WEBVTT - Pop Culture 0:00:02.160 --> 0:00:05.560 Zone Media. Well, I know that assholes grow on trees, 0:00:05.559 --> 0:00:07.800 but I'm here to trim the leaves. This is Better 0:00:07.880 --> 0:00:22.800 off Line. I'm your host ed dait tron. It's bullshit 0:00:22.840 --> 0:00:25.800 Week here at Better Offline headquarters. The thousands of elves 0:00:25.840 --> 0:00:28.560 that work for me have been finding me things that 0:00:28.640 --> 0:00:30.680 say bullshit for no reason. When I wrote the episode 0:00:30.720 --> 0:00:33.040 weeks ago. I don't know what to do with them. 0:00:33.080 --> 0:00:35.639 But anyway, we're starting with one of the largest financial 0:00:35.640 --> 0:00:40.120 institutions of the world calling bullshit on Generative AI as usual. 0:00:40.400 --> 0:00:42.920 Don't take my word for it. Check the episode notes 0:00:42.920 --> 0:00:45.519 for links that map to everything I'm talking about. I've 0:00:45.520 --> 0:00:47.440 tried my best to map them to exactly what I'm 0:00:47.479 --> 0:00:49.240 saying too, and you can feel free to yell about 0:00:49.240 --> 0:00:51.080 it like yell out the words like the Beastie Boys, 0:00:51.440 --> 0:00:53.640 or just follow along, because I think it's important for 0:00:53.680 --> 0:00:55.760 you to know where I'm getting all of this from, 0:00:55.920 --> 0:00:59.400 versus just assuming that I made it up somehow. That 0:00:59.480 --> 0:01:02.120 is an accus I've had made. I don't really know 0:01:02.680 --> 0:01:05.679 how I would do that anyway, the episode I'm very sorry. 0:01:06.720 --> 0:01:09.240 At the tail end of June, Goldman Sachs, one of 0:01:09.240 --> 0:01:12.320 the largest global investment banks put out the thirty one 0:01:12.360 --> 0:01:16.399 page report titled JENAI too much, spend too little benefit 0:01:16.800 --> 0:01:19.040 that's with the question mark at the end. That includes 0:01:19.080 --> 0:01:21.400 some of the most damning literature on generative AI that 0:01:21.480 --> 0:01:25.399 I've ever seen. And yeah, that weep sound in the background. 0:01:25.480 --> 0:01:28.160 You here is the slow, painful deflation of the bubble 0:01:28.200 --> 0:01:31.839 I've been warning you about since March. The report covers 0:01:31.840 --> 0:01:36.600 AI's productivity benefits, which Goldman remarks are likely limited, AI's returns, 0:01:36.640 --> 0:01:39.600 which are likely to be significantly more limited than anticipated, 0:01:39.800 --> 0:01:43.080 and AI's power demands, which are likely so significant that 0:01:43.200 --> 0:01:46.160 utility companies will have to spend nearly forty percent more 0:01:46.360 --> 0:01:48.120 in the next three years to keep up with the 0:01:48.160 --> 0:01:51.760 demand from hyperscalers and rot economists like Google and Microsoft. 0:01:53.240 --> 0:01:56.280 The report is so significant because Goldman sacks, like any 0:01:56.320 --> 0:02:00.000 investment bank doesn't care about your feelings or your happy 0:02:00.360 --> 0:02:04.640 or emotions or anything unless doing so is profitable. It'll 0:02:04.640 --> 0:02:07.120 gladly hypob anything it thinks it will make a bug. 0:02:07.680 --> 0:02:10.120 Back in May, it published a report which claimed that AI, 0:02:10.360 --> 0:02:13.600 not just generative AI, was showing very positive signs of 0:02:13.639 --> 0:02:18.000 eventually boosting GDP and productivity. Even though said report buried 0:02:18.000 --> 0:02:21.240 within it, constant reminders that AI had yet to impact 0:02:21.280 --> 0:02:24.720 productivity growth and states that only five percent of companies 0:02:24.919 --> 0:02:29.040 report using generative AI in regular production. For Goldman to 0:02:29.120 --> 0:02:32.799 suddenly turn on the AI movement suggests that it's extremely 0:02:32.880 --> 0:02:36.399 anxious about the future of generative AI, with almost everybody 0:02:36.440 --> 0:02:39.560 agreeing on one core point in the report that the 0:02:39.639 --> 0:02:42.960 longer the generative AI takes to make people money, the 0:02:43.000 --> 0:02:46.160 more money that it's going to need to make. The 0:02:46.240 --> 0:02:49.760 report also includes an interview with economist Darren Smoglu of MIT, 0:02:50.160 --> 0:02:51.760 which can be found by the way on page four 0:02:51.760 --> 0:02:53.560 of the document, which you can find in this episode 0:02:53.639 --> 0:02:57.240 spreadsheet of links. An Institute professor who published a paperback 0:02:57.280 --> 0:03:00.600 in May called the Simple Macroeconomics of AI that argued 0:03:00.639 --> 0:03:03.880 that and I quote the upside to US productivity and 0:03:04.080 --> 0:03:07.880 consequently GDP growth from genera IVAI will likely prove much 0:03:07.919 --> 0:03:11.680 more limited than many forecasts that expect a month has 0:03:11.760 --> 0:03:16.200 only made Asmoglue more pessimistic, declaring that truly transformative changes 0:03:16.240 --> 0:03:19.800 won't happen quickly and few, if any, will likely occur 0:03:19.919 --> 0:03:23.440 within the next ten years, and that genera ivai's ability 0:03:23.480 --> 0:03:27.440 to affect global productivity is low because and I quote again, 0:03:27.800 --> 0:03:31.440 many of the tasks that humans currently perform are multifaceted 0:03:31.480 --> 0:03:34.920 and require real world interaction, which AI won't be able 0:03:34.960 --> 0:03:39.720 to materially improve anytime soon. What makes this interview, and 0:03:39.960 --> 0:03:44.000 really the whole report so remarkable is how thoroughly and 0:03:44.040 --> 0:03:47.920 aggressively it attacks every bit of marketing collateral that the 0:03:48.040 --> 0:03:54.400 artificial intelligence movement has. Asmoglu specifically questions the belief that 0:03:54.440 --> 0:03:57.440 AI models will simply get more powerful as we throw 0:03:57.520 --> 0:04:00.920 more data and GPU graphics process unit the thing that 0:04:01.080 --> 0:04:05.000 crunches the numbers capacity at them, and specifically asks a question, 0:04:06.000 --> 0:04:10.040 what does it mean to double AI's capabilities? How does 0:04:10.080 --> 0:04:15.320 that make something like say a customer service rap better? Seriously, though, 0:04:15.840 --> 0:04:19.359 what does better really get them less errors? How do 0:04:19.400 --> 0:04:23.159 you quantify less errors without factoring in the current state 0:04:23.440 --> 0:04:27.159 introducing them? It's not great And this really is a 0:04:27.200 --> 0:04:32.680 specific problem with the whole GENII fantasists bullshit spiel. They 0:04:32.720 --> 0:04:35.000 heavily rely on the idea that not only will these 0:04:35.080 --> 0:04:38.880 large language models llms like chat GPT get more powerful, 0:04:39.200 --> 0:04:41.919 but the getting more powerful will somehow grant it the 0:04:41.920 --> 0:04:47.520 power to do something. As Asimoglu says, what does it 0:04:47.600 --> 0:04:52.120 mean the double it What does that do? No? Really, 0:04:52.160 --> 0:04:55.480 what does that more actually mean? We've heard people talk 0:04:55.560 --> 0:04:59.480 about this for eighteen months, saying the more powerful chat 0:04:59.520 --> 0:05:02.880 GBT gets, the better it will get. But what does 0:05:02.960 --> 0:05:06.760 better even look like? GPT four Oh, the latest version 0:05:06.800 --> 0:05:10.839 of chat GPT, other than accepting more inputs, is kind 0:05:10.880 --> 0:05:14.360 of the same thing. The capabilities have not really changed, 0:05:15.839 --> 0:05:19.680 and while one might argue that more powerful will mean 0:05:19.880 --> 0:05:23.960 faster generative processes, there really is no barometer for what 0:05:24.040 --> 0:05:27.960 better looks like, and perhaps that's why Chat GBT, Claude 0:05:27.960 --> 0:05:30.159 and other lms have yet to take a leap beyond 0:05:30.160 --> 0:05:33.160 being able to generate pictures of Garfield and Lingerie are 0:05:33.360 --> 0:05:37.359 Scooby Doo with a gun anthropics. Claude LM might be 0:05:37.520 --> 0:05:40.040 and I quote, best in class according to tech Crunch, 0:05:40.279 --> 0:05:42.800 but that only means that it's faster and more accurate, 0:05:42.839 --> 0:05:47.600 which is cool, but not really the future or revolutionary 0:05:47.680 --> 0:05:52.080 or necessarily good in all cases. I should add that 0:05:52.120 --> 0:05:54.880 these are questions that I and other people writing about 0:05:54.960 --> 0:05:57.560 AI kind of should have been asking the whole time. 0:05:58.560 --> 0:06:02.480 General FAI generates outputs based on text based inputs and requests, 0:06:02.480 --> 0:06:05.240 and eventually multimodal will mean be able to look at 0:06:05.320 --> 0:06:08.320 something and generate an answer too, and these requests can 0:06:08.360 --> 0:06:11.640 be equally specific and intricate. Yet the answer is always 0:06:11.680 --> 0:06:16.039 as obvious as it sounds generated fresh, meaning that there's 0:06:16.080 --> 0:06:19.839 no actual knowledge or indeed intelligence operating in any part 0:06:19.960 --> 0:06:23.440 of the process. As a result, it's easy to see 0:06:23.440 --> 0:06:30.320 how this gets better ish faster, but much much harder to, 0:06:30.520 --> 0:06:33.440 if not impossible, to see how generative AI leads any 0:06:33.560 --> 0:06:37.200 further than where we're already at. And by that, I 0:06:37.240 --> 0:06:40.960 mean what does it do other than what it's doing today? 0:06:40.960 --> 0:06:44.839 What does chat GBT do today? That's radically different to 0:06:44.880 --> 0:06:50.200 eighteen months ago. The jump between say, two generations of iPhones, 0:06:50.200 --> 0:06:52.600 from the first iPhone to what it would be the 0:06:52.640 --> 0:06:55.159 iPhone three g I think it would be someone's going 0:06:55.240 --> 0:06:57.480 to email and say I'm wrong, and I'll say to them, 0:06:57.520 --> 0:07:00.320 you're very rude anyway, But that jump was huge, which 0:07:00.480 --> 0:07:02.960 faster internet meant you could do more with the phone. 0:07:03.040 --> 0:07:05.880 There was the app store, These were new functionalities added 0:07:06.640 --> 0:07:09.400 two years after the first iPhone, two years after the 0:07:09.400 --> 0:07:13.360 first chat GBT, and uh, I guess, I guess you 0:07:13.360 --> 0:07:16.160 could talk to it in a response that's not really 0:07:16.480 --> 0:07:20.840 great How does GPT, a transformer based model that generates 0:07:20.880 --> 0:07:23.520 answers probabilistically, as in, what the next part of the 0:07:23.560 --> 0:07:26.440 generation is most likely to be based on what's been inputed, 0:07:27.160 --> 0:07:30.920 based entirely on training data, How does it do anything 0:07:31.000 --> 0:07:35.080 more than generate paragraphs of occasionally accurate text or images, 0:07:35.760 --> 0:07:38.520 maybe a picture of Scooby Doo in lingerie. I don't know. 0:07:38.800 --> 0:07:41.960 But how do any of these models even differentiate from 0:07:42.000 --> 0:07:43.760 each other when most of them are trained on the 0:07:43.800 --> 0:07:47.720 same training data that they're already running out of? And 0:07:47.760 --> 0:07:50.680 what's crazy? Is I mentioned that I should have asked this, 0:07:50.760 --> 0:07:53.040 I really should have every single episode of mentioned AI. 0:07:53.120 --> 0:07:55.360 What does better even look like? What does more powerful 0:07:55.400 --> 0:07:59.760 even look like? Because when we say better, does that 0:08:00.000 --> 0:08:04.400 mean faster? Does that mean quicker to generate things? I 0:08:04.400 --> 0:08:06.400 mean it can generate more things? And the answer is 0:08:06.480 --> 0:08:09.240 usually not. It's not that it can actually function in 0:08:09.240 --> 0:08:11.960 a different way. It's just that you can do more, 0:08:12.200 --> 0:08:15.040 it can grow more. Hey, remember the idea of growth 0:08:15.200 --> 0:08:20.360 or cost Jesus right, But seriously, though, this feels like 0:08:20.400 --> 0:08:23.480 the question to ask Sam Ultiman or Mirror Marati the 0:08:23.600 --> 0:08:27.920 CTO of open AI's like, what's next? What's next, Oh, 0:08:27.960 --> 0:08:30.120 you can generate videos? When can I use that? Cool? 0:08:30.320 --> 0:08:34.920 What does better look like? They're other than not horrible looking? Okay, 0:08:35.000 --> 0:08:38.480 but what what more can GPT do? Because this thing, 0:08:38.840 --> 0:08:42.360 this thing can't think it's generating stuff, It doesn't know anything. 0:08:42.679 --> 0:08:45.480 It has training data. It eats up and then craps 0:08:45.520 --> 0:08:48.400 out an answer. How is that going to lead to 0:08:48.480 --> 0:08:52.880 even automation? I just and then how do you even 0:08:52.960 --> 0:08:55.520 deal with the fact that it really is running out 0:08:55.559 --> 0:08:58.640 of training data? And I've argued before the training data 0:08:58.679 --> 0:09:00.920 crisis is one that does not get enough attention, But 0:09:00.960 --> 0:09:03.760 it's sufficiently dire now that it has the potential to 0:09:03.800 --> 0:09:07.000 halt or at least dramatically slow any AI development in 0:09:07.000 --> 0:09:10.480 the future, by which I mean generative AI. As one 0:09:10.520 --> 0:09:13.079 paper published in the Journal of Computer Vision and Pattern 0:09:13.160 --> 0:09:16.160 Recognition found, in order to achieve a linear improvement in 0:09:16.480 --> 0:09:20.520 model improvements, you need an exponentially large amount of data. 0:09:21.000 --> 0:09:24.520 So it's not like they read something and then work 0:09:24.600 --> 0:09:27.160 something out from there. They need so much more to 0:09:27.280 --> 0:09:31.439 learn one thing kind of right, it's not very good 0:09:31.960 --> 0:09:34.920 and maybe there's another way to put it. Each additional 0:09:35.000 --> 0:09:38.680 step of training becomes increasingly and exponentially more expensive to 0:09:38.679 --> 0:09:41.800 take two. It requires you to get a bunch of 0:09:41.880 --> 0:09:44.839 training data and process more of it, but it also 0:09:44.880 --> 0:09:50.000 requires extremely expensive technology GPUs and a ton of energy 0:09:50.040 --> 0:09:53.800 to actually do so, and this infers a steep financial cost, 0:09:53.920 --> 0:09:56.439 not merely in just obtaining the data, but also, as 0:09:56.480 --> 0:09:59.760 I mentioned, the compute required to process it. And Anthropics 0:10:00.000 --> 0:10:03.000 CEO Dario Amadel said that the AI models currently in 0:10:03.040 --> 0:10:05.760 development will cost as much as a billion dollars to train, 0:10:06.040 --> 0:10:08.560 and within the next three years we may see models 0:10:08.679 --> 0:10:12.600 that cost ten or one hundred billion dollars to train, 0:10:12.679 --> 0:10:15.640 which is just insane, as that's roughly three times the 0:10:15.679 --> 0:10:21.640 GDP of Estonia, a beautiful, quite cold country. Back to 0:10:21.720 --> 0:10:25.080 mister Darren Smoblo, who doubts that llms can even become 0:10:25.200 --> 0:10:28.880 super intelligent and that even his most conservative estimates of 0:10:28.920 --> 0:10:32.000 productivity gains and I quote may turn out to be 0:10:32.080 --> 0:10:35.079 too large if AI models prove less successful in improving 0:10:35.160 --> 0:10:38.920 upon more complex tasks, and I think that's really the 0:10:39.000 --> 0:10:43.480 root of the problem. All of this excitement, every second 0:10:43.520 --> 0:10:46.000 of breathless beat off hype has been built on this 0:10:46.160 --> 0:10:50.440 idea that the artificial intelligence industry, led by General AVAI, 0:10:50.760 --> 0:10:54.640 will somehow revolutionize and automate everything from robotics to the 0:10:54.679 --> 0:10:57.599 supply chain, despite the fact that general IVAI is not 0:10:57.760 --> 0:11:00.480 actually going to solve these problems because it is not 0:11:00.640 --> 0:11:03.800 built to do so. I willn't have a calculator to 0:11:03.880 --> 0:11:07.240 drive my fucking car Jesus Christ. And While Asamoglu may 0:11:07.240 --> 0:11:09.800 have some positive things to say, for example, that AI 0:11:09.840 --> 0:11:12.240 models could be trained to help scientists conceive of and 0:11:12.280 --> 0:11:15.920 test new materials, which actually already happened thanks to Google 0:11:15.960 --> 0:11:20.560 DeepMind researchers, his general verdict is kind of harsh that 0:11:20.679 --> 0:11:24.000 using generative AI and quote too much automation too soon 0:11:24.280 --> 0:11:26.760 could create bottlenecks and other problems for firms that no 0:11:26.840 --> 0:11:30.480 longer have the flexibility in troubleshooting capabilities that human capital provides. 0:11:31.080 --> 0:11:34.520 In essence, replacing humans with AI might break everything. If 0:11:34.520 --> 0:11:36.920 you're one of those bosses that doesn't actually know what 0:11:36.960 --> 0:11:40.440 the fuck it is they're talking about, But you know what, 0:11:41.040 --> 0:11:44.680 I'm sure the following advertisements from people who know exactly 0:11:44.840 --> 0:11:55.640 what they're talking about, and we're back. The report also 0:11:55.679 --> 0:11:58.640 includes a palate cleanser for the quirked up AI hype fiends, 0:11:59.000 --> 0:12:01.720 and you'll find it on page where Goldman sax As. 0:12:01.800 --> 0:12:04.840 Joseph Briggs argues that generative AI will and I quote 0:12:05.120 --> 0:12:08.120 likely lead to significant economic upside based and I shit 0:12:08.240 --> 0:12:11.440 you not entirely on the idea that AI will replace 0:12:11.520 --> 0:12:13.600 workers in some jobs and then allow them to get 0:12:13.679 --> 0:12:17.000 jobs in other fields. Briggs also argues that, and I 0:12:17.080 --> 0:12:21.280 quote again, the full AI automation of AI exposed tasks 0:12:21.320 --> 0:12:23.800 that are likely to occur over a longer horizon could 0:12:23.840 --> 0:12:28.040 generate significant cost savings, which assumes that generative AI or 0:12:28.040 --> 0:12:30.920 AI itself will actually replace these tasks. But also, that's 0:12:30.960 --> 0:12:34.560 such a funny thing to say, Hey, you know, auto 0:12:34.600 --> 0:12:39.280 mating tasks over a long time could save money. Fucking hell, 0:12:39.320 --> 0:12:42.200 I should get a job at Goldmann Sex anyway, anyway. Sorry. 0:12:42.440 --> 0:12:45.120 I should also add that, unlike every single other interview 0:12:45.160 --> 0:12:49.199 in the report, Briggs continually mixes up AI and generative AI, 0:12:49.480 --> 0:12:52.480 and at one point suggests that recent generative AI advances 0:12:52.480 --> 0:12:57.800 are foreshadowing the emergence of a superintelligence. This is like 0:12:57.880 --> 0:13:01.480 suggesting that because I squeeze my washing up liquid and 0:13:01.480 --> 0:13:03.920 it's sprayed across the wall, that I'm one step close 0:13:03.960 --> 0:13:08.160 to becoming Picasso. I included this part of the report 0:13:08.160 --> 0:13:11.880 because sometimes, very rarely I get somebody suggesting that I'm 0:13:11.920 --> 0:13:15.520 not considering both sides. The reason I don't generally include 0:13:15.520 --> 0:13:17.760 both sides of this argument is that the AI hype 0:13:17.760 --> 0:13:20.760 side generally makes arguments based on the assumption that things 0:13:20.800 --> 0:13:24.800 will happen, such as a transformer model that probabilistically generates 0:13:24.840 --> 0:13:27.640 the next part of a sentence or a picture somehow 0:13:27.679 --> 0:13:33.320 gaining sentience. I wish my calculator gained sentience. During high school, 0:13:33.360 --> 0:13:35.040 I was a very lonely child. I did not have 0:13:35.120 --> 0:13:37.040 many friends. But no, all it would do is just 0:13:37.080 --> 0:13:39.160 let me kind of make it look like it said boobs. 0:13:39.360 --> 0:13:41.600 Then I'd get in a lot of trouble. But anyway. 0:13:42.440 --> 0:13:45.760 Francois Chole, an AI researcher at Google, recently argued that 0:13:45.880 --> 0:13:49.680 large language models like CHET GPT can't lead to average 0:13:49.720 --> 0:13:54.720 general intelligence, the sentient AI that everyone's excited about, explaining 0:13:54.760 --> 0:13:57.240 in detail in an interview with a podcast that I've 0:13:57.280 --> 0:14:00.640 linked in there, that models like GPT A simply not 0:14:00.760 --> 0:14:03.880 capable of the kind of reasoning and theorizing that makes 0:14:03.880 --> 0:14:08.360 a human brain work. Schley also argues that even models 0:14:08.400 --> 0:14:11.520 specifically built to complete the tasks of his abstraction and 0:14:11.559 --> 0:14:14.120 reasoning corpus a benchmark test for AI skills and true 0:14:14.120 --> 0:14:16.920 intelligence that he invented are only doing so because they've 0:14:16.960 --> 0:14:19.800 been fed millions of data points of people solving the test, 0:14:20.640 --> 0:14:23.200 which is kind of like measuring somebody's IQ based on 0:14:23.240 --> 0:14:27.040 them studying really hard to complete an IQ test except 0:14:27.040 --> 0:14:30.200 even dumber. But the reason that I'm suddenly bringing up 0:14:30.240 --> 0:14:35.120 these superintelligences or AGI artificial general intelligence average general intelligence, 0:14:35.360 --> 0:14:38.280 lots of people say different things. The reason I'm saying 0:14:38.280 --> 0:14:41.160 it is because throughout every single defense of generative AI 0:14:41.640 --> 0:14:44.640 is a really nasty, deliberate attempt to get around the 0:14:44.720 --> 0:14:49.720 problem that GENERATIVAI doesn't actually automate many tasks. While it's 0:14:49.720 --> 0:14:53.240 good at generating answers, sometimes even correct ones, or at 0:14:53.280 --> 0:14:56.000 creating things based on a request, sometimes with the right 0:14:56.080 --> 0:14:59.440 number of fingers, there's no real interaction with the task 0:14:59.560 --> 0:15:02.640 or the perse and giving the task, or considering what 0:15:02.760 --> 0:15:05.320 the task needs at all. Just the abstraction of things 0:15:05.360 --> 0:15:08.600 said to output generated orbit in quite a complex way. 0:15:09.400 --> 0:15:11.760 Tasks like taking someone's order and relaying it to the 0:15:11.840 --> 0:15:14.680 kitchen at a fast food restaurant might seem elementary to 0:15:14.720 --> 0:15:17.200 most people and I won't write easy, because working in 0:15:17.240 --> 0:15:20.480 fast food is a hard and horrible job. It might 0:15:20.560 --> 0:15:23.560 seem elementary, though, but it isn't for an AI model 0:15:23.560 --> 0:15:26.480 that generates answers without really understanding the meaning of any 0:15:26.560 --> 0:15:28.040 of the words. And really, I shouldn't have said the 0:15:28.040 --> 0:15:32.000 word really that it doesn't understand anything. Last year, Wendy's, 0:15:32.000 --> 0:15:34.520 a burger chain here in America, announced that it would 0:15:34.520 --> 0:15:38.200 integrate its generative Fresh AI ordering system into some restaurants. 0:15:38.840 --> 0:15:41.520 In late June, it revealed that the system requires human 0:15:41.520 --> 0:15:45.680 intervention on fourteen percent of the orders. On one redditor's post, 0:15:45.920 --> 0:15:49.160 they noted that Wendy's AI regularly required three attempts to 0:15:49.160 --> 0:15:51.280 get it to understand them, and would sometimes cut you 0:15:51.320 --> 0:15:54.400 off if you weren't speaking fast enough. Hey, it's two, 0:15:54.440 --> 0:15:58.120 am I connecting the words chicken sandwich, the diet codes 0:15:58.200 --> 0:16:02.840 difficult for me. Another burger place here in America, White Castle, 0:16:03.520 --> 0:16:07.040 which implemented a similar system in partnership with Samsung and SoundHound, 0:16:07.400 --> 0:16:09.960 fared a little better, with a remarkable ten percent of 0:16:10.040 --> 0:16:14.600 orders requiring human intervention. Last month, McDonald's discontinued its own 0:16:14.680 --> 0:16:17.280 AI ordering system, which it built with IBM and deployed 0:16:17.520 --> 0:16:21.080 to more than one hundred restaurants, likely because it wasn't 0:16:21.200 --> 0:16:25.000 very good, with one customer rang up for literally hundreds 0:16:25.000 --> 0:16:28.240 of chicken nuggets like the I think you should leave sketch. However, 0:16:28.280 --> 0:16:31.560 to be clear, McDonald's system wasn't actually based on GENERATIVAI, 0:16:32.160 --> 0:16:34.840 and if nothing else, all of these examples illustrate this 0:16:34.880 --> 0:16:38.840 disconnect between those building AI systems and how much, or 0:16:39.000 --> 0:16:42.160 really how little they understand the jobs they wish to eliminate. 0:16:43.000 --> 0:16:47.400 Little humility or doing a real job goes a long way. 0:16:48.240 --> 0:16:50.560 Another thing to note is that, on top of generative 0:16:50.600 --> 0:16:54.000 AI generally cocking up these orders, Whendy still requires human 0:16:54.000 --> 0:16:56.800 beings to make the fucking food. Despite all of this hype, 0:16:56.880 --> 0:17:00.000 all of this media attention, all of this incredible investment, 0:17:00.200 --> 0:17:03.560 the supposed innovations don't even seem capable of replacing the 0:17:03.640 --> 0:17:06.439 jobs that all of these horny capitalists have been planning 0:17:06.480 --> 0:17:09.440 for them to do. Not that they think they should, 0:17:09.480 --> 0:17:12.760 just I'm being told that this future is inevitable or indeed, 0:17:12.920 --> 0:17:16.720 here it's starting to really accept what the problem is 0:17:16.760 --> 0:17:20.280 with generative AI, though it isn't good at replacing the 0:17:20.359 --> 0:17:23.960 kind of jobs that actually affect the economy but commoditizing 0:17:24.040 --> 0:17:27.480 distinct acts of labor, and in the process the early 0:17:27.520 --> 0:17:30.359 creative jobs that help people build portfolios to advance in 0:17:30.359 --> 0:17:35.080 their industries. The freelancers having their livelihoods replaced by bosses 0:17:35.160 --> 0:17:38.680 using generative AI aren't being replaced so much as they're 0:17:38.720 --> 0:17:41.800 being shown how little respect many bosses have for their 0:17:41.800 --> 0:17:46.000 craft or the customer they allegedly serve. Copy editors and 0:17:46.040 --> 0:17:49.120 concept artists provide far more valuable work than any generative 0:17:49.119 --> 0:17:52.680 AI can yeat. An economy dominated by managers who don't 0:17:52.680 --> 0:17:55.840 appreciate or participate in labor means that these jobs are 0:17:55.840 --> 0:17:59.119 constantly under assault from large language models pumping out stuff 0:17:59.160 --> 0:18:01.240 that all looks and it sounds the same, to the 0:18:01.240 --> 0:18:03.680 point that the BBC reports that copywriters are now being 0:18:03.680 --> 0:18:06.479 paid to help them make the ais sound more human. 0:18:08.440 --> 0:18:12.160 One of the most fundamental misunderstandings of the bosses replacing 0:18:12.160 --> 0:18:14.560 these workers with generative AI is that you are not 0:18:14.800 --> 0:18:17.760 just asking for a thing, but outsourcing the risk and 0:18:17.800 --> 0:18:21.639 responsibility for delivering it. When I hire an artist to 0:18:21.640 --> 0:18:24.159 make a logo, my expectation is that they'll listen to 0:18:24.200 --> 0:18:26.280 me theyn add their own flare then will go back 0:18:26.320 --> 0:18:28.639 and forth with drafts until we have something I like 0:18:28.680 --> 0:18:31.919 and that they're proud of too. I'm paying them not 0:18:32.040 --> 0:18:35.400 just for their time, but for their years learning their 0:18:35.440 --> 0:18:39.080 craft and the output itself, and so the ultimate burden 0:18:39.160 --> 0:18:42.200 of production is not just my own, and that their 0:18:42.240 --> 0:18:45.040 experience means that they can adapt to circumstances that I 0:18:45.160 --> 0:18:48.359 might not have thought of. These are not things that 0:18:48.400 --> 0:18:50.679 you can train in a data set, because they're derived 0:18:50.680 --> 0:18:55.240 from experiences inside and outside of the creative process. While 0:18:55.240 --> 0:18:57.840 one can teach a generative AI what a billion images 0:18:57.880 --> 0:19:00.800 look like, AI doesn't get handcrams or a call at 0:19:00.800 --> 0:19:04.159 eight pm saying that something needs to pop more. It 0:19:04.160 --> 0:19:06.879 doesn't have moods, nor can it infer them from written 0:19:06.920 --> 0:19:10.760 or visual media, because human emotions are extremely weird, as 0:19:10.760 --> 0:19:14.640 our moods, our bodies, and our general existences. We're disgusting 0:19:14.680 --> 0:19:18.399 and weird and beautiful. And I realize all of this 0:19:18.440 --> 0:19:21.520 is a little flowery, but even the most mediocre copy 0:19:21.800 --> 0:19:25.040 ever written is on some level a collection of experiences, 0:19:25.640 --> 0:19:29.199 and fully replacing any creative is so very unlikely if 0:19:29.240 --> 0:19:31.399 you're doing so based on copying a million pieces of 0:19:31.440 --> 0:19:35.800 someone else's homework. Now, if I was looking for something creative, 0:19:36.359 --> 0:19:39.160 I get it from one of the following advertisements, which 0:19:39.200 --> 0:19:41.880 I am sure aligned perfectly with the things that I'm 0:19:41.920 --> 0:19:44.800 talking about. Won't make me look weird, won't have people 0:19:44.880 --> 0:19:48.760 email me. No, you're going to love the advertisements, and 0:19:48.800 --> 0:20:01.360 then you're not going to email me about them. All Right, 0:20:01.400 --> 0:20:04.879 we're back in the room. The most fascinating part of 0:20:04.920 --> 0:20:07.360 the Goldman Sex Report, and you'll find it on page ten, 0:20:07.720 --> 0:20:10.120 is an interview with Jim Cavello Goldman sax Is head 0:20:10.119 --> 0:20:13.280 of Global Equity Research. Cavello isn't a name you'll have 0:20:13.320 --> 0:20:16.560 heard unless you are, for whatever reason, a big semiconductor head, 0:20:16.960 --> 0:20:19.439 but he's consistently been on the right side of history, 0:20:19.640 --> 0:20:22.440 named as one of the top semiconductor analysts by LI 0:20:22.680 --> 0:20:26.320 Research for years, successfully catching the downturn in fundamentals in 0:20:26.400 --> 0:20:29.800 multiple major chip firms far before others did, at times 0:20:29.880 --> 0:20:33.000 mocked for quote being wrong and then turning out to 0:20:33.040 --> 0:20:38.359 be so very very right. And Jim, Jim Cavello, in 0:20:38.440 --> 0:20:42.840 no uncertain terms, thinks that GENERATIVEAI the whole bubble, this 0:20:42.960 --> 0:20:48.840 whole generative monstrosity kind of for shit. Cavello believes that 0:20:48.880 --> 0:20:51.879 the combined expenditure of all parts of the GENERATIVEAI boom, 0:20:52.040 --> 0:20:56.560 data centers, utilities, applications will cost a trillion dollars in 0:20:56.600 --> 0:20:59.480 the next several years alone, and he asks one very 0:20:59.520 --> 0:21:03.399 simple quest question, what trillion dollar problem will AI solve? 0:21:04.200 --> 0:21:07.520 He notes that replacing low wage jobs with tremendously costly 0:21:07.680 --> 0:21:11.119 technology is basically the polar opposite of the prior technology 0:21:11.119 --> 0:21:14.679 transitions that he's witnessed in the last thirty years. Just 0:21:14.720 --> 0:21:17.880 to be consistent with my AI to national GDP rubric, 0:21:18.160 --> 0:21:20.959 one trillion dollars is roughly half the GDP of Italy 0:21:21.200 --> 0:21:23.720 or the entire Danish economy multiplied by two and a 0:21:23.760 --> 0:21:27.880 half times. Yes, Denmark an EU and NATO member state 0:21:27.920 --> 0:21:30.600 that gave the world, among other things, a zempic, Lego 0:21:30.640 --> 0:21:33.760 and Aqua, the artist behind nineteen ninety seven Saccharin song 0:21:33.800 --> 0:21:38.600 of the Year Barbie Girl, one amazing country, and apparently 0:21:38.640 --> 0:21:42.199 their GDP is just a footnote in comparison to how 0:21:42.280 --> 0:21:46.000 much we have to spend on generating pictures of well 0:21:46.359 --> 0:21:49.000 Garfield in lingerie shooting at Scooby Doo. At this point, 0:21:49.080 --> 0:21:52.280 I'm just going to keep escalating one particular myth that 0:21:52.320 --> 0:21:55.760 Cavello dispels, and this was my favorite part is that 0:21:55.800 --> 0:21:58.760 many people compare generative AI to the early days of 0:21:58.760 --> 0:22:00.760 the Internet, and he knows so. It's that, even if 0:22:00.800 --> 0:22:03.440 it's in its infancy, the Internet was a low cost 0:22:03.440 --> 0:22:06.600 technology solution that enabled things like e commas to replace 0:22:06.800 --> 0:22:11.400 costly incumbent solutions, and that AI technology is exceptionally expensive, 0:22:11.440 --> 0:22:14.560 and to justify those costs, the technology must be able 0:22:14.600 --> 0:22:17.360 to solve complex problems, which it isn't designed to do. 0:22:17.440 --> 0:22:19.840 And that is a quote, by the way, hello's a beast. 0:22:20.560 --> 0:22:23.600 He also dismisses the suggestion that tech starts off expensive 0:22:23.640 --> 0:22:26.440 and gets cheaper over time as revisionist history, and that's 0:22:26.480 --> 0:22:29.280 a quote, and that quote again, the tech world is 0:22:29.320 --> 0:22:32.199 too complacent in the assumption that AI costs will decline 0:22:32.240 --> 0:22:36.200 substantially over time. He specifically notes that the only reason 0:22:36.240 --> 0:22:39.280 that Moore's law was capable of enabling smaller, faster, jeeper 0:22:39.320 --> 0:22:43.080 chips was because competitors like AMD forced Intel and other 0:22:43.160 --> 0:22:46.560 companies to compete, a thing that doesn't really seem to 0:22:46.600 --> 0:22:49.359 be happening with Nvidia, which is a near stranglehold on 0:22:49.400 --> 0:22:53.639 GPUs required to handle generative AI. And indeed, all of 0:22:53.640 --> 0:22:57.280 these other companies aren't really making competitive products. They're just 0:22:57.359 --> 0:23:00.360 kind of making the same thing. Google has their own 0:23:00.600 --> 0:23:05.080 video and text and image generator, sodas Lama, the Metawe, 0:23:05.280 --> 0:23:09.199 sodas chat GPT. It's all the same. There's no competition here. 0:23:10.119 --> 0:23:13.440 Kind of look a a cartail. Almost good idea for 0:23:13.480 --> 0:23:16.880 an episode. And while there are companies making these graphics 0:23:16.920 --> 0:23:19.960 process units aimed at the AI market, especially in China, 0:23:20.000 --> 0:23:23.160 where US trade restrictions prevent local companies from buying high 0:23:23.200 --> 0:23:27.040 powered cards like the A one hundred and video makes 0:23:27.040 --> 0:23:29.160