Unknown

So I think the one narrative of like, here's a hard problem, like AI is going to just provide an end-to-end solution to it, I think is often overly optimistic and I think ignores a lot of the realities of deploying these solutions. The other fallacy is if we just kind of advance AI generically, if we just like scale generic AI, then somehow like scaling large generative language models is now going to kind of solve everything in the power sector. And we really do see that these models don't kind of meet the requirements that you actually have on power grids and they're not actually always the right fit for the problem. From Latitude Media, this is Open Circuit. I'm Stephen Lacey. In the last few years, we've flipped from asking, can AI help us solve climate change? And now we're asking, can AI help us solve climate change? And now we're asking, can AI change? To how much worse will AI make climate change? And even more alarming, can we even keep AI under control? It's easy to understand why enthusiasm swung so radically. A few years ago, we were promised rapid breakthroughs in materials, a business model revolution, and a radically more efficient energy economy. And all those things may very well happen, but many people feel like we got hulking gas plants, higher electricity prices, and broken emissions promises instead. The real story is nuanced. AI is working on a new technology, and it's working on a new technology. And it's working its way into every corner of the clean energy business, across forecasting, product design, resource dispatch, and more. And there are still researchers around the world who are advancing these tools for the benefit of clean solutions. So what is the real story? Few people have thought harder about both sides of this than Priya Danti. Priya is an MIT professor and co-founder of the nonprofit Climate Change AI. She's a researcher focused on machine learning for optimizing the power system. And back in 2019, she co-wrote a paper with a group of leading AI researchers on how the technology can be used for climate. And they explicitly warned not to treat it as a silver bullet. So now that we're in a totally different AI era, I sat down with Priya to talk about where AI is delivering, why claims got so inflated, the backlash, AI risk, and how the AI for good movement has evolved. ♪ ♪ OpenCircuit is brought to you by Energy Hub, the flexibility platform for the modern grid. More than 230 utilities trust Energy Hub to build and operate virtual power plants, delivering 3.6 gigawatts of flexible capacity. As the resources grow, okay, this is really great. This is an amazing resource. I can use it. It's really valuable. But here's what I really need to do to make it as good as or better than my existing conventional capacity. Paul Hines is the chief scientist at Energy Hub, where his team uses algorithms that turn millions of DERs into flexible capacity utilities can stake their grid planning on. My team is working on the algorithms that allow us to meet the utilities' goals for shaping these virtual power plants. Machine learning keeps raising the ceiling on what VPPs can deliver. Energy Hub's platform is trained on years of data from millions of devices, optimized for utility needs across every season and every grid condition. We're working with utilities not to just enable that capacity, but also make it more flexible. So that utilities can use it, you know, more times per year for longer events in ways that are valuable to both customer and utility. Later in the show, Paul explains how Energy Hub's platform helps utilities scale VPPs for their unique needs as grid demands evolve. Learn more at energyhub.com. OpenCircuit is supported by Everyday Electric. You might know them by their former name, Renew Home. We are at a pivotal moment. Energy demand is growing faster than the economy. We're at a pivotal moment. farm somewhere, you actually want to know what the wind conditions might look like that time in the future. And then now in the power sector, there's a bigger conversation now even than then about kind of power system dynamics. So when you have more renewables and inverter-based resources, are you seeing kind of difficult subsynchronous oscillations on the system or kind of your inverters doing kind of weird dynamical things that have implications for the grid? And of course that, you know, when we had the Spain blackout recently, that kind of became a big point of conversation. And there has been some machine learning proposed for things like, you know, detection, control, and modeling in this context, but we're still in the early days even now. And then the last thing I would say is that now I think in the modern conversation, there's a lot more about where AI and machine learning are maybe providing kind of better user interfaces or interaction with existing tools. So you obviously hear today a lot about it. agentic systems and trying to do things like stitch together. You have kind of analysis tool one that outputs an output and you need to massage that output before putting it into another analysis tool. In an ideal world, we would just develop nice power system software that has really good data formats and APIs where these softwares could talk to each other more natively. But I think in filling that gap, we've definitely seen at least now this kind of agentic systems and kind of user experience style AI that we definitely didn't talk about back in 2019. And over the years, I've been really paying attention to how this gets adopted in a business context, right? And so when you bring these technologies out into the real world, there's so much friction in the energy systems. In the energy system, it comes down to how organizations operate, comfort level, change of culture. Again, there's a lot of risks in the power system and also risks around data sharing. There's just so much friction in the system. Has the pace of adoption met or fallen behind your expectations over the last six or seven years? I think it's been about as expected in the sense that I think these like forecasting, monitoring, and predictive maintenance style applications are definitely getting, you know, are very maturely deployed. And I think there's more to do in the sense that I think data sharing between entities becomes a bit of a bottleneck sometimes for like how high of quality the data is. And what volume of data you can put into your monitoring technology if you have these silos between different utilities and different industries. But I think barring that, we are seeing those deployments. Whereas for grid optimization, things in sort of safety critical settings, that has been much slower. And I think that was anticipated in the sense that there's a lot to sort out, not just on the technology side, but also as you're mentioning in the deployment context, kind of when you are kind of putting a new technology into kind of a, an, um, a fast moving or automated system. And you know that even in our current context, the answer that your algorithm puts out may be imperfect because maybe the input it got about what the state of the grid looked like is imperfect, or maybe the algorithm itself didn't output something properly. What do you do and what system do you build around your algorithm to make sure that the overall system is behaving as desired? I think that conversation is longer running and you sort of see it reflected then in the adoption curve of AI algorithms. In that context, you were really careful in this 2019 paper to, uh, make sure that people understand that AI machine learning is not a silver bullet for climate solutions. When this gets out in the world, the narratives get shaped. I think a lot of people do make pretty grandiose claims about how AI is going to be a climate savior. I'm curious about how much you think the hype has actually outrun the, uh, technical capabilities. And where have you seen the most hype? I, I continue to think that AI can be a very kind of helpful and like useful accelerator of climate and energy solutions. But I do think a lot of the hype tends to center around kind of two specific narratives. One of which seems to suggest that somehow AI is an end-to-end solution to these problems. Like AI will, because it's solve fusion, solve grids. And these are complex systems where kind of AI algorithms are usually playing like a particular kind of role. And I think that's a really good piece of that. And you then have to integrate that piece into a broader system. Like even if you have, um, an AI algorithm for forecasting, you still need to figure out how you're going to use that forecast. Or if you have an AI algorithm that allows you to do scenario generation of kind of extreme event, extreme weather, you might see because of shifting climatic conditions, somebody in the grid operator has to believe and care enough to plan against those conditions in order to do something. And I think that's something. So I think the one narrative of like, here's a hard problem. Like AI is going to just provide an end-to-end solution to it, I think is often overly optimistic. And I think ignores a lot of the realities of deploying these solutions. The other, I think is that if we just, the other fallacy is, um, if we just kind of advance AI generically, if we just like scale generic AI, then somehow like all of our problems are solved, like kind of scaling, large generative language models is now going to kind of solve everything in the power sector. And we really do see that these models don't kind of meet the requirements that you actually have on power grids. And they're not actually always the right fit for the problem. So for example, now this is zooming out to kind of AI for climate change as a whole, looking at the kind of publications in the climate change AI workshop series that we hold at the machine learning conferences, like over the kind of, you know, years that we've held them over the seven years, only like six to 7% of those submissions are in the area of natural language processing, which arguably is like a reflection of only a small slice of problems in these sectors are really NLP problems that then like a kind of a large language model style solution in natural language processing would be well-suited to address. So what about the rest of the problems that are not those? I think these two kind of types of hype of like end-to-end solutions and somehow like scaling generic technology has kind of all of these benefits for the power grid are basically like very overhyped. Whereas I continue to believe in, I guess, approaches that really start from the power grid problem. Like we have a hard problem, like how do we optimize a grid at scale with a lot of renewables? We have problems like how do we think about uncertainty over long planning horizons? And when we, we kind of develop methods that are really informed by the problem, I think you start to both get beyond the hype and really get at the heart of what's hard, both from the methods perspective and from the deployment perspective. Yeah, that's so interesting. And I'm kind of curious about what you're seeing in large language models in general purpose AI. I think it was back in 2017 when Dario Amadei wrote this piece called Big Blob of Compute. And it was really foundational to how a lot of the AI labs think about scaling intelligence. They were, you know, focused, you know, open AI initially was focused on more specific models. And then they realized that, like, you should just throw a bunch of compute at it, focus on a generalized model, and it will get smarter about everything else. And that will lead you to super intelligence, and it will be good at mathematics and scientific research and language and whatever. And so we see, we have seen these models improve greatly. But it sounds like what you're seeing is that the specialized models are still much better at these tasks in, say, the power sector. So do you see, like, more general purpose AI starting to have a greater impact in this space? It's a great question. And I think also here, there's, like, maybe also even, like, two different definitions of general purpose AI I might kind of pose. One of which is, like, I think what the frontier labs are shooting for, which is, like, AI that can be used, like, in any context, like, kind of do anything. And there, I think you can see, these models have undoubtedly gotten much more powerful. But you can also see some steering in the sense of, like, when these models have been developed, there have been, like, some domains that maybe it was thought they would be good at that haven't happened. And yet things like coding and mathematical proofs, they sort of have seen a lot of progress since you sort of see, okay, now we're going to focus on that. So even this notion of these models are going to be good at everything, I think they, we don't know what we don't know. It's possible they'll also be, like, really good in domains that they haven't panned out in yet. But there's been a little bit of, like, a moving. sort of application target a little bit in terms of kind of where the focus has been based on what has actually been good. Whereas I would say the other definition of general purpose I would give is, like, imagine that you're in the context of, like, power grid applications, but you want to understand whether, like, can I develop a model that can do power flow and optimal power flow, like multiple types of power grid tasks, or that can work across grid A, grid B, and grid C, without having to be retrained across all of those different grids. There, I think there's definitely been kind of effort to kind of create what have been called, like, grid foundation models, for example. And I think it's not clear yet kind of which paradigm of, like, kind of train a more, like, general purpose model across different power grid tasks and grids versus train task-specific models for these, for an individual task or an individual grid. It's not clear yet kind of which one will win out. I would say right now, I think it's going to be a little bit more difficult to right now, training a task-specific model for a particular task on a specific grid, you're going to get the best model for that particular task. Like, a general purpose model, even in the context of power grids, doesn't necessarily zero-shot general. visualize well to that particular task in context. And even when you fine-tune these models, they don't always reach the performance of whether when you kind of just trained a specific model for that task. That may change, but I think in the context of general purpose, if I had to say like maybe there's one general purpose direction that I think could maybe win out a little bit on power grids, but it's a bit unclear. I think it's just like general purpose models for power grids. But again, I think it's nuanced in terms of when those will be better versus task-specific ones. I think we don't know. Earlier, we met Paul Hines, Energy Hub's chief scientist. His team develops a machine learning that orchestrates 3.6 gigawatts of virtual power plants. When you see a utility using algorithms that we wrote on the whiteboard running 250 megawatt events, I think that's pretty amazing. Utilities need capacity they can deploy now. And they're finding it in virtual power plants, which are precise and configurable. I've got customers that have time of use rates. We need to make sure we respect that. I need to pre-cool during these hours. Maybe there's some hours that have a lot of solar power, so let's make sure that we soak up as much of that solar as we can. But I need to really manage my peak during this other period of time. And so figuring out how do we take all those goals and then come up with schedules for all the resources, whether it's batteries, thermostats, electric vehicles. Energy Hub maps that evolution through a four-level, maturity curve. Level one looks like traditional demand response. At levels three and four, a VPP is a real-time resource that can meet many different needs on the system. It's actually better than a power plant because you're able to solve problems in the distribution network, which is where a lot of the cost is right now. You're able to solve problems in an automated way, whereas power plants actually are fairly slow to respond to issues. It's the future of how the grid operates. And we just need to connect them in the right ways, give them the right algorithms. Energy Hub's GridAware DERMS is the flexibility platform trusted by 230 utilities across North America. Open Circuit is brought to you by Everyday Electric. After years proving the market was ready, they're delivering a new kind of VPP. So how does it work? Starting with features built into the products people already own and use. Things like smart thermostats, electric vehicles, and home batteries, adjust automatically based on everyday routines, helping lower bills and add capacity. People can get meaningful savings and earn rewards from the grid. And this contributes to a reliable, resilient system that lowers costs for everyone across the grid. That is Everyday Electric's vision. It's the next great addition to your program portfolio. You can learn more at everydayelectric.com. What if you gave your climate donations the same due diligence you give your investments? That's what Giving Green does. Giving Green researches the strategies and non-profits best positioned to unlock the next climate breakthrough. Climate impact is difficult to quantify, but Giving Green's research suggests a donation to one of their top climate non-profits is roughly 10 times more effective than a high-quality carbon offset. And the Giving Green fund sends 100% of every dollar straight to them. Maximized impact per dollar, no cuts or guesswork. Make a first gift before the new year, and it will be matched up to $500. Go to givinggreen.earth, slash, circuit, to claim it. I want to talk a little bit about the consequences of this technology out in the world. We're obviously at a very different place within the AI safety conversation right now. This summer, we saw a lot of big stories coming out of the AI labs, the most prominent one being agents breaking out of OpenAI's sandbox and hacking. A lot of websites, they since revealed that there were hundreds of other hacks that they didn't catch while training the model. You have talked a lot about consequences for the power sector. In the past, you and I chatted on stage at our Transition AI conference, and you expressed concern that if something goes wrong on the power sector, we won't be able to interpret the thinking within the model, and you might not know who to hold responsible, for example. So as we get more concerned about the safety of AI, we're going to have to be more careful about the safety of AI. In terms of large language models, are there any scenarios that you're worried about in the power sector within the models that you're working on? Could we see a similar problem, just in a different system? And I think there are a couple of concepts that it almost makes sense to untangle here. So there's kind of interpretability and explainability. So can you understand the mechanisms that led to your output? And then can you explain post-talk why a decision was made? And then kind of in parallel to that, can you explain the mechanisms that led to your you have like safety, verifiability, and alignment? So can you guarantee a particular behavior or provably verify whether it happened or not, or kind of align models to do the things that you kind of want them to do? And I think these kind of play slightly different roles, though they're both important. Like in the context of the open AI hack, for example, some might argue that the models have explainability mechanisms built into them in the sense that you can see the like language the kind of language outputs of these models as they kind of reason through what they're supposed to do. But you still had the break, right? And so I think there there was like an alignment issue, right? Like find a solution to this exploit gym benchmark and, you know, not realizing that doesn't mean don't break laws and cybersecurity protocols to do it. And also like definitely a lack of appropriate like cybersecurity protocols that were in place. Like there wasn't air gapping, there weren't tripwires, there weren't all of these kinds of things. And so in the context of power systems, I think I'm less worried about this in the context of like if we train a specialized model for or like a general purpose model even, but if we train a model for a particular task, like, you know, do a demand forecast given this input, given this output, like here's a demand forecast or learn a surrogate, an approximation for AC optimal power flow. Like here are the inputs to AC optimal power flow. Here are the outputs to AC optimal power flow. There we're specifying like a very specific input output behavior and asking the model to learn like a very specific function. And so there are some of these risks around like alignment don't come into play as much because you're asking the model to do something like quite specific. And in the context of power grids, I think that's where we are likely to go. We actually want kind of individual components where we kind of know what their function is. We know how to verify the behavior and we kind of know what the desired input output mappings look like. Whereas this idea of like kind of getting a generic agent onto the grid and saying like, oh, optimize it. Like we wouldn't want to do that. Right. I mean, and there you could see risk of misalignment. Like what's the best way to minimize the cost of power generation? It's to shed all your load. Right. So like I but I don't necessarily see that as actually being a realistic scenario for how we would actually deploy AI agents into power grids. I think I really see that as a more specific like forecast demand, kind of learn about approximate solution to ACO-PF in places where then we can kind of also use these more targeted notions of verifiability and interpretability where we're like really inspecting the internals of the models and maybe like building mathematical specifications for what we expect an output to look like. Yeah, that's really helpful for you to distinguish between interpretability and alignment. Have you heard anybody like at the labs talking about using these general purpose models to run the power grid, for example, or other critical systems? Like, is that a line of thinking within that world of AI? As far as I know, no. I think even the kind of folks at the kind of, you know, tech companies I named earlier that maybe are working on these these topics, it feels like it's much more these like learn an approximation for optimal power flow, like learn a demand forecast, like much less like throw an agent at the grid. And even in the context where we talk about like agentic systems for power grids, it is much more like, can we chain together analysis tool A and analysis tool B to better talk to each other so you don't have to manually reformat inputs every now and again? And so I think on the side of people who are developing AI for power grids, kind of regardless of the context, I definitely see it being much more that like there's a particular problem that we're trying to somehow approximate or move more quickly on. There have, however, been — oh my gosh, I'm going to forget the name of this initiative — but there was an initiative kind of coming out of the Frontier Labs in the, I think, the wake of the Hugging Face attack, which basically said, you know, should we actually be giving access to more industries, some of these frontier models, so they can use it for the purposes of cyber defense? And I think that my understanding is that the power sector was used as an example. And I think that the power sector was used as an example, at least in that context, of like, if somebody is trying to impose a generic cyber attack on the power grid, are there ways to kind of make sure that power companies have access to these frontier models for the purposes of cyber defense? But I think it's much more about these generic worries about cyber threats and maybe that kind of thing becoming easier with some of these models rather than someone who's working on the power grid suggesting that we deploy very general-purpose agents onto these to do kind of whatever they want. How have you evolved on safety generally? As I can tell, tell you that over, you know, over the last five months or so, I have become more freaked out. And I wouldn't say that I am someone who is really concerned about extinction risk, but to see that the safety mechanisms are breaking down inside these companies, we're in this rapid commercial arms race where there's no incentive to truly slow down. Are you more worried about risk? And where do you fall on the risk spectrum? Yeah, so I likewise, I'm sort of not like a kind of existential risk worried kind of person. I definitely still kind of worry about some of the things I kind of have been worrying about for a while, like bias and fairness and privacy violations, and even just things like AI accelerates the systems we deployed. And so if we develop it in kind of bias systems or systems that have unintended harm, that gets accelerated. I am kind of more worried about cybersecurity attacks. And I think it's interesting to follow the public narratives on these, right? You almost have one side that's a bit like, oh, like, these models are really powerful. And so all you have to do is align them in order to like, prevent this. And kind of others being like, there are also just like common sense cybersecurity things that should have happened, like air gapping, like, when you see a bug in the system, or you see a bunch of files being generated, you don't just go, oh, that's weird, delete them and kind of move on. Right? Like you kind of have some protocols in place to inspect these. And so I think this combination of things and like, as you're saying, this like arms race we're in, I think there's a lot less care around things like cybersecurity than I think there should be. And I do worry about it's like, almost less the technology itself, but like the human systems and the culture we'll be building around these systems to not necessarily be careful in the in the ways that we should be. And I think another big thing that is definitely on my mind is, we're definitely seeing the way that these sort of AI systems, or general purpose AI systems are being developed, I think is leading to worries of a present and a future where we're doing a lot of cognitive offloading. Like there was something really funny I saw on Twitter the other day where someone was like, man, like I wanted AI to like, you know, help me unload my dishwasher so that I could do more spend more time on math proofs. And it looks like we're going the other way. Totally. And yeah, and I think that we're seeing these implications in education where like, it's so much, easier to like GPT your way through like an assignment and to like reduce friction. But friction is what helps you learn. It's like what helps you become more capable. And I struggle that I think, and there are all these studies that show like, that are starting to show that the kind of when you have like people with already really deep expertise use things like large language models to accelerate progress on tasks that they become more effective. But then you also have like a whole set of people who become less effective, who become less effective, who become less effective, who become less effective, or like less knowledgeable, because maybe they didn't have the baseline expertise. And yeah, I basically worry about these systemic effects where we're making ourselves like much less expert by virtue of offloading cognitively a lot of things to these tools. And then as a result, like productivity games that we're seeing from people who build expertise before these LLMs existed, like those won't exist in the same way in the future. So I think all of these things about what it means to think what it means to be human, what it means for society to be intelligent, like that actually is on my mind. quite a lot these days as well. I want to go back to this commercial race that we're in the middle of right now. You know, we're building these models, and the infrastructure at a ridiculous pace. What worries you about the trajectory we are on, either with the infrastructure or with the models themselves? Yeah, so I think there's there's two big things I'm worried about. One of them is consolidation. So I think as we kind of push towards this world where we're kind of being convinced through public narrative, and also through where investment is going, that we should have kind of these bigger models, which fewer and fewer entities have the ability to actually even develop, that fewer and fewer entities are controlling. And then these narratives around it's so dangerous that only we should do it, right? I think that really leads to this consolidation, where you get, you're seeing these like advances on things like coding and math, right? You see these advances on this particular set of things. But like, not all problems across society are being served by those narrow advances. And so there can be this like, almost like, pull away of funding and capacity from making like more problem focused progress. Like if we were to be like climate and energy are the most important thing, what would we build? And consolidation means that there's like less heterogeneity in what people are exploring in this respect. And also, it's it's sort of, it's generally like antithetical to, you know, democratization, and kind of this idea that really, like, in order to make progress on like diverse climate and energy problems across the world, where there are different contracts, context, different data realities, etc. We need to be kind of equipping more people with knowledge, we need to be creating more high quality data in various places, because both task specific and general purpose models don't do well in situations that aren't well represented in in their data. So there, I think, starts to be this like investment away from building knowledge from building ground truth data, from empowering more people. And I think the result a little bit is that you actually get like, crappier solutions for things that you really needed very good solutions for. Because your general purpose model might be kind of good at many things, but it's not really good at the one thing you really needed to be really good at, unless we've chosen to focus our resources there. So I'm really worried about Yeah, essentially, that that capture of values that kind of homogenous thinking, homogeneity of where the field goes in a world where we have so many heterogeneous problems to solve. And then the other thing along with that is I think that I worry about maybe the extractive nature of this. So for example, like, with some of these, like AR for math results that we're seeing where there are lots of proofs going out into into the world, I think like, there's definitely a lot of debate and tension around, you know, have some of these results been stolen? Or, you know, have some of these results been stolen? Are they attributed properly? Are they actually correct? But even in a world where like, those are gone, this question of like, you need generally, like people who are mathematicians to articulate, like, what are the interesting problems to solve, and also on the other side, be able to like, check whether problems are good. And in this world where like, a lot of AI for math results are going on, there's a lot less motivation to be that person who like, dedicates a bunch of your time and education to solving problems. And so I think that's a really good question. to like, getting a PhD in math, to be then your job be like, check the outputs of an LLM. So I think there's these like, almost like, this extractive nature where it's like, we do this one thing, and then like the whole ecosystem around it collapses in a way that that ecosystem then doesn't exist anymore. And I just, I'm not sure that we're really cognizant of like, some of the systemic implications of this, like very fast, like narrowly focused progress either. Yeah, it's interesting. And it's very relevant this morning. I mean, OpenAI just really, at least a bunch of examples of the models making huge progress on millennium math problems. And some mathematicians are saying that like, this would, if a person did this, it would be Nobel Prize winning work. And I know that the mathematics field is very split on this. So very interesting. What about the data center build out? What do you worry about there? I mean, I'm sure that you hear a lot about this when you say you work in AI, everyone's focused on data centers. What are your worries about the cleanliness of this industry, about how we're doing data center build out? Is this something that's on the mind of AI researchers like yourself? Absolutely. And I think like, yeah, generally, right, like, of course, building like AI and data center infrastructure comes with its sort of energy use implications, it comes with hardware and kind of minerals extraction implications, water implications. And I think the challenge here, is that the effects are often very local, right? Like when you have kind of, you know, water use in water scarce regions, it's the people in that region who really suffer, or when you see changes in electricity prices, you see people in that region really suffer. And this brings, I think, in some sense, a debate that's not new around climate justice and climate equity, and basically how we as a society have screwed over various individual over the years in the name of like macro level progress. And that does not sit well with me. I think this is something that is really not great. And I think we really need to contend with, in addition to the fact that I think the macro level progress, and I think the narratives around the macro level progress are also disingenuous sometimes. So yeah, again, this fallacy of like, if we just grow data centers and train generic AI general purpose models, like climate will be solved we've talked about, I don't buy that claim or believe in that claim. And I think the types of AI you build, like, AI is not a silver bullet, as we've talked about. But even in the context of climate where you use AI, the type of AI you build often looks like very different. And then the type of AI that's generic, and even just to give a given example, the so the kind of energy use of a model is kind of correlated, right with the number of parameters within that model that you're during training and kind of using during the inference and use of that model. And some of the like, biggest like AI for climate models that exist are models that kind of approximate climate models that try to provide like a fast approximation to your climate model. And those models are usually like 10 million parameters, maybe 100 million parameters at most. They're not like a trillion parameters as kind of the next generation of kind of general purpose frontier models is proposed to be. And that's because the size of the model is often proportional to the size of data you're putting in. Like the amount of climate data and climate modeling data we have is nowhere near the size of the data on the internet. So I think also there's this conflation around like, oh, but like general purpose stuff can be good, but it's not the same general purpose stuff. It doesn't have the same model size. And so it's like this combination of like basically kind of leaving certain communities behind and then being really disingenuous about what the impacts are, like the positive gains on the other side, like all of that does not sit well with me. I'm curious about how this plays out within the AI research community. I think we have this understanding of sort of two sides of this, the accelerationists and the doomers, right? That's how it's often portrayed. You're within this community of folks who are working on AI for good, for solving lots of big, hairy problems for humanity and for the planet. How do you see the spectrum of how AI is viewed within the community? And how has the AI, for good community evolved in this moment when it's, you know, people are really, really hostile to AI or people just think that we should hit the accelerator no matter what? Yeah. And of course the AI for good community itself is very heterogeneous. So this is also like in some sense hard to answer as a blanket form, like as in any community, you definitely see some people basically kind of running it and being like, okay, well, these tools are being developed. Like how can we best leverage them for, you know, for, you know, for, you know, for, you know, for good things and others that I think maybe are worried about some of the kind of systemic implications of not just the technologies themselves, but also maybe the misleading mental models that somehow the way the narrative is shaped out have led to. So because there is so much like when you now kind of run into a generic person and you talk about AI, they think AI is this point about like heterogeneous ways of approaching the technology in order to serve different applications. I think that causes like investments to go in the wrong places. I think it leads to some amount of kind of greenwashing, like can we work on AI and then justify it via climate benefit, which is like not the right viewpoint in terms of like if you care about climate, it's more about like how do we align society's overall activities around like a climate positive future. So I think basically it just becomes a bit difficult to navigate kind of a landscape where you're having to disentangle so many mental models that are driving where money goes, that are driving how people perceive your work. So I think it's just, it's complicated. It's become complicated to navigate the space, certainly. And where have you landed? Do you still think that AI can be a net positive for the environment? So actually, it's funny that you say still there too. So in some sense, I think, I think AI can be a net positive for the environment. I think AI can be a positive for the climate. I think AI could be a net positive for the climate if we align our overall societal activities to be net positive for the climate. But fundamentally, AI is an accelerator of the systems it is enacted in. So if we as a society are still in like a primarily fossil fuel and polluting kind of system, AI is going to disproportionately accelerate that and be net negative for the climate. So I think I continue to basically think it's like it's sort of not about the AI. Exactly here, like the, there was this really nice analysis by Alpine et al in NPJ Climate Action, where they actually, they modeled AI as like basically providing productivity shocks within an energy economy model where you're basically like, okay, imagine that AI translates to like more productivity for renewables or more productivity for oil and gas. And they basically found that like, even if you had a lot of progress on AI for renewables, if you had even a modest amount of productivity shock towards AI for oil and gas, then the net impacts were negative in the sense that there were more emissions being produced, which kind of emphasizes this point that it's like, if we keep using AI to accelerate not great things, then of course, the impact of AI will be net emitting. Yes. And that acceleration in oil and gas extraction with the use of AI is definitely happening. It's real, right? It's like huge. And so I think it really comes down to both. Like, are we really serious as a society about like aligning our activities overall towards climate goals that includes AI? And I think that the answer to that question is sort of what dominates like whether AI is net positive or not. Well, it's really good to talk to you, Priya. You're one of the great minds in this space. And thanks for your contributions. Really great to talk to you. Likewise. Really great to talk to you too, Stephen. Thanks for having me. So if you're liking the show, be sure to subscribe. Thanks a lot for being here. I'm Stephen Lacey. We'll catch you next week.