Best of: Searching for climate solutions amid the AI hype
30:51 Share

This is an ad by BetterHelp. Let's talk numbers. Traditional in-person therapy can cost anywhere from $100 to $250 per session, and that adds up. But with BetterHelp Online Therapy, you can save up to 50% per session and pay a flat fee for weekly sessions, making therapy more affordable and easier to fit into your life.

Therapy should feel accessible, not like a luxury. BetterHelp connects you with licensed professionals who can help with everything from anxiety to everyday stress at a price that makes sense. Your mental health is worth it. And now it's within reach. Whether you're facing big life changes or just need support, therapy provides the tools to navigate it all.

Hi.

Hi, it's Akshat. This week, we are replaying a conversation I had with Priya Danti of MIT about the role AI can play in tackling climate change. No one is more knowledgeable or realistic about this area than Priya. And no one is doing more to ensure that AI's climate applications are steered responsibly. Enjoy the episode. Welcome to Xero. I am Akshat Rati. This week, what good can AI do?

Remember when data was the new oil? I'm specifically thinking of a 2017 cover of The Economist showing Google, Amazon and other tech giants as big offshore oil rigs. The idea being that data was a new critical resource and it was going to reshape the world.

In some ways, that has already happened. Maybe this podcast was suggested to you by Spotify or Apple based on your listening history. Just a small example of big data at work. Artificial intelligence, the latest buzzword, of course thrives on data.

That devouring of data is energy and resource intensive. It's something we discussed in last week's episode with Microsoft President Brad Smith. The company wants to be carbon negative, but is instead seeing its emissions grow. But of course, fed the right data, AI can do amazing things. Even help tackle climate change. But how exactly?

If there's one person taking the lead on that question, it's MIT's Priya Danti. She's a professor of electrical engineering and AI and the co-founder of Climate Change AI, an organization bringing together academics and industry leaders interested in how AI can be used for climate solutions.

Her group funds independent projects and fieldwork tackling everything from mangrove restoration for Indonesian shrimp farmers to the study of nanoporous separations in the chemical industry. And it also thinks hard about how to avoid AI being used to increase emissions and worsen human suffering. I asked Priya about some of the AI applications she's most excited about.

and why the conceptual framework we build around AI is just as important as the technology itself. Now, before we get into the heart of some of the work you do at Climate Change AI, I think it would be helpful to define the terms because there's just so many of them and that's just a mixing and muddling when people think about AI.

For most people, the biggest point of entry for AI is ChatGPT. ChatGPT is what people have played with. People kind of know it's based on this thing called an LLM, a large language model. But it's just one example of types of AI. So if you start at the very top, how would you define AI and what types of AI are there?

Yeah, so there isn't kind of one universally agreed upon definition for AI, but roughly you can think about AI as referring to systems that perform some kind of complex task. And there are two big branches of AI. One is rule-based systems.

which is when you kind of know how to do something, like you know how to play chess in some sense, you could write down the rules, but actually reasoning over those rules to figure out how to be a good chess player is the hard part. And so rule-based systems are places where you write down the rules and reason over them automatically. Oh, and that means...

Deep Blue beating Garry Kasparov for the first time in 1997, that would be classified technically as a rule-based AI system. That's right. Even with those primitive computers. Oh, yes. So AI has been around for a long time, actually. And so another type of AI is machine learning. And machine learning is often used in situations where

You might have intuition for something, but it's really hard to write down rules to codify your intuition. So if I gave you, Akshat, an image of a dog, you could probably tell me that it's a dog. But if I asked you to write down a set of rules that characterize exactly why it's a dog, it would be really hard for you to write down that set of rules exactly.

And so machine learning is a paradigm where you actually infer some of these rules automatically from examples or data. So I give you a bunch of images. Maybe I tell you which ones are dogs or cats. And the machine learning algorithm learns how to map between the images and the labels of is this a dog or cat and kind of infer the rules that cause that to be true.

And so if you take the type of AI that most people know, which is large language models, that's machine learning. That is machine learning. And large language models are basically...

one type of machine learning model that basically looks a specific way, has a particular specification of how you update it. And that type of model can be used in various different ways. And roughly the three kinds of ways they can be used are called supervised learning, unsupervised learning, and reinforcement learning. Well, it all sounds like you're trying to

teach a child it's either through supervision or through play or through punishment

Yeah, and in some sense, a lot of machine learning algorithms and ways of trying to learn these things are vaguely inspired by some notion of how humans learn, although the practicalities of how we actually do this might be quite different. And so we talk about AI in the climate context for two big reasons. One is because of the energy cost of creating AI models and using AI models. And second is that

These models, again, different types of them, can have different applications that could make solving for climate change, deploying these solutions, easier.

And I would add in a third pillar, which is that AI is also used for many types of applications that make solving climate harder. So when we talk about the good and the bad, we should think about the fact that AI has its own footprint and AI is used in both good and bad ways. Yeah. And so let's address the footprint part, because the size of footprint that may come from AI will be dependent on the type of AI.

And large language models are in the news because these are the models that try and train themselves on the entire corpus of the Internet. And that just requires a ton of computing power, which is why companies like Microsoft and Alphabet and Meta are all now

in this rush to build more data centers, consume a lot more power in the process, and blow past some of their own set climate goals as we found out with Microsoft's recent update that its emissions are actually 30% higher rather than 30% lower last year. Does that mean all kinds of AI is doomed to have a higher footprint? Because all kinds of AI will want as much data as possible.

So there's definitely a diversity in the types of AI that exist, and as a result, the kind of energy usage of these. So there has been in long history, you know, AI and machine learning models that use, you know, a reasonable amount of data, but much less than the entirety of the internet. And where the models themselves are also much smaller, they have fewer parameters. And as a result, you don't need as much computational power to actually update and get these models to learn.

And so, you know, some of the models that we develop, even in my research group, can run on a laptop. But then, of course, you have these, you know, large data intensive state of the art algorithms that are kind of being deployed through products like chat GPT and definitely the kind of energy consumption and, you know, water consumption from data centers, the materiality impacts of actually getting the computational hardware in place that is starting to get worrying.

Right. Now, climate applications themselves don't have to go down the LLM route of having to consume that much data. You know, you say models on a laptop can work. Let's start with that because you got into AI through trying to figure out how to make the grid work better, right? That's right. So basically, as we start to integrate, you know, more and more renewables into power grids,

Many of these renewables, their output varies based on the weather. So it varies over time. Think about solar, think about wind. And yet on a power grid, you're having to maintain this exact delicate balance between how much power is put into the grid and how much is consumed, which gets harder when you have a lot of variations coming onto the grid.

And so AI and machine learning can be helpful in terms of doing things like first, I mean, just giving us better predictions of what your solar power output, wind power output, electricity demand will look like. But also in actually helping to speed up some of the existing physics based and engineering based algorithms that are used to manage the power grid in the back end to maintain that balance.

And so one of the challenges with trying to understand AI as an application to try and help solve some of the climate problems is that it becomes really abstract very quickly. So you say, oh yeah, we have a number of data points and there's an intelligent way in which we can use them and that gives us an output, but we don't usually know why we have that output, but that output is better, so we use it and that's the solution.

And it just does not feel satisfying. You know, as a science reporter, to me, the joy of an invention is to actually break down the steps to try and figure out why this step led to that step led to that step. And finally, you have something that is really useful. Can we do that with AI?

Yeah, so I think that there are a couple of categories of ways we can think about AI and machine learning being used for climate that can help maybe give a mental model for what's actually going on under the hood.

So one of these categories is, you know, taking large streams of raw data and distilling it into actionable information. So one project we're funding through Climate Change AI is actually a project that tries to improve the sustainability of shrimp aquaculture.

practices. So kind of shrimp aquaculture is currently, it can be harmful to, you know, coastal mangrove forests. And that has implications for climate change adaptation in terms of kind of flood resilience, as well as climate change mitigation in terms of the sequestration potential of mangroves. And so we're currently funding a team from Arizona State Conservation International and Thinking Machines, a data science firm in the Philippines, to actually use satellite imagery to assess mangrove

aquaculture farms that actually might be able to benefit from better aquaculture practices. The intervention here is that you can actually do things like if you have an aquaculture farm, you can intensify how much you're farming on one part of the farm and then you can kind of conserve on another part of the farm. And so without impacting your overall productivity, you can just farm in a way that's better for the mangroves. And so Conservation International has a program where they're working with farmers to try to

kind of help them do this, but actually identifying which farms are amenable to this type of intervention at scale is difficult. So they use a combination of, you know, satellite imagery, data on like sea level rise and sea risk and things like this in order to then actually pinpoint at scale which farms might be amenable to this intervention and then actually go work with them to do that. And you said that was just one approach. What are some other approaches? There's a couple of other ways, actually. So one is, you know,

predicting and forecasting. So taking, you know, historical data where you have relationships between some input and some quantity you would want to predict. So things like I want to predict electricity demand on the power grid. So I can take historical data about what electricity demand looked like. I can take historical weather data and I can learn relationships between those so that in the future, when I have a weather prediction,

but I don't know what the electricity demand would be based on that weather prediction. I can just go ahead and predict that. And you have kind of, for example, nonprofits like Open Climate Fix that are working with the UK power system operator to actually improve their electricity demand forecast. And they've been able to use machine learning to half the error of those forecasts. After the break, why it's important for all of us to be involved in the development of AI.

By the way, if you're enjoying this episode, please do take a moment to rate and review the show on Apple Podcasts or Spotify. It helps other listeners find it.

And for digital banks, it means staying ahead and keeping up with evolving cyber attacks. Microsoft Security equips you with deeper insights to help you pinpoint vulnerabilities, see around corners, and innovate confidently. They scan trillions of signals daily, giving you the guidance, expertise, and tools to protect your business without sacrificing speed for safety.

Security is your job, and it's also theirs. With Microsoft Security, you have a partner that looks deeper, keeps you ahead, and helps your business move forward securely. To learn more, visit Microsoft.com slash CISO. That's Microsoft.com slash C-I-S-O.

Some companies, one that my colleague wrote about called Climate AI, is using AI to try and improve weather prediction models because currently you are starting to get better and better predictions

And that has, at least for them, been a profitable enterprise because then they are working with these large agriculture companies that want to figure out when should we start to put the seed down or when should we start to harvest because we have a better understanding of the weather, not just over the next two weeks, but over the next three months.

Yeah, so I think this idea of kind of medium to long term forecasting is also really cool. But often to do good forecasting in these settings, you want to use a combination of physical models and data. So, for example, one of the teams that we're funding at Climate Change AI spread between a couple of U.S. universities and an Indian university. They're basically trying to figure out how do I actually make longer term predictions of weather

in order to foster how we actually build out power grids for the future. And the difficulty here is if you just use past data, what machine learning does is it learns patterns in that past data and just projects them forward.

but the climate is changing, which means that the patterns in how weather is occurring are changing. And so you can't just use a pure data-driven technique to do this. And so what this team does is they say, well, we have climate models. The issue with climate models is that they don't give you very granular information on exactly what's going to happen at a particular place, just because they're very computationally intensive to run. But

If we can, quote unquote, back cast the climate model, so run and say, what would the climate model say now? And we already have really fine grained weather data. Historically, we can learn a mapping between what the climate model said and what the weather data would be.

And then in the future, where we only have a climate model prediction, we can use our learned relationship to say, oh, and this is what the weather would be in a more fine grained way in the future. With a lot of these kind of climate downscaling techniques, you often want to think about who is the user of these techniques and as a result, what aspects of your downscaled

predictions have to be good. So here they're actually doing this for the power grid planning context where they're saying, can we produce fine-grained data sets of what electricity usage will look like, wind power production and solar power production might look like in order to facilitate power grid planning. All of this sort of was something that you published in a paper titled Tackling Climate Change with Machine Learning. Why did you need to write this paper?

Yeah, so I'd say back in 2019, we definitely saw a combination of a lot of people in the AI and machine learning space who wanted to leverage their skills to help in facilitating climate action, but didn't necessarily know how. And on the other side, many people in the climate change related space who are seeing, you know, things like larger streams of data becoming available and saying, but how do I actually utilize this? And so we really felt like there was a need to really leverage

put forward for the community, where is it that AI is well matched to climate change related problems in order to then help AI people get into the space and to help climate people understand, okay, for some of these complex problems we're seeing, is the bottleneck potentially solvable via AI and machine learning? And there were two kind of big aims through that work. One is again, to lay out the space of applications, but the second is to try to provide some kind of

mental model and guidance for how to do this work in a sound, impactful, and responsible way. Because there are lots of places where AI is not the right fit, and it can be a huge distraction. Or there are ways that, for example, because you often have data and computational power concentrated in certain geographies versus others, where the practice of AI can exacerbate some of these inequities by basically causing people who already have access to

compute data to be able to do a lot more and leave others behind. And so there's a lot kind of in there to make sure we're actually moving the space forward in a way that makes sense for climate and for equity. Yeah. So recently we spoke to the president of Microsoft, Brad Smith, and he was talking about how he would like AI to be available to everybody. He doesn't want AI haves and AI have nots. And I interpreted that to mean that

You know, we've had technological leaps in the past and when they have been more widely available, that has been beneficial to humanity, mobile phones, internet. Do you see AI as being essential for unlocking technology?

human potential, like Microsoft president is saying. So I think there are two things I'd like to unpack in there. One, I think that AI can be a really powerful kind of support and accelerator for

for many different climate change related applications and others. There are some where I think it is essential. I, for example, don't know how we will manage power grids with lots of variability and large amounts of renewables without AI. There are other places where it can be helpful, but I don't necessarily think it's the critical bottleneck. One thing I'd like to also mention is that with certain things like mobile technologies and such,

So democratization has been used to sort of indicate, OK, a few people created a thing and it was pushed onto the rest of the world. That's not actually in some ways democratization, especially in the context of AI, where actually the type of AI you build and the way you do it.

It fundamentally needs to look very different depending on the context you're in. You have different amounts of data in different contexts. You have different amounts of compute in different contexts. You have different amounts of kind of existing knowledge that can be integrated into systems. And so if we kind of develop AI among a small set of entities and then push that onto the rest of the world, it's actually not going to serve the needs of the full world.

world. And so democratization really means enabling more people to contribute to the trajectory of actually developing AI, not just sort of being users of a product that a few people developed on the other side.

Can you give a specific way in which that might play out, say, through the development of chat GPT or cloud or these other types of generative AI products? Yes. So basically, if you think about something like GPT, it needs a huge amount of data to train. It needs a huge amount of compute to run. And

Most entities in the world do not have the ability to curate or collect that amount of data, nor do they have the ability to pay for or procure the amount of compute needed to run those models. So those models are being developed by a small set of people and then kind of packaged and sent out in a kind of interface like ChatDBT that many people can use. And that can be helpful for a certain set of use cases, but...

There are lots of use cases that don't necessarily fit that mold. Imagine that you're trying to train your own weather prediction model in a situation where you have some amount of data and also some amount of knowledge of just how kind of weather physics works.

If you do this in a fully data-driven way, there does exist a reality in which you're able to just purely from data figure out how to predict weather. But you often need much more data and a much bigger model if you're basically not embedding the rules of physics and as a result, learning them fully from scratch.

And so that leads to a situation where you, again, have a bigger model that fewer people can use and fewer people can train. And it can also lead to situations where people say, oh, I don't have a lot of data. Is the thing I'm supposed to do collect tons and tons of data? So they invest a bunch of money into setting up data infrastructure and data collection. On the other side, though, you might-

it maps kind of one-to-one to the wealth there is in the world. America and Europe is littered with weather stations, whereas Africa is empty. And so if you go down that route, the answer would be just deploy more weather stations, but it isn't the right answer. Yeah, I mean, and in some sense...

Fixing the data inequity problem is obviously a great thing. It would be great to have more weather stations in Africa than there are today. But there are kind of additional ways to contend with this problem, which include take the data you have and take some knowledge of the physical rules that govern weather, combine them together in a clever way so you don't need as much data.

to still get good answers. And so that really informs how you think about as an organization, as a country, where you invest your resources. If you're just assuming that you invest them in collecting a maximal amount of data necessary, that might actually be a misinvestment of resources if you assume that AI just means maximal data collection and learning only on data.

In addition to sort of taking in data and producing insight, there are also situations in which AI and machine learning can actually help us to more efficiently optimize a complex system in order to improve its efficiency. So, for example, if we think about buildings, right?

There are lots of ways in which we can actually better control, for example, the heating and cooling systems in buildings, both to kind of reduce the amount of energy they're actually using while kind of maintaining something like thermal comfort in the building, and also be responsive to things like how much renewable energy is actually available on the grid at this particular time, this concept of demand response. What's

Kind of interesting is when you start to think not just about individual building performance, but also how this connects up to the power grid and when renewable energy is available. You sometimes want to start thinking about this, not just at the individual building level, but for example, at the neighborhood level, where you actually might want to co-optimize what's going on in different buildings to jointly be doing the best thing for overall efficiency and the power grid.

grid. And so one of the projects we're funding through Climate Change AI is called the City Learn Challenge. And they actually created a simulation environment that actually tries to provide some structure of, okay, there are a bunch of buildings, they're connected up to a neighborhood grid in this particular way, here's some data on how they're consuming energy. And they're putting this forward as a challenge to the machine learning community to say, can you come up with better ways to actually optimize this neighborhood to improve its energy efficiency? Yeah.

Yeah, that is cool. I feel like one other thing that AI could be helping is speeding up innovation with these solutions in places where otherwise you would have required more time, more skills, more people with the skills, especially in developing countries where you really want to speed up the solution set. AI could allow for these sort of optimization techniques to come through more quickly than it would otherwise have done.

Yeah. So across the projects that we kind of are funding and kind of facilitating through Climate Change AI, they are, you know, happening all around the world. So for example, one of the projects we're funding is a team of researchers working with the government of Fiji to actually better map the damages from floods that occur in Fiji in order to facilitate Fijian disaster response efforts. The idea being that

When you actually are trying to figure out, OK, in a flood, what happened? Who was affected? It's really hard to kind of systematically and fully collect that on the ground data. And so one of the teams that we're funding is actually alongside the government of Fiji developing algorithms to kind of map from satellite imagery data.

to targeted information about what the impacts were after a flood and to be able to kind of continuously update these maps based on satellite imagery in order to aid disaster response efforts. So that's one example, but a lot of this work is going on, you know, all around the world.

And so going back to the start of the conversation where you said there's also how you can use those same tools, but to actually increase emissions, you could optimize for how you can extract oil and gas in a cheaper way or go to places that previously were not found or not reachable. Is that the biggest concern? Is that the biggest downside of AI, even more so than the resource use?

Yeah, to me, I think that we obviously need to be thinking about both the resource use and the applications. But the applications are very concerning to me because I think they're having an outsized negative impact, some of these applications, while also not being centered in the conversations about how we actually align the use of AI with climate action.

So oil and gas is one example, but there are other things like, you know, AI being the driver behind targeted advertising and increases of consumption in ways that don't always make us happier, but do increase our resource use, use efficiency.

AI also drives in many ways the information that we actually consume online. And that has really a lot of ties to the spread of climate information or misinformation in ways that could be harmful or helpful, depending on how we're actually shaping those particular trends of AI induced information spread. And then there are also things like

AI for autonomous vehicles, which we don't often talk about in the context of climate, but where the choices we're making are affecting the transportation sector in ways that could be good or bad for the climate. If you're kind of facilitating private fossil fuel transportation, then you're potentially increasing energy usage and emissions. Whereas if you're using autonomous vehicles to facilitate, you know, public multimodal transit, you're potentially bringing the emissions of the sector down. So, yeah.

I think the applications really can have an outsized impact, and it's really important to not leave them out of the conversation. And my exposure to AI actually went back a decade when I was in grad school at Oxford, and it wasn't really the models or the applications, but it was the ethics. There was a lot of conversations that were happening around the ethics of how you would put AI to use.

Do you think we're doing substantial work on the ethical side to ensure that the applications are beneficial to humanity? Or are we just in this race to develop new AI products and have kind of forgotten that there are huge ethical implications here?

So ethics is a really, really important part of the conversation. And I think there's been a lot of great work done on it, but there's a lot more that needs to be done. So you have things like UNESCO's AI ethics recommendations, which were actually adopted very widely and were really extensive in terms of thinking about things like bias, equity, privacy, transparency, environmental impact, which I would also count as a part of ethics.

And so I think there's been some really great thinking done on this, but that there's a lot more that needs to be done to sort of operationalize this and also incentivize people to actually do work in the ethical way rather than the way that kind of leaves ethics behind and just, you know, you run forward. So when we talk about AI ethics, we historically have been talking about issues like

fairness, equity, transparency, privacy, and so forth. Or friendly AI that we shouldn't create something that would then want to try and destroy humanity. And that's the kind of part that has come kind of into the conversation really recently, this idea of, you know, AI existential risk, AI existential threat, and so forth. And

I would say that that's not an unimportant part of the conversation. We really should be thinking about the full range of risks that AI can pose in addressing them. But it's become maybe an outsized part of the conversation. We should think about AI ethics holistically and make sure that we're not letting kind of one particular sub part of AI ethics dominate the conversation at the expense of really thinking about the rest of AI ethics as well.

Really, there's a huge need to democratize literacy skills and expertise on AI so that more people are able to engage in a way that is kind of informed by knowledge of the strengths, limitations, risks associated with technology. And so I think really enabling more people to participate by having that literacy skills and expertise is really, I think, the huge thing that we need to achieve at the moment.

I did enjoy this conversation a lot. Thank you. Thanks so much. Thank you for listening to Xero. If you liked this episode, please take a moment to rate or review the show on Apple Podcasts and Spotify. Share this episode with a friend or with someone who fears our robot overlords. You can get in touch at xeropod at bloomberg.net. Xero's producer is Maithili Rao. Our theme music is composed by Wonderly. Special thanks to Keira Bindram and Alicia Clanton. I am Akshat Rati. Back soon.