit of 0:17:11.880 --> 0:17:14.359 doing so, big tech invested hundreds of billions of dollars 0:17:14.400 --> 0:17:17.920 into infrastructure specifically to follow one goal and put AI 0:17:18.040 --> 0:17:20.280 front and center of their businesses, claiming it was the 0:17:20.280 --> 0:17:24.280 future without ever checking if it was, if it actually 0:17:24.320 --> 0:17:28.600 did anything useful. And really they, as I've said in 0:17:28.600 --> 0:17:30.440 the rock com bubble a few months ago, they don't 0:17:30.520 --> 0:17:33.679 have anything else. They don't have another growth market, and 0:17:33.720 --> 0:17:35.800 maybe that's why they're shoving all the cash into this, 0:17:36.119 --> 0:17:37.000 because they don't. 0:17:36.760 --> 0:17:38.159 Have it anywhere. Else to put it. 0:17:38.359 --> 0:17:40.760 They don't want to admit they've got nothing else. They 0:17:40.760 --> 0:17:43.080 don't want to admit that their businesses are crumbling under 0:17:43.480 --> 0:17:47.520 scrutiny from antitrust. They don't want to admit their services 0:17:47.560 --> 0:17:50.680 are kind of deteriorating, especially in the case of Meta. 0:17:50.800 --> 0:17:53.760 They don't have anything else. The revenue isn't coming, the 0:17:53.800 --> 0:17:58.040 products aren't coming. Oryan open AI's next model, it will underwhelm, 0:17:58.119 --> 0:18:00.119 as will its competitors model, and at some point, think 0:18:00.200 --> 0:18:02.560 somebody is going to blink in one of the hyperscalers 0:18:02.560 --> 0:18:06.399 and the AI era will end. Almost every single GENERAI 0:18:06.520 --> 0:18:09.040 company that you've heard of is deeply unprofitable, and there 0:18:09.040 --> 0:18:11.639 are very few innovations coming to save them from the 0:18:11.680 --> 0:18:14.960 atrophy of the Foundation model system, which by the way, 0:18:15.400 --> 0:18:17.760 is the main large language models that they're training in 0:18:17.800 --> 0:18:21.960 all of these cases. Now, I've tried to keep my 0:18:22.000 --> 0:18:24.400 emotions in check, but you might you might be able 0:18:24.400 --> 0:18:26.760 to tell that this pisses me off a little. I 0:18:26.840 --> 0:18:30.040 just feel sad and exhausted about it all. I feel 0:18:30.160 --> 0:18:32.000 drained as I look at how many times I've tried 0:18:32.000 --> 0:18:34.399 to warm people. I'm pissed off at the many members 0:18:34.400 --> 0:18:36.000 of the media that failed to push back against the 0:18:36.040 --> 0:18:38.760 over promises and outright lies of people like Sam Altman 0:18:38.840 --> 0:18:41.600 and Dario Ami Day. Warrio ami Day is what I'm 0:18:41.640 --> 0:18:44.159 calling him now. It's not particularly funny ira acurate, but 0:18:44.160 --> 0:18:46.320 it's funny to say for me. And I'm just full 0:18:46.359 --> 0:18:48.679 of dread as I consider the economic ramifications of this 0:18:48.720 --> 0:18:51.280 industry collapsing, as well as the damage it's already done 0:18:51.400 --> 0:18:54.720 to our fucking environment. Once the AI bubble pops, there 0:18:54.760 --> 0:18:57.720 are no hyper growth markets left, which will in turn 0:18:57.800 --> 0:18:59.679 lead to a blood bath in big tech stocks as 0:18:59.680 --> 0:19:02.000 they really that they're out of big ideas to convince 0:19:02.040 --> 0:19:05.000 the street that they're going to grow forever. There are 0:19:05.040 --> 0:19:07.360 some that will boast about being right here, and yes, 0:19:07.400 --> 0:19:10.840 there is some satisfaction in being so, of course, of 0:19:10.880 --> 0:19:12.960 course you take the moment, you do the victory lap. 0:19:13.040 --> 0:19:16.159 Whatever I said a thing, and I was right. I 0:19:16.160 --> 0:19:18.560 don't know if I'm I'm really doing the joker kick 0:19:18.640 --> 0:19:20.720 down the stairs here. I don't know if I'm really 0:19:21.280 --> 0:19:24.560 happy about being right here. I think generative AI is 0:19:24.600 --> 0:19:26.439 a huge waste of money. I think the damage it's 0:19:26.480 --> 0:19:29.800 doing to the environment is disgusting. I think the stealing 0:19:30.000 --> 0:19:33.280 required to make the training data side is disgusting. I 0:19:33.280 --> 0:19:35.880