WEBVTT - The Case Against Generative AI (Part 1) 0:00:02.440 --> 0:00:06.280 Zone Media, Hello, and welcome to Better Offline. I am, 0:00:06.360 --> 0:00:20.360 of course your host ed zitron what and after a 0:00:20.360 --> 0:00:22.600 few three part episodes, I had an idea, what if 0:00:22.640 --> 0:00:25.200 I did a four parter? In all seriousness, I know 0:00:25.239 --> 0:00:27.320 that this is a little bit long, but the topic 0:00:27.360 --> 0:00:29.560 we're about to explore demand's quite a bit of depth 0:00:29.600 --> 0:00:31.600 and it isn't something I could really do justice to 0:00:31.760 --> 0:00:34.240 in a one part or two parts, or I guess 0:00:34.280 --> 0:00:37.760 even three parter, but let's get into him. Over the 0:00:37.840 --> 0:00:40.760 last few months, we felt the vibes shift downward in 0:00:40.800 --> 0:00:43.720 an aggressive way, with both Marg Zuckerberg and Clammy Sam 0:00:43.760 --> 0:00:46.480 Mortman saying that we're in a bubble. In the latter 0:00:46.560 --> 0:00:48.960 case said, warnings of a bubble are always couched in 0:00:49.000 --> 0:00:52.159 rank hypocrisy is it's always implied that whoever it is 0:00:52.240 --> 0:00:54.880 and the companies they represent aren't part of that bubble, 0:00:54.920 --> 0:00:58.600 but rather it's other people and other companies making unfortunate decisions. 0:00:59.080 --> 0:01:02.160 The thing is, there's really no escape for either of 0:01:02.200 --> 0:01:05.360 these guys, not for Zaken, definitely not for Sam Ortmon. 0:01:05.760 --> 0:01:07.760 And over the next four episodes, I'm going to make 0:01:07.760 --> 0:01:09.760 a comprehensive case for the fact that we're in a 0:01:09.800 --> 0:01:13.160 bubble and condense everything I've been talking about into one series. 0:01:13.600 --> 0:01:15.280 And I know I've been all over the place, and 0:01:15.319 --> 0:01:17.080 I get a lot of people saying, oh, well, where 0:01:17.080 --> 0:01:18.959 did you talk about this? And where'd you talk about that? 0:01:19.520 --> 0:01:21.520 And that's kind of fair when you put out as 0:01:21.560 --> 0:01:23.800 much as I do. But I'm going to break this 0:01:23.880 --> 0:01:25.640 down in four episodes. I'm going to give you a 0:01:25.720 --> 0:01:29.319 comprehensive argument against the bubble. Well I mean that for 0:01:29.440 --> 0:01:31.679 a bubble, I guess, but against generative AI in general. 0:01:31.720 --> 0:01:34.040 But in this episode, I think it's good to start 0:01:34.080 --> 0:01:36.520 from the beginning and work our way forward to track 0:01:36.560 --> 0:01:38.920 the thread from the origins of chat GPT to the 0:01:38.920 --> 0:01:41.880 billion spurned building data centers all over the world and 0:01:41.920 --> 0:01:44.880 the weak business justifications for burning and nearly a trillion 0:01:44.920 --> 0:01:48.360 dollars to keep this hollow industry alive. Now, in twenty 0:01:48.400 --> 0:01:51.400 twenty two, a kind of company called open Ai surprise 0:01:51.440 --> 0:01:53.880 the world with a website called chat GPT. They could 0:01:53.920 --> 0:01:56.800 generate text that sort of sounded like a person using 0:01:56.800 --> 0:02:00.360 a technology called large language models LLMS, which can also 0:02:00.400 --> 0:02:02.760 be used to generate images, video, and computer code, or 0:02:02.760 --> 0:02:06.680 at least would eventually last. Language models require entire clusters 0:02:06.680 --> 0:02:09.880 of servers connected with high speed networking or containing this 0:02:09.919 --> 0:02:13.880 thing called a GPU. Graphics processing units. These are difference 0:02:13.919 --> 0:02:16.640 to the GPUs in your xbox or laptop or gaming PC. 0:02:17.000 --> 0:02:19.040 They cost much much more, and they're good at doing 0:02:19.120 --> 0:02:21.600 the processes of inference, the creation of an output of 0:02:21.639 --> 0:02:24.720 any LLM and training, feeding masses of training data to 0:02:24.760 --> 0:02:27.000 the models, or feeding them information about what a good 0:02:27.080 --> 0:02:29.320 output might look like so they can later identify a 0:02:29.360 --> 0:02:33.040 thing or replicate it. These models showed some immediate promise 0:02:33.080 --> 0:02:36.280 in their ability to articulate concepts or generate video visuals, 0:02:36.280 --> 0:02:39.800 audio text, and code. They also immediately had one glaring 0:02:39.880 --> 0:02:44.040 obvious problem because their propabilistic, meaning that they're just guessing 0:02:44.080 --> 0:02:46.920 whatever the right output might be. These models can't actually 0:02:46.919 --> 0:02:49.919 be relied upon to do exactly the same thing every 0:02:49.960 --> 0:02:53.200 single time. So if you generated a picture of a 0:02:53.200 --> 0:02:56.400 person that you wanted to, for example, using this story book, 0:02:56.600 --> 0:02:58.720 every time you created a new page using the same 0:02:58.760 --> 0:03:01.320 prompt to describe the protragton, the person would look different, 0:03:01.560 --> 0:03:03.600 and that difference could be minor of something that a 0:03:03.639 --> 0:03:05.640 reader could shrug off, or it could make the character 0:03:05.680 --> 0:03:08.440 look like a completely different person. Now None of this, 0:03:08.560 --> 0:03:11.520 by the way, is me validating or saying that any 0:03:11.560 --> 0:03:14.880 of this stuff is good. I'm just describing him. Moreover, 0:03:14.960 --> 0:03:17.760 the probabilistic nature of generative AI meant that whenever you 0:03:17.800 --> 0:03:20.000 asked it a question, it would guess as to the answer, 0:03:20.040 --> 0:03:21.920 not because it knew the answer, but rather because it 0:03:21.960 --> 0:03:23.480 was guessing on the right word to add in a 0:03:23.560 --> 0:03:26.920 sentence based on previous training data. As a result, these 0:03:26.960 --> 0:03:30.200 models would frequently make mistakes, something which we later referred 0:03:30.240 --> 0:03:33.440 to as hallucinations. And that's not even mentioning the cost 0:03:33.480 --> 0:03:35.520 of training these models, the cost of running them, the 0:03:35.600 --> 0:03:38.720 vast amounts of computational power they required, the fact that 0:03:38.760 --> 0:03:41.080 the legality of using materials straight from books in the 0:03:41.080 --> 0:03:44.400 web without the owner's permission was and remains legally dubious, 0:03:44.640 --> 0:03:46.240 or the fact that nobody seemed to know how to 0:03:46.320 --> 0:03:50.400 use these models that actually create profitable businesses. These problems 0:03:50.400 --> 0:03:53.160 were overshadowed by something flashy and new, and something that 0:03:53.200 --> 0:03:56.120 investors and the tech media believed would eventually aut to 0:03:56.160 --> 0:03:59.800 make jobs that have proven most resistance towardstion knowledge work 0:03:59.840 --> 0:04:03.320 and the creative economy. The newness and hyper and these 0:04:03.400 --> 0:04:06.040 expectations sent the market into a frenzy, with every hyper 0:04:06.040 --> 0:04:09.560 scaler immediately creating the most aggressive market for one supplier 0:04:09.600 --> 0:04:12.840 I've ever seen. Nvidia has sold over two hundred billion 0:04:12.840 --> 0:04:15.240 dollars of GPUs since the beginning of twenty twenty three, 0:04:15.480 --> 0:04:18.440 becoming the largest company on the American stock market and 0:04:18.480 --> 0:04:20.920 trading over one hundred and seventy dollars as of writing 0:04:20.920 --> 0:04:23.119 this sentence, only a few years off to being worth 0:04:23.440 --> 0:04:26.560 nineteen dollars and fifty two cents a share. Now there's 0:04:26.560 --> 0:04:29.039 a stock split that happened there, but it works out 0:04:29.040 --> 0:04:31.400 that way. Now. While I've talked about some of the 0:04:31.440 --> 0:04:35.120 propelling factors behind the AI wave, automation and novelty, that's 0:04:35.160 --> 0:04:38.320 not really the complete picture. A huge reason why everybody 0:04:38.360 --> 0:04:41.719 decided to do AI was because the software industry's growth 0:04:41.800 --> 0:04:45.240 was slowing and SaaS software as a service company valuations 0:04:45.720 --> 0:04:48.880 were stalling or dropping, resulting in the terrifying prospect of 0:04:48.880 --> 0:04:52.240 companies having to underpromise and over deliver and be efficient, 0:04:52.320 --> 0:04:56.680 you know, gross things like running sustainable businesses. Things that 0:04:56.839 --> 0:04:59.919 normal companies, those whose valuations aren't contingent on ever increase, 0:05:00.120 --> 0:05:02.360 ever constant growth, don't have to worry about because they're 0:05:02.400 --> 0:05:06.760 normal companies. Suddenly there was a new promise of new technology, 0:05:06.880 --> 0:05:10.279 large language models that were getting exponentially more powerful, which 0:05:10.800 --> 0:05:13.320 was mostly a lie but hard to disprove because powerful 0:05:13.360 --> 0:05:16.719 can mean basically anything, and the definition of powerful depended 0:05:16.880 --> 0:05:19.400 entirely on whoever you asked at any given time and 0:05:19.400 --> 0:05:23.560 what that person's motivations were. The media also immediately started 0:05:23.560 --> 0:05:26.719 tripping over its own feet, mistakenly claiming open ais GPT 0:05:26.880 --> 0:05:29.320 form model tricked a task grab it into solving a capture. 0:05:29.480 --> 0:05:32.120 It didn't. This never happened, or saying that, and I 0:05:32.240 --> 0:05:34.440 quote people who don't know how to code already used 0:05:34.440 --> 0:05:37.239 bots to produce full fledged games. And if you weren't 0:05:37.320 --> 0:05:39.640 wondering what the New York Times was referring to when 0:05:39.640 --> 0:05:42.480 they said full fledged, there it meant Pong and are 0:05:42.520 --> 0:05:45.719 coupled together rolling demo of Skyroads game from nineteen ninety three, 0:05:46.160 --> 0:05:48.440 likely because a bunch of that training data was fed 0:05:48.440 --> 0:05:51.760 into the models. Now, the media and investors helped pedal 0:05:51.800 --> 0:05:54.520 the narrative that AI was always getting better, could basically 0:05:54.560 --> 0:05:56.880 do anything, and that any problems you saw today would 0:05:56.880 --> 0:06:00.000 inevitably be solved in a few short months or years. 0:06:00.200 --> 0:06:03.720 Or that some point. I guess not really sure when 0:06:03.760 --> 0:06:06.440 that point is, but damn do they think it's coming. 0:06:06.960 --> 0:06:09.600 And llms were touted as a kind of digital panacea, 0:06:09.680 --> 0:06:12.400 and the companies building them off a traditional software companies 0:06:12.400 --> 0:06:15.000 the chance to plug these models into their software using 0:06:15.000 --> 0:06:18.040 an API, thus allowing them to write the same generative 0:06:18.040 --> 0:06:21.440 AI wave that every other company was riding. The model 0:06:21.480 --> 0:06:25.040 companies similarly started going after individual and business customers, offering 0:06:25.040 --> 0:06:28.080 software and subscriptions that promised the world, though this mostly 0:06:28.120 --> 0:06:30.760 boiled down to chatbots that could generate stuff, and then 0:06:30.960 --> 0:06:33.880 doubled down with the promise of agents, a marketing term 0:06:33.880 --> 0:06:36.719 that's meant to make you think autonomous digital worker, but 0:06:36.920 --> 0:06:40.159 really means broken digital chatbot of some sort or just 0:06:40.240 --> 0:06:43.000 broken digital product. It really depends how you're feeling that day. 0:06:44.520 --> 0:06:47.000 Throughout this era, investors in the media spoke with a 0:06:47.040 --> 0:06:50.279 sense of inevitability that they never really backed up with data. 0:06:50.839 --> 0:06:54.320 It was an era based on confidently asserted vibes. Everything 0:06:54.440 --> 0:06:57.360 was always getting better and more powerful, even though there 0:06:57.360 --> 0:07:00.840 was never much proof that this was truly disruptive technology 0:07:01.120 --> 0:07:03.400 other than in its ability to disrupt apps you were 0:07:03.480 --> 0:07:07.080 using with AI, making them worse, For example, suggesting questions 0:07:07.080 --> 0:07:09.400 on every Facebook post that you could ask meta AI, 0:07:09.640 --> 0:07:13.240 but which METAAI couldn't answer. And I mean on memes, 0:07:13.760 --> 0:07:17.600 on just random posts. It's really not useful in any way, 0:07:17.600 --> 0:07:21.440 shape or form. AI became omnipresent, and it eventually grew 0:07:21.480 --> 0:07:24.040 to mean everything and nothing. Open AI would see. It's 0:07:24.040 --> 0:07:26.880 every move loaded over like a gifted child. It's CEO 0:07:26.960 --> 0:07:29.480 Sam Allman called the Oppenheimer of our age, even if 0:07:29.480 --> 0:07:32.840 it wasn't really obvious why everybody was impressed. GPT four 0:07:32.880 --> 0:07:36.000 felt like something a bit different, but was it actually meaningful? 0:07:36.400 --> 0:07:39.440 The thing is our official intelligence is built and sold 0:07:39.520 --> 0:07:41.600 or not just faith, but a series of myths that 0:07:41.640 --> 0:07:44.240 AI boosters expect us to believe, but the same certainty 0:07:44.280 --> 0:07:46.920 that we treat things like gravity or the boiling point 0:07:46.920 --> 0:07:51.200 of water. Can large language models actually replace coders? Not really, No, 0:07:51.320 --> 0:08:05.160 and I'll get into way later in this series. Consora 0:08:05.320 --> 0:08:08.640 Open AI's video creation tool replace actors or animators. No, 0:08:08.840 --> 0:08:10.560 not at all, But it still fills the air full 0:08:10.600 --> 0:08:13.440 of tension because you can immediately see who is preregistered 0:08:13.440 --> 0:08:17.120 to replace everyone that works for them. AI is apparently 0:08:17.160 --> 0:08:20.880 replacing workers, but nobody seems able to prove it at scale. 0:08:21.120 --> 0:08:23.680 But every few weeks a story runs where everybody tries 0:08:23.720 --> 0:08:25.960 to pretend that AI is replacing workers with some sort 0:08:26.000 --> 0:08:30.440 of poorly sourced and incomprehensible study, never actually saying somebody's 0:08:30.520 --> 0:08:33.520 job got replaced by AI, because it isn't happening at scale, 0:08:33.640 --> 0:08:36.080 and because if you provide real world examples, people can 0:08:36.120 --> 0:08:38.959 actually check if they're true. Now, I want to be clear, 0:08:39.360 --> 0:08:42.520 some people have lost jobs to AI, just not really 0:08:42.520 --> 0:08:45.280 white collar workers, software engineers, or really any of the 0:08:45.320 --> 0:08:48.000 career paths that the mainstream media and AI investors would 0:08:48.040 --> 0:08:51.520 have you believe. Brian Merchant has done excellent work covering 0:08:51.520 --> 0:08:54.640 how llms have devoured the work of translators, using cheap, 0:08:54.760 --> 0:08:57.720 almost good automation to lower already stagnant wages in a 0:08:57.720 --> 0:09:00.280 field that has already been hurting before the add to 0:09:00.440 --> 0:09:03.160 generative AI, with some having to abandon the field and 0:09:03.200 --> 0:09:06.560 others pushed into bankruptcy. I've heard the same for art directors, 0:09:06.720 --> 0:09:09.400 SEO experts, and copy editors, and Christopher Mims of The 0:09:09.400 --> 0:09:12.920 Wall Street Journal covered these last year. These fields all 0:09:12.920 --> 0:09:15.480 have something in common shitty bosses with little regard for 0:09:15.520 --> 0:09:18.080 their customers. Who have been eagerly waiting for the opportunity 0:09:18.080 --> 0:09:21.720 to slash labor to quote Merchant, the drum beat marketing 0:09:21.760 --> 0:09:24.720 and pop culture of powerful AI encourages and permits management 0:09:24.760 --> 0:09:27.040 to replace with the great jobs they might not otherwise have. 0:09:27.679 --> 0:09:30.440 Across the board, the people being replaced by AI are 0:09:30.440 --> 0:09:33.160 the victims of lazy, incompetent cost cutters who don't care 0:09:33.160 --> 0:09:36.520 if they ship poorly translated text to quote Merchant. Again, 0:09:36.920 --> 0:09:40.079 AI hypers created the cover necessary to justify slashing rates 0:09:40.080 --> 0:09:43.800 and accepting just good enough automation output for video games 0:09:43.800 --> 0:09:47.560 and media products. Yet the jobs crisis facing translators speaks 0:09:47.559 --> 0:09:49.760 to the larger flaws of the large language model era 0:09:49.960 --> 0:09:52.240 and why other careers aren't seeing this kind of disruption. 0:09:53.200 --> 0:09:57.000 Generative AI creates outputs, and by extension, defines all labour 0:09:57.040 --> 0:09:59.640 as some kind of output creative from a request. In 0:09:59.679 --> 0:10:02.040 the case translation, it's possible for a company to get 0:10:02.040 --> 0:10:04.600 by with a shitty version because many customers see translation 0:10:04.679 --> 0:10:07.280 as what do these words say, even though, as one 0:10:07.320 --> 0:10:12.040 worker told Brian Merchant, translation is about conveying meaning. Nevertheless, 0:10:12.120 --> 0:10:14.600 translation workers already started to condense to a world where 0:10:14.679 --> 0:10:17.320 humans would at times clean up machine generated text, and 0:10:17.360 --> 0:10:19.599 the same worker warned that the same might come for 0:10:19.720 --> 0:10:23.000 other industries. Yet the problem is that translation is a 0:10:23.040 --> 0:10:27.120 heavily output driven industry, one where idiot bosses can say, oh, yeah, 0:10:27.160 --> 0:10:29.520 that's fine because they ran an output back through Google 0:10:29.600 --> 0:10:32.480 Translate and it seemed fine in their native tongue. The 0:10:32.520 --> 0:10:35.360 problems of a poor translation are obvious, but customers of 0:10:35.400 --> 0:10:38.320 translation are it seems, often capable of getting by with 0:10:38.360 --> 0:10:41.120 a shitty product. The problem is that most jobs are 0:10:41.160 --> 0:10:43.400 not output driven at all, and what we're buying from 0:10:43.400 --> 0:10:45.920 a human beings a person's ability to think and do. 0:10:47.000 --> 0:10:50.640 Every CEO talking about replacing workers with AI is an 0:10:50.640 --> 0:10:53.319 example of the real problem that most companies are run 0:10:53.360 --> 0:10:56.360 by people who don't understand or experience the problems they're solving, 0:10:56.640 --> 0:10:59.280 don't do any real work, don't face any real problems, 0:10:59.280 --> 0:11:01.760 and thus can never be trusted to solve them. In 0:11:01.800 --> 0:11:04.040 the Era of the Business Idio, which is a piece 0:11:04.080 --> 0:11:05.880 I wrote a while ago, I talked about how this 0:11:06.000 --> 0:11:09.040 was the result of letting management consultants and neoliberal free 0:11:09.080 --> 0:11:12.600 market sociopaths take over everything, leaving us with companies run 0:11:12.640 --> 0:11:14.600 by people who don't know how the companies make money, 0:11:14.640 --> 0:11:17.800 just that they must always make more without fail and 0:11:17.840 --> 0:11:20.240 when you're a big stupid asshole. Every job that you 0:11:20.240 --> 0:11:22.319 see is condensed to its outputs, and not the stuff 0:11:22.320 --> 0:11:24.920 that leads up to the output, or the small nuances 0:11:24.960 --> 0:11:27.360 and conscious decisions that make an output good as opposed 0:11:27.360 --> 0:11:30.960 to simply acceptable or even bad. What does the software 0:11:30.960 --> 0:11:33.800 engineer do? They write code right, What does a rater do? 0:11:33.880 --> 0:11:36.560 They write words right? What does a hairdresser do they 0:11:36.600 --> 0:11:39.839 can't hear? Yeah, that's of course not actually the case. 0:11:40.280 --> 0:11:42.200 As I'll get into later in the series, the software 0:11:42.240 --> 0:11:44.400 engineer does far more than just code. And when they 0:11:44.480 --> 0:11:46.600 write code, they're not just saying what would solve this 0:11:46.640 --> 0:11:48.960 problem with a big smile on their face. They're taking 0:11:49.000 --> 0:11:51.800 into account their years of experience, what code does, what 0:11:51.880 --> 0:11:54.360 code could do, and all the things that my break 0:11:54.400 --> 0:11:55.920 is a result, and all of the things that you 0:11:55.960 --> 0:11:57.960 can't really tell from just looking at the code, like 0:11:58.080 --> 0:12:00.520 whether there's a reason things are made in a particle way, 0:12:01.280 --> 0:12:03.400 And a good coder doesn't just hammer at the keyboard 0:12:03.400 --> 0:12:05.720 with the aim of doing a particular task. They factor 0:12:05.760 --> 0:12:08.199 in questions like how does this functionality fit into the 0:12:08.240 --> 0:12:11.080 code that's already there? Or if someone has to update 0:12:11.160 --> 0:12:12.840 this code in the future, how do I make it 0:12:12.880 --> 0:12:14.679 easy for them to understand what I've written and make 0:12:14.760 --> 0:12:17.720 changes without breaking a bunch of other stuff. A writer 0:12:17.840 --> 0:12:20.800 doesn't just write words. They just ideas and ideas and 0:12:20.800 --> 0:12:23.080 emotions and thoughts and facts and feelings into a condensed 0:12:23.080 --> 0:12:25.240 piece of text. They sit up late at night typing 0:12:25.320 --> 0:12:27.600 thousands and thousands of words, and it drives them in say, 0:12:29.280 --> 0:12:31.360 it's often quite a motive or at the very least 0:12:31.440 --> 0:12:34.000 driven who or inspired by a given emotion, which is 0:12:34.000 --> 0:12:36.040 something that an AI simply can't replicate in a way 0:12:36.040 --> 0:12:40.240 that's authentic or believable. And a hairdresser doesn't just cut hair. 0:12:40.440 --> 0:12:43.679 They cut your hair, which may be wiry, dry, oily long, 0:12:43.800 --> 0:12:46.760 sure healthy, unhealthy on a scout with particular issues at 0:12:46.760 --> 0:12:48.559 a time of year, perhaps you want to change length 0:12:48.720 --> 0:12:50.360 at a time that fits you in the way you 0:12:50.559 --> 0:12:53.160 like it, which may be impossible to actually write down. 0:12:53.200 --> 0:12:55.520 But they get it just right, and they make conversation 0:12:55.640 --> 0:12:57.720 making you feel at ease while they snip and clip away. 0:12:57.760 --> 0:13:00.320 It tresses with you never having to think for a second. Fuck, 0:13:00.320 --> 0:13:02.040 does this person know what they're doing? Are they going 0:13:02.080 --> 0:13:05.439 to listen to me. This is the true nature of labor. 0:13:05.559 --> 0:13:08.640 The executives fail to comprehend at scale that the things 0:13:08.640 --> 0:13:12.640 we do are not units of work, but extrapolations of experience, emotion, 0:13:12.760 --> 0:13:15.960 and context that cannot be condensed in written meaning or 0:13:16.000 --> 0:13:19.520 bunches of trading material. Business idiots see our labor as 0:13:19.559 --> 0:13:22.640 the results of a smart manager saying do this, rather 0:13:22.640 --> 0:13:25.480 than human ingenuity interpreting both the requests and the shit 0:13:25.520 --> 0:13:31.080 the manager didn't say. Now, what does the CEO do? Well? 0:13:32.040 --> 0:13:35.040 I did look, and a Harvard study said that they 0:13:35.120 --> 0:13:38.240 spend twenty five percent of their time on people and relationships, 0:13:38.280 --> 0:13:41.840 twenty five percent on functional and business unit reviews, sixteen 0:13:41.880 --> 0:13:45.240 percent on organization and culture, and twenty one percent on 0:13:45.360 --> 0:13:47.360 just strategy, with a few percent here and there for 0:13:47.480 --> 0:13:52.840 things like professional development. Hmm. That's who runs the vast 0:13:52.880 --> 0:13:55.920 majority of companies. People that describe their work predominantly as 0:13:55.920 --> 0:13:58.640 looking at staff, talking to people, thinking what we do next, 0:13:59.160 --> 0:14:02.480 and go to lunch. The most highly paid jobs in 0:14:02.520 --> 0:14:04.920 the world are impossible to describe their labour described in 0:14:04.920 --> 0:14:08.559 a mishmash of linked inspiration. Yet everybody else's labor is 0:14:08.559 --> 0:14:11.560 an output that can be automated. As a result, large 0:14:11.600 --> 0:14:14.720 language models must seem like magic to these dickheads. When 0:14:14.720 --> 0:14:16.960 you see everything as an outcome, an outcome, you may 0:14:17.040 --> 0:14:19.960 or may not understand it. Definitely don't understand the process behind, 0:14:20.080 --> 0:14:22.560 let alone care about. You've kind of already see your 0:14:22.600 --> 0:14:26.080 workers as llms. You create a stratification of the workforce 0:14:26.120 --> 0:14:29.680 that goes beyond the normal organizational chart, with senior executives 0:14:29.720 --> 0:14:32.840 those closer to the class level of CEO acting as 0:14:33.000 --> 0:14:35.680 those who have risen above the doldrums of doing things, 0:14:35.720 --> 0:14:38.440 to the level of decision making, a fuzzy term that 0:14:38.480 --> 0:14:41.200 can mean everything from making nuance decisions with input from 0:14:41.240 --> 0:14:44.640 multiple different subject matter experts to as service now Bill 0:14:44.720 --> 0:14:47.280 McDermott did in twenty twenty two and I quote make 0:14:47.360 --> 0:14:50.120 it clear to everybody in a boardroom of other executives 0:14:50.120 --> 0:14:54.840 that everything they do must be AIAIAIAIAI. And that's five 0:14:54.880 --> 0:14:57.760 of those the same extents that some members of the 0:14:57.800 --> 0:14:59.920 business and tech media that have for the most part 0:15:00.120 --> 0:15:02.000 gotten by without having to think too hard about the 0:15:02.000 --> 0:15:05.840 actual things the companies are saying. Look, I realize this 0:15:05.880 --> 0:15:08.560 sounds a little mean, and it's not a unilateral statement, 0:15:08.960 --> 0:15:11.120 and I must must be clear. It doesn't mean that 0:15:11.160 --> 0:15:13.560 these people know nothing, just that it's been possible that 0:15:13.600 --> 0:15:15.880 scoot through the world without thinking too hard about whether 0:15:15.960 --> 0:15:19.040 or not something is true, just because an executive said up. 0:15:19.640 --> 0:15:22.000 When Salesforce said back in twenty twenty four that it's 0:15:22.040 --> 0:15:25.360 Einstein Trust Layer and AI would be transformational for jobs, 0:15:25.560 --> 0:15:28.080 the media dutifully wrote it down and published it without 0:15:28.080 --> 0:15:31.080 a second thought. It fully trusted Mark Benioff when he 0:15:31.120 --> 0:15:33.800 said that Agent Force agents would replace human workers, and 0:15:33.800 --> 0:15:35.720 then again when he said that AI agents were doing 0:15:35.800 --> 0:15:38.960 thirty to fifty percent of all the work in Salesforce itself, 0:15:39.200 --> 0:15:43.280 even though that's an unproven and nakedly ridiculous statement. Salesforce 0:15:43.320 --> 0:15:45.320 is CFO, by the way, said earlier in this year 0:15:45.320 --> 0:15:48.440 that AI wouldn't boost sales growth in twenty twenty five. 0:15:48.800 --> 0:15:51.480 One would think this would change how Salesforce was covered 0:15:51.560 --> 0:15:54.120 or how seriously one takes Mark Benioff, But it hasn't 0:15:54.200 --> 0:15:57.720 because nobody's really paying attention. In fact, nobody seems to 0:15:57.720 --> 0:16:13.480 want to do their job in this case. And this 0:16:13.760 --> 0:16:15.840 is how the core myths of jenerit if AI were 0:16:15.840 --> 0:16:18.600 built by executives saying stuff and the media publishing it 0:16:18.640 --> 0:16:21.920 without thinking about him. AI is replacing workers. AI is 0:16:21.920 --> 0:16:26.000 writing entire computer programs. AI is getting exponentially more powerful. 0:16:26.200 --> 0:16:28.320 What does powerful mean? Well, it means that the models 0:16:28.360 --> 0:16:30.520 are getting better on benchmarks that are rigged in their favor. 0:16:30.640 --> 0:16:33.360 But because nobody fucking explains what the benchmarks are, regular 0:16:33.400 --> 0:16:35.640 people are regularly told that AI is powerful and getting 0:16:35.640 --> 0:16:38.640 more powerful every single day. The only thing powerful about 0:16:38.680 --> 0:16:42.880 generifai is its pathology. The world's executives entirely disconnect different 0:16:42.920 --> 0:16:45.360 labor and natural production, are doing the only thing they 0:16:45.400 --> 0:16:47.320 know how to spend a bunch of money and say 0:16:47.400 --> 0:16:51.920 vague stuff about AI being the future. There are people journalists, investors, 0:16:51.960 --> 0:16:54.480 and analysts that have built entire careers on filling in 0:16:54.480 --> 0:16:56.480 the gaps for the powerful as they splurge billions of 0:16:56.520 --> 0:16:59.640 dollars in repeat with increasing desperation that the future is here, 0:16:59.680 --> 0:17:06.159 and then well, absolutely nothing else happens. You've likely seen 0:17:06.320 --> 0:17:09.160 a few ridiculous headlines recently, though. One of the most 0:17:09.200 --> 0:17:11.720 recent and most absurd is that open Ai will pay 0:17:11.720 --> 0:17:14.440 Oracle three hundred billion dollars over the next four years, 0:17:14.480 --> 0:17:17.320 closely followed with the claim that Invidia will invest and 0:17:17.320 --> 0:17:19.520 I put that in air, quotes one hundred billion dollars 0:17:19.560 --> 0:17:22.960 in open Ai to build ten gigawats of ai data centers, 0:17:23.200 --> 0:17:25.200 though the deal is structured in a way that means 0:17:25.240 --> 0:17:28.120 open Ai is paid progressively as each giga what is deployed, 0:17:28.640 --> 0:17:31.600 and also apparently open Ai will be leasing the chips 0:17:31.680 --> 0:17:34.240 rather than buying them out right. I must be clear 0:17:34.240 --> 0:17:36.400 that these deals are intentionally made to continue the myth 0:17:36.440 --> 0:17:38.639 of generat ifai to pump in video and to make 0:17:38.680 --> 0:17:41.320 sure open ai insiders can sell ten point three billion 0:17:41.320 --> 0:17:43.680 dollars worth of shares, which they're currently trying to do 0:17:43.880 --> 0:17:47.480 at evaluation of five hundred billion goddamn dollars. I want 0:17:47.480 --> 0:17:50.119 to be clear about something open Ai cannot afford the 0:17:50.119 --> 0:17:53.000 three hundred billion dollars. Open Ai has not received a 0:17:53.080 --> 0:17:55.720 dollar from Nvidia and won't do so for at least 0:17:55.760 --> 0:17:57.400 a month when I think they're going to receive ten 0:17:57.440 --> 0:18:00.400 billion dollars. But the rest of that ninety's only coming 0:18:00.440 --> 0:18:02.840 when they build those data centers, which open Ai can't 0:18:02.840 --> 0:18:05.800 afford to do. In Vidia needs this myth to continue 0:18:05.840 --> 0:18:07.960 because in truth, all of these data centers are being 0:18:07.960 --> 0:18:10.560 built for demand that doesn't exist, or that if it 0:18:10.920 --> 0:18:14.480 did exist, doesn't necessarily translate into business customers paying huge 0:18:14.480 --> 0:18:17.840 amount amounts for access to open AI's generative AI services. 0:18:18.400 --> 0:18:21.159 In Vidia, open Ai, Core Weave, and other AI related 0:18:21.200 --> 0:18:24.240 companies hope that by announcing theoretical billions of dollars or 0:18:24.440 --> 0:18:27.680 hundreds of billions of dollars of these strange, vague, and 0:18:27.680 --> 0:18:30.639 impossible seeming deals, they can keep pretending that the demand 0:18:30.720 --> 0:18:33.080 is there, because why else would they build all these 0:18:33.160 --> 0:18:37.080 data centers? Right, Well, there's there's that, and the entire 0:18:37.080 --> 0:18:40.000 stock market races on. In video's back. It accounts for 0:18:40.040 --> 0:18:41.800 seven to eight percent of the value of the S 0:18:41.880 --> 0:18:44.520 and P five hundred, and Jensen Huang needs to keep 0:18:44.560 --> 0:18:48.679 selling those fucking GPUs. I intend to explain later how 0:18:48.720 --> 0:18:50.840 all of this works and how brittle all of this 0:18:50.880 --> 0:18:53.879 really is. But the intention of these deals is simple 0:18:53.960 --> 0:18:56.320 to make you think this much money can't be wrong, 0:18:56.720 --> 0:18:59.359 and I assure you it can. These people need you 0:18:59.400 --> 0:19:01.879 to believe this is inevitable. But they are being proven 0:19:01.960 --> 0:19:04.439 wrong again and again and again, and today I'm going 0:19:04.480 --> 0:19:07.840 to continue to do so. Underpinning these stories about huge 0:19:07.880 --> 0:19:10.560 amounts of money and endless opportunity lies a dark secret. 0:19:10.760 --> 0:19:12.320 The none of this is working, and all of this 0:19:12.400 --> 0:19:15.320 money has been invested in a technology that doesn't make 0:19:15.440 --> 0:19:17.720 much revenue and loves to burn millions or billions or 0:19:17.960 --> 0:19:21.480 hundreds of billions of dollars. Over half a trillion dollars 0:19:21.520 --> 0:19:24.159 in fact, has gone into an entire industry without a 0:19:24.200 --> 0:19:27.400 single profitable company developing models of products built on top 0:19:27.440 --> 0:19:31.240 of these AI models. By my estimates, there's about forty 0:19:31.280 --> 0:19:33.359 four billion dollars of revenue in general of AI this 0:19:33.480 --> 0:19:35.679 year when you add in anthropic and open AIS revenue 0:19:35.680 --> 0:19:37.879 to the part, along with other stragglers, and most of 0:19:37.880 --> 0:19:40.600 that number has been gathered from reporting from outlets like 0:19:40.640 --> 0:19:43.320 The Information. Because none of these companies share their revenues, 0:19:43.560 --> 0:19:45.600 all of them lose shit tons of money, and their 0:19:45.640 --> 0:19:50.000 actual revenues are really, really, really small. Only one member 0:19:50.000 --> 0:19:53.080 of the Magnificent seven outside of Nvidia, has ever disclosed 0:19:53.080 --> 0:19:56.359 its AI revenue, Microsoft, which has stopped reporting, in January 0:19:56.400 --> 0:19:59.320 twenty twenty five, when it reported it would have thirteen 0:19:59.320 --> 0:20:02.080 billion in aniized revenue this well, I guess that will 0:20:02.160 --> 0:20:04.800 be for the month because it's month times twelve, about 0:20:04.800 --> 0:20:07.960 one point eight three billion a month. I'm I know 0:20:08.040 --> 0:20:11.480 that sounds like a lot. But Microsoft is a sales machine. 0:20:11.880 --> 0:20:16.199 It's built specifically to create or exploit software markets, suffocating 0:20:16.200 --> 0:20:18.960