Latest / kill switch / is using AI worse than driving a car?
Transcript
- 0:09Speaker 1: I've lived in LA for a decade, and this whole time,
- 0:12Speaker 1: I haven't owned a car. When I tell people that,
- 0:15Speaker 1: they usually look at me weird. And yes, riding the
- 0:19Speaker 1: bus and walking and using a bike is less convenient,
- 0:22Speaker 1: but at this point I'm used to it. But sometimes
- 0:26Speaker 1: I do wonder if I should just give in and
- 0:28Speaker 1: buy a car like everyone else. So to help me decide,
- 0:31Speaker 1: I did what a lot of people do recently when
- 0:33Speaker 1: they're weighing options. I asked AI. I opened up Claude
- 0:38Speaker 1: dot AI and I input all my current transportation costs.
- 0:42Speaker 1: I put in my bus fares, my bike, the cost
- 0:45Speaker 1: of my occasional uber and I asked it to compare
- 0:48Speaker 1: that to the average costs of car ownership in Los Angeles,
- 0:51Speaker 1: so parking, gas, insurance, repairs, all that sort of thing.
- 0:56Speaker 1: And I asked it to give me the pros and
- 0:58Speaker 1: cons on either side, and it did. The big con
- 1:02Speaker 1: is convenience, which I already knew. And on the pro side,
- 1:06Speaker 1: it said that I was saving thousands of dollars per year,
- 1:09Speaker 1: but it added one extra thing. It said that I
- 1:12Speaker 1: could have the nice feeling of knowing that I was
- 1:14Speaker 1: also being eco friendly. And I thought, hold on, wait
- 1:18Speaker 1: a second, eco friendly, I just spent half an hour
- 1:22Speaker 1: running scenarios through an LM, which I know is built
- 1:26Speaker 1: off the back of a massive amount of computing, which
- 1:29Speaker 1: in turn means a massive amount of energy. So am
- 1:32Speaker 1: I actually helping the environment here? Or am I hurting it?
- 1:37Speaker 1: So this week I set out to answer what seems
- 1:40Speaker 1: like a pretty simple question, how bad is AI's environmental impact? Really?
- 1:46Speaker 1: And yes, before you ask, I did consider asking Claude
- 1:49Speaker 1: and maybe chat GBT about AI's own impact on the environment.
- 1:53Speaker 1: But then I figured, you know what, maybe this is
- 1:56Speaker 1: a question I should ask actual human beings. And I
- 2:00Speaker 1: found a couple of people who've been studying this stuff
- 2:02Speaker 1: for a while to help me parse all of this.
- 2:05Speaker 1: Is there a way that I can compare? Is my
- 2:08Speaker 1: AI usage worse or better than car usage, or worse
- 2:13Speaker 1: or better than my impact on the environment from eating
- 2:16Speaker 1: meat or something like that? Are we able to make
- 2:18Speaker 1: those kinds of comparisons.
- 2:20Speaker 2: That's what carbon footprints were kind of invented for, so
- 2:22Speaker 2: you can make this type of comparison. If you're driving
- 2:25Speaker 2: in fossil fuel based car, you know exactly how much
- 2:27Speaker 2: gas you're using and what that might mean in terms
- 2:30Speaker 2: of carbon emissions. That's pretty straightforward. It's much harder to
- 2:34Speaker 2: do this for AI.
- 2:39Speaker 1: I'm afraid from Kaleidoscope and iHeart podcasts, this is kill switch.
- 2:49Speaker 1: I'm Jackster Thomas.
- 2:52Speaker 2: I'm sorry.
- 3:21Speaker 1: If you use social media, you've probably seen people being
- 3:24Speaker 1: criticized for using AI, and depending on who you hang
- 3:27Speaker 1: out with, that criticism can be kind of different in
- 3:30Speaker 1: how it shows up. When I see someone post an
- 3:33Speaker 1: AI generated image or some AI generated text and there's
- 3:37Speaker 1: angry comments in the comment section, it's usually one of
- 3:40Speaker 1: two types. The first is people saying that it's disrespectful
- 3:45Speaker 1: that by posting AI generated poetry or drawings you're devaluing
- 3:49Speaker 1: the original artists who didn't consent to having their work
- 3:52Speaker 1: fed into an l l M. That one's pretty easy
- 3:55Speaker 1: to understand, even if you don't agree with it or
- 3:57Speaker 1: think it's a big deal. The other comment I see
- 4:00Speaker 1: a lot of is people saying that using AI is
- 4:03Speaker 1: destroying the environment. Figuring out whether that's a big deal
- 4:07Speaker 1: or not is a little bit less straightforward.
- 4:09Speaker 2: What I tried to do in my research, I tried
- 4:12Speaker 2: to keep track of how the global electricity consumption of
- 4:16Speaker 2: AI is developing.
- 4:18Speaker 1: Alex Deviries is the founder of digit Economists and a
- 4:20Speaker 1: PhD candidate at the VU Amsterdam. He's been researching the
- 4:24Speaker 1: sustainability of new technologies for about a decade.
- 4:27Speaker 2: The way I do that is by looking at how
- 4:29Speaker 2: many machines of specialized AI's devices are being produced by
- 4:33Speaker 2: the AI hardware supply chain, and then considering that power
- 4:37Speaker 2: consumption profile how much power is now being consumed by
- 4:41Speaker 2: all of these devices. Which is a very imperfect way
- 4:44Speaker 2: of keeping track of this, but it's kind of like
- 4:46Speaker 2: the only rule you have available at the moment.
- 4:48Speaker 1: And even Alex is having a hard time keeping up
- 4:50Speaker 1: with this. What I called him. He was in the
- 4:53Speaker 1: middle of putting together new research. Back in twenty twenty three.
- 4:56Speaker 1: His data showed that by twenty twenty seven, new AI
- 4:59Speaker 1: service sold could use the same amount of energy annually
- 5:02Speaker 1: as the yearly energy consumption of a country like Argentina
- 5:06Speaker 1: or the Netherlands. But things have accelerated. His current research
- 5:10Speaker 1: shows that it won't take until twenty twenty seven for
- 5:13Speaker 1: that to happen. At this rate, we're going to hit
- 5:16Speaker 1: that mark sometime this year.
- 5:18Speaker 2: Simply because now the amount of devices that's being produced
- 5:21Speaker 2: by the AI hardware supply chain is way higher than
- 5:24Speaker 2: it was two years ago.
- 5:26Speaker 1: So it's even exceeding your pretty bleak estimations that you
- 5:31Speaker 1: made a while ago.
- 5:32Speaker 2: Oh yeah, it's just that the hype is so big,
- 5:35Speaker 2: and then the mod for this type of hardware is
- 5:37Speaker 2: so big that the numbers are going up much faster
- 5:40Speaker 2: than could be anticipated just two years ago.
- 5:43Speaker 1: But hold on, before we get too much further, let's
- 5:45Speaker 1: just clarify what we're even talking about when we say AI.
- 5:50Speaker 1: If you could break it down for me, how does
- 5:53Speaker 1: artificial intelligence use natural resources?
- 5:57Speaker 3: Heah, it's a general umbrella term that includes many different things.
- 6:01Speaker 3: But right now, if you're talking to a person on
- 6:04Speaker 3: the streets random like when they say AI, they're referring
- 6:08Speaker 3: to large languine models, or maybe you meage generation models.
- 6:11Speaker 3: So these are the generaryty AI models.
- 6:14Speaker 1: Chale Wren is an associate professor of electrical and computer
- 6:17Speaker 1: engineering at the University of California, Riverside, and he's kind
- 6:21Speaker 1: of a colleague of mine. Our fields are completely different.
- 6:25Speaker 1: But a couple of years ago I taught a class
- 6:27Speaker 1: in the building right next to his. I'd had no
- 6:30Speaker 1: idea that on the same campus there was an expert
- 6:33Speaker 1: who'd been researching the environmental impact of generative AI the
- 6:36Speaker 1: whole time, and I thought, perfect, this guy's kind of
- 6:40Speaker 1: a colleague, so I can stop doing all this research
- 6:43Speaker 1: on my own and just go ask him. Can you
- 6:46Speaker 1: give me an idea of how, say, car usage compares
- 6:51Speaker 1: to usage of an AI model.
- 6:54Speaker 3: I would say having a large language or medium sized
- 6:58Speaker 3: language model. Right, roughly ten short emails could be consuming
- 7:05Speaker 3: a quarter of the electric kill what hour energy, So
- 7:08Speaker 3: that's roughly enough to drive a test in a Model three.
- 7:11Speaker 2: For one mile, or as Alex puts it, Chat GPT
- 7:15Speaker 2: must be running on like something like five hundred mega
- 7:17Speaker 2: what hours a day, which is enough to power a
- 7:20Speaker 2: small city.
- 7:21Speaker 1: Basically, chat GPT's overall daily energy use, it's about the
- 7:26Speaker 1: same as powering every home, every grocery store, every street
- 7:30Speaker 1: light in a small city like San Luis, Opistpo in
- 7:33Speaker 1: California or Ithaca in upstate New York. But what does
- 7:37Speaker 1: that actually mean for me and you? How much energy
- 7:40Speaker 1: does it take to just ask chat GBT one question
- 7:44Speaker 1: on a.
- 7:44Speaker 2: Per interaction basis? It's actually not that much. You're talking
- 7:48Speaker 2: about something like three one hours maybe per interaction that's
- 7:51Speaker 2: something like a low lumin led build that you have
- 7:54Speaker 2: running for one hour. It's not a lot of power,
- 7:56Speaker 2: but it's it's nevertheless significantly more than a standard Google.
- 8:00Speaker 1: Search in one second. Just as an aside here, we
- 8:03Speaker 1: usually don't think of something like a Google Search as
- 8:05Speaker 1: using electricity. I mean, your phone or your computer is
- 8:08Speaker 1: already on, so what does it matter if you're typing
- 8:11Speaker 1: stuff into it or not. But on the other end
- 8:14Speaker 1: of that Google search you typed in, they're servers and
- 8:17Speaker 1: those are using energy. So as we keep going in
- 8:20Speaker 1: this episode, maybe think about that that on your end
- 8:23Speaker 1: you're not seeing any energy used or environmental effects, but
- 8:27Speaker 1: doing a Google search, watching a video, or even downloading
- 8:32Speaker 1: this podcast that does use some amount of energy.
- 8:35Speaker 2: Even Google's CEO at some point commented, like, hey, interacting
- 8:39Speaker 2: with these large language models, it takes ten times more
- 8:43Speaker 2: power than the standard Google search. So and that would
- 8:45Speaker 2: mean that if you're talking about three one hours for
- 8:48Speaker 2: interaction in a large language model, for a standard Google search,
- 8:51Speaker 2: it would be like zero point three one hours, which
- 8:53Speaker 2: is a very very tiny amount.
- 8:56Speaker 1: Just to explain here, a what hour is a unit
- 8:58Speaker 1: that tells you how much energy device uses over time.
- 9:02Speaker 1: For example, a sixty watt light bulb running for an
- 9:04Speaker 1: hour uses sixty watt hours. A single Google search uses
- 9:09Speaker 1: about zero point three watt hours. That's enough to power
- 9:12Speaker 1: that same light bulb for around eighteen seconds. But now
- 9:16Speaker 1: there's that AI add on that comes stacked by default
- 9:19Speaker 1: on top of every Google search, which takes that number
- 9:22Speaker 1: up ten x, up to three full watt hours per search.
- 9:26Speaker 1: That's a little different now you're running that same light
- 9:29Speaker 1: bulb for three full minutes and then but.
- 9:32Speaker 2: It's of course in the number of interactions where these
- 9:36Speaker 2: numbers start to stack up quickly, Because if you're talking
- 9:38Speaker 2: about Google Skill, you're talking about nine billion interactions a day,
- 9:42Speaker 2: going as three one hours per interaction. Then, interestingly, the
- 9:46Speaker 2: whole company Google would require as much power as Ireland
- 9:50Speaker 2: just to serve as a search engine.
- 9:51Speaker 1: If that was the case, Wow, using as much power
- 9:55Speaker 1: as a small country sounds wild, but if we think
- 9:59Speaker 1: about it, it kind of makes sense. We've default ten
- 10:02Speaker 1: xed our energy use overnight across nine billion searches a day.
- 10:08Speaker 1: That energy use is going to add up pretty fast.
- 10:11Speaker 1: But there's another thing to consider when we talk about
- 10:13Speaker 1: AI's energy use. The difference between training the model or
- 10:17Speaker 1: giving it a bunch of data to teach it how
- 10:19Speaker 1: to work, and using it like when you ask it
- 10:22Speaker 1: to write a cover letter or I ask it if
- 10:25Speaker 1: I should buy a car. When we talk about AI
- 10:28Speaker 1: and the energy consumption that can go into AI, there's
- 10:30Speaker 1: different phases, right. There's the training phase. There's me actually
- 10:34Speaker 1: sitting down and asking you an agent, a question. Can
- 10:38Speaker 1: you break that down for me?
- 10:39Speaker 3: The training part we call it learning, so based on
- 10:43Speaker 3: the data we try to optimize the parameters so that
- 10:46Speaker 3: when we see some new queries from the users, we
- 10:49Speaker 3: can give you as equate an answer as possible. And
- 10:51Speaker 3: training is really one time. Of course, later we're going
- 10:55Speaker 3: to do some update fine tuning. Inference is when the
- 10:58Speaker 3: users actually interact with the model, and depending on the
- 11:01Speaker 3: popularity of the model, but once it gets trained, it
- 11:04Speaker 3: could be used by many hundreds of millions of times
- 11:07Speaker 3: or even billions of times if you train the large
- 11:10Speaker 3: lunguine model like LAMA three point one. According to the
- 11:13Speaker 3: data released by Meta Training, a large lan grain model
- 11:16Speaker 3: like that air pllutant we gamerated through the training will
- 11:20Speaker 3: be roughly equivalent to more than ten thousand round trips
- 11:24Speaker 3: by cart LA and New.
- 11:25Speaker 1: York City ten thousand round trips by car. Yeah, so
- 11:29Speaker 1: that sounds bad, that sounds like a lot. But is
- 11:34Speaker 1: that a one time and it's just the one time.
- 11:36Speaker 3: It's a one time.
- 11:39Speaker 1: Let's clear something up here. That number ten thousand round
- 11:42Speaker 1: trips from LA to New York by car. It's not
- 11:45Speaker 1: just about carbon it's about air pollution, specifically things like
- 11:49Speaker 1: nitrogen oxides and fine particles that come from power plants
- 11:52Speaker 1: and can get deep into your lungs. This isn't theoretical.
- 11:56Speaker 1: This is stuff that raises risks of diseases like cancer,
- 12:00Speaker 1: and it doesn't just affect people next to the place
- 12:02Speaker 1: where all those computers are. Pollution travels and it lingers.
- 12:07Speaker 1: So what Challet's talking about here isn't just numbers. His
- 12:11Speaker 1: calculations are showing that training a single model the size
- 12:14Speaker 1: of Metaslama three point one can produce that level of
- 12:17Speaker 1: pollution on its own. So yes, training these models is
- 12:21Speaker 1: a one time hit, but it's a big one. If
- 12:25Speaker 1: we're talking just about energy usage, using an LM to say,
- 12:29Speaker 1: write ten emails might be like driving an electric vehicle
- 12:32Speaker 1: for a mile. And since an electric vehicle was maybe
- 12:35Speaker 1: three times more efficient than a gas vehicle. Figure, those
- 12:38Speaker 1: ten emails might get you a quarter to a third
- 12:40Speaker 1: of a mile in a regular car. And yeah, maybe
- 12:43Speaker 1: these are relevant numbers for me and my decision about
- 12:45Speaker 1: whether or not my AI usage is counterbalanced by me
- 12:49Speaker 1: not having a car. But these numbers are just estimates,
- 12:52Speaker 1: and we are going to get to that. But the
- 12:55Speaker 1: bigger issue here is that running those data centers doesn't
- 12:58Speaker 1: just use electricity. And this is where Chalet's research comes
- 13:02Speaker 1: in because we've heard about AI's carbon footprint, but what
- 13:06Speaker 1: about its water footprint, which could be a much bigger
- 13:09Speaker 1: concern for us living here on earth. That's after the break.
- 13:26Speaker 1: So you had a study come out last year called
- 13:30Speaker 1: making AI Less Thirsty, uncovering and addressing the secret water
- 13:33Speaker 1: footprint of AI models. What made you want to look
- 13:36Speaker 1: at this?
- 13:37Speaker 3: Maybe that was due to my childhood experience. I spent
- 13:39Speaker 3: a couple of years in a small town in China
- 13:42Speaker 3: where we only had water access for half an hour
- 13:45Speaker 3: each day, so we just had to think about how
- 13:48Speaker 3: to use water wisely and to every possible means to
- 13:51Speaker 3: save water. Then in twenty thirteen, oh I saw this issue.
- 13:55Speaker 3: I want you to find out more about it. What
- 13:57Speaker 3: about the water consumption and nobody new at that time.
- 14:01Speaker 1: A big environmental impact we don't talk about as often
- 14:04Speaker 1: its carbon emissions is water usage. And the impact that
- 14:08Speaker 1: water usage has on all of us depends on where
- 14:11Speaker 1: that water comes from and where it goes when it
- 14:14Speaker 1: comes to AI. A main use of water is to
- 14:17Speaker 1: cool down the data centers, which, as we know, use
- 14:20Speaker 1: a lot of energy. This is how they make sure
- 14:22Speaker 1: that they don't overheat.
- 14:23Speaker 3: To prevent syrup from overheating, usually we use water evaporation,
- 14:28Speaker 3: and that's a very efficient way to move the heat,
- 14:30Speaker 3: to dissipate the heat to the environment, and this water
- 14:33Speaker 3: evaporation could be in the cooling towers. That is essentially
- 14:37Speaker 3: evaporating water twenty four to seven.
- 14:38Speaker 1: When water evaporates from a data center's cooling system, it
- 14:42Speaker 1: goes out into the air and is basically considered gone,
- 14:45Speaker 1: at least from the local supply. You might be thinking
- 14:47Speaker 1: of the water that you use when you take a shower.
- 14:50Speaker 1: How that water goes down the drain, It gets treated
- 14:53Speaker 1: and it can be reused. But evaporated water rises up
- 14:56Speaker 1: into the atmosphere and you can't reuse it. It can
- 15:00Speaker 1: eventually come back down as rain, but that takes a while.
- 15:04Speaker 3: Some tech companies they can use over twenty billion liters
- 15:07Speaker 3: of water each year.
- 15:09Speaker 1: Twenty billion, Yeah, that.
- 15:11Speaker 3: Number basically is the same as some major beverage companies
- 15:15Speaker 3: annual water consumption, the water they put into their product,
- 15:18Speaker 3: basically the water we drink from a bottled water. Those
- 15:22Speaker 3: are the water consumption for the beverage industry. So in
- 15:25Speaker 3: some sense, this AI is turning these tech companies into
- 15:29Speaker 3: a beverage company in term of water consumption.
- 15:32Speaker 1: Nobody's drinking that bottled water or those sodas is just evaporating.
- 15:38Speaker 3: Yes, yes, yes.
- 15:40Speaker 1: One important thing here is that when Challte's talking about water,
- 15:43Speaker 1: he's talking about a specific kind of water. For example,
- 15:47Speaker 1: you might have heard that for every kilogram of beef,
- 15:50Speaker 1: it needs fifteen thousand liters of water. But ninety percent
- 15:54Speaker 1: of that water is what's called green water. That's water
- 15:57Speaker 1: that's naturally stored in soil and used by plants like
- 16:00Speaker 1: rain water. It doesn't have to be clean enough for
- 16:03Speaker 1: people to drink it. It would be nice if data
- 16:06Speaker 1: centers could use that, but that's not really practical for
- 16:09Speaker 1: their usage. They rely on what's called blue water, the
- 16:13Speaker 1: stuff that's clean enough for humans to drink. So when
- 16:16Speaker 1: Chalet is comparing a tech company's usage of water to
- 16:19Speaker 1: say like Pepsi's global use of water, this is a
- 16:23Speaker 1: pretty direct comparison. Use the phrase when you're evaluating GPT three,
- 16:28Speaker 1: that GBT three needs to drink a certain amount of water.
- 16:32Speaker 3: A rap flight ten to fifty queries for five hundred
- 16:35Speaker 3: million digits of water, so basically a bottle water.
- 16:38Speaker 1: Let's pause on that number for a second. Ten to
- 16:41Speaker 1: fifty queries the kind of thing you might do in
- 16:43Speaker 1: a single session using chat GPT that could drink half
- 16:47Speaker 1: a liter of water. I'm pretty sure that me going
- 16:50Speaker 1: back and forth about buying a car, I probably used
- 16:53Speaker 1: about a leader, and that's using conservative estimates. Challeat and
- 16:58Speaker 1: his team. We're focusing on GPT three, which was released
- 17:01Speaker 1: back in twenty twenty. Even five years later, OpenAI hasn't
- 17:05Speaker 1: released all the details researchers would need to give us
- 17:08Speaker 1: a clear picture of its environmental impact. Do the companies
- 17:12Speaker 1: know how much water that they're using?
- 17:13Speaker 3: Of course I can't really speak on their behalf, but
- 17:15Speaker 3: I think they do. They could figure out the water
- 17:18Speaker 3: consumption easily because they know their energy they know their
- 17:22Speaker 3: water efficiency of the courting system, they know where they
- 17:24Speaker 3: build the data centers, so they have the information, but
- 17:27Speaker 3: we're not seeing their own disclosure.
- 17:29Speaker 1: By this point you might be picking up on a
- 17:32Speaker 1: recurring theme here. Putting a specific number on the impact
- 17:35Speaker 1: of AI is basically impossible, and it's not because the
- 17:39Speaker 1: math is too difficult.
- 17:41Speaker 2: The thing is the tech companies are also refusing to
- 17:43Speaker 2: tell us exactly what's going on. So if you take
- 17:46Speaker 2: Google's environmental report, it will show you the numbers are
- 17:48Speaker 2: bad because in twenty three they show that their carbon
- 17:52Speaker 2: emations were up like fifty percent compared to five years before,
- 17:55Speaker 2: and they were pointing to AI as the main culprit.
- 17:58Speaker 2: They were saying, Okay, data center infrastructure is adding to
- 18:02Speaker 2: our combon emissions, we're using more electricity. And at the
- 18:06Speaker 2: same time they just don't specify exactly what's going on
- 18:11Speaker 2: with regard to AI. They say that making distinctions is
- 18:14Speaker 2: not meaningful at all, even though weirdly, Google was the
- 18:17Speaker 2: company that just three years ago was in fact making
- 18:21Speaker 2: this distinction. They were disclosing the ten to fifteen percent
- 18:24Speaker 2: of their total energy costs were related to artificial intelligence.
- 18:27Speaker 2: Now they stop doing that they don't want to tell
- 18:29Speaker 2: us anymore.
- 18:30Speaker 1: All of a sudden, they it seems like something changed there.
- 18:33Speaker 1: What do you think changed?
- 18:35Speaker 2: The numbers got big, that's what's changed.
- 18:38Speaker 1: Okay, not to spoil the end here, but it looks
- 18:41Speaker 1: like I'm not going to get a direct answer to
- 18:43Speaker 1: my question. But at least I have something of a ballpark,
- 18:46Speaker 1: even if it's a conservative one. And I also know
- 18:49Speaker 1: that we're using AI every day for everything. We might
- 18:53Speaker 1: not know the exact environmental impact of AI, but we
- 18:56Speaker 1: do know that it's increasing, So what do we do
- 19:00Speaker 1: about it? That's after the break. So in this episode,
- 19:13Speaker 1: we've been having some trouble figuring out the exact environmental
- 19:16Speaker 1: costs of AI. But this is a pretty common problem.
- 19:19Speaker 1: I mean, my friend Matthew Galt wrote up an article
- 19:22Speaker 1: at four or four Media explaining that the Government Accountability Office,
- 19:26Speaker 1: which is a nonpartisan group that answers to Congress, is
- 19:29Speaker 1: struggling with the exact same thing. They came up with
- 19:33Speaker 1: roughly the same numbers that we talked about earlier. They
- 19:36Speaker 1: put together a forty seven page report that acknowledges that
- 19:39Speaker 1: even after interviewing agency officials, researchers, experts, they're still left
- 19:44Speaker 1: with having to do estimates because, as they said, quote,
- 19:49Speaker 1: generative AI uses significant energy and water resources, but companies
- 19:53Speaker 1: are generally not reporting details of these uses. So even
- 19:57Speaker 1: the US government has no idea exactly how much carbon
- 20:02Speaker 1: we're pumping out or how much water we're pouring into
- 20:04Speaker 1: the sand. And this is an issue because when researchers
- 20:09Speaker 1: like Chalet and Alex were first looking into AI's environmental impact,
- 20:13Speaker 1: the biggest concern was training. That's the one time process
- 20:17Speaker 1: of feeding those massive data sets in the powerful machines.
- 20:20Speaker 1: That's what was making headlines for energy use. But then
- 20:24Speaker 1: came chat GBT three and suddenly people weren't just training models,
- 20:29Speaker 1: they were using them all the time, and that shift
- 20:34Speaker 1: changed everything.
- 20:35Speaker 2: As an end user, you can't even manage it properly
- 20:38Speaker 2: because the companies are not telling you. So it's not
- 20:41Speaker 2: like when you're interacting with chut GPT that judge GPT
- 20:44Speaker 2: is gonna tell you, Okay, be aware, now the carbon
- 20:47Speaker 2: footprint of this conversation has already exceeded this amount. Open
- 20:51Speaker 2: AI knows this kind of stuff. They could tell you,
- 20:53Speaker 2: but they won't, And then other people are left trying
- 20:56Speaker 2: to make some kind of customer to figure out what
- 20:59Speaker 2: might be going on. We also see that they are
- 21:01Speaker 2: kind of downplaying the impact of what they're doing here.
- 21:03Speaker 2: I mean, we see their environmental reports or disasters. The
- 21:07Speaker 2: carbon emitions are shooting up, and the only thing they're
- 21:09Speaker 2: saying is like, Okay, don't worry about it. AI will
- 21:12Speaker 2: solve this in a couple of years from now.
- 21:14Speaker 1: So the thing that's causing the problem is going to
- 21:17Speaker 1: solve the problem.
- 21:18Speaker 2: Also, yeah, that's the excuse they're using. AI is going
- 21:22Speaker 2: to solve it. It's bad right now, but everything will
- 21:24Speaker 2: be better in a couple of years, trust us. But
- 21:27Speaker 2: it's one hundred percent wishful thinking. And to be honest,
- 21:31Speaker 2: if you look at the whole history of technological developments,
- 21:34Speaker 2: even if we do end up realizing a lot of
- 21:37Speaker 2: efficiency gains with AI, this is definitely not a given.
- 21:40Speaker 2: It doesn't mean that our resource uses in total is
- 21:43Speaker 2: going to go down. This is the infamous Jevins paradox.
- 21:46Speaker 1: Jevins paradox is a concept that comes up a lot
- 21:49Speaker 1: in AI recently. Basically, in the Industrial Revolution, cold powered
- 21:54Speaker 1: engines started to get more efficient, and some people assume that, Okay,
- 21:58Speaker 1: this is going to mean that now we're going to
- 22:00Speaker 1: use less coal overall, but an economist named William Jevins said, no,
- 22:05Speaker 1: this is going to have the opposite effect. As coal
- 22:07Speaker 1: powered energy gets cheaper, demand will increase, and total consumption
- 22:12Speaker 1: of coal won't go down, it'll go up. He was right,
- 22:16Speaker 1: and that effect seems to keep repeating.
- 22:19Speaker 2: Despite all the efficiency gains that we had. We're not
- 22:22Speaker 2: saving on resources, we are using more resources.
- 22:26Speaker 1: And essentially you're saying here is even if we are
- 22:29Speaker 1: able to make AI more efficient, we're just going to
- 22:32Speaker 1: use it more, and so any efficiency gains are going
- 22:35Speaker 1: to be offset by the fact that we're just constantly
- 22:37Speaker 1: using this more and more and more.
- 22:39Speaker 2: One thing that's extra annoying with AI is that there's
- 22:42Speaker 2: also this bigger is better dynamic going on, whereas if
- 22:45Speaker 2: you make the models bigger, you'll actually end up with
- 22:48Speaker 2: a better performing model, but it just means that your
- 22:50Speaker 2: efficiency gains are completely negated all the time.
- 22:53Speaker 1: Every chat, every prompt, every AI generated jibbli image adds up.
- 22:58Speaker 1: We just don't see that impact directly, So let's all
- 23:03Speaker 1: just stop using AI. Right, Well, that's probably not realistic
- 23:07Speaker 1: at this point, and that's not necessarily what everyone's recommending.
- 23:10Speaker 3: So I work on optimization, and I think this is
- 23:13Speaker 3: a problem. We can optimize it, we can make it better,
- 23:15Speaker 3: reduce the cost, and there are a lot of opportunities,
- 23:18Speaker 3: so we should definitely not panic. I hope the model
- 23:21Speaker 3: developer can disclosee that cost to the users. I will
- 23:24Speaker 3: figure out should I use it now or should I
- 23:26Speaker 3: use it later.
- 23:27Speaker 1: Let's say that I log in the chat GBT and
- 23:30Speaker 1: it says this query is going to use this much energy,
- 23:35Speaker 1: this much carbon, and this much water. And if I
- 23:38Speaker 1: have that information up front, then I, the user might
- 23:41Speaker 1: decide maybe I don't need to have it summarize the
- 23:45Speaker 1: entirety of the collective works of Shakespeare today.
- 23:47Speaker 2: Yeah.
- 23:48Speaker 3: Maybe. Or they could tell you if you do it later,
- 23:51Speaker 3: in one hour or in the evening, the cost will
- 23:54Speaker 3: be different. And do you figure it out? Do you
- 23:57Speaker 3: want to do it now or do it later.
- 24:00Speaker 1: What Shelley is proposing here is that developers could build
- 24:03Speaker 1: in a system that would alert users that their query
- 24:06Speaker 1: is coming in at a high impact time of day,
- 24:08Speaker 1: and it could suggest that there might be a better
- 24:11Speaker 1: time to make that request when data centers have lower usage.
- 24:14Speaker 1: They can use optimization techniques to reduce energy consumption. This
- 24:19Speaker 1: concept isn't totally new. Google flights shows carbon emissions estimates
- 24:23Speaker 1: for flights and it will show you which option has
- 24:26Speaker 1: the least impact. So something like this for AI is
- 24:29Speaker 1: definitely possible, but I'm not totally convinced people would actually care.
- 24:34Speaker 1: The last time I booked a flight, I saw the
- 24:36Speaker 1: most carbon friendly option, but I didn't pick it because
- 24:40Speaker 1: it had a long layover. I didn't want to deal
- 24:42Speaker 1: with that. Putting the responsibility on users can sound good
- 24:46Speaker 1: in theory, but the flip side of that is it
- 24:48Speaker 1: can just be a way for companies to avoid doing
- 24:50Speaker 1: anything themselves. So should this responsibility really fall on us?
- 24:56Speaker 1: I mean, sure, you could decide to skip the chatbot
- 24:58Speaker 1: and take notes by hand, and that only really works
- 25:01Speaker 1: if you know what the trade off actually is, and
- 25:04Speaker 1: right now we don't, because the companies building these tools
- 25:08Speaker 1: aren't giving us the data that we would need to
- 25:10Speaker 1: make informed decisions in the first place. So maybe the
- 25:14Speaker 1: responsibility should fall elsewhere. Like policy makers, Shelley is already
- 25:19Speaker 1: thinking about what this could look like and how much
- 25:21Speaker 1: of a difference it could make.
- 25:22Speaker 3: We're informing the policy makers, hopefully when they make decisions
- 25:26Speaker 3: they could take into account this public health burden, water consumption,
- 25:31Speaker 3: power strain on their infrastructures. These are the cost the
- 25:35Speaker 3: local people will be paying for the companies. I think,
- 25:38Speaker 3: especially for those big techs, they already have the systems
- 25:42Speaker 3: ready to do this type of optimization. They are doing
- 25:45Speaker 3: it for carbon orware computing. And we use the math
- 25:48Speaker 3: as location as an example. If they factor in the
- 25:51Speaker 3: public health burden into their decision making, for example, where
- 25:55Speaker 3: they route to their workload, they can reduce the public
- 25:57Speaker 3: health cost by about twenty five percent really and reduce
- 26:01Speaker 3: the energy bo by about two percent and also cut
- 26:05Speaker 3: the carbon by about one point three percent.
- 26:09Speaker 1: So just by being more intentional where they route digital traffic,
- 26:12Speaker 1: a company like Meta could reduce detrimental impacts on public
- 26:16Speaker 1: health and they'd be saving some cash at the same time.
- 26:19Speaker 1: This is called geographic load balancing, and for the user
- 26:23Speaker 1: it's totally seamless. You log in your feed loads, you
- 26:27Speaker 1: don't notice anything, but behind the scenes, your request is
- 26:31Speaker 1: going somewhere where it's cleaner, cheaper, and less harmful to process.
- 26:36Speaker 1: Even beyond where companies route traffic, they can also consider
- 26:39Speaker 1: where they build the data centers from a public health perspective.
- 26:43Speaker 3: When they built data centers in the future, they can
- 26:45Speaker 3: take into account this of factors because the decision that
- 26:48Speaker 3: we make today will be impacting the public health, the
- 26:51Speaker 3: water consumption, the power infrastructure for many years to come.
- 26:55Speaker 1: Shelle is thinking about the future and research on future
- 26:58Speaker 1: optimization is a big deal because the AI boom is
- 27:02Speaker 1: already here. Big tech companies are projected to spend three
- 27:06Speaker 1: hundred and twenty billion dollars on AI technology and data
- 27:10Speaker 1: centers this year, which is nearly one hundred billion more
- 27:12Speaker 1: than last year. So where we put these data centers
- 27:16Speaker 1: and where we route the traffic really matters.
- 27:19Speaker 3: Something that I was not expecting to be widespread because
- 27:23Speaker 3: I was thinking, if I leave, let's say, five miles
- 27:26Speaker 3: away from a data center or power plan, I wouldn't
- 27:28Speaker 3: be affected. That was wrong. These air pollutants are what
- 27:32Speaker 3: EPA defines as cross state air pollutants. They do travel
- 27:37Speaker 3: hundreds of miles along with a wind. We're going to
- 27:40Speaker 3: have a significant impact just by strategically placing the data
- 27:45Speaker 3: centers for the public health.
- 27:46Speaker 1: What that really highlights is something that we don't think
- 27:49Speaker 1: about with tech infrastructure. It doesn't just impact the people
- 27:52Speaker 1: who live next door. When air pollution travels hundreds of miles,
- 27:56Speaker 1: it turns these data centers into regional issues, not just
- 28:00Speaker 1: local ones. I'll give you an example right here. As
- 28:03Speaker 1: we're working on this episode, I saw this article in
- 28:05Speaker 1: Politico and I just want to read you the first
- 28:07Speaker 1: sentence quote Elon Musk's artificial intelligence company is belching smog
- 28:12Speaker 1: forming pollution into an area of South Memphis that already
- 28:16Speaker 1: leads the state in emergency department visits for ASTHMA end quote.
- 28:21Speaker 1: That's probably enough to give you the idea. But just
- 28:23Speaker 1: to explain more, XAI, which is the company behind groc
- 28:27Speaker 1: which is the AI chatbot that you use on Twitter,
- 28:30Speaker 1: set up shop in Memphis with enough methane gas turbines
- 28:33Speaker 1: to power two hundred and eighty thousand homes. The company
- 28:38Speaker 1: didn't get the required air pollution permits. They're run without
- 28:41Speaker 1: the emission controls that federal law usually requires, and in
- 28:45Speaker 1: under a year of operation, XAI is now one of
- 28:48Speaker 1: the largest emitters of smog producing nitrogen oxides in the
- 28:52Speaker 1: entire county. And this facility is located near predominantly black
- 28:56Speaker 1: neighborhoods that are already dealing with high levels of indust pollution.
- 29:01Speaker 1: These inequalities already existed, and tech development is not making
- 29:06Speaker 1: it better, it's making it worse. It is often like this.
- 29:12Speaker 1: There are absolutely people who are feeling the impacts of
- 29:14Speaker 1: this right now, and there's people who will feel it
- 29:18Speaker 1: in the future. Maybe somebody will write an article about them,
- 29:23Speaker 1: maybe not. So I was hoping that I could use
- 29:26Speaker 1: this podcast to solve all my personal problems. But apparently
- 29:30Speaker 1: we're over one here, because when I started working on
- 29:32Speaker 1: this episode, I was thinking that this section right here,
- 29:35Speaker 1: the outro is where I'd say, Wow, now I know
- 29:39Speaker 1: exactly what impact my use of AI is having on
- 29:41Speaker 1: the planet. But I don't. And that's pretty annoying because,
- 29:47Speaker 1: and I guess this is as close to an answer
- 29:49Speaker 1: as we're going to get. It's not really about how
- 29:52Speaker 1: often I personally decide to use Chat, GPT or Gemini
- 29:56Speaker 1: or Claude or whatever. It's about what happens when companies
- 29:59Speaker 1: build systems that are this powerful but also this resource hungry,
- 30:04Speaker 1: and they refuse to tell us what it really costs.
- 30:07Speaker 1: And I think we deserve to know, not just so
- 30:10Speaker 1: that we can make individual choices about how often to
- 30:13Speaker 1: use Chat, GIBT or Gemini or whatever, but so that
- 30:16Speaker 1: we can hold the right people accountable, because if AI
- 30:19Speaker 1: is really going to change the future like they say
- 30:21Speaker 1: it will, we shouldn't know how much that future costs.
- 30:35Speaker 1: Thank you so much for listening to kill Switch. If
- 30:37Speaker 1: you got any ideas or thoughts about the show, you
- 30:40Speaker 1: could hit us at kill Switch at Kaleidoscope dot NYC,
- 30:44Speaker 1: or you could hit me at dex Digi that's d
- 30:46Speaker 1: e x d ig I on Instagram or on Blue
- 30:49Speaker 1: Sky if that's more your thing. And if you like
- 30:52Speaker 1: this episode, if you're on Apple Podcasts or Spotify, take
- 30:56Speaker 1: your phone out your pocket and leave us a review.
- 30:59Speaker 1: It really helps people find the show, and in turn,
- 31:01Speaker 1: that helps us keep doing our thing. Kill Switch is
- 31:04Speaker 1: hosted by Me Dexter Thomas. It's produced by sen Ozaki,
- 31:08Speaker 1: darl Luk Potts, and Kate Osborne. Our theme song is
- 31:11Speaker 1: by me and Kyle Murdoch, and Kyle also mixed the
- 31:14Speaker 1: show from Kaleidoscope. Our executive producers are Ozma Lashin, Mungesh Hatigadour,
- 31:20Speaker 1: and Kate Osborne from iHeart. Our executive producers are Katrina
- 31:25Speaker 1: Norville and Nikki e. Tour. See you on the next
- 31:27Speaker 1: one