1 00:00:02,160 --> 00:00:05,560 Speaker 1: Zone Media. Well, I know that assholes grow on trees, 2 00:00:05,559 --> 00:00:07,800 Speaker 1: but I'm here to trim the leaves. This is Better 3 00:00:07,880 --> 00:00:22,800 Speaker 1: off Line. I'm your host ed dait tron. It's bullshit 4 00:00:22,840 --> 00:00:25,800 Speaker 1: Week here at Better Offline headquarters. The thousands of elves 5 00:00:25,840 --> 00:00:28,560 Speaker 1: that work for me have been finding me things that 6 00:00:28,640 --> 00:00:30,680 Speaker 1: say bullshit for no reason. When I wrote the episode 7 00:00:30,720 --> 00:00:33,040 Speaker 1: weeks ago. I don't know what to do with them. 8 00:00:33,080 --> 00:00:35,639 Speaker 1: But anyway, we're starting with one of the largest financial 9 00:00:35,640 --> 00:00:40,120 Speaker 1: institutions of the world calling bullshit on Generative AI as usual. 10 00:00:40,400 --> 00:00:42,920 Speaker 1: Don't take my word for it. Check the episode notes 11 00:00:42,920 --> 00:00:45,519 Speaker 1: for links that map to everything I'm talking about. I've 12 00:00:45,520 --> 00:00:47,440 Speaker 1: tried my best to map them to exactly what I'm 13 00:00:47,479 --> 00:00:49,240 Speaker 1: saying too, and you can feel free to yell about 14 00:00:49,240 --> 00:00:51,080 Speaker 1: it like yell out the words like the Beastie Boys, 15 00:00:51,440 --> 00:00:53,640 Speaker 1: or just follow along, because I think it's important for 16 00:00:53,680 --> 00:00:55,760 Speaker 1: you to know where I'm getting all of this from, 17 00:00:55,920 --> 00:00:59,400 Speaker 1: versus just assuming that I made it up somehow. That 18 00:00:59,480 --> 00:01:02,120 Speaker 1: is an accus I've had made. I don't really know 19 00:01:02,680 --> 00:01:05,679 Speaker 1: how I would do that anyway, the episode I'm very sorry. 20 00:01:06,720 --> 00:01:09,240 Speaker 1: At the tail end of June, Goldman Sachs, one of 21 00:01:09,240 --> 00:01:12,320 Speaker 1: the largest global investment banks put out the thirty one 22 00:01:12,360 --> 00:01:16,399 Speaker 1: page report titled JENAI too much, spend too little benefit 23 00:01:16,800 --> 00:01:19,040 Speaker 1: that's with the question mark at the end. That includes 24 00:01:19,080 --> 00:01:21,400 Speaker 1: some of the most damning literature on generative AI that 25 00:01:21,480 --> 00:01:25,399 Speaker 1: I've ever seen. And yeah, that weep sound in the background. 26 00:01:25,480 --> 00:01:28,160 Speaker 1: You here is the slow, painful deflation of the bubble 27 00:01:28,200 --> 00:01:31,839 Speaker 1: I've been warning you about since March. The report covers 28 00:01:31,840 --> 00:01:36,600 Speaker 1: AI's productivity benefits, which Goldman remarks are likely limited, AI's returns, 29 00:01:36,640 --> 00:01:39,600 Speaker 1: which are likely to be significantly more limited than anticipated, 30 00:01:39,800 --> 00:01:43,080 Speaker 1: and AI's power demands, which are likely so significant that 31 00:01:43,200 --> 00:01:46,160 Speaker 1: utility companies will have to spend nearly forty percent more 32 00:01:46,360 --> 00:01:48,120 Speaker 1: in the next three years to keep up with the 33 00:01:48,160 --> 00:01:51,760 Speaker 1: demand from hyperscalers and rot economists like Google and Microsoft. 34 00:01:53,240 --> 00:01:56,280 Speaker 1: The report is so significant because Goldman sacks, like any 35 00:01:56,320 --> 00:02:00,000 Speaker 1: investment bank doesn't care about your feelings or your happy 36 00:02:00,360 --> 00:02:04,640 Speaker 1: or emotions or anything unless doing so is profitable. It'll 37 00:02:04,640 --> 00:02:07,120 Speaker 1: gladly hypob anything it thinks it will make a bug. 38 00:02:07,680 --> 00:02:10,120 Speaker 1: Back in May, it published a report which claimed that AI, 39 00:02:10,360 --> 00:02:13,600 Speaker 1: not just generative AI, was showing very positive signs of 40 00:02:13,639 --> 00:02:18,000 Speaker 1: eventually boosting GDP and productivity. Even though said report buried 41 00:02:18,000 --> 00:02:21,240 Speaker 1: within it, constant reminders that AI had yet to impact 42 00:02:21,280 --> 00:02:24,720 Speaker 1: productivity growth and states that only five percent of companies 43 00:02:24,919 --> 00:02:29,040 Speaker 1: report using generative AI in regular production. For Goldman to 44 00:02:29,120 --> 00:02:32,799 Speaker 1: suddenly turn on the AI movement suggests that it's extremely 45 00:02:32,880 --> 00:02:36,399 Speaker 1: anxious about the future of generative AI, with almost everybody 46 00:02:36,440 --> 00:02:39,560 Speaker 1: agreeing on one core point in the report that the 47 00:02:39,639 --> 00:02:42,960 Speaker 1: longer the generative AI takes to make people money, the 48 00:02:43,000 --> 00:02:46,160 Speaker 1: more money that it's going to need to make. The 49 00:02:46,240 --> 00:02:49,760 Speaker 1: report also includes an interview with economist Darren Smoglu of MIT, 50 00:02:50,160 --> 00:02:51,760 Speaker 1: which can be found by the way on page four 51 00:02:51,760 --> 00:02:53,560 Speaker 1: of the document, which you can find in this episode 52 00:02:53,639 --> 00:02:57,240 Speaker 1: spreadsheet of links. An Institute professor who published a paperback 53 00:02:57,280 --> 00:03:00,600 Speaker 1: in May called the Simple Macroeconomics of AI that argued 54 00:03:00,639 --> 00:03:03,880 Speaker 1: that and I quote the upside to US productivity and 55 00:03:04,080 --> 00:03:07,880 Speaker 1: consequently GDP growth from genera IVAI will likely prove much 56 00:03:07,919 --> 00:03:11,680 Speaker 1: more limited than many forecasts that expect a month has 57 00:03:11,760 --> 00:03:16,200 Speaker 1: only made Asmoglue more pessimistic, declaring that truly transformative changes 58 00:03:16,240 --> 00:03:19,800 Speaker 1: won't happen quickly and few, if any, will likely occur 59 00:03:19,919 --> 00:03:23,440 Speaker 1: within the next ten years, and that genera ivai's ability 60 00:03:23,480 --> 00:03:27,440 Speaker 1: to affect global productivity is low because and I quote again, 61 00:03:27,800 --> 00:03:31,440 Speaker 1: many of the tasks that humans currently perform are multifaceted 62 00:03:31,480 --> 00:03:34,920 Speaker 1: and require real world interaction, which AI won't be able 63 00:03:34,960 --> 00:03:39,720 Speaker 1: to materially improve anytime soon. What makes this interview, and 64 00:03:39,960 --> 00:03:44,000 Speaker 1: really the whole report so remarkable is how thoroughly and 65 00:03:44,040 --> 00:03:47,920 Speaker 1: aggressively it attacks every bit of marketing collateral that the 66 00:03:48,040 --> 00:03:54,400 Speaker 1: artificial intelligence movement has. Asmoglu specifically questions the belief that 67 00:03:54,440 --> 00:03:57,440 Speaker 1: AI models will simply get more powerful as we throw 68 00:03:57,520 --> 00:04:00,920 Speaker 1: more data and GPU graphics process unit the thing that 69 00:04:01,080 --> 00:04:05,000 Speaker 1: crunches the numbers capacity at them, and specifically asks a question, 70 00:04:06,000 --> 00:04:10,040 Speaker 1: what does it mean to double AI's capabilities? How does 71 00:04:10,080 --> 00:04:15,320 Speaker 1: that make something like say a customer service rap better? Seriously, though, 72 00:04:15,840 --> 00:04:19,359 Speaker 1: what does better really get them less errors? How do 73 00:04:19,400 --> 00:04:23,159 Speaker 1: you quantify less errors without factoring in the current state 74 00:04:23,440 --> 00:04:27,159 Speaker 1: introducing them? It's not great And this really is a 75 00:04:27,200 --> 00:04:32,680 Speaker 1: specific problem with the whole GENII fantasists bullshit spiel. They 76 00:04:32,720 --> 00:04:35,000 Speaker 1: heavily rely on the idea that not only will these 77 00:04:35,080 --> 00:04:38,880 Speaker 1: large language models llms like chat GPT get more powerful, 78 00:04:39,200 --> 00:04:41,919 Speaker 1: but the getting more powerful will somehow grant it the 79 00:04:41,920 --> 00:04:47,520 Speaker 1: power to do something. As Asimoglu says, what does it 80 00:04:47,600 --> 00:04:52,120 Speaker 1: mean the double it What does that do? No? Really, 81 00:04:52,160 --> 00:04:55,480 Speaker 1: what does that more actually mean? We've heard people talk 82 00:04:55,560 --> 00:04:59,480 Speaker 1: about this for eighteen months, saying the more powerful chat 83 00:04:59,520 --> 00:05:02,880 Speaker 1: GBT gets, the better it will get. But what does 84 00:05:02,960 --> 00:05:06,760 Speaker 1: better even look like? GPT four Oh, the latest version 85 00:05:06,800 --> 00:05:10,839 Speaker 1: of chat GPT, other than accepting more inputs, is kind 86 00:05:10,880 --> 00:05:14,360 Speaker 1: of the same thing. The capabilities have not really changed, 87 00:05:15,839 --> 00:05:19,680 Speaker 1: and while one might argue that more powerful will mean 88 00:05:19,880 --> 00:05:23,960 Speaker 1: faster generative processes, there really is no barometer for what 89 00:05:24,040 --> 00:05:27,960 Speaker 1: better looks like, and perhaps that's why Chat GBT, Claude 90 00:05:27,960 --> 00:05:30,159 Speaker 1: and other lms have yet to take a leap beyond 91 00:05:30,160 --> 00:05:33,160 Speaker 1: being able to generate pictures of Garfield and Lingerie are 92 00:05:33,360 --> 00:05:37,359 Speaker 1: Scooby Doo with a gun anthropics. Claude LM might be 93 00:05:37,520 --> 00:05:40,040 Speaker 1: and I quote, best in class according to tech Crunch, 94 00:05:40,279 --> 00:05:42,800 Speaker 1: but that only means that it's faster and more accurate, 95 00:05:42,839 --> 00:05:47,600 Speaker 1: which is cool, but not really the future or revolutionary 96 00:05:47,680 --> 00:05:52,080 Speaker 1: or necessarily good in all cases. I should add that 97 00:05:52,120 --> 00:05:54,880 Speaker 1: these are questions that I and other people writing about 98 00:05:54,960 --> 00:05:57,560 Speaker 1: AI kind of should have been asking the whole time. 99 00:05:58,560 --> 00:06:02,480 Speaker 1: General FAI generates outputs based on text based inputs and requests, 100 00:06:02,480 --> 00:06:05,240 Speaker 1: and eventually multimodal will mean be able to look at 101 00:06:05,320 --> 00:06:08,320 Speaker 1: something and generate an answer too, and these requests can 102 00:06:08,360 --> 00:06:11,640 Speaker 1: be equally specific and intricate. Yet the answer is always 103 00:06:11,680 --> 00:06:16,039 Speaker 1: as obvious as it sounds generated fresh, meaning that there's 104 00:06:16,080 --> 00:06:19,839 Speaker 1: no actual knowledge or indeed intelligence operating in any part 105 00:06:19,960 --> 00:06:23,440 Speaker 1: of the process. As a result, it's easy to see 106 00:06:23,440 --> 00:06:30,320 Speaker 1: how this gets better ish faster, but much much harder to, 107 00:06:30,520 --> 00:06:33,440 Speaker 1: if not impossible, to see how generative AI leads any 108 00:06:33,560 --> 00:06:37,200 Speaker 1: further than where we're already at. And by that, I 109 00:06:37,240 --> 00:06:40,960 Speaker 1: mean what does it do other than what it's doing today? 110 00:06:40,960 --> 00:06:44,839 Speaker 1: What does chat GBT do today? That's radically different to 111 00:06:44,880 --> 00:06:50,200 Speaker 1: eighteen months ago. The jump between say, two generations of iPhones, 112 00:06:50,200 --> 00:06:52,600 Speaker 1: from the first iPhone to what it would be the 113 00:06:52,640 --> 00:06:55,159 Speaker 1: iPhone three g I think it would be someone's going 114 00:06:55,240 --> 00:06:57,480 Speaker 1: to email and say I'm wrong, and I'll say to them, 115 00:06:57,520 --> 00:07:00,320 Speaker 1: you're very rude anyway, But that jump was huge, which 116 00:07:00,480 --> 00:07:02,960 Speaker 1: faster internet meant you could do more with the phone. 117 00:07:03,040 --> 00:07:05,880 Speaker 1: There was the app store, These were new functionalities added 118 00:07:06,640 --> 00:07:09,400 Speaker 1: two years after the first iPhone, two years after the 119 00:07:09,400 --> 00:07:13,360 Speaker 1: first chat GBT, and uh, I guess, I guess you 120 00:07:13,360 --> 00:07:16,160 Speaker 1: could talk to it in a response that's not really 121 00:07:16,480 --> 00:07:20,840 Speaker 1: great How does GPT, a transformer based model that generates 122 00:07:20,880 --> 00:07:23,520 Speaker 1: answers probabilistically, as in, what the next part of the 123 00:07:23,560 --> 00:07:26,440 Speaker 1: generation is most likely to be based on what's been inputed, 124 00:07:27,160 --> 00:07:30,920 Speaker 1: based entirely on training data, How does it do anything 125 00:07:31,000 --> 00:07:35,080 Speaker 1: more than generate paragraphs of occasionally accurate text or images, 126 00:07:35,760 --> 00:07:38,520 Speaker 1: maybe a picture of Scooby Doo in lingerie. I don't know. 127 00:07:38,800 --> 00:07:41,960 Speaker 1: But how do any of these models even differentiate from 128 00:07:42,000 --> 00:07:43,760 Speaker 1: each other when most of them are trained on the 129 00:07:43,800 --> 00:07:47,720 Speaker 1: same training data that they're already running out of? And 130 00:07:47,760 --> 00:07:50,680 Speaker 1: what's crazy? Is I mentioned that I should have asked this, 131 00:07:50,760 --> 00:07:53,040 Speaker 1: I really should have every single episode of mentioned AI. 132 00:07:53,120 --> 00:07:55,360 Speaker 1: What does better even look like? What does more powerful 133 00:07:55,400 --> 00:07:59,760 Speaker 1: even look like? Because when we say better, does that 134 00:08:00,000 --> 00:08:04,400 Speaker 1: mean faster? Does that mean quicker to generate things? I 135 00:08:04,400 --> 00:08:06,400 Speaker 1: mean it can generate more things? And the answer is 136 00:08:06,480 --> 00:08:09,240 Speaker 1: usually not. It's not that it can actually function in 137 00:08:09,240 --> 00:08:11,960 Speaker 1: a different way. It's just that you can do more, 138 00:08:12,200 --> 00:08:15,040 Speaker 1: it can grow more. Hey, remember the idea of growth 139 00:08:15,200 --> 00:08:20,360 Speaker 1: or cost Jesus right, But seriously, though, this feels like 140 00:08:20,400 --> 00:08:23,480 Speaker 1: the question to ask Sam Ultiman or Mirror Marati the 141 00:08:23,600 --> 00:08:27,920 Speaker 1: CTO of open AI's like, what's next? What's next, Oh, 142 00:08:27,960 --> 00:08:30,120 Speaker 1: you can generate videos? When can I use that? Cool? 143 00:08:30,320 --> 00:08:34,920 Speaker 1: What does better look like? They're other than not horrible looking? Okay, 144 00:08:35,000 --> 00:08:38,480 Speaker 1: but what what more can GPT do? Because this thing, 145 00:08:38,840 --> 00:08:42,360 Speaker 1: this thing can't think it's generating stuff, It doesn't know anything. 146 00:08:42,679 --> 00:08:45,480 Speaker 1: It has training data. It eats up and then craps 147 00:08:45,520 --> 00:08:48,400 Speaker 1: out an answer. How is that going to lead to 148 00:08:48,480 --> 00:08:52,880 Speaker 1: even automation? I just and then how do you even 149 00:08:52,960 --> 00:08:55,520 Speaker 1: deal with the fact that it really is running out 150 00:08:55,559 --> 00:08:58,640 Speaker 1: of training data? And I've argued before the training data 151 00:08:58,679 --> 00:09:00,920 Speaker 1: crisis is one that does not get enough attention, But 152 00:09:00,960 --> 00:09:03,760 Speaker 1: it's sufficiently dire now that it has the potential to 153 00:09:03,800 --> 00:09:07,000 Speaker 1: halt or at least dramatically slow any AI development in 154 00:09:07,000 --> 00:09:10,480 Speaker 1: the future, by which I mean generative AI. As one 155 00:09:10,520 --> 00:09:13,079 Speaker 1: paper published in the Journal of Computer Vision and Pattern 156 00:09:13,160 --> 00:09:16,160 Speaker 1: Recognition found, in order to achieve a linear improvement in 157 00:09:16,480 --> 00:09:20,520 Speaker 1: model improvements, you need an exponentially large amount of data. 158 00:09:21,000 --> 00:09:24,520 Speaker 1: So it's not like they read something and then work 159 00:09:24,600 --> 00:09:27,160 Speaker 1: something out from there. They need so much more to 160 00:09:27,280 --> 00:09:31,439 Speaker 1: learn one thing kind of right, it's not very good 161 00:09:31,960 --> 00:09:34,920 Speaker 1: and maybe there's another way to put it. Each additional 162 00:09:35,000 --> 00:09:38,680 Speaker 1: step of training becomes increasingly and exponentially more expensive to 163 00:09:38,679 --> 00:09:41,800 Speaker 1: take two. It requires you to get a bunch of 164 00:09:41,880 --> 00:09:44,839 Speaker 1: training data and process more of it, but it also 165 00:09:44,880 --> 00:09:50,000 Speaker 1: requires extremely expensive technology GPUs and a ton of energy 166 00:09:50,040 --> 00:09:53,800 Speaker 1: to actually do so, and this infers a steep financial cost, 167 00:09:53,920 --> 00:09:56,439 Speaker 1: not merely in just obtaining the data, but also, as 168 00:09:56,480 --> 00:09:59,760 Speaker 1: I mentioned, the compute required to process it. And Anthropics 169 00:10:00,000 --> 00:10:03,000 Speaker 1: CEO Dario Amadel said that the AI models currently in 170 00:10:03,040 --> 00:10:05,760 Speaker 1: development will cost as much as a billion dollars to train, 171 00:10:06,040 --> 00:10:08,560 Speaker 1: and within the next three years we may see models 172 00:10:08,679 --> 00:10:12,600 Speaker 1: that cost ten or one hundred billion dollars to train, 173 00:10:12,679 --> 00:10:15,640 Speaker 1: which is just insane, as that's roughly three times the 174 00:10:15,679 --> 00:10:21,640 Speaker 1: GDP of Estonia, a beautiful, quite cold country. Back to 175 00:10:21,720 --> 00:10:25,080 Speaker 1: mister Darren Smoblo, who doubts that llms can even become 176 00:10:25,200 --> 00:10:28,880 Speaker 1: super intelligent and that even his most conservative estimates of 177 00:10:28,920 --> 00:10:32,000 Speaker 1: productivity gains and I quote may turn out to be 178 00:10:32,080 --> 00:10:35,079 Speaker 1: too large if AI models prove less successful in improving 179 00:10:35,160 --> 00:10:38,920 Speaker 1: upon more complex tasks, and I think that's really the 180 00:10:39,000 --> 00:10:43,480 Speaker 1: root of the problem. All of this excitement, every second 181 00:10:43,520 --> 00:10:46,000 Speaker 1: of breathless beat off hype has been built on this 182 00:10:46,160 --> 00:10:50,440 Speaker 1: idea that the artificial intelligence industry, led by General AVAI, 183 00:10:50,760 --> 00:10:54,640 Speaker 1: will somehow revolutionize and automate everything from robotics to the 184 00:10:54,679 --> 00:10:57,599 Speaker 1: supply chain, despite the fact that general IVAI is not 185 00:10:57,760 --> 00:11:00,480 Speaker 1: actually going to solve these problems because it is not 186 00:11:00,640 --> 00:11:03,800 Speaker 1: built to do so. I willn't have a calculator to 187 00:11:03,880 --> 00:11:07,240 Speaker 1: drive my fucking car Jesus Christ. And While Asamoglu may 188 00:11:07,240 --> 00:11:09,800 Speaker 1: have some positive things to say, for example, that AI 189 00:11:09,840 --> 00:11:12,240 Speaker 1: models could be trained to help scientists conceive of and 190 00:11:12,280 --> 00:11:15,920 Speaker 1: test new materials, which actually already happened thanks to Google 191 00:11:15,960 --> 00:11:20,560 Speaker 1: DeepMind researchers, his general verdict is kind of harsh that 192 00:11:20,679 --> 00:11:24,000 Speaker 1: using generative AI and quote too much automation too soon 193 00:11:24,280 --> 00:11:26,760 Speaker 1: could create bottlenecks and other problems for firms that no 194 00:11:26,840 --> 00:11:30,480 Speaker 1: longer have the flexibility in troubleshooting capabilities that human capital provides. 195 00:11:31,080 --> 00:11:34,520 Speaker 1: In essence, replacing humans with AI might break everything. If 196 00:11:34,520 --> 00:11:36,920 Speaker 1: you're one of those bosses that doesn't actually know what 197 00:11:36,960 --> 00:11:40,440 Speaker 1: the fuck it is they're talking about, But you know what, 198 00:11:41,040 --> 00:11:44,680 Speaker 1: I'm sure the following advertisements from people who know exactly 199 00:11:44,840 --> 00:11:55,640 Speaker 1: what they're talking about, and we're back. The report also 200 00:11:55,679 --> 00:11:58,640 Speaker 1: includes a palate cleanser for the quirked up AI hype fiends, 201 00:11:59,000 --> 00:12:01,720 Speaker 1: and you'll find it on page where Goldman sax As. 202 00:12:01,800 --> 00:12:04,840 Speaker 1: Joseph Briggs argues that generative AI will and I quote 203 00:12:05,120 --> 00:12:08,120 Speaker 1: likely lead to significant economic upside based and I shit 204 00:12:08,240 --> 00:12:11,440 Speaker 1: you not entirely on the idea that AI will replace 205 00:12:11,520 --> 00:12:13,600 Speaker 1: workers in some jobs and then allow them to get 206 00:12:13,679 --> 00:12:17,000 Speaker 1: jobs in other fields. Briggs also argues that, and I 207 00:12:17,080 --> 00:12:21,280 Speaker 1: quote again, the full AI automation of AI exposed tasks 208 00:12:21,320 --> 00:12:23,800 Speaker 1: that are likely to occur over a longer horizon could 209 00:12:23,840 --> 00:12:28,040 Speaker 1: generate significant cost savings, which assumes that generative AI or 210 00:12:28,040 --> 00:12:30,920 Speaker 1: AI itself will actually replace these tasks. But also, that's 211 00:12:30,960 --> 00:12:34,560 Speaker 1: such a funny thing to say, Hey, you know, auto 212 00:12:34,600 --> 00:12:39,280 Speaker 1: mating tasks over a long time could save money. Fucking hell, 213 00:12:39,320 --> 00:12:42,200 Speaker 1: I should get a job at Goldmann Sex anyway, anyway. Sorry. 214 00:12:42,440 --> 00:12:45,120 Speaker 1: I should also add that, unlike every single other interview 215 00:12:45,160 --> 00:12:49,199 Speaker 1: in the report, Briggs continually mixes up AI and generative AI, 216 00:12:49,480 --> 00:12:52,480 Speaker 1: and at one point suggests that recent generative AI advances 217 00:12:52,480 --> 00:12:57,800 Speaker 1: are foreshadowing the emergence of a superintelligence. This is like 218 00:12:57,880 --> 00:13:01,480 Speaker 1: suggesting that because I squeeze my washing up liquid and 219 00:13:01,480 --> 00:13:03,920 Speaker 1: it's sprayed across the wall, that I'm one step close 220 00:13:03,960 --> 00:13:08,160 Speaker 1: to becoming Picasso. I included this part of the report 221 00:13:08,160 --> 00:13:11,880 Speaker 1: because sometimes, very rarely I get somebody suggesting that I'm 222 00:13:11,920 --> 00:13:15,520 Speaker 1: not considering both sides. The reason I don't generally include 223 00:13:15,520 --> 00:13:17,760 Speaker 1: both sides of this argument is that the AI hype 224 00:13:17,760 --> 00:13:20,760 Speaker 1: side generally makes arguments based on the assumption that things 225 00:13:20,800 --> 00:13:24,800 Speaker 1: will happen, such as a transformer model that probabilistically generates 226 00:13:24,840 --> 00:13:27,640 Speaker 1: the next part of a sentence or a picture somehow 227 00:13:27,679 --> 00:13:33,320 Speaker 1: gaining sentience. I wish my calculator gained sentience. During high school, 228 00:13:33,360 --> 00:13:35,040 Speaker 1: I was a very lonely child. I did not have 229 00:13:35,120 --> 00:13:37,040 Speaker 1: many friends. But no, all it would do is just 230 00:13:37,080 --> 00:13:39,160 Speaker 1: let me kind of make it look like it said boobs. 231 00:13:39,360 --> 00:13:41,600 Speaker 1: Then I'd get in a lot of trouble. But anyway. 232 00:13:42,440 --> 00:13:45,760 Speaker 1: Francois Chole, an AI researcher at Google, recently argued that 233 00:13:45,880 --> 00:13:49,680 Speaker 1: large language models like CHET GPT can't lead to average 234 00:13:49,720 --> 00:13:54,720 Speaker 1: general intelligence, the sentient AI that everyone's excited about, explaining 235 00:13:54,760 --> 00:13:57,240 Speaker 1: in detail in an interview with a podcast that I've 236 00:13:57,280 --> 00:14:00,640 Speaker 1: linked in there, that models like GPT A simply not 237 00:14:00,760 --> 00:14:03,880 Speaker 1: capable of the kind of reasoning and theorizing that makes 238 00:14:03,880 --> 00:14:08,360 Speaker 1: a human brain work. Schley also argues that even models 239 00:14:08,400 --> 00:14:11,520 Speaker 1: specifically built to complete the tasks of his abstraction and 240 00:14:11,559 --> 00:14:14,120 Speaker 1: reasoning corpus a benchmark test for AI skills and true 241 00:14:14,120 --> 00:14:16,920 Speaker 1: intelligence that he invented are only doing so because they've 242 00:14:16,960 --> 00:14:19,800 Speaker 1: been fed millions of data points of people solving the test, 243 00:14:20,640 --> 00:14:23,200 Speaker 1: which is kind of like measuring somebody's IQ based on 244 00:14:23,240 --> 00:14:27,040 Speaker 1: them studying really hard to complete an IQ test except 245 00:14:27,040 --> 00:14:30,200 Speaker 1: even dumber. But the reason that I'm suddenly bringing up 246 00:14:30,240 --> 00:14:35,120 Speaker 1: these superintelligences or AGI artificial general intelligence average general intelligence, 247 00:14:35,360 --> 00:14:38,280 Speaker 1: lots of people say different things. The reason I'm saying 248 00:14:38,280 --> 00:14:41,160 Speaker 1: it is because throughout every single defense of generative AI 249 00:14:41,640 --> 00:14:44,640 Speaker 1: is a really nasty, deliberate attempt to get around the 250 00:14:44,720 --> 00:14:49,720 Speaker 1: problem that GENERATIVAI doesn't actually automate many tasks. While it's 251 00:14:49,720 --> 00:14:53,240 Speaker 1: good at generating answers, sometimes even correct ones, or at 252 00:14:53,280 --> 00:14:56,000 Speaker 1: creating things based on a request, sometimes with the right 253 00:14:56,080 --> 00:14:59,440 Speaker 1: number of fingers, there's no real interaction with the task 254 00:14:59,560 --> 00:15:02,640 Speaker 1: or the perse and giving the task, or considering what 255 00:15:02,760 --> 00:15:05,320 Speaker 1: the task needs at all. Just the abstraction of things 256 00:15:05,360 --> 00:15:08,600 Speaker 1: said to output generated orbit in quite a complex way. 257 00:15:09,400 --> 00:15:11,760 Speaker 1: Tasks like taking someone's order and relaying it to the 258 00:15:11,840 --> 00:15:14,680 Speaker 1: kitchen at a fast food restaurant might seem elementary to 259 00:15:14,720 --> 00:15:17,200 Speaker 1: most people and I won't write easy, because working in 260 00:15:17,240 --> 00:15:20,480 Speaker 1: fast food is a hard and horrible job. It might 261 00:15:20,560 --> 00:15:23,560 Speaker 1: seem elementary, though, but it isn't for an AI model 262 00:15:23,560 --> 00:15:26,480 Speaker 1: that generates answers without really understanding the meaning of any 263 00:15:26,560 --> 00:15:28,040 Speaker 1: of the words. And really, I shouldn't have said the 264 00:15:28,040 --> 00:15:32,000 Speaker 1: word really that it doesn't understand anything. Last year, Wendy's, 265 00:15:32,000 --> 00:15:34,520 Speaker 1: a burger chain here in America, announced that it would 266 00:15:34,520 --> 00:15:38,200 Speaker 1: integrate its generative Fresh AI ordering system into some restaurants. 267 00:15:38,840 --> 00:15:41,520 Speaker 1: In late June, it revealed that the system requires human 268 00:15:41,520 --> 00:15:45,680 Speaker 1: intervention on fourteen percent of the orders. On one redditor's post, 269 00:15:45,920 --> 00:15:49,160 Speaker 1: they noted that Wendy's AI regularly required three attempts to 270 00:15:49,160 --> 00:15:51,280 Speaker 1: get it to understand them, and would sometimes cut you 271 00:15:51,320 --> 00:15:54,400 Speaker 1: off if you weren't speaking fast enough. Hey, it's two, 272 00:15:54,440 --> 00:15:58,120 Speaker 1: am I connecting the words chicken sandwich, the diet codes 273 00:15:58,200 --> 00:16:02,840 Speaker 1: difficult for me. Another burger place here in America, White Castle, 274 00:16:03,520 --> 00:16:07,040 Speaker 1: which implemented a similar system in partnership with Samsung and SoundHound, 275 00:16:07,400 --> 00:16:09,960 Speaker 1: fared a little better, with a remarkable ten percent of 276 00:16:10,040 --> 00:16:14,600 Speaker 1: orders requiring human intervention. Last month, McDonald's discontinued its own 277 00:16:14,680 --> 00:16:17,280 Speaker 1: AI ordering system, which it built with IBM and deployed 278 00:16:17,520 --> 00:16:21,080 Speaker 1: to more than one hundred restaurants, likely because it wasn't 279 00:16:21,200 --> 00:16:25,000 Speaker 1: very good, with one customer rang up for literally hundreds 280 00:16:25,000 --> 00:16:28,240 Speaker 1: of chicken nuggets like the I think you should leave sketch. However, 281 00:16:28,280 --> 00:16:31,560 Speaker 1: to be clear, McDonald's system wasn't actually based on GENERATIVAI, 282 00:16:32,160 --> 00:16:34,840 Speaker 1: and if nothing else, all of these examples illustrate this 283 00:16:34,880 --> 00:16:38,840 Speaker 1: disconnect between those building AI systems and how much, or 284 00:16:39,000 --> 00:16:42,160 Speaker 1: really how little they understand the jobs they wish to eliminate. 285 00:16:43,000 --> 00:16:47,400 Speaker 1: Little humility or doing a real job goes a long way. 286 00:16:48,240 --> 00:16:50,560 Speaker 1: Another thing to note is that, on top of generative 287 00:16:50,600 --> 00:16:54,000 Speaker 1: AI generally cocking up these orders, Whendy still requires human 288 00:16:54,000 --> 00:16:56,800 Speaker 1: beings to make the fucking food. Despite all of this hype, 289 00:16:56,880 --> 00:17:00,000 Speaker 1: all of this media attention, all of this incredible investment, 290 00:17:00,200 --> 00:17:03,560 Speaker 1: the supposed innovations don't even seem capable of replacing the 291 00:17:03,640 --> 00:17:06,439 Speaker 1: jobs that all of these horny capitalists have been planning 292 00:17:06,480 --> 00:17:09,440 Speaker 1: for them to do. Not that they think they should, 293 00:17:09,480 --> 00:17:12,760 Speaker 1: just I'm being told that this future is inevitable or indeed, 294 00:17:12,920 --> 00:17:16,720 Speaker 1: here it's starting to really accept what the problem is 295 00:17:16,760 --> 00:17:20,280 Speaker 1: with generative AI, though it isn't good at replacing the 296 00:17:20,359 --> 00:17:23,960 Speaker 1: kind of jobs that actually affect the economy but commoditizing 297 00:17:24,040 --> 00:17:27,480 Speaker 1: distinct acts of labor, and in the process the early 298 00:17:27,520 --> 00:17:30,359 Speaker 1: creative jobs that help people build portfolios to advance in 299 00:17:30,359 --> 00:17:35,080 Speaker 1: their industries. The freelancers having their livelihoods replaced by bosses 300 00:17:35,160 --> 00:17:38,680 Speaker 1: using generative AI aren't being replaced so much as they're 301 00:17:38,720 --> 00:17:41,800 Speaker 1: being shown how little respect many bosses have for their 302 00:17:41,800 --> 00:17:46,000 Speaker 1: craft or the customer they allegedly serve. Copy editors and 303 00:17:46,040 --> 00:17:49,120 Speaker 1: concept artists provide far more valuable work than any generative 304 00:17:49,119 --> 00:17:52,680 Speaker 1: AI can yeat. An economy dominated by managers who don't 305 00:17:52,680 --> 00:17:55,840 Speaker 1: appreciate or participate in labor means that these jobs are 306 00:17:55,840 --> 00:17:59,119 Speaker 1: constantly under assault from large language models pumping out stuff 307 00:17:59,160 --> 00:18:01,240 Speaker 1: that all looks and it sounds the same, to the 308 00:18:01,240 --> 00:18:03,680 Speaker 1: point that the BBC reports that copywriters are now being 309 00:18:03,680 --> 00:18:06,479 Speaker 1: paid to help them make the ais sound more human. 310 00:18:08,440 --> 00:18:12,160 Speaker 1: One of the most fundamental misunderstandings of the bosses replacing 311 00:18:12,160 --> 00:18:14,560 Speaker 1: these workers with generative AI is that you are not 312 00:18:14,800 --> 00:18:17,760 Speaker 1: just asking for a thing, but outsourcing the risk and 313 00:18:17,800 --> 00:18:21,639 Speaker 1: responsibility for delivering it. When I hire an artist to 314 00:18:21,640 --> 00:18:24,159 Speaker 1: make a logo, my expectation is that they'll listen to 315 00:18:24,200 --> 00:18:26,280 Speaker 1: me theyn add their own flare then will go back 316 00:18:26,320 --> 00:18:28,639 Speaker 1: and forth with drafts until we have something I like 317 00:18:28,680 --> 00:18:31,919 Speaker 1: and that they're proud of too. I'm paying them not 318 00:18:32,040 --> 00:18:35,400 Speaker 1: just for their time, but for their years learning their 319 00:18:35,440 --> 00:18:39,080 Speaker 1: craft and the output itself, and so the ultimate burden 320 00:18:39,160 --> 00:18:42,200 Speaker 1: of production is not just my own, and that their 321 00:18:42,240 --> 00:18:45,040 Speaker 1: experience means that they can adapt to circumstances that I 322 00:18:45,160 --> 00:18:48,359 Speaker 1: might not have thought of. These are not things that 323 00:18:48,400 --> 00:18:50,679 Speaker 1: you can train in a data set, because they're derived 324 00:18:50,680 --> 00:18:55,240 Speaker 1: from experiences inside and outside of the creative process. While 325 00:18:55,240 --> 00:18:57,840 Speaker 1: one can teach a generative AI what a billion images 326 00:18:57,880 --> 00:19:00,800 Speaker 1: look like, AI doesn't get handcrams or a call at 327 00:19:00,800 --> 00:19:04,159 Speaker 1: eight pm saying that something needs to pop more. It 328 00:19:04,160 --> 00:19:06,879 Speaker 1: doesn't have moods, nor can it infer them from written 329 00:19:06,920 --> 00:19:10,760 Speaker 1: or visual media, because human emotions are extremely weird, as 330 00:19:10,760 --> 00:19:14,640 Speaker 1: our moods, our bodies, and our general existences. We're disgusting 331 00:19:14,680 --> 00:19:18,399 Speaker 1: and weird and beautiful. And I realize all of this 332 00:19:18,440 --> 00:19:21,520 Speaker 1: is a little flowery, but even the most mediocre copy 333 00:19:21,800 --> 00:19:25,040 Speaker 1: ever written is on some level a collection of experiences, 334 00:19:25,640 --> 00:19:29,199 Speaker 1: and fully replacing any creative is so very unlikely if 335 00:19:29,240 --> 00:19:31,399 Speaker 1: you're doing so based on copying a million pieces of 336 00:19:31,440 --> 00:19:35,800 Speaker 1: someone else's homework. Now, if I was looking for something creative, 337 00:19:36,359 --> 00:19:39,160 Speaker 1: I get it from one of the following advertisements, which 338 00:19:39,200 --> 00:19:41,880 Speaker 1: I am sure aligned perfectly with the things that I'm 339 00:19:41,920 --> 00:19:44,800 Speaker 1: talking about. Won't make me look weird, won't have people 340 00:19:44,880 --> 00:19:48,760 Speaker 1: email me. No, you're going to love the advertisements, and 341 00:19:48,800 --> 00:20:01,360 Speaker 1: then you're not going to email me about them. All Right, 342 00:20:01,400 --> 00:20:04,879 Speaker 1: we're back in the room. The most fascinating part of 343 00:20:04,920 --> 00:20:07,360 Speaker 1: the Goldman Sex Report, and you'll find it on page ten, 344 00:20:07,720 --> 00:20:10,120 Speaker 1: is an interview with Jim Cavello Goldman sax Is head 345 00:20:10,119 --> 00:20:13,280 Speaker 1: of Global Equity Research. Cavello isn't a name you'll have 346 00:20:13,320 --> 00:20:16,560 Speaker 1: heard unless you are, for whatever reason, a big semiconductor head, 347 00:20:16,960 --> 00:20:19,439 Speaker 1: but he's consistently been on the right side of history, 348 00:20:19,640 --> 00:20:22,440 Speaker 1: named as one of the top semiconductor analysts by LI 349 00:20:22,680 --> 00:20:26,320 Speaker 1: Research for years, successfully catching the downturn in fundamentals in 350 00:20:26,400 --> 00:20:29,800 Speaker 1: multiple major chip firms far before others did, at times 351 00:20:29,880 --> 00:20:33,000 Speaker 1: mocked for quote being wrong and then turning out to 352 00:20:33,040 --> 00:20:38,359 Speaker 1: be so very very right. And Jim, Jim Cavello, in 353 00:20:38,440 --> 00:20:42,840 Speaker 1: no uncertain terms, thinks that GENERATIVEAI the whole bubble, this 354 00:20:42,960 --> 00:20:48,840 Speaker 1: whole generative monstrosity kind of for shit. Cavello believes that 355 00:20:48,880 --> 00:20:51,879 Speaker 1: the combined expenditure of all parts of the GENERATIVEAI boom, 356 00:20:52,040 --> 00:20:56,560 Speaker 1: data centers, utilities, applications will cost a trillion dollars in 357 00:20:56,600 --> 00:20:59,480 Speaker 1: the next several years alone, and he asks one very 358 00:20:59,520 --> 00:21:03,399 Speaker 1: simple quest question, what trillion dollar problem will AI solve? 359 00:21:04,200 --> 00:21:07,520 Speaker 1: He notes that replacing low wage jobs with tremendously costly 360 00:21:07,680 --> 00:21:11,119 Speaker 1: technology is basically the polar opposite of the prior technology 361 00:21:11,119 --> 00:21:14,679 Speaker 1: transitions that he's witnessed in the last thirty years. Just 362 00:21:14,720 --> 00:21:17,880 Speaker 1: to be consistent with my AI to national GDP rubric, 363 00:21:18,160 --> 00:21:20,959 Speaker 1: one trillion dollars is roughly half the GDP of Italy 364 00:21:21,200 --> 00:21:23,720 Speaker 1: or the entire Danish economy multiplied by two and a 365 00:21:23,760 --> 00:21:27,880 Speaker 1: half times. Yes, Denmark an EU and NATO member state 366 00:21:27,920 --> 00:21:30,600 Speaker 1: that gave the world, among other things, a zempic, Lego 367 00:21:30,640 --> 00:21:33,760 Speaker 1: and Aqua, the artist behind nineteen ninety seven Saccharin song 368 00:21:33,800 --> 00:21:38,600 Speaker 1: of the Year Barbie Girl, one amazing country, and apparently 369 00:21:38,640 --> 00:21:42,199 Speaker 1: their GDP is just a footnote in comparison to how 370 00:21:42,280 --> 00:21:46,000 Speaker 1: much we have to spend on generating pictures of well 371 00:21:46,359 --> 00:21:49,000 Speaker 1: Garfield in lingerie shooting at Scooby Doo. At this point, 372 00:21:49,080 --> 00:21:52,280 Speaker 1: I'm just going to keep escalating one particular myth that 373 00:21:52,320 --> 00:21:55,760 Speaker 1: Cavello dispels, and this was my favorite part is that 374 00:21:55,800 --> 00:21:58,760 Speaker 1: many people compare generative AI to the early days of 375 00:21:58,760 --> 00:22:00,760 Speaker 1: the Internet, and he knows so. It's that, even if 376 00:22:00,800 --> 00:22:03,440 Speaker 1: it's in its infancy, the Internet was a low cost 377 00:22:03,440 --> 00:22:06,600 Speaker 1: technology solution that enabled things like e commas to replace 378 00:22:06,800 --> 00:22:11,400 Speaker 1: costly incumbent solutions, and that AI technology is exceptionally expensive, 379 00:22:11,440 --> 00:22:14,560 Speaker 1: and to justify those costs, the technology must be able 380 00:22:14,600 --> 00:22:17,360 Speaker 1: to solve complex problems, which it isn't designed to do. 381 00:22:17,440 --> 00:22:19,840 Speaker 1: And that is a quote, by the way, hello's a beast. 382 00:22:20,560 --> 00:22:23,600 Speaker 1: He also dismisses the suggestion that tech starts off expensive 383 00:22:23,640 --> 00:22:26,440 Speaker 1: and gets cheaper over time as revisionist history, and that's 384 00:22:26,480 --> 00:22:29,280 Speaker 1: a quote, and that quote again, the tech world is 385 00:22:29,320 --> 00:22:32,199 Speaker 1: too complacent in the assumption that AI costs will decline 386 00:22:32,240 --> 00:22:36,200 Speaker 1: substantially over time. He specifically notes that the only reason 387 00:22:36,240 --> 00:22:39,280 Speaker 1: that Moore's law was capable of enabling smaller, faster, jeeper 388 00:22:39,320 --> 00:22:43,080 Speaker 1: chips was because competitors like AMD forced Intel and other 389 00:22:43,160 --> 00:22:46,560 Speaker 1: companies to compete, a thing that doesn't really seem to 390 00:22:46,600 --> 00:22:49,359 Speaker 1: be happening with Nvidia, which is a near stranglehold on 391 00:22:49,400 --> 00:22:53,639 Speaker 1: GPUs required to handle generative AI. And indeed, all of 392 00:22:53,640 --> 00:22:57,280 Speaker 1: these other companies aren't really making competitive products. They're just 393 00:22:57,359 --> 00:23:00,360 Speaker 1: kind of making the same thing. Google has their own 394 00:23:00,600 --> 00:23:05,080 Speaker 1: video and text and image generator, sodas Lama, the Metawe, 395 00:23:05,280 --> 00:23:09,199 Speaker 1: sodas chat GPT. It's all the same. There's no competition here. 396 00:23:10,119 --> 00:23:13,440 Speaker 1: Kind of look a a cartail. Almost good idea for 397 00:23:13,480 --> 00:23:16,880 Speaker 1: an episode. And while there are companies making these graphics 398 00:23:16,920 --> 00:23:19,960 Speaker 1: process units aimed at the AI market, especially in China, 399 00:23:20,000 --> 00:23:23,160 Speaker 1: where US trade restrictions prevent local companies from buying high 400 00:23:23,200 --> 00:23:27,040 Speaker 1: powered cards like the A one hundred and video makes 401 00:23:27,040 --> 00:23:29,160 Speaker 1: for fears that they'll be diverted to the Chinese military, 402 00:23:29,359 --> 00:23:31,680 Speaker 1: they're not doing so at the same scale as in video, 403 00:23:31,800 --> 00:23:34,480 Speaker 1: and Cavello notes that and I quote, the market is 404 00:23:34,520 --> 00:23:38,560 Speaker 1: too complacent about the certainty of cost declients. He also 405 00:23:38,640 --> 00:23:41,160 Speaker 1: notes that costs are so high that even if they 406 00:23:41,200 --> 00:23:43,760 Speaker 1: were to come down, they'd have to do so dramatically, 407 00:23:44,119 --> 00:23:46,280 Speaker 1: and that the comparison to the early days of the Internet, 408 00:23:46,520 --> 00:23:49,280 Speaker 1: where businesses often relied on sixty four thousand dollars service 409 00:23:49,320 --> 00:23:52,560 Speaker 1: from Sun Microsystems and there was no live cloud stories 410 00:23:52,680 --> 00:23:56,840 Speaker 1: like Amazon Web Services or linod or Azure, those days 411 00:23:56,880 --> 00:24:00,000 Speaker 1: paled in comparison to the current costs of generative AI, 412 00:24:00,520 --> 00:24:03,199 Speaker 1: and that's even before you include the replacement of the 413 00:24:03,240 --> 00:24:05,960 Speaker 1: power grid, which he says is a necessity to keep 414 00:24:05,960 --> 00:24:08,720 Speaker 1: the boom going. I could probably just read you the 415 00:24:08,880 --> 00:24:13,320 Speaker 1: entirety of Cavello's interview because it's just so nasty. I 416 00:24:13,400 --> 00:24:16,520 Speaker 1: want to roll around on the floor on it. It's amazing 417 00:24:17,040 --> 00:24:20,359 Speaker 1: and he's so good, and he even attacks one of 418 00:24:20,359 --> 00:24:25,040 Speaker 1: my least favorite things, where he believes that when people say, oh, 419 00:24:25,080 --> 00:24:27,200 Speaker 1: people didn't really think the iPhone was a big deal, 420 00:24:27,280 --> 00:24:30,080 Speaker 1: the Internet was a big deal. Put specifically the iPhone 421 00:24:30,080 --> 00:24:32,000 Speaker 1: and smartphones, they didn't think it was going to be big, 422 00:24:32,000 --> 00:24:34,400 Speaker 1: and thus generative will be big. And he just says 423 00:24:34,400 --> 00:24:38,120 Speaker 1: it's complete nonsense. He says that he sat through hundreds 424 00:24:38,119 --> 00:24:40,640 Speaker 1: of presentations in the early two thousands, many of them 425 00:24:40,640 --> 00:24:44,680 Speaker 1: including road maps that accurately fear how smartphones eventually rolled out, 426 00:24:44,800 --> 00:24:49,480 Speaker 1: specifically noting things like GPS. So when GPS technology came down, 427 00:24:49,520 --> 00:24:51,440 Speaker 1: of course that would be in a smartphone. That makes 428 00:24:51,440 --> 00:24:54,600 Speaker 1: perfect sense. Just to be clear, this guy is there 429 00:24:54,640 --> 00:24:56,880 Speaker 1: and he's a software and hardware analyst. He sits through 430 00:24:56,920 --> 00:24:59,200 Speaker 1: presentations about what the future might look like all the time, 431 00:25:00,240 --> 00:25:02,560 Speaker 1: and he said there's no such roadmap for generative AI, 432 00:25:03,680 --> 00:25:07,040 Speaker 1: and there's no killer app either. He also notes the 433 00:25:07,040 --> 00:25:09,680 Speaker 1: big tech companies now have no choice but to engage 434 00:25:09,720 --> 00:25:12,640 Speaker 1: in an artificial intelligence ARM race given the hype, which 435 00:25:12,680 --> 00:25:16,080 Speaker 1: will continue the trend of massive spending. And he believes 436 00:25:16,080 --> 00:25:19,040 Speaker 1: that there are and I quote low odds of AI 437 00:25:19,160 --> 00:25:22,159 Speaker 1: related revenue expansion, in part because he doesn't believe that 438 00:25:22,200 --> 00:25:25,199 Speaker 1: generative AAI will make workers smarter, just more capable of 439 00:25:25,240 --> 00:25:29,760 Speaker 1: finding better information faster, which is not great, by the way, 440 00:25:30,520 --> 00:25:33,199 Speaker 1: and that any advantages that generative AI gives can be 441 00:25:33,320 --> 00:25:35,720 Speaker 1: arbitraged the way because the tech can be used everywhere, 442 00:25:35,720 --> 00:25:39,399 Speaker 1: and thus you can't as a company really raise prices. 443 00:25:39,760 --> 00:25:42,360 Speaker 1: In fact, and bridging off from what he said there, 444 00:25:43,359 --> 00:25:47,000 Speaker 1: what might happen is a race to the bottom, or 445 00:25:47,040 --> 00:25:49,320 Speaker 1: maybe when things get too expensive, everyone will get more 446 00:25:49,320 --> 00:25:52,879 Speaker 1: expensive too. None of this is good, by the way 447 00:25:53,320 --> 00:25:58,120 Speaker 1: for them, but in plain English, just saying it, I'm 448 00:25:58,119 --> 00:25:59,960 Speaker 1: just going to put it out there, you'll be supper. 449 00:26:00,280 --> 00:26:03,200 Speaker 1: But what I'm about to say, generative AI isn't making 450 00:26:03,200 --> 00:26:06,040 Speaker 1: any money for anybody because it doesn't actually make companies 451 00:26:06,040 --> 00:26:09,399 Speaker 1: that use it any extra money. Efficiency is useful, but 452 00:26:09,480 --> 00:26:13,800 Speaker 1: it's not company defining. Cavello also adds that hyperscalers like 453 00:26:13,840 --> 00:26:18,119 Speaker 1: Google and Microsoft will also garner incremental revenue from AI, 454 00:26:18,560 --> 00:26:21,200 Speaker 1: not the huge returns that they're perhaps counting on given 455 00:26:21,240 --> 00:26:24,119 Speaker 1: their vast AI related expenditures over the last two years 456 00:26:24,400 --> 00:26:27,360 Speaker 1: and the ridiculous things they've been doing, like putting generative 457 00:26:27,400 --> 00:26:31,520 Speaker 1: AI in search. Jesus Christ Sundar. And this is damning 458 00:26:31,560 --> 00:26:34,000 Speaker 1: for many reasons, chief of which is that the biggest 459 00:26:34,080 --> 00:26:37,240 Speaker 1: thing that artificial intelligence is meant to do is be 460 00:26:37,480 --> 00:26:41,680 Speaker 1: smart and make you smarter. Being able to access information 461 00:26:41,800 --> 00:26:44,200 Speaker 1: faster might make you better at your job, but that's 462 00:26:44,200 --> 00:26:47,200 Speaker 1: efficiency rather than allowing you to do something new. You're 463 00:26:47,240 --> 00:26:51,760 Speaker 1: not actually really even being enhanced. This is one step 464 00:26:51,760 --> 00:26:54,760 Speaker 1: above Google. But maybe it's an internal Google search. I 465 00:26:54,880 --> 00:26:58,399 Speaker 1: guess you can generate really crappy looking ah. I'm just 466 00:26:58,480 --> 00:27:00,800 Speaker 1: not sure where it is that. I don't think Jim 467 00:27:00,920 --> 00:27:04,560 Speaker 1: is either. He ends with one important and brutal note 468 00:27:05,080 --> 00:27:08,320 Speaker 1: that the more time that passes without significant AI applications, 469 00:27:08,600 --> 00:27:11,840 Speaker 1: the more challenging and I quote the AI story will become, 470 00:27:12,200 --> 00:27:15,320 Speaker 1: with corporate profitability likely floating the bubble as long as 471 00:27:15,359 --> 00:27:17,159 Speaker 1: it takes for the tech industry to hit a more 472 00:27:17,200 --> 00:27:20,000 Speaker 1: difficult economic period, kind of like we saw in the 473 00:27:20,000 --> 00:27:23,240 Speaker 1: middle of twenty twenty two. He also add his own prediction, 474 00:27:23,800 --> 00:27:26,879 Speaker 1: investor enthusiasm may begin to fade if important use cases 475 00:27:26,920 --> 00:27:29,560 Speaker 1: don't start to become more apparent in the next twelve 476 00:27:29,600 --> 00:27:32,960 Speaker 1: to eighteen months. And I think he's being a little optimistic. 477 00:27:37,480 --> 00:27:39,720 Speaker 1: While I won't recount the rest of the report, one 478 00:27:39,840 --> 00:27:42,840 Speaker 1: theme brought up repeatedly is the idea that America's power 479 00:27:42,840 --> 00:27:47,119 Speaker 1: grid is literally not ready for GENERAIVAI. In anew in 480 00:27:47,160 --> 00:27:51,080 Speaker 1: an interview with former Microsoft VP of Energy Brian Janus 481 00:27:51,440 --> 00:27:55,080 Speaker 1: on page fifteen, the report details numerous nightmarish problems that 482 00:27:55,119 --> 00:27:57,320 Speaker 1: the growth of GENERAIVAI is causing to the power grid, 483 00:27:57,640 --> 00:28:01,560 Speaker 1: such as hyperscalas like Microsoft and Google increasing their power 484 00:28:01,600 --> 00:28:04,119 Speaker 1: demands from a few hundred megawatts in the early twenty 485 00:28:04,160 --> 00:28:07,000 Speaker 1: to tens to a few gigawatts by twenty thirty, enough 486 00:28:07,000 --> 00:28:10,520 Speaker 1: to power multiple American cities, or the centralization of data 487 00:28:10,600 --> 00:28:13,760 Speaker 1: center operations from multiple big tech companies in northern Virginia, 488 00:28:14,359 --> 00:28:18,840 Speaker 1: potentially requiring a doubling of grid capacity over the next decade. 489 00:28:19,000 --> 00:28:22,320 Speaker 1: Utilities they've not experienced a period of load growth as 490 00:28:22,320 --> 00:28:25,439 Speaker 1: in a significant increase in power draw in nearly twenty years, 491 00:28:25,560 --> 00:28:27,919 Speaker 1: which is a problem because power infrastructure is slow to 492 00:28:27,920 --> 00:28:32,639 Speaker 1: build and involves numerous onerous permitting in bureaucratic measures to 493 00:28:32,640 --> 00:28:35,560 Speaker 1: make sure it's done properly or at all, and the 494 00:28:35,600 --> 00:28:37,960 Speaker 1: total capacity of power projects waiting to connect to the 495 00:28:38,000 --> 00:28:40,560 Speaker 1: grid grew thirty percent in the last year and wait 496 00:28:40,600 --> 00:28:44,920 Speaker 1: times of forty to seventy months. Expanding the grid is 497 00:28:45,040 --> 00:28:48,920 Speaker 1: no easy or quick task, and Mark Zuckerberg said that 498 00:28:48,920 --> 00:28:51,120 Speaker 1: these power constraints are the biggest thing in the way 499 00:28:51,160 --> 00:28:54,480 Speaker 1: of AI, which is sort of true and remarkable for 500 00:28:54,560 --> 00:28:58,520 Speaker 1: maguy who's are often full of shit in essence. On 501 00:28:58,600 --> 00:29:01,160 Speaker 1: top of generative AI not having any killer apps, not 502 00:29:01,240 --> 00:29:06,400 Speaker 1: meaningfully or helpfully increasing productivity, or GDP not generating any revenue, 503 00:29:06,440 --> 00:29:10,240 Speaker 1: not creating any new jobs, or massively changing existing industries, 504 00:29:10,560 --> 00:29:13,440 Speaker 1: it also requires America to rebuild its power grid, which 505 00:29:13,480 --> 00:29:17,440 Speaker 1: is a good thing which Brian Jane has regrettably had 506 00:29:17,680 --> 00:29:20,000 Speaker 1: that the US has kind of forgotten how to do. 507 00:29:20,200 --> 00:29:24,840 Speaker 1: US doesn't really do be give infrastructure anymore. It's not 508 00:29:24,360 --> 00:29:28,000 Speaker 1: a great scene. I don't know. Perhaps Sam Mortman's energy 509 00:29:28,000 --> 00:29:30,680 Speaker 1: breakthrough could be these fucking AI companies being made to 510 00:29:30,720 --> 00:29:33,200 Speaker 1: pay for new power infrastructure. And I don't mean them 511 00:29:33,240 --> 00:29:36,920 Speaker 1: oaning it. I mean tax them, tax their asses they're 512 00:29:36,920 --> 00:29:40,200 Speaker 1: burning the world, make them pay for it. The reason 513 00:29:40,200 --> 00:29:43,680 Speaker 1: I so agonizingly picked apart this report is that if 514 00:29:43,680 --> 00:29:48,360 Speaker 1: Goldman's sex is saying this, things are very, very bad. 515 00:29:48,960 --> 00:29:52,280 Speaker 1: And it also directly attacks the specific hype tactics of 516 00:29:52,320 --> 00:29:55,360 Speaker 1: AI freaks. The sense that generative AI will create new 517 00:29:55,400 --> 00:29:58,720 Speaker 1: jobs really hasn't in the last eighteen months. The sense 518 00:29:58,760 --> 00:30:01,040 Speaker 1: that costs will come down they haven't in there doesn't 519 00:30:01,040 --> 00:30:02,280 Speaker 1: seem to be a part of them to do so 520 00:30:02,320 --> 00:30:05,480 Speaker 1: in a way that matters. And that there's this incredible 521 00:30:05,520 --> 00:30:09,040 Speaker 1: demand for these products they claim exists, and there really isn't, 522 00:30:09,360 --> 00:30:12,600 Speaker 1: and I don't see a path to it even Goldman Sachs, 523 00:30:12,840 --> 00:30:16,240 Speaker 1: when describing the efficiency benefits of AI, added that it 524 00:30:16,320 --> 00:30:19,040 Speaker 1: was able to create an AI that updated historical data 525 00:30:19,040 --> 00:30:21,600 Speaker 1: in its company models more quickly than doing so manually 526 00:30:22,160 --> 00:30:25,240 Speaker 1: with one problem it costs six times as much to do. 527 00:30:27,120 --> 00:30:30,760 Speaker 1: I love the AI future. I love artificial intelligence. I 528 00:30:30,840 --> 00:30:33,959 Speaker 1: think it's so good that it's saving so many I 529 00:30:34,040 --> 00:30:41,680 Speaker 1: feel crazy. I feel crazy, I feel sorry. Moving on now, 530 00:30:41,680 --> 00:30:43,760 Speaker 1: there is a remaining defense, and it's also one of 531 00:30:43,760 --> 00:30:48,560 Speaker 1: the most annoying. People sometimes argue that perhaps open ai 532 00:30:48,920 --> 00:30:51,840 Speaker 1: has something we don't know about, some sort of big, sexy, 533 00:30:51,880 --> 00:30:54,520 Speaker 1: secret technology that will break the bones of every hater 534 00:30:54,600 --> 00:30:57,160 Speaker 1: that will bring me to my knees and beg the 535 00:30:57,200 --> 00:31:01,360 Speaker 1: machine God for mercy. And I have account point. No, 536 00:31:01,520 --> 00:31:07,040 Speaker 1: they don't. They don't have shit. Seriously. Mirror Marati, cto 537 00:31:07,080 --> 00:31:09,840 Speaker 1: of open ai, said in early June that the models 538 00:31:09,880 --> 00:31:12,040 Speaker 1: that open ai has in its labs are not much 539 00:31:12,040 --> 00:31:14,800 Speaker 1: more advanced than those that are publicly available. And that's 540 00:31:14,840 --> 00:31:17,560 Speaker 1: my answer to all of this. There is no magic trick. 541 00:31:17,800 --> 00:31:20,320 Speaker 1: There is no secret thing that Sam Ortman has that 542 00:31:20,360 --> 00:31:22,080 Speaker 1: he's going to reveal to us in the next few 543 00:31:22,120 --> 00:31:24,680 Speaker 1: months that makes me eat cro or some magical tool 544 00:31:24,800 --> 00:31:27,640 Speaker 1: the Microsoft or Google pops out that makes all of 545 00:31:27,680 --> 00:31:30,760 Speaker 1: this worth it. There isn't. I'm telling you, there isn't. 546 00:31:30,800 --> 00:31:33,640 Speaker 1: We would have seen a sign. There's not even a tinkle. 547 00:31:33,760 --> 00:31:36,480 Speaker 1: There's not a rumor, there's not a leak. There's nothing. 548 00:31:37,520 --> 00:31:40,320 Speaker 1: Generative AI, as I said a few months ago, is peaking. 549 00:31:40,440 --> 00:31:43,360 Speaker 1: If it hasn't already peaked, it can't do more than 550 00:31:43,360 --> 00:31:46,719 Speaker 1: it's currently doing. At least not much more other than 551 00:31:46,760 --> 00:31:49,640 Speaker 1: maybe doing it faster with some new inputs. It isn't 552 00:31:49,640 --> 00:31:53,400 Speaker 1: getting more efficient. SEQUOIA hype man David Kahn gleefully mentioned 553 00:31:53,400 --> 00:31:56,520 Speaker 1: in a recent blog that Nvideo's B one hundred chips, 554 00:31:56,520 --> 00:31:59,120 Speaker 1: the kinds used for training and running these models, will 555 00:31:59,160 --> 00:32:02,360 Speaker 1: have two point five x better performance for only twenty 556 00:32:02,400 --> 00:32:05,480 Speaker 1: five percent more cost, which doesn't mean a goddamn thing, 557 00:32:05,520 --> 00:32:08,800 Speaker 1: because generative AI isn't going to gain sentience or intelligence 558 00:32:08,880 --> 00:32:12,200 Speaker 1: or consciousness because it's able to run faster. I don't 559 00:32:12,240 --> 00:32:14,360 Speaker 1: even think it's going to be able to do more things. 560 00:32:15,120 --> 00:32:19,040 Speaker 1: Generative AI is not going to become AGI, nor will 561 00:32:19,040 --> 00:32:21,680 Speaker 1: it become the kind of artificial intelligence you've seen in 562 00:32:21,720 --> 00:32:25,400 Speaker 1: science fiction. Ultra smart assistance like Jarvis from Iron Man 563 00:32:25,560 --> 00:32:29,360 Speaker 1: would require a kind of consciousness that no technology currently 564 00:32:29,480 --> 00:32:31,920 Speaker 1: or may ever be able to produce, which is the 565 00:32:31,960 --> 00:32:35,840 Speaker 1: ability to both process and understand information flawlessly and make 566 00:32:35,880 --> 00:32:39,440 Speaker 1: decisions based on experience, which, if I haven't been clear enough, 567 00:32:39,760 --> 00:32:43,400 Speaker 1: are all entirely distinct things. And generative AI and AI 568 00:32:43,480 --> 00:32:48,000 Speaker 1: in general doesn't have experiences. They don't have shades of gray, 569 00:32:48,360 --> 00:32:52,520 Speaker 1: They only have shades of brown. Poops, poop, joke. Generative 570 00:32:52,520 --> 00:32:55,760 Speaker 1: AI at best processes of information when it trains on data, 571 00:32:55,800 --> 00:32:58,440 Speaker 1: but at no point does it learn or understand because 572 00:32:58,440 --> 00:33:01,479 Speaker 1: everything it's doing is based on ingesting, training, data, and 573 00:33:01,520 --> 00:33:05,040 Speaker 1: developing answers based on a mathematical sense of probability rather 574 00:33:05,120 --> 00:33:09,880 Speaker 1: than any appreciation or comprehension of the material itself. Llms 575 00:33:09,960 --> 00:33:13,000 Speaker 1: are entirely different pieces of technology to that of an 576 00:33:13,080 --> 00:33:16,360 Speaker 1: artificial intelligence in the sense that the AI bubble is 577 00:33:16,400 --> 00:33:20,200 Speaker 1: hyping and it's disgraceful that the AI industry has taken 578 00:33:20,240 --> 00:33:24,320 Speaker 1: so much money and attention with such a flagrant offensive lie. 579 00:33:24,560 --> 00:33:27,600 Speaker 1: The jobs market isn't going to change because of generative AI, 580 00:33:27,960 --> 00:33:31,400 Speaker 1: because generative AI can't actually do many jobs, and it's 581 00:33:31,440 --> 00:33:34,440 Speaker 1: mediocre at the things that it's capable of doing, which 582 00:33:34,480 --> 00:33:36,760 Speaker 1: is why it's so shocking. There are even people who'd 583 00:33:36,800 --> 00:33:40,200 Speaker 1: replace real people with it. While it's a useful efficiency 584 00:33:40,240 --> 00:33:43,680 Speaker 1: tool in specific contexts, said efficiency is based off of 585 00:33:43,840 --> 00:33:47,440 Speaker 1: technology that's extremely expensive, and I believe at some point 586 00:33:47,520 --> 00:33:50,320 Speaker 1: AI companies like Anthropic and OpenAI are going to have 587 00:33:50,400 --> 00:33:52,960 Speaker 1: to increase prices or begin to collapse under the weight 588 00:33:52,960 --> 00:33:56,560 Speaker 1: of a technology that has no path to profitability. If 589 00:33:56,600 --> 00:33:58,880 Speaker 1: there was some secret way that this would all get fixed. 590 00:33:58,920 --> 00:34:02,640 Speaker 1: Wouldn't Microsoft or Meta or Google or maybe Amazon, who's 591 00:34:02,960 --> 00:34:06,080 Speaker 1: CEO of their cloud platform compared GENERATIVEAI to dot com 592 00:34:06,120 --> 00:34:09,279 Speaker 1: bubble in February. Wouldn't they have taken advantage of this big, 593 00:34:09,320 --> 00:34:12,800 Speaker 1: sexy secret. Why am I hearing the open AI is 594 00:34:12,840 --> 00:34:15,879 Speaker 1: already trying to raise another multi billion dollar round after 595 00:34:15,960 --> 00:34:18,720 Speaker 1: raising an indeterminate amount at an eighty billion dollar valuation 596 00:34:18,800 --> 00:34:23,359 Speaker 1: in February. Isn't open AI's annualized revenue three point four 597 00:34:23,400 --> 00:34:27,040 Speaker 1: billion dollars? Why do they need more money? How much 598 00:34:27,040 --> 00:34:30,839 Speaker 1: are they burning? Because I'm guessing that if it's more 599 00:34:30,880 --> 00:34:33,239 Speaker 1: than three point four billion dollars, it's got to be 600 00:34:33,280 --> 00:34:35,520 Speaker 1: a lot more if they're still trying to raise capital, 601 00:34:35,800 --> 00:34:37,759 Speaker 1: and if they're still going out there lying about what 602 00:34:37,840 --> 00:34:40,680 Speaker 1: CHATGBT could do in the future. But I'll give you 603 00:34:40,719 --> 00:34:44,640 Speaker 1: an educated guess, because whatever they open AI and other 604 00:34:44,680 --> 00:34:49,400 Speaker 1: generative AI hucksters have today is obviously painfully not the future. 605 00:34:50,480 --> 00:34:54,480 Speaker 1: Generative AI is not the future but a regurgitation of 606 00:34:54,520 --> 00:34:57,520 Speaker 1: the past, a useful, yet not groundbreaking way to quickly 607 00:34:57,560 --> 00:35:00,719 Speaker 1: generate new data from old that costs too much to 608 00:35:00,719 --> 00:35:03,840 Speaker 1: make the compute and the energy demands worth it. Google 609 00:35:03,880 --> 00:35:06,200 Speaker 1: grew its emissions by forty eight percent in the last 610 00:35:06,280 --> 00:35:08,880 Speaker 1: five years, chasing a technology that made its search engine 611 00:35:08,920 --> 00:35:12,400 Speaker 1: even worse than it already was, and they've got nothing 612 00:35:12,400 --> 00:35:14,399 Speaker 1: to show for it other than a bunch of very 613 00:35:14,400 --> 00:35:18,440 Speaker 1: funny headlines. It's genuinely remarkable how many people have been 614 00:35:18,480 --> 00:35:22,680 Speaker 1: won over by this insane con this unscrupulous manipulation of 615 00:35:22,719 --> 00:35:27,720 Speaker 1: the capital markets, the media, and brainless executives disconnected from production. 616 00:35:27,960 --> 00:35:30,400 Speaker 1: They're all buying it, and it's all thanks to a 617 00:35:30,440 --> 00:35:35,520 Speaker 1: tech industry that's disconnected itself from building useful things. I've 618 00:35:35,520 --> 00:35:37,840 Speaker 1: been asked a few times when I think the bubble 619 00:35:37,840 --> 00:35:40,160 Speaker 1: will burst, by the way, and I maintain that part 620 00:35:40,200 --> 00:35:43,120 Speaker 1: of the collapse will be investor descent, punishing one of 621 00:35:43,120 --> 00:35:46,680 Speaker 1: the major providers, Microsoft or Google for a massive investment 622 00:35:46,680 --> 00:35:51,279 Speaker 1: in an industry that produces little actual revenue. However, I 623 00:35:51,320 --> 00:35:53,880 Speaker 1: think what'll really get the bubble poppin will be a 624 00:35:53,920 --> 00:35:57,000 Speaker 1: succession of bad events, like say, Figma pausing its new 625 00:35:57,080 --> 00:36:00,839 Speaker 1: AI feature after it immediately plagiarized Apple's weather, likely because 626 00:36:00,880 --> 00:36:02,920 Speaker 1: it was trained on it as part of its training data, 627 00:36:03,360 --> 00:36:06,640 Speaker 1: crested by one large, nasty one such as a major 628 00:36:06,680 --> 00:36:09,680 Speaker 1: AI company like chatbox company character ai, which raised one 629 00:36:09,719 --> 00:36:11,799 Speaker 1: hundred and fifty million dollars in funding a few years ago, 630 00:36:11,840 --> 00:36:15,000 Speaker 1: and the information great publication claims might sell to one 631 00:36:15,000 --> 00:36:17,840 Speaker 1: of the bigger tech companies. I think someone like character 632 00:36:17,880 --> 00:36:20,759 Speaker 1: ai could collapse under the way of a non sustainable 633 00:36:20,800 --> 00:36:25,600 Speaker 1: business model and unprofitable technology and just also, do you 634 00:36:25,680 --> 00:36:27,440 Speaker 1: really need a chat this is what this company does. 635 00:36:27,480 --> 00:36:30,720 Speaker 1: You can speak to like sotorro Gojo from jiu jitsu Chism, 636 00:36:30,719 --> 00:36:32,880 Speaker 1: which is an anime and amonga. By the way, if 637 00:36:32,880 --> 00:36:36,560 Speaker 1: you're reading the margarets, take it too long and it's weird. 638 00:36:36,600 --> 00:36:38,239 Speaker 1: It's one of these weird companies you can talk to 639 00:36:38,360 --> 00:36:40,360 Speaker 1: historical figures and you're kind of like, okay, I assume 640 00:36:40,360 --> 00:36:43,960 Speaker 1: you have a vague enterprise product. Oh no, it's all 641 00:36:44,040 --> 00:36:47,719 Speaker 1: just nonsense. It's just when I see these companies and 642 00:36:47,760 --> 00:36:49,440 Speaker 1: want they raise them, just like, what the fuck are 643 00:36:49,480 --> 00:36:51,839 Speaker 1: you even doing all day? But it could be a 644 00:36:51,880 --> 00:36:55,480 Speaker 1: slightly more useful sounding one like Cognition Ai, which raised 645 00:36:55,480 --> 00:36:58,319 Speaker 1: one hundred and fifty seventy five million dollars a two 646 00:36:58,400 --> 00:37:01,720 Speaker 1: billion dollar valuation in April to make an AI software 647 00:37:01,719 --> 00:37:07,520 Speaker 1: engineer with one problem. Well, I mean it's just a 648 00:37:07,560 --> 00:37:09,319 Speaker 1: small one. It was that they had to fake a 649 00:37:09,560 --> 00:37:11,960 Speaker 1: demo of it working. They had to fake the demo, 650 00:37:12,200 --> 00:37:14,120 Speaker 1: and I've included a link in the notes to a 651 00:37:14,160 --> 00:37:16,719 Speaker 1: YouTube of a really great software engineer just sitting there 652 00:37:16,760 --> 00:37:19,280 Speaker 1: saying I'm actually a fan of AI, but this is crap, 653 00:37:19,320 --> 00:37:22,280 Speaker 1: and he just goes through it line by line. Basically, 654 00:37:22,520 --> 00:37:24,600 Speaker 1: there's going to be a moment that spooks the venture 655 00:37:24,600 --> 00:37:27,280 Speaker 1: capital firms, and it's going to spook them into pushing 656 00:37:27,320 --> 00:37:30,400 Speaker 1: one of their startups to sell, which will lead to 657 00:37:30,440 --> 00:37:33,920 Speaker 1: a sudden and unexpected yet very obvious collapse of a 658 00:37:33,920 --> 00:37:37,680 Speaker 1: major player. As valuations drop and people get desperate for 659 00:37:37,760 --> 00:37:41,279 Speaker 1: open AI and anthropic there really is no path profitability, 660 00:37:41,880 --> 00:37:44,480 Speaker 1: only one that includes burning further billions of dollars in 661 00:37:44,480 --> 00:37:47,160 Speaker 1: the hope that they discover something, anything, that might be 662 00:37:47,160 --> 00:37:50,880 Speaker 1: truly innovative or indicative of the future, rather than further 663 00:37:50,960 --> 00:37:53,920 Speaker 1: iterations of generative AI, which is at best an extremely 664 00:37:53,960 --> 00:37:57,759 Speaker 1: expensive new way to process data. I just see no 665 00:37:57,920 --> 00:38:01,120 Speaker 1: situation where open ai and anthropy continue to iterate on 666 00:38:01,200 --> 00:38:05,320 Speaker 1: large language models in perpetuity, because at some point, Microsoft, Amazon, 667 00:38:05,400 --> 00:38:08,560 Speaker 1: and Google will decide that cloud compute welfare isn't a 668 00:38:08,600 --> 00:38:11,440 Speaker 1: business model. And by the way, all of those companies 669 00:38:11,440 --> 00:38:14,239 Speaker 1: I listed, Microsoft, Amazon, and Google have either invested in 670 00:38:14,320 --> 00:38:18,040 Speaker 1: open AI or Anthropic. In the case of Amazon and Google, 671 00:38:18,080 --> 00:38:20,680 Speaker 1: they've only invested in anthropic, and a lot of that 672 00:38:20,760 --> 00:38:23,920 Speaker 1: is in cloud credits, meaning that it's just welfare. It's hilarious, 673 00:38:24,040 --> 00:38:26,279 Speaker 1: very right wing people in the tech industry. You think 674 00:38:26,280 --> 00:38:28,480 Speaker 1: that this kind of welfa would bother them. I know 675 00:38:28,560 --> 00:38:32,520 Speaker 1: what the difference is anyway, Without a real tangible breakthrough, 676 00:38:32,560 --> 00:38:34,600 Speaker 1: one that would require them to leave the world of 677 00:38:34,680 --> 00:38:38,320 Speaker 1: large language models entirely, it's unclear how generative AI companies 678 00:38:38,360 --> 00:38:42,319 Speaker 1: can even survive. GENERATIVEAI is locked in the Red Queen's race, 679 00:38:42,719 --> 00:38:45,200 Speaker 1: burning money to make money in an attempt to prove 680 00:38:45,239 --> 00:38:47,600 Speaker 1: that they one day will make more money, despite their 681 00:38:47,840 --> 00:38:50,800 Speaker 1: being no clear path to making more money than they spend. 682 00:38:51,960 --> 00:38:54,360 Speaker 1: It's all a little wild, a little bit upsetting, and 683 00:38:54,400 --> 00:38:56,960 Speaker 1: a little bit crazy making. And I feel a little 684 00:38:56,960 --> 00:39:00,359 Speaker 1: bit crazy every time I put together one of these 685 00:39:00,360 --> 00:39:06,400 Speaker 1: episodes because it's also patently ridiculous. GENERATIVEAI is unprofitable, unsustainable, 686 00:39:06,440 --> 00:39:08,880 Speaker 1: and fundamentally limited in what it can do thanks to 687 00:39:08,920 --> 00:39:12,960 Speaker 1: the fact that it's probabilistically generating an answer. It's not 688 00:39:13,120 --> 00:39:16,839 Speaker 1: learning anything. It doesn't know anything. It isn't intelligent, it's 689 00:39:16,880 --> 00:39:22,799 Speaker 1: only artificial. It's been eighteen months since this bubble started inflating, 690 00:39:22,960 --> 00:39:27,000 Speaker 1: and since then very little has actually happened involving technology 691 00:39:27,080 --> 00:39:31,040 Speaker 1: doing new stuff, just iterative explorations of the very clear 692 00:39:31,040 --> 00:39:33,880 Speaker 1: limits of what an AI model that generates answers based 693 00:39:33,880 --> 00:39:36,920 Speaker 1: on training data can produce, with the answer being something 694 00:39:36,960 --> 00:39:40,719 Speaker 1: that at times is sort of good. It's obvious, it's 695 00:39:40,760 --> 00:39:45,080 Speaker 1: well documented. GENERATIVEAI costs far too much, it isn't getting cheaper, 696 00:39:45,280 --> 00:39:47,759 Speaker 1: it uses too much power, and doesn't do enough to 697 00:39:47,920 --> 00:39:50,960 Speaker 1: justify its existence. There are no killer apps, and no 698 00:39:51,080 --> 00:39:53,840 Speaker 1: killer apps on the horizon, and there are no answers 699 00:39:53,880 --> 00:39:57,719 Speaker 1: to these problems. I don't know why more people aren't 700 00:39:57,760 --> 00:40:00,400 Speaker 1: saying this as loudly as I am. I don't now 701 00:40:00,800 --> 00:40:04,520 Speaker 1: why everyone's not freaking out. I understand that big tech 702 00:40:04,600 --> 00:40:07,080 Speaker 1: desperately needs this to be the next hyper growth market, 703 00:40:07,160 --> 00:40:09,839 Speaker 1: as they haven't got any others. But pursuing this cause 704 00:40:09,880 --> 00:40:14,279 Speaker 1: our useful environmental disaster causing cloud efficiency boondoggle will send 705 00:40:14,320 --> 00:40:17,319 Speaker 1: shock waves through the industry when it all collapses and 706 00:40:17,360 --> 00:40:20,640 Speaker 1: what's frustrating to me about this is it's becoming obvious, 707 00:40:20,640 --> 00:40:23,640 Speaker 1: and you're seeing people lately kind of come out their 708 00:40:23,640 --> 00:40:25,719 Speaker 1: shells and saying, I'm not sure this is going to 709 00:40:25,800 --> 00:40:28,200 Speaker 1: be good. I'm not sure are we hitting PKI as 710 00:40:28,200 --> 00:40:30,080 Speaker 1: a great piece in the New York Times by someone 711 00:40:30,120 --> 00:40:32,200 Speaker 1: recently saying that we hit PKI and it used a 712 00:40:32,200 --> 00:40:35,120 Speaker 1: bunch of my links. Thank you, Julia. Anyway, putting aside 713 00:40:35,160 --> 00:40:38,480 Speaker 1: my bitterness, the problem here is all of the money 714 00:40:38,480 --> 00:40:41,720 Speaker 1: and power going into this, but also the big lie 715 00:40:41,760 --> 00:40:46,120 Speaker 1: being perpetrated by people. In my next episode, I'm going 716 00:40:46,160 --> 00:40:48,000 Speaker 1: to go into some more of that. I'm going to 717 00:40:48,040 --> 00:40:51,200 Speaker 1: talk about the fact that a lot of these companies 718 00:40:51,239 --> 00:40:54,440 Speaker 1: are not really doing anything that they're kind of doing 719 00:40:54,520 --> 00:40:57,160 Speaker 1: marketing exercises as a means of manipulating the media in 720 00:40:57,200 --> 00:41:02,440 Speaker 1: the markets. It's all just the discrace waste. The people 721 00:41:02,440 --> 00:41:05,760 Speaker 1: getting hurt by Generative AI are people that are working 722 00:41:05,800 --> 00:41:07,920 Speaker 1: contract jobs because it's harder than ever to get a 723 00:41:07,920 --> 00:41:11,759 Speaker 1: full time job. They're being replaced by shitty software that 724 00:41:11,800 --> 00:41:14,680 Speaker 1: cost too much money that will invariably go away when 725 00:41:14,680 --> 00:41:18,359 Speaker 1: this all collapses. And I feel that the media has 726 00:41:18,400 --> 00:41:22,800 Speaker 1: some responsibility here. The markets too, but definitely the media. 727 00:41:22,880 --> 00:41:25,080 Speaker 1: I understand that it's difficult to look at the tech 728 00:41:25,120 --> 00:41:27,920 Speaker 1: industry and think, are they all kind of leading us on? 729 00:41:28,200 --> 00:41:30,279 Speaker 1: But they've done it so many times and this is 730 00:41:30,320 --> 00:41:35,080 Speaker 1: the worst one I've seen. This is ridiculous. Billions of 731 00:41:35,120 --> 00:41:39,319 Speaker 1: dollars wasted, so much time, so much energy, And yes, 732 00:41:39,640 --> 00:41:41,840 Speaker 1: we can cover this, Yes we can talk about this. 733 00:41:42,280 --> 00:41:44,520 Speaker 1: We don't owe them, are both sides. Why do we 734 00:41:44,520 --> 00:41:46,840 Speaker 1: owe them that they wouldn't give us one? Have you 735 00:41:46,880 --> 00:41:49,279 Speaker 1: ever seen a tech entrepreneur go, oh, that was a nice, 736 00:41:49,280 --> 00:41:51,480 Speaker 1: fair interview? Or do they usually piss and moan the 737 00:41:51,520 --> 00:41:55,279 Speaker 1: moment anyone says anything negative. No, the right way to 738 00:41:55,280 --> 00:41:58,879 Speaker 1: look at this now is with suspicion and frankly most innovations, 739 00:42:00,239 --> 00:42:04,920 Speaker 1: but in particular this one, because despite an alarming amount 740 00:42:05,120 --> 00:42:08,680 Speaker 1: people in the media saying things like, oh yeah, well, 741 00:42:08,840 --> 00:42:12,600 Speaker 1: generative I will of course lead to superintelligence, it's never 742 00:42:12,640 --> 00:42:15,040 Speaker 1: going to do that. And we all I don't know 743 00:42:15,080 --> 00:42:17,080 Speaker 1: if you consider me a reporter or a journalist. This 744 00:42:17,120 --> 00:42:19,720 Speaker 1: is a conversation to have with me via my email, 745 00:42:19,760 --> 00:42:23,719 Speaker 1: I guess, but whatever I am, I should not be 746 00:42:24,040 --> 00:42:26,480 Speaker 1: the only one that's really going out this hard, and 747 00:42:26,520 --> 00:42:28,360 Speaker 1: I know I'm not. There are other people, some people 748 00:42:28,400 --> 00:42:31,719 Speaker 1: right about Brian Merchant on the Machine, fantastic Riyer another 749 00:42:31,760 --> 00:42:34,360 Speaker 1: great guy, Paris Mark's another great guy doing great work. 750 00:42:34,800 --> 00:42:37,399 Speaker 1: And Ellen Hewitt the Bloomberger was on a few episodes ago. 751 00:42:37,480 --> 00:42:40,680 Speaker 1: Great coverage of Sam Hortmaan. But nevertheless, it is time 752 00:42:40,760 --> 00:42:45,480 Speaker 1: to start saying when we cover generative AI, specifically open 753 00:42:45,520 --> 00:42:50,239 Speaker 1: AI and Anthropic, hey, where are we with superintelligence? And 754 00:42:50,320 --> 00:42:52,640 Speaker 1: I don't mean what level we're at, I mean what 755 00:42:52,800 --> 00:42:55,319 Speaker 1: is the technology that you have that is going to 756 00:42:55,400 --> 00:42:57,760 Speaker 1: lead to that? And when they refuse to give an answer, 757 00:42:57,960 --> 00:42:59,960 Speaker 1: what should be covered is that they don't have one 758 00:43:00,680 --> 00:43:04,880 Speaker 1: make them dance? Why do we have to do the 759 00:43:05,040 --> 00:43:10,000 Speaker 1: work to make AI important? Why doesn't generative AI impress 760 00:43:10,080 --> 00:43:14,040 Speaker 1: us right now? It's taking money out of people's hands, 761 00:43:14,640 --> 00:43:20,920 Speaker 1: It's killing freelancers. It's really horrible. And the thing to 762 00:43:20,960 --> 00:43:24,359 Speaker 1: cover is that you can cover fun apps. I don't 763 00:43:24,360 --> 00:43:28,000 Speaker 1: have a problem with that, but when it comes to Google, Microsoft, Amazon, 764 00:43:28,360 --> 00:43:31,000 Speaker 1: Open AI, Anthropic, all of these companies, we need to 765 00:43:31,040 --> 00:43:33,200 Speaker 1: treat them as kind of interlopers at this point because 766 00:43:33,239 --> 00:43:35,520 Speaker 1: they need to fucking show us what they're working on them. 767 00:43:35,560 --> 00:43:37,640 Speaker 1: They need I don't care about the safety side. They 768 00:43:37,640 --> 00:43:39,959 Speaker 1: don't care about it either. We don't need to worry 769 00:43:39,960 --> 00:43:43,200 Speaker 1: about the safety issue right now, because the only safety 770 00:43:43,239 --> 00:43:46,400 Speaker 1: issue is the problem we are facing today, which is 771 00:43:46,400 --> 00:43:49,360 Speaker 1: that millions of people's works being stolen and millions of 772 00:43:49,360 --> 00:43:51,879 Speaker 1: other people are having their jobs taken in the most 773 00:43:51,920 --> 00:43:54,600 Speaker 1: mediocre way by shitty bosses who don't do any work. 774 00:43:55,520 --> 00:43:58,400 Speaker 1: When Sam Altman gets on stage and claims that AI 775 00:43:58,480 --> 00:44:01,319 Speaker 1: will solve all the physics, the appropriate answer is, what 776 00:44:01,320 --> 00:44:03,920 Speaker 1: a fuck are you talking about? You sound like an idiot. 777 00:44:04,880 --> 00:44:07,960 Speaker 1: When Brad light Caps, COO of Open AI says, I 778 00:44:08,000 --> 00:44:10,160 Speaker 1: don't know if it's trained on YouTube, you should say 779 00:44:10,280 --> 00:44:13,600 Speaker 1: why don't you know? And then when he bubbles away, goes, oh, 780 00:44:13,800 --> 00:44:16,800 Speaker 1: I'm not sure. Say it's really weird that a company 781 00:44:17,480 --> 00:44:20,160 Speaker 1: worth eighty billion dollars as a COO who does not 782 00:44:20,320 --> 00:44:23,000 Speaker 1: know what's happening, saying goes from Mira Marati when she 783 00:44:23,080 --> 00:44:26,840 Speaker 1: can't answer whether Sura was trained on YouTube. These people 784 00:44:26,960 --> 00:44:31,880 Speaker 1: must be treated, if not as criminals, as suspicious types, 785 00:44:32,200 --> 00:44:36,880 Speaker 1: as con artists, as people that have absorbed so much attention, 786 00:44:37,440 --> 00:44:43,040 Speaker 1: so much money, so much time, so much adulation. They 787 00:44:43,080 --> 00:44:46,719 Speaker 1: should be held accountable and holding them accountable starts with 788 00:44:46,800 --> 00:44:50,000 Speaker 1: a real evaluation of generative AI. And as I've shown 789 00:44:50,040 --> 00:44:53,719 Speaker 1: you today, this bubble is full of shit. I'm gonna 790 00:44:53,800 --> 00:44:56,279 Speaker 1: end on a happier note, though I looked it up 791 00:44:56,360 --> 00:44:58,839 Speaker 1: yesterday and as of speaking this out, it's only been 792 00:44:58,880 --> 00:45:00,440 Speaker 1: four and a half months of this show show. I 793 00:45:00,520 --> 00:45:02,920 Speaker 1: thought it had been six months or a year. I 794 00:45:02,960 --> 00:45:05,520 Speaker 1: have time dilation issues in my brain and I will 795 00:45:05,520 --> 00:45:08,080 Speaker 1: be seeing a doctor. But I genuinely want to thank 796 00:45:08,120 --> 00:45:09,960 Speaker 1: all of you. I want to thank everyone who's been 797 00:45:10,000 --> 00:45:13,080 Speaker 1: there from last episode or the first episode, everyone who's 798 00:45:13,080 --> 00:45:16,040 Speaker 1: come through through different ways. This show's evolving and it 799 00:45:16,040 --> 00:45:19,120 Speaker 1: will continue to evolve. I know I'm angry, I'm pissy, 800 00:45:20,320 --> 00:45:23,000 Speaker 1: but I think at this time in history, there's a 801 00:45:23,000 --> 00:45:25,719 Speaker 1: reason to be. I'm not feeling myself full of it. 802 00:45:25,719 --> 00:45:28,759 Speaker 1: It's naturally there. It's naturally there. When I look at 803 00:45:28,760 --> 00:45:31,200 Speaker 1: an industry, at tech industry that I genuinely love, that 804 00:45:31,320 --> 00:45:33,840 Speaker 1: genuinely made me the person I am today, and I 805 00:45:33,960 --> 00:45:36,319 Speaker 1: see what's being done to it, and it makes me 806 00:45:36,400 --> 00:45:40,560 Speaker 1: so pissed off. Because technology is everything. We do. To 807 00:45:40,840 --> 00:45:44,200 Speaker 1: discard tech as something that happens just on the computer 808 00:45:44,239 --> 00:45:47,040 Speaker 1: and then there's real life. It's kind of silly. They're 809 00:45:47,040 --> 00:45:49,520 Speaker 1: both the same thing. Now we are all influenced by 810 00:45:49,600 --> 00:45:53,680 Speaker 1: offline and online culture, times far more by online culture. 811 00:45:54,360 --> 00:45:56,440 Speaker 1: As this show progresses, I'm going to get more into that, 812 00:45:56,600 --> 00:45:58,320 Speaker 1: and the upcoming episodes I have are going to be 813 00:45:58,360 --> 00:46:01,120 Speaker 1: a lot of fun. The seconds of this week is 814 00:46:01,800 --> 00:46:04,040 Speaker 1: just a real laugh, and it kind of showed me 815 00:46:04,080 --> 00:46:07,239 Speaker 1: how this show can grow into a more conversational, fun, 816 00:46:07,400 --> 00:46:10,600 Speaker 1: interesting thing where I talk to all sorts of folks 817 00:46:10,600 --> 00:46:12,160 Speaker 1: about all sorts of things in tech and how it 818 00:46:12,160 --> 00:46:15,879 Speaker 1: affects them. I'm really grateful for you all listening. You'll 819 00:46:15,920 --> 00:46:17,480 Speaker 1: hear my email after that. And by the way, it's 820 00:46:17,480 --> 00:46:20,200 Speaker 1: easy as an letter E letter Z or z from 821 00:46:20,200 --> 00:46:24,760 Speaker 1: my British listeners. Email me easy abotter offline dot com, 822 00:46:25,120 --> 00:46:27,680 Speaker 1: message me on Twitter or Facebook or Instagram. I'm always 823 00:46:27,680 --> 00:46:29,160 Speaker 1: happy to hear from people. I want to know how 824 00:46:29,160 --> 00:46:31,800 Speaker 1: to make this better, but I know one way is 825 00:46:31,840 --> 00:46:34,200 Speaker 1: to just keep doing it. And I'm really excited for 826 00:46:34,239 --> 00:46:36,719 Speaker 1: the future and I'm very very thankful to have all 827 00:46:36,719 --> 00:46:39,440 Speaker 1: of you there, even the people who are very mad 828 00:46:39,480 --> 00:46:43,000 Speaker 1: at me if you're not like in Generative AI. But seriously, though, 829 00:46:44,200 --> 00:46:48,120 Speaker 1: one final thing. People will see this stuff and they 830 00:46:48,120 --> 00:46:50,319 Speaker 1: do contact me fairly regularly, and they say, what can 831 00:46:50,360 --> 00:46:51,680 Speaker 1: I do? And I said this at the end of 832 00:46:51,680 --> 00:46:54,000 Speaker 1: the show Older Supremacy as well, but I'll say it 833 00:46:54,000 --> 00:46:57,239 Speaker 1: here as well. You can't do a ton. What you 834 00:46:57,280 --> 00:47:00,720 Speaker 1: can do is say people's names, Mirror Marathi, Sam Prabagar, 835 00:47:00,840 --> 00:47:04,279 Speaker 1: ragavans Under Pashi. I know it sounds silly, but a 836 00:47:04,320 --> 00:47:07,200 Speaker 1: lot of these ultra rich people, they don't care about experiences, 837 00:47:07,560 --> 00:47:10,800 Speaker 1: don't really have predators. What they have is their reputations. 838 00:47:11,880 --> 00:47:15,279 Speaker 1: But also what they have is their lies. I know 839 00:47:15,320 --> 00:47:17,520 Speaker 1: it sounds a bit dramatic to say they're liars, but 840 00:47:17,600 --> 00:47:20,319 Speaker 1: look at what Sam Moltman's saying. The only reason Sam 841 00:47:20,360 --> 00:47:22,120 Speaker 1: Mortman has been able to get where he is today 842 00:47:22,400 --> 00:47:25,719 Speaker 1: is a series of lies and cons. The way to 843 00:47:25,760 --> 00:47:29,000 Speaker 1: break generative AI, which it should be broken until it 844 00:47:29,000 --> 00:47:32,520 Speaker 1: can prove itself profitable and not environmentally destructive, is to 845 00:47:32,520 --> 00:47:34,880 Speaker 1: push back, to constantly talk about how bad it is. 846 00:47:34,920 --> 00:47:38,120 Speaker 1: To share your bad experiences with it, share every hallucination, 847 00:47:38,440 --> 00:47:42,200 Speaker 1: share them publicly on social media channels. Tag Sam Multman 848 00:47:42,280 --> 00:47:46,200 Speaker 1: at Samma on Twitter, Sama do it. I know this 849 00:47:46,280 --> 00:47:49,600 Speaker 1: sounds silly, but guess what these people have grown so 850 00:47:49,760 --> 00:47:53,480 Speaker 1: rich and powerful from nobody calling out their shit. And 851 00:47:53,520 --> 00:47:55,440 Speaker 1: there are people up me. They're journalists who've done it. 852 00:47:55,719 --> 00:47:58,040 Speaker 1: And I'm not saying nobody does it. Don't get mad 853 00:47:58,080 --> 00:48:01,600 Speaker 1: at me. What I'm saying is thousands, tens of thousands, 854 00:48:01,719 --> 00:48:04,280 Speaker 1: hundreds of thousands of people listening while this stuff sucks, 855 00:48:04,480 --> 00:48:09,000 Speaker 1: why it's bad. We'll get to the tech people and 856 00:48:09,040 --> 00:48:11,200 Speaker 1: eventually one of those messages will get to some of 857 00:48:11,239 --> 00:48:14,480 Speaker 1: these financial people, and when they really turn against this, 858 00:48:15,120 --> 00:48:19,480 Speaker 1: it will collapse. And if it doesn't, if I'm wrong, 859 00:48:19,640 --> 00:48:21,640 Speaker 1: well I'll do an episode about how wrong I am, 860 00:48:21,719 --> 00:48:24,920 Speaker 1: because you know what this is meant to be informative, 861 00:48:24,920 --> 00:48:27,959 Speaker 1: it's meant to be entertainment. But also I don't mind 862 00:48:28,000 --> 00:48:32,560 Speaker 1: being wrong, just kind of yet to be so someone's 863 00:48:32,560 --> 00:48:43,840 Speaker 1: gonna be mad at that anyway. Thank you, Thank you 864 00:48:43,840 --> 00:48:46,560 Speaker 1: for listening to Better Offline. The editor and composer of 865 00:48:46,600 --> 00:48:49,560 Speaker 1: the Better Offline theme song is Metasowski. You can check 866 00:48:49,560 --> 00:48:52,480 Speaker 1: out more of his music and audio projects at Matasowski 867 00:48:52,560 --> 00:48:58,480 Speaker 1: dot com, m A. T. T OsO w Ski dot com. 868 00:48:58,560 --> 00:49:01,160 Speaker 1: You can email me an easy, Better Offline dot com, 869 00:49:01,280 --> 00:49:03,879 Speaker 1: or visit Better Offline dot com to find more podcast links, 870 00:49:03,880 --> 00:49:07,000 Speaker 1: and of course my newsletter. I also really recommend you 871 00:49:07,080 --> 00:49:09,200 Speaker 1: go to chat dot where's youreaed dot at to visit 872 00:49:09,239 --> 00:49:11,880 Speaker 1: the discord, and go to our slash Better Offline to 873 00:49:11,960 --> 00:49:15,040 Speaker 1: check out our reddit. Thank you so much for listening. 874 00:49:15,880 --> 00:49:18,600 Speaker 1: Better Offline is a production of cool Zone Media. For 875 00:49:18,680 --> 00:49:21,840 Speaker 1: more from cool Zone Media, visit our website cool Zonemedia 876 00:49:21,920 --> 00:49:24,759 Speaker 1: dot com, or check us out on the iHeartRadio app, 877 00:49:24,800 --> 00:49:27,520 Speaker 1: Apple Podcasts, or wherever you get your podcasts.