1 00:00:00,080 --> 00:00:03,360 Speaker 1: Hi, it's Aksha. This week we are replaying our conversation 2 00:00:03,520 --> 00:00:07,200 Speaker 1: I had with Priyadante of MIT about the role AI 3 00:00:07,320 --> 00:00:10,760 Speaker 1: can play in tackling climate change. No one is more 4 00:00:10,800 --> 00:00:16,040 Speaker 1: knowledgeable or realistic about this area than Priya, and no 5 00:00:16,079 --> 00:00:19,400 Speaker 1: one is doing more to ensure that AI's climate applications 6 00:00:19,440 --> 00:00:24,479 Speaker 1: are steered responsibly. Enjoy the episode Welcome to Zero. I 7 00:00:24,520 --> 00:00:41,120 Speaker 1: am Kshatrati this week what good can AI do? Remember 8 00:00:41,159 --> 00:00:44,920 Speaker 1: when data was the new oil? I'm specifically thinking of 9 00:00:44,960 --> 00:00:48,800 Speaker 1: a twenty seventeen cover of The Economists showing Google, Amazon 10 00:00:48,920 --> 00:00:52,440 Speaker 1: and other tech giants as big offshore oil rigs. The 11 00:00:52,560 --> 00:00:56,400 Speaker 1: idea being that data was a new critical resource and 12 00:00:56,920 --> 00:01:01,360 Speaker 1: it was going to reshape the world in some ways 13 00:01:01,720 --> 00:01:05,280 Speaker 1: that has already happened. Maybe this podcast was suggested to 14 00:01:05,319 --> 00:01:08,759 Speaker 1: you by Spotify or Apple based on your listening history. 15 00:01:09,200 --> 00:01:14,280 Speaker 1: Just a small example of big data at work. Artificial intelligence, 16 00:01:14,440 --> 00:01:19,000 Speaker 1: the latest buzzword, of course, thrives on data. That devouring 17 00:01:19,000 --> 00:01:22,880 Speaker 1: of data is energy and resource intensive. It's something we 18 00:01:22,959 --> 00:01:27,200 Speaker 1: discussed in last week's episode with Microsoft president Bradsmith. The 19 00:01:27,240 --> 00:01:30,720 Speaker 1: company wants to be cover negative but is instead seeing 20 00:01:30,800 --> 00:01:35,760 Speaker 1: its emissions grow, but of course fed the right data, 21 00:01:36,160 --> 00:01:40,280 Speaker 1: EI can do amazing things, even help tackle climate change, 22 00:01:41,200 --> 00:01:45,800 Speaker 1: but how exactly. If there's one person taking the lead 23 00:01:45,840 --> 00:01:50,040 Speaker 1: on that question, it's MIT's Priadante. She's a professor of 24 00:01:50,080 --> 00:01:53,600 Speaker 1: electrical engineering and AI and the co founder of Climate 25 00:01:53,680 --> 00:01:58,320 Speaker 1: Change AI, an organization bringing together academics and industry leaders 26 00:01:58,720 --> 00:02:03,120 Speaker 1: interested in how A can be used for climate solutions. 27 00:02:03,520 --> 00:02:07,800 Speaker 1: Her group funds independent projects and fieldwork tackling everything from 28 00:02:07,880 --> 00:02:11,760 Speaker 1: mangrove restoration for Indonesian shrimp farmers to the study of 29 00:02:11,880 --> 00:02:16,720 Speaker 1: nanoporous separations in the chemical industry, and it also thinks 30 00:02:16,840 --> 00:02:20,919 Speaker 1: hard about how to avoid AI being used to increase 31 00:02:20,919 --> 00:02:25,440 Speaker 1: emissions and worsen human suffering. I ask Pria about some 32 00:02:25,520 --> 00:02:28,880 Speaker 1: of the AI applications he's most excited about, and why 33 00:02:28,919 --> 00:02:32,720 Speaker 1: the conceptual framework we build around AI is just as 34 00:02:32,800 --> 00:02:48,119 Speaker 1: important as the technology itself. Now, before we get into 35 00:02:48,160 --> 00:02:50,280 Speaker 1: the heart of some of the work you do at 36 00:02:50,280 --> 00:02:53,800 Speaker 1: Climate Change AI, I think it would be helpful to 37 00:02:53,880 --> 00:02:57,480 Speaker 1: define the terms, because there's just so many of them, 38 00:02:57,800 --> 00:03:00,560 Speaker 1: and there's just a mixing and muddling when people think 39 00:03:00,600 --> 00:03:04,640 Speaker 1: about AI. For most people, the biggest point of entry 40 00:03:05,280 --> 00:03:09,480 Speaker 1: for AI is CHAD GPT. CHADGPT is what people have 41 00:03:09,560 --> 00:03:13,079 Speaker 1: played with people kind of know. It's based on this 42 00:03:13,080 --> 00:03:16,440 Speaker 1: thing called an LM, a large language model. But it's 43 00:03:16,520 --> 00:03:21,280 Speaker 1: just one example of types of AI. So if you 44 00:03:21,320 --> 00:03:24,120 Speaker 1: start at the very top, how would you define AI 45 00:03:24,280 --> 00:03:26,800 Speaker 1: and what types of AI are there? 46 00:03:27,240 --> 00:03:30,400 Speaker 2: Yeah, so there isn't kind of one universally agreed upon 47 00:03:30,560 --> 00:03:34,360 Speaker 2: definition for AI. But roughly you can think about AI 48 00:03:34,720 --> 00:03:38,440 Speaker 2: as referring to systems that perform some kind of complex task. 49 00:03:39,000 --> 00:03:43,040 Speaker 2: And there are two big branches of AI. One is 50 00:03:43,240 --> 00:03:47,440 Speaker 2: rule based systems, which is when you kind of know 51 00:03:47,880 --> 00:03:50,000 Speaker 2: how to do something, like you know how to play 52 00:03:50,160 --> 00:03:52,280 Speaker 2: chess in some sense, you could write down the rules, 53 00:03:52,680 --> 00:03:55,080 Speaker 2: but actually reasoning over those rules to figure out how 54 00:03:55,080 --> 00:03:57,320 Speaker 2: to be a good chess player is the hard part. 55 00:03:57,680 --> 00:03:59,800 Speaker 2: And so rule based systems are places where you write 56 00:03:59,840 --> 00:04:02,400 Speaker 2: down the rules and reason over them automatically. 57 00:04:03,000 --> 00:04:07,440 Speaker 1: Oh and that means Deep Blue beating Gary Kasperov for 58 00:04:07,480 --> 00:04:10,360 Speaker 1: the first time in nineteen ninety seven. That would be 59 00:04:10,440 --> 00:04:14,560 Speaker 1: classified technically as a rule based AI system. That's right, 60 00:04:14,920 --> 00:04:16,720 Speaker 1: even with those primitive computers. 61 00:04:16,800 --> 00:04:19,880 Speaker 2: Oh yes, so AI has been around for a long 62 00:04:19,920 --> 00:04:25,120 Speaker 2: time actually, and so another type of AI is machine learning, 63 00:04:25,480 --> 00:04:29,359 Speaker 2: and machine learning is often used in situations where you 64 00:04:29,440 --> 00:04:32,680 Speaker 2: might have intuition for something, but it's really hard to 65 00:04:32,720 --> 00:04:34,880 Speaker 2: write down rules to cadify your intuition. 66 00:04:35,320 --> 00:04:36,160 Speaker 3: So if I gave you. 67 00:04:36,160 --> 00:04:39,480 Speaker 2: Actually at an image of a dog, you could probably 68 00:04:39,480 --> 00:04:41,760 Speaker 2: tell me that it's a dog. But if I asked 69 00:04:41,760 --> 00:04:44,000 Speaker 2: you to write down a set of rules that characterize 70 00:04:44,040 --> 00:04:46,159 Speaker 2: exactly why it's a dog, and it would be really 71 00:04:46,240 --> 00:04:47,320 Speaker 2: hard for you to write. 72 00:04:47,080 --> 00:04:48,600 Speaker 3: Down that set of rules exactly. 73 00:04:49,160 --> 00:04:52,240 Speaker 2: And so machine learning is a paradigm where you actually 74 00:04:52,760 --> 00:04:57,360 Speaker 2: infer some of these rules automatically from examples or data. 75 00:04:57,400 --> 00:04:59,720 Speaker 2: So I give you a bunch of images, maybe I 76 00:04:59,800 --> 00:05:03,040 Speaker 2: tell which ones are dogs or cats, and the machine 77 00:05:03,080 --> 00:05:05,800 Speaker 2: learning algorithm learns how to map between the images and 78 00:05:05,839 --> 00:05:07,680 Speaker 2: the labels of is this a dog or cat and 79 00:05:07,800 --> 00:05:11,240 Speaker 2: kind of infer the rules that cause that to be true. 80 00:05:11,920 --> 00:05:14,839 Speaker 1: And so if it take the type of AI that 81 00:05:14,880 --> 00:05:19,640 Speaker 1: most people know, which is large language models, that's machine learning. 82 00:05:19,839 --> 00:05:23,760 Speaker 2: That is machine learning, and large language models are basically 83 00:05:24,880 --> 00:05:29,800 Speaker 2: one type of machine learning model that basically looks a 84 00:05:29,880 --> 00:05:33,760 Speaker 2: specific way, has a particular specification of how you update it, 85 00:05:34,240 --> 00:05:37,120 Speaker 2: and that type of model can be used in various 86 00:05:37,160 --> 00:05:40,320 Speaker 2: different ways, and roughly the three kinds of ways they 87 00:05:40,360 --> 00:05:43,840 Speaker 2: can be used are called supervised learning, unsupervised learning, and 88 00:05:43,880 --> 00:05:44,880 Speaker 2: reinforcement learning. 89 00:05:45,160 --> 00:05:48,919 Speaker 1: Well, it all sounds like you're trying to teach a child. 90 00:05:49,560 --> 00:05:54,560 Speaker 1: It's either through supervision, or through play, or through punishment. 91 00:05:56,320 --> 00:05:58,880 Speaker 2: Yeah, and in some sense, a lot of machine learning 92 00:05:59,160 --> 00:06:02,640 Speaker 2: algorithms and ways of trying to learn these things are 93 00:06:03,080 --> 00:06:05,279 Speaker 2: vaguely inspired by. 94 00:06:05,120 --> 00:06:07,919 Speaker 3: Some notion of how humans learn, although. 95 00:06:07,760 --> 00:06:10,080 Speaker 2: The practicalities of how we actually do this might be 96 00:06:10,160 --> 00:06:10,920 Speaker 2: quite different. 97 00:06:11,320 --> 00:06:14,479 Speaker 1: And so we talk about AI in the climate context 98 00:06:14,480 --> 00:06:18,560 Speaker 1: for two big reasons. One is because of the energy 99 00:06:18,640 --> 00:06:24,039 Speaker 1: cost of creating AI models and using AI models. And 100 00:06:24,120 --> 00:06:27,479 Speaker 1: second is that these models, again different types of them, 101 00:06:27,720 --> 00:06:31,600 Speaker 1: can have different applications that could make solving for climate 102 00:06:31,680 --> 00:06:33,880 Speaker 1: change deploying these solutions easier. 103 00:06:34,720 --> 00:06:37,359 Speaker 2: And I would add in a third pillar, which is 104 00:06:37,360 --> 00:06:41,240 Speaker 2: that AI is also used for many types of applications 105 00:06:41,279 --> 00:06:44,640 Speaker 2: that make solving climate harder. So when we talk about 106 00:06:44,680 --> 00:06:46,599 Speaker 2: the good and the bad, we should think about the 107 00:06:46,640 --> 00:06:47,240 Speaker 2: fact that. 108 00:06:47,160 --> 00:06:50,680 Speaker 3: AI has its own footprint. ANAI is used in both 109 00:06:50,720 --> 00:06:51,520 Speaker 3: good and bad ways. 110 00:06:51,720 --> 00:06:54,240 Speaker 1: Yeah, And so let's address the footprint part. Because the 111 00:06:54,279 --> 00:06:57,000 Speaker 1: size of footprint that may come from AI will be 112 00:06:57,120 --> 00:07:01,839 Speaker 1: dependent on the type of AI and large language models 113 00:07:01,880 --> 00:07:04,000 Speaker 1: are in the news because these are the models that 114 00:07:04,040 --> 00:07:07,400 Speaker 1: try and train themselves on the entire corpus of the Internet, 115 00:07:07,640 --> 00:07:10,640 Speaker 1: and that just requires a ton of computing power, which 116 00:07:10,720 --> 00:07:15,200 Speaker 1: is why companies like Microsoft and Alphabet and Meta are 117 00:07:15,240 --> 00:07:19,160 Speaker 1: all now in this rush to build more data centers, 118 00:07:19,640 --> 00:07:23,360 Speaker 1: consume a lot more power in the process and blow 119 00:07:23,400 --> 00:07:25,560 Speaker 1: past some of their own set climate goals. As we 120 00:07:25,640 --> 00:07:29,800 Speaker 1: found out with Microsoft's recent update that it's emissions are 121 00:07:29,800 --> 00:07:33,640 Speaker 1: actually thirty percent higher rather than thirty percent lower last year. 122 00:07:34,080 --> 00:07:37,440 Speaker 1: Does that mean all kinds of AI is doomed to 123 00:07:37,600 --> 00:07:40,080 Speaker 1: have a higher footprint because all kinds of AI will 124 00:07:40,240 --> 00:07:42,320 Speaker 1: want as much data as possible. 125 00:07:43,040 --> 00:07:46,200 Speaker 2: So there's definitely a diversity in the types of AI 126 00:07:46,240 --> 00:07:48,520 Speaker 2: that exist and as a result, the kind of energy 127 00:07:48,600 --> 00:07:49,480 Speaker 2: usage of these. 128 00:07:49,560 --> 00:07:52,240 Speaker 3: So there has been in long history. 129 00:07:52,320 --> 00:07:55,040 Speaker 2: You know, AI and machine learning models that use you know, 130 00:07:55,440 --> 00:07:58,160 Speaker 2: a reasonable amount of data, but much less than the 131 00:07:58,280 --> 00:08:02,600 Speaker 2: entirety of the Internet. The models themselves are also much smaller, 132 00:08:02,640 --> 00:08:05,080 Speaker 2: they have fewer parameters, and as a result, you don't 133 00:08:05,120 --> 00:08:08,080 Speaker 2: need as much computational power to actually update and get. 134 00:08:07,920 --> 00:08:08,880 Speaker 3: These models to learn. 135 00:08:09,640 --> 00:08:13,160 Speaker 2: And so, you know, some of the models that we develop, 136 00:08:13,240 --> 00:08:15,400 Speaker 2: even in my research group, can run on a laptop. 137 00:08:16,000 --> 00:08:17,560 Speaker 3: But then of course you have these you. 138 00:08:17,480 --> 00:08:22,280 Speaker 2: Know, large data intensive state of the art algorithms that 139 00:08:22,320 --> 00:08:25,720 Speaker 2: are kind of being deployed through products like chat, GPT, 140 00:08:26,440 --> 00:08:29,960 Speaker 2: and definitely the kind of energy consumption and you know, 141 00:08:30,360 --> 00:08:34,600 Speaker 2: water consumption from data centers, the materiality impacts of actually 142 00:08:34,640 --> 00:08:37,760 Speaker 2: getting the computational hardware in place, that is starting to 143 00:08:37,760 --> 00:08:38,680 Speaker 2: get worrying. 144 00:08:39,200 --> 00:08:43,640 Speaker 1: Right now, Climate applications themselves don't have to go down 145 00:08:43,800 --> 00:08:46,480 Speaker 1: the LLM route of having to consume that much data. 146 00:08:46,520 --> 00:08:49,120 Speaker 1: You know, you say models on a laptop can work. 147 00:08:49,640 --> 00:08:54,000 Speaker 1: Let's start with that. Because you got into AI through 148 00:08:54,360 --> 00:08:57,960 Speaker 1: trying to figure out how to make the grid work better, right. 149 00:08:58,000 --> 00:08:58,439 Speaker 3: That's right. 150 00:08:58,679 --> 00:09:02,400 Speaker 2: So basically, we as we start to integrate you know, 151 00:09:02,480 --> 00:09:07,520 Speaker 2: more and more renewables into power grids, many of these renewables, 152 00:09:07,520 --> 00:09:10,360 Speaker 2: their output varies based on the weather, so it varies 153 00:09:10,400 --> 00:09:13,280 Speaker 2: over time. Think about solar, think about wind. And yet 154 00:09:13,320 --> 00:09:16,120 Speaker 2: on a power grid you're having to maintain this exact, 155 00:09:16,240 --> 00:09:18,760 Speaker 2: delicate balance between how much power is put into the 156 00:09:18,760 --> 00:09:21,319 Speaker 2: grid and how much is consumed, which gets harder when 157 00:09:21,320 --> 00:09:23,559 Speaker 2: you have a lot of variations coming onto the grid. 158 00:09:24,000 --> 00:09:26,640 Speaker 2: And so AI and machine learning can be helpful in 159 00:09:26,720 --> 00:09:29,120 Speaker 2: terms of doing things like first, I mean just giving 160 00:09:29,160 --> 00:09:32,360 Speaker 2: us better predictions of what your solar power output, wind 161 00:09:32,360 --> 00:09:35,840 Speaker 2: power output, electricity demand will look like, but also in 162 00:09:35,880 --> 00:09:39,400 Speaker 2: actually helping to speed up some of the existing physics 163 00:09:39,400 --> 00:09:42,200 Speaker 2: based and engineering based algorithms that are used to manage 164 00:09:42,240 --> 00:09:44,760 Speaker 2: the power grid in the back end to maintain that balance. 165 00:09:45,200 --> 00:09:47,880 Speaker 1: And so one of the challenges with trying to understand 166 00:09:48,600 --> 00:09:50,640 Speaker 1: as an application to try and help solve some of 167 00:09:50,640 --> 00:09:55,319 Speaker 1: the climate problems is that it becomes really abstract very quickly. 168 00:09:55,440 --> 00:09:58,080 Speaker 1: So you say, oh, yeah, we have a number of 169 00:09:58,120 --> 00:10:00,680 Speaker 1: data points and there's an intelligent way in which we 170 00:10:00,720 --> 00:10:03,680 Speaker 1: can use them, and that gives us an output, but 171 00:10:03,760 --> 00:10:06,360 Speaker 1: we don't usually know why we have that output. But 172 00:10:06,400 --> 00:10:08,080 Speaker 1: that output is better, so we use it and that's 173 00:10:08,080 --> 00:10:11,560 Speaker 1: the solution, and it just does not feel satisfying, you know, 174 00:10:11,600 --> 00:10:15,160 Speaker 1: as a science reporter. To me, the joy off an 175 00:10:15,200 --> 00:10:18,720 Speaker 1: invention is to actually break down the steps to try 176 00:10:18,720 --> 00:10:21,200 Speaker 1: and figure out why this step led to that step, 177 00:10:21,280 --> 00:10:23,719 Speaker 1: let to that step, and finally you have something that 178 00:10:23,920 --> 00:10:27,880 Speaker 1: is really useful. Can we do that with AI? 179 00:10:29,200 --> 00:10:31,319 Speaker 2: Yeah, So I think that there are a couple of 180 00:10:31,960 --> 00:10:35,040 Speaker 2: categories of ways we can think about AI and machine 181 00:10:35,120 --> 00:10:37,679 Speaker 2: learning being used for climate that can help maybe give 182 00:10:37,679 --> 00:10:40,319 Speaker 2: a mental model for what's actually going on under the hood. 183 00:10:41,840 --> 00:10:45,480 Speaker 2: So one of these categories is you know, taking large 184 00:10:45,520 --> 00:10:49,559 Speaker 2: streams of broad data and distilling it into actionable information. 185 00:10:50,080 --> 00:10:52,720 Speaker 2: So one project we're funding through Climate Change AI is 186 00:10:53,280 --> 00:10:57,440 Speaker 2: actually a project that tries to improve the sustainability of 187 00:10:57,440 --> 00:11:02,520 Speaker 2: shrimp aquaculture practices. So kind of shrimp acoculture is currently 188 00:11:03,160 --> 00:11:05,679 Speaker 2: it can be harmful to you know, coastal mangrove forests, 189 00:11:05,679 --> 00:11:08,400 Speaker 2: and that has implications for climate change adaptation in terms 190 00:11:08,440 --> 00:11:11,160 Speaker 2: of kind of flood resilience as well as climate change 191 00:11:11,160 --> 00:11:14,640 Speaker 2: mitigation in terms of the sequestration potential of mangroves. And 192 00:11:14,679 --> 00:11:18,400 Speaker 2: so we're currently funding a team from Arizona State, Conservation 193 00:11:18,480 --> 00:11:22,199 Speaker 2: International and Thinking Machines Data Science from in the Philippines 194 00:11:22,640 --> 00:11:28,480 Speaker 2: to actually use satellite imagery to assess aquaculture farms that 195 00:11:28,559 --> 00:11:32,480 Speaker 2: actually might be able to benefit from better aquaculture practices. 196 00:11:32,720 --> 00:11:36,280 Speaker 2: The intervention here is that you can actually do things 197 00:11:36,360 --> 00:11:39,280 Speaker 2: like if you have an aquaculture farm, you can intensify 198 00:11:39,400 --> 00:11:41,880 Speaker 2: how much you're farming on one part of the farm, 199 00:11:41,920 --> 00:11:43,960 Speaker 2: and then you can kind of conserve on another part 200 00:11:43,960 --> 00:11:47,439 Speaker 2: of the farm and so without impacting your overall productivity, 201 00:11:47,480 --> 00:11:49,280 Speaker 2: you can just farm in a way that's better for 202 00:11:49,360 --> 00:11:53,559 Speaker 2: the mangroves. And so Conservation International has a program where 203 00:11:53,559 --> 00:11:56,640 Speaker 2: they're working with farmers to try to kind of help 204 00:11:56,720 --> 00:11:59,840 Speaker 2: them do this. But actually identifying which farms are amenable 205 00:11:59,840 --> 00:12:02,400 Speaker 2: to this type of intervention at scale is difficult, So 206 00:12:02,440 --> 00:12:06,439 Speaker 2: they use a combination of you know, satellite imagery data 207 00:12:06,520 --> 00:12:09,040 Speaker 2: on like sea level rise and sea risk and things 208 00:12:09,080 --> 00:12:12,959 Speaker 2: like this in order to then actually pinpoint at scale 209 00:12:13,160 --> 00:12:15,560 Speaker 2: which farms might be amenable to this CT invention and 210 00:12:15,559 --> 00:12:17,160 Speaker 2: then actually go work with them to do that. 211 00:12:17,280 --> 00:12:19,959 Speaker 1: And you said that was just one approach, what are 212 00:12:20,000 --> 00:12:21,400 Speaker 1: some other approaches? 213 00:12:21,720 --> 00:12:23,360 Speaker 3: There's a couple of other ways actually. 214 00:12:23,400 --> 00:12:26,840 Speaker 2: So one is you know, predicting and forecasting, so taking 215 00:12:27,000 --> 00:12:30,480 Speaker 2: you know, historical data where you have relationships between some 216 00:12:30,679 --> 00:12:33,280 Speaker 2: input and some quantity you would want to predict. So 217 00:12:33,520 --> 00:12:36,679 Speaker 2: things like I want to predict electricity demand on the 218 00:12:36,720 --> 00:12:40,160 Speaker 2: power grid, so I can take historical data about what 219 00:12:40,240 --> 00:12:43,880 Speaker 2: electricity demand looked like. I can take historical weather data 220 00:12:44,360 --> 00:12:47,120 Speaker 2: and I can learn relationships between those so that in 221 00:12:47,160 --> 00:12:49,400 Speaker 2: the future, when I have a weather prediction but I 222 00:12:49,440 --> 00:12:51,520 Speaker 2: don't know what the electricity demand would be based on 223 00:12:51,520 --> 00:12:54,160 Speaker 2: that weather prediction, I can just go ahead and predict 224 00:12:54,200 --> 00:12:56,800 Speaker 2: that and you have kind of for example, nonprofits like 225 00:12:56,840 --> 00:13:00,080 Speaker 2: open Climate Fix that are working with the UK Powers 226 00:13:00,280 --> 00:13:03,920 Speaker 2: Operator to actually improve their electricity demand forecasts and they've 227 00:13:03,920 --> 00:13:06,080 Speaker 2: been able to use machine learning to have the error 228 00:13:06,320 --> 00:13:07,600 Speaker 2: of those forecasts. 229 00:13:09,880 --> 00:13:12,960 Speaker 1: After the break, why it's important for all of us 230 00:13:13,000 --> 00:13:16,920 Speaker 1: to be involved in the development of AI. By the way, 231 00:13:17,000 --> 00:13:19,520 Speaker 1: if you're enjoying this episode, please do take a moment 232 00:13:19,559 --> 00:13:22,239 Speaker 1: to rate and review the show on Apple Podcasts or Spotify. 233 00:13:22,720 --> 00:13:34,400 Speaker 1: It helps other listeners find it. Some companies, one that 234 00:13:34,480 --> 00:13:38,319 Speaker 1: my colleague wrote about called Climate AI is using AI 235 00:13:38,400 --> 00:13:41,600 Speaker 1: to try and improve weather prediction models because currently you 236 00:13:41,640 --> 00:13:45,720 Speaker 1: are starting to get better and better predictions, and that 237 00:13:45,760 --> 00:13:49,559 Speaker 1: has at least for them, been a profitable enterprise because 238 00:13:49,559 --> 00:13:53,600 Speaker 1: then they're working with these large agriculture companies that want 239 00:13:53,600 --> 00:13:56,040 Speaker 1: to figure out when should we start to put the 240 00:13:56,080 --> 00:13:58,280 Speaker 1: seed down or when should we start to harvest because 241 00:13:58,320 --> 00:14:02,200 Speaker 1: we have a better understanding of the weather not just 242 00:14:02,320 --> 00:14:04,440 Speaker 1: over the next two weeks, but over the next three months. 243 00:14:04,920 --> 00:14:07,679 Speaker 2: Yeah. So I think this idea of kind of medium 244 00:14:07,720 --> 00:14:11,439 Speaker 2: to long term forecasting is also really cool, but often 245 00:14:11,559 --> 00:14:14,160 Speaker 2: to do good forecasting. In these settings, you want to 246 00:14:14,240 --> 00:14:19,080 Speaker 2: use a combination of physical models and data. So, for example, 247 00:14:19,400 --> 00:14:21,760 Speaker 2: one of the teams that we're funding at Climate Change 248 00:14:21,760 --> 00:14:25,160 Speaker 2: AI spread between a couple of US universities and an 249 00:14:25,200 --> 00:14:28,800 Speaker 2: Indian university. They're basically trying to figure out how do 250 00:14:28,880 --> 00:14:33,600 Speaker 2: I actually make longer term predictions of weather in order 251 00:14:33,640 --> 00:14:36,480 Speaker 2: to foster how we actually build out power grands for 252 00:14:36,520 --> 00:14:39,800 Speaker 2: the future. And the difficulty here is if you just 253 00:14:40,000 --> 00:14:43,280 Speaker 2: use past data. What machine learning does is it learns 254 00:14:43,360 --> 00:14:46,120 Speaker 2: patterns in that past data and just projects them forward. 255 00:14:47,280 --> 00:14:49,920 Speaker 2: But the climate is changing, which means that the patterns 256 00:14:49,960 --> 00:14:52,920 Speaker 2: in how weather is occurring are changing, and so you 257 00:14:52,960 --> 00:14:55,280 Speaker 2: can't just use a pure data driven technique to do this. 258 00:14:55,680 --> 00:14:58,760 Speaker 2: And so what this team does is they say, well, 259 00:14:58,800 --> 00:15:01,960 Speaker 2: we have climate models. The issue with climate models is 260 00:15:01,960 --> 00:15:05,280 Speaker 2: that they don't give you very granular information on exactly 261 00:15:05,320 --> 00:15:07,400 Speaker 2: what's going to happen at a particular place, just because 262 00:15:07,400 --> 00:15:11,560 Speaker 2: they're very computationally intensive to run. But if we can 263 00:15:11,800 --> 00:15:14,960 Speaker 2: quote unquote back cast the climate models, so run and 264 00:15:15,040 --> 00:15:17,680 Speaker 2: say what would the climate model say now, and we 265 00:15:17,760 --> 00:15:21,600 Speaker 2: already have really fine grained weather data historically, we can 266 00:15:21,680 --> 00:15:25,040 Speaker 2: learn a mapping between what the climate model said, and 267 00:15:25,040 --> 00:15:27,800 Speaker 2: what the weather data would be. And then in the future, 268 00:15:27,800 --> 00:15:30,400 Speaker 2: where we only have a climate model prediction, we can 269 00:15:30,520 --> 00:15:32,520 Speaker 2: use our learned relationship to say, oh, and this is 270 00:15:32,520 --> 00:15:34,360 Speaker 2: what the weather would be in a more fine grained 271 00:15:34,360 --> 00:15:36,600 Speaker 2: way in the future. With a lot of these kind 272 00:15:36,600 --> 00:15:39,520 Speaker 2: of climate downscaling techniques, you often want to think about 273 00:15:39,600 --> 00:15:42,040 Speaker 2: who is the user of these techniques and as a result, 274 00:15:42,080 --> 00:15:45,040 Speaker 2: what aspects of your downscaled predictions have to be good. 275 00:15:45,320 --> 00:15:47,440 Speaker 2: So here they're actually doing this for the power grid 276 00:15:47,480 --> 00:15:50,200 Speaker 2: planning context, where they're saying, can we produce fine grained 277 00:15:50,240 --> 00:15:52,720 Speaker 2: data sets of what electricity usage will look like, wind 278 00:15:52,760 --> 00:15:55,360 Speaker 2: power production, and solar power production might look like in 279 00:15:55,440 --> 00:15:57,160 Speaker 2: order to facilitate power grid planning. 280 00:15:57,520 --> 00:16:01,160 Speaker 1: All of this sort of was something that you published 281 00:16:01,160 --> 00:16:04,880 Speaker 1: in a paper title tackling Climate Change with Machine Learning. 282 00:16:05,240 --> 00:16:06,880 Speaker 1: Why do you need to write this paper? 283 00:16:08,240 --> 00:16:11,680 Speaker 2: Yeah, so, I'd say back in twenty nineteen, we definitely 284 00:16:11,720 --> 00:16:14,320 Speaker 2: saw a combination of a lot of people in the 285 00:16:14,360 --> 00:16:16,840 Speaker 2: AI and machine learning space who wanted to leverage their 286 00:16:16,840 --> 00:16:20,400 Speaker 2: skills to help in facilitating climate action but didn't necessarily 287 00:16:20,440 --> 00:16:23,640 Speaker 2: know how. And on the other side, many people in 288 00:16:23,680 --> 00:16:26,600 Speaker 2: the climate change related space who are seeing you know, 289 00:16:26,680 --> 00:16:29,400 Speaker 2: things like larger streams of data becoming available and saying, 290 00:16:29,400 --> 00:16:32,440 Speaker 2: but how do I actually utilize this? And so we 291 00:16:32,560 --> 00:16:35,840 Speaker 2: really felt like there was a need to really put 292 00:16:35,960 --> 00:16:38,080 Speaker 2: fore in for the community. You know, where is it 293 00:16:38,160 --> 00:16:41,000 Speaker 2: that AI is well matched to climate change related problems 294 00:16:41,040 --> 00:16:43,280 Speaker 2: in order to then help AI people get into the 295 00:16:43,320 --> 00:16:46,120 Speaker 2: space and to help climate people understand okay for some 296 00:16:46,160 --> 00:16:49,440 Speaker 2: of these complex problems we're seeing, is the bottleneck potentially 297 00:16:50,120 --> 00:16:53,640 Speaker 2: solvable VIAI and machine learning. And there were two kind 298 00:16:53,640 --> 00:16:55,680 Speaker 2: of big aims through that work. One is again to 299 00:16:55,760 --> 00:16:58,680 Speaker 2: lay out the space of applications, but the second. 300 00:16:58,360 --> 00:17:00,400 Speaker 3: Is to try to provide some of. 301 00:17:02,120 --> 00:17:04,600 Speaker 2: Mental model and guidance for how to do this work 302 00:17:04,720 --> 00:17:10,240 Speaker 2: in a sound, impactful and responsible way, because there are 303 00:17:10,240 --> 00:17:13,040 Speaker 2: lots of places where AI is not the right fit 304 00:17:13,520 --> 00:17:16,560 Speaker 2: and it can be a huge distraction or there are 305 00:17:16,560 --> 00:17:19,679 Speaker 2: ways that for example, because you often have you know, 306 00:17:19,840 --> 00:17:24,159 Speaker 2: data and computational power concentrated in certain geographies versus others, 307 00:17:24,320 --> 00:17:26,920 Speaker 2: where the practice of AI can exacerbate some of these 308 00:17:26,920 --> 00:17:30,239 Speaker 2: inequities by basically causing people who already have access to 309 00:17:30,440 --> 00:17:32,320 Speaker 2: compute data to be able to do a lot more 310 00:17:32,359 --> 00:17:35,240 Speaker 2: and leave others behind. And so there's a lot kind 311 00:17:35,280 --> 00:17:37,119 Speaker 2: of in there to make sure we're actually moving the 312 00:17:37,600 --> 00:17:41,120 Speaker 2: space forward in a way that makes sense for climate 313 00:17:41,280 --> 00:17:42,080 Speaker 2: and for equity. 314 00:17:42,920 --> 00:17:43,120 Speaker 3: Yeah. 315 00:17:43,160 --> 00:17:46,960 Speaker 1: So recently we spoke to the president of Microsoft, Brad Smith, 316 00:17:47,160 --> 00:17:51,240 Speaker 1: and he was talking about how he would like AI 317 00:17:51,280 --> 00:17:54,400 Speaker 1: to be available to everybody. He doesn't want AI haves 318 00:17:54,720 --> 00:17:58,240 Speaker 1: and AI have nots, And I interpreted that to mean, 319 00:17:59,160 --> 00:18:01,960 Speaker 1: you know, we've had technological leaps in the past, and 320 00:18:02,000 --> 00:18:05,200 Speaker 1: when they have been more widely available that has been 321 00:18:05,240 --> 00:18:11,480 Speaker 1: beneficial to humanity mobile phones, Internet. Do you see AI 322 00:18:11,560 --> 00:18:19,200 Speaker 1: as being essential for unlocking human potential like Microsoft president 323 00:18:19,280 --> 00:18:19,680 Speaker 1: is saying? 324 00:18:20,040 --> 00:18:22,080 Speaker 3: So, I think there are two things I'd like to 325 00:18:22,160 --> 00:18:22,920 Speaker 3: unpack in there. 326 00:18:23,359 --> 00:18:28,440 Speaker 2: One, I think that AI can be a really powerful 327 00:18:28,640 --> 00:18:33,000 Speaker 2: kind of support and accelerator for many different climate change 328 00:18:33,000 --> 00:18:34,880 Speaker 2: related applications and others. 329 00:18:35,440 --> 00:18:38,040 Speaker 3: There are some where I think it is essential. 330 00:18:38,280 --> 00:18:41,760 Speaker 2: For example, don't know how we will manage power grids 331 00:18:41,800 --> 00:18:44,560 Speaker 2: with lots of variability and large amounts of renewables without AI. 332 00:18:45,119 --> 00:18:47,560 Speaker 2: There are other places where it can be helpful, but 333 00:18:47,600 --> 00:18:50,960 Speaker 2: I don't necessarily think it's the critical bottleneck. One thing 334 00:18:51,000 --> 00:18:55,080 Speaker 2: I'd like to also mention is that with certain things 335 00:18:55,280 --> 00:18:56,399 Speaker 2: like mobile. 336 00:18:56,040 --> 00:18:57,520 Speaker 3: Technologies, and Stutch. 337 00:18:59,000 --> 00:19:02,520 Speaker 2: Democratization has been used to sort of indicate, Okay, a 338 00:19:02,560 --> 00:19:05,000 Speaker 2: few people created a thing and it was pushed onto 339 00:19:05,040 --> 00:19:08,840 Speaker 2: the rest of the world. That's not actually in some 340 00:19:08,920 --> 00:19:13,480 Speaker 2: ways democratization, especially in the context of AI, where actually 341 00:19:13,520 --> 00:19:16,040 Speaker 2: the type of AI you build and that the way 342 00:19:16,080 --> 00:19:19,240 Speaker 2: you do it it fundamentally needs to look very different 343 00:19:19,280 --> 00:19:21,840 Speaker 2: depending on the context you're in. You have different amounts 344 00:19:21,880 --> 00:19:23,960 Speaker 2: of data in different contexts, you have different amounts of 345 00:19:23,960 --> 00:19:27,000 Speaker 2: compute in different contexts, you have different amounts of kind 346 00:19:27,000 --> 00:19:30,560 Speaker 2: of existing knowledge that can be integrated into systems. And 347 00:19:30,640 --> 00:19:33,320 Speaker 2: so if we kind of develop AI among a small 348 00:19:33,320 --> 00:19:35,600 Speaker 2: set of entities and then push that onto the rest 349 00:19:35,600 --> 00:19:37,399 Speaker 2: of the world, it's actually not going to serve the 350 00:19:37,480 --> 00:19:40,440 Speaker 2: needs of the full world. And so democratization really means 351 00:19:40,560 --> 00:19:43,520 Speaker 2: enabling more people to contribute to the trajectory of actually 352 00:19:43,600 --> 00:19:46,720 Speaker 2: developing AI, not just sort of being users of a 353 00:19:46,800 --> 00:19:48,640 Speaker 2: product that a few people developed. 354 00:19:48,920 --> 00:19:51,720 Speaker 1: On the other side, can you give a specific way 355 00:19:51,960 --> 00:19:55,960 Speaker 1: in which that might play out, say through the development 356 00:19:55,960 --> 00:19:58,679 Speaker 1: of CHA, GPT or cloud or these other types of 357 00:19:58,680 --> 00:19:59,880 Speaker 1: generative AI products. 358 00:20:00,359 --> 00:20:04,560 Speaker 2: Yes, So basically, if you think about something like GPT, 359 00:20:05,040 --> 00:20:07,560 Speaker 2: it needs a huge amount of data to train. It 360 00:20:07,600 --> 00:20:11,679 Speaker 2: needs a huge amount of compute to run, and most 361 00:20:11,880 --> 00:20:14,119 Speaker 2: entities in the world do not have the ability to 362 00:20:14,200 --> 00:20:16,800 Speaker 2: curate or collect that amount of data, nor do they 363 00:20:16,800 --> 00:20:19,639 Speaker 2: have the ability to pay for or procure the amount 364 00:20:19,680 --> 00:20:20,399 Speaker 2: of compute needed to. 365 00:20:20,400 --> 00:20:22,920 Speaker 3: Run those models. So those models are being. 366 00:20:22,760 --> 00:20:25,360 Speaker 2: Developed by a small set of people and then kind 367 00:20:25,400 --> 00:20:28,000 Speaker 2: of packaged and sent out in a kind of interface 368 00:20:28,080 --> 00:20:31,720 Speaker 2: like CHATDPT that many people can use, and that can 369 00:20:31,760 --> 00:20:34,840 Speaker 2: be helpful for a certain set of use cases, but 370 00:20:35,800 --> 00:20:37,880 Speaker 2: there are lots of use cases that don't necessarily fit 371 00:20:37,920 --> 00:20:44,080 Speaker 2: that mold. Imagine that you're trying to train your own weather. 372 00:20:43,880 --> 00:20:45,400 Speaker 3: Prediction model. 373 00:20:46,720 --> 00:20:49,480 Speaker 2: In a situation where you have some amount of data 374 00:20:49,560 --> 00:20:52,880 Speaker 2: and also some amount of knowledge of just how kind of. 375 00:20:52,840 --> 00:20:56,080 Speaker 3: Weather physics works. If you do this in a. 376 00:20:56,000 --> 00:20:59,560 Speaker 2: Fully data driven way, there does exist the reality in 377 00:20:59,560 --> 00:21:02,000 Speaker 2: which you're well to just purely from data figure out 378 00:21:02,040 --> 00:21:05,600 Speaker 2: how to predict weather. But you often need much more 379 00:21:05,720 --> 00:21:09,439 Speaker 2: data and a much bigger model if you're basically not 380 00:21:09,760 --> 00:21:12,320 Speaker 2: embedding the rules of physics and as a result learning 381 00:21:12,359 --> 00:21:14,800 Speaker 2: them fully from scratch, And so that leads to a 382 00:21:14,840 --> 00:21:17,280 Speaker 2: situation where you again have a bigger model that fewer 383 00:21:17,280 --> 00:21:19,720 Speaker 2: people can use and fewer people can train, and it 384 00:21:19,760 --> 00:21:22,320 Speaker 2: can also lead to situations where people say, oh, I 385 00:21:22,359 --> 00:21:23,919 Speaker 2: don't have a lot of data. Is the thing I'm 386 00:21:23,920 --> 00:21:25,960 Speaker 2: supposed to do collect tons and tons of data, so 387 00:21:25,960 --> 00:21:29,880 Speaker 2: they invest a bunch of money into setting up data 388 00:21:29,920 --> 00:21:33,400 Speaker 2: infrastructure and data collection. On the other side, though. 389 00:21:34,400 --> 00:21:36,520 Speaker 1: And this is a real limitation because if you looked 390 00:21:36,520 --> 00:21:41,119 Speaker 1: at the map of weather stations in the world, it 391 00:21:41,320 --> 00:21:44,160 Speaker 1: maps kind of one to one to the wealth there 392 00:21:44,280 --> 00:21:47,600 Speaker 1: is in the world. America and Europe is littered with 393 00:21:47,720 --> 00:21:52,399 Speaker 1: weather stations, whereas Africa is empty. And so if you 394 00:21:52,520 --> 00:21:55,480 Speaker 1: go down that route, the answer would be just deploying 395 00:21:55,480 --> 00:21:58,080 Speaker 1: more weather stations. But it is in the right answer. 396 00:21:58,280 --> 00:22:00,920 Speaker 2: Yeah, I mean, and in some sense, fixing the data 397 00:22:00,920 --> 00:22:03,320 Speaker 2: in equity problem is obviously great thing. It would be 398 00:22:03,320 --> 00:22:05,479 Speaker 2: great to have more weather stations in Africa than there 399 00:22:05,480 --> 00:22:09,119 Speaker 2: are today. But there are kind of additional ways to 400 00:22:09,119 --> 00:22:11,879 Speaker 2: contend with this problem, which include take the data you 401 00:22:12,000 --> 00:22:14,879 Speaker 2: have and take some knowledge of the physical rules that 402 00:22:14,960 --> 00:22:18,080 Speaker 2: govern weather, combine them together in a clever way so 403 00:22:18,160 --> 00:22:21,440 Speaker 2: you don't need as much data to still get good answers, 404 00:22:21,800 --> 00:22:24,400 Speaker 2: And so that really informs how you think about as 405 00:22:24,880 --> 00:22:27,840 Speaker 2: an organization, as a country where you invest your resources. 406 00:22:27,880 --> 00:22:30,880 Speaker 2: If you're just assuming that you invest them in collecting 407 00:22:30,880 --> 00:22:34,000 Speaker 2: a maximal amount of data necessary, actually that might actually 408 00:22:34,000 --> 00:22:36,680 Speaker 2: be a misinvestment of resources if you assume that AI 409 00:22:37,119 --> 00:22:41,040 Speaker 2: just means maximal data collection and learning only on data. 410 00:22:41,119 --> 00:22:41,880 Speaker 3: In addition to. 411 00:22:41,880 --> 00:22:45,240 Speaker 2: Sort of taking in data and producing insight, there are 412 00:22:45,240 --> 00:22:48,159 Speaker 2: also situations in which AI and machine learning can actually 413 00:22:48,200 --> 00:22:52,960 Speaker 2: help us to more efficiently optimize a complex system in order. 414 00:22:52,800 --> 00:22:54,080 Speaker 3: To improve its efficiency. 415 00:22:54,520 --> 00:22:58,320 Speaker 2: So, for example, if we think about buildings, there are 416 00:22:58,359 --> 00:23:01,040 Speaker 2: lots of ways in which we can actually better control, 417 00:23:01,200 --> 00:23:04,440 Speaker 2: for example, the heating and cooling systems in buildings, both 418 00:23:04,480 --> 00:23:07,000 Speaker 2: to kind of reduce the amount of energy they're actually 419 00:23:07,119 --> 00:23:09,960 Speaker 2: used in while kind of maintaining something like thermal comfort 420 00:23:09,960 --> 00:23:13,240 Speaker 2: in the building, and also be responsive to things like 421 00:23:13,280 --> 00:23:15,920 Speaker 2: how much renewable energy is actually available on the grid 422 00:23:15,960 --> 00:23:17,000 Speaker 2: at this particular time. 423 00:23:17,080 --> 00:23:18,520 Speaker 3: This concept of demand response. 424 00:23:18,960 --> 00:23:21,720 Speaker 2: What's kind of interesting is when you start to think 425 00:23:21,760 --> 00:23:25,000 Speaker 2: not just about individual building performance, but also how this 426 00:23:25,080 --> 00:23:28,199 Speaker 2: connects up to the power grid and when renewable energy 427 00:23:28,240 --> 00:23:30,840 Speaker 2: is available. You sometimes want to start thinking about this 428 00:23:30,920 --> 00:23:33,320 Speaker 2: not just at the individual building level, but for example, 429 00:23:33,359 --> 00:23:35,600 Speaker 2: at the neighborhood level, where you actually might want to 430 00:23:35,600 --> 00:23:38,440 Speaker 2: co optimize what's going on in different buildings to jointly 431 00:23:38,480 --> 00:23:41,439 Speaker 2: be doing the best thing for overall efficiency. 432 00:23:40,880 --> 00:23:41,960 Speaker 3: And the power grid. 433 00:23:41,960 --> 00:23:44,520 Speaker 2: And so one of the projects we're funding through Climate 434 00:23:44,600 --> 00:23:47,000 Speaker 2: Change AI is called the City Learn Challenge, and they 435 00:23:47,000 --> 00:23:52,160 Speaker 2: actually created a simulation environment that actually tries to provide 436 00:23:52,200 --> 00:23:54,880 Speaker 2: some structure of Okay, there are a bunch of buildings, 437 00:23:54,920 --> 00:23:57,520 Speaker 2: they're connected up to a neighborhood grid in this particular way. 438 00:23:57,600 --> 00:24:00,360 Speaker 2: Here's some data on how they're consuming energy. And they're 439 00:24:00,359 --> 00:24:02,720 Speaker 2: putting this forward as a challenge to the machine learning 440 00:24:02,720 --> 00:24:04,720 Speaker 2: community to say, can you come up with better ways 441 00:24:04,760 --> 00:24:07,600 Speaker 2: to actually optimize this neighborhood to improve its energy efficiency. 442 00:24:08,720 --> 00:24:10,880 Speaker 1: Yeah, that is cool. I feel like one other thing 443 00:24:10,960 --> 00:24:15,960 Speaker 1: that I could be helping is speeding up innovation with 444 00:24:16,080 --> 00:24:19,800 Speaker 1: these solutions in places where otherwise you would have required 445 00:24:19,840 --> 00:24:24,000 Speaker 1: more time, more skills, more people with the skills, especially 446 00:24:24,040 --> 00:24:26,280 Speaker 1: in developing countries where you really want to speed up 447 00:24:26,280 --> 00:24:29,439 Speaker 1: the solution set. I could allow for these sort of 448 00:24:29,480 --> 00:24:33,280 Speaker 1: optimization techniques to come through more quickly than it would 449 00:24:33,320 --> 00:24:34,480 Speaker 1: otherwise have done. 450 00:24:35,280 --> 00:24:35,600 Speaker 3: Yeah. 451 00:24:35,680 --> 00:24:38,440 Speaker 2: So across the projects that we kind of are funding 452 00:24:38,480 --> 00:24:41,240 Speaker 2: and kind of facilitating through Climate Change AI, they are 453 00:24:42,160 --> 00:24:46,119 Speaker 2: happening all around the world. So, for example, one of 454 00:24:46,160 --> 00:24:50,240 Speaker 2: the projects we're funding is a team of researchers working 455 00:24:50,280 --> 00:24:54,439 Speaker 2: with the Government of Fiji to actually better map the 456 00:24:54,560 --> 00:24:57,240 Speaker 2: damages from floods that occur in Fiji in order to 457 00:24:57,280 --> 00:25:01,480 Speaker 2: facilitate Fijian disaster response efforts. The idea being that when 458 00:25:01,520 --> 00:25:03,639 Speaker 2: you actually are trying to figure out, Okay, in a flood, 459 00:25:03,800 --> 00:25:07,120 Speaker 2: what happened, who was affected, it's really hard to kind 460 00:25:07,119 --> 00:25:09,920 Speaker 2: of systematically and fully collect that on the ground data. 461 00:25:10,440 --> 00:25:13,399 Speaker 2: And so one of the teams that we're funding is 462 00:25:13,440 --> 00:25:17,200 Speaker 2: actually alongside the Government of Fiji developing algorithms to kind 463 00:25:17,200 --> 00:25:21,560 Speaker 2: of map from satellite imagery to targeted information about what 464 00:25:21,600 --> 00:25:23,680 Speaker 2: the impacts were after a flood, and to be able 465 00:25:23,680 --> 00:25:26,760 Speaker 2: to kind of continuously update these maps based on satellite 466 00:25:26,760 --> 00:25:29,920 Speaker 2: imagery in order to aid disaster response efforts. So that's 467 00:25:29,920 --> 00:25:31,840 Speaker 2: one example, but a lot of this work is going 468 00:25:31,920 --> 00:25:34,360 Speaker 2: on all around the world. 469 00:25:34,680 --> 00:25:36,879 Speaker 1: And so going back to the start of the conversation 470 00:25:36,960 --> 00:25:40,200 Speaker 1: where you said, there's also how you can use those 471 00:25:40,240 --> 00:25:44,480 Speaker 1: same tools but to actually increase emissions. You could optimize 472 00:25:44,520 --> 00:25:48,320 Speaker 1: for how you can extract oil and gas in a 473 00:25:48,400 --> 00:25:51,720 Speaker 1: cheaper way, or go to places that previously were not 474 00:25:52,160 --> 00:25:56,439 Speaker 1: found or not reachable. Is that the biggest concern is 475 00:25:56,440 --> 00:25:59,720 Speaker 1: that the biggest downside of AI, even more so than 476 00:25:59,840 --> 00:26:00,760 Speaker 1: the resource use. 477 00:26:01,520 --> 00:26:04,439 Speaker 2: Yeah, to me, I think that we obviously need to 478 00:26:04,440 --> 00:26:07,879 Speaker 2: be thinking about both the resource use and the applications. 479 00:26:08,160 --> 00:26:10,679 Speaker 2: But the applications are very concerning to me because I 480 00:26:10,680 --> 00:26:14,240 Speaker 2: think they're having an outsized negative impact some of these 481 00:26:14,240 --> 00:26:18,440 Speaker 2: applications while also not being centered in the conversations about 482 00:26:18,440 --> 00:26:20,760 Speaker 2: how we actually align the use of AI with climate action. 483 00:26:21,560 --> 00:26:24,280 Speaker 2: So oil and gas is one example, but there are 484 00:26:24,280 --> 00:26:27,240 Speaker 2: other things like you know, AI being the driver behind 485 00:26:27,280 --> 00:26:30,920 Speaker 2: targeted advertising and increases of consumption in ways that don't 486 00:26:30,920 --> 00:26:32,360 Speaker 2: always make us happier but do. 487 00:26:32,400 --> 00:26:33,720 Speaker 3: Increase our resource use. 488 00:26:35,160 --> 00:26:39,040 Speaker 2: AI also drives in many ways the information that we 489 00:26:39,119 --> 00:26:42,560 Speaker 2: actually consume online, and that has really a lot of 490 00:26:42,560 --> 00:26:46,840 Speaker 2: ties to the spread of climate information or misinformation in 491 00:26:46,880 --> 00:26:49,800 Speaker 2: ways that could be harmful or helpful, depending on how 492 00:26:49,800 --> 00:26:53,800 Speaker 2: we're actually shaping those particular trends of AI induced information spread. 493 00:26:53,960 --> 00:26:58,080 Speaker 2: And then there are also things like AI for autonomous vehicles, 494 00:26:58,080 --> 00:27:00,880 Speaker 2: which we don't often talk about in the txt of climate, 495 00:27:01,240 --> 00:27:04,800 Speaker 2: but where the choices we're making are affecting the transportation 496 00:27:04,960 --> 00:27:07,479 Speaker 2: sector in ways that could be good or bad for 497 00:27:07,520 --> 00:27:13,480 Speaker 2: the climate. If you're kind of facilitating private fossil fuel transportation, 498 00:27:13,680 --> 00:27:17,320 Speaker 2: then you're potentially increasing energy usage and emissions, Whereas if 499 00:27:17,320 --> 00:27:21,800 Speaker 2: you're using autonomous vehicles to facilitate public multimodal transit, you're 500 00:27:21,840 --> 00:27:25,720 Speaker 2: potentially bringing the emissions of the sector down. So I 501 00:27:25,760 --> 00:27:29,000 Speaker 2: think the applications really can have an outsized impact and 502 00:27:29,119 --> 00:27:31,440 Speaker 2: it's really important to not leave them out of the conversation. 503 00:27:32,560 --> 00:27:35,280 Speaker 1: And my exposure to AI actually went back a decade 504 00:27:35,280 --> 00:27:37,520 Speaker 1: when I was in grad school at Oxford, and it 505 00:27:37,560 --> 00:27:40,080 Speaker 1: wasn't really the models or the applications, but it was 506 00:27:40,119 --> 00:27:43,000 Speaker 1: the ethics. There was a lot of conversations that were 507 00:27:43,000 --> 00:27:46,199 Speaker 1: happening around the ethics of how you would put AI 508 00:27:46,280 --> 00:27:49,600 Speaker 1: to use. Do you think we're doing substantial work on 509 00:27:49,720 --> 00:27:54,800 Speaker 1: the ethical side to ensure that the applications are beneficial 510 00:27:54,800 --> 00:27:58,040 Speaker 1: to humanity or are we just in this race to 511 00:27:58,119 --> 00:28:01,160 Speaker 1: develop new AI products and have kind of forgotten that 512 00:28:01,200 --> 00:28:03,800 Speaker 1: there are huge ethical implications here. 513 00:28:04,359 --> 00:28:07,560 Speaker 2: So ethics is a really, really important part of the conversation, 514 00:28:07,760 --> 00:28:09,520 Speaker 2: and I think there's been a lot of great work 515 00:28:09,560 --> 00:28:11,840 Speaker 2: done on it, but there's a lot more that needs 516 00:28:11,880 --> 00:28:15,000 Speaker 2: to be done. So you have things like UNESCO's AI 517 00:28:15,080 --> 00:28:18,600 Speaker 2: Ethics recommendations, which were actually you know, adopted very widely. 518 00:28:18,680 --> 00:28:21,280 Speaker 2: We're really extensive in terms of thinking about things like 519 00:28:21,520 --> 00:28:26,560 Speaker 2: you know, bias, equity, privacy, transparency, environmental impact, which I 520 00:28:26,560 --> 00:28:29,919 Speaker 2: would also count as a part of ethics. And so 521 00:28:30,000 --> 00:28:32,359 Speaker 2: I think there's been some really great thinking done on this, 522 00:28:32,560 --> 00:28:35,679 Speaker 2: but that there's a lot more that needs to be 523 00:28:35,680 --> 00:28:39,600 Speaker 2: done to sort of operationalize this and also incentivize people 524 00:28:39,640 --> 00:28:42,240 Speaker 2: to actually do work in the ethical way rather than 525 00:28:42,360 --> 00:28:44,520 Speaker 2: the way that kind of leaves ethics behind and just 526 00:28:44,600 --> 00:28:46,840 Speaker 2: you know, you run forward. So when we talk about 527 00:28:46,880 --> 00:28:52,920 Speaker 2: AI ethics, we historically have been talking about issues like fairness, equity, transparency, 528 00:28:52,960 --> 00:28:54,560 Speaker 2: privacy and so forth. 529 00:28:55,080 --> 00:28:58,280 Speaker 1: Or friendly AI that we shouldn't create something that would 530 00:28:58,280 --> 00:28:59,880 Speaker 1: then want to try and destroy him. 531 00:29:00,400 --> 00:29:02,400 Speaker 2: And that's the kind of part that has come kind 532 00:29:02,440 --> 00:29:05,320 Speaker 2: of into the conversation really recently, this idea of you know, 533 00:29:05,560 --> 00:29:09,760 Speaker 2: AI existential risk, AI existential threat and so forth. And 534 00:29:10,680 --> 00:29:13,760 Speaker 2: I would say that that's not an unimportant part of 535 00:29:13,760 --> 00:29:16,160 Speaker 2: the conversation. We really should be thinking about the full 536 00:29:16,240 --> 00:29:18,760 Speaker 2: range of risks that AI can pose and addressing them. 537 00:29:19,120 --> 00:29:22,520 Speaker 2: But it's become maybe an outsized part of the conversation. 538 00:29:23,080 --> 00:29:25,960 Speaker 2: We should think about AI ethics holistically and make sure 539 00:29:26,000 --> 00:29:28,800 Speaker 2: that we're not letting kind of one particular sub part 540 00:29:28,880 --> 00:29:32,080 Speaker 2: of AI ethics dominate the conversation at the expense of 541 00:29:32,120 --> 00:29:34,840 Speaker 2: really thinking about the rest of AI ethics as well. Really, 542 00:29:34,880 --> 00:29:41,080 Speaker 2: there's a huge need to democratize literacy skills and expertise 543 00:29:41,240 --> 00:29:44,400 Speaker 2: on AI so that more people are able to engage 544 00:29:44,440 --> 00:29:47,960 Speaker 2: in a way that is kind of informed by knowledge 545 00:29:48,000 --> 00:29:51,640 Speaker 2: of the strength's limitations risks associated with the technology. And 546 00:29:51,640 --> 00:29:56,360 Speaker 2: so I think really enabling more people to participate by 547 00:29:56,400 --> 00:29:58,680 Speaker 2: having that literacy skills and expertise is. 548 00:29:58,640 --> 00:30:01,000 Speaker 3: Really, I think the huge thing that we need to 549 00:30:01,040 --> 00:30:01,920 Speaker 3: achieve at the moment. 550 00:30:02,600 --> 00:30:04,240 Speaker 1: I did enjoy this conversation a lot. 551 00:30:04,280 --> 00:30:05,480 Speaker 3: Thank you, Thanks so much. 552 00:30:12,200 --> 00:30:14,640 Speaker 1: Thank you for listening to Zero. If you liked this episode, 553 00:30:14,720 --> 00:30:16,760 Speaker 1: please take a moment to rate or review the show 554 00:30:16,800 --> 00:30:20,120 Speaker 1: on Apple Podcasts and Spotify. Share this episode with a 555 00:30:20,160 --> 00:30:24,920 Speaker 1: friend or with someone who fears our robot overlords. You 556 00:30:24,960 --> 00:30:27,520 Speaker 1: can get in touch at zero pod at Bloomberg dot Net. 557 00:30:27,960 --> 00:30:31,080 Speaker 1: Zero's producer is Mighty Lee raw. Our theme music is 558 00:30:31,080 --> 00:30:35,480 Speaker 1: composed by Wondering Special Thanks to Kira Bendrim and Alicia 559 00:30:35,520 --> 00:30:38,520 Speaker 1: clanton I am Akshatrati back soon