Will AI strain the grid, help it, or both?

Concern over how much artificial intelligence will strain the power grid just continues to mount. In fact, from 2022 to 2023, utilities across the country doubled projections for load growth in the next five years. 

But it’s unclear precisely how much electricity AI actually needs. 

• The top line: Experts point out that the correlation between AI and energy use isn’t direct: more of one doesn’t inevitably mean more of the other. According to Alp Kucukelbir, co-founder and chief scientist at Fero Labs, on a recent episode of Columbia Energy Exchange the technology’s applications in the power sector could — if utilities themselves play their cards right — result in a cleaner, more efficient, and reliable grid. 
• The outlook: Globally, power demand from generative AI could increase at a 70% compound annual growth rate, said David Sandalow, who is a former official with the Department of Energy and is currently a distinguished visiting fellow at Columbia University researching the overlap of energy use and AI, on the same episode. As AI creeps into almost every industry, it is driving more demand for data centers and computing power. However, there’s still room for the supporting hardware and software to increase in efficiency. And the technology can also create efficiencies in the power system, and help integrate more clean resources to meet rising demand. 

Forecasts vary on how much power AI will require in the next decade. 

“The recent rise of AI is…putting pressure on data centers that large technology companies are operating,” explained Kucukelbir on Columbia Energy Exchange, adding that “this type of computation has been traditionally serving things like the e-commerce sector, search in general, and other…’commodity applications’.” 

An “enormous” amount of growth is already being forecast in these areas, especially for large language models. That application, he said, “is new and is not fundamentally based on computation,” unlike applications like cryptocurrency mining. (Bitcoin specifically relies on a principle called proof of work, which requires a certain amount of computational power to be used to verify a new “block” on the chain.)

That isn’t the only way that AI is fundamentally different from crypto, though: “We already see the academic community, scientific community, working towards reducing the energy required to achieve similar outcomes in AI,” Kucukelbir said. “This is a statement I can't make in crypto.”

He added that hardware manufacturers are already thinking about how new electronics and chips will reduce energy consumption and achieve similar outcomes to today’s large language models.

“Think about how computation looked…in the 1970s, 1980s,” he said. “We can replicate, if not far exceed, the supercomputers of those decades with the devices in our pockets, who draw, I don't know how many orders of magnitude less energy.” 

Applications for the power system

The intersection of artificial intelligence and energy is not limited to power demand, though. The technology can also help solve problems related to siting and permitting, renewable intermittency, and power flow, according to both Kucukelbir and Sandalow. 

“I think we need to be ambitious and creative about using AI tools to help get over the challenges that we're facing right now in managing the electric grid,” said Sandalow. 

The three main ways the technology can be used are pattern recognition, forecasting or predicting, and optimization. 

Pattern recognition is useful for sifting through large amounts of data. Today, it’s commonly used by search engines to find commonalities among all the words typed into a search bar. This type of data is characterized as unstructured, and Kucukelbir says it “physically cannot” be structured. But machine learning can identify patterns in that lack of structure instantaneously. And that could help, for instance, make buildings more energy-efficient by making sense of their energy use tendencies. 

Meanwhile, the use cases of forecasting and predicting are frequently used interchangeably. But Kucukelbir said there is a key difference between the two: the time period over which the prediction occurs. Forecasting involves looking further into the future, and is essentially predicting over a specific time horizon. Prediction is commonly used today at solar and wind farms to maximize the output of variable resources, for example via tilting a solar panel toward the sun or turning a turbine to catch more wind. 

Forecasting further into the future, though, is useful for determining where to build those renewables, said Sandalow: “Any siting of electricity generation assets can benefit from AI technologies in terms of both weather and power demand in the area.” 

Optimization, however, complicates things. When predicting, a user feeds inputs to a model and asks for an outcome. But with optimization, a user creates an outcome, and asks the model what inputs are needed to achieve it. 

This, Kuckelbir said, is “arguably” the most valuable to users, at least when applied well. For instance, this type of machine learning can be used to optimize how power flows through the grid. Current algorithms can take hours or longer to run power flow optimizations, but AI can both speed things up and make them more accurate, Kucukelbir added. 

“Optimal power flow, I would say is…a problem space that would make the typical champions of AI very afraid,” Kucukelbir said. “You've got physical constraints of how energy is going to flow over a particular grid, network, or topology, and you need to satisfy physical requirements,” which is why approximations need to be as close to a perfect solution as possible. He added that AI is particularly well-suited to factoring in complexities like variable power supply or unmodeled demand.

 That technology is already progressing rapidly, according to Sandalow. 

“All across the power sector this technology can make a big difference,” he said. “And it's already starting to happen, but it's going to I think progress dramatically in the years ahead.”

Listen to the full episode:

Columbia Energy Exchange is a co-production of the Columbia Center on Global Energy Policy and Latitude Studios. Listen to all episodes here

Transcript

Alp Kucukelbir: It's important to delineate the two different camps of how AI really makes a difference today. There is a camp of use cases where AI helps us do things that we already do better, a little bit faster, a little bit more efficiently. Then there are these transformational applications. These are cases where we are doing something that we previously couldn't, or we've embarked on doing something in a different way that we previously couldn't.

Jason Bordoff: Artificial intelligence is a revolutionary technology with the potential to transform a wide range of sectors. For the energy transition, the applicability of this technology is broad, from methane monitoring, to integrating more renewables into the power mix. It can also be used to reduce emissions from food systems and in hard-to-abate sectors like steel and cement manufacturing. But the amount of energy AI will require is also a source of much interest, uncertainty, and concern, coming on top of the need for more and more electricity to help decarbonize sectors, from transportation to buildings.

So what are the high potential opportunities for using AI to combat climate change and what are the risks? How will AI exacerbate existing stress on the power sector, and what are some of the opportunities to lower costs and to increase efficiencies? This is Columbia Energy Exchange, a weekly podcast from the Center on Global Energy Policy at Columbia University. I'm Jason Bordoff.

Today on the show, David Sandalow and Alp Kucukelbir. David is the inaugural fellow here at the Center on Global Energy Policy. He founded and directs the Center's US China program. Previously, David served at the US Department of Energy and was a senior fellow at the Brookings Institution. He also served as an Assistant Secretary of State and as a senior director on the staff of the National Security Council.

Alp is the co-founder and chief scientist at Fero Labs. He's an adjunct professor of computer science here at Columbia University, and he leads the entrepreneurship efforts at Climate Change AI. David and Alp helped co-author the Roadmap on Artificial Intelligence for Climate Change Mitigation, published last year in Innovation for Cool Earth Forum. So I brought them on the show to discuss the report's findings, and the potential for AI to drive down emissions across a range of sectors and its range of applications for the energy system. I hope you enjoy our conversation.

David Sandalow, Alp Kucukelbir, thank you for joining us on Columbia Energy Exchange. Great to have you both with us to talk about your latest piece of research.

‍David Sandalow: Thanks for inviting us, Jason.

Alp Kucukelbir: Pleasure to be here.

Jason Bordoff: Okay, so you wrote an important report several months ago that The New York Times wrote up and several others have been referring to, and has been of interest to many in the policy world, about this broad topic everyone is talking about, artificial intelligence. What it means for the clean energy transition, what it means for our response to climate change opportunities and risks. And so that's what I want to talk about now.

David, we were just at CERAWeek in Houston, biggest energy conference, about 10,000 people. And it seemed like every panel included some reference to AI in multiple respects, and we'll come to different pieces of that. But as you know, because you got this question often, one of those was first, how much electricity this is all going to take? Decarbonization means we're going to increase electricity use for things like transportation and heat, and then also now, we have this new dimension where data centers and training machine learning models potentially, estimates seem to be all over the place, is going to take an enormous amount of electricity.

Tell us what you think about what we know about that question today, understanding there's surely a decent amount of uncertainty at least.

David Sandalow: Well, you're absolutely right about two things, Jason. First, there's a lot of attention to this topic, and second, there's a lot of uncertainty. But we do know some things. We know that power demand in the United States is projected to increase in the next several years at a much faster rate than it's increased in the past several decades. To be specific, in December 2022, US utilities submitting their five-year load growth projections for the Federal Energy Regulatory Commission projected 2.6% increase in next five years. By a year later, December 2023, their projections had almost doubled to around 5% growth.

And many equity analysts and other research shops are doing studies on this topic. I saw one research shop that projected that in the next four years, power demand from generative AI would increase at a 70% compound annual growth rate. That's a global figure. But there are other figures that are lower than that, and I'd underscore two things. First, there is significant uncertainty in this area. We know the generative AI, AI demand is going to increase significantly in the years ahead, and that's going to have an impact on power demand. But we also know that the efficiency of the hardware that's used for generative AI is going to improve dramatically, and that algorithmic efficiency, the efficiency of software, is likely to improve as well.

So there's a lot of uncertainty in this area. I think it's also worth noting that AI is by no means the only reason the power demand is growing in the United States. Under the Biden administration, there is a significant growth of manufacturing in the United States, reshoring and manufacturing. There's also growth in electric vehicles. Those factors, as well as others, are also driving the increase in power demand in the United States.

Jason Bordoff: Alp, let's step back for everyone listening and just be clear about what we're talking about, because I feel like a lot of times these things get all lumped together. There's data centers for cloud computing. There's new tools of artificial intelligence and the amount of energy it takes to train large language models. So again, you were one of the co-authors, along with David, who chaired this roadmap project on Artificial Intelligence for Climate Change Mitigation. So for everyone listening, when we talk about artificial intelligence, what are we talking about?

Alp Kucukelbir: Yeah, that's a great question, Jason. So I think AI has captured the imagination of the public, especially with the rise of large language models and ChatGPT, and how easy it is to use. So that's really created this landscape that's ripe for misinterpreting what this might mean. So when we think about what the recent rise of AI is really talking about, it is putting pressure on data centers that large technology companies are operating. And this type of computation has been traditionally serving things like the e-commerce sector, search in general, and other, let's call them commodity applications.

And that is where we see an enormous amount of growth being forecast. Parallel to this is also the amount of computation that we are looking at coming more from the scientific community through the realm of supercomputers. So these types of supercomputing clusters are typically outside of what we're talking about. This type of hardware is being used mostly by scientists and academia and government labs to really continue studying the climate and so on and so forth. So when we talk about AI, we're really talking about any type of computer application that is approximating what we think of as actions and planning, and other types of activities that we associate with human intelligence.

The most recent incarnation is large language models, but AI is not limited to large language models. So in our report, we do talk about applications of AI across a variety of sectors, be it the power sector, manufacturing, agriculture, where it's not just large language models making this kind of big difference. So when we are really trying to explore what the energy demands for AI are going to look like, that is part of the uncertainty. Large language models are very new. We're exploring where they can be helpful. We're trying to anticipate demand for this particular type of AI, and all of that is factoring into the uncertainty in these forecasts.

Jason Bordoff: And just on the question I asked, David, your take on it, give people a sense of how much electricity it really takes to train some of these models. As you said, we're just getting started. I suspect, like with the Internet in the early days of the digital technology revolution, we've barely scratched the surface for the use cases that we will find for tools like AI. So you see some projections that are really quite staggering, not just going from 2% to 3 or 4%, like David was talking about, but dramatic growth in electricity use. What's your best sense of where we're headed right now?

Alp Kucukelbir: Yeah, I am less concerned, I would say, of seeing that kind of exponential growth. Of course, it's possible, but I'll tell you why. People frequently think about crypto as an equivalent industry that grew very quickly and whose energy demands have remained extremely high, and that's by design. So the crypto effort, this entire activity, the societal activity is driven based off of requiring computation. The whole idea of mining currencies requires computation. It's baked into the activity itself. AI, especially with this enormous large language model, huge data crunching form of AI, is new and is not fundamentally based on computation.

What I mean by that is we see already the academic community, scientific community, working towards reducing the energy required to achieve similar outcomes in AI. This is a statement I can't make in crypto. We haven't really found a way of saying, "Let's get the equivalent value from crypto but at a lower energy footprint." It is just fundamentally not possible. So where I am hopeful is I'm seeing a lot of software development that will reduce the amount of energy required to achieve similar outcomes. I'm seeing a lot of hardware development, where hardware manufacturers are thinking about new electronics, new chips, new computing setups that will reduce the amount of energy that is required to achieve these similar outcomes.

Jason Bordoff: Meaning that the next, every generation of Nvidia chips or whatever, whoever's making these, you can get more computational power with the same energy input. They'd just get more efficient over time, that's part of the trend you're talking about.

Alp Kucukelbir: Precisely, yeah. Think about how computation looked like in the 1970s, 1980s. We can replicate, if not far exceed, the super computers of those decades with the devices in our pockets, who draw, I don't know how many orders of magnitude less energy than those did back in the day.

Jason Bordoff: And just so everyone knows, and I'll come to David in a minute, but I think people who have been following this podcast or the Center on Global Energy for a while know David and his background. Of course, I'll give your bios at the start of this podcast. But sometimes, from someone's title you don't get a sense of it. Just explain for people listening the work that you do day to day in computer science.

Alp Kucukelbir: I wear two hats. In the computer science department at Columbia, I'm an adjunct faculty. I teach a class called machine learning and the climate, and I research a branch of machine learning that I like to call explainable machine learning. This is a contrast to the black box machine learning technology that has really catapulted the e-commerce and technology sectors to drive the wonderful applications that we have, large language models being one of them. Explainability seeks to combat the issue that we have now recognized, for example, with large language models, the issue of hallucination, the inability to explain why the program is delivering the output that it is delivering given a specific query.

So this is an exciting area that has applications that I've been exploring in the manufacturing sector. In my call it other hat that I wear, which is as co-founder and chief scientist of a company called Fero Labs, which builds factory optimization software that's powered with AI. We sell software into the steel, cement, and chemicals sectors, which are, by share, the largest emitting sectors in heavy manufacturing. Our software is being used to help them reduce their energy consumption, reduce their waste, and reduce the variability of their production.

Jason Bordoff: David, just coming back to you on some of the numbers you gave and what you heard Alp say about these offsetting factors of significant growth in electricity, potentially because we're going to have more and more applications for AI, and on the other hand, the technology gets better and more efficient. We haven't seen electricity demand growth in the US for quite some time. It's been pretty flat. Getting ready to meet that demand, the permitting of infrastructure, transmission lines, which is so much a topic of discussion now.

So we're trying to meet that growth in demand at the same time that we're putting more and more intermittent sources of electricity onto the grid to meet the challenge of decarbonization. How do you think about our ability to handle and manage that right now? What do we need to be doing from a policy standpoint or otherwise to meet that challenge?

David Sandalow: I think it's a very significant challenge, Jason. We need to pull together all the forces within the policy world and the technical world to try to meet it. I think AI offers some helpful tools, interestingly. So AI can help us, for example, optimize production from solar and wind farms. That's pretty well-established technology actually, because AI is very good at predicting patterns. So we need to deploy the technology to help optimize the production from clean energy using AI tools. AI can help us optimize as well. It can help with optimal power flow problems and citing of transmission lines, and other issues like that.

And AI can help interpret in interesting ways. For example, some companies have put the databases from FERC orders and NERC orders, Federal Energy Regulatory Commission and Nuclear Energy Regulatory Commission online. And that allows querying of those databases in ways that may facilitate permitting over the years ahead. So I think we need to be ambitious and creative about using AI tools to help get over the challenges that we're facing right now in managing the electric grid, and growing the electric grid to address these problems.

And then things like basic permitting reform, which has been almost over the hump in the US congress in the past years. If we could possibly get that over the hump, that would make a big difference as well.

Jason Bordoff: Yeah, and so I want to come to those opportunities, because that was the focus of the report that you did. Obviously, there's a lot of interest in how much additional electricity demand these technologies will create. But as you say, there's an opportunity to help with integrating more low carbon sources of electricity onto the grid, or understanding other issues we need to understand, like emissions. So let's just talk about some of those. I think greenhouse gas emissions monitoring was the first one you talked about.

Maybe start with David and then go to Alp, if you could talk a little bit about, we're putting satellites up into space, we're trying to better understand where emissions, methane leaks are coming from, trying to get better granularity in what greenhouse gas emissions actually look like year to year. How will AI change that and how much of a difference will it... Give us a sense of magnitude. Is it helps a little bit or is this a real sea change in our ability to pay attention to emissions?

David Sandalow: I think it's a sea change, let's see what Alp thinks. But I think that we have the ability right now to understand greenhouse gas emissions in real time in a way that historically we haven't. So right now, vast amounts of data are being thrown off of satellites, aerial monitoring, drones, ground-based sensors. And that data is incredibly useful, but there's no way of interpreting and understanding it without machine learning and AI tools. Historically, we've relied upon voluntary self-reporting, to some extent fossil fuel data, consumption analysis, in order to understand the emissions that are coming from individual plants.

I think in the decade ahead, the combination of this new sensor data plus machine learning may give us the opportunity to really understand at a granular level what's happening in different places. And in fact, this is already making a difference in methane emissions policy, as I think you alluded to, Jason. The global methane pledge and the commitment to reduce methane 30% by 2030, and all the activity that's going on around methane policy right now really wouldn't be possible without these machine learning and AI tools. And in the years ahead, with spectroscopy and other types of tools, I think we can actually make a difference with these machine learning. And understanding where greenhouse gas emissions are coming from, and therefore in policy development.

Jason Bordoff: Alp, [inaudible 00:17:51] and you could help elaborate on that. We have a lot more data coming in on methane emissions, for example, and as David said, the tools we've been using historically to measure greenhouse gas emissions are imperfect. What's different about a world with AI tools than would've been true a decade or so ago to use all of that data and make sure that we have a clear understanding of what the emissions picture looks like.

Alp Kucukelbir: Yeah, it's a great point, Jason, and David set the groundwork really well here. So it's important to delineate the two different camps of how AI really makes a difference today. There is a camp of use cases where AI helps us do things that we already do better, a little bit faster, a little bit more efficiently. There, we're always talking about, let's say a 10 to 20% gain in whatever we're achieving. Then there are these transformational applications, these are cases where we are doing something that we previously couldn't or we've embarked on doing something in a different way that we previously couldn't.

And greenhouse gas emissions monitoring falls into the second camp, this transformational camp. So it doesn't come as a surprise, I'm sure, to any of your listeners that one property of AI is its ability to ingest enormous amounts of data. Large language models are, already there's articles being written today in major news outlets talking about companies that develop large language models not having enough information to feed them after having fed the entire internet to these models, just that statement alone is extraordinary.

So a decade ago, we could write software that was extremely expensive and difficult to build, to integrate data coming from these various satellite systems, let alone integrating data from different sources of data, such as drones and measurements on the ground. And putting all that together into one picture, into one system that allows us to just objectively say, "Here's the big picture." AI is ideally suited to be able to do that at scale, to do that in a way that allows us to really find the signal from the noise and really give this trust into a one source of truth, the actual ground truth of where are we emitting, how much are we emitting, when was it emitted, and really getting all stakeholders around this to agree and say, "Yes, this is the source of truth."

That was not possible with AI before. And so I would argue that that is the linchpin in taking all of these extraordinary hardware advances, where we've developed these satellites, we've developed these sensors, we've collectively managed to get them into orbit, drones, ground sensors, to then say, "Here's how we're going to bring it all together."

‍Jason Bordoff: So that's a helpful frame, that transformational or I'll call it marginal, although again, 10 to 20% change, that can be a big number too. So that's not a small thing. But just that framing is helpful. So from the standpoint of greenhouse gas accounting, David, how does that change the way we engage with the challenge of climate change, the policy response into international coordination? Does it just give us a higher degree of confidence? Your expertise is China, I don't think we're going to suddenly learn China's not the largest emitter. Maybe the numbers get revised slightly.

Or is it something like that ability to really pinpoint where the methane leaks are coming from that simply wouldn't have been possible with all the data coming in from satellites, et cetera, without this technology? How does it change what we know about the problem?

David Sandalow: So look, 10% of greenhouse gas emissions is five gigatons. So if we have a 10% difference, that's a heck of a lot. I think it makes a big difference in, for example, methane emissions regulation. We have learned over the course of the past 10 years what we didn't know more than a decade ago. For example, that a huge amount of methane emissions are coming from super emitting events or from large releases, and that affects the policy development and the response. And that's absolutely central, I think, to methane control, which is going to be fundamental to addressing the greenhouse gas problems. So that's just, I think, one example.

Another area, I think, is in research and development and innovation. And one area that I'm particularly excited about from a transformational standpoint is in materials innovation. I recall that when I got to the US Department of Energy in 2009, I remember receiving a briefing from the staff there saying that offshore wind would probably never be feasible because the marine environment's too corrosive, and the steel that you'd have to put out there couldn't really withstand the corrosivity of the marine environment.

My understanding is that, as the result of materials innovation, super lightweight materials and materials that withstand the corrosivity of the marine environment, we've now moved dramatically forward in terms of our ability to deploy offshore wind. Artificial intelligence can make an enormous difference in terms of accelerating the pace of materials development. And just by way of example, when Thomas Edison was inventing the modern light bulb 150 years ago, he took months to take dozens of different types of materials and run electric charges to them to find out what would happen. Today with AI tools, we can simulate a million of those types of interactions in a second. And that allows us to both [inaudible 00:23:17] select much more quickly and choose among different materials to find out what's best, but it also actually allows us to test materials that don't actually exist right now but might exist using simulation and chemical structural constraints.

And then if a hypothetical material seems to have good properties, synthesize it, create it, and move forward. So I think this is going to affect research and development budgets and research and development agendas around the world on clean energy dramatically in the years ahead.

Jason Bordoff: Alp, maybe you could comment on that point David made about materials. And particularly with, as you said, your background as the co-founder of Fero Labs, talk a little more about how that relates to these challenges, for example, in hard-to-abate sectors.

Alp Kucukelbir: Absolutely. So I recently learned, and I want to share this on this podcast, that Thomas Edison's famous expression of "Genius is 1% inspiration and 99% perspiration" talks about this point, but actually is attributed to Kate Sanborn, who's an American author, teacher, and lecturer. It's misattributed to Edison. Nevertheless, it does capture Edison's method. And so let's talk about how material science intersects with the energy transition. So if we think about lithium-ion battery technology, where we've come from the 1970s where we initially identified certain lithium-based materials for anodes, cathodes, electrolytes, we have improved on the efficiency of those batteries in terms of the chemistry and the design over the next 50 years.

What AI really allows us to do from a transformational perspective is to say, "How do we do that 50 year process of making lithium-ion batteries more efficient? How do we compress that to five years for sodium-ion technology, for solid-state batteries, for the next type of energy storage method that we're going to start exploring? How do we do this more rapidly?" And that is enormous, we're in a race against time here. And the ability to use AI to quickly accelerate the progress of science is enormous.

But transformational opportunities, as you described, Jason, do not lie solely in just these handpicked, cherry-picked sectors. There are applications of these across the board. So I'll give an example in steel, which is considered a hard-to-abate heavy manufacturing sector, alone responsible for anywhere between 6 to 8% of the global carbon footprint. Electrifying steel involves using more and more scrap metal as the feedstock. So steel manufacturing traditionally has been designed to take virgin material mined from across the earth, process those to make high quality steel. To electrify steel in a very straightforward way, you want to use old cars, old appliances, scrap metal from rail as your feedstock, use electricity to melt that and make new steel.

The issue there is, every batch of steel that you melt, today, it's a bunch of Hondas, tomorrow, it's a bunch of Chryslers, slightly different. And so what steel manufacturers do currently is that they cap the amount of recycled steel that they use. They cap at 25%, 30%. They still rely on the high quality, pure ingredients to make high quality steel. But with AI actually coming in and giving guidance to the operators, "Every five minutes, every 10 minutes for this batch of steel, this is how you need to operate your plant to make sure you get the high-quality product." We see steel manufacturers in the United States being able to push the boundary of how much recycled steel they're using, 50%, 75%, 80%, even for high-quality, high-grade steel that they need to produce for a variety of applications. That's transformational. So that's an application of AI where you're doing something, you're operating your factory in a way that previously you couldn't. And without this technology, you can't do that.

Jason Bordoff: Really, really interesting. So again, coming back to what's 10 or 20%, which makes a big difference, and what you might call transformational. You talk in the report about the impact this just has on the power sector more broadly, the ability to manage demand, energy efficiency, what it means for renewables to forecast weather better, predictive maintenance, things like that. So talk a little bit more about what this technology... We've talked about how much more electricity it could use, but as you said earlier, David, in passing, that it can help with how we build this grid and hopefully make it cleaner. Say more about that. And is that 10, 20%, or is that transformational? Again, 10, 20%'s a big number, so I don't want to downplay that, but you know what I mean. Give us a sense of the magnitude of the impact.

‍David Sandalow: It's very important. We didn't try to quantify it in our report, but every stage of the power sector can be significantly affected and actually is already being affected by this technology. We've already mentioned that solar and wind farms use AI today frequently in order to better predict the solar, the variable solar and wind resource and maximize output. Any siting of electricity generation assets can benefit from AI technologies in terms of both weather and power demand in the area, a variety of other different factors.

Geothermal power can benefit a lot from AI in terms of understanding subsurface conditions, and it seems to be important for the development of geothermal power. One area there's a lot of interest in right now is in nuclear power innovation, the ability of AI to simulate and do what's called digital twinning. And dramatically accelerate the pace at which we understand new nuclear technologies without having to invest hundreds of millions of dollars in building or billions of dollars in building new facilities. And then there's transmission. On the transmission side, AI can help with what's called dynamic line rating and other types of ways of improving the operation of transmission lines.

At the end use stage, AI can certainly help with building energy efficiency, better understanding patterns within a building and energy use patterns. And then with virtual power plants, right now we have distributed resources around the electric grid, machine learning AI tools are really fundamental to using those. Vehicle-to-grid technologies, which I think is a hugely important area in VPPs, in virtual power plants, are going to depend upon AI technologies. So basically all across the power sector this technology can make a big difference. And it's already starting to happen, but it's going to, I think, progress dramatically in the years ahead.

Jason Bordoff: Alp, anything you want to add to that? And then if you could, for people for whom, when we talk about AI, I may have used ChatGPT for something, but it's still somewhat abstract concept, like maybe a concrete example or two on the generation side or the demand side to help people understand how this is really going to be deployed.

Alp Kucukelbir: So I think maybe that point is worth digging a little bit deeper, and maybe I'll start with a little bit of abstraction and give a little bit of a more specific example as we go along. So AI can do a variety of things, and it's important to distinguish between them. So one is just sifting through large amounts of data. Think about this as pattern recognition. It's search. Search done in a way that you have messy data, you have a lot of data, it's unstructured, the work hasn't gone in. It physically cannot go in to make it structured. AI can go through this and identify those patterns.

Another set of things that AI can do is forecasting, predicting. Frequently these are used interchangeably. Forecasting typically involves some notion of time, thinking about a time horizon over which you're forecasting. Predicting might just be the next hour, it might be just simulating some sort of scenario and predicting what would happen under that scenario. That's another camp of activities there. And then the most complex, but arguably when applied well the most high value add, is optimization. So this is when you're giving a very complex problem that is very difficult to solve with traditional software methods, AI frequently can either come up with approximations so that they're fast, that help solve the true problem, the hard problem exactly. Maybe an approximate solution is fine. So that's another example.

So optimal power flow, I would say is one of those cases where you have a, let me put it this way, you have a problem space that would make the typical champions of AI very afraid. So the folks who are in the e-commerce and the ad tech and the technology space, they have some trepidations trying to solve problems like these, really hard problems. You've got physical constraints of how energy is going to flow over a particular grid, network, or topology, and you need to satisfy physical requirements.

You can't just say, "Oh, ChatGPT, give me your best estimate of going, what's..." No, certain equations of physics need to be satisfied if you're going to adopt this type of technology. So AI is actually quite good at once you iron out the wrinkles and really figure things out of answering questions like, "Okay, if I have variable supply or I have some sort of unmodeled demand, how do I use this existing network to optimally route my power?"

Jason Bordoff: David, you mentioned how it can help with geothermal just as one example because you can better map the subsurface, do directional drilling more accurately. I take it all of those are consistent with the point that this technology can be transformational to advance clean energy, but I presume also oil gas as well?

David Sandalow: No question, and the oil and gas industry is using it pretty extensively today for all kinds of purposes, and it has been for a number of years.

‍Jason Bordoff: And you also in the report, we don't work quite as much on this at the energy center, although Columbia does more broadly and it's a very important climate issue, what impact this technology can have on food systems globally. I was wondering, David, if you could talk a little about that, and then Alp.

‍David Sandalow: You got some very important applications here. So first of all, AI tools can help in addressing the impact of the food system on the climate for reducing greenhouse gas emissions, the food system for example. Optimizing the application of nitrous oxide on fields and fertilizers that will emit nitrous oxides from fields, and also in developing new crops, new innovative crops that may have better properties with some of the same type of simulation tools that we were talking about before. And then absolutely AI tools can help in better predicting weather patterns and other types of climate phenomenon that will have an impact on the food system that may damage the food system.

Jason Bordoff: Alp, anything to add to that? Again, giving people an example, maybe a use case there.

Alp Kucukelbir: Absolutely. So again, keeping our framework in mind, you've got the 10, 20% doing things better. So this is, how do I use less fertilizer to achieve the same crop yields? How do I use better forecasting to not waste the amount of how I'm using my land? How do I use my land a little bit more optimally to reduce waste and things like that? How do I integrate data from different sensors? Now you're no longer talking about satellites, but maybe you have five ground sensors and two drones. How do I get the data between those two sources together?

Then you've got the transformational applications. How do I build a more drought-resistant strain of a particular grain? How do I think about a more heat-resistant alternative of a specific staple? These are the more transformational. We can do this today. We do it trial and error. It takes time. How do we get ahead of that?

Jason Bordoff: And when you think about, this applies to everything we've been talking about, but as you were describing that particularly for agriculture, you think about the potential, but then you think about the access that any individual landowner has to that kind of information. What you're describing sounds like pretty high-tech stuff, and you're talking about in some cases large multinational companies, but often individuals, families.

Then when you think about around the world where agriculture is done, where the emissions are coming from, it raises the question of the accessibility of these tools you were talking about. How restricted will they be? How complex? Can they just be on someone's mobile device, or is it much harder and more expensive than that? And what are the risks for how we think about what this could mean? Everything you're talking about, food or otherwise, could this be transformational say in wealthy or developed countries and leave others behind?

David Sandalow: Such a key point. How long have you got, Jason? We've been talking about a number of tremendously high potential applications for artificial intelligence in addressing climate change problems, energy system problems. None of those outcomes are inevitable. There are barriers to achieving all of them, and there are risks as you started to point out. So just to talk about the barriers for a minute, we highlight two barriers on our report as being probably the most significant, people and data. And with respect to people. None of this happens unless we have people who are trained across a whole range of disciplines. And we certainly need the computer scientists who can develop the algorithms that do this type of work, but we need a lot more than that actually. We need climate experts who understand enough about AI to understand how their field can benefit from the application of AI tools, and the same in the energy system.

And we just need people generally to understand how this type of work can integrate into their institutions. Actually, we recommend in our report that every institution with a role in climate change mitigation have a top advisor for AI to the CEO or to the minister. And we're really pleased that last week the Biden administration announced that every federal agency will have a chief AI officer, and that's exactly, I think the type of direction that makes sense in addressing the people issues around this. But then there's also data issues, and you were starting to get this at your question. Making sure that there is available and accessible data to prepare these AI models, to train these AI models is going to be incredibly important. And then making sure that the results are accessible to people of different types all around the world is going to be key as well. A lot to say about that.

Jason Bordoff: Yeah. Alp, anything you want to add? Either the particular question of exacerbating or narrowing North-South divides, or more broadly some of the risks David talked about?

Alp Kucukelbir: Yeah, so maybe start with the risks and end on a higher note. So on the risk side we're talking about almost every segment of the economy, and we're talking about applications that we can really transform how we do things. So human health and safety is typically a concern that AI has not had, or the sector in general has not been applied to problems where that is a concern. Security, applying technology like AI to the grid involves security concerns that go beyond the current applications of Ai. So all of this needs to be top of mind. These are legitimate risks of the adoption of technology like AI at scale.

Jason Bordoff: You're talking about cybersecurity risks, it can exacerbate those.

Alp Kucukelbir: Correct.

Jason Bordoff: Okay.

Alp Kucukelbir: Correct, correct. But on a positive note, if you look at the barriers, there is a world in which the exploration of what is needed to make this technology productive is done in wealthier nations, that allow developing nations to leapfrog. And not waste the time that's needed to explore and develop this technology on their own. So if best practices are developed in terms of the application of AI, let's say to agriculture, could that help accelerate the development of developing nations to get to those results quicker?

‍Jason Bordoff: And Alp, when you, again, thinking about what'll have a smaller or larger difference, we've talked about a number of things just in the last 30 minutes. If you had to spotlight, there are many potential impacts on clean energy, on our understanding of climate change, but there are really one or two that strike you as being order of magnitude larger than the rest. Is that true or not? And if so, just help people listening understand what might be the biggest thing in this somewhat extensive list of potential impacts and opportunities you identified.

Alp Kucukelbir: Yeah, here's how I think about it. So if you think about that, let's call it marginal, the 10, 20% benefit, I just apply that to the size of the pie, so the slice of the pie that the sector occupies. So in that regard, I see power and manufacturing as being the two sectors that will benefit net numbers the most from technology that we know works right now. And we just need to get out. We need the right incentives, we need the right bureaucracy, we need the right support mechanisms to deploy this technology at scale. And if so, we will reap those benefits, and that is positive.

On the transformational side, there's higher uncertainty. I don't know if material science innovation will lead to a [inaudible 00:41:12] that will make carbon capture 10 times cheaper or 10 times more effective, or that will make sodium ion battery technology a hundred times more efficient in the next five years. But if it does, these will have huge impacts. So these transformational applications of AI have higher uncertainty, but potentially could have a higher impact if they're successful too.

Jason Bordoff: Is that, David, how you see it? These are all exciting opportunities, but if you had to say the ones that strike you as the biggest.

David Sandalow: Exactly, I agree exactly what Alp just said. I think there's potential for tremendous transformational benefits in the materials innovation space. We don't know if that'll happen. It's high uncertainty, but very high reward if it does. But I've just underscored the point that these so-called incremental improvements we've been talking about are actually huge in the context of climate and energy policy. If we're talking five to 10 gigatons of reductions a year, that makes a big, big difference.

Jason Bordoff: And I just, let me play devil's advocate and just push you on some of this, because the report lays out so many exciting and promising areas where this could make a transformational difference. And in preparing for this, I went back and tried... I had a vague recollection, it was a long time ago, but you can go back to the early days of the internet and find Xerox researchers saying, "We'll never use paper again." You can find myriad reports saying internet and digital technology will increase efficiency and reduce emissions by allowing for telecommuting.

Of course, global energy demand's risen about 50% in the last 25 years. There were reports from World Resources Institute and others predicting we would democratize access to information and build awareness around the world for strong environmental action. You find all of these things that it's like, here's the opportunity, here's all the things it could do. And in retrospect, it didn't necessarily have the impact. Why do you think AI might be different, or will it?

David Sandalow: I'm so glad you raised this point, Jason, because none of these results are inevitable, and there's enormous uncertainty. First, we're at the beginning stages of transformational technology having an impact in the world, so the directions are quite uncertain. But I think it underscores the need for policy, that policy guidance is hugely important. Innovation can happen in a variety of different areas. Innovation can happen with respect to technologies that are not good for the planet, or innovation can happen with technologies that are good for the planet. And so we need policy that helps guide us.

And that's why it's so important in this area that we have government step in and do things like bring together the communities that are working on this topic in order to better understand each other, help to develop training programs, help to support research in areas that are going to make a difference. Help to standardize data and make data more accessible, help to address bias issues that we've touched upon in this. All these things are important. The role of policy in this area is absolutely central if we're going to get the results that we hoped for.

Jason Bordoff: Alp, you've been doing this for a long time. Is that question fair? Do we suffer from optimism bias with new technology? And is that potentially applicable here and what needs to be done to make sure we realize some of these opportunities?

‍Alp Kucukelbir: Yeah, I completely agree. So techno optimism, if that's the term to use here, is potentially doesn't really achieve the end goal. But at the end of the day, artificial intelligence, just like many of these things that you mentioned, Jason, like the internet, it's just the general purpose technology. And so recognizing that not buying too much into the hype, trying to sift through the noise to find the signal is important. And I agree with David here that we need incentive structures to really guide towards the outcome that we are looking for here.

Jason Bordoff: David, can you say more about what your policy recommendations would be to achieve what you just described?

David Sandalow: Yes, thank you, Jason. We start with some institutional recommendations like the one I mentioned, which is that every institution should be paying attention to this with people at the top. Then government should use its convening powers. One of the, I think, lowest cost abilities of government is to bring different communities together, and that's very much needed here. For example, in this area, bringing together climate experts and AI experts. Focusing research and development dollars on applications that will have a difference in this area. There's been a tendency in some areas to focus on the next breakthrough in AI innovation. We recommend focusing research and development dollars on how AI can be applied for benefits.

A big area which is beyond in some ways the scope of AI, but hugely important in this area are utility incentives, utilities, [inaudible 00:45:55] and have incentives that cut against improving energy efficiency, for example, against investing in certain clean technologies. It's important to align utility incentives here with the outcomes that we want. And then we recommend as well international cooperation in this area. Institutions like the UN Framework Convention on Climate Change, the Clean Energy Ministerial and others can provide a platform for sharing information in this area that can be very helpful globally.

Jason Bordoff: Alp, what are you most worried about? One risk might be we don't realize opportunities, that's a missed opportunity. But in terms of the things that could be even worse than that, what are you worried about with how this technology might get deployed or misused, and what risks should we be paying attention to, particularly in the energy and climate space?

Alp Kucukelbir: Yeah, the biggest technical risk, I'll add to David's list here, is bias. And bias means something specific in the AI community. It means the data that is used to train the machine learning and AI software systems do not reflect the full picture of what we're trying to solve. So it's quite easy to fall into that trap of saying, "Hey, we found a way to predict climate patterns, forecast patterns really, really well, but it only works for the northern hemisphere well. And we omit the bias of not having a good enough data for, let's say, another part of the world."

Similarly, crops that get developed that only work with specific soils, that those soils are not available. We don't take the full picture into account. There's increasing awareness in this. I think the applications of AI on human data with privacy and various other inequalities have created awareness around this, but it's a little bit less prominent when we think about applications in these novel sectors.

Jason Bordoff: And what's the solution when you do in fact have a much larger data set for one application or one area than another?

Alp Kucukelbir: Yeah, twofold. One is technical, one's much more about awareness. Awareness is the education, just people need to be aware that this is a risk. We need to have guidelines and frameworks in place to understand whether the bias risk has been mitigated in any AI technology that we are developing, and then ideally deploying. On the technical side from the research community, there is very active work towards quantifying this bias, detecting it, better methods to be able to understand whether there are gaps in how an AI system has been built. These are all tools that will make this process a little bit easier.

Jason Bordoff: David, you mentioned international coordination and cooperation on this, and of course you have deep expertise in China and how it's approaching the clean energy revolution and climate change. The US and China announced a new bilateral channel for consultation on AI in November of last year. Do you see this as an area of potential cooperation? What is happening? How do you view the role of China and AI? Is this going to exacerbate tensions that are already quite high between the US and China?

‍David Sandalow: Technology cooperation is probably the principle area of tension in the bilateral relationship right now along with Taiwan. I think the disputes over semiconductor chips have risen to the top of the agenda in the bilateral relationship. So I think genuine cooperation on AI in the US China relationship is going to be challenging. I think communication is incredibly important, and as you said, there's a channel right now for that discussion. I think that's incredibly important.

I hope it will grow and continue. But China's capabilities in this area are very, very significant. Enormous amount of the peer-reviewed literature coming out globally is coming from China. There's a lot of technical development happening in China in this area right now. The world will be a better place if we can find a way to maintain at least some amount of communication and cooperate, but it's not going to be easy given broader geopolitical tensions in the years ahead.

Jason Bordoff: Alp, not just with China, but more broadly, you're a computer scientist, academic, cooperating with people all over the world. How do you see the impact of potential geopolitical tensions on our ability to do that and work together?

Alp Kucukelbir: Yeah, I think one example is how Europe is approaching the challenge versus the US, Europe taking much more of a stick and US taking much more of a carrot approach and incentivizing how this technology gets adopted. And I think even that is an interesting divide, where technology providers in Europe have to really think about how does their development fit into the regulatory environment in Europe. GDPR is much more on the personal privacy side, but the carbon border adjustment mechanism, so on and so forth, it's a very different landscape.

And so that coupled with the bias, again now, a different type of bias of how far ahead the United States is relative to Europe in the development of AI technology is leading to even tension among friendly and allied nation states.

Jason Bordoff: We're just about out of time, but I just want to ask each of you in closing, we've talked about there's a huge amount of interest on the potential for significant electricity demand from this technology. We've talked about some of the opportunities to lower costs, improve efficiency. What is most misunderstood, unappreciated? What's coming around the corner no one's talking about, that in the course of your work, you would highlight for people, people should be more aware of and paying closer attention to in the broad space of AI and the energy transition and climate? Alp, maybe I'll start with you.

Alp Kucukelbir: I think this hallucination problem is exacerbating, hopefully making it clear that these black boxes are not suitable for adoption into these hard to bait sectors, these high risk sectors.

Jason Bordoff: Do you want to just explain for everyone what you're referring to?

Alp Kucukelbir: Yep, absolutely. So if you take your ride sharing, favorite ride sharing app and it tells you, "Hey, it's going to take 17 minutes to get to your destination," you don't really need an explanation for that. Or if your favorite media app gives you a recommendation of a television show to watch, you don't really need an explanation. You don't benefit from it. If your ride hailing app gives you a confidence interval, it says, "You're going to get to your destination 17 minutes plus or minus five," that's not valuable to you. You just want to get there as quickly as possible. These are all just applications of machine learning and AI, where black boxes are, "Okay, we don't need explanations." There are low risk applications.

When we talk about the power sector, we talk about manufacturing, we talk about agriculture. The risks are too high for just adopting these types of just black mystery, black boxes. So explainability, tackling the problem of hallucinations, being able to understand when an AI system can be trusted and plugged into a workflow is essential.

Jason Bordoff: David, same thing, question for you. Just in your sense, what's most misunderstood or people have the least awareness of with this technology?

David Sandalow: I found Jason, in the past couple of months when I've had conversations with people about this, I would say roughly 80% of the commentary and questions I get are all about how AI is going to drive up power demand and cause problems. And I think the larger picture, the potential for AI to deliver enormous benefits here is just not getting the attention that I think it deserves. And so I hope that we'll see much more of that dialogue going forward.

As we've said in this discussion, those benefits are not inevitable, but they're absolutely achievable if we pay attention to them, if we have the right policy framework, and we have people dedicated to working on them.

Jason Bordoff: Al Kucukelbir, David Sandalow, thank you so much for your work on this report, and thanks for sharing your insights with us here on Columbia Energy Exchange today. I appreciate it.

David Sandalow: Thanks for having us, Jason.

Alp Kucukelbir: Thank you, Jason. Pleasure to be here.

Jason Bordoff: Thank you again, David and Alp, and thank you for listening to this week's episode of Columbia Energy Exchange. The show is brought to you by the Center on Global Energy Policy at Columbia University School of International and Public Affairs. The show is hosted by me, Jason Bordoff, and by Bill Loveless. The show is produced by Erin Hardick from Latitude Studios. Additional support from Paul DeBarre, Lily Lee, Caroline Pittman, Victoria Prado, and Kyu Lee. Roy Campanella engineered the show.

For more information about the podcast or the Center on Global Energy Policy, please visit us online at energypolicy.columbia.edu or follow us on social media at Columbia U Energy. And please, if you feel inclined, give us a rating on Apple Podcasts, it really helps us out. Thanks again for listening, we'll see you next week