1 00:00:00,640 --> 00:00:04,560 Speaker 1: Welcome to Zero. I am Akshatrati. This week, what good 2 00:00:04,760 --> 00:00:20,040 Speaker 1: can AI do? Remember when data was the new oil? 3 00:00:20,880 --> 00:00:23,639 Speaker 1: I'm specifically thinking of a twenty seventeen cover of The 4 00:00:23,680 --> 00:00:27,920 Speaker 1: Economist showing Google, Amazon, and other tech giants as big 5 00:00:28,000 --> 00:00:31,920 Speaker 1: offshore oil rigs. The idea being that data was a 6 00:00:31,960 --> 00:00:35,559 Speaker 1: new critical resource and it was going to reshape the 7 00:00:35,600 --> 00:00:41,280 Speaker 1: world in some ways. That has already happened. Maybe this 8 00:00:41,360 --> 00:00:44,760 Speaker 1: podcast was suggested to you by Spotify or Apple based 9 00:00:44,800 --> 00:00:48,000 Speaker 1: on your listening history. Just a small example of big 10 00:00:48,120 --> 00:00:53,440 Speaker 1: data at work. Artificial intelligence, the latest buzzword, of course, 11 00:00:53,600 --> 00:00:58,000 Speaker 1: thrives on data. That devouring of data is energy and 12 00:00:58,120 --> 00:01:01,880 Speaker 1: resource intensive. It's something we discussed in last week's episode 13 00:01:01,960 --> 00:01:05,600 Speaker 1: with Microsoft president Bradsmith. The company wants to be cover 14 00:01:05,720 --> 00:01:11,759 Speaker 1: negative but is instead seeing its emissions grow. But of course, 15 00:01:12,040 --> 00:01:16,040 Speaker 1: fed the right data, AI can do amazing things, even 16 00:01:16,120 --> 00:01:21,760 Speaker 1: help tackle climate change. But how exactly. If there's one 17 00:01:21,800 --> 00:01:25,920 Speaker 1: person taking the lead on that question, it's Mits Priadante. 18 00:01:26,400 --> 00:01:29,679 Speaker 1: She's a professor of electrical engineering and AI and the 19 00:01:29,720 --> 00:01:33,720 Speaker 1: co founder of Climate Change AI, an organization bringing together 20 00:01:33,840 --> 00:01:38,120 Speaker 1: academics and industry leaders interested in how AI can be 21 00:01:38,240 --> 00:01:42,880 Speaker 1: used for climate solutions. Her group funds independent projects and 22 00:01:42,959 --> 00:01:47,720 Speaker 1: fieldwork tackling everything from mangrove restoration for Indonesian shrimp farmers 23 00:01:48,160 --> 00:01:51,800 Speaker 1: to the study of nanoporous separations in the chemical industry, 24 00:01:52,920 --> 00:01:56,360 Speaker 1: and it also thinks hard about how to avoid AI 25 00:01:56,640 --> 00:02:01,720 Speaker 1: being used to increase emissions and worsen humans suffering. I 26 00:02:01,760 --> 00:02:04,440 Speaker 1: asked Pria about some of the AI applications he's most 27 00:02:04,480 --> 00:02:08,560 Speaker 1: excited about, and why the conceptual framework we build around 28 00:02:08,600 --> 00:02:24,560 Speaker 1: AI is just as important as the technology itself. Now, 29 00:02:24,600 --> 00:02:26,440 Speaker 1: before we get into the heart of some of the 30 00:02:26,480 --> 00:02:29,640 Speaker 1: work you do at climate change AI, I think it 31 00:02:29,639 --> 00:02:33,680 Speaker 1: would be helpful to define the terms because there's just 32 00:02:33,840 --> 00:02:36,400 Speaker 1: so many of them, and there's just a mixing and 33 00:02:36,520 --> 00:02:40,600 Speaker 1: muddling when people think about AI. For most people, the 34 00:02:40,639 --> 00:02:45,320 Speaker 1: biggest point of entry for AI is chad GPT. Chad 35 00:02:45,360 --> 00:02:49,080 Speaker 1: REPT is what people have played with. People kind of 36 00:02:49,200 --> 00:02:51,720 Speaker 1: know it's based on this thing called an LM, a 37 00:02:51,880 --> 00:02:57,000 Speaker 1: large language model, But it's just one example of types 38 00:02:57,080 --> 00:02:59,680 Speaker 1: of AI. So if you start at the very top, 39 00:03:00,120 --> 00:03:04,080 Speaker 1: would you define AI and what types of AI are there? 40 00:03:04,520 --> 00:03:07,680 Speaker 2: Yeah, so there isn't kind of one universally agreed upon 41 00:03:07,840 --> 00:03:11,600 Speaker 2: definition for AI, but roughly You can think about AI 42 00:03:12,000 --> 00:03:15,760 Speaker 2: as referring to systems that perform some kind of complex task. 43 00:03:16,320 --> 00:03:20,320 Speaker 2: And there are two big branches of AI. One is 44 00:03:20,520 --> 00:03:24,720 Speaker 2: rule based systems, which is when you kind of know 45 00:03:25,160 --> 00:03:27,760 Speaker 2: how to do something, like you know how to play chess. 46 00:03:27,800 --> 00:03:30,040 Speaker 2: In some sense, you could write down the rules, but 47 00:03:30,120 --> 00:03:32,480 Speaker 2: actually reasoning over those rules to figure out how to 48 00:03:32,480 --> 00:03:35,080 Speaker 2: be a good chess player is the hard part. And 49 00:03:35,160 --> 00:03:37,360 Speaker 2: so rule based systems are places where you write down 50 00:03:37,400 --> 00:03:39,680 Speaker 2: the rules and reason over them automatically. 51 00:03:40,280 --> 00:03:44,720 Speaker 1: Oh and that means Deep Blue beating Gary Kasperov for 52 00:03:44,760 --> 00:03:47,640 Speaker 1: the first time in nineteen ninety seven. That would be 53 00:03:47,720 --> 00:03:51,839 Speaker 1: classified technically as a rule based AI system. That's right, 54 00:03:52,200 --> 00:03:54,000 Speaker 1: even with those primitive computers. 55 00:03:54,080 --> 00:03:57,160 Speaker 2: Oh yes, so AI has been around for a long 56 00:03:57,240 --> 00:04:02,400 Speaker 2: time actually, And so another type of AI is machine learning. 57 00:04:02,760 --> 00:04:03,960 Speaker 3: And machine learning is. 58 00:04:03,920 --> 00:04:08,840 Speaker 2: Often used in situations where you might have intuition for something, 59 00:04:08,920 --> 00:04:11,480 Speaker 2: but it's really hard to write down rules to codify 60 00:04:11,560 --> 00:04:14,440 Speaker 2: your intuition. So if I gave you actually an image 61 00:04:14,800 --> 00:04:17,640 Speaker 2: of a dog, you could probably tell me that it's 62 00:04:17,640 --> 00:04:19,680 Speaker 2: a dog. But if I asked you to write down 63 00:04:19,680 --> 00:04:22,680 Speaker 2: a set of rules that characterize exactly why it's a dog, 64 00:04:22,760 --> 00:04:23,960 Speaker 2: and it would be really hard for. 65 00:04:23,920 --> 00:04:25,880 Speaker 3: You to write down that set of rules exactly. 66 00:04:26,480 --> 00:04:29,520 Speaker 2: And so machine learning is a paradigm where you actually 67 00:04:30,080 --> 00:04:34,640 Speaker 2: infer some of these rules automatically from examples or data. 68 00:04:34,680 --> 00:04:37,080 Speaker 2: So I give you a bunch of images, maybe I 69 00:04:37,160 --> 00:04:39,840 Speaker 2: tell you which ones are dogs or cats, and the 70 00:04:39,960 --> 00:04:42,960 Speaker 2: machine learning algorithm learns how to map between the images 71 00:04:43,000 --> 00:04:44,839 Speaker 2: and the labels of is this a dog or cat 72 00:04:44,880 --> 00:04:47,000 Speaker 2: and kind of infer the rules that cause that to 73 00:04:47,040 --> 00:04:47,640 Speaker 2: be true. 74 00:04:49,240 --> 00:04:52,120 Speaker 1: And so if it take the type of AI that 75 00:04:52,160 --> 00:04:56,920 Speaker 1: most people know, which is large language models, that's machine learning. 76 00:04:57,160 --> 00:05:01,040 Speaker 2: That is machine learning and large language model are basically 77 00:05:02,160 --> 00:05:07,080 Speaker 2: one type of machine learning model that basically looks a 78 00:05:07,160 --> 00:05:11,080 Speaker 2: specific way, has a particular specification of how you update it, 79 00:05:11,520 --> 00:05:14,440 Speaker 2: and that type of model can be used in various 80 00:05:14,440 --> 00:05:17,640 Speaker 2: different ways, and roughly the three kinds of ways they 81 00:05:17,640 --> 00:05:21,120 Speaker 2: can be used are called supervised learning, unsupervised learning, and 82 00:05:21,160 --> 00:05:22,159 Speaker 2: reinforcement learning. 83 00:05:22,480 --> 00:05:26,280 Speaker 1: Well, it all sounds like you're trying to teach a child. 84 00:05:26,839 --> 00:05:31,840 Speaker 1: It's either through supervision or through play, or through punishment. 85 00:05:33,640 --> 00:05:36,200 Speaker 2: Yeah, And in some sense, a lot of machine learning 86 00:05:36,440 --> 00:05:39,919 Speaker 2: algorithms and ways of trying to learn these things are 87 00:05:40,360 --> 00:05:42,560 Speaker 2: vaguely inspired by. 88 00:05:42,400 --> 00:05:45,200 Speaker 3: Some notion of how humans learn, although. 89 00:05:45,040 --> 00:05:47,400 Speaker 2: The practicalities of how we actually do this might be 90 00:05:47,480 --> 00:05:48,200 Speaker 2: quite different. 91 00:05:48,600 --> 00:05:51,760 Speaker 1: And so we talk about AI in the climate context 92 00:05:51,760 --> 00:05:55,840 Speaker 1: for two big reasons. One is because of the energy 93 00:05:55,920 --> 00:06:01,320 Speaker 1: cost of creating AI models and use AI models. And 94 00:06:01,400 --> 00:06:04,760 Speaker 1: second is that these models, again different types of them, 95 00:06:05,000 --> 00:06:08,880 Speaker 1: can have different applications that could make solving for climate 96 00:06:08,960 --> 00:06:11,160 Speaker 1: change deploying these solutions easier. 97 00:06:12,000 --> 00:06:14,640 Speaker 2: And I would add in a third pillar, which is 98 00:06:14,640 --> 00:06:18,520 Speaker 2: that AI is also used for many types of applications 99 00:06:18,560 --> 00:06:20,640 Speaker 2: that make solving climate harder. 100 00:06:20,920 --> 00:06:21,839 Speaker 3: So when we talk. 101 00:06:21,680 --> 00:06:23,760 Speaker 2: About the good and the bad, we should think about 102 00:06:23,800 --> 00:06:25,159 Speaker 2: the fact that AI has. 103 00:06:25,040 --> 00:06:25,960 Speaker 3: Its own footprint. 104 00:06:26,360 --> 00:06:28,839 Speaker 2: ANAI is used in both good and bad ways. 105 00:06:29,000 --> 00:06:31,560 Speaker 1: Yeah, and so let's address the footprint part, because the 106 00:06:31,600 --> 00:06:34,280 Speaker 1: size of footprint that may come from AI will be 107 00:06:34,400 --> 00:06:39,120 Speaker 1: dependent on the type of AI. And large language models 108 00:06:39,160 --> 00:06:41,279 Speaker 1: are in the news because these are the models that 109 00:06:41,320 --> 00:06:44,680 Speaker 1: try and train themselves on the entire corpus of the Internet, 110 00:06:44,920 --> 00:06:47,960 Speaker 1: and that just requires a ton of computing power, which 111 00:06:48,000 --> 00:06:52,479 Speaker 1: is why companies like Microsoft and Alphabet and Meta are 112 00:06:52,520 --> 00:06:56,479 Speaker 1: all now in this rush to build more data centers 113 00:06:56,960 --> 00:07:00,640 Speaker 1: consume a lot more power in the process can blow 114 00:07:00,680 --> 00:07:02,840 Speaker 1: past some of their own set climate goals, as we 115 00:07:02,920 --> 00:07:07,120 Speaker 1: found out with Microsoft's recent update that it's emissions are 116 00:07:07,120 --> 00:07:10,920 Speaker 1: actually thirty percent higher rather than thirty percent lower last year. 117 00:07:11,360 --> 00:07:14,720 Speaker 1: Does that mean all kinds of AI is doomed to 118 00:07:14,880 --> 00:07:17,400 Speaker 1: have a higher footprint because all kinds of AI will 119 00:07:17,520 --> 00:07:19,600 Speaker 1: want as much data as possible. 120 00:07:20,320 --> 00:07:23,480 Speaker 2: So there's definitely a diversity in the types of AI 121 00:07:23,560 --> 00:07:25,800 Speaker 2: that exist and as a result, the kind of energy 122 00:07:25,880 --> 00:07:29,520 Speaker 2: usage of these. So there has been in long history, 123 00:07:29,600 --> 00:07:32,320 Speaker 2: you know, AI and machine learning models that use you know, 124 00:07:32,720 --> 00:07:35,440 Speaker 2: a reasonable amount of data, but much less than the 125 00:07:35,560 --> 00:07:38,680 Speaker 2: entirety of the Internet, and where the models themselves are 126 00:07:38,720 --> 00:07:42,040 Speaker 2: also much smaller, they have fewer parameters and as a result, 127 00:07:42,080 --> 00:07:44,880 Speaker 2: you don't need as much computational power to actually update 128 00:07:44,920 --> 00:07:48,000 Speaker 2: and get these models to learn. And so you know, 129 00:07:48,120 --> 00:07:50,920 Speaker 2: some of the models that we develop, even in my 130 00:07:51,000 --> 00:07:52,680 Speaker 2: research group, can run on a laptop. 131 00:07:53,280 --> 00:07:54,880 Speaker 3: But then of course you have these. 132 00:07:54,680 --> 00:07:59,440 Speaker 2: You know, large data intensive state of the art algorithms 133 00:07:59,480 --> 00:08:03,040 Speaker 2: that are kind of being deployed through products like chat GPT, 134 00:08:03,720 --> 00:08:07,240 Speaker 2: and definitely the kind of energy consumption and you know, 135 00:08:07,640 --> 00:08:11,880 Speaker 2: water consumption from data centers, the materiality impacts of actually 136 00:08:11,960 --> 00:08:15,040 Speaker 2: getting the computational hardware in place that is starting to 137 00:08:15,080 --> 00:08:15,600 Speaker 2: get worrying. 138 00:08:16,480 --> 00:08:20,960 Speaker 1: Right now, climate applications themselves don't have to go down 139 00:08:21,120 --> 00:08:23,760 Speaker 1: the LLM route of having to consume that much data. 140 00:08:23,800 --> 00:08:26,400 Speaker 1: You know, you say, models on a laptop can work. 141 00:08:26,920 --> 00:08:31,320 Speaker 1: Let's start with that, because you got into AI through 142 00:08:31,680 --> 00:08:35,319 Speaker 1: trying to figure out how to make the grid work better, right. 143 00:08:35,280 --> 00:08:35,839 Speaker 3: That's right. 144 00:08:35,960 --> 00:08:40,040 Speaker 2: So basically, as we start to integrate you know, more 145 00:08:40,240 --> 00:08:44,760 Speaker 2: and more renewables into power grids, many of these renewables, 146 00:08:44,800 --> 00:08:47,640 Speaker 2: their output varies based on the weather, so it varies 147 00:08:47,679 --> 00:08:50,560 Speaker 2: over time. Think about solar, think about wind, and yet 148 00:08:50,600 --> 00:08:53,400 Speaker 2: on a power grid you're having to maintain this exact, 149 00:08:53,520 --> 00:08:56,040 Speaker 2: delicate balance between how much power is put into the 150 00:08:56,040 --> 00:08:58,600 Speaker 2: grid and how much is consumed, which gets harder when 151 00:08:58,640 --> 00:09:00,839 Speaker 2: you have a lot of variations coming onto the grid. 152 00:09:01,280 --> 00:09:03,960 Speaker 2: And so AI and machine learning can be helpful in 153 00:09:04,000 --> 00:09:06,400 Speaker 2: terms of doing things like first, I mean just giving 154 00:09:06,440 --> 00:09:09,559 Speaker 2: us better predictions of what your solar power output, when 155 00:09:09,640 --> 00:09:13,120 Speaker 2: power output, electricity demand will look like, but also in 156 00:09:13,160 --> 00:09:16,680 Speaker 2: actually helping to speed up some of the existing physics 157 00:09:16,720 --> 00:09:19,480 Speaker 2: based and engineering based algorithms that are used to manage 158 00:09:19,520 --> 00:09:22,040 Speaker 2: the power grid in the back end to maintain that balance. 159 00:09:22,480 --> 00:09:25,040 Speaker 1: And so one of the challenges with trying to understand 160 00:09:25,679 --> 00:09:27,920 Speaker 1: as an application to try and help solve some of 161 00:09:27,920 --> 00:09:32,600 Speaker 1: the climate problems is that it becomes really abstract very quickly. 162 00:09:32,760 --> 00:09:35,319 Speaker 1: So you say, oh, yeah, we have a number of 163 00:09:35,400 --> 00:09:37,960 Speaker 1: data points and there's an intelligent way in which we 164 00:09:38,000 --> 00:09:40,960 Speaker 1: can use them, and that gives us an output, but 165 00:09:41,040 --> 00:09:43,640 Speaker 1: we don't usually know why we have that output. But 166 00:09:43,679 --> 00:09:45,360 Speaker 1: that output is better, so we use it and that's 167 00:09:45,400 --> 00:09:48,800 Speaker 1: the solution, and it just does not feel satisfying, you know, 168 00:09:48,880 --> 00:09:52,480 Speaker 1: as a science reporter. To me, the joy of an 169 00:09:52,480 --> 00:09:56,000 Speaker 1: invention is to actually break down the steps, to try 170 00:09:56,000 --> 00:09:58,480 Speaker 1: and figure out why this step led to that step 171 00:09:58,559 --> 00:10:01,000 Speaker 1: led to that step, and finally you have something that 172 00:10:01,200 --> 00:10:05,120 Speaker 1: is really useful. Can we do that with AI? 173 00:10:06,480 --> 00:10:08,600 Speaker 2: Yeah, So I think that there are a couple of 174 00:10:09,240 --> 00:10:12,319 Speaker 2: categories of ways we can think about AI and machine 175 00:10:12,400 --> 00:10:14,959 Speaker 2: learning being used for climate that can help maybe give 176 00:10:14,960 --> 00:10:16,160 Speaker 2: a mental model for what's. 177 00:10:16,000 --> 00:10:17,880 Speaker 3: Actually going on under the hood. 178 00:10:19,120 --> 00:10:22,800 Speaker 2: So one of these categories is, you know, taking large 179 00:10:22,800 --> 00:10:26,840 Speaker 2: streams of broad data and distilling it into actionable information. 180 00:10:27,360 --> 00:10:30,000 Speaker 2: So one project we're funding through Climate Change AI is 181 00:10:30,559 --> 00:10:34,720 Speaker 2: actually a project that tries to improve the sustainability of 182 00:10:34,720 --> 00:10:37,040 Speaker 2: shrimp aquaculture practices. 183 00:10:37,200 --> 00:10:38,880 Speaker 3: So kind of shrimp baco. 184 00:10:38,679 --> 00:10:41,560 Speaker 2: Culture is currently it can be harmful to you know, 185 00:10:41,600 --> 00:10:44,680 Speaker 2: coastal mangrove forests, and that has implications for climate change 186 00:10:44,679 --> 00:10:47,600 Speaker 2: adaptation in terms of kind of flood resilience as well 187 00:10:47,600 --> 00:10:50,600 Speaker 2: as climate change mitigation in terms of the sequestration potential 188 00:10:50,640 --> 00:10:54,040 Speaker 2: of mangroves. And so we're currently funding a team from 189 00:10:54,360 --> 00:10:58,679 Speaker 2: Arizona State, Conservation International and Thinking Machines Data Science from 190 00:10:58,679 --> 00:11:03,360 Speaker 2: in the Philippines to actually use sad light imagery to 191 00:11:03,520 --> 00:11:07,720 Speaker 2: assess aquaculture farms that actually might be able to benefit 192 00:11:07,880 --> 00:11:12,280 Speaker 2: from better aquaculture practices. The intervention here is that you 193 00:11:12,280 --> 00:11:15,480 Speaker 2: can actually do things like if you have an aquaculture farm, 194 00:11:15,520 --> 00:11:18,640 Speaker 2: you can intensify how much you're farming on one part 195 00:11:18,679 --> 00:11:20,439 Speaker 2: of the farm, and then you can kind of conserve 196 00:11:20,520 --> 00:11:23,160 Speaker 2: on another part of the farm, and so without impacting 197 00:11:23,200 --> 00:11:26,000 Speaker 2: your overall productivity, you can just farm in a way 198 00:11:26,000 --> 00:11:30,160 Speaker 2: that's better for the mangroves. And so Conservation International has 199 00:11:30,200 --> 00:11:32,800 Speaker 2: a program where they're working with farmers to try to 200 00:11:33,480 --> 00:11:36,280 Speaker 2: kind of help them do this, but actually identifying which 201 00:11:36,280 --> 00:11:38,720 Speaker 2: farms are amenable to this type of intervention at scale 202 00:11:38,800 --> 00:11:41,360 Speaker 2: is difficult, so they use a combination of you know, 203 00:11:41,559 --> 00:11:44,160 Speaker 2: sad light imagery data on. 204 00:11:44,360 --> 00:11:46,600 Speaker 3: Sea level rise and sea risk and things like. 205 00:11:46,520 --> 00:11:50,680 Speaker 2: This, in order to then actually pinpoint at scale which 206 00:11:50,720 --> 00:11:53,360 Speaker 2: farms might be amenable to this invention and then actually 207 00:11:53,440 --> 00:11:54,480 Speaker 2: go work with them to do that. 208 00:11:54,559 --> 00:11:57,240 Speaker 1: And you said that was just one approach, what are 209 00:11:57,320 --> 00:11:58,680 Speaker 1: some other approaches? 210 00:11:59,000 --> 00:12:01,520 Speaker 2: There's a couple of other actually. So one is you know, 211 00:12:02,080 --> 00:12:06,080 Speaker 2: predicting and forecasting. So taking you know, historical data where 212 00:12:06,120 --> 00:12:09,520 Speaker 2: you have relationships between some input and some quantity you 213 00:12:09,559 --> 00:12:12,000 Speaker 2: would want to predict. So things like I want to 214 00:12:12,000 --> 00:12:15,839 Speaker 2: predict electricity demand on the power grid, so I can 215 00:12:15,880 --> 00:12:18,920 Speaker 2: take historical data about what electricity demand looked like. 216 00:12:19,360 --> 00:12:20,839 Speaker 3: I can take historical. 217 00:12:20,480 --> 00:12:24,120 Speaker 2: Weather data and I can learn relationships between those so 218 00:12:24,160 --> 00:12:26,240 Speaker 2: that in the future, when I have a weather prediction 219 00:12:26,480 --> 00:12:28,440 Speaker 2: but I don't know what the electricity demand would be 220 00:12:28,440 --> 00:12:30,959 Speaker 2: based on that weather prediction, I can just go ahead 221 00:12:30,960 --> 00:12:33,320 Speaker 2: and predict that. And you have kind of for example, 222 00:12:33,400 --> 00:12:36,520 Speaker 2: nonprofits like open Climate Fix that are working with the 223 00:12:36,640 --> 00:12:40,960 Speaker 2: UK Power System Operator to actually improve their electricity demand forecasts, 224 00:12:40,960 --> 00:12:42,720 Speaker 2: and they've been able to use machine learning to have 225 00:12:42,920 --> 00:12:44,120 Speaker 2: the error of those. 226 00:12:43,920 --> 00:12:50,000 Speaker 1: Forecasts after the break. Why it's important for all of 227 00:12:50,120 --> 00:12:53,920 Speaker 1: us to be involved in the development of AI. By 228 00:12:53,960 --> 00:12:56,400 Speaker 1: the way, if you're enjoying this episode, please do take 229 00:12:56,400 --> 00:12:58,319 Speaker 1: a moment to rate and review the show on Apple 230 00:12:58,320 --> 00:13:10,959 Speaker 1: Podcasts or Spotify. It helps other listeners find it. Some companies, 231 00:13:11,360 --> 00:13:14,400 Speaker 1: one that my colleague wrote about called Climate AI is 232 00:13:14,760 --> 00:13:18,200 Speaker 1: using AI to try and improve weather prediction models because 233 00:13:18,400 --> 00:13:21,520 Speaker 1: currently you are starting to get better and better predictions, 234 00:13:22,080 --> 00:13:25,800 Speaker 1: and that has at least for them, been a profitable 235 00:13:26,040 --> 00:13:30,240 Speaker 1: enterprise because then they're working with these large agriculture companies 236 00:13:30,440 --> 00:13:33,040 Speaker 1: that want to figure out when should we start to 237 00:13:33,040 --> 00:13:34,800 Speaker 1: put the seed down or when should we start to 238 00:13:34,840 --> 00:13:38,480 Speaker 1: harvest because we have a better understanding of the weather 239 00:13:39,080 --> 00:13:40,880 Speaker 1: not just over the next two weeks, but over the 240 00:13:40,960 --> 00:13:41,840 Speaker 1: next three months. 241 00:13:42,200 --> 00:13:44,920 Speaker 2: Yeah. So I think this idea of kind of medium 242 00:13:45,000 --> 00:13:48,720 Speaker 2: to long term forecasting is also really cool. But often 243 00:13:48,840 --> 00:13:51,439 Speaker 2: to do good forecasting in these settings you want to 244 00:13:51,559 --> 00:13:56,360 Speaker 2: use a combination of physical models and data. So, for example, 245 00:13:56,679 --> 00:13:59,000 Speaker 2: one of the teams that we're funding at Climate Change 246 00:13:59,040 --> 00:14:02,280 Speaker 2: AI spread Bitch between a couple of US universities and 247 00:14:02,320 --> 00:14:06,000 Speaker 2: an Indian university. They're basically trying to figure out how 248 00:14:06,000 --> 00:14:10,520 Speaker 2: do I actually make longer term predictions of weather in 249 00:14:10,640 --> 00:14:13,680 Speaker 2: order to foster how we actually build out power grids 250 00:14:13,679 --> 00:14:16,800 Speaker 2: for the future. And the difficulty here is if you 251 00:14:17,000 --> 00:14:20,200 Speaker 2: just use past data. What machine learning does is it 252 00:14:20,280 --> 00:14:24,160 Speaker 2: learns patterns in that past data and just projects them forward. 253 00:14:24,560 --> 00:14:27,200 Speaker 2: But the climate is changing, which means that the patterns 254 00:14:27,240 --> 00:14:30,200 Speaker 2: in how weather is occurring are changing, and so you 255 00:14:30,240 --> 00:14:32,560 Speaker 2: can't just use a pure data driven technique to do this. 256 00:14:33,000 --> 00:14:36,080 Speaker 2: And so what this team does is they say, well, 257 00:14:36,080 --> 00:14:39,240 Speaker 2: we have climate models. The issue with climate models is 258 00:14:39,240 --> 00:14:42,560 Speaker 2: that they don't give you very granular information on exactly 259 00:14:42,600 --> 00:14:44,680 Speaker 2: what's going to happen at a particular place, just because 260 00:14:44,680 --> 00:14:48,880 Speaker 2: they're very computationally intensive to run. But if we can 261 00:14:49,080 --> 00:14:52,280 Speaker 2: quote unquote back cast the climate models, so run and 262 00:14:52,320 --> 00:14:54,960 Speaker 2: say what would the climate model say now, and we 263 00:14:55,040 --> 00:14:58,880 Speaker 2: already have really fine grained weather data historically, we can 264 00:14:58,960 --> 00:15:02,320 Speaker 2: learn a mapping between what the climate model said and 265 00:15:02,360 --> 00:15:03,600 Speaker 2: what the weather data would be. 266 00:15:04,240 --> 00:15:04,760 Speaker 3: And then in the. 267 00:15:04,720 --> 00:15:07,520 Speaker 2: Future, where we only have a climate model prediction, we 268 00:15:07,560 --> 00:15:09,640 Speaker 2: can use our learned relationship to say, oh, and this 269 00:15:09,720 --> 00:15:11,320 Speaker 2: is what the weather would be in a more fine 270 00:15:11,360 --> 00:15:13,680 Speaker 2: grained way in the future. With a lot of these 271 00:15:13,720 --> 00:15:16,560 Speaker 2: kind of climate downscaling techniques, you often want to think 272 00:15:16,560 --> 00:15:18,920 Speaker 2: about who is the user of these techniques and as 273 00:15:18,920 --> 00:15:21,920 Speaker 2: a result, what aspects of your downscaled predictions have to 274 00:15:21,920 --> 00:15:24,240 Speaker 2: be good. So here they're actually doing this for the 275 00:15:24,240 --> 00:15:26,960 Speaker 2: power grid planning context, where they're saying, can we produce 276 00:15:27,040 --> 00:15:29,680 Speaker 2: fine grained data sets of what electricity usage will look like, 277 00:15:29,760 --> 00:15:32,520 Speaker 2: wind power production, and solar power production might look like 278 00:15:32,560 --> 00:15:34,440 Speaker 2: in order to facilitate power grid planning. 279 00:15:34,800 --> 00:15:38,440 Speaker 1: All of this sort of was something that you published 280 00:15:38,480 --> 00:15:42,200 Speaker 1: in a paper titled Tackling Climate Change with Machine Learning. 281 00:15:42,520 --> 00:15:44,160 Speaker 1: Why do you need to write this paper? 282 00:15:45,520 --> 00:15:48,960 Speaker 2: Yeah, so, I'd say back in twenty nineteen, we definitely 283 00:15:49,000 --> 00:15:51,600 Speaker 2: saw a combination of a lot of people in the 284 00:15:51,640 --> 00:15:54,120 Speaker 2: AI and machine learning space who wanted to leverage their 285 00:15:54,160 --> 00:15:57,720 Speaker 2: skills to help in facilitating climate action but didn't necessarily 286 00:15:57,720 --> 00:16:00,920 Speaker 2: know how. And on the other side, many people in 287 00:16:00,960 --> 00:16:03,880 Speaker 2: the climate change related space who are seeing, you know, 288 00:16:03,960 --> 00:16:06,640 Speaker 2: things like larger streams of data becoming available and saying 289 00:16:06,680 --> 00:16:09,720 Speaker 2: that how do I actually utilize this? And so we 290 00:16:09,840 --> 00:16:13,120 Speaker 2: really felt like there is a need to really put 291 00:16:13,240 --> 00:16:15,600 Speaker 2: forward for the community, you know, where is it that 292 00:16:15,640 --> 00:16:18,400 Speaker 2: AI is well matched to climate change related problems? In 293 00:16:18,480 --> 00:16:20,960 Speaker 2: order to then help AI people get into the space 294 00:16:21,240 --> 00:16:23,680 Speaker 2: and help climate people understand, okay, for some of these 295 00:16:23,680 --> 00:16:28,400 Speaker 2: complex problems we're seeing is the bottleneck potentially solvable via 296 00:16:28,520 --> 00:16:31,200 Speaker 2: and machine learning. And there were two kind of big 297 00:16:31,240 --> 00:16:33,360 Speaker 2: aims through that work. One is again to lay out 298 00:16:33,360 --> 00:16:36,120 Speaker 2: the space of applications, but the second is to try 299 00:16:36,160 --> 00:16:40,800 Speaker 2: to provide some kind of mental model and guidance for 300 00:16:40,880 --> 00:16:44,800 Speaker 2: how to do this work in a sound, impactful and 301 00:16:44,880 --> 00:16:48,720 Speaker 2: responsible way, because there are lots of places where AI 302 00:16:48,880 --> 00:16:51,800 Speaker 2: is not the right fit and it can be a 303 00:16:51,880 --> 00:16:55,760 Speaker 2: huge distraction, or there are ways that, for example, because 304 00:16:55,920 --> 00:16:59,560 Speaker 2: you often have you know, data and computational power concentrated 305 00:16:59,640 --> 00:17:02,440 Speaker 2: in certain and geographies versus others, where the practice of 306 00:17:02,480 --> 00:17:05,639 Speaker 2: AI can exacerbate some of these inequities by basically causing 307 00:17:05,680 --> 00:17:08,600 Speaker 2: people who already have access to compute data to be 308 00:17:08,640 --> 00:17:10,719 Speaker 2: able to do a lot more and leave others behind. 309 00:17:11,119 --> 00:17:13,239 Speaker 2: And so there's a lot kind of in there to 310 00:17:13,359 --> 00:17:15,840 Speaker 2: make sure we're actually moving the space forward in a 311 00:17:15,880 --> 00:17:19,360 Speaker 2: way that makes sense for climate and for equity. 312 00:17:20,200 --> 00:17:23,600 Speaker 1: Yeah, So recently we spoke to the president of Microsoft, 313 00:17:23,640 --> 00:17:27,480 Speaker 1: Brad Smith, and he was talking about how he would 314 00:17:27,640 --> 00:17:30,720 Speaker 1: like AI to be available to everybody. He doesn't want 315 00:17:30,800 --> 00:17:35,040 Speaker 1: AI haves and AI have nods and I interpreted that 316 00:17:35,119 --> 00:17:38,919 Speaker 1: to mean, you know, we've had technological leaves in the past, 317 00:17:38,960 --> 00:17:42,280 Speaker 1: and when they have been more widely available that has 318 00:17:42,320 --> 00:17:47,840 Speaker 1: been beneficial to humanity mobile phones, Internet. Do you see 319 00:17:48,359 --> 00:17:56,879 Speaker 1: AI as being essential for unlocking human potential? Like Microsoft president. 320 00:17:56,560 --> 00:17:59,160 Speaker 3: Is saying, So, I think there are two things I'd 321 00:17:59,200 --> 00:17:59,760 Speaker 3: like to unpack. 322 00:17:59,800 --> 00:18:04,960 Speaker 2: And one, I think that AI can be a really 323 00:18:05,200 --> 00:18:09,600 Speaker 2: powerful kind of support and accelerator for many different you know, 324 00:18:09,640 --> 00:18:12,160 Speaker 2: climate change related applications and others. 325 00:18:12,720 --> 00:18:15,320 Speaker 3: There are some where I think it is essential. 326 00:18:15,560 --> 00:18:19,040 Speaker 2: For example, don't know how we will manage power grids 327 00:18:19,080 --> 00:18:21,800 Speaker 2: with lots of variability and large amounts of renewables without AI. 328 00:18:22,400 --> 00:18:24,840 Speaker 2: There are other places where it can be helpful, but 329 00:18:24,880 --> 00:18:28,240 Speaker 2: I don't necessarily think it's the critical bottleneck. One thing 330 00:18:28,280 --> 00:18:32,359 Speaker 2: I'd like to also mention is that with certain things 331 00:18:32,560 --> 00:18:38,520 Speaker 2: like mobile technologies and stutch, democratization has been used to 332 00:18:38,640 --> 00:18:41,320 Speaker 2: sort of indicate, Okay, a few people created a thing 333 00:18:41,359 --> 00:18:43,240 Speaker 2: and it was pushed onto the rest of the world. 334 00:18:43,840 --> 00:18:48,719 Speaker 2: That's not actually, in some ways democratization, especially in the 335 00:18:48,720 --> 00:18:52,119 Speaker 2: context of AI, where actually the type of AI you 336 00:18:52,200 --> 00:18:55,359 Speaker 2: build and that the way you do it it fundamentally 337 00:18:55,400 --> 00:18:58,040 Speaker 2: needs to look very different depending on the context you're in, 338 00:18:58,160 --> 00:19:00,440 Speaker 2: you have different amounts of data, and different context you 339 00:19:00,520 --> 00:19:01,640 Speaker 2: have different amounts of compute. 340 00:19:01,640 --> 00:19:02,800 Speaker 3: In different contexts, you. 341 00:19:02,800 --> 00:19:05,840 Speaker 2: Have different amounts of kind of existing knowledge that can 342 00:19:05,840 --> 00:19:08,760 Speaker 2: be integrated into systems. And so if we kind of 343 00:19:08,800 --> 00:19:11,760 Speaker 2: develop AI among a small set of entities and then 344 00:19:11,880 --> 00:19:13,840 Speaker 2: push that onto the rest of the world, it's actually 345 00:19:13,840 --> 00:19:15,760 Speaker 2: not going to serve the needs of the full world. 346 00:19:15,840 --> 00:19:19,560 Speaker 2: And so democratization really means enabling more people to contribute 347 00:19:19,560 --> 00:19:22,920 Speaker 2: to the trajectory of actually developing AI, not to sort 348 00:19:22,920 --> 00:19:25,920 Speaker 2: of being users of a product that a few people developed. 349 00:19:26,200 --> 00:19:26,800 Speaker 3: On the other. 350 00:19:26,640 --> 00:19:30,280 Speaker 1: Side, can you give a specific way in which that 351 00:19:30,359 --> 00:19:34,040 Speaker 1: might play out, say through the development of Chad, GPT 352 00:19:34,359 --> 00:19:37,320 Speaker 1: or cloud or these other types of generative AI products. 353 00:19:37,640 --> 00:19:41,840 Speaker 2: Yes, So basically, if you think about something like GPT, 354 00:19:42,320 --> 00:19:44,840 Speaker 2: it needs a huge amount of data to train, it 355 00:19:44,880 --> 00:19:49,000 Speaker 2: needs a huge amount of compute to run, and most 356 00:19:49,160 --> 00:19:51,439 Speaker 2: entities in the world do not have the ability to 357 00:19:51,480 --> 00:19:54,040 Speaker 2: curate or collect that amount of data, nor do they 358 00:19:54,119 --> 00:19:56,960 Speaker 2: have the ability to pay for or procure the amount 359 00:19:56,960 --> 00:19:57,760 Speaker 2: of compute needed. 360 00:19:57,600 --> 00:20:00,520 Speaker 3: To run those models. So those models are being developed 361 00:20:00,520 --> 00:20:02,560 Speaker 3: by a small set of people and then. 362 00:20:02,480 --> 00:20:04,760 Speaker 2: Kind of packaged and sent out in a kind of 363 00:20:04,800 --> 00:20:08,760 Speaker 2: interface like CHATDPT that many people can use, and that 364 00:20:08,840 --> 00:20:11,600 Speaker 2: can be helpful for a certain set of use cases, 365 00:20:12,040 --> 00:20:14,960 Speaker 2: but there are lots of use cases that don't necessarily 366 00:20:15,000 --> 00:20:15,760 Speaker 2: fit that mold. 367 00:20:16,480 --> 00:20:17,879 Speaker 3: Imagine that you're trying. 368 00:20:17,640 --> 00:20:24,800 Speaker 2: To train your own weather prediction model in a situation 369 00:20:24,920 --> 00:20:28,040 Speaker 2: where you have some amount of data and also some 370 00:20:28,119 --> 00:20:29,960 Speaker 2: amount of knowledge of just how. 371 00:20:29,800 --> 00:20:31,520 Speaker 3: Kind of weather physics works. 372 00:20:32,200 --> 00:20:34,680 Speaker 2: If you do this in a fully data driven way, 373 00:20:35,160 --> 00:20:37,600 Speaker 2: there does exist the reality in which you're able to 374 00:20:37,720 --> 00:20:40,200 Speaker 2: just purely from data figure out how to predict weather. 375 00:20:40,720 --> 00:20:44,200 Speaker 2: But you often need much more data and a much 376 00:20:44,280 --> 00:20:48,199 Speaker 2: bigger model if you're basically not embedding the rules of 377 00:20:48,200 --> 00:20:50,920 Speaker 2: physics and as a result learning them fully from scratch, 378 00:20:51,200 --> 00:20:53,119 Speaker 2: And so that leads to a situation where you again 379 00:20:53,160 --> 00:20:55,399 Speaker 2: have a bigger model that fewer people can use and 380 00:20:55,440 --> 00:20:57,800 Speaker 2: fewer people can train. And it can also lead to 381 00:20:57,840 --> 00:21:00,320 Speaker 2: situations where people say, oh, I don't have a lot 382 00:21:00,320 --> 00:21:02,119 Speaker 2: of data. Is the thing I'm supposed to do collect 383 00:21:02,160 --> 00:21:04,280 Speaker 2: tons and tons of data, So they invest a bunch 384 00:21:04,320 --> 00:21:08,560 Speaker 2: of money into setting up data infrastructure and data collection. 385 00:21:09,720 --> 00:21:12,119 Speaker 1: On the other side, though, this is and this is 386 00:21:12,119 --> 00:21:15,080 Speaker 1: a real limitation because if you looked at the map 387 00:21:15,359 --> 00:21:19,320 Speaker 1: of weather stations in the world, it maps kind of 388 00:21:19,359 --> 00:21:22,320 Speaker 1: one to one to the wealth there is in the world. 389 00:21:22,720 --> 00:21:27,200 Speaker 1: America and Europe is littered with weather stations, whereas Africa 390 00:21:27,600 --> 00:21:30,840 Speaker 1: is empty. And so if you go down that route, 391 00:21:31,119 --> 00:21:33,840 Speaker 1: the answer would be just deploy more weather stations. But 392 00:21:34,240 --> 00:21:35,399 Speaker 1: it is in the right answer. 393 00:21:35,560 --> 00:21:38,200 Speaker 2: Yeah, I mean, and in some sense, fixing the data 394 00:21:38,200 --> 00:21:40,600 Speaker 2: in equity problem is obviously great thing. It would be 395 00:21:40,600 --> 00:21:42,760 Speaker 2: great to have more weather stations in Africa than there 396 00:21:42,760 --> 00:21:46,399 Speaker 2: are today. But there are kind of additional ways to 397 00:21:46,440 --> 00:21:49,159 Speaker 2: contend with this problem, which include take the data you 398 00:21:49,280 --> 00:21:52,199 Speaker 2: have and take some knowledge of the physical rules that 399 00:21:52,240 --> 00:21:55,399 Speaker 2: govern weather, combine them together in a clever way so 400 00:21:55,440 --> 00:21:58,760 Speaker 2: you don't need as much data to still get good answers. 401 00:21:59,080 --> 00:22:01,680 Speaker 2: And so that really forms how you think about as 402 00:22:01,760 --> 00:22:05,119 Speaker 2: an organization, as a country, where you invest your resources. 403 00:22:05,160 --> 00:22:07,119 Speaker 3: If you're just assuming that you invest them. 404 00:22:06,960 --> 00:22:10,760 Speaker 2: In collecting a maximal amount of data necessary, actually that 405 00:22:10,840 --> 00:22:13,399 Speaker 2: might actually be a misinvestment of resources if you assume 406 00:22:13,480 --> 00:22:16,720 Speaker 2: that AI just means maximal data collection and learning only 407 00:22:16,760 --> 00:22:20,600 Speaker 2: on data. In addition to sort of taking in data 408 00:22:20,640 --> 00:22:24,160 Speaker 2: and producing insight, there are also situations in which AI 409 00:22:24,160 --> 00:22:27,280 Speaker 2: and machine learning can actually help us to more efficiently 410 00:22:27,400 --> 00:22:30,240 Speaker 2: optimize a complex system in order. 411 00:22:30,080 --> 00:22:31,360 Speaker 3: To improve its efficiency. 412 00:22:31,800 --> 00:22:35,600 Speaker 2: So, for example, if we think about buildings, there are 413 00:22:35,640 --> 00:22:38,320 Speaker 2: lots of ways in which we can actually better control, 414 00:22:38,480 --> 00:22:41,720 Speaker 2: for example, the heating and cooling systems in buildings, both 415 00:22:41,760 --> 00:22:44,960 Speaker 2: to reduce the amount of energy they're actually using while 416 00:22:45,040 --> 00:22:47,880 Speaker 2: kind of maintaining something like thermal comfort in the building, 417 00:22:48,359 --> 00:22:51,400 Speaker 2: and also be responsive to things like how much renewable 418 00:22:51,480 --> 00:22:54,320 Speaker 2: energy is actually available on the grid at this particular time. 419 00:22:54,359 --> 00:22:58,040 Speaker 2: This concept of demand response. What's kind of interesting is 420 00:22:58,200 --> 00:23:01,159 Speaker 2: when you start to think not just about individual building performance, 421 00:23:01,200 --> 00:23:04,320 Speaker 2: but also how this connects up to the power grid 422 00:23:04,320 --> 00:23:07,280 Speaker 2: and when renewable energy is available, you sometimes want to 423 00:23:07,280 --> 00:23:10,000 Speaker 2: start thinking about this not just at the individual building level, 424 00:23:10,040 --> 00:23:12,440 Speaker 2: but for example, at the neighborhood level, where you actually 425 00:23:12,480 --> 00:23:14,600 Speaker 2: might want to co optimize what's going on in different 426 00:23:14,640 --> 00:23:17,359 Speaker 2: buildings to jointly be doing the best thing for overall 427 00:23:17,400 --> 00:23:20,399 Speaker 2: efficiency and the power grid. And so one of the 428 00:23:20,440 --> 00:23:22,760 Speaker 2: projects we're funding through Climate Change AI is called the 429 00:23:22,760 --> 00:23:27,680 Speaker 2: City Learn Challenge, and they actually created a simulation environment 430 00:23:27,720 --> 00:23:31,239 Speaker 2: that actually tries to provide some structure of Okay, there 431 00:23:31,240 --> 00:23:33,040 Speaker 2: are a bunch of buildings they're connected up to a 432 00:23:33,080 --> 00:23:35,600 Speaker 2: neighborhood grid in this particular way, here's some data on 433 00:23:35,600 --> 00:23:38,479 Speaker 2: how they're consuming energy, and they're putting this forward as 434 00:23:38,520 --> 00:23:40,959 Speaker 2: a challenge to the machine learning community to say, can 435 00:23:41,000 --> 00:23:43,000 Speaker 2: you come up with better ways to actually optimize this 436 00:23:43,000 --> 00:23:44,879 Speaker 2: neighborhood to improve its energy efficiency? 437 00:23:46,000 --> 00:23:48,200 Speaker 1: Yeah, that is cool. I feel like one other thing 438 00:23:48,240 --> 00:23:53,240 Speaker 1: that I could be helping is speeding up innovation with 439 00:23:53,359 --> 00:23:57,080 Speaker 1: these solutions in places where otherwise you would have required 440 00:23:57,119 --> 00:24:01,399 Speaker 1: more time, more skills, more people with this, especially in 441 00:24:01,400 --> 00:24:03,639 Speaker 1: developing countries where you really want to speed up the 442 00:24:03,680 --> 00:24:07,480 Speaker 1: solution set. I could allow for these sort of optimization 443 00:24:07,640 --> 00:24:11,120 Speaker 1: techniques to come through more quickly than it would otherwise 444 00:24:11,160 --> 00:24:12,199 Speaker 1: have done. 445 00:24:12,560 --> 00:24:12,919 Speaker 3: Yeah. 446 00:24:12,960 --> 00:24:15,719 Speaker 2: So across the projects that we kind of are funding 447 00:24:15,760 --> 00:24:18,560 Speaker 2: and kind of facilitating through Climate change AI, they are 448 00:24:18,680 --> 00:24:22,880 Speaker 2: you know, happening all around the world. So, for example, 449 00:24:23,119 --> 00:24:26,639 Speaker 2: one of the projects we're funding is a team of 450 00:24:26,680 --> 00:24:30,680 Speaker 2: researchers working with the government of Fiji to actually better 451 00:24:30,960 --> 00:24:34,159 Speaker 2: map the damages from floods that occur in Fiji in 452 00:24:34,240 --> 00:24:36,920 Speaker 2: order to facilitate Fijian disaster response efforts. 453 00:24:37,000 --> 00:24:37,800 Speaker 3: The idea being. 454 00:24:37,680 --> 00:24:40,320 Speaker 2: That when you actually are trying to figure out, Okay, 455 00:24:40,320 --> 00:24:43,720 Speaker 2: in a flood, what happened, who was affected, It's really 456 00:24:43,800 --> 00:24:46,439 Speaker 2: hard to kind of systematically and fully collect that on 457 00:24:46,480 --> 00:24:49,840 Speaker 2: the ground data. And so one of the teams that 458 00:24:49,840 --> 00:24:53,480 Speaker 2: we're funding is actually alongside the Government of Fiji developing 459 00:24:53,520 --> 00:24:57,880 Speaker 2: algorithms to kind of map from satellite imagery to targeted 460 00:24:57,920 --> 00:25:00,639 Speaker 2: information about what the impacts were after flood and to 461 00:25:00,680 --> 00:25:03,480 Speaker 2: be able to kind of continuously update these maps based 462 00:25:03,520 --> 00:25:06,600 Speaker 2: on satellite imagery in order to aid disaster response efforts. 463 00:25:06,600 --> 00:25:08,840 Speaker 2: So that's one example, but a lot of this work 464 00:25:08,880 --> 00:25:11,600 Speaker 2: is going on all around the world. 465 00:25:12,000 --> 00:25:14,160 Speaker 1: And so going back to the start of the conversation 466 00:25:14,240 --> 00:25:17,480 Speaker 1: where you said, there's also how you can use those 467 00:25:17,520 --> 00:25:21,760 Speaker 1: same tools, but to actually increase emissions. You could optimize 468 00:25:21,800 --> 00:25:25,600 Speaker 1: for how you can extract oil and gas in a 469 00:25:25,680 --> 00:25:29,000 Speaker 1: cheaper way, or go to places that previously were not 470 00:25:29,440 --> 00:25:33,719 Speaker 1: found or not reachable. Is that the biggest concern is 471 00:25:33,720 --> 00:25:37,000 Speaker 1: that the biggest downside of AI, even more so than 472 00:25:37,160 --> 00:25:38,040 Speaker 1: the resource use. 473 00:25:38,840 --> 00:25:41,720 Speaker 2: Yeah, to me, I think that we obviously need to 474 00:25:41,720 --> 00:25:45,159 Speaker 2: be thinking about both the resource use and the applications. 475 00:25:45,440 --> 00:25:47,920 Speaker 2: But the applications are very concerning to me because I 476 00:25:47,960 --> 00:25:51,520 Speaker 2: think they're having an outsized negative impact some of these 477 00:25:51,560 --> 00:25:55,720 Speaker 2: applications while also not being centered in the conversations about 478 00:25:55,720 --> 00:25:57,040 Speaker 2: how we actually align the use. 479 00:25:56,920 --> 00:25:58,040 Speaker 3: Of AI with climate action. 480 00:25:58,880 --> 00:26:01,800 Speaker 2: So oil and gas is example, but there are other 481 00:26:01,920 --> 00:26:04,960 Speaker 2: things like you know, AI being the driver behind targeted 482 00:26:05,000 --> 00:26:08,480 Speaker 2: advertising and increases of consumption in ways that don't always 483 00:26:08,520 --> 00:26:12,720 Speaker 2: make us happier but do increase our resource use. AI 484 00:26:13,000 --> 00:26:16,840 Speaker 2: also drives in many ways the information that we actually 485 00:26:16,880 --> 00:26:20,200 Speaker 2: consume online, and that has really a lot of ties 486 00:26:20,240 --> 00:26:24,520 Speaker 2: to the spread of climate information or misinformation in ways 487 00:26:24,520 --> 00:26:27,280 Speaker 2: that could be harmful or helpful, depending on how we're 488 00:26:27,320 --> 00:26:31,080 Speaker 2: actually shaping those particular trends of AI induced information spread. 489 00:26:31,240 --> 00:26:35,360 Speaker 2: And then there are also things like AI for autonomous vehicles, 490 00:26:35,400 --> 00:26:38,160 Speaker 2: which we don't often talk about in the context of climate, 491 00:26:38,520 --> 00:26:42,080 Speaker 2: but where the choices we're making are affecting the transportation 492 00:26:42,240 --> 00:26:44,760 Speaker 2: sector in ways that could be good or bad for 493 00:26:44,800 --> 00:26:50,760 Speaker 2: the climate. If you're kind of facilitating private fossil fuel transportation, 494 00:26:50,960 --> 00:26:54,600 Speaker 2: then you're potentially increasing energy usage and emissions, whereas if 495 00:26:54,600 --> 00:26:58,840 Speaker 2: you're using autonomous vehicles to facilitate you know, public multimodal transit, 496 00:26:58,920 --> 00:27:01,840 Speaker 2: you're potentially bringing the missions of the sector down. So 497 00:27:02,920 --> 00:27:06,119 Speaker 2: I think the applications really can have an outsized impact, 498 00:27:06,200 --> 00:27:08,119 Speaker 2: and it's really important to not leave them out of 499 00:27:08,119 --> 00:27:08,720 Speaker 2: the conversation. 500 00:27:09,840 --> 00:27:12,560 Speaker 1: And my exposure to AI actually went back a decade 501 00:27:12,560 --> 00:27:14,800 Speaker 1: when I was in grad school at Oxford, and it 502 00:27:14,840 --> 00:27:17,359 Speaker 1: wasn't really the models or the applications, but it was 503 00:27:17,400 --> 00:27:20,280 Speaker 1: the ethics. There was a lot of conversations that were 504 00:27:20,280 --> 00:27:23,439 Speaker 1: happening around the ethics of how you would put AI 505 00:27:23,560 --> 00:27:26,879 Speaker 1: to use. Do you think we're doing substantial work on 506 00:27:27,000 --> 00:27:32,040 Speaker 1: the ethical side to ensure that the applications are beneficial 507 00:27:32,119 --> 00:27:35,320 Speaker 1: to humanity or are we just in this race to 508 00:27:35,400 --> 00:27:38,440 Speaker 1: develop new AI products and have kind of forgotten that 509 00:27:38,480 --> 00:27:41,080 Speaker 1: there are huge ethical implications here. 510 00:27:41,640 --> 00:27:44,880 Speaker 2: So ethics is a really really important part of the conversation, 511 00:27:45,040 --> 00:27:46,800 Speaker 2: and I think there's been a lot of great work 512 00:27:46,880 --> 00:27:49,120 Speaker 2: done on it, but there's a lot more that needs 513 00:27:49,160 --> 00:27:52,320 Speaker 2: to be done. So you have things like UNESCO's AI 514 00:27:52,359 --> 00:27:55,879 Speaker 2: ethics recommendations, which were actually you know, adopted very widely. 515 00:27:55,960 --> 00:27:58,720 Speaker 3: We're really extensive in terms of thinking about things. 516 00:27:58,440 --> 00:28:03,760 Speaker 2: Like you know, bias, equity, privacy, transparency, environmental impact, which 517 00:28:03,800 --> 00:28:06,160 Speaker 2: I would also count as a part of ethics. 518 00:28:06,680 --> 00:28:09,480 Speaker 3: And so I think there's been some really great thinking. 519 00:28:09,200 --> 00:28:12,520 Speaker 2: Down on this, but that there's a lot more that 520 00:28:12,600 --> 00:28:15,600 Speaker 2: needs to be done to sort of operationalize this and 521 00:28:15,640 --> 00:28:18,600 Speaker 2: also incentivize people to actually do work in the ethical 522 00:28:18,640 --> 00:28:21,199 Speaker 2: way rather than the way that kind of leaves ethics 523 00:28:21,200 --> 00:28:23,640 Speaker 2: behind and just you know, you run forward. So when 524 00:28:23,640 --> 00:28:26,119 Speaker 2: we talk about AI ethics, we historically have been talking 525 00:28:26,119 --> 00:28:31,840 Speaker 2: about issues like fairness, equity, transparency, privacy, and so forth. 526 00:28:32,359 --> 00:28:35,560 Speaker 1: Or friendly AI that we shouldn't create something that would 527 00:28:35,560 --> 00:28:37,600 Speaker 1: then want to try and destroy humanity. 528 00:28:37,720 --> 00:28:38,280 Speaker 3: And that's the. 529 00:28:38,280 --> 00:28:40,320 Speaker 2: Kind of part that has come kind of into the 530 00:28:40,320 --> 00:28:44,280 Speaker 2: conversation really recently, this idea of you know, AI existential risk, 531 00:28:44,360 --> 00:28:48,520 Speaker 2: AI existential threat and so forth. And I would say 532 00:28:48,520 --> 00:28:52,040 Speaker 2: that that's not an unimportant part of the conversation. We 533 00:28:52,080 --> 00:28:54,320 Speaker 2: really should be thinking about the full range of risks 534 00:28:54,320 --> 00:28:57,000 Speaker 2: that AI can pose and addressing them, but it's become 535 00:28:57,040 --> 00:29:00,840 Speaker 2: maybe an outsized part of the conversation. We should think 536 00:29:00,840 --> 00:29:03,719 Speaker 2: about AI ethics holistically and make sure that we're not 537 00:29:03,880 --> 00:29:07,000 Speaker 2: letting kind of one particular sub part of AI ethics 538 00:29:07,120 --> 00:29:10,200 Speaker 2: dominate the conversation at the expense of really thinking about 539 00:29:10,240 --> 00:29:11,640 Speaker 2: the rest of AI ethics as well. 540 00:29:11,800 --> 00:29:14,200 Speaker 3: Really, there's a huge need to. 541 00:29:14,320 --> 00:29:20,320 Speaker 2: Democratize literacy skills and expertise on AI, so that more 542 00:29:20,360 --> 00:29:22,880 Speaker 2: people are able to engage in a way that is 543 00:29:23,240 --> 00:29:28,720 Speaker 2: kind of informed by knowledge of the strength's limitations risks associated. 544 00:29:28,200 --> 00:29:28,840 Speaker 3: With the technology. 545 00:29:28,880 --> 00:29:32,640 Speaker 2: And so I think really enabling more people to participate 546 00:29:33,280 --> 00:29:36,360 Speaker 2: by having that literacy, skills and expertise is really, I 547 00:29:36,360 --> 00:29:38,800 Speaker 2: think the huge thing that we need to achieve at 548 00:29:38,800 --> 00:29:39,200 Speaker 2: the moment. 549 00:29:39,920 --> 00:29:42,320 Speaker 1: I did enjoy this conversation a lot. Thank you, Thanks 550 00:29:42,360 --> 00:29:51,160 Speaker 1: so much, Thank you for listening to Zero. If you 551 00:29:51,240 --> 00:29:53,440 Speaker 1: liked this episode, please take a moment to rate or 552 00:29:53,440 --> 00:29:56,719 Speaker 1: review the show on Apple Podcasts and Spotify. Share this 553 00:29:56,760 --> 00:29:59,880 Speaker 1: episode with a friend or with someone who fears our 554 00:30:00,080 --> 00:30:03,680 Speaker 1: robot overlords. You can get in touch at zero pod 555 00:30:03,680 --> 00:30:07,520 Speaker 1: at Bloomberg dot Net. Zero's producer is Mightily Round. Our 556 00:30:07,600 --> 00:30:11,200 Speaker 1: theme music is composed by Wondering Special thanks to Kira 557 00:30:11,240 --> 00:30:15,720 Speaker 1: Bendram and Alicia Clanton. I am Akshatrati back soon.