1 00:00:02,480 --> 00:00:03,560 Speaker 1: Fall Zone Media. 2 00:00:05,559 --> 00:00:08,280 Speaker 2: Hello and welcome to Better Offline. I'm your host ed 3 00:00:08,360 --> 00:00:22,200 Speaker 2: z tron. It's been just under a year and a 4 00:00:22,239 --> 00:00:25,960 Speaker 2: half since chat gpt, an AI powered chatbot launched by 5 00:00:26,079 --> 00:00:29,360 Speaker 2: so called nonprofit open Ai, ushered in a new investor 6 00:00:29,360 --> 00:00:32,680 Speaker 2: in media hype cycle around how AI would change the world. 7 00:00:33,040 --> 00:00:36,760 Speaker 2: Chat GPT's instant success made both open Ai and its CEO, 8 00:00:36,920 --> 00:00:40,840 Speaker 2: Sam Mortman overnight celebrities. As a result of chat GPT's 9 00:00:40,920 --> 00:00:44,839 Speaker 2: alleged intelligence, which seemingly allowed you to do everything from 10 00:00:44,880 --> 00:00:47,879 Speaker 2: generate an entire essay from a simple prompt to writing 11 00:00:48,360 --> 00:00:51,640 Speaker 2: entire reams of software code. You can theoretically ask it 12 00:00:51,680 --> 00:00:54,680 Speaker 2: anything and it will spit out an intelligent sounding response 13 00:00:54,880 --> 00:00:57,360 Speaker 2: thanks to being trained on terabides of text data, like 14 00:00:57,400 --> 00:01:00,560 Speaker 2: a search engine that's able to think for itself. Ah. 15 00:01:00,920 --> 00:01:04,000 Speaker 2: Big problem is chat gpt doesn't think at all. It 16 00:01:04,040 --> 00:01:09,080 Speaker 2: doesn't know anything. Chat gpt is actually probabilistic. It uses 17 00:01:09,080 --> 00:01:11,720 Speaker 2: a statistical model to generate the next piece of information 18 00:01:11,760 --> 00:01:14,360 Speaker 2: in a sequence. If you ask it what an elephant is, 19 00:01:14,560 --> 00:01:16,880 Speaker 2: it'll guess that the most likely answer to that prompt 20 00:01:17,080 --> 00:01:19,000 Speaker 2: is that an elephant is a large mammal, and then 21 00:01:19,040 --> 00:01:22,480 Speaker 2: perhaps describe its features, such as a long trunk. Chat 22 00:01:22,600 --> 00:01:25,120 Speaker 2: GPT doesn't know what an elephant is, or what a 23 00:01:25,160 --> 00:01:28,880 Speaker 2: trunk is, or what the alefant Haantai family is. It's 24 00:01:28,880 --> 00:01:32,240 Speaker 2: simply ingested enough information to reliably guess what an elephant 25 00:01:32,280 --> 00:01:35,240 Speaker 2: is meant to be, or indeed know that that's what 26 00:01:35,240 --> 00:01:38,880 Speaker 2: you're asking it. This is the technology underpinning the latest 27 00:01:38,959 --> 00:01:43,920 Speaker 2: artificial intelligence boom. It's called generative artificial intelligence, and it's 28 00:01:43,920 --> 00:01:47,840 Speaker 2: powered by large language models, and they underpin tools like 29 00:01:47,880 --> 00:01:52,559 Speaker 2: open ais, chat, GPT, Anthropics, claude x dot COM's horrifying 30 00:01:52,640 --> 00:01:57,200 Speaker 2: chatbot Grock, and of course Google's Gemini. Essentially, they're AI 31 00:01:57,280 --> 00:02:01,200 Speaker 2: systems that ingest vast quantities of written text or other data, 32 00:02:01,760 --> 00:02:05,160 Speaker 2: and then, through mathematics, try and identify patterns of relationships 33 00:02:05,200 --> 00:02:08,720 Speaker 2: between words or symbols, or basically any meaning from the 34 00:02:08,760 --> 00:02:12,600 Speaker 2: text or thing they're being fed. And it almost seems 35 00:02:12,639 --> 00:02:15,760 Speaker 2: like magic because it's able to generate this plausible, seeming 36 00:02:15,919 --> 00:02:19,920 Speaker 2: almost human content at this remarkable speed. These models are 37 00:02:19,960 --> 00:02:23,280 Speaker 2: now capable of generating text, images, and even video in 38 00:02:23,320 --> 00:02:25,959 Speaker 2: response to simple chat prompts or by learning the patterns 39 00:02:26,000 --> 00:02:29,239 Speaker 2: and structures of their data. Yet underneath the hood, there's 40 00:02:29,280 --> 00:02:34,160 Speaker 2: always something a bit wrong. Generative AI, at times authoritatively 41 00:02:34,200 --> 00:02:37,440 Speaker 2: spits out incorrect information, which can range from funny like 42 00:02:37,480 --> 00:02:40,160 Speaker 2: telling you that you can melt an egg, to outright dangerous, 43 00:02:40,280 --> 00:02:42,919 Speaker 2: like when the recently launched AI powered New York City 44 00:02:43,360 --> 00:02:46,480 Speaker 2: chatbot for small business owners started telling them that it 45 00:02:46,520 --> 00:02:49,760 Speaker 2: was legal to fire somebody for refusing to cut their dreadlocks. 46 00:02:50,560 --> 00:02:53,679 Speaker 2: This is why you'll see strange glitches in images generated 47 00:02:53,720 --> 00:02:57,200 Speaker 2: by AI hands with too many fingers, horrifying looking people 48 00:02:57,240 --> 00:02:59,600 Speaker 2: in the back of realistic looking photos, and so on 49 00:02:59,639 --> 00:03:03,160 Speaker 2: and so far. Because these models don't actually know what 50 00:03:03,200 --> 00:03:08,600 Speaker 2: anything is. They don't have meaning, they don't have consciousness 51 00:03:08,800 --> 00:03:14,320 Speaker 2: or intelligence. They're guessing, and when they guess, they sometimes hallucinate, 52 00:03:14,520 --> 00:03:17,919 Speaker 2: which I'll get too soon. And while they might be 53 00:03:18,000 --> 00:03:22,000 Speaker 2: really really good at guessing, there are effectively a very 54 00:03:22,120 --> 00:03:25,120 Speaker 2: very powerful version of auto complete. I don't know anything. 55 00:03:25,720 --> 00:03:29,320 Speaker 2: I really mean that these things aren't even intelligent. But 56 00:03:29,400 --> 00:03:31,799 Speaker 2: because these models seem like they know stuff, and they 57 00:03:31,840 --> 00:03:33,760 Speaker 2: seem to be able to do stuff, and the things 58 00:03:33,760 --> 00:03:36,880 Speaker 2: that they create almost seem right. The media and the 59 00:03:36,920 --> 00:03:40,240 Speaker 2: Vocal Investor class on Twitter have declared that large language 60 00:03:40,240 --> 00:03:44,680 Speaker 2: models would change everything to them. Llms like chat GPT 61 00:03:44,960 --> 00:03:49,760 Speaker 2: would upend entire business models, render once unassailable tech giants nunerable, 62 00:03:49,960 --> 00:03:53,680 Speaker 2: and rewrite our entire economic playbook by turning entire industries 63 00:03:53,840 --> 00:03:56,240 Speaker 2: into something you tell a chatbot to do. In a sentence, 64 00:03:57,280 --> 00:03:59,240 Speaker 2: you know, it doesn't really matter that genera ive AA 65 00:03:59,320 --> 00:04:02,480 Speaker 2: is mostly good at reams of generic slop, and that 66 00:04:02,520 --> 00:04:05,840 Speaker 2: it's also clogging services like Amazon's Kindle eBookstore. And I 67 00:04:05,880 --> 00:04:08,400 Speaker 2: guess the rest of the Internet with generative content that 68 00:04:08,480 --> 00:04:13,520 Speaker 2: doesn't matter at all, because it's kind of good. Obviously, 69 00:04:13,560 --> 00:04:17,520 Speaker 2: I'm being sarcastic. This is all very, very bad. In 70 00:04:17,560 --> 00:04:20,760 Speaker 2: the last year, there have been hundreds of muling articles 71 00:04:20,760 --> 00:04:23,960 Speaker 2: about how AI will replace everything from drive through workers 72 00:04:23,960 --> 00:04:27,680 Speaker 2: to medical professionals. Theoretically, you could just feed whatever information 73 00:04:27,720 --> 00:04:30,640 Speaker 2: a potential customer could ask for into a vast database 74 00:04:31,040 --> 00:04:33,920 Speaker 2: and have an AI chew it up, and then they 75 00:04:34,200 --> 00:04:37,240 Speaker 2: could just generate exactly the answer you'd need. AI would 76 00:04:37,320 --> 00:04:40,880 Speaker 2: just naturally slip into areas of disorganization and inefficiency and 77 00:04:40,960 --> 00:04:44,760 Speaker 2: spit out remarkable new ideas, all with minimum human input. 78 00:04:46,360 --> 00:04:48,840 Speaker 2: In every one of these stories carries with them a 79 00:04:48,920 --> 00:04:52,200 Speaker 2: shared belief one might even call it a shared hallucination. 80 00:04:53,560 --> 00:04:56,240 Speaker 2: And they all believe that generative AI will actually be 81 00:04:56,279 --> 00:04:58,560 Speaker 2: able to do these things, that it will actually be 82 00:04:58,560 --> 00:05:01,080 Speaker 2: able to replace people. And what we're seeing today is 83 00:05:01,160 --> 00:05:05,080 Speaker 2: just the beginning of our glorious automated future. But what 84 00:05:05,160 --> 00:05:09,560 Speaker 2: if it's not. What if generative AI can't actually do 85 00:05:09,760 --> 00:05:13,680 Speaker 2: much more than it can today? What if we're actually 86 00:05:13,720 --> 00:05:18,320 Speaker 2: at peak AI. In the next two episodes, I'm going 87 00:05:18,360 --> 00:05:20,200 Speaker 2: to tell you why I think that is. I'm going 88 00:05:20,279 --> 00:05:22,320 Speaker 2: to tell you how I think this whole thing falls apart. 89 00:05:33,200 --> 00:05:36,560 Speaker 2: AI's media hype train has been fairly relentless since November 90 00:05:36,560 --> 00:05:40,160 Speaker 2: twenty twenty two, when chat GBT launched, and AI champions 91 00:05:40,200 --> 00:05:43,479 Speaker 2: like open Ai CEO Sam Mortman have proven all too 92 00:05:43,520 --> 00:05:46,400 Speaker 2: willing to grease its wheels, making bold promises about what 93 00:05:46,440 --> 00:05:50,000 Speaker 2: AI could do and how today's problems are so easily overcome. 94 00:05:50,640 --> 00:05:53,400 Speaker 2: It's also helped that the tech media has largely accepted 95 00:05:53,400 --> 00:05:57,520 Speaker 2: these promises without asking basic questions like how and when 96 00:05:57,560 --> 00:06:00,320 Speaker 2: will it do this stuff? And can it do this stuff? 97 00:06:00,880 --> 00:06:03,960 Speaker 2: Don't believe me, Go and look at any interview with 98 00:06:04,080 --> 00:06:07,719 Speaker 2: Sam Moltman from the last few years, watch any of them. 99 00:06:07,760 --> 00:06:10,520 Speaker 2: In fact, just look at any AI figurehead getting interviewed 100 00:06:10,680 --> 00:06:13,080 Speaker 2: and count the amount of times they've actually received any 101 00:06:13,120 --> 00:06:17,200 Speaker 2: pushback or been asked to elaborate on any specific issue. 102 00:06:17,279 --> 00:06:20,000 Speaker 2: It's actually very rare. Let me play you one of 103 00:06:20,000 --> 00:06:23,280 Speaker 2: the few times that anyone's actually interrogated an AI person. 104 00:06:23,720 --> 00:06:26,400 Speaker 2: Specifically Joanna Stern of The Wall Street Journal, who you 105 00:06:26,480 --> 00:06:29,160 Speaker 2: might remember from the Vision Pro episode A Better Offline. 106 00:06:29,520 --> 00:06:33,800 Speaker 2: She interviewed open AI's Chief Technology Officer, Mirror Marathi about Sora, 107 00:06:33,880 --> 00:06:37,159 Speaker 2: which is open aiy's video based version of chat GPT 108 00:06:37,320 --> 00:06:40,640 Speaker 2: where you can theoretically ask it to generate videos. Just 109 00:06:40,640 --> 00:06:42,920 Speaker 2: to be clear, it's unreleased and unclear whether it'll ever 110 00:06:42,960 --> 00:06:45,040 Speaker 2: actually get released, and the videos look good at first, 111 00:06:45,279 --> 00:06:47,880 Speaker 2: then they look really weird. But just listen to this 112 00:06:47,960 --> 00:06:52,600 Speaker 2: particular question. It's Joanna asking Mirror, the CTO of open 113 00:06:52,640 --> 00:06:56,400 Speaker 2: ai and eighty billion dollar AI company, Hey, did you 114 00:06:56,520 --> 00:06:59,040 Speaker 2: train on YouTube? What data was. 115 00:06:59,040 --> 00:06:59,640 Speaker 3: Used to train? 116 00:07:01,200 --> 00:07:07,479 Speaker 1: We used publicly available data and licensed data, so videos 117 00:07:07,520 --> 00:07:08,000 Speaker 1: on YouTube? 118 00:07:09,720 --> 00:07:11,520 Speaker 2: Now, I encourage you to go and look up this 119 00:07:11,600 --> 00:07:16,000 Speaker 2: clip because at this point Maurti makes the strangest face 120 00:07:16,280 --> 00:07:18,000 Speaker 2: I've ever seen in a tech interview. 121 00:07:19,040 --> 00:07:19,840 Speaker 1: I'm actually not. 122 00:07:19,760 --> 00:07:26,520 Speaker 3: Sure about that okay, videos from Facebook, Instagram. 123 00:07:26,640 --> 00:07:32,360 Speaker 1: You know, if they were publicly available available, publicly available 124 00:07:32,400 --> 00:07:38,200 Speaker 1: to use, there might be the data, but I'm not sure. 125 00:07:38,280 --> 00:07:39,480 Speaker 1: I'm not confident about it. 126 00:07:39,560 --> 00:07:41,880 Speaker 3: What about Shutterstock, I know you guys have a deal 127 00:07:41,960 --> 00:07:44,200 Speaker 3: with them. 128 00:07:44,280 --> 00:07:46,080 Speaker 1: I'm just not going to go into the details of 129 00:07:46,400 --> 00:07:49,360 Speaker 1: the data that was that was used, but it was 130 00:07:49,400 --> 00:07:51,240 Speaker 1: publicly available or licensed data. 131 00:07:52,160 --> 00:07:54,640 Speaker 2: The remarkable part about this interview is that it's a 132 00:07:54,760 --> 00:07:58,120 Speaker 2: relatively simple question. You, as the CTO of an eighty 133 00:07:58,160 --> 00:08:01,960 Speaker 2: billion dollar AI company, what training data did you use 134 00:08:02,000 --> 00:08:05,280 Speaker 2: to train your model? Did you use YouTube? So yes 135 00:08:05,400 --> 00:08:09,560 Speaker 2: or no? Question? Mirror mirror, answer the bloody question, mirror 136 00:08:11,280 --> 00:08:13,880 Speaker 2: all right, right? The answer is, of course Open Ai 137 00:08:14,400 --> 00:08:17,560 Speaker 2: likely trained it's video generating model, Sora on YouTube videos, 138 00:08:18,240 --> 00:08:21,680 Speaker 2: which might be why they're yet to launch it. And 139 00:08:21,920 --> 00:08:26,560 Speaker 2: the videos generated by Soora also feature some remarkably similar 140 00:08:26,560 --> 00:08:30,440 Speaker 2: images to say, SpongeBob SquarePants, and I wouldn't be surprised 141 00:08:30,480 --> 00:08:33,960 Speaker 2: if they carry with them multiple weird biases about race 142 00:08:34,000 --> 00:08:37,559 Speaker 2: and gender that we'll see in the future. But also 143 00:08:37,640 --> 00:08:41,360 Speaker 2: when you watch these videos, much like most generative AI content, 144 00:08:41,600 --> 00:08:44,120 Speaker 2: there's something a bit off about them. In The Wall 145 00:08:44,120 --> 00:08:46,760 Speaker 2: Street Journal's interview. You get to see some of the 146 00:08:46,960 --> 00:08:48,920 Speaker 2: prompts that were used and some of the videos that 147 00:08:48,960 --> 00:08:51,920 Speaker 2: came out, and you see crazy things happening, like a 148 00:08:52,000 --> 00:08:56,440 Speaker 2: robot completely changing shape as it turns, cars disappearing and 149 00:08:56,520 --> 00:09:00,560 Speaker 2: appearing behind the robot. It's not very good. It seems 150 00:09:00,559 --> 00:09:02,960 Speaker 2: cool at first. If you squint really hard, it looks real, 151 00:09:03,320 --> 00:09:06,160 Speaker 2: but there's always something off. And that's because, as I've 152 00:09:06,160 --> 00:09:08,360 Speaker 2: said before, these models don't know anything and don't know 153 00:09:08,360 --> 00:09:10,040 Speaker 2: what at robot looks. They can make a really good 154 00:09:10,080 --> 00:09:16,400 Speaker 2: guess though. Anyway, Sterne's interview with Marati of Open AI 155 00:09:16,760 --> 00:09:19,480 Speaker 2: is a great example of how the entire AI artifst 156 00:09:19,520 --> 00:09:23,199 Speaker 2: falls apart at the slightest touch, because it's fundamentally flawed 157 00:09:23,480 --> 00:09:27,000 Speaker 2: and not actually able to deliver the society defining promises 158 00:09:27,160 --> 00:09:29,680 Speaker 2: that Sam Morltman and the venture capital sect would have 159 00:09:29,760 --> 00:09:33,400 Speaker 2: you believe. In a year and a half, despite billions 160 00:09:33,400 --> 00:09:37,320 Speaker 2: of dollars of investment, despite every major media outlet claiming otherwise, 161 00:09:37,559 --> 00:09:42,400 Speaker 2: generative artificial intelligence has proven itself incapable of replacing or 162 00:09:42,440 --> 00:09:46,920 Speaker 2: even meaningfully enhancing human work. And the thing is, all 163 00:09:46,920 --> 00:09:49,600 Speaker 2: of these problems I'm talking about with generative AI all 164 00:09:49,640 --> 00:09:52,800 Speaker 2: of these hallucinations, all of these weird artifacts that are 165 00:09:52,800 --> 00:09:56,080 Speaker 2: popping up throughout these videos, the weird mistakes that the 166 00:09:56,120 --> 00:09:59,640 Speaker 2: texts that are popped out by chat GPT have, all 167 00:09:59,679 --> 00:10:03,679 Speaker 2: of these. The problems are problems that aren't necessarily just technological. 168 00:10:04,120 --> 00:10:10,080 Speaker 2: Their physics, they're mathematics. These aren't things you can just outrun. 169 00:10:11,040 --> 00:10:14,760 Speaker 2: And I believe that there are four intractable problems that 170 00:10:14,840 --> 00:10:18,080 Speaker 2: will stop generative AI from progressing much further than it 171 00:10:18,160 --> 00:10:21,440 Speaker 2: is today. The first is, of course, its energy demands, 172 00:10:21,480 --> 00:10:25,120 Speaker 2: the massive amounts of power requires. The second are its 173 00:10:25,120 --> 00:10:29,640 Speaker 2: computational demands, the amount of compute power it requires to 174 00:10:29,679 --> 00:10:34,280 Speaker 2: even crunch the simplest things out of chat GPT, It's hallucinations, 175 00:10:34,320 --> 00:10:37,680 Speaker 2: the authoritative failures it makes when it spits out nonsense 176 00:10:37,760 --> 00:10:41,520 Speaker 2: or creates a human hand with eighteen fingers, and of 177 00:10:41,559 --> 00:10:44,160 Speaker 2: course the fact that these large language models have an 178 00:10:44,160 --> 00:10:49,240 Speaker 2: insatiable hunger for more training data. Now, let me break 179 00:10:49,240 --> 00:10:54,560 Speaker 2: that down. Large language models are extremely technologically and environmentally demanding. 180 00:10:55,040 --> 00:10:57,160 Speaker 2: The New York Are reported in March twenty twenty four 181 00:10:57,320 --> 00:11:00,280 Speaker 2: the CHATGBT uses more than half a million kis what 182 00:11:00,360 --> 00:11:03,240 Speaker 2: hours of electricity to respond to the two hundred million 183 00:11:03,280 --> 00:11:07,000 Speaker 2: requests it receives in a day, or seventeen thousand times 184 00:11:07,040 --> 00:11:09,839 Speaker 2: the amount that the average American household uses in a day, 185 00:11:10,080 --> 00:11:12,600 Speaker 2: and others have suggested it might be as high as 186 00:11:12,679 --> 00:11:18,079 Speaker 2: thirty three thousand households worth. Generative AI models demand specialist 187 00:11:18,120 --> 00:11:21,760 Speaker 2: chips called graphics processing units, typically a souped up version 188 00:11:21,800 --> 00:11:23,959 Speaker 2: of the technology used to drive the graphics in a 189 00:11:24,000 --> 00:11:27,480 Speaker 2: gaming console, albeit at a much higher cost, with each 190 00:11:27,640 --> 00:11:31,199 Speaker 2: one costing tens of thousands of dollars each. They do 191 00:11:31,240 --> 00:11:34,600 Speaker 2: this because large language models like chat GPT are highly 192 00:11:34,640 --> 00:11:38,400 Speaker 2: computationally intensive. I'm going to break that down. Don't worry. 193 00:11:38,880 --> 00:11:41,760 Speaker 2: When you ask chat GPT a question, it tokenizes it, 194 00:11:42,240 --> 00:11:45,359 Speaker 2: breaking it down into smaller parts for the model to understand. 195 00:11:46,760 --> 00:11:50,120 Speaker 2: It then feeds these tokens into various mechanisms that help 196 00:11:50,160 --> 00:11:52,480 Speaker 2: it understand the meaning of the thing you asked it 197 00:11:52,520 --> 00:11:56,040 Speaker 2: to do based on the parameters that it learned in training. 198 00:11:56,240 --> 00:11:58,920 Speaker 2: Chat GPT generates a response by predicting the most likely 199 00:11:58,960 --> 00:12:00,800 Speaker 2: sequence of things that you might want it to do. 200 00:12:01,160 --> 00:12:03,720 Speaker 2: An answer to a question, an image, so on, and 201 00:12:03,760 --> 00:12:07,560 Speaker 2: so forth. Each one of these steps is extremely demanding, 202 00:12:07,720 --> 00:12:11,360 Speaker 2: processing hundreds of billions of these parameters learned patterns from 203 00:12:11,520 --> 00:12:14,440 Speaker 2: ingesting training data such as how the English language works 204 00:12:14,480 --> 00:12:16,800 Speaker 2: or what a dog looks like to produce even the 205 00:12:16,840 --> 00:12:21,280 Speaker 2: simplest thing. Training these models is equally intensive, requiring chat 206 00:12:21,280 --> 00:12:24,440 Speaker 2: GPT to process massive amounts of data. Another problem I'll 207 00:12:24,440 --> 00:12:27,320 Speaker 2: get to in a bit adjusting those hundreds of billions 208 00:12:27,320 --> 00:12:30,000 Speaker 2: of parameters and developing new ones based on what the 209 00:12:30,120 --> 00:12:33,439 Speaker 2: data says as it quote unquote learns more. Though as 210 00:12:33,440 --> 00:12:37,120 Speaker 2: we're clear, chap GPT doesn't learn anything. It just makes 211 00:12:37,160 --> 00:12:40,960 Speaker 2: new parameters to read things. A model like chat GBT grows, 212 00:12:41,000 --> 00:12:44,199 Speaker 2: making it more complex, which in turn requires more data 213 00:12:44,240 --> 00:12:47,520 Speaker 2: to train on and more compute power to both ingest 214 00:12:47,559 --> 00:12:50,439 Speaker 2: the data, create more parameters and turn it into something 215 00:12:50,480 --> 00:12:54,240 Speaker 2: resembling an answer, And because it doesn't know anything, it's 216 00:12:54,280 --> 00:12:57,200 Speaker 2: suggesting the most likely to be correct answer, which leads 217 00:12:57,200 --> 00:13:00,360 Speaker 2: it to hallucinating and correct things that, based on proper ability, 218 00:13:00,600 --> 00:13:03,920 Speaker 2: kind of seem like the right thing to say. These 219 00:13:03,960 --> 00:13:07,200 Speaker 2: hallucinations are the dirty little secret of generative AI, and 220 00:13:07,240 --> 00:13:09,640 Speaker 2: are impossible to avoid thanks to the fact that every 221 00:13:09,720 --> 00:13:13,480 Speaker 2: single thing these models say is a mathematical equation rather 222 00:13:13,559 --> 00:13:17,520 Speaker 2: than any kind of intellectual exercise. If you ask CHATGPT 223 00:13:17,840 --> 00:13:19,560 Speaker 2: how many days there are in a week. It doesn't 224 00:13:19,600 --> 00:13:21,720 Speaker 2: know that there are seven days, but it's been trained 225 00:13:21,720 --> 00:13:24,360 Speaker 2: on patterns of language and generates a result based on 226 00:13:24,400 --> 00:13:27,600 Speaker 2: those patterns, which at times can be correct and can 227 00:13:27,640 --> 00:13:32,240 Speaker 2: also be wrong. There's no way of fixing this problem. 228 00:13:33,000 --> 00:13:36,200 Speaker 2: You can mitigate it, you can make it less likely 229 00:13:36,200 --> 00:13:39,640 Speaker 2: it will mess up, but hallucinations will happen because there 230 00:13:39,679 --> 00:13:43,160 Speaker 2: is no consciousness. It is not learning anything. This thing 231 00:13:43,320 --> 00:13:48,120 Speaker 2: has no knowledge. More computing power would allow it more 232 00:13:48,160 --> 00:13:50,720 Speaker 2: parameters to give it more rules, so that a generative 233 00:13:50,720 --> 00:13:53,200 Speaker 2: AI will be more likely to give a correct answer, 234 00:13:53,679 --> 00:13:56,880 Speaker 2: but there's no eliminating them, and doing so may require 235 00:13:56,920 --> 00:14:01,160 Speaker 2: more computing power than actually exists or is possible without 236 00:14:01,160 --> 00:14:04,400 Speaker 2: an AI of consciousness, an impossible dream known as average 237 00:14:04,440 --> 00:14:07,760 Speaker 2: generalized intelligence that Sam Mortman would have you believe is imminent. 238 00:14:08,280 --> 00:14:13,280 Speaker 2: There's really no solving hallucinations. When you answer questions using probability, 239 00:14:13,920 --> 00:14:16,880 Speaker 2: you're always going to have mistakes because you're not actually 240 00:14:16,960 --> 00:14:21,560 Speaker 2: answering them using knowledge, intellect, or experience. You're using dice rolls. 241 00:14:22,200 --> 00:14:25,520 Speaker 2: It's a bloody game of dungeons and dragons. We turned 242 00:14:25,520 --> 00:14:30,200 Speaker 2: in Carter into dungeons and dragons anyway. A newly published 243 00:14:30,240 --> 00:14:33,280 Speaker 2: paper by Tepo Felon and Matthias Holweg of the University 244 00:14:33,280 --> 00:14:36,440 Speaker 2: of Oxford agrees funding that large language models like chat 245 00:14:36,480 --> 00:14:40,680 Speaker 2: GPT are incapable of generating new knowledge. It's a remarkably 246 00:14:40,760 --> 00:14:44,280 Speaker 2: in depth rundown of the fundamental differences between a large 247 00:14:44,320 --> 00:14:47,160 Speaker 2: language model and a human brain, and it combines both 248 00:14:47,200 --> 00:14:51,080 Speaker 2: psychological and mathematical research going back to child psychology as 249 00:14:51,120 --> 00:14:54,480 Speaker 2: well the basic building blocks of how we consume and 250 00:14:54,560 --> 00:14:57,080 Speaker 2: learn things and how we make decisions. As a result, 251 00:14:57,720 --> 00:15:01,440 Speaker 2: the paper title theory is or You Need AI Human 252 00:15:01,480 --> 00:15:05,360 Speaker 2: Cognition and Decision Making argues that AI's data and prediction 253 00:15:05,440 --> 00:15:09,920 Speaker 2: based orientation is an incomplete view of human cognition and 254 00:15:10,160 --> 00:15:13,560 Speaker 2: the forward thinking theorizing of the human mind. In Layman's terms, 255 00:15:13,880 --> 00:15:16,200 Speaker 2: the mess of the information we've learned over our lives, 256 00:15:16,200 --> 00:15:19,560 Speaker 2: our experiences, and our ability to look forward and consider 257 00:15:19,880 --> 00:15:23,560 Speaker 2: the future is just fundamentally different to a model that 258 00:15:23,640 --> 00:15:27,240 Speaker 2: predicts things only based off of past data. Think of 259 00:15:27,280 --> 00:15:30,480 Speaker 2: it like this, If you've read a book and you 260 00:15:30,560 --> 00:15:33,160 Speaker 2: might think about writing a new book based on those ideas, 261 00:15:33,600 --> 00:15:36,320 Speaker 2: you're not remembering every part of the book. You don't 262 00:15:36,360 --> 00:15:40,560 Speaker 2: have a perfect memory, and you're also constantly thinking about 263 00:15:40,600 --> 00:15:43,040 Speaker 2: things as your day goes on. The human brain is 264 00:15:43,040 --> 00:15:47,080 Speaker 2: a goddamn mess. Generative AI is in some level stuck 265 00:15:47,120 --> 00:15:50,960 Speaker 2: in amber. Though the billions of parameters might change, the 266 00:15:51,080 --> 00:15:54,080 Speaker 2: data never does the way it consumes the data. May 267 00:15:54,160 --> 00:15:58,800 Speaker 2: be that the data doesn't change. In essence, generative AI 268 00:15:58,920 --> 00:16:01,240 Speaker 2: is held back by the fact that it can't consider 269 00:16:01,280 --> 00:16:04,760 Speaker 2: the future and is actually permanently mired in the data 270 00:16:04,800 --> 00:16:09,480 Speaker 2: of the past. Their largest problem might be a fast, 271 00:16:09,520 --> 00:16:12,760 Speaker 2: simpler one, a farcillio, or a kind of an ironic one. 272 00:16:13,920 --> 00:16:16,240 Speaker 2: There might not be enough data for these bloody things 273 00:16:16,280 --> 00:16:31,080 Speaker 2: to actually train on. While the Internet may at times 274 00:16:31,120 --> 00:16:34,440 Speaker 2: feel limitless, a researcher recently told The Wall Street Journal 275 00:16:35,120 --> 00:16:37,840 Speaker 2: that only a tenth of the most commonly used web 276 00:16:37,920 --> 00:16:40,800 Speaker 2: data set, the common Crawl, are freely available. Two hundred 277 00:16:40,800 --> 00:16:43,880 Speaker 2: and fifty billion page dump of the web's information is 278 00:16:43,920 --> 00:16:47,520 Speaker 2: actually of high enough quality data for large language models 279 00:16:47,520 --> 00:16:51,080 Speaker 2: like CHATGBT to actually train on. Putting aside the fact 280 00:16:51,080 --> 00:16:52,880 Speaker 2: that I can't find a single definition of what high 281 00:16:52,960 --> 00:16:58,400 Speaker 2: quality actually means, the researcher pat Blow Vilobos suggested that 282 00:16:58,440 --> 00:17:01,360 Speaker 2: the next version of chat GBT would require more than 283 00:17:01,480 --> 00:17:04,400 Speaker 2: five times the amount of data it took to train 284 00:17:04,400 --> 00:17:07,159 Speaker 2: in its previous version, GPT four. The new one is 285 00:17:07,200 --> 00:17:10,240 Speaker 2: called GPT five. By the way, and other researchers have 286 00:17:10,320 --> 00:17:12,520 Speaker 2: suggested that AI companies are going to run out of 287 00:17:12,560 --> 00:17:17,960 Speaker 2: training data in the next two years. Now. That sounds dire, 288 00:17:18,040 --> 00:17:20,280 Speaker 2: but don't worry. They've come up with a very funny 289 00:17:20,359 --> 00:17:24,440 Speaker 2: and extremely stupid idea to fix it. One specifically posed 290 00:17:24,480 --> 00:17:27,200 Speaker 2: by The Wall Street Journal is that the AR companies 291 00:17:27,200 --> 00:17:29,800 Speaker 2: are going to create their own synthetic data to train 292 00:17:29,880 --> 00:17:33,440 Speaker 2: their models, a computer science version of inbreeding. The researcher 293 00:17:33,680 --> 00:17:38,920 Speaker 2: Jason Stadowski calls habsburg AI. This is, of course, an 294 00:17:38,960 --> 00:17:43,560 Speaker 2: absolutely terrible idea. A research paper from last year found 295 00:17:43,600 --> 00:17:47,200 Speaker 2: that feeding model generated data into models to train them 296 00:17:47,240 --> 00:17:51,000 Speaker 2: creates something called model collapse, a degenerative learning process where 297 00:17:51,040 --> 00:17:54,040 Speaker 2: models start forgetting improbable events over time as the model 298 00:17:54,040 --> 00:17:57,560 Speaker 2: becomes poison with its own projection of reality. The paper, 299 00:17:57,800 --> 00:18:00,640 Speaker 2: called the Curse of recursion. Training on generated data makes 300 00:18:00,680 --> 00:18:03,960 Speaker 2: models forgam highlights an example where feeding a generative AI 301 00:18:04,040 --> 00:18:07,400 Speaker 2: its own data eventually destroys its ability to answer questions, 302 00:18:07,720 --> 00:18:10,840 Speaker 2: and within nine generations. One answered a simple prompt about 303 00:18:10,920 --> 00:18:14,560 Speaker 2: architecture with an insane screed about jack rabbits full of 304 00:18:14,840 --> 00:18:19,800 Speaker 2: at symbols and weird characters. So not to worry again. 305 00:18:20,280 --> 00:18:22,560 Speaker 2: The tech overlords have come up with a great idea 306 00:18:22,600 --> 00:18:25,680 Speaker 2: to fix this problem. Their common retort to the problem 307 00:18:25,760 --> 00:18:29,400 Speaker 2: of synthetic data is that you could use another generative 308 00:18:29,400 --> 00:18:32,600 Speaker 2: AI to monitor the synthetic data being fed into all 309 00:18:32,680 --> 00:18:36,400 Speaker 2: model to make sure it's right. At this point, I'd 310 00:18:36,440 --> 00:18:39,080 Speaker 2: like to get slightly angry. Are you kidding me? Are 311 00:18:39,119 --> 00:18:41,320 Speaker 2: you fucking kidding me? You're saying that the way to 312 00:18:41,359 --> 00:18:44,720 Speaker 2: make sure the data generated by an unreliable generative AI 313 00:18:45,200 --> 00:18:48,560 Speaker 2: is to use another generative AI, one with the same 314 00:18:48,640 --> 00:18:53,240 Speaker 2: goddamn problems, which also hallucinates information that knows nothing. You're 315 00:18:53,240 --> 00:18:55,359 Speaker 2: going to use that AI to monitor whether the data 316 00:18:55,640 --> 00:18:58,000 Speaker 2: that is created by an AI is any good? Are 317 00:18:58,000 --> 00:19:01,800 Speaker 2: you completely insane? You insane? You're going to feed the 318 00:19:01,840 --> 00:19:04,240 Speaker 2: crap from the crap machine into another crap machine to 319 00:19:04,280 --> 00:19:07,480 Speaker 2: make it not make crap? Why am I reading journalists 320 00:19:07,520 --> 00:19:10,960 Speaker 2: credulously printing this ridiculous solution in the New York Goddamn 321 00:19:11,000 --> 00:19:14,440 Speaker 2: Times every time? Every time, these bubbles are inflated because 322 00:19:14,480 --> 00:19:17,200 Speaker 2: tech executives are able to get their half ass, half 323 00:19:17,240 --> 00:19:20,399 Speaker 2: baked solutions parroted by reporters who should know better. You 324 00:19:20,440 --> 00:19:22,199 Speaker 2: don't have to give them the better affair of the 325 00:19:22,200 --> 00:19:24,840 Speaker 2: goddamn fucking doubt. This is how we got the bloody 326 00:19:24,840 --> 00:19:31,520 Speaker 2: meta us. Pardon me, I've calmed down now. Anyway. Anyway, 327 00:19:31,920 --> 00:19:34,920 Speaker 2: if you're worried about model collapse, you're already too late, 328 00:19:35,200 --> 00:19:37,960 Speaker 2: as these models are likely already being fed their own data. 329 00:19:38,840 --> 00:19:41,679 Speaker 2: You see, these models are trained on the web, as 330 00:19:41,720 --> 00:19:45,119 Speaker 2: I previously told you, and they're desperate. They need data. 331 00:19:45,200 --> 00:19:47,960 Speaker 2: They need more stuff. They need more stuff to ingest 332 00:19:48,040 --> 00:19:51,439 Speaker 2: so they can spit out more stuff. The problem is 333 00:19:51,440 --> 00:19:54,240 Speaker 2: that these machines are purpose built to make a lot 334 00:19:54,240 --> 00:19:58,600 Speaker 2: of content, and so the Web's already being filled with 335 00:19:58,680 --> 00:20:03,560 Speaker 2: generative AI. Generative AI is already spamming the Internet. A 336 00:20:03,680 --> 00:20:06,840 Speaker 2: report from four oh four Media from last week said 337 00:20:06,880 --> 00:20:11,280 Speaker 2: that Google Books has already started to index several different 338 00:20:11,280 --> 00:20:14,360 Speaker 2: works that were potentially written by AI, featuring the hallmark 339 00:20:14,400 --> 00:20:17,720 Speaker 2: generic writing tropes of these models. Four or four Media 340 00:20:17,760 --> 00:20:19,879 Speaker 2: also reports that the same thing is happening over at 341 00:20:19,880 --> 00:20:22,879 Speaker 2: Google Scholar their index of scientific papers with one hundred 342 00:20:22,880 --> 00:20:25,840 Speaker 2: and fifteen different articles featuring the phrase as of my 343 00:20:25,920 --> 00:20:30,280 Speaker 2: last knowledge update a specific phrase spat out by generative models. 344 00:20:32,080 --> 00:20:34,560 Speaker 2: This is really bad, by the way, and this is 345 00:20:34,640 --> 00:20:36,719 Speaker 2: only going to get worse. When you have an Internet 346 00:20:36,800 --> 00:20:39,960 Speaker 2: economy that is built so that the people that can 347 00:20:40,000 --> 00:20:42,280 Speaker 2: put the most out there will probably get the most traffic. 348 00:20:42,840 --> 00:20:44,800 Speaker 2: They're going to use these tools. These tools are great 349 00:20:44,800 --> 00:20:46,720 Speaker 2: for that. If you don't give a rap fuck about 350 00:20:46,720 --> 00:20:48,680 Speaker 2: the quality, this is the best thing in the world 351 00:20:48,680 --> 00:20:54,160 Speaker 2: for you. And that's the thing. This is a problem 352 00:20:54,920 --> 00:20:58,119 Speaker 2: both created and caused by these models. You see. The 353 00:20:58,160 --> 00:21:00,880 Speaker 2: other dirty little secret of generative II is that these 354 00:21:00,920 --> 00:21:04,920 Speaker 2: models unashamedly plagiarize the entire web, leading outlets like The 355 00:21:04,960 --> 00:21:07,480 Speaker 2: New York Times and authors like John Grisham to sue 356 00:21:07,520 --> 00:21:11,719 Speaker 2: open ai for plagiarism. While open ai won't reveal exactly 357 00:21:11,720 --> 00:21:14,000 Speaker 2: what their training data is, the New York Times is 358 00:21:14,080 --> 00:21:18,400 Speaker 2: able to successfully make chat gpt reproduce content from the newspaper, 359 00:21:18,640 --> 00:21:21,119 Speaker 2: and the company has repeatedly said that it trains on 360 00:21:21,160 --> 00:21:24,960 Speaker 2: publicly available data from the Internet, which will naturally include 361 00:21:25,040 --> 00:21:29,080 Speaker 2: things like Google scholar and Google Books. The Times also 362 00:21:29,160 --> 00:21:31,840 Speaker 2: reports that open ai has become so desperate for data 363 00:21:31,920 --> 00:21:35,040 Speaker 2: that they've used their whisper tool to transcribe YouTube videos 364 00:21:35,240 --> 00:21:39,600 Speaker 2: into texts to feed into chat GPT's training data. Pretty sure, 365 00:21:39,640 --> 00:21:44,840 Speaker 2: that's plagiarism, but who am I to tell you? And 366 00:21:44,920 --> 00:21:47,840 Speaker 2: as the web gets increasingly pumped full of this generative content, 367 00:21:47,920 --> 00:21:50,639 Speaker 2: these models are going to just start eating their own swill, 368 00:21:51,119 --> 00:21:56,080 Speaker 2: slowly corrupting themselves in a kind of ironic death. According 369 00:21:56,080 --> 00:21:58,640 Speaker 2: to Zakhar Schumelov, one of the authors of the model 370 00:21:58,680 --> 00:22:01,639 Speaker 2: collapsed paper at the University of Cambridge, the unique problem 371 00:22:01,640 --> 00:22:05,000 Speaker 2: that synthetic data creates is that it lacks human errors. 372 00:22:05,760 --> 00:22:08,080 Speaker 2: Human made training data, by the nature of it being 373 00:22:08,080 --> 00:22:10,879 Speaker 2: written by a human, includes errors and imperfections, and models 374 00:22:10,960 --> 00:22:13,919 Speaker 2: need to be robust to such errors. So what do 375 00:22:13,960 --> 00:22:15,960 Speaker 2: we do if models are trained off of content created 376 00:22:16,000 --> 00:22:20,000 Speaker 2: without them? Do we introduce the errors ourselves? How many 377 00:22:20,080 --> 00:22:22,680 Speaker 2: errors are there? How do we can introduce them more? 378 00:22:22,800 --> 00:22:25,959 Speaker 2: And indeed, what are the errors? What do they look like? 379 00:22:26,040 --> 00:22:28,680 Speaker 2: Do we even know? Are we conscious of the errors 380 00:22:28,680 --> 00:22:33,960 Speaker 2: in the human language that make us human? The models aren't? Well, 381 00:22:33,960 --> 00:22:37,520 Speaker 2: maybe they are. It's kind of unclear. It's kind of 382 00:22:37,560 --> 00:22:41,399 Speaker 2: tough to express how deeply dangerous the synthetic data idea 383 00:22:41,600 --> 00:22:45,520 Speaker 2: is for AI models like chat, GPT and Claude are 384 00:22:45,520 --> 00:22:48,919 Speaker 2: deeply dependent on training data to improve their outputs, and 385 00:22:48,960 --> 00:22:52,040 Speaker 2: their very existence is actively impeding the creation of the 386 00:22:52,160 --> 00:22:55,080 Speaker 2: very thing they need to survive. While publishers like Axel 387 00:22:55,160 --> 00:22:57,600 Speaker 2: Springer have cut deals to license their companies data to 388 00:22:57,640 --> 00:23:00,480 Speaker 2: chat GPT for training purposes, this money is flowing to 389 00:23:00,480 --> 00:23:02,960 Speaker 2: the writers that create the content that open Ai and 390 00:23:03,040 --> 00:23:07,000 Speaker 2: Anthropic need to grow their models much further. In fact, 391 00:23:07,440 --> 00:23:10,000 Speaker 2: I don't think you're going to see more journalists get 392 00:23:10,080 --> 00:23:12,879 Speaker 2: hired as a result of these deals, which kind of 393 00:23:12,920 --> 00:23:17,919 Speaker 2: makes them a little bit stupid. This puts AI companies 394 00:23:17,920 --> 00:23:20,359 Speaker 2: in a kind of Kafka esque bind, but they can't 395 00:23:20,359 --> 00:23:23,040 Speaker 2: really improve a tool for automating the creation of content 396 00:23:23,240 --> 00:23:26,639 Speaker 2: without human beings creating more content than they've ever created before, 397 00:23:28,080 --> 00:23:31,359 Speaker 2: just as said tool actively crowds out human made data. 398 00:23:32,320 --> 00:23:36,719 Speaker 2: It's a little silly. The solution to these problems, if 399 00:23:36,760 --> 00:23:39,679 Speaker 2: you ask open AI's Sam Altman, is always more money 400 00:23:39,680 --> 00:23:42,480 Speaker 2: and power, which is why the information reports he is 401 00:23:42,480 --> 00:23:45,600 Speaker 2: trying to convince open ai investor Microsoft to build him 402 00:23:45,800 --> 00:23:49,120 Speaker 2: and I'm not kidding, and one hundred billion dollars supercomputer 403 00:23:49,440 --> 00:23:54,760 Speaker 2: called Stargate. This massive series of interconnected machines will require 404 00:23:55,119 --> 00:23:58,480 Speaker 2: entirely new ways to mountain cool processing units and is 405 00:23:58,640 --> 00:24:03,159 Speaker 2: entirely contingent on pen AI's ability to meaningfully improve chet GPT, 406 00:24:04,080 --> 00:24:08,280 Speaker 2: something Sam Mortman claims isn't possible without more computing power. 407 00:24:09,720 --> 00:24:12,840 Speaker 2: To be clear, open Air already failed to build a 408 00:24:12,880 --> 00:24:20,800 Speaker 2: more efficient model, dubbed Arakis, which ended up getting mothballed 409 00:24:20,800 --> 00:24:24,800 Speaker 2: because it wasn't more efficient. It's also important to note 410 00:24:24,800 --> 00:24:28,560 Speaker 2: that every major cloud company now has inextricably tied themselves 411 00:24:28,720 --> 00:24:32,480 Speaker 2: to the generative AI movement. Google and Amazon have invested 412 00:24:32,560 --> 00:24:36,359 Speaker 2: billions into chat GPT competitor Anthropic, and both claim to 413 00:24:36,359 --> 00:24:40,480 Speaker 2: be Anthropic's primary cloud provider, though isn't really obvious which 414 00:24:40,520 --> 00:24:44,000 Speaker 2: one is. In doing so, they've guaranteed, according to a 415 00:24:44,000 --> 00:24:47,000 Speaker 2: source of mind, about seven hundred and fifty million dollars 416 00:24:47,040 --> 00:24:49,879 Speaker 2: a year of revenue for Google's cloud and eight hundred 417 00:24:49,960 --> 00:24:53,520 Speaker 2: million dollars a year of revenue for Amazon Web Services 418 00:24:53,560 --> 00:24:57,040 Speaker 2: the cloud service from Amazon by mandating that Anthropic uses 419 00:24:57,119 --> 00:25:00,639 Speaker 2: their services to power their clawed model. This similar to 420 00:25:00,680 --> 00:25:04,040 Speaker 2: the thirteen billion dollar investment that Microsoft gave open AI 421 00:25:04,480 --> 00:25:06,840 Speaker 2: last year, most of which was made up of credits 422 00:25:06,840 --> 00:25:09,960 Speaker 2: for Microsoft's cy or cloud, And I somehow doubt that 423 00:25:10,000 --> 00:25:13,280 Speaker 2: Microsoft is going to be the noble party that goes 424 00:25:13,280 --> 00:25:15,240 Speaker 2: on the earnings and says, well, we don't want to 425 00:25:15,320 --> 00:25:17,960 Speaker 2: count the credits that we gave open out want to 426 00:25:17,960 --> 00:25:19,600 Speaker 2: be fair. No, they're going to mash that shit right 427 00:25:19,640 --> 00:25:23,840 Speaker 2: back into their revenue. Kind of a con kind of 428 00:25:23,880 --> 00:25:25,639 Speaker 2: makes me angry when I think about it too. Anyway, 429 00:25:25,720 --> 00:25:27,520 Speaker 2: let me just put that aside and not going to 430 00:25:27,560 --> 00:25:30,959 Speaker 2: get pissed off again. Look, I'm surprised more people aren't 431 00:25:31,359 --> 00:25:35,040 Speaker 2: really upset about this very incestuous relationship between big tech 432 00:25:35,119 --> 00:25:40,280 Speaker 2: and this supposedly independent generative AI movement. Microsoft, Google, and 433 00:25:40,320 --> 00:25:44,359 Speaker 2: Amazon have effectively handed cash to one or two companies 434 00:25:44,800 --> 00:25:46,800 Speaker 2: that will eventually hand the cash back to them in 435 00:25:46,840 --> 00:25:49,240 Speaker 2: exchange for cloud services that are necessary to make their 436 00:25:49,240 --> 00:25:52,560 Speaker 2: companies work, and all three big tech firms are spending 437 00:25:52,640 --> 00:25:56,719 Speaker 2: billions to expand their data center operations to capture this 438 00:25:57,160 --> 00:26:02,440 Speaker 2: theoretical demand from generative AI. Every penny the open AI 439 00:26:02,640 --> 00:26:05,080 Speaker 2: or anthropic makes will now flow back to one of 440 00:26:05,200 --> 00:26:08,280 Speaker 2: three big tech firms, even more so in the case 441 00:26:08,280 --> 00:26:12,080 Speaker 2: of open AI, because Microsoft's investment entitles Microsoft to a 442 00:26:12,119 --> 00:26:18,040 Speaker 2: share of any future profits from open ai and chat GBT. Yeah, 443 00:26:18,080 --> 00:26:20,120 Speaker 2: it doesn't even really matter if they make one, because 444 00:26:20,119 --> 00:26:23,520 Speaker 2: Big tech wins either way. Anthropic has to use Google 445 00:26:23,560 --> 00:26:28,600 Speaker 2: Cloud and Amazon Web services, open Ai has to use 446 00:26:28,800 --> 00:26:32,080 Speaker 2: Microsoft's z your Cloud, and Microsoft is actively selling open 447 00:26:32,119 --> 00:26:35,160 Speaker 2: AI's models to their zero cloud customers. And every time 448 00:26:35,640 --> 00:26:38,560 Speaker 2: somebody uses open AI's models, that model is being run 449 00:26:38,600 --> 00:26:43,080 Speaker 2: on a zero cloud, generating revenue for Microsoft. This is 450 00:26:43,200 --> 00:26:45,760 Speaker 2: the rot economy in action, by the way, Big tech 451 00:26:45,800 --> 00:26:49,879 Speaker 2: has funded its biggest customers for their next growth revenue stream, 452 00:26:49,960 --> 00:26:53,680 Speaker 2: justifying this massive expansion of their data center operations because 453 00:26:53,760 --> 00:26:57,800 Speaker 2: ai is the future and they're telegraphing growth to these 454 00:26:57,840 --> 00:27:00,280 Speaker 2: brainless drones in the market who will buy any thing, 455 00:27:00,280 --> 00:27:04,280 Speaker 2: who never think too hard about what they're actually investing in. Hey, 456 00:27:04,280 --> 00:27:08,199 Speaker 2: i's this big, sexy, exciting and theoretically powerful way to 457 00:27:08,280 --> 00:27:13,479 Speaker 2: centralized labor. And it's innovative sounding enough that it allows 458 00:27:13,480 --> 00:27:16,360 Speaker 2: people to dream big about how it might change their lives. Now, 459 00:27:16,359 --> 00:27:21,600 Speaker 2: it might help them not pay real people to do shit. Yeah, 460 00:27:21,600 --> 00:27:24,320 Speaker 2: here's the biggest worry I have. Here's the real pickle, 461 00:27:24,359 --> 00:27:27,840 Speaker 2: here's the thing that keeps me up at night. None 462 00:27:27,840 --> 00:27:30,480 Speaker 2: of these companies seem to have appeared to consider something. 463 00:27:31,800 --> 00:27:34,119 Speaker 2: What if generative AI can't actually do any of the 464 00:27:34,160 --> 00:27:39,040 Speaker 2: things they're excited about. What if genera avii's content, as 465 00:27:39,040 --> 00:27:42,680 Speaker 2: you probably seen from anything chet GBT spits out, isn't 466 00:27:42,760 --> 00:27:48,840 Speaker 2: really good enough. Hey, is anyone checked if anyone's actually 467 00:27:48,920 --> 00:27:53,080 Speaker 2: using these tools, if they're helpful to anyone? Is this 468 00:27:53,160 --> 00:28:00,120 Speaker 2: actually replacing anyone's work? Huh, that's a bit worrying, mate, 469 00:28:00,680 --> 00:28:03,480 Speaker 2: I didn't think about that before. Just kidding, I've been 470 00:28:03,480 --> 00:28:07,679 Speaker 2: thinking about it for months. Look, here's the thing. I 471 00:28:07,800 --> 00:28:10,919 Speaker 2: think that the big problem here is that Sam Altman 472 00:28:11,000 --> 00:28:13,879 Speaker 2: and his cronies have allowed the media, the markets, and 473 00:28:13,920 --> 00:28:16,720 Speaker 2: big tech to fill in the gaps of their specious messaging. 474 00:28:17,240 --> 00:28:21,080 Speaker 2: They've allowed everybody to think that open AI can do 475 00:28:21,240 --> 00:28:27,920 Speaker 2: whatever anyone dreams. Yeah, I don't think that generative AI 476 00:28:28,000 --> 00:28:30,800 Speaker 2: can do much more than it is today. And also, 477 00:28:31,720 --> 00:28:34,560 Speaker 2: from what I've seen, none of these generative AI companies 478 00:28:34,560 --> 00:28:38,280 Speaker 2: actually make a profit, and with each new model they 479 00:28:38,360 --> 00:28:42,640 Speaker 2: become less profitable, and I don't see that changing in 480 00:28:42,680 --> 00:28:45,960 Speaker 2: the future. And so I've dug in a little more 481 00:28:46,560 --> 00:28:51,160 Speaker 2: looking under the hood. All the demand that's spurring Microsoft, Google, 482 00:28:51,200 --> 00:28:55,360 Speaker 2: and Amazon's data center operations might not actually be there. 483 00:28:57,120 --> 00:28:59,480 Speaker 2: My friends, I think we're in the next tech bubble, 484 00:28:59,680 --> 00:29:02,239 Speaker 2: and the next episode, I'm gonna walk you through how 485 00:29:02,280 --> 00:29:14,480 Speaker 2: I think it might pop. Thank you for listening to 486 00:29:14,480 --> 00:29:15,320 Speaker 2: Better Offline. 487 00:29:15,480 --> 00:29:17,880 Speaker 4: The editor and composer of the Better Offline theme song 488 00:29:17,960 --> 00:29:20,640 Speaker 4: is Matasowski. You can check out more of his music 489 00:29:20,640 --> 00:29:24,280 Speaker 4: and audio projects at Matasowski dot com, M A T 490 00:29:24,280 --> 00:29:28,560 Speaker 4: T O, S O W s ki dot com. You 491 00:29:28,560 --> 00:29:31,080 Speaker 4: can email me at easy at Better offline dot com, 492 00:29:31,160 --> 00:29:33,120 Speaker 4: or check out Better Offline dot com to find my 493 00:29:33,160 --> 00:29:35,800 Speaker 4: newsletter and more links to this podcast. Thank you so 494 00:29:35,880 --> 00:29:36,560 Speaker 4: much for listening. 495 00:29:37,720 --> 00:29:40,200 Speaker 2: Better Offline is a production of cool Zone Media. 496 00:29:40,320 --> 00:29:43,160 Speaker 4: For more from cool Zone Media, visit our website cool 497 00:29:43,240 --> 00:29:46,360 Speaker 4: Zonemedia dot com, or check us out on the iHeartRadio 498 00:29:46,360 --> 00:30:07,200 Speaker 3: App, Apple Podcasts, or wherever you get your podcasts.