WEBVTT 1 00:00:09.400 --> 00:00:12.359 All right, welcome back everyone. I'm here with Ryan. We're 2 00:00:12.359 --> 00:00:15.160 going to be discussing AI once again. We actually discussed 3 00:00:15.160 --> 00:00:18.879 this in our pilot episode, which I believe was in November, 4 00:00:19.600 --> 00:00:22.920 And obviously this is a developing story all the time. 5 00:00:22.960 --> 00:00:25.839 There's a lot to talk about. But I guess, first off, Ryan, 6 00:00:27.000 --> 00:00:29.559 looking back on what we discussed in November, what you 7 00:00:29.600 --> 00:00:31.960 remember of it, how do you think that's worked out 8 00:00:31.960 --> 00:00:33.840 as far as where we said this was going what 9 00:00:33.960 --> 00:00:34.520 to expect. 10 00:00:34.719 --> 00:00:39.119 I think we've been vindicated so far. The hype cycle continues, 11 00:00:39.759 --> 00:00:45.920 the same faces at the helm. You know, the original 12 00:00:46.359 --> 00:00:50.399 open AI group still quarreling. Elon Musk is going at 13 00:00:50.439 --> 00:00:55.039 it with Sam Altman on on X and you know 14 00:00:55.520 --> 00:00:59.759 open ayes under severe pressure. But I guess we'll get 15 00:00:59.759 --> 00:01:02.920 into the numbers of it as far as we can 16 00:01:03.119 --> 00:01:07.079 work out, because there's a huge amount of obviousigation in 17 00:01:07.120 --> 00:01:13.079 this market. But the build out is absolutely remarkable. The 18 00:01:13.159 --> 00:01:17.879 numbers that are being talked about, the you know, the 19 00:01:18.319 --> 00:01:22.959 demands that is being calculated based on labor displacement. It's 20 00:01:23.000 --> 00:01:29.040 being pitched as a battle for global supremacy between America 21 00:01:29.040 --> 00:01:34.799 and China particularly, and you know, a new economic paradigm, 22 00:01:35.040 --> 00:01:38.719 which Musk at least is saying we'll be work free 23 00:01:38.799 --> 00:01:44.560 and abundant. So yeah, that's uh, let's get into it. 24 00:01:44.599 --> 00:01:47.040 I guess yeah. The numbers are staggering. We went over 25 00:01:47.120 --> 00:01:48.719 some of them the last time we spoke. I mean, 26 00:01:48.920 --> 00:01:52.280 six hundred and fifty billion is the numbers. Six hundred 27 00:01:52.280 --> 00:01:55.840 and thirty billion direoboats of capex that's just going to 28 00:01:55.879 --> 00:02:01.640 be spent by four companies on AI this year, Amazon, Microsoft, Alphabet, 29 00:02:01.760 --> 00:02:04.680 Meta and obviously these numbers, you know, they just keep 30 00:02:04.799 --> 00:02:08.000 blowing up during unbelievable territory. But there's no sign of 31 00:02:08.000 --> 00:02:10.759 it's slown down. And there's definitely more of a sense 32 00:02:10.840 --> 00:02:13.800 now of people recognizing that this is some kind of bubble. 33 00:02:14.199 --> 00:02:17.719 But I guess the question is how does that come 34 00:02:17.759 --> 00:02:18.919 to an end? Because I think a lot a lot 35 00:02:18.960 --> 00:02:22.199 of people have a sense, okay, it's a bubble to degree, 36 00:02:22.240 --> 00:02:24.840 but there is a worthwhile product there, something will come 37 00:02:24.879 --> 00:02:27.639 of it, some kind of productivity gains. We just don't 38 00:02:27.680 --> 00:02:30.439 know how that will manifest or where the money needs 39 00:02:30.479 --> 00:02:34.560 to go to benefit from that. But also last time 40 00:02:34.639 --> 00:02:38.319 we discussed this, we talked about China, which sometimes doesn't 41 00:02:38.319 --> 00:02:41.360 get talked about so much in these discussions, and you 42 00:02:41.400 --> 00:02:45.159 were kind of bullish on Chinese manufacturing and China's potential 43 00:02:45.199 --> 00:02:48.360 to catch up, and that's suddenly something that's coming to 44 00:02:48.479 --> 00:02:50.840 the fore neat. There was a couple of models released 45 00:02:50.879 --> 00:02:54.319 in the last couple of weeks from China that again 46 00:02:54.840 --> 00:02:58.360 has caused alarm in Silicon Valley because they're so close 47 00:02:58.599 --> 00:03:03.479 to the top models being it out by entropic open Ai, 48 00:03:05.120 --> 00:03:08.039 and the head of strategy for open Ai actually had 49 00:03:08.199 --> 00:03:10.639 an interesting post I think he made on x where 50 00:03:10.680 --> 00:03:14.240 he warned that these open source models being put out 51 00:03:14.240 --> 00:03:16.759 by the Chinese that this was going to lead to 52 00:03:17.000 --> 00:03:20.639 an era of AI communism, I think he called it, 53 00:03:21.560 --> 00:03:24.960 And he was basically warning people of this terrible potential 54 00:03:25.039 --> 00:03:29.080 future where you can just access AI models for free. 55 00:03:29.120 --> 00:03:31.919 I'm not sure why he thought that would be such 56 00:03:31.919 --> 00:03:34.240 a terrifying prospect of people, that you would be able 57 00:03:34.240 --> 00:03:38.240 to access these models for free, rather than the brilliant, 58 00:03:38.360 --> 00:03:42.800 liberal capitalist future of paying a subscription to scam Altmann 59 00:03:42.840 --> 00:03:46.240 to access the same tech. But apparently this to him 60 00:03:46.360 --> 00:03:51.599 is a dangerous specter of AI communism. Interestingly enough, I 61 00:03:51.599 --> 00:03:55.240 think David Sachs, who's the AI advisor to the US 62 00:03:55.319 --> 00:03:58.960 government actually pushed back on that and warned about regulatory captures. 63 00:03:58.960 --> 00:04:02.199 So there's there's some interesting dynamics happening there as far 64 00:04:02.240 --> 00:04:04.800 as how these tech companies now are actually going to 65 00:04:04.840 --> 00:04:07.840 have to deal with this tread that you warn't a 66 00:04:07.840 --> 00:04:10.879 bode of China coming in and eating their lunch. And 67 00:04:10.879 --> 00:04:12.520 it's suddenly seemed to look more realistic. 68 00:04:13.759 --> 00:04:18.000 Yeah, I mean, let's go back to the beginning of AI. 69 00:04:18.560 --> 00:04:23.439 Twenty fifteen, Open AI is born. It's a collaboration between 70 00:04:23.720 --> 00:04:27.160 likes of Vilo mass Musks, Sam Altman, a bunch of 71 00:04:27.160 --> 00:04:35.600 other tech nerds, mathematicians, whatever. Twenty nineteen, that project, which 72 00:04:35.680 --> 00:04:39.240 was meant to benefit all humanity, open source AI for all, 73 00:04:39.879 --> 00:04:45.040 becomes a private company because they realized that to scale 74 00:04:45.120 --> 00:04:49.120 this up into something brilliant, something resembling a brain, is 75 00:04:49.160 --> 00:04:53.399 going to require a lot of computing power. And then 76 00:04:53.720 --> 00:04:58.839 along comes in Nvidia, who historically were just a company 77 00:04:58.839 --> 00:05:01.639 that provided graphics for gamers like me when I was 78 00:05:01.680 --> 00:05:07.680 a teenager, and then became a bitcoin mining company or 79 00:05:07.879 --> 00:05:14.759 you know, bitcoin mining compute provider. You know, the the 80 00:05:14.920 --> 00:05:18.639 tech optimists of AI decided that they could do brilliant 81 00:05:18.720 --> 00:05:23.879 things with GPUs, and that's you know, that's when the 82 00:05:24.000 --> 00:05:28.199 kind of American rent seeking model was born that they 83 00:05:28.199 --> 00:05:30.959 were going to, you know, create all these data centers, 84 00:05:31.040 --> 00:05:34.639 create these huge clusters of compute, and then they were 85 00:05:34.639 --> 00:05:40.000 going to sell it back to consumers enterprises after they've 86 00:05:40.120 --> 00:05:44.560 scraped all of Western civilization by training their models on 87 00:05:45.319 --> 00:05:50.639 Reddit and you know, every document, every pdf on the Internet. 88 00:05:51.519 --> 00:05:54.439 Because I read it was the pinnacle of civilization. 89 00:05:54.759 --> 00:05:59.160 I mean, to be fair to them, they have or 90 00:05:59.399 --> 00:06:04.560 here's an exact sample, anthropic. They were actually scared of 91 00:06:05.920 --> 00:06:12.720 copyright infringement for all the sources they effectively stole, so 92 00:06:12.800 --> 00:06:17.639 they used their seed capital to buy millions of physical books, 93 00:06:18.160 --> 00:06:22.319 tore them apart by removing the bindings and copy in 94 00:06:22.360 --> 00:06:27.439 every page. So they have a phenomenal database of real 95 00:06:27.480 --> 00:06:33.319 world books. But I think AI is bundled together in 96 00:06:33.399 --> 00:06:37.360 this kind of very vague tech optimism, and I think 97 00:06:37.399 --> 00:06:40.319 we should separate the kind of processes that are actually 98 00:06:40.399 --> 00:06:44.439 going on here. So you've got the LLM, which is 99 00:06:45.199 --> 00:06:51.720 basically predicting the next word, and it happens to be 100 00:06:51.759 --> 00:06:56.680 great for coding. So you've got these two frontier labs, 101 00:06:56.839 --> 00:07:00.560 Open AI and the spinoff of open Ai because Oltman 102 00:07:00.639 --> 00:07:05.319 and Amedi fell out. Anthropic and Athropic is the market 103 00:07:05.399 --> 00:07:11.399 leader encoding. You know, they you know, have basically taken 104 00:07:11.399 --> 00:07:14.839 the software development well by storm. I think pretty much 105 00:07:14.879 --> 00:07:17.279 everyone's using them. 106 00:07:18.040 --> 00:07:21.439 I should point out, actually, since we last spoke, I 107 00:07:21.560 --> 00:07:24.319 was banned by Claude. I really, I don't know if 108 00:07:24.319 --> 00:07:27.319 I told you this, I guess not yet. So I 109 00:07:27.319 --> 00:07:29.759 I'd only used Claude a couple of times, wasn't really 110 00:07:29.839 --> 00:07:32.480 feeling chat GPT, but so I decided to test out 111 00:07:32.560 --> 00:07:34.319 the you know, every want to talk about who great 112 00:07:34.319 --> 00:07:36.319 Claude is compared to chat GUPT. So I bought the 113 00:07:36.600 --> 00:07:41.279 premium subscription. Immediately banned, and I got an email from 114 00:07:42.079 --> 00:07:47.720 Claude saying that their algorithm had I forget how they 115 00:07:47.800 --> 00:07:49.639 worded it. It was signals. That was it. That there 116 00:07:49.639 --> 00:07:53.759 were It's algorithm had generated signals associated with me as 117 00:07:53.759 --> 00:07:56.920 a user. And on the basis of this, decided that 118 00:07:57.519 --> 00:08:00.240 I may be in violation of their term or I 119 00:08:00.240 --> 00:08:04.040 shouldn't be using their product, and banned me. And then 120 00:08:04.160 --> 00:08:06.639 you know, centered to review human reviewed it and yes 121 00:08:06.720 --> 00:08:08.720 I'm banned. So that was that was the extent of 122 00:08:08.759 --> 00:08:11.920 the explanation. I got as there are signals associated with 123 00:08:11.959 --> 00:08:14.279 you as a user that means you can't use Claude AI. 124 00:08:14.480 --> 00:08:17.600 So I haven't. I haven't been able to experience and 125 00:08:17.759 --> 00:08:21.079 of any of the Entropic models. But this is there's 126 00:08:21.120 --> 00:08:22.920 your future, you know. I guess I'll be cheering on 127 00:08:23.040 --> 00:08:27.160 for the Chinese open models. I guess we're facing. 128 00:08:26.839 --> 00:08:30.040 You're banned from vibe coding now if that was interest 129 00:08:30.079 --> 00:08:34.559 of yours? So yeah, where to go next? Yeah, Andthropic 130 00:08:34.720 --> 00:08:37.919 is brilliant coding every development. 131 00:08:38.200 --> 00:08:41.240 But and Tropic does seem like the most liberal of 132 00:08:41.279 --> 00:08:44.200 all these companies in terms of most the other tech 133 00:08:44.240 --> 00:08:48.879 companies seem pretty amoral. But wasn't Entropics kind of modus 134 00:08:48.879 --> 00:08:50.759 operando when they started out? It's like, oh, we're going 135 00:08:50.840 --> 00:08:51.919 to be ethical AI. 136 00:08:52.080 --> 00:08:53.960 Wasn't that kind of the Well yeah they said that, 137 00:08:54.039 --> 00:08:55.679 but I just think it's pr no. 138 00:08:55.759 --> 00:08:58.159 But but I'm I think maybe they're their idea of 139 00:08:58.200 --> 00:09:01.360 ethics is I thought Drew's posted something that he was 140 00:09:01.399 --> 00:09:05.000 asking what the different types of head speech laws are 141 00:09:05.000 --> 00:09:07.639 in Europe and a Claude wouldn't tell them, you know, 142 00:09:07.799 --> 00:09:10.000 bands me, so I think, like, you know, their idea 143 00:09:10.039 --> 00:09:14.879 of ethics is you know, even society leftism'm. 144 00:09:14.879 --> 00:09:17.559 I mean, I basically think they're all in it together. 145 00:09:17.679 --> 00:09:24.399 They have a pr mantra, and then in reality they're 146 00:09:24.480 --> 00:09:28.519 ducing their own stock so they can cash out. So 147 00:09:29.000 --> 00:09:33.720 Amadi at Davos earlier in the year, he basically announced 148 00:09:34.159 --> 00:09:36.679 that twenty percent of the workforce was going to be 149 00:09:36.799 --> 00:09:40.799 unemployed in a couple of years. And then he announced 150 00:09:40.879 --> 00:09:44.639 Mithos and said this is so good. Everyone's going to 151 00:09:44.639 --> 00:09:47.600 have their cybersecurity compromise, and this could bring down the 152 00:09:47.600 --> 00:09:51.519 world economy. So he's fallen out with Department of Defense, 153 00:09:52.200 --> 00:09:57.840 like the American government basically hate him. They basically said 154 00:09:57.840 --> 00:10:00.759 that there had to be severe restriction on the use 155 00:10:00.960 --> 00:10:05.200 of myth foss because it was such a risk. And 156 00:10:05.240 --> 00:10:07.639 you've even got the likes of Alex carp A Palenteer 157 00:10:07.759 --> 00:10:10.279 coming out and saying, this guy is a clown. He's 158 00:10:10.480 --> 00:10:13.159 you know, scaring the voters. He's putting us in danger 159 00:10:13.200 --> 00:10:17.559 of you know, electing communists and Luddites and all this 160 00:10:17.679 --> 00:10:19.679 kind of red scare stuff. 161 00:10:19.759 --> 00:10:23.600 So there doesn't seem to be that. 162 00:10:23.639 --> 00:10:28.279 Much consensus between the various oligarchs associated with AI. They're 163 00:10:28.320 --> 00:10:30.399 all kind of feuding with each other and they're trying 164 00:10:30.399 --> 00:10:35.000 to reduce their own stock, and you know, Sam Altman 165 00:10:35.080 --> 00:10:41.399 says that AGI is coming. There are different, so many 166 00:10:41.399 --> 00:10:43.559 different layers to this. And then you've also got the 167 00:10:43.639 --> 00:10:48.639 hyperscalers who own equity in the frontier labs open Ai 168 00:10:49.360 --> 00:10:54.799 and and Thropics, so they're incentivized for these two ipo 169 00:10:54.960 --> 00:10:59.799 at very high evaluations. Meanwhile, there's Nvidia in the backgrounds 170 00:10:59.799 --> 00:11:02.240 there just trying to ship out as much Compute as 171 00:11:02.240 --> 00:11:05.919 possible to you know, all all buyers, which is basically 172 00:11:06.039 --> 00:11:11.320 just the hyperscalers, and to keep everything afloat, they're doing 173 00:11:11.360 --> 00:11:17.279 all these really dodgy circular finance deals, very dodgy accounting 174 00:11:17.759 --> 00:11:22.559 where they're putting things on the revenue statements that really 175 00:11:22.559 --> 00:11:27.440 shouldn't be there, and they're hiding the debt and so 176 00:11:27.600 --> 00:11:31.639 much is obviouscated. But if you look at the raw 177 00:11:31.720 --> 00:11:36.679 numbers and what these companies are saying, they're basically selling 178 00:11:36.720 --> 00:11:39.279 Compute at a loss, and Thropic and open Ai are 179 00:11:39.320 --> 00:11:42.320 selling at a loss. And you know, you've got all 180 00:11:42.320 --> 00:11:46.440 these stories of enterprises token maxing. They're trying to get 181 00:11:46.480 --> 00:11:49.200 productivity out their workers, they end up spending too much, 182 00:11:49.240 --> 00:11:54.000 they cancel it that happened to Uber and they're having 183 00:11:54.039 --> 00:11:57.000 to go through so many rounds of raising new capital. 184 00:11:58.279 --> 00:12:03.279 Open Ai has you know, raised capital from Massi Yoshi 185 00:12:03.399 --> 00:12:06.440 Son to the tune of about sixty five billion dollars. 186 00:12:07.000 --> 00:12:11.039 His last investment was we work lost about sixteen billion 187 00:12:11.080 --> 00:12:13.679 dollars on that, So I wouldn't say he's a very 188 00:12:13.720 --> 00:12:19.639 good investor himself. And now the hyperscale is a juice 189 00:12:19.639 --> 00:12:23.399 in the high whole cycle. There are plans for huge 190 00:12:23.399 --> 00:12:27.120 amounts of data centers all across America, one hundred plus gigawatts. 191 00:12:27.120 --> 00:12:31.799 They've already broken ground on twenty and the kind of 192 00:12:31.799 --> 00:12:34.559 predictions they're making are you know, quite phenomenal. 193 00:12:34.919 --> 00:12:38.519 Yeah, one hundred and fifty gigawatts and just proposed new 194 00:12:38.600 --> 00:12:44.600 data centers across the US. One of the but, you know, 195 00:12:44.639 --> 00:12:48.799 one of the limitations on this is the compute. But 196 00:12:48.919 --> 00:12:51.919 one of the limitations as well is just what is 197 00:12:51.960 --> 00:12:55.240 the actual limit of an LM as far as intelligence? 198 00:12:55.440 --> 00:12:57.200 And this is kind of a lot of what I 199 00:12:57.320 --> 00:13:00.799 discussed when we covered this topic last is the way 200 00:13:00.840 --> 00:13:07.559 people misunderstand intelligence, the embodied nature of intelligence, and not 201 00:13:08.279 --> 00:13:12.080 kind of accepting that LLM is especiically an advanced predictive 202 00:13:12.120 --> 00:13:15.279 text model that can't jump into what is, you know, 203 00:13:15.440 --> 00:13:20.840 genuine learning. I think since then, though, I think there 204 00:13:20.879 --> 00:13:23.759 are people in the field that are kind of putting 205 00:13:23.799 --> 00:13:27.440 this out there of you know, the AGI thing has 206 00:13:27.480 --> 00:13:30.399 been very overhyped, and it often comes with you know, 207 00:13:30.720 --> 00:13:34.679 this can give us great productivity gains, and it can 208 00:13:34.720 --> 00:13:39.399 be very useful in specific fields like coding directed by humans. 209 00:13:40.120 --> 00:13:42.799 But a lot of this kind of sci fi AGI hype, 210 00:13:43.120 --> 00:13:46.799 you know, where Elon Musay in artificial general intelligence was 211 00:13:46.919 --> 00:13:49.639 just around the corner for years, I think you see 212 00:13:49.679 --> 00:13:53.519 some more mature voices kind of damping that. Jan Lakun, 213 00:13:53.600 --> 00:13:57.039 who's the head of AI development at Facebook, says, look, 214 00:13:57.639 --> 00:14:00.039 llms are not going to develop. 215 00:13:59.679 --> 00:14:01.080 A g I. 216 00:14:01.159 --> 00:14:04.919 They are very useful in a certain way, certain limited way, 217 00:14:05.000 --> 00:14:09.240 but we have kind of reached the limits of this paradigm. Like, 218 00:14:09.279 --> 00:14:11.759 if there's going to be another leap in AI, it's 219 00:14:11.799 --> 00:14:13.799 going to be from some kind of model that isn't 220 00:14:13.840 --> 00:14:17.480 just an LLM. And I think that's clear. You know, 221 00:14:17.519 --> 00:14:20.360 we get we're obviously in this race of new AI 222 00:14:20.399 --> 00:14:22.679 models coming out all the time, and some of the 223 00:14:22.799 --> 00:14:25.519 releases have been pretty impressive, but it is kind of 224 00:14:25.840 --> 00:14:28.440 you know, it's a more advanced version of the same thing, 225 00:14:28.519 --> 00:14:30.879 and a lot of the core problems that were there 226 00:14:31.440 --> 00:14:34.399 with lllms are still there. You know, these models still hallucinate. 227 00:14:36.000 --> 00:14:38.720 You know, you still can't rely on it on supervised 228 00:14:38.799 --> 00:14:43.879 for serious work. Right, it really is still basically a 229 00:14:43.960 --> 00:14:46.720 guide because it does make mistakes. It's still hallucinates. They 230 00:14:46.720 --> 00:14:50.200 haven't been able to solve that. There's increasingly a problem 231 00:14:51.000 --> 00:14:56.600 of AI critting this ecosystem of unsourced or badly sourced 232 00:14:56.720 --> 00:15:00.559 and just fake material. Some people have been doc coumenting this. 233 00:15:00.639 --> 00:15:04.759 You know that a YouTuber uses AI, it throws him 234 00:15:04.759 --> 00:15:09.080 out a fake statistic or a fake scientific fact. He 235 00:15:09.159 --> 00:15:11.360 puts it in his YouTube video, the YouTube video gets 236 00:15:11.399 --> 00:15:14.240 a million views. That then becomes a source on this 237 00:15:14.360 --> 00:15:18.919 maybe niche topic that other lllms use people posted on Reddit, 238 00:15:20.039 --> 00:15:22.679 and then it kind of enters into this ecosystem of 239 00:15:22.759 --> 00:15:26.159 knowledge within the AI. Sometimes you have you know, one 240 00:15:26.200 --> 00:15:29.759 AI is sited in Krokipedia. We're not really sure where 241 00:15:29.840 --> 00:15:34.000 Rokipedia is getting the information from that itself as LLM generated. 242 00:15:35.320 --> 00:15:37.840 So like you're running into limits with these lms where 243 00:15:37.840 --> 00:15:41.639 they can't really self correct. It's still very much human 244 00:15:41.720 --> 00:15:45.039 guided and there's been all sorts of hype about all 245 00:15:45.159 --> 00:15:47.879 these are learning in new ways that we hadn't expected, 246 00:15:47.919 --> 00:15:50.240 and that we're moving into this new kind of paradigm 247 00:15:50.240 --> 00:15:54.200 where they learn in ways similar to neural networks and 248 00:15:54.240 --> 00:15:57.960 it's not just the standard token prediction. And as far 249 00:15:58.000 --> 00:15:59.399 as I can tell, this is all nonsense. I mean, 250 00:15:59.440 --> 00:16:01.919 antime I've looked into these, it's it's kind of a 251 00:16:01.919 --> 00:16:03.759 slide of hand trick where it's like, you know, one 252 00:16:03.799 --> 00:16:07.159 extra step between the same kind of token prediction that's 253 00:16:07.159 --> 00:16:09.519 always been happening. So it seems like we're kind of 254 00:16:09.559 --> 00:16:12.960 trapped in this paradigm where they're going to add more 255 00:16:12.960 --> 00:16:15.559 compute and it may look somewhat more impressive, But as 256 00:16:15.559 --> 00:16:18.639 far as breaking into that next step where it gives 257 00:16:18.679 --> 00:16:21.639 you something like, you know, scientific discoveries, which we still 258 00:16:21.679 --> 00:16:24.240 don't get from these lllms, you know, again they're helpful 259 00:16:24.919 --> 00:16:28.720 their replace jobs and work when you kind of know 260 00:16:28.759 --> 00:16:31.039 what the parameters are and it can do it using 261 00:16:31.120 --> 00:16:33.600 existing material that's out there and put it together. But 262 00:16:33.679 --> 00:16:36.200 as far as any kind of intelligence breakthroughs, you know 263 00:16:36.240 --> 00:16:39.840 that they're going to be providing incredible medical discoveries or 264 00:16:39.919 --> 00:16:43.639 things that are really just going to you know, advanced capitalism, 265 00:16:43.799 --> 00:16:45.960 like like Musk says, that make us all post scarcity. 266 00:16:46.039 --> 00:16:48.919 I mean, it's just not there in this LLM paradigm. 267 00:16:49.120 --> 00:16:52.840 And you know, someone like Jean Lacun acknowledges that, and 268 00:16:52.879 --> 00:16:54.919 he kind of says it with the suggestion that something 269 00:16:54.960 --> 00:16:56.440 that's going to come after it, that's going to be 270 00:16:56.440 --> 00:16:59.840 the real deal. But you know, I'm not sure. I'm 271 00:16:59.879 --> 00:17:02.759 not sure that there's just this bridge. Everyone is assuming 272 00:17:02.840 --> 00:17:06.039 that and them gets so complicated and then we get 273 00:17:06.079 --> 00:17:09.880 this kind of self sufficient component in general intelligence. 274 00:17:10.519 --> 00:17:13.559 Well, I can't even get my Gemini to make a 275 00:17:13.599 --> 00:17:16.640 basic meme and it hallucinates all the time, So I 276 00:17:16.680 --> 00:17:21.440 have to be quite careful about the sourcing. But let's 277 00:17:21.480 --> 00:17:25.440 talk about another of the big powerbrokers of Ai Demissabis 278 00:17:25.799 --> 00:17:28.160 of Google Deep Mind. 279 00:17:29.759 --> 00:17:30.400 He was. 280 00:17:31.839 --> 00:17:34.519 Talking about AGI as. 281 00:17:34.400 --> 00:17:39.119 You know, this brilliant you know, something that's equal to 282 00:17:39.200 --> 00:17:43.000 human intelligence. That's now the new definition. The previous cope 283 00:17:43.240 --> 00:17:45.960 was there's going to be recursive learning, it's going to 284 00:17:45.960 --> 00:17:49.440 become super intelligent and be able to solve anything. That's 285 00:17:49.480 --> 00:17:54.279 now been dialed back down to oh AGI means human 286 00:17:54.400 --> 00:17:59.400 level intelligence across a number of domains, you know exactly 287 00:17:59.480 --> 00:18:04.960 as a intelligent person, I might correct you a little 288 00:18:04.960 --> 00:18:08.839 bit there. Demester Sarvice with Google Deep Mind did solve 289 00:18:08.920 --> 00:18:13.039 the protein folding problem and he won the. 290 00:18:12.880 --> 00:18:14.519 Nobel Prize for Chemistry for that. 291 00:18:15.279 --> 00:18:18.559 But the way that works different to an LM is 292 00:18:18.680 --> 00:18:21.599 they use the GPU compute to do what they call 293 00:18:21.680 --> 00:18:26.920 reinforcement learning, where they basically give it like a biological problem, 294 00:18:27.160 --> 00:18:32.519 chemical problem, run like, you know, all these huge amounts 295 00:18:32.559 --> 00:18:36.920 of computations, and then they gradually work out how the 296 00:18:37.599 --> 00:18:41.119 you know, natural biology works and. 297 00:18:41.119 --> 00:18:43.759 They were able to discover right, But that's the point. 298 00:18:43.799 --> 00:18:48.559 Did the did the LLLM intelligently make a scientific discovery 299 00:18:49.680 --> 00:18:53.000 intuite and how to you know, synthesize these different fields 300 00:18:53.039 --> 00:18:58.119 and theories, or did someone that had this proposition using 301 00:18:58.200 --> 00:19:00.119 to process a ton of data and then. 302 00:19:01.079 --> 00:19:04.039 Yeah, they knew where there were unknowns that were hard 303 00:19:04.079 --> 00:19:07.039 to record, and then they were able to use the 304 00:19:07.079 --> 00:19:12.400 compute to solve the unknowns which were already known byrilogists. 305 00:19:14.480 --> 00:19:19.480 But Hasabis is now talking about the use cases being 306 00:19:19.559 --> 00:19:25.200 things like automated labs and we're going to use AI 307 00:19:25.440 --> 00:19:31.519 to you know, solve scientific problems. So overall, the projections 308 00:19:31.559 --> 00:19:36.759 have been downgraded significantly, and there have been signals among 309 00:19:36.839 --> 00:19:40.039 these people to you know, stop telling the public they're 310 00:19:40.039 --> 00:19:45.200 going to be unemployed and things like this. And you know, 311 00:19:45.279 --> 00:19:48.880 the AI build out doesn't get that much attention in 312 00:19:48.960 --> 00:19:49.640 the mainstream. 313 00:19:49.839 --> 00:19:51.279 It's kind of you. 314 00:19:51.200 --> 00:19:55.640 Know, talked about on finance podcasts and you know, tech 315 00:19:55.720 --> 00:20:01.400 nerd podcasts, but it doesn't receive much attention and you know, 316 00:20:01.799 --> 00:20:09.200 in the news or politically. But yeah, Hasabis is probably 317 00:20:09.240 --> 00:20:10.400 the most trustworthy figure. 318 00:20:10.519 --> 00:20:13.880 I would say, Well, it's interesting how there is this 319 00:20:14.000 --> 00:20:21.000 kind of leftist, populist sentiment of anti AI prolog erics. 320 00:20:21.359 --> 00:20:24.039 You know, those kinds of people are especially disliked, the 321 00:20:24.640 --> 00:20:28.880 ailig erics. But also this like anti data center trend, 322 00:20:29.279 --> 00:20:31.440 which is kind of it's kind of cross left and 323 00:20:31.519 --> 00:20:34.680 right wing. You know, like you've got this guy James 324 00:20:34.680 --> 00:20:39.599 Fishback who's running in Florida as a Republican and he's 325 00:20:39.680 --> 00:20:42.839 kind of griperus candidate, but but he's he's he's doing 326 00:20:42.880 --> 00:20:48.599 this populous like I'm going to ban data centers in Florida. Yeah, 327 00:20:48.599 --> 00:20:51.000 it's it's kind of it's kind of like a slopuloust 328 00:20:52.119 --> 00:20:56.359 slopulous meme like the it's kind of like the you know, 329 00:20:56.480 --> 00:20:59.160 the the most populist elements of the left and right 330 00:20:59.200 --> 00:21:02.240 are both doing this out a center thing. And sometimes 331 00:21:02.039 --> 00:21:05.759 it's not really clear why the general public poses these 332 00:21:05.839 --> 00:21:07.759 data centers. A lot of times, I mean, I think 333 00:21:07.759 --> 00:21:10.039 they have a sense that makes things more expensive and 334 00:21:10.319 --> 00:21:12.720 maybe it's going to be damaging for the infrastructure around. 335 00:21:13.319 --> 00:21:17.480 But there hasn't really been a well articulated political case. 336 00:21:17.480 --> 00:21:20.240 I mean, it's basically just nimbism, isn't it. Like, let's 337 00:21:20.559 --> 00:21:23.960 banner in our state and you know, still have the 338 00:21:24.000 --> 00:21:26.920 gains of AI productivity, but is built it somewhere else? 339 00:21:27.359 --> 00:21:31.960 I think there's the cost concern of increasing electricity prices. 340 00:21:32.079 --> 00:21:35.880 There's the environmental concern, you know, things like you know, 341 00:21:35.920 --> 00:21:41.440 the ugliness of a data center, the water usage, et cetera. 342 00:21:43.799 --> 00:21:44.720 But yeah, there are. 343 00:21:44.799 --> 00:21:49.240 Thousands of these things planned all across America. There's been 344 00:21:49.279 --> 00:21:52.599 a lot of resistance to it. These tech bros are 345 00:21:52.640 --> 00:21:56.200 even suggesting that the Chinese might be involved in leading 346 00:21:56.240 --> 00:22:00.640 this kind of movement, and Sam Altman has two attempts 347 00:22:00.680 --> 00:22:04.359 on his life this year. You know, a lot of 348 00:22:04.359 --> 00:22:07.559 people are very skeptical about it, and you know, the 349 00:22:07.599 --> 00:22:09.640 tech bros are blaming the likes of Amma Day for, 350 00:22:10.559 --> 00:22:13.559 you know, saying that they're going to be unemployed, and 351 00:22:13.640 --> 00:22:17.640 there is this big question that if the predictions do 352 00:22:17.759 --> 00:22:20.559 come to pass and it really is as intelligent and 353 00:22:20.559 --> 00:22:25.200 disruptive as they are claiming, and as it needs to 354 00:22:25.240 --> 00:22:29.680 be for all the debt that's being accumulated, then yeah, 355 00:22:29.680 --> 00:22:31.960 people will be unemployed and we will be in a 356 00:22:32.039 --> 00:22:37.680 you know, semi communistic economic paradigm. 357 00:22:38.519 --> 00:22:39.880 I mean, we should, we should talk about it. 358 00:22:39.960 --> 00:22:42.440 I mean, well, one question is what's what's funding all this? 359 00:22:42.720 --> 00:22:46.640 Then you know, you can get into the numbers, but 360 00:22:46.680 --> 00:22:49.039 I mean, I mean, the essential thing is that the 361 00:22:49.319 --> 00:22:52.799 US economy basically is not growth outside of AI companies, 362 00:22:52.960 --> 00:22:55.960 Like that's entirely what's driving everything, and even growth in 363 00:22:56.039 --> 00:22:58.759 other areas of the economy of the economy you can 364 00:22:58.799 --> 00:23:04.079 say is very downstream of the growth generated by the 365 00:23:04.119 --> 00:23:07.920 promise of AI. But it's basically the entirety of what 366 00:23:07.960 --> 00:23:11.519 the US economy is built on right now in serious trouble. 367 00:23:13.039 --> 00:23:16.799 But there's like certain there's certain signs that investors are 368 00:23:16.880 --> 00:23:19.759 getting a little bit concerned about how deep they are 369 00:23:19.799 --> 00:23:23.640 in this. And the AI companies have used all sorts 370 00:23:23.680 --> 00:23:27.400 of interesting account and tricks to kind of hide how 371 00:23:27.440 --> 00:23:30.720 deep they're spending is. But I think currently the number 372 00:23:30.799 --> 00:23:34.279 is that they're they're spending one point four trillion this 373 00:23:34.480 --> 00:23:39.079 year and all sorts of expenditure, and the revenue is 374 00:23:39.359 --> 00:23:43.000 less than half that. So I guess at this point, 375 00:23:43.559 --> 00:23:46.920 if we are looking at a serious correction that is 376 00:23:46.960 --> 00:23:50.839 at some point going to come somehow, who is initially 377 00:23:50.880 --> 00:23:52.599 going to bear the brunt of that? Like, where is 378 00:23:52.599 --> 00:23:55.960 the majority of this funding coming from? 379 00:23:56.000 --> 00:23:59.359 Well, there's different types of funding. So the hyperscale is 380 00:23:59.400 --> 00:24:04.920 the big tech companies. They've basically issued a lot of debt. 381 00:24:05.960 --> 00:24:10.240 Google recently did an equity raise, which implies that they 382 00:24:10.240 --> 00:24:13.839 don't think their stock is particularly good value. They raised 383 00:24:13.839 --> 00:24:19.240 eighty five billion dollars with that, So yeah, they've created 384 00:24:19.279 --> 00:24:23.200 all these bonds equity raises. Other people talking about equity raises. 385 00:24:23.960 --> 00:24:30.279 Open Ai got funding from Massiyoshi Son Andthropic got funding 386 00:24:30.480 --> 00:24:37.680 from Amazon. Microsoft has seed funding in open Ai, and 387 00:24:37.720 --> 00:24:43.440 then Google also has interest in the Frontier Labs. So 388 00:24:44.680 --> 00:24:49.119 that's kind of mostly for purchasing compute. But there are 389 00:24:49.160 --> 00:24:52.640 also these other vehicles which are off balance sheet called 390 00:24:52.920 --> 00:24:57.759 SPV special purpose vehicles, and these are mainly funding the 391 00:24:57.839 --> 00:25:03.400 data center build out, and that usually involves a minority 392 00:25:03.440 --> 00:25:08.079 stake by a hyperscaler around twenty percent, but then they 393 00:25:08.160 --> 00:25:14.160 keep the debt obligations off their balance sheet. And it's 394 00:25:14.200 --> 00:25:19.319 basically private credit that is providing all the additional funding 395 00:25:19.759 --> 00:25:26.640 um so basically American boomers. They're also Japanese banks like 396 00:25:28.279 --> 00:25:33.599 Mitsubishi also funding a lot of this. And the scale 397 00:25:33.799 --> 00:25:37.799 of the buildout is remarkable. They're talking about ten trillion 398 00:25:37.839 --> 00:25:42.960 dollars by twenty thirty, and that implies needing to be 399 00:25:43.359 --> 00:25:46.359 taking two trillion dollars revenue or one and a half 400 00:25:46.440 --> 00:25:51.079 trillion dollar dollars revenue by that time. And it's not 401 00:25:51.279 --> 00:25:55.720