WEBVTT - is using AI worse than driving a car? 0:00:09.960 --> 0:00:12.640 I've lived in LA for a decade, and this whole time, 0:00:12.760 --> 0:00:15.680 I haven't owned a car. When I tell people that, 0:00:15.840 --> 0:00:19.279 they usually look at me weird. And yes, riding the 0:00:19.320 --> 0:00:22.360 bus and walking and using a bike is less convenient, 0:00:22.800 --> 0:00:25.800 but at this point I'm used to it. But sometimes 0:00:26.000 --> 0:00:28.000 I do wonder if I should just give in and 0:00:28.120 --> 0:00:31.720 buy a car like everyone else. So to help me decide, 0:00:31.840 --> 0:00:33.879 I did what a lot of people do recently when 0:00:33.920 --> 0:00:38.680 they're weighing options. I asked AI. I opened up Claude 0:00:38.680 --> 0:00:42.240 dot AI and I input all my current transportation costs. 0:00:42.280 --> 0:00:45.160 I put in my bus fares, my bike, the cost 0:00:45.240 --> 0:00:48.159 of my occasional uber and I asked it to compare 0:00:48.200 --> 0:00:51.280 that to the average costs of car ownership in Los Angeles, 0:00:51.400 --> 0:00:56.240 so parking, gas, insurance, repairs, all that sort of thing. 0:00:56.360 --> 0:00:58.279 And I asked it to give me the pros and 0:00:58.360 --> 0:01:02.320 cons on either side, and it did. The big con 0:01:02.600 --> 0:01:06.119 is convenience, which I already knew. And on the pro side, 0:01:06.200 --> 0:01:08.759 it said that I was saving thousands of dollars per year, 0:01:09.440 --> 0:01:12.120 but it added one extra thing. It said that I 0:01:12.120 --> 0:01:14.680 could have the nice feeling of knowing that I was 0:01:14.720 --> 0:01:18.360 also being eco friendly. And I thought, hold on, wait 0:01:18.400 --> 0:01:22.840 a second, eco friendly, I just spent half an hour 0:01:22.920 --> 0:01:26.240 running scenarios through an LM, which I know is built 0:01:26.280 --> 0:01:29.160 off the back of a massive amount of computing, which 0:01:29.280 --> 0:01:32.600 in turn means a massive amount of energy. So am 0:01:32.600 --> 0:01:35.880 I actually helping the environment here? Or am I hurting it? 0:01:37.360 --> 0:01:40.120 So this week I set out to answer what seems 0:01:40.160 --> 0:01:45.560 like a pretty simple question, how bad is AI's environmental impact? Really? 0:01:46.160 --> 0:01:49.840 And yes, before you ask, I did consider asking Claude 0:01:49.840 --> 0:01:53.720 and maybe chat GBT about AI's own impact on the environment. 0:01:53.960 --> 0:01:56.480 But then I figured, you know what, maybe this is 0:01:56.480 --> 0:02:00.160 a question I should ask actual human beings. And I 0:02:00.160 --> 0:02:02.240 found a couple of people who've been studying this stuff 0:02:02.280 --> 0:02:04.160 for a while to help me parse all of this. 0:02:05.400 --> 0:02:08.120 Is there a way that I can compare? Is my 0:02:08.240 --> 0:02:13.120 AI usage worse or better than car usage, or worse 0:02:13.200 --> 0:02:16.280 or better than my impact on the environment from eating 0:02:16.320 --> 0:02:18.799 meat or something like that? Are we able to make 0:02:18.840 --> 0:02:20.000 those kinds of comparisons. 0:02:20.200 --> 0:02:22.880 That's what carbon footprints were kind of invented for, so 0:02:22.960 --> 0:02:25.720 you can make this type of comparison. If you're driving 0:02:25.720 --> 0:02:27.799 in fossil fuel based car, you know exactly how much 0:02:27.840 --> 0:02:30.320 gas you're using and what that might mean in terms 0:02:30.320 --> 0:02:34.640 of carbon emissions. That's pretty straightforward. It's much harder to 0:02:34.720 --> 0:02:35.640 do this for AI. 0:02:39.400 --> 0:02:48.440 I'm afraid from Kaleidoscope and iHeart podcasts, this is kill switch. 0:02:49.800 --> 0:02:51.360 I'm Jackster Thomas. 0:02:52.040 --> 0:02:52.640 I'm sorry. 0:03:21.400 --> 0:03:24.400 If you use social media, you've probably seen people being 0:03:24.400 --> 0:03:27.760 criticized for using AI, and depending on who you hang 0:03:27.800 --> 0:03:30.720 out with, that criticism can be kind of different in 0:03:30.800 --> 0:03:33.680 how it shows up. When I see someone post an 0:03:33.680 --> 0:03:37.280 AI generated image or some AI generated text and there's 0:03:37.360 --> 0:03:40.480 angry comments in the comment section, it's usually one of 0:03:40.560 --> 0:03:44.720 two types. The first is people saying that it's disrespectful 0:03:45.200 --> 0:03:49.760 that by posting AI generated poetry or drawings you're devaluing 0:03:49.800 --> 0:03:52.680 the original artists who didn't consent to having their work 0:03:52.720 --> 0:03:55.280 fed into an l l M. That one's pretty easy 0:03:55.360 --> 0:03:57.800 to understand, even if you don't agree with it or 0:03:57.880 --> 0:04:00.320 think it's a big deal. The other comment I see 0:04:00.320 --> 0:04:03.280 a lot of is people saying that using AI is 0:04:03.320 --> 0:04:07.160 destroying the environment. Figuring out whether that's a big deal 0:04:07.280 --> 0:04:09.640 or not is a little bit less straightforward. 0:04:09.960 --> 0:04:12.000 What I tried to do in my research, I tried 0:04:12.000 --> 0:04:16.200 to keep track of how the global electricity consumption of 0:04:16.320 --> 0:04:18.320 AI is developing. 0:04:18.080 --> 0:04:20.919 Alex Deviries is the founder of digit Economists and a 0:04:20.960 --> 0:04:24.560 PhD candidate at the VU Amsterdam. He's been researching the 0:04:24.600 --> 0:04:27.359 sustainability of new technologies for about a decade. 0:04:27.640 --> 0:04:29.760 The way I do that is by looking at how 0:04:29.800 --> 0:04:33.640 many machines of specialized AI's devices are being produced by 0:04:33.680 --> 0:04:37.400 the AI hardware supply chain, and then considering that power 0:04:37.440 --> 0:04:41.040 consumption profile how much power is now being consumed by 0:04:41.080 --> 0:04:44.479 all of these devices. Which is a very imperfect way 0:04:44.480 --> 0:04:46.200 of keeping track of this, but it's kind of like 0:04:46.240 --> 0:04:48.680 the only rule you have available at the moment. 0:04:48.520 --> 0:04:50.800 And even Alex is having a hard time keeping up 0:04:50.839 --> 0:04:53.240 with this. What I called him. He was in the 0:04:53.240 --> 0:04:56.400 middle of putting together new research. Back in twenty twenty three. 0:04:56.760 --> 0:04:59.679 His data showed that by twenty twenty seven, new AI 0:04:59.680 --> 0:05:02.360 service sold could use the same amount of energy annually 0:05:02.560 --> 0:05:05.720 as the yearly energy consumption of a country like Argentina 0:05:06.040 --> 0:05:10.799 or the Netherlands. But things have accelerated. His current research 0:05:10.839 --> 0:05:13.479 shows that it won't take until twenty twenty seven for 0:05:13.520 --> 0:05:16.360 that to happen. At this rate, we're going to hit 0:05:16.360 --> 0:05:17.840 that mark sometime this year. 0:05:18.120 --> 0:05:21.719 Simply because now the amount of devices that's being produced 0:05:21.760 --> 0:05:24.600 by the AI hardware supply chain is way higher than 0:05:24.640 --> 0:05:25.960 it was two years ago. 0:05:26.200 --> 0:05:31.760 So it's even exceeding your pretty bleak estimations that you 0:05:31.800 --> 0:05:32.480 made a while ago. 0:05:32.839 --> 0:05:35.320 Oh yeah, it's just that the hype is so big, 0:05:35.320 --> 0:05:37.080 and then the mod for this type of hardware is 0:05:37.120 --> 0:05:40.599 so big that the numbers are going up much faster 0:05:40.680 --> 0:05:42.960 than could be anticipated just two years ago. 0:05:43.080 --> 0:05:45.720 But hold on, before we get too much further, let's 0:05:45.800 --> 0:05:49.200 just clarify what we're even talking about when we say AI. 0:05:50.160 --> 0:05:52.839 If you could break it down for me, how does 0:05:53.880 --> 0:05:57.080 artificial intelligence use natural resources? 0:05:57.400 --> 0:06:01.480 Heah, it's a general umbrella term that includes many different things. 0:06:01.839 --> 0:06:04.680 But right now, if you're talking to a person on 0:06:04.720 --> 0:06:08.159 the streets random like when they say AI, they're referring 0:06:08.200 --> 0:06:11.600 to large languine models, or maybe you meage generation models. 0:06:11.680 --> 0:06:14.320 So these are the generaryty AI models. 0:06:14.400 --> 0:06:17.840 Chale Wren is an associate professor of electrical and computer 0:06:17.960 --> 0:06:21.640 engineering at the University of California, Riverside, and he's kind 0:06:21.640 --> 0:06:25.359 of a colleague of mine. Our fields are completely different. 0:06:25.720 --> 0:06:27.719 But a couple of years ago I taught a class 0:06:27.720 --> 0:06:30.360 in the building right next to his. I'd had no 0:06:30.520 --> 0:06:33.560 idea that on the same campus there was an expert 0:06:33.560 --> 0:06:36.840 who'd been researching the environmental impact of generative AI the 0:06:36.880 --> 0:06:40.640 whole time, and I thought, perfect, this guy's kind of 0:06:40.640 --> 0:06:43.240 a colleague, so I can stop doing all this research 0:06:43.279 --> 0:06:46.640 on my own and just go ask him. Can you 0:06:46.680 --> 0:06:51.640 give me an idea of how, say, car usage compares 0:06:51.760 --> 0:06:54.480 to usage of an AI model. 0:06:54.920 --> 0:06:58.640 I would say having a large language or medium sized 0:06:58.760 --> 0:07:05.000 language model. Right, roughly ten short emails could be consuming 0:07:05.440 --> 0:07:08.640 a quarter of the electric kill what hour energy, So 0:07:08.720 --> 0:07:12.000 that's roughly enough to drive a test in a Model three. 0:07:11.880 --> 0:07:14.960 For one mile, or as Alex puts it, Chat GPT 0:07:15.120 --> 0:07:17.480 must be running on like something like five hundred mega 0:07:17.520 --> 0:07:20.320 what hours a day, which is enough to power a 0:07:20.360 --> 0:07:21.400 small city. 0:07:21.960 --> 0:07:26.360 Basically, chat GPT's overall daily energy use, it's about the 0:07:26.400 --> 0:07:30.000 same as powering every home, every grocery store, every street 0:07:30.040 --> 0:07:33.160 light in a small city like San Luis, Opistpo in 0:07:33.200 --> 0:07:37.160 California or Ithaca in upstate New York. But what does 0:07:37.160 --> 0:07:40.320 that actually mean for me and you? How much energy 0:07:40.360 --> 0:07:43.480 does it take to just ask chat GBT one question 0:07:44.280 --> 0:07:44.600 on a. 0:07:44.560 --> 0:07:48.440 Per interaction basis? It's actually not that much. You're talking 0:07:48.480 --> 0:07:51.800 about something like three one hours maybe per interaction that's 0:07:51.840 --> 0:07:54.080 something like a low lumin led build that you have 0:07:54.160 --> 0:07:56.600 running for one hour. It's not a lot of power, 0:07:56.760 --> 0:08:00.160 but it's it's nevertheless significantly more than a standard Google. 0:08:00.240 --> 0:08:03.160 Search in one second. Just as an aside here, we 0:08:03.280 --> 0:08:05.720 usually don't think of something like a Google Search as 0:08:05.800 --> 0:08:08.800 using electricity. I mean, your phone or your computer is 0:08:08.840 --> 0:08:11.320 already on, so what does it matter if you're typing 0:08:11.320 --> 0:08:13.960 stuff into it or not. But on the other end 0:08:14.000 --> 0:08:17.160 of that Google search you typed in, they're servers and 0:08:17.280 --> 0:08:20.280 those are using energy. So as we keep going in 0:08:20.360 --> 0:08:23.640 this episode, maybe think about that that on your end 0:08:23.920 --> 0:08:27.480 you're not seeing any energy used or environmental effects, but 0:08:27.760 --> 0:08:32.040 doing a Google search, watching a video, or even downloading 0:08:32.080 --> 0:08:35.120 this podcast that does use some amount of energy. 0:08:35.559 --> 0:08:39.479 Even Google's CEO at some point commented, like, hey, interacting 0:08:39.600 --> 0:08:42.960 with these large language models, it takes ten times more 0:08:43.040 --> 0:08:45.640 power than the standard Google search. So and that would 0:08:45.720 --> 0:08:47.880 mean that if you're talking about three one hours for 0:08:48.000 --> 0:08:51.000 interaction in a large language model, for a standard Google search, 0:08:51.040 --> 0:08:53.120 it would be like zero point three one hours, which 0:08:53.160 --> 0:08:55.120 is a very very tiny amount. 0:08:56.360 --> 0:08:58.880 Just to explain here, a what hour is a unit 0:08:58.920 --> 0:09:01.559 that tells you how much energy device uses over time. 0:09:02.120 --> 0:09:04.720 For example, a sixty watt light bulb running for an 0:09:04.760 --> 0:09:09.079 hour uses sixty watt hours. A single Google search uses 0:09:09.120 --> 0:09:12.240 about zero point three watt hours. That's enough to power 0:09:12.280 --> 0:09:16.360 that same light bulb for around eighteen seconds. But now 0:09:16.400 --> 0:09:19.280 there's that AI add on that comes stacked by default 0:09:19.360 --> 0:09:22.200 on top of every Google search, which takes that number 0:09:22.280 --> 0:09:25.600 up ten x, up to three full watt hours per search. 0:09:26.320 --> 0:09:28.960 That's a little different now you're running that same light 0:09:29.000 --> 0:09:32.559 bulb for three full minutes and then but. 0:09:32.640 --> 0:09:35.800 It's of course in the number of interactions where these 0:09:36.120 --> 0:09:38.520 numbers start to stack up quickly, Because if you're talking 0:09:38.559 --> 0:09:42.400 about Google Skill, you're talking about nine billion interactions a day, 0:09:42.679 --> 0:09:46.440 going as three one hours per interaction. Then, interestingly, the 0:09:46.480 --> 0:09:50.040 whole company Google would require as much power as Ireland 0:09:50.120 --> 0:09:51.840 just to serve as a search engine. 0:09:51.880 --> 0:09:55.480 If that was the case, Wow, using as much power 0:09:55.520 --> 0:09:58.959 as a small country sounds wild, but if we think 0:09:59.000 --> 0:10:02.760 about it, it kind of makes sense. We've default ten 0:10:02.880 --> 0:10:07.560 xed our energy use overnight across nine billion searches a day. 0:10:08.080 --> 0:10:10.600 That energy use is going to add up pretty fast. 0:10:11.120 --> 0:10:13.880 But there's another thing to consider when we talk about 0:10:13.920 --> 0:10:17.840 AI's energy use. The difference between training the model or 0:10:17.920 --> 0:10:19.600 giving it a bunch of data to teach it how 0:10:19.600 --> 0:10:22.840 to work, and using it like when you ask it 0:10:22.920 --> 0:10:25.640 to write a cover letter or I ask it if 0:10:25.720 --> 0:10:28.240 I should buy a car. When we talk about AI 0:10:28.440 --> 0:10:30.840 and the energy consumption that can go into AI, there's 0:10:30.840 --> 0:10:34.640 different phases, right. There's the training phase. There's me actually 0:10:34.640 --> 0:10:38.280 sitting down and asking you an agent, a question. Can 0:10:38.320 --> 0:10:39.320 you break that down for me? 0:10:39.600 --> 0:10:43.080 The training part we call it learning, so based on 0:10:43.120 --> 0:10:45.920 the data we try to optimize the parameters so that 0:10:46.200 --> 0:10:49.120 when we see some new queries from the users, we 0:10:49.120 --> 0:10:51.679 can give you as equate an answer as possible. And 0:10:51.920 --> 0:10:54.960 training is really one time. Of course, later we're going 0:10:55.000 --> 0:10:58.520 to do some update fine tuning. Inference is when the 0:10:58.600 --> 0:11:01.880 users actually interact with the model, and depending on the 0:11:01.880 --> 0:11:04.480 popularity of the model, but once it gets trained, it 0:11:04.559 --> 0:11:07.480 could be used by many hundreds of millions of times 0:11:07.880 --> 0:11:10.120 or even billions of times if you train the large 0:11:10.200 --> 0:11:13.440 lunguine model like LAMA three point one. According to the 0:11:13.520 --> 0:11:16.600 data released by Meta Training, a large lan grain model 0:11:16.760 --> 0:11:20.200 like that air pllutant we gamerated through the training will 0:11:20.240 --> 0:11:24.200 be roughly equivalent to more than ten thousand round trips 0:11:24.200 --> 0:11:25.720 by cart LA and New. 0:11:25.679 --> 0:11:29.000 York City ten thousand round trips by car. Yeah, so 0:11:29.040 --> 0:11:34.040 that sounds bad, that sounds like a lot. But is 0:11:34.600 --> 0:11:36.920 that a one time and it's just the one time. 0:11:36.960 --> 0:11:37.640 It's a one time. 0:11:39.000 --> 0:11:42.400 Let's clear something up here. That number ten thousand round 0:11:42.480 --> 0:11:45.200 trips from LA to New York by car. It's not 0:11:45.240 --> 0:11:49.360 just about carbon it's about air pollution, specifically things like 0:11:49.440 --> 0:11:52.840 nitrogen oxides and fine particles that come from power plants 0:11:52.880 --> 0:11:56.160 and can get deep into your lungs. This isn't theoretical. 0:11:56.360 --> 0:12:00.240 This is stuff that raises risks of diseases like cancer, 0:12:00.280 --> 0:12:02.439 and it doesn't just affect people next to the place 0:12:02.440 --> 0:12:07.080 where all those computers are. Pollution travels and it lingers. 0:12:07.720 --> 0:12:11.160 So what Challet's talking about here isn't just numbers. His 0:12:11.240 --> 0:12:14.160 calculations are showing that training a single model the size 0:12:14.160 --> 0:12:17.480 of Metaslama three point one can produce that level of 0:12:17.480 --> 0:12:21.839 pollution on its own. So yes, training these models is 0:12:21.880 --> 0:12:25.000 a one time hit, but it's a big one. If 0:12:25.040 --> 0:12:29.120 we're talking just about energy usage, using an LM to say, 0:12:29.240 --> 0:12:32.480 write ten emails might be like driving an electric vehicle 0:12:32.559 --> 0:12:35.240 for a mile. And since an electric vehicle was maybe 0:12:35.280 --> 0:12:38.120 three times more efficient than a gas vehicle. Figure, those 0:12:38.120 --> 0:12:40.440 ten emails might get you a quarter to a third 0:12:40.440 --> 0:12:43.000 of a mile in a regular car. And yeah, maybe 0:12:43.040 --> 0:12:45.720 these are relevant numbers for me and my decision about 0:12:45.720 --> 0:12:48.960 whether or not my AI usage is counterbalanced by me 0:12:49.120 --> 0:12:52.400 not having a car. But these numbers are just estimates, 0:12:52.480 --> 0:12:54.960 and we are going to get to that. But the 0:12:55.000 --> 0:12:58.280 bigger issue here is that running those data centers doesn't 0:12:58.320 --> 0:13:02.160 just use electricity. And this is where Chalet's research comes 0:13:02.160 --> 0:13:06.040 in because we've heard about AI's carbon footprint, but what 0:13:06.160 --> 0:13:09.720 about its water footprint, which could be a much bigger 0:13:09.760 --> 0:13:13.520 concern for us living here on earth. That's after the break. 0:13:26.480 --> 0:13:29.760 So you had a study come out last year called 0:13:30.000 --> 0:13:33.760 making AI Less Thirsty, uncovering and addressing the secret water 0:13:33.800 --> 0:13:36.520 footprint of AI models. What made you want to look 0:13:36.559 --> 0:13:36.800 at this? 0:13:37.400 --> 0:13:39.760 Maybe that was due to my childhood experience. I spent 0:13:39.840 --> 0:13:42.080 a couple of years in a small town in China 0:13:42.640 --> 0:13:45.520 where we only had water access for half an hour 0:13:45.600 --> 0:13:48.120 each day, so we just had to think about how 0:13:48.160 --> 0:13:51.560 to use water wisely and to every possible means to 0:13:51.600 --> 0:13:55.280 save water. Then in twenty thirteen, oh I saw this issue. 0:13:55.559 --> 0:13:57.640 I want you to find out more about it. What 0:13:57.720 --> 0:14:01.000 about the water consumption and nobody new at that time. 0:14:01.320 --> 0:14:04.160 A big environmental impact we don't talk about as often 0:14:04.200 --> 0:14:08.000 its carbon emissions is water usage. And the impact that 0:14:08.040 --> 0:14:11.080 water usage has on all of us depends on where 0:14:11.120 --> 0:14:14.600 that water comes from and where it goes when it 0:14:14.600 --> 0:14:17.240 comes to AI. A main use of water is to 0:14:17.240 --> 0:14:19.920 cool down the data centers, which, as we know, use 0:14:20.040 --> 0:14:22.200 a lot of energy. This is how they make sure 0:14:22.240 --> 0:14:23.440 that they don't overheat. 0:14:23.720 --> 0:14:27.720 To prevent syrup from overheating, usually we use water evaporation, 0:14:28.040 --> 0:14:30.760 and that's a very efficient way to move the heat, 0:14:30.880 --> 0:14:33.360 to dissipate the heat to the environment, and this water 0:14:33.400 --> 0:14:36.840 evaporation could be in the cooling towers. That is essentially 0:14:37.000 --> 0:14:38.600 evaporating water twenty four to seven. 0:14:38.960 --> 0:14:41.920 When water evaporates from a data center's cooling system, it 0:14:42.040 --> 0:14:45.120 goes out into the air and is basically considered gone, 0:14:45.360 --> 0:14:47.920 at least from the local supply. You might be thinking 0:14:47.920 --> 0:14:49.920 of the water that you use when you take a shower. 0:14:50.120 --> 0:14:53.160 How that water goes down the drain, It gets treated 0:14:53.520 --> 0:14:56.680 and it can be reused. But evaporated water rises up 0:14:56.680 --> 0:14:59.920 into the atmosphere and you can't reuse it. It can 0:15:00.160 --> 0:15:03.680 eventually come back down as rain, but that takes a while. 0:15:04.320 --> 0:15:07.600 Some tech companies they can use over twenty billion liters 0:15:07.600 --> 0:15:08.480 of water each year. 0:15:09.400 --> 0:15:11.720 Twenty billion, Yeah, that. 0:15:11.720 --> 0:15:15.040 Number basically is the same as some major beverage companies 0:15:15.120 --> 0:15:18.720 annual water consumption, the water they put into their product, 0:15:18.960 --> 0:15:22.320 basically the water we drink from a bottled water. Those 0:15:22.360 --> 0:15:25.080 are the water consumption for the beverage industry. So in 0:15:25.120 --> 0:15:29.520 some sense, this AI is turning these tech companies into 0:15:29.960 --> 0:15:32.320 a beverage company in term of water consumption. 0:15:32.880 --> 0:15:37.840 Nobody's drinking that bottled water or those sodas is just evaporating. 0:15:38.520 --> 0:15:39.840 Yes, yes, yes. 0:15:40.080 --> 0:15:43.080 One important thing here is that when Challte's talking about water, 0:15:43.360 --> 0:15:46.960 he's talking about a specific kind of water. For example, 0:15:47.080 --> 0:15:49.600 you might have heard that for every kilogram of beef, 0:15:50.040 --> 0:15:54.240 it needs fifteen thousand liters of water. But ninety percent 0:15:54.280 --> 0:15:57.560 of that water is what's called green water. That's water 0:15:57.600 --> 0:16:00.760 that's naturally stored in soil and used by plants like 0:16:00.960 --> 0:16:03.880 rain water. It doesn't have to be clean enough for 0:16:03.920 --> 0:16:06.520 people to drink it. It would be nice if data 0:16:06.520 --> 0:16:09.280 centers could use that, but that's not really practical for 0:16:09.320 --> 0:16:12.960 their usage. They rely on what's called blue water, the 0:16:13.000 --> 0:16:16.520 stuff that's clean enough for humans to drink. So when 0:16:16.600 --> 0:16:19.640 Chalet is comparing a tech company's usage of water to 0:16:19.720 --> 0:16:23.200 say like Pepsi's global use of water, this is a 0:16:23.240 --> 0:16:28.320 pretty direct comparison. Use the phrase when you're evaluating GPT three, 0:16:28.880 --> 0:16:32.000 that GBT three needs to drink a certain amount of water. 0:16:32.120 --> 0:16:35.840 A rap flight ten to fifty queries for five hundred 0:16:35.840 --> 0:16:38.120 million digits of water, so basically a bottle water. 0:16:38.360 --> 0:16:41.280 Let's pause on that number for a second. Ten to 0:16:41.280 --> 0:16:43.640 fifty queries the kind of thing you might do in 0:16:43.640 --> 0:16:47.760 a single session using chat GPT that could drink half 0:16:47.800 --> 0:16:50.800 a liter of water. I'm pretty sure that me going 0:16:50.840 --> 0:16:53.520 back and forth about buying a car, I probably used 0:16:53.560 --> 0:16:58.080 about a leader, and that's using conservative estimates. Challeat and 0:16:58.120 --> 0:17:01.280 his team. We're focusing on GPT three, which was released 0:17:01.320 --> 0:17:05.520 back in twenty twenty. Even five years later, OpenAI hasn't 0:17:05.520 --> 0:17:08.359 released all the details researchers would need to give us 0:17:08.359 --> 0:17:12.040 a clear picture of its environmental impact. Do the companies 0:17:12.080 --> 0:17:13.400 know how much water that they're using? 0:17:13.480 --> 0:17:15.600 Of course I can't really speak on their behalf, but 0:17:15.680 --> 0:17:18.800 I think they do. They could figure out the water 0:17:18.960 --> 0:17:22.240 consumption easily because they know their energy they know their 0:17:22.240 --> 0:17:24.719 water efficiency of the courting system, they know where they 0:17:24.760 --> 0:17:27.600 build the data centers, so they have the information, but 0:17:27.720 --> 0:17:29.399 we're not seeing their own disclosure. 0:17:29.800 --> 0:17:31.960 By this point you might be picking up on a 0:17:32.000 --> 0:17:35.720 recurring theme here. Putting a specific number on the impact 0:17:35.720 --> 0:17:39.679 of AI is basically impossible, and it's not because the 0:17:39.760 --> 0:17:40.919 math is too difficult. 0:17:41.080 --> 0:17:43.720 The thing is the tech companies are also refusing to 0:17:43.800 --> 0:17:46.119 tell us exactly what's going on. So if you take 0:17:46.160 --> 0:17:48.840 Google's environmental report, it will show you the numbers are 0:17:48.880 --> 0:17:52.000 bad because in twenty three they show that their carbon 0:17:52.040 --> 0:17:55.440 emations were up like fifty percent compared to five years before, 0:17:55.560 --> 0:17:58.439 and they were pointing to AI as the main culprit. 0:17:58.480 --> 0:18:02.760 They were saying, Okay, data center infrastructure is adding to 0:18:02.800 --> 0:18:06.159 our combon emissions, we're using more electricity. And at the 0:18:06.160 --> 0:18:10.920 same time they just don't specify exactly what's going on 0:18:11.000 --> 0:18:14.480 with regard to AI. They say that making distinctions is 0:18:14.520 --> 0:18:17.399 not meaningful at all, even though weirdly, Google was the 0:18:17.480 --> 0:18:21.240 company that just three years ago was in fact making 0:18:21.240 --> 0:18:24.200 this distinction. They were disclosing the ten to fifteen percent 0:18:24.240 --> 0:18:27.720 of their total energy costs were related to artificial intelligence. 0:18:27.800 --> 0:18:29.640 Now they stop doing that they don't want to tell 0:18:29.720 --> 0:18:30.320 us anymore. 0:18:30.560 --> 0:18:33.800 All of a sudden, they it seems like something changed there. 0:18:33.800 --> 0:18:35.240 What do you think changed? 0:18:35.480 --> 0:18:37.479 The numbers got big, that's what's changed. 0:18:38.800 --> 0:18:41.760 Okay, not to spoil the end here, but it looks 0:18:41.760 --> 0:18:43.439 like I'm not going to get a direct answer to 0:18:43.480 --> 0:18:46.200 my question. But at least I have something of a ballpark, 0:18:46.320 --> 0:18:49.159 even if it's a conservative one. And I also know 0:18:49.240 --> 0:18:53.120 that we're using AI every day for everything. We might 0:18:53.160 --> 0:18:56.600 not know the exact environmental impact of AI, but we 0:18:56.680 --> 0:19:00.159 do know that it's increasing, So what do we do 0:19:00.200 --> 0:19:12.880 about it? That's after the break. So in this episode, 0:19:13.000 --> 0:19:16.159 we've been having some trouble figuring out the exact environmental 0:19:16.160 --> 0:19:19.680 costs of AI. But this is a pretty common problem. 0:19:19.960 --> 0:19:22.280 I mean, my friend Matthew Galt wrote up an article 0:19:22.280 --> 0:19:26.320 at four or four Media explaining that the Government Accountability Office, 0:19:26.600 --> 0:19:29.919 which is a nonpartisan group that answers to Congress, is 0:19:29.960 --> 0:19:33.359 struggling with the exact same thing. They came up with 0:19:33.480 --> 0:19:35.960 roughly the same numbers that we talked about earlier. They 0:19:36.000 --> 0:19:39.119 put together a forty seven page report that acknowledges that 0:19:39.359 --> 0:19:44.920 even after interviewing agency officials, researchers, experts, they're still left 0:19:44.960 --> 0:19:48.880 with having to do estimates because, as they said, quote, 0:19:49.040 --> 0:19:53.520 generative AI uses significant energy and water resources, but companies 0:19:53.600 --> 0:19:57.840 are generally not reporting details of these uses. So even 0:19:57.880 --> 0:20:02.000 the US government has no idea exactly how much carbon 0:20:02.040 --> 0:20:04.800 we're pumping out or how much water we're pouring into 0:20:04.840 --> 0:20:09.200 the sand. And this is an issue because when researchers 0:20:09.240 --> 0:20:12.680 like Chalet and Alex were first looking into AI's environmental impact, 0:20:13.200 --> 0:20:17.160 the biggest concern was training. That's the one time process 0:20:17.200 --> 0:20:20.160 of feeding those massive data sets in the powerful machines. 0:20:20.640 --> 0:20:24.200 That's what was making headlines for energy use. But then 0:20:24.440 --> 0:20:29.120 came chat GBT three and suddenly people weren't just training models, 0:20:29.480 --> 0:20:33.719 they were using them all the time, and that shift 0:20:34.000 --> 0:20:35.200 changed everything. 0:20:35.840 --> 0:20:38.480 As an end user, you can't even manage it properly 0:20:38.560 --> 0:20:41.160 because the companies are not telling you. So it's not 0:20:41.280 --> 0:20:44.800 like when you're interacting with chut GPT that judge GPT 0:20:44.920 --> 0:20:47.760 is gonna tell you, Okay, be aware, now the carbon 0:20:47.800 --> 0:20:51.840 footprint of this conversation has already exceeded this amount. Open 0:20:51.880 --> 0:20:53.840 AI knows this kind of stuff. They could tell you, 0:20:53.880 --> 0:20:56.679 but they won't, And then other people are left trying 0:20:56.800 --> 0:20:58.919 to make some kind of customer to figure out what 0:20:59.040 --> 0:21:01.160 might be going on. We also see that they are 0:21:01.240 --> 0:21:03.880 kind of downplaying the impact of what they're doing here. 0:21:03.920 --> 0:21:07.480 I mean, we see their environmental reports or disasters. The 0:21:07.560 --> 0:21:09.920 carbon emitions are shooting up, and the only thing they're 0:21:09.920 --> 0:21:12.480 saying is like, Okay, don't worry about it. AI will 0:21:12.520 --> 0:21:14.480 solve this in a couple of years from now. 0:21:14.720 --> 0:21:17.560 So the thing that's causing the problem is going to 0:21:17.600 --> 0:21:18.439 solve the problem. 0:21:18.440 --> 0:21:22.480 Also, yeah, that's the excuse they're using. AI is going 0:21:22.520 --> 0:21:24.920 to solve it. It's bad right now, but everything will 0:21:24.920 --> 0:21:26.879 be better in a couple of years, trust us. But 0:21:27.119 --> 0:21:31.320 it's one hundred percent wishful thinking. And to be honest, 0:21:31.320 --> 0:21:34.880 if you look at the whole history of technological developments, 0:21:34.960 --> 0:21:37.320 even if we do end up realizing a lot of 0:21:37.359 --> 0:21:40.879 efficiency gains with AI, this is definitely not a given. 0:21:40.960 --> 0:21:43.639 It doesn't mean that our resource uses in total is 0:21:43.680 --> 0:21:46.439 going to go down. This is the infamous Jevins paradox. 0:21:46.920 --> 0:21:49.680 Jevins paradox is a concept that comes up a lot 0:21:49.680 --> 0:21:54.359 in AI recently. Basically, in the Industrial Revolution, cold powered 0:21:54.440 --> 0:21:58.399 engines started to get more efficient, and some people assume that, Okay, 0:21:58.400 --> 0:21:59.960 this is going to mean that now we're going to 0:22:00.200 --> 0:22:04.600 use less coal overall, but an economist named William Jevins said, no, 0:22:05.040 --> 0:22:07.840 this is going to have the opposite effect. As coal 0:22:07.880 --> 0:22:12.520 powered energy gets cheaper, demand will increase, and total consumption 0:22:12.600 --> 0:22:16.320 of coal won't go down, it'll go up. He was right, 0:22:16.600 --> 0:22:18.840 and that effect seems to keep repeating. 0:22:19.520 --> 0:22:22.639 Despite all the efficiency gains that we had. We're not 0:22:22.800 --> 0:22:26.320 saving on resources, we are using more resources. 0:22:26.440 --> 0:22:29.400