WEBVTT - Drilling Deep: Karen Hao on How Big AI Is Gambling with the Planet’s Chips 0:00:17.242 --> 0:00:17.722 Pushkin. 0:00:23.922 --> 0:00:27.562 Hello and welcome back to Drilled. I'm Amy Westerbelt. Today 0:00:27.642 --> 0:00:32.122 we're bringing you another installment of our ongoing series Drilling Deep, 0:00:32.642 --> 0:00:35.922 where we speak to authors of recent books that are 0:00:35.962 --> 0:00:39.362 either about climate or about things that intersect with climate 0:00:39.442 --> 0:00:43.162 in a big way. Today is a super timely one. 0:00:43.682 --> 0:00:47.882 We're going to talk about AI. Specifically, we're going to 0:00:47.882 --> 0:00:52.282 talk about Karen Howe's new book Empire of AI, Dreams 0:00:52.322 --> 0:00:57.962 and Nightmares in Sam Altman's Open AI. In this episode, 0:00:58.122 --> 0:01:03.962 Adam Lowenstein interviews journalist Howe, who argues that AI and 0:01:04.002 --> 0:01:09.922 the profit driven infrastructure that surrounds it is a colonial project. 0:01:10.242 --> 0:01:14.082 What openI boss Altman and his fellow ideologues in Silicon 0:01:14.122 --> 0:01:18.441 Valley are pursuing, how says, is not just corporate power, 0:01:18.521 --> 0:01:22.722 but imperial power. It makes sense when you think about 0:01:22.762 --> 0:01:27.842 the fact that AI, at its core, its most basic 0:01:27.961 --> 0:01:32.401 and ongoing purpose is about twenty four hour surveillance to 0:01:32.521 --> 0:01:36.802 keep all us plates in line. These guys are building empires, 0:01:36.842 --> 0:01:40.602 and as history shows, empires are built on resource extraction, 0:01:40.961 --> 0:01:47.202 particularly the old fashioned kind labor, energy, minerals, land, and water. 0:01:48.082 --> 0:01:51.882 It seems like almost overnight, big techts feel good climate 0:01:51.922 --> 0:01:58.002 promises have evaporated. They've been swapped seamlessly for slippery promises 0:01:58.082 --> 0:02:03.602 that so called artificial general intelligence will solve climate, never 0:02:03.722 --> 0:02:07.962 mind that it's a fantastic concept that has no agreed 0:02:08.042 --> 0:02:12.802 upon definition, or more fundamentally, that appears nowhere close to 0:02:12.922 --> 0:02:17.922 actually existing. In big tech's frenzied pursuit the quote unquote 0:02:18.162 --> 0:02:23.642 hyper scale AI dominance that evangelists claim will unlock AGI, 0:02:24.122 --> 0:02:28.562 as well as its expanding alliances with fossil fuel backed 0:02:28.602 --> 0:02:33.802 petro states, fossil fuel companies, and authoritarian political movements, it's 0:02:33.842 --> 0:02:38.482 become an increasingly central contributor to the climate crisis. In 0:02:38.522 --> 0:02:42.402 an October conversation with Drilled, How discussed how Silicon Valley 0:02:42.402 --> 0:02:45.402 giants appear to be following the oil and gas industry's 0:02:45.442 --> 0:02:50.882 playbook of disinformation and deceit, how Altman and Open AI's 0:02:50.882 --> 0:02:55.242 secrecy and disingenuous rhetoric transformed the field of AI research 0:02:55.362 --> 0:02:59.642 into corporate pr and why the destructive trajectory of AI 0:02:59.802 --> 0:03:04.722 scale and commercialization is not inevitable, no matter what its 0:03:04.762 --> 0:03:08.842 power hungry proponents would have You. Believe that conversation is 0:03:08.842 --> 0:03:12.482 coming up after this quick break and find a condensed 0:03:12.802 --> 0:03:15.802 written version of this interview on our website at drill 0:03:15.882 --> 0:03:22.842 dot Media. 0:03:31.002 --> 0:03:32.962 How does it feel having the book out? Because it's 0:03:33.122 --> 0:03:36.562 it's a monumental achievement. I know you've gotten lots of 0:03:36.602 --> 0:03:39.482 praise for them already, but it's just it's just epic 0:03:39.802 --> 0:03:42.402 and I can't imagine the amount of work that went 0:03:42.442 --> 0:03:44.522 into producing it. 0:03:45.402 --> 0:03:48.922 Yeah, I was the most intense and painful thing that 0:03:48.962 --> 0:03:53.162 I've ever worked on me, like emotionally and physically painful 0:03:53.202 --> 0:03:56.522 because I was typing so much that I developed tendonitis 0:03:56.602 --> 0:04:02.202 on my hands. But yeah, I know, it's I mean, 0:04:02.242 --> 0:04:06.802 it's I think it's been interesting. Have you ever written 0:04:06.802 --> 0:04:07.642 a book before? 0:04:08.442 --> 0:04:10.642 It was not as long as Empire of Ai, but 0:04:10.722 --> 0:04:14.322 I have some understanding of the like the long term 0:04:14.402 --> 0:04:16.681 nature of a writing project like this, but not to 0:04:16.722 --> 0:04:19.082 the same extent with the amount of reporting in the 0:04:19.242 --> 0:04:21.402 yashion that you were wrestling with here. 0:04:22.282 --> 0:04:25.882 The thing that I didn't really prepare for was like 0:04:26.162 --> 0:04:30.642 the one to eighty flip from being completely alone and 0:04:30.882 --> 0:04:34.322 isolated working on the book to like suddenly talking with 0:04:34.362 --> 0:04:38.322 like thousands of people about the same thing once the 0:04:38.322 --> 0:04:42.482 book was launched, And that has been both like a 0:04:42.602 --> 0:04:47.042 very fun and new experience, and one that comes with 0:04:47.082 --> 0:04:48.962 like a lot of relief that like people are resonating 0:04:48.962 --> 0:04:53.122 with the thing, and also like a very uh like 0:04:53.202 --> 0:04:57.722 slightly not slightly extremely chaotic transition, because like you go 0:04:57.882 --> 0:05:02.042 from like mode of like total silence and not having 0:05:02.202 --> 0:05:05.722 really feedback from anyone about anything that you're doing, to 0:05:05.802 --> 0:05:10.922 suddenly getting bombarded with everyone's thoughts about what you've been 0:05:11.082 --> 0:05:14.601 working on. So yeah, so it's like it's like a 0:05:14.842 --> 0:05:17.322 it's like a strange mix of relief and and like 0:05:17.442 --> 0:05:20.002 new sources of stress since the price. 0:05:21.122 --> 0:05:25.802 Have you heard from either sources or industry players with 0:05:25.882 --> 0:05:27.042 feedback about the book. 0:05:28.082 --> 0:05:32.762 I've heard from a few sources who really loved the book, 0:05:32.762 --> 0:05:34.362 which is always a relief. M M. 0:05:34.842 --> 0:05:37.362 Yeah, you want to do them justice for sharing their stories. 0:05:37.602 --> 0:05:41.962 Yeah, And I've I've heard from some like people within 0:05:42.082 --> 0:05:45.802 industry that are not like exacts, but you know, employees 0:05:45.802 --> 0:05:49.082 at these various companies that I've mentioned that were not sources, 0:05:49.122 --> 0:05:53.362 but that have mentioned that they really appreciated the book. 0:05:53.562 --> 0:05:56.442 But in terms of like formal responses from Open AI 0:05:56.922 --> 0:06:01.601 for other companies, there's been nothing, which I much prefer 0:06:02.082 --> 0:06:06.681 to like a lawsuit, So I'll take silence any day. Yeah, 0:06:06.722 --> 0:06:07.162 for sure. 0:06:07.802 --> 0:06:12.282 One of the things I found most troubling eye opening 0:06:12.722 --> 0:06:17.242 the way that you convey how artificial intelligence or the 0:06:17.322 --> 0:06:21.722 concept of it is an ideology in Silicon Valley and 0:06:21.762 --> 0:06:24.802 in these companies. And one of the things that the 0:06:24.842 --> 0:06:28.562 book does such a brilliant job communicating, and that I 0:06:28.602 --> 0:06:32.002 wanted to convey in this conversation is the fact that 0:06:32.402 --> 0:06:35.921 these are not like normal companies, and this is not 0:06:36.042 --> 0:06:39.882 a normal technology, whatever normal means. Yeah, can you talk 0:06:39.922 --> 0:06:44.041 a bit about how AI is an ideology and then 0:06:44.442 --> 0:06:50.162 how these companies are already or are pursuing this status 0:06:50.242 --> 0:06:52.242 as what you describe as empires. 0:06:53.042 --> 0:06:55.642 Yeah, So if you think about just the fact that, 0:06:57.402 --> 0:06:58.962 like if we look at the history of AI and 0:06:59.842 --> 0:07:02.762 the founding of the discipline, the scientific discipline in nineteen 0:07:02.802 --> 0:07:06.202 fifty six, even back then, it was already an ideological 0:07:06.802 --> 0:07:11.242 project because it was a group of scientists deciding they 0:07:11.242 --> 0:07:15.242 wanted to try and replicate human intelligence and computers. And 0:07:15.482 --> 0:07:21.162 that's a very political choice, Like why do that? You know, 0:07:21.642 --> 0:07:24.842 from a it's not really clear what the scientific reason is. 0:07:24.842 --> 0:07:28.802 It is more of a of just like a there 0:07:28.882 --> 0:07:35.122 was an intention behind that choosing that path and setting 0:07:35.202 --> 0:07:40.562 up a field that defined its goal as attempting to 0:07:40.802 --> 0:07:48.002 mimic and therefore ultimately potentially replace humans. And Joseph Weisenbauman, 0:07:48.402 --> 0:07:51.122 MIT professor, who was kind of part of the initial 0:07:51.882 --> 0:07:54.642 co word of scientists who did start researching AI, then 0:07:54.842 --> 0:07:58.682 later quickly became a critic of the entire endeavor, still 0:07:58.762 --> 0:08:01.402 back in the nineteen sixties, and he said at the time, 0:08:01.442 --> 0:08:06.122 in a very prophetic way, there is something inherently disturbing 0:08:06.282 --> 0:08:10.522 about a goal that intends to replace a thing that 0:08:10.522 --> 0:08:15.362 we already have, because ultimately, all this is going to 0:08:15.362 --> 0:08:21.602 do is allow people empower politicians and corporate CEOs to 0:08:21.762 --> 0:08:29.122 impose their will through algorithms and use the AI system 0:08:29.442 --> 0:08:34.322 as the load bearer of responsibility, essentially, like it allows 0:08:34.362 --> 0:08:36.122 them to do whatever they want, but to say that 0:08:36.242 --> 0:08:39.842 something else did it. And so he was already recognizing 0:08:39.881 --> 0:08:43.722 back then that there's an ideological project behind the creation 0:08:43.921 --> 0:08:47.321 of AI, and so fast forward all the way to today, 0:08:47.962 --> 0:08:51.761 that ideology has not only continued to permeate and drive 0:08:52.202 --> 0:08:58.362 the industry and its goals, it has now become much 0:08:58.401 --> 0:09:02.401 more extreme because they're layered on top of this fundamental 0:09:02.401 --> 0:09:05.122 assumption that it is somehow inherently good to recreate human 0:09:05.122 --> 0:09:08.522 intelligence and computers. We now have a particular approach to 0:09:08.641 --> 0:09:13.041 doing that that these companies index on, which is scaling 0:09:13.082 --> 0:09:16.441 at these models at all costs, which is also an ideology. 0:09:16.482 --> 0:09:21.761 There is no scientific basis for choosing that approach, and 0:09:21.801 --> 0:09:25.161 there's no scientific reason for why these companies are trying 0:09:25.202 --> 0:09:30.042 to dominate as monopolies. This is also based on these 0:09:30.681 --> 0:09:34.242 worldviews and values that a certain small group of people 0:09:34.641 --> 0:09:39.241 have chosen, and that group of people have an extraordinary 0:09:39.281 --> 0:09:44.081 amount of power to make that choice affect everyone. I'm 0:09:44.122 --> 0:09:46.482 really glad that you're asking about the ideological aspect of it, 0:09:46.521 --> 0:09:50.201 because I think this is a really huge part of 0:09:50.242 --> 0:09:53.602 the story of AI that has often missed. People keep 0:09:53.641 --> 0:09:56.282 thinking that this is just a business story. It's you know, 0:09:56.362 --> 0:10:01.002 Silicon value doing commercial things and ultimately trying to drive 0:10:01.362 --> 0:10:05.081 for profit maximization. But that is only half the story 0:10:05.082 --> 0:10:09.321 when it comes to AI. There is a deep ideological 0:10:09.401 --> 0:10:14.562 drive to recreate human intelligence too, and a belief in 0:10:14.842 --> 0:10:18.202 the idea that this is somehow going to create to 0:10:18.281 --> 0:10:21.921 solve all of our problems or potentially destroy us, and 0:10:21.962 --> 0:10:23.562 that in order to do all these things that we 0:10:23.681 --> 0:10:28.041 have to consume the entire planet's resources and all of 0:10:28.082 --> 0:10:30.522 those that need to be scrutinized for what they are, 0:10:30.602 --> 0:10:34.362 which is actually just an extremely narrow view about how 0:10:34.401 --> 0:10:36.761 the world is and how it should be. 0:10:38.482 --> 0:10:42.802 The idea of artificial general intelligence or AGI, I think, 0:10:42.921 --> 0:10:46.201 is something people hear about a lot these days because 0:10:46.561 --> 0:10:49.722 obviously it's great for headlines, and it's also great for 0:10:50.641 --> 0:10:53.921 CEOs of these companies to throw out as an idea, 0:10:54.161 --> 0:10:56.682 you know, implying that it's already here or it's just 0:10:56.722 --> 0:10:59.482 around the corner is a great way to boost a 0:10:59.482 --> 0:11:02.362 stock price. But they also, as you show in the book, 0:11:02.401 --> 0:11:05.402 they also use the idea of AGI is basically a 0:11:05.442 --> 0:11:09.201 pretext for whatever they want to do. Can you talk 0:11:09.242 --> 0:11:12.722 a little bit about what AGE is or as close 0:11:12.761 --> 0:11:15.482 as we can come to a definition, and then how 0:11:15.602 --> 0:11:22.602 they use this hypothetical concept to basically justify whatever it 0:11:22.641 --> 0:11:23.322 is they want to do. 0:11:24.041 --> 0:11:29.082 Yeah. So, the loosest definition of AGI is the point 0:11:29.122 --> 0:11:33.002 at which an AI system fully encapsulates all the dimensions 0:11:33.002 --> 0:11:37.402 of human intelligence, which was the original definition of AI. 0:11:37.561 --> 0:11:39.881 And the reason why we now use the term AGI 0:11:39.921 --> 0:11:43.362 instead of AI is because over the decades, the term 0:11:43.362 --> 0:11:45.881 AI has sort of been cheapened by the fact that 0:11:45.921 --> 0:11:49.322 companies keep using it to market existing products and services, 0:11:49.401 --> 0:11:52.202 and so now AI has come to mean what we 0:11:52.281 --> 0:11:57.161 have currently and AGI is supposed to indicate a return 0:11:57.281 --> 0:12:02.802 to that original premise of this scientific discipline. But the 0:12:02.881 --> 0:12:07.962 problem why AGI is so ill defined is because we 0:12:08.002 --> 0:12:12.362 don't have scientific cosensus around what human intelligence is. There's 0:12:12.442 --> 0:12:19.681 no signs, there's no neurological, biological, psychological definition for this 0:12:20.242 --> 0:12:23.041 special characteristic that humans seem to have more of than 0:12:23.082 --> 0:12:27.242 other animal species. And in fact, like when you think 0:12:27.242 --> 0:12:32.962 about the history of trying to measure, quantify rank human intelligence, 0:12:33.122 --> 0:12:37.722 it has always been driven by extremely dark, dark motives eugenics, 0:12:39.202 --> 0:12:42.921 the justification of oppressing different groups of people. And so 0:12:43.362 --> 0:12:47.602 this is why AGI remains an extremely nebulous milestone because 0:12:47.881 --> 0:12:51.682 different people have totally different ideas of what does it 0:12:51.761 --> 0:12:55.961 mean that we've finally achieved human level intelligence and open 0:12:56.002 --> 0:12:58.641 eye jokes about like this has been a long standing 0:12:58.681 --> 0:13:02.041 joke within the company. If you ask thirteen AI researchers 0:13:02.722 --> 0:13:06.242 what AGI is, you'll get fifteen definitions, and so you know, 0:13:06.281 --> 0:13:09.042 they're kind of self aware about the fact that this 0:13:09.242 --> 0:13:14.921 lacks consensus, but it becomes a tool that the executives 0:13:15.002 --> 0:13:17.722 use to their advantage, and Altman in particular in that 0:13:17.881 --> 0:13:21.442 because there is no neither a definition within the company 0:13:21.482 --> 0:13:24.641 that has full consensus nor externals of the company that 0:13:24.641 --> 0:13:29.002 has full consensus, they can just paint these visions of 0:13:29.082 --> 0:13:35.322 AGI based on whatever is necessary to overcome a particular 0:13:35.401 --> 0:13:38.801 hurdle that they're facing in that moment. So when Altman 0:13:38.881 --> 0:13:41.122 is speaking in front of Congress, you'll paint AGI as 0:13:41.161 --> 0:13:45.362 this magical system that is like a magical want that 0:13:45.362 --> 0:13:48.161 you can waive that's going to solve climate change, cure cancer, 0:13:48.322 --> 0:13:50.881 do all these things that people have wanted for a 0:13:50.962 --> 0:13:53.682 very very long time. And then when you're talking to 0:13:53.801 --> 0:13:59.242 like CEOs, AGI suddenly an employee like something that they 0:13:59.242 --> 0:14:01.881 can buy as a service so that they don't have 0:14:01.921 --> 0:14:04.562 to hire real humans. And then when you're talking to 0:14:04.801 --> 0:14:07.921 you know, an average consumer, AGI suddenly a friend or 0:14:08.041 --> 0:14:11.281 it's an assistant or whatever it is that you know 0:14:11.401 --> 0:14:14.321 compels that person to put down money or compels the 0:14:14.921 --> 0:14:20.322 regulators to not put in regulation. And so AGI just 0:14:20.402 --> 0:14:23.362 shape shifts and morphs in these ways that when you 0:14:23.442 --> 0:14:28.042 actually list out all of the possible things that supposedly 0:14:28.082 --> 0:14:31.321 AGI is that Altman or any other company executive has 0:14:31.322 --> 0:14:36.802 ever said it's a completely incoherent vision of a technology, 0:14:37.082 --> 0:14:40.602 like it just doesn't make sense, and like opening eye uses. 0:14:41.522 --> 0:14:46.162 They've also written down in documents completely different definitions of 0:14:46.202 --> 0:14:50.122 this as well, Like they just they officially say on 0:14:50.162 --> 0:14:54.282 their website that AGI is highly autonomous systems that outperform 0:14:54.402 --> 0:14:57.962 humans and most economically valuable work, So they are defining 0:14:58.002 --> 0:15:01.802 it as a labor automating technology. But then in their 0:15:01.842 --> 0:15:05.642 contracts with Microsoft, as reported by the Information, AGI is 0:15:05.642 --> 0:15:07.962 apparently an AI system that can generate one hundred billion 0:15:08.002 --> 0:15:12.681 dollars in revenue, which is like a completely different definition. 0:15:14.322 --> 0:15:18.802 So yeah, it's just it's just a rhetorical tool forgetting 0:15:18.802 --> 0:15:19.321 what they want. 0:15:19.602 --> 0:15:22.202 On a side note, and I think this is probably 0:15:22.242 --> 0:15:23.961 obvious to a lot of people, but I think still 0:15:23.962 --> 0:15:30.082 bears emphasizing is the hunger that corporate CEOs have to 0:15:30.162 --> 0:15:32.802 get rid of their needy employees. And by needy, I 0:15:32.802 --> 0:15:40.042 mean like healthcare benefits, vacation, parental leave, potential union organizing. 0:15:40.442 --> 0:15:44.442 They salivate at the idea of replacing workers with machines, 0:15:44.442 --> 0:15:47.522 and so you can see how the concept of AGI 0:15:47.602 --> 0:15:49.961 would be so appealing to a CEO. 0:15:50.681 --> 0:15:57.482 Absolutely yeah, and this is why this has continued, in 0:15:57.522 --> 0:16:00.682 part to be a self perpetuating myth. I guess you 0:16:00.681 --> 0:16:04.682 could say that it is possible to build AGI is 0:16:04.722 --> 0:16:10.002 because there is such a high demand for that magical 0:16:10.242 --> 0:16:14.922 thing that a rep sense that it is just all 0:16:15.002 --> 0:16:19.042 of the market incentives and all of the political incentives 0:16:20.002 --> 0:16:24.321 push Silicon Valley to continue propping that myth up and 0:16:25.282 --> 0:16:27.522 using it to dominate the narrative. 0:16:28.561 --> 0:16:32.122 Something you mentioned earlier and that you trace throughout the 0:16:32.122 --> 0:16:36.602 book is the way that the promise of AGI solving 0:16:36.642 --> 0:16:42.962 the climate crisis has been used to rationalize really unfathomable 0:16:43.482 --> 0:16:46.362 environmental costs. And you describe it at one point in 0:16:46.402 --> 0:16:48.842 the book as a gamble that are basically saying, like, 0:16:49.282 --> 0:16:54.602 it doesn't really matter resource consumption, the land grabbing, not 0:16:54.681 --> 0:16:58.441 to mention the human exploitation that it requires right now, 0:16:59.042 --> 0:17:03.762 because this magical mystery technology will solve this crisis that 0:17:03.802 --> 0:17:06.522 we're contributing to in a massive way at some point 0:17:06.561 --> 0:17:08.922 in the future. Can you talk a bit about that 0:17:09.042 --> 0:17:12.562 gamble that they are making the planet's chips, even though 0:17:12.922 --> 0:17:15.081 no one actually gave them those chips to play with, 0:17:15.121 --> 0:17:18.042 but they're basically the extent to which they do not 0:17:18.121 --> 0:17:21.042 seem to care about any of the impacts of the 0:17:21.081 --> 0:17:26.321 technologies they're building because of this future promise to solve 0:17:26.482 --> 0:17:30.321 this existential crisis. It's really pretty astonishing. 0:17:31.202 --> 0:17:35.522 It is incredibly astonishing. Yeah, I think the first to 0:17:35.561 --> 0:17:39.401 just quantify what level of environmental impact we're talking about, 0:17:39.482 --> 0:17:42.121 there was this McKenzie report that came out earlier this 0:17:42.202 --> 0:17:46.202 year that did an estimation on how much energy would 0:17:46.242 --> 0:17:50.882 actually be required to sustain the current projections of AI 0:17:50.922 --> 0:17:54.962 industry growth, and we would need to add two to 0:17:55.042 --> 0:17:57.842 six times the amount of energy consumed by California onto 0:17:57.882 --> 0:18:01.282 the global grid in five years just to sustain the 0:18:01.321 --> 0:18:04.042 AA industry and its state of center of demands. That 0:18:04.242 --> 0:18:08.922 is insane. That is six possibly six times the fifth 0:18:09.002 --> 0:18:11.282 largest economy in the world, and there have been, you know, 0:18:11.321 --> 0:18:14.522 other projections that have said that's more energy than all 0:18:14.561 --> 0:18:18.802 of India, the largest energy consumer at this point globally, 0:18:19.561 --> 0:18:21.961 and most of that will come from fossil fuels. So 0:18:22.002 --> 0:18:25.962 we are not only talking about an acceleration of climate change, 0:18:26.121 --> 0:18:29.962 we are also talking about an acceleration of air pollution. 0:18:30.242 --> 0:18:34.642 We are already seeing reports of methane gas turbines, for example, 0:18:34.642 --> 0:18:38.641 being used to power Colossus, the supercomputer that Elon Musk 0:18:38.682 --> 0:18:42.682 built for XAI and for training Rock in Memphis, Tennessee. 0:18:43.081 --> 0:18:47.882 That's pumping thousands of tons of pollutants into working class 0:18:48.402 --> 0:18:52.682 communities that have already had a long history of being 0:18:52.722 --> 0:18:57.442 denied the fundamental right to clean air for decades. And 0:18:57.962 --> 0:19:02.361 we also see the exacerbation of the water the clean 0:19:02.442 --> 0:19:07.321 water crisis. There is a huge crisis of people getting 0:19:07.361 --> 0:19:10.601 access to clean drinking water around the world, in part 0:19:10.682 --> 0:19:15.601 due to the acceleration of climate change, and AI also 0:19:15.682 --> 0:19:19.282 exacerbates that. There was this investigation from Bloomberg that found 0:19:19.321 --> 0:19:21.841 that two thirds of the data center is being built 0:19:21.841 --> 0:19:25.602 for the AI industry are going into places that are 0:19:25.601 --> 0:19:32.242 already scarce on freshwater resources, and data centers require freshwater 0:19:32.282 --> 0:19:35.841 resources to cool in order to make sure that these 0:19:36.361 --> 0:19:40.921 extremely expensive computer chips don't overheat and bust. And so 0:19:41.442 --> 0:19:48.122 you have this multi layered environmental public health crisis kind 0:19:48.121 --> 0:19:53.842 of being pushed to the brank and undermining people's fundamental 0:19:53.882 --> 0:20:01.042 ability to live a decent quality life. And that's happening now. 0:20:01.242 --> 0:20:06.282 This is not speculative, like this is literally playing out currently, 0:20:07.361 --> 0:20:11.361 and the AA industry justifies all of this base on 0:20:11.722 --> 0:20:16.441 a speculative assumption that maybe we will get to a 0:20:16.482 --> 0:20:19.802 point where these AI systems can then fix all of 0:20:19.841 --> 0:20:24.201 those problems. And that is the thing that is so 0:20:24.482 --> 0:20:28.522 mind blowing to me, is that in this AI era, 0:20:29.282 --> 0:20:35.802 somehow future speculation can be used to justify present day, 0:20:35.962 --> 0:20:41.682 real world substantial harm. And the question that I always 0:20:41.922 --> 0:20:45.042 ask people is like, how long do we wait for 0:20:45.081 --> 0:20:49.522 this potential speculative future. How much harm do we suffer 0:20:50.002 --> 0:20:53.522 before we realize that future might not arrive and we 0:20:53.642 --> 0:20:57.881 might not survive these harms. And of course, like the 0:20:57.882 --> 0:21:00.242 people who are ultimately making the decisions, they're not the 0:21:00.242 --> 0:21:03.081 ones that they're the brunt of these harms. They're not 0:21:03.121 --> 0:21:05.802 the ones that are bearing the brunt of climate having 0:21:06.002 --> 0:21:10.722 changed already, and they're not the ones breathing in this 0:21:10.841 --> 0:21:15.721 toxic air from fossil fuels burning in their communities. And 0:21:15.762 --> 0:21:18.882 so for THEMB the question of how long can we 0:21:19.121 --> 0:21:23.161 kind of bear to do all these things before we 0:21:23.361 --> 0:21:27.601 potentially reach euphoria is really really long compared to most 0:21:27.642 --> 0:21:29.481 of the rest of the world. 0:21:30.242 --> 0:21:32.881 There's a lot of ways in which the tech industry 0:21:32.922 --> 0:21:36.322 broadly but especially AI, if there is still any separation 0:21:36.402 --> 0:21:38.881 between the two seems to be following the fossil fuel 0:21:38.962 --> 0:21:42.161 industry playbook, and one of those ways seems to be 0:21:43.162 --> 0:21:46.402 this narrative of you know, A, this is inevitable, It's 0:21:46.442 --> 0:21:48.962 happening whether we like it or not, and so we 0:21:49.081 --> 0:21:51.522 might as well might as well be the ones doing it. 0:21:52.081 --> 0:21:54.762 And b this is the you know, the price of progress, 0:21:54.802 --> 0:21:57.682 and if you are against this, then you're against human progress. 0:21:58.202 --> 0:22:02.441 These really powerful narratives of inevitability and progress really struck 0:22:02.442 --> 0:22:05.561 me because it's the exact same rhetoric that the fossil 0:22:05.561 --> 0:22:08.081 fuel industry has been deploying in some form or another 0:22:08.162 --> 0:22:08.922 for decades. 0:22:09.561 --> 0:22:11.482 Yeah. That's such a great point, I mean, and it 0:22:12.002 --> 0:22:16.482 is the trump card for any industry that knows that 0:22:16.522 --> 0:22:23.522 they're engaging in something that's deeply incorrect. And yeah, I mean, 0:22:24.002 --> 0:22:27.522 these technologies are not inevitable. Like the thing that I 0:22:27.561 --> 0:22:31.282 spend a lot of time trying to do with my 0:22:31.361 --> 0:22:36.602 book is document the moments in which open AI employees 0:22:37.042 --> 0:22:42.201 made decisions that fundamentally shape the trajectory of how the 0:22:42.202 --> 0:22:46.201 technology was introduced into society based purely on you know, 0:22:46.361 --> 0:22:49.802 whims of judgment, Like technology is always a product of 0:22:49.841 --> 0:22:54.802 human choices and sometimes those choices are actually seem really 0:22:54.882 --> 0:22:59.122 minor in the moment. Yeah, So, like, how can it 0:22:59.162 --> 0:23:02.802 be inevitable when the shape of a technology can be 0:23:03.002 --> 0:23:05.561 swayed one way or the other based on just like 0:23:05.682 --> 0:23:09.361 a person sitting in a room deciding on the fly 0:23:09.882 --> 0:23:13.162 to do one thing or another and the other thing 0:23:13.202 --> 0:23:16.361 with progress. So yeah, this is like the thing that 0:23:16.442 --> 0:23:20.401 I hope we can have a more nuanced conversation about 0:23:20.482 --> 0:23:25.522 moving forward. Is there's always this idea that the tech 0:23:25.561 --> 0:23:31.121 industry perpetuates that AI is there's only one form of AI, 0:23:31.601 --> 0:23:34.481 and so if we want progress from AI, we just 0:23:34.561 --> 0:23:38.522 have to accept the costs of this progress, meaning the 0:23:38.561 --> 0:23:41.722 colossal environmental impact, the colossal public health impacts, and so 0:23:41.802 --> 0:23:45.242 on and so forth. AI is actually a collection of 0:23:45.361 --> 0:23:49.282 many different technologies that work in very different ways, that 0:23:49.321 --> 0:23:53.242 have very different costs benefit trade offs. And I often 0:23:53.282 --> 0:23:56.522 make the analogy that it's like transportation transportation. There's many 0:23:56.561 --> 0:23:59.442 different types of transportation, and when we talk about how 0:23:59.482 --> 0:24:03.601 we need to improve our transportation options as part of 0:24:04.121 --> 0:24:07.962 client resiliency and fighting climate change, we're not saying like 0:24:08.042 --> 0:24:11.522 we just need more transportation. We're saying we need more bikes, 0:24:11.561 --> 0:24:15.881 we need more public transit. We need less gas guzzling trucks. 0:24:15.882 --> 0:24:18.081 We need to electrify cars. There's like a much more 0:24:18.162 --> 0:24:20.921 nuanced conversation about like which mode of transportation are we 0:24:20.962 --> 0:24:25.802 talking about, and how to tweak and tune and adjust 0:24:25.922 --> 0:24:28.442 the supply and demand of each of these different types 0:24:28.482 --> 0:24:32.601 of transportation in order to get that progress. And we 0:24:32.682 --> 0:24:36.081 need to have a similar nuanced conversation about AI. The 0:24:36.321 --> 0:24:39.522 large scale AI models that Silicon Valley has imposed on 0:24:39.601 --> 0:24:44.561 everyone that I describe as an imperial consolidation project. That 0:24:44.841 --> 0:24:47.682 is the form of AI that is the most costly 0:24:48.361 --> 0:24:53.841 and has very unclear benefits that just do not justify 0:24:54.402 --> 0:24:58.442 those costs. But when talking about the kinds of progress 0:24:58.442 --> 0:25:01.841 that we would actually want in general, putting technology aside, 0:25:01.882 --> 0:25:05.482 like things like wanting client change to be resolved, wanting 0:25:05.841 --> 0:25:10.442 better drugs, better healthcare for treating different types of diseases 0:25:10.482 --> 0:25:13.282 like these are there are actually types of progress that 0:25:13.361 --> 0:25:16.841 we already know how to build AI systems for that 0:25:16.882 --> 0:25:21.282 look fundamentally different from these large SCALEI systems. Those types 0:25:21.321 --> 0:25:25.321 of AI systems they're task specific, they're often very very 0:25:25.361 --> 0:25:29.202 small and cost extremely small amounts of energy to develop. 0:25:29.682 --> 0:25:32.641 They're trained on highly curated data sets that do not 0:25:32.762 --> 0:25:36.961 require the type of labor exploitation and content moderation that 0:25:37.002 --> 0:25:42.202 goes into something like Tatgybut and they have already documented. 0:25:42.922 --> 0:25:46.202 We know that when we build these systems, we get 0:25:46.442 --> 0:25:49.601 benefits on the end, rather than this speculative maybe in 0:25:49.642 --> 0:25:52.402 the future will reach AGI one day and it will 0:25:52.442 --> 0:25:57.121 solve everything. And so once you kind of realize that, 0:25:57.162 --> 0:26:00.282 once you realize there's actually a portfolio of different options 0:26:00.882 --> 0:26:04.722 for the shape of AI and what it can do, 0:26:05.361 --> 0:26:09.161 it suddenly becomes like blatantly obvious that we're actually in 0:26:09.202 --> 0:26:13.482 this colossal capital misallocation problem where we're putting all of 0:26:13.482 --> 0:26:17.802 our capital in the one type of AI that has 0:26:17.841 --> 0:26:21.641 the absolute worst trade offs, and we should be rapidly 0:26:22.002 --> 0:26:24.922 urgently shifting that capital to all of the other types 0:26:24.962 --> 0:26:28.002 of AI that have the correct trade offs very little 0:26:28.002 --> 0:26:29.601 cost for maximal benefit. 0:26:30.682 --> 0:26:34.081 The question of scale, or the idea or the obsession 0:26:34.561 --> 0:26:37.841 with scale, I guess you describe as a doctrine in 0:26:37.922 --> 0:26:40.722 Silicon Valley and in these companies that the only way 0:26:40.722 --> 0:26:43.441 to do this, as you're just referencing, is to scale 0:26:43.482 --> 0:26:46.681 as big and as fast as possible, And something I 0:26:46.682 --> 0:26:49.202 see a lot in headlines these days is in press 0:26:49.242 --> 0:26:53.681 releases is talk of hyper scaling and mega campuses and 0:26:53.722 --> 0:26:56.282 all of this stuff, which is obviously touted is a 0:26:56.321 --> 0:26:58.882 good thing. When you see hyper scale and mega campus 0:26:58.962 --> 0:27:01.202 and where bigger is better, so we must be doing 0:27:01.242 --> 0:27:07.361 something revolutionary and profitable. But it seems like the biggest 0:27:07.561 --> 0:27:10.841 certainly from an environmental perspective, the biggest costs come from 0:27:11.162 --> 0:27:11.722 the scale. 0:27:11.841 --> 0:27:12.081