Mind and Iron: AI's climate-change paradox
It could solve the water crisis but bring down the grid. Also, Swifties vs. The Machine
Hi and welcome back to Mind and Iron. I'm Steve Zeitchik, veteran of The Washington Post and Los Angeles Times and pilot-in-command of this Cessna Skyhawk.
Every Thursday we bring you the ways innovations will shape our future — all you need to know about what’s changing, how we should feel about it, what we should do about it. Please consider pledging support for our mission.
Speaking of that Cessna, I'm writing this from a plane high above the eastern seaboard, because nothing says our global future like the below image. (Musk is a complicated guy, but maybe inhabiting other planets is a good idea if they could look like this.)
Last week we got a little forecast-y on you, offering a variety of thoughts about what 2024 might bring (tl;dr, AI medical breakthroughs, high-speed trains and a bad actor detonating nuclear weapons or AI deepfakes).
This week we're back in the thick of future-news. And the photo is apt because the view we take of nature determines what that future looks like. It’s seeming more and more probable that machine-intelligence can help us address crises like water insecurity, as a new AI nonprofit launched at Davos illustrates. But AI won't be able to pull off that environmental magic trick — or anything of major substance, really — without depleting our current supply to feed its hungry maw. Sam Altman himself just admitted that.
Which puts us in a fine how-do-you-do: AI can address the climate crisis only while simultaneously worsening it.
Our issue this Thursday examines this paradox. We also talk with one of the leading watchdogs on computing power and energy. The issue is important, so we should hear from the person who’s doing some of the most important work on it.
First, our future-quote of the week. It comes from said watchdog.
"If everyone wants to use ChatGPT, we have a problem.”
—Alex de Vries, sustainability researcher and founder of tech-consequence site Digiconomist, telling Mind and Iron what we’re facing, energy-wise
Let's get to the messy business of building the future.
IronSupplement
Everything you do — and don’t — need to know in future-world this week
AI addresses the water crisis; energy can’t solve the AI crisis; Swifties battle the pink robots
1. FIRST THE GOOD NEWS.
The statistics about water are as many as they are scary: about a quarter of the Earth's population doesn't have access to safe drinking water, according to a recent U.N. report, while double that number undergo some kind of water scarcity every year.
This is not only a catastrophe for the billions directly affected. It creates a trickle-down effect, literally, for the billions more who depend on the crops that such water scarcity impacts. Run that dishwasher and toss that case of Poland Spring in the trunk, as I did earlier today. But don’t forget how many of the eight billion people on the planet don’t have easy access to clean water.
Climate change is of course a big aggravator here, creating harmful algae blooms in drinkable water, never mind the capricious effect that floods and droughts have on ecosystems. (Floods are generally bad because they ensure that water picks up a lot of contaminants.) In Florida, Lake Okeechobee is a foot higher than normal this winter, setting off a host of environmental fears.
That’s obviously not the good news. The good news is that AI might have a little something to say about this.
I should note that I’m mostly skeptical that AI can save us from the worst effects of climate-change. The idea that algorithms will figure out how to build a better world for us sounds nice, but so much of climate-change is due to human irresponsibility, and machines can’t do a thing about that.
But at least some water challenges are matters of logistics — really complex logistics. If we knew where droughts would strike or floods would flash, we can prepare accordingly; if we understood at a deep analytics level the best use of limited water resources, we can mitigate the damage. Now, this kind of forecasting and modeling is something machines ARE good at. And a couple of companies are making news doing just this.
One such outfit is Pani. The British Columbia software firm approaches bottlers with the proposition it could both help them use less energy (a whole other vegan sandwich) and less water for every beverage it churns out. Yeah, that two-liter bottle of Coke ironically requires more than 100 gallons of water to produce. The company also says it could help plants that purify wastewater — typically at 20 times the cost and energy it takes to process freshwater — become more efficient. That lowers those costs and ultimately heightens access to clean water.
Fortune just ran a profile of Pani, and while not all of the company’s claims are slam-dunks — I’m not sure software sending information back to the managers of a water-treatment plant is really “just like a coach, watching, analyzing all the data,” as co-founder Devesh Bharadwaj says — there’s enough here to feel optimistic.
The World Economic Forum in Davos last week also brought the launch of another promising watery startup. Earth05, a nonprofit out of Barcelona, has the same use-the-little-details-to-achieve-big-goals gestalt of Pani. AI, the outfit says, can detect leaks, predict contaminations and, most interestingly, match crops to a region so that the wrong ones aren’t planted (if they are, it increases the likelihood of ruined crops and wasted water). "AI can give us these predictabilities that that we need in order to see the possibilities of what could be done," Earth05 founder Maria Dahrieh told ABC News.
I’m more dubious of some of Dahrieh’s other claims, like her contention that AI will help us find water deep underground, a kind of digital-age divining. But the takeaway of these types of efforts is that AI can be an ally in the climate crisis. For all the ways people might get machine intelligence to swipe our copyrighted work, atrophy our cognitive abilities, try to replace human connection or actually replace human jobs, AI can also, you know, improve the world in tangible ways.
If we want to build a future that gives billions of deprived people in some parts of the world access to clean water — and avoids turning water into a war-instigating have-and-have-not commodity everywhere else — we’re gonna need some help. Machines are asking ‘Why not us?’ It’s a question worth answering.
2. NOW FOR THE CONCERNING PART — WHAT EXACTLY IS POWERING AI IN THE FIRST PLACE?
Energy, of course. But what kind? And how much?
Sam Altman made headlines in Davos last week for saying that AI as he envisions it can’t be realized with current energy capacities.
“There's no way to get there without a breakthrough," he said. "It motivates us to go invest more in fusion.”
We don’t presently have the energy to build a shiny AI future, he says. Only something like nuclear fusion — that mythic, long hoped-for source of endless clean power — can do that. So where does this leave AI? And OpenAI, as it’s supposed to be the one building this future?
Altman’s agenda with his comment is complex; it could be everything from trying to court investment in his own nuclear-fusion startup to getting a regulatory hall pass when OpenAI starts gobbling up the grid. (“If you don’t buy us cheap power, we’ll kill this GPT.”)
But his underlying premise isn’t wrong and is backed up by science. On the trajectory it currently sits, AI could, in just a few years, be consuming as much energy as a mid-sized country. On Wednesday the International Energy Agency released a new report about energy consumption and technology, and it was sobering to say the least.
“Electricity consumption from data centers, artificial intelligence and the cryptocurrency sector could double by 2026,” the organization wrote, and went on to say, “After globally consuming an estimated 460 terawatt-hours in 2022, data centers total electricity consumption could reach more than 1 000 TWh in 2026. This demand is roughly equivalent to the electricity consumption of Japan.”
[A quick explainer on why energy is so crucial in the first place. AI operates at the pleasure of the data centers that enable so much of our lives. Data centers need power to run them hot and water to keep them cool. Super-powerful AI really needs power to run them hot and water to keep them cool. The source of all this new power and water that society has never needed before is something most of us don’t think about but probably should.]
The IEA report actually based some of its findings on the work of Alex de Vries, a Dutch researcher who is one of the leading people holding Big Tech to account on these issues. de Vries had for several years focused on all the energy that crypto-mining requires but has lately shifted his focus to AI. In October he published a paper in the scientific journal Joule that outlined these risks, including terawatt estimates (the energy consumption of Argentina or Sweden, is what just AI alone would equate to). He found that, if the supply chain cooperates, AI could cause a ten-fold increase in data-center activity. Not great, since data centers already account for about one percent of global energy consumption.
(Don’t get too excited btw that some of this energy could be renewable; our ability to satisfy green-energy demand is already at 100 percent, which means using renewables for AI would just shift other venues from green energy back to fossil fuels.)
Making projections about AI isn’t easy since the companies that traffic in it don’t give up much data. de Vries used a clever workaround involving the rate of production of NVIDIA chips, which at the moment power pretty much all AI platforms.
I reached de Vries at his home in the Netherlands Wednesday to ask where his head was in the wake of the IEA report and Altman’s comments. He said there was plenty of reason for concern given 1) how strongly big tech companies are committed to AI and 2) how much the whole system is premised on the notion that you need to keep building more powerful machines with more parameters; that’s after all what allows AI to do ever-sharper thinking for us.
"‘Bigger is better’ is what AI is built on. So we’re going to keep needing more and more power,” he said. (You can read our longer conversation with de Vries below.) Not to mention that as this tech improves it will get deployed in more places — smart-home apps, decision-making bots, AI Assistants. Which in turn makes more people use it. So what happens then? Energy consumption gone wild.
de Vries estimates that if it keeps growing at the current rate, the amount of power AI needs for both training and usage could hit 130 terawatts per year by 2027, which, taking the average carbon-dioxide emissions that a terawatt generates, would mean an additional 60 million tons of emissions, which all just sounds like a lot of numbers, except—
“For perspective that’s the total amount of emissions that electric vehicles around the world save compared to internal-combustion engines,” de Vries said.
Yes, AI would offset the gain of every single EV.
Of course AI is also what helps make EVs possible in the first place. Aaaaaaaah!
In fact, AI can help with the energy crisis in many other ways. Google, for instance, now uses neural networks to predict when wind might blow, allowing it to sell wind power in advance and making for much more efficient usage. Startups harness AI for wildfire detection, energy-demand forecasts and a slew of other potential applications.
This all amounts to a kind of innovation collision course: the more we’re doing nifty AI things to improve our environment, the more our environment gets screwed. We have enough of an energy crisis without adding to it with AI, no matter how much AI is also solving that energy crisis. (That makes my head spin like a turbine.)
But to simply opt out of AI because of these costs is wrong too; we need the tech to chip away at some of our planet’s largest problems. So this isn’t an all-or-nothing proposition, and please eye warily anyone who says it is. What should — and hopefully one day will — abide is a responsible use of such systems. Of calculating the benefits of AI with the costs of using it. (A common theme of this newsletter.)
The future will be made up of — indeed, ensured by — machine systems calculating the best uses of water, power and other critical resources. But all of this needs to be balanced with the climate cost of using these systems in the first place.
I mean, people still take airplanes. But they do so more judiciously, and often attempt carbon-offsets when doing so.
“AI is not like cryptocurrency, where all the data center does is create equations that it just throws away. We get something back with AI,” de Vries said. “But we’re going to need to think more carefully about what we’re getting because there is a sustainability cost. We can’t just go around looking for problems to put this technology on.”
As much as it appears invisible to us, massive computing operations require resources. And before we thrust headlong into deploying them, we should be considering the consequences. Because if something is going to resolve environmental challenges, it would be nice if it didn’t create new ones.
3. ON WEDNESDAY TAYLOR SWIFT BECAME THE LATEST VICTIM OF THE CELEBRITY DEEPFAKE EPIDEMIC.
This is the scourge in which pictures of female celebrities are manipulated, with the help of image generators, into authentic-seeming explicit images they of course never consented to. This week deepfakes of Swift began popping up, with one such image receiving 24,000 reposts on X before the platform finally removed it.
According to the independent tech-journalism outlet 404 Media, the manipulations originated on an abusive Telegram channel where a Microsoft image-generator is used (in case you by some chance still thought Big Tech could claim clean hands on any of this).
But there was a twist this time. “Taylor Swift AI” began trending on social media Wednesday night not just because of the deepfakes but because Swifties banded together to post real and respectful images of the pop star under the same hashtag. That basically flooded sites with legitimate images and sometimes even drowned the nefarious ones.
If you were looking for a perfect example of how humans can come together to defeat the worst of the machines, this was it. The Swifties had no help from image generators, which had been weaponized against an innocent person. Nor were they supported by the social platforms themselves, which for a (not-so-) shockingly long time left a lot of these deepfakes up on their sites; content moderation, who needs content moderation. The Swifties simply protected their human by using these tech companies’ own tools against them.
Yet as feelgood as this moment is, the larger reality is depressing. Plenty of celebrities don’t have a millions-strong stan army. And right now a deepfake still comes with telltale signs. In a Silicon Valley minute, though, models will improve and it’ll be that much tougher to tell what’s real and what isn’t. Which might make mobilization — never mind moderation — a lot harder to pull off.
Legislation is the only real answer here, stiff punishments that weed out the offenders who are doing this primarily because they can. New York State just enacted such a law. We’ll need more efforts that seek to protect the truth — and quickly, before we forget that it existed.
IronClad
Conversations with people living the deep side of tech
How AI could create a new energy crisis
A talk with Alex de Vries
Alex de Vries has a cool background. Which befits a guy trying to keep the planet from heating up.
A one-time consultant for PricewaterhouseCoopers, the 34-year-old has more recently turned to academic research. He has conducted fine-grained studies of how cryptomining, AI and other tech trends will affect the environment — the kind of stuff not a lot of people are doing but should be. It’s part of his Ph.D work at the Vrije Universiteit in the Netherlands and the basis of Digiconomist, his important site about unintended tech consequences. Oh yes, he has a day job as data scientist at the central bank of the Netherlands.
We talked to de Vries this week about the energy toll that AI could take and what hope there might be in reducing it. The conversation was edited for brevity and clarity.
--
Mind and Iron: So give it to me neat — how much should we be worried AI is harming the planet?
Alex de Vries: Overall, things are not looking too great. Cryptocurrency is already a pretty serious problem today. The biggest source of growth now will be AI combined with regular non-AI data centers. We could be looking at adding well over 100 terawatts just by 2027. We’re talking entire new countries’ worth of electricity just to power ChatGPT and our apps.
M&I: There are people who say that it’s ok, these data centers we’ll get better and not gobble up as much energy.
AdV: The problem is that we’re actually not making those efficiency gains with the hardware anymore. Before 2020 data centers’ global electric consumption was basically staying flat at one percent even with all the new data centers and cryptocurrency mining over the course of the decade. It all got offset by efficiencies of equipment. And now we’re reaching the end of that; Moore’s Law doesn’t really apply anymore. The demand for all these digital products is going up but without any efficiency increases at the data centers. So we’re really going to see some rises in consumption.
M&I: Do you think this ultimately will be worse than crypto? Because crypto has been pretty bad.
AdV: Well with Bitcoin mining they’re doing all of this for about $10 billion [of revenue] ever year. And they’re responsible for half a percent of global electricity. Compare that to the big tech companies that are all in on AI, who have a lot more to spend than $10 billion. And then imagine what you get.
M&I: Plus it’s baked in to what they want to do, right? Bitcoin mining doesn’t really keep growing year-to-year. AI is about building faster and faster microprocessors so the models can do more and more.
AdV: ‘Bigger is better’ is what AI is built on. So we’re going to keep needing more and more power. Also we have to realize that all of this is based on consumer habits now, where people might use some AI and then stop or other people might not use any at all. But that could change. If everyone wants to use ChatGPT, we have a problem.
M&I: You actually estimated that if every one of Google’s nine billion daily queries became chatbot queries we’d literally be adding as much energy consumption as the country of Ireland.
AdV: Yes, that’s true. It’s based on the Alphabet CEO himself saying that if you convert every Google search query into a chatbot query the energy consumption will go up tenfold. And we haven’t even talked about the electronic waste. Or the water needed to cool these data centers.
M&I: You’re bumming me out here! Should we have any reason for hope?
AdV: The supply chain. The only good news is that it’s constrained — to do all of this added AI there will have to be real changes to its capacity. The industry just can’t supply the amount of chips to meet this presumed demand.
M&I: So we’re basically hoping the tech companies’ production capabilities get throttled.
AdV [laughs]: Yes.
M&I: The International Energy Agency released a report today drawing from and confirming a lot of the work you’ve done. Does this help move the needle?
AdV: We have seen how the attention on cryptocurrency has led to requirements by the EU for very specific disclosures on the environmental impact. But that took a long time — people were talking about transparency back in 2018. So I think it’s going to take a lot of time before we could hope to get those requirements for AI. We’re now where we were with cryptocurrency six years ago. People are first starting to realize ‘hey, this is really energy-intensive.’
M&I: Is disclosure enough? I feel like most people would just skate by those.
AdV: The next thing you need to do is make policy decisions based on the data. Theoretically disclosure can change people’s behavior a little but I’m a firm believer we need a more forceful push.
M&I: Is there also hope that maybe operating AI itself can become more efficient? That as these models get better we can do the same things with them using less data and thus reduce the energy footprint?
AdV: Actually efficiency gains in a good or service usually doesn’t mean that. There’s a classic paradox known as Jevons Paradox based on this economist William Stanley Jevons from 1865 which says that when you realize efficiency gains in a good or service it doesn’t mean you use less resources. You use more, because now more people want it. So to reduce the footprint you need to actually lower the demand. Which I don’t see happening with AI.
M&I: Finally, is there an argument that even with all the energy guzzling this is worth it? I mean, if AI can do half the things its makers believe with regard to climate solutions or medicine, then maybe, some would say, we should live with the environmental fallout?
AdV: There’s no doubt we’re getting something back with AI. But we should start asking more questions about what that is. In the end we know we’re facing a very tangible climate impact. That’s on one side. So then what is AI doing that something else can’t do? Sometimes the answer is a lot. But it’s not a magic cure-all, and sometimes the answer is not much. And if we’re talking about using resources that are very scarce, we really need to think hard about whether we should use them. Let’s not go around looking for problems to put this technology on. When you walk around with a hammer, everything you see is a nail.
The Mind and Iron Totally Scientific Apocalypse Score
Every week we bring you the TSAS — the TOTALLY SCIENTIFIC APOCALYPSE SCORE (tm). It’s a barometer of the biggest future-world news of the week, from a sink-to-our-doom -5 or -6 to a life-is-great +5 or +6 the other way. Last year ended with a score of -21.5 — gulp. But it’s a new year, so we’re starting fresh — a big, welcoming zero to kick off 2024. Let’s hope it gets into (and stays) in plus territory for a long while to come.
AI CAN GIVE A BIG BOOST TO OUR WATER SECURITY GOALS: Yes! +2.5
AI SUCKS UP WAY TOO MUCH POWER, CONTRIBUTING TO THE CLIMATE CRISIS IT PURPORTS TO SOLVE: No. -3
TAYLOR SWIFT FANS BEAT BACK THE DEEPFAKES, BUT FOR HOW LONG? -1.5