To some, it might seem like our societies are obsessed with power.
No, not the kind of power that rests in a board room or a government building, although that’s fairly important to people, too, but energy. The power on which we run our businesses, and our lives. Power to make the machines go, air conditioners, cars, equipment. With a seemingly endless appetite for energy, humans have been wrestling with their need for consumption for decades, as we see more of the problems and consequences of relying on fossil fuels.
Then in comes AI.
I was watching a presentation on AI 2027, the often-cited narrative prophecy of what we’ll perhaps encounter during the next few years. An analyst, drawing on a whiteboard, was explaining that we’re used to things moving in linear ways, not exponential ones. That is a chilling idea when you look at the acceleration of AI, and correspondingly, its power draws.
These prognostications aren’t hard to come by, either. In addition to AI 2027, there are many other reports of a growing need for energy in the sector. (Here’s one from Axios supposing a 1000% increase by 2030). If you talk to people on the street, one of their biggest concerns is how much power it takes to ask GPT a question.
So this is, by any account, a big deal.
More on Energy, AI, and the Climate
I was at a panel at Imagination in Action at Stanford this fall where a group of presenters took all of this on, and made some insights around what it will take to deal with the energy issue.
Part of the problem, panelist Vijay Gadepally of Bay Compute pointed out, is that we need “power for everything,” even the most quotidian tasks.
“If you’re an enterprise interested in AI, you can’t get your hands on enough power,” he said. “That is the new problem of the day’s access to power…. I think because of that, coupling the fact that the usage has just blown up, you can’t do anything without AI… we can’t do anything without power consumption at this point. And so I think the answer is yes, we are making it worse, but I do think there’s a lot of opportunities to mitigate that or repair it.”
Tammana Sait of Crusoe gave some suggestions for energy-first approaches that don’t sacrifice sustainability. Carl Page of the Anthropocene Institute suggested that the problem is bigger than just applying band-aid solutions, calling for a more comprehensive project, like a “thermostat for our world.”
“If we zero out our carbon footprint, it does not make our world sustainable,” he said. “Global warming will accelerate anyway. We need to pull carbon out of our environment, and we need good thinking to help us do that. So sustainability requires making sure that our spacecraft is actually under control. …We need to actually use engineering to control our environment and make us safe and prosperous.”
Conservation Tactics
Along with the above ideas, Sait explained how bringing the data to where the power is generated makes a difference. This kind of colocation is being applied to some modern data centers, including the one that Crusoe is working on in Abilene, Texas, under the ambitious Project Stargate plan.
“Traditionally, you would have your power come to your data centers, but at Crusoe, we flipped the script a little bit,” she said. “We said we will take our data centers, and we will build data centers at sources of energy, where it is abundant and it is viable. We actually felt it was easier to move the data than to move the power, so we actually don’t consume power. We actually source it, we generate it, and then we optimize it, and optimize it at a site which is very viable, and then also at the same time (the site) has different sources of energy, sources that we could actually go to.”
Another plan, brought up by LL, is to “optimize for carbon-intensity.” For instance, Gadepally explained, perhaps a company doesn’t need its AI engine to provide long, detailed, constrained answers. Maybe half as good is good enough, and in some cases, more thinking could actually hurt the result.
‘We’re building these gigantic models to get a small increase in accuracy,” he said. “We’ve started to hit that (threshold of) diminishing returns with the growth of these models, and so we need to be a lot smarter with how we’re using them. And in many cases, we worked with a lot of enterprises that honestly are fine with a little less accuracy if it means that it could be half the computing, or half the energy budget. We see this all the time.”
Gadepally cited a study that Bay published earlier this year, based on such a project.
“All we told chat GPT was just: reduce the size of the answer based on carbon emissions, right? Based on the carbon intensity. Just give me a shorter answer, or give me an unconstrained answer. We saw a reduction of carbon emissions over 48 hours by 80%, but while still maintaining about the same accuracy, sometimes the same the shorter answer gave a better answer.”
Sait noted that a diverse portfolio helps, and fixing the misallocation of power could save a lot of resources.
Cleaner, Safer Energy Production
Amid this debate, America is investing in new nuclear power, which to some, seems like a timely solution.
“Fortunately, we have a fuel that’s a million times more efficient than coal, and it’s quite easy to use that fuel, nuclear, to provide abundant energy for our data centers, and for human needs,” Page said. “You know, Los Angeles recently had a big fire that consumed, two fires (consumed) $150 billion worth of stuff. And even if we bought Grandpa’s nuclear power plants, at the outrageous prices we took to build those cathedrals from the 1970s, we could have powered all of California three times over with the money we’re spending to fix up LA.”
“I am concerned about nuclear safety,” noted moderator Nina Gregory. “I’m not saying that we haven’t made advances in nuclear power, but I don’t think it’s fair to sort of poo-poo these concerns.”
Later in the talk, the topic turned back to nuclear power and its role.
“I think the most important objective is safety, human safety and climate safety, and it’s easy for nuclear power to prove it’s a million times safer than other sources of energy,” Page said. “For instance, (suppose) there’s someone trying to kill you, and they have a one out of 1000 chance of succeeding this year, and that goes for every one of you. And the person I’m talking about is the fossil fuel industry, because air pollution is a one in 1000 chance of killing everybody on earth each year. Nuclear power hasn’t killed anyone for a decade, so that’s much more than a million times safer, but we’re dealing with a world that’s run by PR flacks for the fossil fuel industry.”
Page continued, creating a pretty compelling and detailed argument for unachieved gains with nuclear.
“The Nuclear Regulatory Commission, for example, did not approve any significant change to nuclear power for half a century, 50 years,” he said. “If it had been an ordinary industry not even trying hard, it would have gotten 5% better every year. What’s 1.05 to the 50?”
This, he said, feeds the notion that AI can help us with this challenge.
“We need AI to help us with the institutional problems,” he said. “I think if we can get better institutional decisions, great. I’m not as worried as some people are about AI taking over the world, because I feel like institutions are always stupider than humans, and they are doing a great job at helping Wall Street and hurting humans, and we need to get past that, and I’m really looking forward to AIs helping us make better decisions for human beings.”
Guideposts for Change
Panelists went back to discussing how to optimize systems according to principles. Sait stressed reliability.
“Reliability and security is always your number one priority,” she said. “It does not matter what an innovative infrastructure you build, if you cannot use it, if it’s not always available, it’s not up and running 24/7 … reliability and security is most important.”
“My general suggestion (is): focus on things that have the human touch,” Gadepally said. “Those are things that are going to be very difficult for machines to take over. Yes, they can do information retrieval. All of that, yes, is fine, but … I still believe that people want to talk to humans…. in general, I think anything that has that social aspect and touch, I think it’s going to become more valuable as these tools come in and take away the little interactions that you used to have with the postal worker or delivery person, as these … drones (are) dropping stuff off at your house.”
There were some good tips here, about looking for energy in new ways, about optimizing brand new systems, and for engineering our interactions with AI. Stay tuned for more.