WEBVTT - The Subprime AI Crisis 0:00:02.840 --> 0:00:07.600 Zone Media. Hello and welcome to Better Offline. I'm your 0:00:07.600 --> 0:00:22.119 host and chief romance officer, ed Ze Tron. In the 0:00:22.200 --> 0:00:25.200 last episode, I dug into the fundamental weaknesses and open 0:00:25.239 --> 0:00:28.280 AI the supposed leader in the genital of AI boom, 0:00:28.400 --> 0:00:30.440 and today I'm going to get into a much larger, 0:00:30.720 --> 0:00:34.200 more systemic, more terminal problem and the signs that things 0:00:34.240 --> 0:00:37.320 are really really falling apart. And as ever, I will 0:00:37.320 --> 0:00:39.920 have links to everything I'm talking about in the episode notes, 0:00:40.080 --> 0:00:42.000 so you know I'm not making it up, which one 0:00:42.040 --> 0:00:44.600 person suggested I did once and it bothered me a 0:00:44.640 --> 0:00:48.479 great deal. But back to the actual stuff. The problems 0:00:48.479 --> 0:00:50.640 that open ai is facing are those faced by the 0:00:50.800 --> 0:00:54.320 entire generative AI industry. One's born of their sole focused 0:00:54.360 --> 0:00:57.560 on the transformer based architecture underlying large language models like 0:00:57.680 --> 0:01:01.640 chat GPT open A issues besides the fact that they're 0:01:01.640 --> 0:01:04.400 in a terrible business as discussed in the last episode, 0:01:04.600 --> 0:01:07.280 is that generative AI, and by extension, the model GPT 0:01:07.360 --> 0:01:10.720 in the product Chat GPT, doesn't really solve complex problems 0:01:10.760 --> 0:01:13.720 that would justify the massive costs behind it. It is 0:01:13.760 --> 0:01:16.800 these massive intractable challenges that are a result of these 0:01:16.800 --> 0:01:19.880 models being probabilistic, meaning that they don't know anything. They're 0:01:19.880 --> 0:01:22.840 just generating an answer based on maths and training data, 0:01:23.200 --> 0:01:25.280 something that model developers are running out of at an 0:01:25.280 --> 0:01:29.880 incredible pace. Hallucinations, which occur when models authoritatively state something 0:01:29.920 --> 0:01:31.679 that isn't true, or, in the case of an image 0:01:31.720 --> 0:01:35.000 or a video, makes something that just looks wrong. Well, 0:01:35.600 --> 0:01:38.560 they're impossible to resolve without new branches of maths, and 0:01:38.600 --> 0:01:40.560 while you might be able to reduce or mitigate them, 0:01:40.600 --> 0:01:43.399 their existence makes it hard for business critical applications to 0:01:43.440 --> 0:01:46.040 truly rely on this kind of AI. I don't even 0:01:46.080 --> 0:01:48.920 know if i'd call it an AI, but regardless, we 0:01:49.000 --> 0:01:52.880 go forward, and even tech's most dominant players can't seem 0:01:52.960 --> 0:01:57.000 to turn generative AI into any kind of real business line. 0:01:57.080 --> 0:02:00.000 The Information reported in early September that customers of Micross 0:02:00.280 --> 0:02:03.360 three sixty five Suite are barely adopting its AI powered 0:02:03.440 --> 0:02:06.960 copilot products, with somewhere between zero point one percent and 0:02:07.040 --> 0:02:09.959 one percent of the four hundred and forty million paying 0:02:10.040 --> 0:02:12.799 people who pay for Microsoft three sixty five, which is 0:02:12.800 --> 0:02:15.200 about thirty to fifty dollars a person, by the way, 0:02:15.720 --> 0:02:18.800 are willing to pay for AI, and just to be clear, 0:02:18.840 --> 0:02:21.840 I muddied that little. It's thirty to fifty bucks per 0:02:22.040 --> 0:02:24.600 person per head to add this stuff. I'll get into 0:02:24.639 --> 0:02:27.440 it in a minute. One firm, according to the information, 0:02:27.840 --> 0:02:30.440 was testing the AI features and was quoted as saying 0:02:30.480 --> 0:02:33.000 that most people don't find it that valuable right now, 0:02:33.080 --> 0:02:35.840 and others are saying that many businesses haven't seen breakthroughs 0:02:35.840 --> 0:02:39.280 in productivity or other benefits, and that they're not sure 0:02:39.320 --> 0:02:42.840 that they will. In an internal presentation provided to be 0:02:42.919 --> 0:02:46.120 by a source, users of Microsoft SharePoint copilot complained that 0:02:46.160 --> 0:02:50.520 Microsoft chatbot kept getting questions wrong, sometimes failing to provide 0:02:50.520 --> 0:02:53.800 references even for correct answers, with another complaining that the 0:02:53.840 --> 0:02:57.080 copilot was and I quote using content not connected as 0:02:57.120 --> 0:02:59.919 a document resource to answer questions. And by the way, 0:03:00.040 --> 0:03:02.119 the whole point of share point is that it's your 0:03:02.200 --> 0:03:06.400 data informing everything. I assume it was drawing from its 0:03:06.440 --> 0:03:11.040 training data or perhaps the internet anyway, genuinely not useful. 0:03:11.560 --> 0:03:14.960 And you'd think that with these new services that don't 0:03:14.960 --> 0:03:18.560 seem that useful, that are questionably useful, that Microsoft would 0:03:18.520 --> 0:03:21.800 be doing people a deal right wrong. How much is 0:03:21.800 --> 0:03:26.119 Microsoft charging for these services? Thirty dollars a seat per 0:03:26.200 --> 0:03:29.040 person on top of what you are already paying, are 0:03:29.080 --> 0:03:31.880 as much as fifty dollars a month extra for specialist 0:03:31.880 --> 0:03:36.000 products like co pilots for sales, Microsoft is effectively asking 0:03:36.040 --> 0:03:39.280 customers to double their spend, and by the way, that's 0:03:39.320 --> 0:03:42.200 with an annual commitment for products that don't seem to 0:03:42.240 --> 0:03:45.600 be that helpful. And really, that is kind of the 0:03:45.720 --> 0:03:49.080 state of generative AI, the literal leader and productivity in 0:03:49.120 --> 0:03:51.920 business software, cannot seem to find a product that will 0:03:51.960 --> 0:03:55.440 make people more productive and that they will then pay for. 0:03:55.800 --> 0:03:58.240 And it's in part because the results are kind of mediocre, 0:03:58.600 --> 0:04:01.000 and also that the costs are so burdensome that there's 0:04:01.000 --> 0:04:05.040 no way for Microsoft to avoid charging a premium. And really, 0:04:05.080 --> 0:04:07.720 if Microsoft needs to charge this much, it's either because 0:04:07.720 --> 0:04:10.320 Sachin Adela is really desperate to hit half a trillion 0:04:10.360 --> 0:04:13.520 dollars in revenue by twenty thirty, or that the costs 0:04:13.520 --> 0:04:16.000 are too high to charge much less. Maybe it's a 0:04:16.040 --> 0:04:19.359 little bit of both. And this all only serves to 0:04:19.400 --> 0:04:22.400 shed further light on just the mediocrity of generative AI 0:04:22.640 --> 0:04:26.200 and how limited large language models are. And all of this, 0:04:26.279 --> 0:04:29.640 by the way, is existentially threatening to Open AI because 0:04:30.200 --> 0:04:33.880 they've coasted to one hundred and fifty seven billion dollar valuation, 0:04:34.240 --> 0:04:38.120 almost entirely based on hype. And so it's that company's 0:04:38.120 --> 0:04:40.080 always tried to tell us that the future of AI 0:04:40.120 --> 0:04:42.240 will blow us away, that the next generation of large 0:04:42.279 --> 0:04:45.080 language models are eminent and they're going to be incredible. 0:04:45.400 --> 0:04:48.360 And the artificial general intelligence, where machines can reason and 0:04:48.440 --> 0:04:52.520 act beyond human capabilities, that's just around the corner. And 0:04:52.560 --> 0:04:54.520 by the way, all of that is in part thanks 0:04:54.560 --> 0:04:57.000 to the media sloping it down and just assuming that 0:04:57.040 --> 0:05:00.280 they get it right. But until now that that's all 0:05:00.279 --> 0:05:03.279 they've really had to do. But I think we're finally 0:05:03.320 --> 0:05:06.480 getting the rubber meeting the road with this. I previously 0:05:06.480 --> 0:05:08.919 said one of the pale horses of the AI apocalypse 0:05:08.960 --> 0:05:11.839 is when a big stupid magic trick was necessary, a 0:05:11.880 --> 0:05:14.160 product that someone shoves out the door in hopes it 0:05:14.200 --> 0:05:17.400 will impress people and keep them believing in the magical future. 0:05:18.400 --> 0:05:20.600 And you'd think that they'd have something really good right 0:05:20.640 --> 0:05:23.400 now because open ai just raised all this money and 0:05:23.440 --> 0:05:28.880 the practical applications just are obviously not there, except well, 0:05:29.240 --> 0:05:31.000 you know, no, no, no, no, this is open AI. 0:05:31.080 --> 0:05:33.920 They wouldn't make a big, stupid mistake, would they. I mean, 0:05:34.279 --> 0:05:36.000 one of the things I always tell clients of mine 0:05:36.000 --> 0:05:37.760 in pr is not to shove a product out the 0:05:37.800 --> 0:05:40.080 door before it's ready, and to also make sure it's 0:05:40.160 --> 0:05:43.760 really obvious why people should pay for it. Otherwise you're 0:05:43.800 --> 0:05:46.400 just kind of launching something into the ether and hope 0:05:46.400 --> 0:05:49.880 people will find a reason to sell it for you. 0:05:51.120 --> 0:05:53.599 And yeah, that's exactly what they did. It happened. On 0:05:53.640 --> 0:05:56.800 September twelfth. Open Ai launched OH one, which had been 0:05:56.880 --> 0:05:59.719 code named Strawberry, with all of the excitement as a 0:05:59.720 --> 0:06:04.120 trip to the Proctologist. Across a series of tweets, CEO 0:06:04.240 --> 0:06:07.320 Samuel and described one as open ayes most capable and 0:06:07.360 --> 0:06:11.000 aligned models, yet then immediately conceded that O one was 0:06:11.040 --> 0:06:14.440 still flawed, still limited, and it still seems more impressive 0:06:14.480 --> 0:06:16.599 on its first use than it does after you spend 0:06:16.600 --> 0:06:20.200 more time with it. Oh my god, he admitted. He 0:06:20.240 --> 0:06:23.279 then promised it would deliver more accurate results when performing 0:06:23.279 --> 0:06:26.040 the kinds of activities where there's a definitive right answer, 0:06:26.160 --> 0:06:30.159 like coding maths or answering science questions. One might think 0:06:30.200 --> 0:06:32.320 that he'd walk in with I don't know, like a 0:06:32.360 --> 0:06:35.000 product built on top of one or like an use 0:06:35.080 --> 0:06:37.880 case or thing that would make the audience go, wow, 0:06:38.000 --> 0:06:41.720 I could build something with this. He didn't. I don't 0:06:41.760 --> 0:06:44.119 think he wants to try. I don't think he hasn't 0:06:44.120 --> 0:06:46.520 had to try that hard. So far people have been 0:06:46.560 --> 0:06:50.040 sloping down his slop happily. This boy may not have 0:06:50.080 --> 0:06:54.279 any tricks left. But let's talk about how O one works. 0:06:54.279 --> 0:06:55.839 And I'm going to introduce you to a bunch of 0:06:55.880 --> 0:06:58.080 new concepts here, but I promise I won't get too 0:06:58.120 --> 0:07:00.760 deep into the weeds. And I really want you to 0:07:00.800 --> 0:07:04.360 know how these machines work. It's critical for critiquing these companies. 0:07:04.640 --> 0:07:06.880 And the big way they take advantage of you is 0:07:06.920 --> 0:07:08.719 that they claim all of this is black magic, that 0:07:08.760 --> 0:07:11.920 you could never possibly understand it. You absolutely can. And 0:07:12.000 --> 0:07:13.880 if you want their explanation, I'm going to have it 0:07:13.920 --> 0:07:18.880 in the show notes. Okay. When presented with a problem, 0:07:19.000 --> 0:07:21.920 OH one breaks it down into individual steps that hopefully 0:07:22.000 --> 0:07:24.400 would lead to a correct answer in a process called 0:07:24.480 --> 0:07:28.040 chain of thought. Again, these things are not thinking. They're 0:07:28.040 --> 0:07:30.960 not thinking, but this is the term. It's also a 0:07:31.000 --> 0:07:33.920 little easier if you think of OH one as two 0:07:34.000 --> 0:07:37.480 parts of one model. On each step, one part of 0:07:37.520 --> 0:07:40.560 the model applies something called reinforcement learning with the other one, 0:07:40.600 --> 0:07:44.320 which is the model actually outputting things rewarded or punished 0:07:44.360 --> 0:07:47.320 based on the perceived correctness of their progress. And this 0:07:47.520 --> 0:07:49.680 is what is called reasoning, by the way, even though 0:07:49.680 --> 0:07:53.320 it really doesn't match human reasoning at all, and then 0:07:53.440 --> 0:07:56.360 based on the reward of the punishment, it generates a 0:07:56.400 --> 0:08:00.560 final answer from this chain of thought consideration. This is 0:08:00.640 --> 0:08:03.000 different to how other large language models work in the 0:08:03.040 --> 0:08:06.600 sense that the model is generating outputs than actually looking 0:08:06.640 --> 0:08:09.280 back at them then ignoring or approving what it thinks 0:08:09.280 --> 0:08:11.320 are good steps to get to an answer, rather than 0:08:11.440 --> 0:08:14.640 just generating one and saying here's the answer. This may 0:08:14.680 --> 0:08:17.200 seem like a big breakthrough or even another step towards 0:08:17.280 --> 0:08:21.000 artificial general intelligence, and it isn't, And you can tell 0:08:21.040 --> 0:08:23.119 that by the fact that open ai opt to release 0:08:23.160 --> 0:08:25.800 O one as its own standalone product rather than something 0:08:25.840 --> 0:08:29.840 built into GPT. It's also telling that the examples demonstrated 0:08:29.880 --> 0:08:32.800 by open AI, like maths and science problems, are the 0:08:32.800 --> 0:08:34.839 ones where the answer can be known ahead of time 0:08:35.040 --> 0:08:37.720 and a solution is either correct or false, thus allowing 0:08:37.760 --> 0:08:39.880 the model to guide the chain of thought through each 0:08:39.920 --> 0:08:43.679 step towards that answer, rather than actually having to produce 0:08:43.720 --> 0:08:47.480 something where they might not necessarily be one. Open ai 0:08:47.600 --> 0:08:50.920 didn't show the one model trying to tackle complex problems 0:08:51.400 --> 0:08:54.520 such as high end mathematical equations or otherwise where the 0:08:54.520 --> 0:08:57.640 solution isn't known in advance by its own admission. Open 0:08:57.679 --> 0:09:00.320 AI has heard reports that one is actually more prone 0:09:00.320 --> 0:09:03.000 to hallucinations than GPT four H, and the model is 0:09:03.080 --> 0:09:05.199 less inclined to admit when it doesn't have the answer 0:09:05.280 --> 0:09:08.240 to a question when compared to other previous models. This 0:09:08.400 --> 0:09:10.960 is because despite there being a model that checks the 0:09:11.000 --> 0:09:14.080 work of the model, the work checking part of the 0:09:14.120 --> 0:09:17.320 model is still capable of hallucinations. It's kind of like 0:09:18.000 --> 0:09:20.280 a kid being taught something by a teacher who just 0:09:20.320 --> 0:09:23.880 occasionally gets things horribly wrong. That child, though they may 0:09:24.200 --> 0:09:28.280 mostly get right answers, will learn bad things. Now, learning 0:09:28.320 --> 0:09:31.520 here isn't really what's happening, but the output of the 0:09:31.640 --> 0:09:35.280 end will be informed by a model that makes the loocinations. 0:09:35.280 --> 0:09:37.560 It's like, I don't know, got a town full of dogs. 0:09:37.600 --> 0:09:39.720 You get a bunch of baboons in to get rid 0:09:39.760 --> 0:09:41.599 of the dogs. The baboons succeed and getting rid of 0:09:41.640 --> 0:09:43.199 the dogs Now you just got a bunch of baboons, 0:09:43.200 --> 0:09:46.800 so you get in, aren't no robots? Robots destroy the baboons. 0:09:46.800 --> 0:09:48.960 At this point, you've got robots. If the robots are autonomous, 0:09:48.960 --> 0:09:50.839 they start taking over the town. So they need to 0:09:50.880 --> 0:09:52.720 find a bigger robot to take over the town from 0:09:52.800 --> 0:09:56.080 the robots. Now you've just got an escalating problem where 0:09:56.080 --> 0:09:58.520 things are only going to get worse. And if you 0:09:58.559 --> 0:10:01.760 work open ai and that sounds accurate, please email me anyway. 0:10:02.600 --> 0:10:05.360 According to open ai, OH one also, thanks to this 0:10:05.559 --> 0:10:08.679 chain of thought process, feels more convincing to human users 0:10:08.679 --> 0:10:12.120 because it provides more detailed answers, and thus people are 0:10:12.120 --> 0:10:16.000 more inclined to trust the outputs even when they're completely wrong. Now, 0:10:16.000 --> 0:10:18.320 if you think I'm being overly hard on open ai, 0:10:18.480 --> 0:10:20.920 consider the ways in which the company is marketed. One 0:10:21.760 --> 0:10:25.760 open ai described O one's reinforcement training as thinking and 0:10:25.880 --> 0:10:28.720 reasoning when it's making guesses and then guessing on the 0:10:28.760 --> 0:10:31.320 correctness of these guesses at each step. Where the end 0:10:31.360 --> 0:10:34.000 destination is often something that can be known in advance. 0:10:34.760 --> 0:10:38.920 Generative AI does not know anything. These are still probabilistic models. 0:10:39.760 --> 0:10:42.840 This thing is not thinking at all. There is no reasoning. 0:10:43.480 --> 0:10:45.480 It's got a model, reading a model, giving a model 0:10:45.480 --> 0:10:48.800 answers from it's a mess, and it's an insult to people, 0:10:48.960 --> 0:10:52.200 actual human beings who, when they think, are acting based 0:10:52.240 --> 0:10:56.120 on many, many complex factors, their experience, their knowledge, the 0:10:56.200 --> 0:11:00.119 knowledge they've accumulated over years of experiences, their brain chemistry, 0:11:00.559 --> 0:11:03.480 so on and so forth. Well, we may to guess 0:11:03.520 --> 0:11:07.000 about the correctness of each thing we're guessing at, and 0:11:07.040 --> 0:11:09.880 we may reason through a complex problem. All of this 0:11:09.960 --> 0:11:13.120 is based on something concrete. When we get something wrong, 0:11:13.160 --> 0:11:18.120 it's based on actual experience versus training data and probabilistic models. 0:11:19.240 --> 0:11:22.559 This shit is not thinking at all, and by god 0:11:22.679 --> 0:11:26.240 is it expensive. Pricing for one preview, which is the 0:11:26.240 --> 0:11:29.560 first model, is fifteen dollars per million input tokens and 0:11:29.600 --> 0:11:32.840 sixteen per million output tokens. In essence, it's three times 0:11:32.880 --> 0:11:35.800 as expensive as their most expensive model GPT four O 0:11:36.160 --> 0:11:39.280 for input and four times is expensive for output. And 0:11:39.320 --> 0:11:42.840 then there's a hidden cost. Data scientist Max Wolf reported 0:11:42.840 --> 0:11:45.520 the open ayes reasoning tokens the output it uses to 0:11:45.520 --> 0:11:47.880 get you to the final answer where it says, Okay, 0:11:47.960 --> 0:11:51.600 I need to find out the solution to this problem. 0:11:51.640 --> 0:11:55.080 So here are the thirty steps I've gone through. Yeah, 0:11:55.400 --> 0:11:58.880 those are actually generated using the most expensive tokens, the 0:11:58.920 --> 0:12:01.439 output tokens. So the more it has to think, the 0:12:01.480 --> 0:12:04.520 more expensive it gets. All of the things it generates 0:12:04.520 --> 0:12:08.040 to consider an answer are also charged for, which means 0:12:08.040 --> 0:12:10.160 the more complex it is, the more expensive it's going 0:12:10.240 --> 0:12:13.800 to be worse. Still, if you integrate this model, open 0:12:13.840 --> 0:12:18.440 ai does not show you what it's reasoning. All of 0:12:18.480 --> 0:12:21.839 that calculation happens in the background, and they still charge 0:12:21.880 --> 0:12:25.240 you for it. You just don't know how much every 0:12:25.440 --> 0:12:29.720 oh one step is charged to you in an indeterminate way, 0:12:30.000 --> 0:12:32.600 and open ai claims that they can't show you because 0:12:32.600 --> 0:12:38.400 of competitive reasons. Ugh, nasty company, really greasy and they're 0:12:38.400 --> 0:12:42.680 still gonna burn Okay, okay, though it's different GPT four. 0:12:42.720 --> 0:12:45.880 Oh and it's really expensive. But is it better? Of 0:12:45.920 --> 0:12:49.240 course it must be better, right, right? It sounds great. 0:12:49.440 --> 0:12:52.000 It's thinking, right, it's reasoning right. No, No, it's not, 0:12:52.360 --> 0:12:57.559 it's not. It's worse. This crap's worse. Let's talk about accuracy. 0:12:57.760 --> 0:13:00.000 On how can news the reddit styles I owned by 0:13:00.000 --> 0:13:03.160 am Horman's former alum y Combinator, one person complained about 0:13:03.160 --> 0:13:06.240 O one hallucinating libraries and functions when presented with a 0:13:06.360 --> 0:13:09.079 programming task, and making mistakes when asked questions where the 0:13:09.080 --> 0:13:13.160 answer isn't readily available in the Internet. On Twitter, Henrik Nyberg, 0:13:13.320 --> 0:13:15.839 a startup founder and former game developer, asked OH one 0:13:15.880 --> 0:13:18.600 to write a Python program that multiplied two numbers, then 0:13:18.600 --> 0:13:21.959 calculated the expected output of said program. While OH one 0:13:22.080 --> 0:13:24.560 correctly wrote the code, although said code could have been 0:13:24.559 --> 0:13:29.880 more succinct, the actual result was wildly incorrect. Karthig Cannon, himself, 0:13:29.920 --> 0:13:32.440 a founder of an AI company, tried a programming task 0:13:32.520 --> 0:13:35.199 on O one where it also hallucinated a non existent 0:13:35.200 --> 0:13:39.920 command for the API he was using. Another person, Sashi Yanshin, 0:13:40.160 --> 0:13:41.640 tried to play a game of chess with O one, 0:13:41.679 --> 0:13:44.360 and it hallucinated an entire piece onto the board and 0:13:44.400 --> 0:13:46.880 then it lost. And because I'm a little shit, I 0:13:46.920 --> 0:13:49.079 also tried asking one to listen a number of states 0:13:49.080 --> 0:13:51.920 with A in the name. After contemplating for eighteen seconds, 0:13:51.960 --> 0:13:55.240 it provided the names of thirty seven states, including Mississippi, 0:13:55.440 --> 0:13:56.960 you know, the classic state with an A in it. 0:13:57.360 --> 0:13:59.679 By the way, there are thirty six states that have 0:14:00.080 --> 0:14:02.520 in them, just in case you're curious. I then asked 0:14:02.520 --> 0:14:04.240 for a list of states with the letter W in 0:14:04.240 --> 0:14:05.880 the name, and then it sat and it thought for 0:14:05.920 --> 0:14:09.440 eleven seconds, and then included North Carolina and North Dakota. 0:14:10.360 --> 0:14:13.000 Great stuff. By the way, I also asked tho one 0:14:13.000 --> 0:14:14.840 to count the number of times the letter R appears 0:14:14.840 --> 0:14:17.480 in the word strawberry, which is the pre release code 0:14:17.559 --> 0:14:20.400 name for this. It's said too, I would have hard 0:14:20.400 --> 0:14:23.080 coded that one. Personally. You can't give me that kind 0:14:23.120 --> 0:14:26.040 of joy now. Open AI claims that I one performs 0:14:26.040 --> 0:14:30.120 similarly to PhD students on challenging benchmark tasks in physics, chemistry, 0:14:30.120 --> 0:14:33.560 and biology, just not in geography, it seems, or basic 0:14:33.640 --> 0:14:38.520 elementary level English, or maths or programming. Also, I mean 0:14:38.720 --> 0:14:42.160 for the PhD listeners. I've met a few PhD people 0:14:42.160 --> 0:14:44.800 who authoritatively state things that are completely untrue that they 0:14:44.800 --> 0:14:47.480 know nothing about. This is not a broad stroke thing, 0:14:47.560 --> 0:14:51.200 but I get the sense that it's true anyway. This 0:14:51.400 --> 0:14:53.960 is I should know, the big stupid magic trick I 0:14:54.000 --> 0:14:57.120 predicted in the past. Open AI is shoving strawberry out 0:14:57.120 --> 0:14:58.840 the door as a means of proving to investors and 0:14:58.880 --> 0:15:01.800 the greater public. But they've still got it that the 0:15:01.840 --> 0:15:05.480 AI revolution is still here, that this thing is thinking, 0:15:06.280 --> 0:15:08.520 and what they actually have is a clunky, on exciting 0:15:08.560 --> 0:15:10.760 and expensive model that doesn't really seem to have any 0:15:10.800 --> 0:15:14.600 measurable improvement. Okay, I'm sorry. It has a measurable improvement. 0:15:15.000 --> 0:15:17.240 You can measure it on the weird rigged test they 0:15:17.240 --> 0:15:19.600 do for all of these things. And the thing is, 0:15:20.840 --> 0:15:24.200 at this point, you'd think that even Apple, when they 0:15:24.200 --> 0:15:26.040 pulled together a new thing, even when they had the 0:15:26.160 --> 0:15:28.240 first Apple Watch and it was not obvious why you 0:15:28.240 --> 0:15:30.440 had to own it, they still had apps that were 0:15:30.440 --> 0:15:33.000 connected to it. They still had things you could point 0:15:33.000 --> 0:15:36.120 out and go, oh, that's cool, I've got four square 0:15:36.120 --> 0:15:38.359 on this four square on there at the time. Nevertheless, 0:15:38.440 --> 0:15:41.360 they had apps to show. I just feel like open 0:15:41.400 --> 0:15:45.680 Ai has this deep contempt for Silicon Valley and for 0:15:45.800 --> 0:15:48.720 the world at large. They don't even have it in 0:15:48.760 --> 0:15:51.640 them to be like, Okay, we have this new model, 0:15:51.800 --> 0:15:53.960 and here is the new thing we built with it, 0:15:54.280 --> 0:15:57.480 and this thing does this and now you will see 0:15:57.480 --> 0:16:01.600 how important this company is. Instead, we get this crap. 0:16:02.680 --> 0:16:06.160 We just get this very boring crap. And sure, I'm 0:16:06.160 --> 0:16:09.280 sure someone technical is going to email me and say, ed, wow, 0:16:09.400 --> 0:16:11.760 chain thought reasoning. There are other companies that have been 0:16:11.760 --> 0:16:14.880 doing it already. Anthropic already had something like this, and 0:16:14.960 --> 0:16:18.120 even then they didn't do shit with it. Where's the product, man, 0:16:18.320 --> 0:16:22.400 where's the thing I meant to care about? Why should 0:16:22.440 --> 0:16:25.880 anybody give a shit about this? Well, sam Altman is 0:16:25.960 --> 0:16:29.480 likely trying to trump up the reasoning abilities of O one. Well, 0:16:29.520 --> 0:16:32.160 people you know, such as the people bankrolling him, will 0:16:32.200 --> 0:16:34.440 actually see he's a ten to twenty second waiting time 0:16:34.480 --> 0:16:37.280 for an answer which may or may not be correct. 0:16:37.640 --> 0:16:39.840 But you have a bit more detail, which isn't even 0:16:39.880 --> 0:16:43.960 the reasoning happening because open AI hides that bit. Nobody 0:16:43.960 --> 0:16:47.640 gives a shit about better answers anymore. They want generative 0:16:47.680 --> 0:16:50.080 AI to do something new, and I don't think the 0:16:50.080 --> 0:16:52.600 open AI has any idea how to make that happen. 0:16:53.080 --> 0:16:56.320 Sam Orman's limp shitty attempts to anthropomorphize I one by 0:16:56.360 --> 0:16:59.520 making it think can use reasoning obvious attempts to suggest 0:16:59.520 --> 0:17:01.960 that this is on how part of the path through AGI. 0:17:02.160 --> 0:17:06.200 But even the most staunch AI advocates, well, they can't 0:17:06.200 --> 0:17:08.920 seem to get excited about this. In fact, I kind 0:17:08.920 --> 0:17:10.800 of argue that O one shows that open AI is 0:17:10.840 --> 0:17:14.119 desperate and out of ideas. Now if you don't have 0:17:14.160 --> 0:17:17.440 any ideas, though, the following advertisements will be more than 0:17:17.480 --> 0:17:20.280 happy to fill your empty little brain with new ideas 0:17:20.359 --> 0:17:23.280 that involve giving someone money or downloading something. And I 0:17:23.400 --> 0:17:28.800 must implore you to just accept everything that follows. I 0:17:28.880 --> 0:17:30.720 don't endorse any of it because I don't know what 0:17:30.760 --> 0:17:45.440 it's going to be, but you must. And we're back. 0:17:47.000 --> 0:17:49.080 So I think now is a good time to get 0:17:49.119 --> 0:17:52.080 back to the root of the generative AI problem. Generative 0:17:52.080 --> 0:17:54.960 AI is being sold to you on multiple lives that 0:17:55.040 --> 0:17:58.760 it's AI, it's actually artificial intelligence, it's going to get better, 0:17:59.119 --> 0:18:02.560 that this will become artificial general intelligence, that this will 0:18:02.600 --> 0:18:06.480 become the thinking computer, and all of this is inevitable. 0:18:07.640 --> 0:18:10.879 Putting aside terms like performance, as they're largely us as 0:18:10.960 --> 0:18:13.920 a means of generating things accurately or faster, rather than 0:18:13.920 --> 0:18:17.520 being good at anything. Large language models have effectively platowed 0:18:18.040 --> 0:18:21.120 more powerful never seems to mean does more, and more 0:18:21.160 --> 0:18:24.760 powerful often means more expensive to run or more expensive 0:18:24.760 --> 0:18:27.080 for you as the user to access, meaning that you've 0:18:27.160 --> 0:18:29.320 just made something that doesn't do more and does cost 0:18:29.359 --> 0:18:32.520 more to run. If the combined forces of every venture 0:18:32.560 --> 0:18:34.879 capitalist and big tech hyperscaler have yet to come up 0:18:34.920 --> 0:18:36.800 with a meaningful use case that lots of people will 0:18:36.800 --> 0:18:39.920 actually pay for. I just don't see one coming. Large 0:18:39.960 --> 0:18:42.320 language models and yes, that's where all of these billions 0:18:42.359 --> 0:18:44.919 of dollars are going. Are not going to magically sprout 0:18:44.920 --> 0:18:47.840 new capabilities a big Tech and open Ai burn another 0:18:47.880 --> 0:18:51.200 one hundred and fifty billion dollars. And yes that number 0:18:51.240 --> 0:18:54.000 isn't hyperbole. It's actually pretty close to the amount being 0:18:54.040 --> 0:18:56.800 plowed into these companies when you include things like investments 0:18:56.800 --> 0:18:59.840 in companies like Anthropic and open Ai. And they're genuinely 0:18:59.840 --> 0:19:02.120 and sane amount of capex from the likes of Google, 0:19:02.160 --> 0:19:06.240 Amazon and Microsoft going into expanding data centers and buying GPUs. 0:19:07.040 --> 0:19:09.760 Nobody seems to be trying to make these things more efficient, 0:19:09.960 --> 0:19:12.199 or at the very least nobody's succeeded in doing so, 0:19:12.359 --> 0:19:14.639 because I think if they had, they'd be shouting it 0:19:14.640 --> 0:19:18.320 from the rooftops and as on the side. By the way, 0:19:18.920 --> 0:19:21.680 the biggest sign that no one's actually making money from 0:19:21.680 --> 0:19:23.760 this is that no one's talking about how much money 0:19:23.760 --> 0:19:28.520 they're making. Microsoft and all of these companies they love 0:19:28.640 --> 0:19:32.680 talking about making profit. They love doing that. Beyond earnings. 0:19:32.720 --> 0:19:36.160 They love talking about it instead whenever they're asked to go, oh, hey, 0:19:36.160 --> 0:19:38.520 I will do some things in the future, I need 0:19:38.560 --> 0:19:39.879 to take a phone CALLT and then they kind of 0:19:39.920 --> 0:19:43.040 disappear from the room. Amy heard CFO of Microsoft classic 0:19:43.040 --> 0:19:48.280 bullshit artists dancing around Yeah, oh net revenue increase checking 0:19:48.280 --> 0:19:52.720 a watch. It's just really sad. It's really sad because 0:19:52.720 --> 0:19:55.320 what we have here is a shared delusion, a shared 0:19:55.320 --> 0:19:58.399 delusion about a dead end technology that runs on copyright theft, 0:19:58.560 --> 0:20:01.040 one that requires a CONTINUALUS supply of capital to keep 0:20:01.119 --> 0:20:03.720 running as it provides services that are at best in 0:20:03.960 --> 0:20:06.080 essential sold to US dressed up as a kind of 0:20:06.119 --> 0:20:09.159 automation that does not exist, and it doesn't provide, costing 0:20:09.200 --> 0:20:11.840 billions and millions of dollars and continuing to do so 0:20:12.200 --> 0:20:16.000 im perpetuity. Generative AI doesn't run on money or cloud 0:20:16.040 --> 0:20:18.399 credit so much as it does on faith. And the 0:20:18.440 --> 0:20:21.240 problem is that faith, like investor capital, is actually a 0:20:21.280 --> 0:20:24.760 finite resource. And that's where I bring you one of 0:20:24.800 --> 0:20:28.320 my biggest anxieties about this industry, because I think we're 0:20:28.359 --> 0:20:30.919 in the midst of a SUBPRIMEI crisis where thousands of 0:20:30.960 --> 0:20:34.880 companies have integrated this stuff into their software at prices 0:20:34.920 --> 0:20:38.240 that are far from stable and even further from profitable 0:20:38.240 --> 0:20:41.240 for the services providing them. This concern, by the way, 0:20:41.320 --> 0:20:44.199 isn't unfounded. At the latest open ai dev Day, they 0:20:44.240 --> 0:20:46.960 said that they'd slash prices for their APIs by ninety 0:20:47.000 --> 0:20:50.080 nine percent over the previous two years, largely as tech crunchies. 0:20:50.119 --> 0:20:53.320 MAXT theorized due to price pressure from Meta and Google, 0:20:53.480 --> 0:20:55.840 both of whom want to take that API access for 0:20:56.480 --> 0:21:00.880 I assume some reason. Anyway, almost every AI powered startup 0:21:01.000 --> 0:21:03.960 uses large language model features is based on some combination 0:21:04.040 --> 0:21:07.360 of GPT or chlaud so open Ai or Anthropics models. 0:21:07.800 --> 0:21:10.680 These models are built by two companies that are deeply unprofitable. 0:21:10.720 --> 0:21:13.720 Open Ai they can lose five billion this year, Anthropic 0:21:13.880 --> 0:21:16.280 is on course to lose two point seven billion this 0:21:16.440 --> 0:21:20.000 year on much less revenue, and they all have pricing 0:21:20.040 --> 0:21:22.240