Episode 3 - AI and Climate Change

Transcription

0:00:01 - Andy B

Can you guys hear the baby? 

0:00:05 - Joe C

I can hear the baby.

0:00:07 - Kiran V

Could you hear the cat? 

0:00:09 - Joe C

Oh, maybe it was the cat. 

0:00:15 - Kiran V

Yeah, Pico was meowing outside my door. 

0:00:18 - Andy B

Incredible. You know this one non-profit made these big fish-looking mouth things with big, big nets on the end to go around the Pacific garbage patch and pick it up. And then they come once a year with big barges and picked it all up. They've already cleaned double-digit percentages of the garbage patch up. 

0:00:42 - Kiran V

Yeah, I know it's crazy, and it was like some hundreds of thousands of tons of material. 

0:00:50 - Andy B

Yeah, it turns out you could just put giant robot fish out there in peace and come back and we're solving it. 

0:00:55 - Kiran V

Yeah, it's crazy. 

0:00:57 - Andy B

Anyway, there's a huge vested interest in making people be nihilistic and pessimistic, because there's a lot of money to be made in making you think it's hopeless and just keep buying shit. Nonsense, propaganda. 

0:01:07 - Joe C

I love it.  All right, sounds good, so I want to. 

Ha ha Andy Yeah whatever. Activisim, climate change. Whatever. 

0:01:24 - Andy B

I'm not saying that, I'm just saying that, like all our research is pointing at, proving our point that AI is just a tool, right, and that tool can be used for incredible good, or it can be used for sh** baggery, but like AI is really incredible because of it's like second and third order potential where, like, if AI can make a material scientist more effective, that material scientist can make better building materials, which makes buildings, concrete, have less carbon. 

0:01:49 - Joe C

There's so many things in the works. It's like can we just wrap this up in five years and like solve climate? Like what's the hold up? 

0:01:58 - Kiran V

Yeah right. 

0:02:00 - Andy B

This was super fun for me because I'm certainly somewhat a climate pessimist and I started the research like fully prepared to leave with a great feeling of shame, and then instead I was like, oh we are fixing it. Nice.

0:02:21 - Joe C

I  thought you made a good point, Kiran, with like a lot of this just comes back to renewable. At least how AI is contributing to it. 

0:02:32 - Kiran V

Yeah, I mean, really the main negative impact is the electricity usage. If it's all solar powered, that's fine. The water is really important, but again, and the mining, but, again, depending on how these systems are built, like the infrastructure is built, you can actually get a lot more efficient in your heat management systems. 

(Intro Music)

0:02:32 - Kiran V

Welcome back everyone to another episode of AI FYI, we are Andy, Joe and Kiran and this is your go to podcast to learn about all things AI and to understand what is AI, what are we doing with it, how is it impacting the world around us and how can I start getting more into AI. Today we have a very exciting topic for you: AI and climate change. We're going to talk about some of the different approaches, applications, techniques. 

0:03:52 - Andy B

And then as soon as we all went off and started researching, we were like this is a big topic, so we're not going to have too much depth on anything right now. I personally could do an easy six hour rant about agriculture and machine learning, but we're going to kind of overview the state of the situation as it currently is. So a couple of things I want to highlight before we get into it. We kind of took an approach in our research of looking at the good, the bad and the ugly. We just needed some way to organize all the information we were collecting. We're not going to go in that order when we discuss with you, but that's the main categories. We also learned in our research process that our hypothesis that AI is just a tool that we've been trying to convince you is absolutely correct. 

In the world of AI and its impact on the climate and the environment, machine learning itself in AI is not as directly involved as you could imagine. There aren't actually robot farmers or oil rigors out there yet, but machine learning has a ton of second and third order effects, by which I mean that if there's some person looking at a situation and being like this could be done better for the environment, there's probably a company that has a scientist who knows how to do that, and that scientist probably requires some input that could be made better with AI. So that's something that's really exciting about this is that the same applications or technologies can be used by a variety of different people in a variety of industries to make small progress in their domain towards fighting climate change. One thing that we want to make sure we do here is just level set a little bit on climate change. 

0:05:44 - Joe C

Obviously, this is a topic that extends far past how AI is contributing to it and how AI is helping. But just to level set our beliefs we believe climate change is real and that humanity is contributing to it and that it is something that humanity can also fix. So a couple higher level stats on climate change, and these are from the EPA 2020 report on sources of global greenhouse gases. So 25% of greenhouse gases comes from electricity and heat, 24% from agriculture and forestry, 21% from industry and then 14% from transportation, and then we have a handful of other smaller categories there. 

0:06:31 - Kiran V

And if you don't believe in climate change and you don't believe that humans are having a huge impact on the world around us, the three of us really do feel for you and wish you all the best in the future. 

0:06:46 - Andy B

But this is not the podcast for you. Goodbye. 

0:06:49 - Joe C

All right, so let's jump into the bad. So we're going to be talking a lot about how AI contributes to climate change, but in our research, we started to realize how sticky this is. We're going to be talking about some more direct contributions, but we can sort of go back and look deeper into how AI is contributing indirectly, and one area, for example, is more at the beginning of the process. AI and compute and technology needs a lot of raw materials, and so we didn't have too many stats on this, but we do believe that the mining and extraction of that is definitely a great contributor to climate change and something that should be a factor into this, even before we get into the more AI parts and how AI is being trained and some of the things that we're going to be getting into. 

0:07:39 - Kiran V

Yeah, and as we go further down the stack. So you know, as Joe was alluding to right, there's a lot of carbon emissions and climate impact that happens just at the mining of the materials that are used to create the machines that we build the AI on top of, and that is a really difficult challenge to actually track the impact on climate. 

Where we can start to have some better metrics and insight into the impact of AI on the climate is really once you get into the actual creation and utilization of that AI, and where that really comes into play is going to be in data centers, because this is where vast amounts of AI are stored. Trained power is used to generate and power those machines, which is what the AI is running on. 

And so you know, when we talk about the impact of AI on climate change and specifically the negative impacts that AI has, we're really going to look at the data centers, things like electricity usage, things like water usage, and that's where you're really going to start to track the carbon footprint for AI itself. 

0:08:58 - Andy B

So I went into my research thinking how high is the carbon footprint really? And full disclosure. I was ready to drag the industry of AI and stay on this podcast and just roast myself, because I'm a self-avowed hippie and I do a lot of things to try and reduce my own carbon consumption, but then I work in this industry. That's so bad for the environment. But the data that I found doesn't quite tell that story. So obviously the reason that training of machine learning models produces a high carbon footprint is because the hardware runs on electricity. So depending on how you source that electricity, that impacts that. 

As mentioned, 25% of global greenhouse gases comes from the production of electricity and heat and that's why you see so much effort going into making sustainable electricity. If all of our electricity came from renewable resources then the carbon footprint of training machine learning models would be a lot lower. But for context, I found a paper of how much it costs in pounds of carbon to train a standard NLP pipeline versus a big transformer and then they benchmark that against how much carbon like a standard across the US flight is and a normal human and an American human carbon footprint in in a single year. The data was really interesting when you train a big transformer, you're using approximately the same amount of carbon as the average American would in 17 years. 

0:10:44 - Kiran V

Wild. Whoa, that's crazy. For one model. 

0:10:48 - Andy B

For one model, one neural architecture, big search. This is much lower than what I thought it was going to be, to be completely honest. So if you look at the actual numbers, in the lifetime of a car, including fuel, it's looking at 126,000 pounds of carbon. One transformer, a GPU trained, one big transformer, is 626,000 pounds of carbon. So it's like what? Like five cars per model trained. Five cars for the whole lifetime (https://www.cbsnews.com/news/artificial-intelligence-carbon-footprint-climate-change/). 

For the whole lifetime, but the sense of scale. There's hundreds of millions, maybe billions, of cars on this earth. This is true. There's not billions of transformers, there's not even millions of transformers. A standard NLP pipeline, including tuning and experimentation, is 80,000 pounds of carbon, which is only three times what the average American uses in one year. It's three years of an American lifetime, If you're wondering. The average human uses 11,000 pounds of carbon a year and the average American uses 36,000 pounds of carbon a year. 

0:12:13 - Kiran V

It's crazy. I just hear these numbers and I think about carbon and it's like that sounds like a lot of carbon Tons and tons of carbon, co2, cubic, something, something I don't know the exact stuff, but yeah, it's a lot. 

0:12:29 - Andy B

But then I found another paper that was even more surprising to me. That came out in March of this year. It's on archive. It's titled the Carbon Emissions of Writing and Illustrating Are Lower for AI Than for Humans. I'm going to quote this. As AI systems proliferate, their greenhouse gas emissions are increasingly important, concerned for human societies, we analyze the emissions of several AI systems -  Chat GPT, Bloom, Dolly-2and Midjourney relative to those of humans completing the same tasks. We find that an AI writing a page of text emits 130 to 1500 times less carbon than a human doing so. Similarly, an AI creating an image emits 310 to 2,900 times less. Emissions analysis do not account for social impact such as professional displacement, legality and rebound effects. In addition, AI is not a substitute for all human tasks. Nevertheless, at present, the use of AI holds the potential to carry out several major activities at much lower emission levels than can humans. 

0:13:40 - Joe C

Does it talk about the medium, like if I were to set up a canvas with actual paints, or is it all digital? 

0:13:49 - Andy B

They basically did this bottoms up analysis and they have a chart for the laptop computer being on for the duration of a human writing one page, a desktop computer, a human just being the one writing it, like they have all the data organized. It's a very bottoms up analysis. Authorship does not exist in a vacuum, and any accounting for the return on energy expenditure is confounded by the impact that the rest of the network in which it is embedded. 

0:14:21 - Joe C

Isn't that interesting. I think we keep coming back to this idea that we hear numbers and stats on the performance of AI, like with self-driving cars, without looking at the same numbers that would be produced by a human. And I've referenced self-driving cars because we've talked about comparing self-driving car accidents to human accidents and would guess and I'm sure there's numbers on this that human accidents are far more so. Here we sort of have the same thing that maybe the output of humans is still exceeding the output of what an AI can do when given an equal task. 

0:15:06 - Kiran V

Yeah, I think that. Another point that's interesting to me is like, while the model say, you know, writing that one page is more efficient when we talk about carbon emissions, I would argue that because we have access to AI, we're actually writing a lot more, or creating a lot more of these images using AI than a normal human would have without AI. So while one page is 300 times more efficient, I would argue we're maybe just creating 300 times the amount of content that's redundant or unnecessary. So it's like how do we actually measure the true impact or outcome? 

0:15:54 - Andy B

I don't know. It takes us back to how sticky this is to measure, from the actual raw mining, like raw materials extraction, all the way through the application of the machine learning AI technology. 

There's so many pieces of data it's impossible to untangle and actually measure but, I was pleasantly surprised by how relatively low even the training numbers were relative to what I was expecting, and I think the reason for that is that I have a lot of knowledge of how many resources go into training models. I've heard the bills of how much it costs self-driving car companies to do simulated driving. I underestimate how many other sources of carbon there are that far outstrip what technology is doing. 

0:16:59 - Kiran V

Fortunately, this is all using electricity, and electricity is 100% renewable. If we want it to be so, we just need solar farms and we should be good. I do wanna add one more stat to this. So we mentioned, all of this is, 99%, for the most part, hosted inside of data centers, and so these data centers do make up, in the US, 2% of all electricity usage across the country, which, if you think about the number of people, the number of homes and buildings and offices, 2% is a lot to just be storing, processing, handling our data and technology. So that's a lot of a lot. 

And one paper it was an unverified research study that was trying to again assess the impact on the climate from AI, and so what they looked at is they actually looked at the amount of water that was being used to train Chat GPT, because, again, these data centers is where the training happens and a lot of data centers use our water cooled, so it's literally there's so much heat being generated from these machines that they need to cool it. Air is very inefficient in cooling water because it doesn't have a particularly high heat capacity, so they use water to cool these data centers, and the water when it's used for cooling. Most of it evaporates so you can't really use that water again unless you're capturing that. But the study found that Chat GPT used about 700,000 liters of fresh water to train Chat GPT. So that's a lot of fresh water and, as a very precious resource that humans require, was fairly surprising. But again, this is all relative to the amount of water, relative to everything else. So how much is that impact is very challenging to really assess accurately. 

0:19:12 - Joe C

If you've ever been burned by a laptop on your lap, you know how hot computers can get. So imagine these hanging out in a data center by the thousands. 

0:19:20 - Andy B

Yeah, and I suspect that a larger percentage of the whole data center footprint is gonna be AI going forward. Right now, there's not that many people training these very, very large models because the expertise isn't there, but it's getting easier. Hopefully, we don't need to train these big models from scratch all the time, and lower resource techniques like few shot learning, fine tuning, will also yield the results we want without having to waste a lake's worth of water every time we do it. 

0:20:00 - Kiran V

Yeah, and I don't know if you guys actually saw the recent announcement from Open AI. They're creating a interface to allow humans to create their own GPT, so you can make like a Kiran GPT or Andy GPT, or a GPT for you know, give me instructions on how to build IKEA furniture, like, so I think and and all of this can be done without writing any code. So now you know, any human that has access to the Internet is going to be able to train their own generative, pre-trained model, which, again, now, as we're increasing the access to these things, it's going to be more and more important to be tracking how this is impacting the climate. 

as we touch on a lot of these topics, we've, you know, hopefully, made it very apparent that AI does truly have an impact for the worse when we talk about the climate. What exactly that impact is and how far it goes is very difficult to measure, again, because of how far up the pipe you can go to start tracing this all the way into. You know, those, the minds that are pulling out the precious metals to create the machines that are used to train the AI, that are delivered, you know, ultimately onto your cell phone, right? So, while the negative impacts are certainly there, there are are actually a lot of amazing cases where AI is being used for good in AI and for good in climate change to actually, you know, improve the lives of this earth and humans and, you know, making sure that we have a truly sustainable future to start off. 

So, you know, talking about the good of AI and climate change, I do actually want to start with something that happened very recently. So, on October 26, the UN announced an AI advisory body, and essentially, what this group of individuals is tasked with is how do we responsibly, responsibly leverage AI in our world, right so, for climate change, for safety, making sure that, you know, again, as we're training models, and so you know we have our episode on training data, where we can get, where we get a lot more into this, but, you know, mitigating bias and models and making sure that this AI is being created very ethically (https://www.un.org/sg/en/content/sg/personnel-appointments/2023-10-26/secretary-generals-advisory-body-members-artificial-intelligence)(https://news.un.org/en/story/2023/11/1143187). 

So there are a lot of initiatives that the UN is putting together to make sure our future with AI is something that is sustainable and can be, you know, brought into this world without, you know, too many negative repercussions, and so what they have put together is actually, you know, starting to capture some of these different efforts on AI and climate change specifically, and understanding the large industries where there is a very big impact on climate change and you know, are there opportunities to start bringing AI into these industries to improve the outcomes on the climate?

And so, you know, just to touch on those and you know we'll dive into a few more or we'll dive into some of them in detail. But you know..

Weather. So being able to predict the weather is actually, you know, can have a very large impact on climate change and safety of humans, disaster prevention. So things like wildfires or hurricanes, tracking pollution, again, we all know. 

You know the climate change and global warming is a result of increased greenhouse gases, which are things which come from burning fossil fuels, like running your internal combustion engine car. Carbon neutrality, which is to how do we get to a world where the amount of carbon that we're putting into that atmosphere is the same amount that we're taking out of that atmosphere? So over time we're not increasing those greenhouse gases. 

And a couple other ones which you know were actually surprising to me, which I didn't really realize how much of an impact it is, but it makes so much sense when you think about it Fast fashion and fast food, which are just such major industries. Fashion is a $2.4 trillion industry in the world and both of these industries have huge impacts on climate change because if you think about you know in farming and agriculture that's necessary to sustain all of the fast food demand, all of the shipping and transportation that's required to bring those clothes that are being made abroad in some of these third world industries over to be sold at your local stores. 

So lots of different industries having huge impacts on climate change and what's really cool is AI is actually being used in a lot of these industries now to help mitigate that and to improve the future of climate change. 

0:25:20 - Joe C

Alright, so Kiran touched on the UN AI advisory body, and that's really just one body that's working to solve this crisis. We have this government organization and many others, many projects within the government. There's a lot of nonprofits out there that are doing this in many ways, including through technology and AI. There's also a whole host of startups that are working to take advantage of our green economy and contribute in their own way, and even the bigger players, such as like Google have a lot of projects going on that are helping to limit their own carbon footprint and, through the technologies that they're exploring, can help other companies and us as a global population reduce carbon emissions. 

So all of these different bodies are really working towards this idea of carbon neutrality, and there's really two ways we can do this. 

We can limit carbon emissions in the first place, so I think governments have a large role to play in this. 

You've probably heard of the idea of carbon credits and other regulations that have been put out, and these are both programs to tell big companies and producers of carbon emissions that they have a limit and make it possible for them to continue to do business while still contributing in hopefully smaller ways to greenhouse gases and carbon emissions. 

And then I want to talk about how we pull carbon emissions out of the atmosphere, carbon capture and I think in a lot of ways this is still a very green or new technology, but there are a couple startups that are really trying to understand this. I'd imagine one of the biggest hurdles for a startup or even a bigger company getting into this business is how do you be profitable without contributing to even more carbon emissions? So I know there's a couple startups that are actually pulling carbon out of the air and then turning them into building materials like concrete that can be sold again, which is really neat. It commoditizes our pollution and I know I'd love to see more of that in the future (https://www.mapmortar.io/)(http://eugenie.ai/). 

0:27:41 - Kiran V

Yeah, and I think just a really basic example of carbon capture for people that don't know is you know you can buy carbon offsets where you know.. I'll buy a flight and then I can buy carbon offsets and there's a company that's going to go and plant trees or go and sequester carbon in some way, using whatever their processes, and so this is again more ways of capturing carbon out of the air. If you plant more trees, they take more carbon in and store that and sequester the carbon inside of their tree trunks. 

0:28:14 - Andy B

So AI has this incredible multiplying force second order, third order impacts on different fields. What we mean by that is that if you think about a single farmer in a field going, hmm, this could be done better, maybe what they need is a better tool. And at a company making that better tool, maybe there's a scientist who's like I can make this, but I need a better material. And there's another scientist at a materials company who's like I can make this material better. And at different steps in that pipeline, you can use AI to make the process better. So I'm really proud of the fact that, like, AI has these kind of second and third order impacts that really enable a lot of cool solutions. 

And it's just about how 25/24% of global greenhouse emissions comes from agriculture and forestry, and what's interesting about that is that we might be able to solve that entire input relatively quickly compared to some of these other inputs. We already know how to do green agriculture. We did it for (back of the napkin) 30,000 years. The problem was we couldn't scale it right. At the time 30,000 years ago, probably 99% of humans were involved in producing food. Now, less than 1% of human beings grow all the food for all of us, in my opinion, a collective responsibility to make sure that we're not blaming farmers for their high carbon footprint when reality we're benefiting from the fact that not all of us have to farm anymore.

And there's so many ways to do something called regenerative agriculture with AI. And regenerative agriculture basically is taking what humans were doing for a long time and figuring out how do we do it at scale to preserve the health of the soil and make sure that every time we farm plants or animals that we are adding back to the environment and improving it. And there's lots of cool examples for this. So, for example, if you have a field of grass and you have cows living on it, those cows will poop. That poop produces methane, cows fart. You can actually drastically reduce the impact of those cows if you just run chickens through the field after the cows, because the chickens like to peck at the poop and eat the worms out of it and they are little clawy feet grind it into the soil and actually make that field perfect for planting on the next year. So there's lots of clever hacks and I want to share how some of these are being implemented with AI. 

So there's a company called Blue River. They're subsidiary of John Deere and they have a product called See and Spray, and the way they market it is they basically tell these big farm companies and farmers that hey, your pesticides are really expensive. And when you have a field laying fallow which is really important for any kind of agriculture or regenerative agriculture is to let the soil rest for a couple of seasons you still have to reduce the weed cover on it and traditionally this was done by just spraying this fallow field with expensive pesticides. The plants don't like it, so they evolved to be pesticide resistant, so the pesticide gets more expensive, so on and so on. They took the sprayer and put cameras on it and implemented some computer vision tooling and they can reduce the amount of chemicals needed by 77%. They'll hit 98% of the weeds on the field, which is basically the same accuracy than if you just sprayed all of the soil, but you use 77% less herbicide or pesticide in the same field, just by, like using a computer vision algorithm to spot spray, spot spray, and all the mechanics is the same. somebody's driving a sprayer arm thing back and forth across the field, but you just you use less toxins. I think that's incredible. 

0:32:47 - Kiran V

Yeah, that's so cool. 

0:32:50 - Andy B

Yeah, they did this test on 75,000 test acres multiple times and, like they have the data to back it up, it's good for the farmer, it's good for the environment, like and it's just making sure that the existing robot arms that spray have camera and computer vision like perfect application of the tech. 

0:33:14 - Joe C

I think it's such a good example. You know, for a lot of startups or companies, this is a good example where it's a win win on so many different levels, and I think you need that. A company needs that in order to be profitable and to be able to operate. They need to have these selling points that say it's going to be less expensive, in addition too, you know, being good for the environment, it's going to be a reduction in your costs and, by the way, you're doing something good for the world. In a few other examples that I've seen, there was sort of that that set up in a way. 

0:33:56 - Kiran V

I mean, and that stat alone that you said 77% less herbicides. That's crazy to me if you look at the other way, that we've been using four times the amount of herbicides that we've actually really needed to use, and been doing this for generations. So the fact that there's this new tool, technology, AI that's coming in that can have orders of magnitudes of impact in so far every single industry that I know of you know can have some application of this. I don't know anything that doesn't or has zero application. So that to me is just like another, you know, like instance of how AI is so incredibly powerful, even within just the category of climate change. 

0:34:53 - Andy B

I have one more incredible agriculture. I'm going to use case to talk to you guys. So a lot of carbon emissions from agriculture come from livestock farming. Right, cows fart, chickens. Poop contains all kinds of stuff in it. It's a process, and part of that challenge is like we need to scale how many animals we are keeping, and at a farm the individual relationship between the farmer, the livestock guardian and the animal doesn't scale that way. 

Like one person can't have a close emotional tie and take good care of 10,000 cows just doesn't work. And so the way this has been handled in the last 100 years as we ramp agricultural production by individuals is but for cows you can put a tag on their ear and it has a number, and then somebody has like a spreadsheet and they've got all the cows numbers and they're trying to track. It quickly becomes overwhelming. So then you do things like you can't tell if a cow is sick or which cow is sick, so you give all the cows antibiotics and this gets orders of magnitude more complicated with all the animals smaller than cows that we eat, like fish and chickens. You cannot tag a chicken. They do, they put bands on their feet but like it's impossible for somebody to go through and figure out and even like weigh each chicken. They don't. They just try to like weigh an entire platform of chicken for an average. But if you can take individual actions for each animal it's so much better for that animal. You use a lot less resources. The farmer's heart hurts a lot less. 

I worry a lot about people in agriculture working with animals because that's got to be damaging. So, for example, when you want to grow chickens, you have to take eggs, make a baby chicken, and if you want those chickens to lay eggs for you, they can only be female chickens. So like tens of millions of male baby chickens are killed around the world every single day. I think like it's the numbers crazy. There are people whose job it is to take a newborn chicken, flip it upside down and guess whether it's male or female, and the females get sent to - you'll be meat or eggs, usually both, and the males get (I kid you not) put into a grinder. 

It's horrible that poor person doing that job and it's completely preventable with AI because you can look at a single egg and scan it at the developing zygote and determine its sex and you can use a computer vision model to do that at scale. So you just take the eggs that will become males and make an omelet. Never make a chicken suffer, never make a human suffer. And then take the eggs that will become females and go make more eggs. A chicken in the production scale can make at least an egg a day, sometimes more than one. Like you literally can't check every egg manually, but a computer vision model can. And so much waste and suffering and harm for animals, the environment, for humans. Can just go away with a really simple, effectively object classification model. 

0:38:35 - Joe C

I also think this is just one example of you know how AI is impacting climate change, but also the meat industry and, like human health, and there are probably a lot of other AI projects in the works to combat, you know the ethics around even eating animals. Even someone is using AI to create lab grown meat. I'm sure AI is part of that process or, you know, to create better tasting vegetarian options. So I don't know. It's really fun to think of this on a collective, larger scale, about how a lot of different AI endeavors are converging to make the world a better place. 

0:39:19 - Kiran V

Yeah, it also sounds like we need a facial recognition algorithm for cows so we can identify individuals and know which one is sick. 

0:39:27 - Andy B

I read a whole thing about that. Basically you can put RFID tags on the cows but then if, like, they don't go near the sensor, your host. So doing like tracking of individual birds and tracking of individual cows lets you give specialized treatment to each animal and that's just so much better for everybody. I think we can all agree. 

I have cut myself down to a simple dozen other examples. I could list, but let's do that in a different episode if people are Interested, because my obsession with agriculture, I realize, is not everyone's interest. But yeah, I just you know love regenerative agriculture. If agriculture is contributing 25% of our carbon emissions and greenhouse gases and there's so many ways to make that better with AI, then I'm really proud of all the work I've done to make computer vision work in that world. 

0:40:27 - Kiran V

So stay tuned for our Christmas special. Andy talks AI and Agriculture. 

0:40:33 - Andy B

It's going to be six hours. Okay, Bring a snack. 

0:40:40 - Joe C

We definitely have to do that episode, so I want to talk about just a couple of other domains. There's certainly a lot to talk about with agriculture, but, Kiran, you touched on a couple other domains. One of them is fast fashion, and so just to talk about some inspiring work I've seen in that realm, I came across a startup called Refibered (https://refiberd.com/). So something like 186 billion pounds of textile waste is discarded globally, and one of the reasons for that is because it's so hard to sort and then recycle properly, and this startup is using AI to do that sorting, so hitting that first step of the process, where then the right textiles that are recyclable can be picked out, and what can't be recycled can go to the landfill, and what can be can go to folks who can reuse it, and with their algorithm, they say that 70% of textiles can be recovered of the textiles that come through their stream. 

0:41:44 - Andy B

I just heard yesterday on a TikTok something about the dyes used to dye fabric have huge environmental impacts. So now some folks are using gene editing and traditional farming, like crop selection gene editing and AI field gene editing, to just grow cotton that doesn't grow white, and then you don't have to die the fabric. 

0:42:10 - Joe C

That's incredible. 

0:42:12 - Andy B

Yeah, that's so cool. 

0:42:13 - Joe C

That's crazy. Yeah, so, Kiran, also, you touched on weather and disaster prevention. There's so much going on in this space. I think this could also be another episode just on, like predicting the weather and predicting other natural disasters not even the weather, but you know, landslides, floods. I guess flooding is technically weather, but wildfires all sorts of things. I was reading about deep mind, which is Google's AI sector. They have a group that is particularly looking into climate change and solutions for it, and they had a recent project called Nowcasting (https://deepmind.google/discover/blog/nowcasting-the-next-hour-of-rain/) that predicts weather within one to two hours, particularly precipitation, which is much better model than anything we've ever come up with in terms of predicting the weather. Now, you can imagine that would be good for you know, checking tomorrow's weather for your, for your road trip or whatever, but also you can imagine this having great impacts on predicting disaster and where there might be a big storm and where there could be flooding. Again, this is just one example. 

0:43:24 - Andy B

And I have to cut in here and bring this back to farming, because this is really important. There's this concept in agriculture called crop insurance. Long story short, farmers get in trouble all the time because they're like today's a nice day, it's spring, they plant a seed and then a freak storm arrives three days later and washes all the seed they just spent tons of money buying and washes it away. And in the case of small scale agriculture in the global south, sometimes people can't afford to repurchase that seed and in the scale of like large scale agriculture this is can cause millions of dollars of waste. So if you can even get just a few days better heads up or statistical probability of like the best day to plant especially since historical farmers almanacs are getting less accurate the way people used to do this because of climate change you can save farms. You can save livelihoods of people who've been farming that land and will know better than anyone what to do correctly. So yes, predicting the weather is hugely important. 

0:44:42 - Joe C

Yeah, and it's really neat just to think that this is one model, this is one teams project, and if this model can be proliferated and given to the world, there are so many applications for it and and yet protecting costly seeds is certainly one of them. A lot of our examples maybe even all of them about how AI is, you know, making the world a better place in terms of climate change just really point to the fact that we can now do things that humans have never been able to do before. It's really adding to efficiency. We may have shared the stat earlier, but 40% of carbon emissions really tie back to like infrastructure and buildings, and even improving buildings themselves and how we do construction and development can have a huge impact on climate change (https://www.economist.com/finance-and-economics/2022/06/15/the-construction-industry-remains-horribly-climate-unfriendly). So a lot of the big players in tech have actually used AI to make their data centers more efficient. 

We touched on that a lot, but you can imagine, when you look at all development and all building construction and how we can apply AI to it to make those buildings more efficient, that's going to be a great reduction in carbon emissions. Another example of this sort of being indirectly affecting climate change is that AI can can be used to do renewable energy better. And nuclear fusion. Google contributed to enabling nuclear fusion using AI. And just that simple effort.. well, I'm not going to say simple, but that effort can mean we use more nuclear fusion and rely on less, you know, carbon based energy sources. 

0:46:30 - Kiran V

Another example of infrastructure improvements that is happening using AI is in mining of lithium. So you know, this is the mineral that is used to create all of our batteries that power our cars and our cell phones or laptops, everything. And currently this is a very expensive, very climate impactful process and there's a company called energy source materials that's actually working on improving the process of mining lithium to again improve the downstream effects of this. Right so there's applications at every layer in all of these industries and chains happening to improve overall right. So you have AI in the car that is improving the efficiency of the car itself, but now we have AI all the way at the top of the funnel that is being used to improve the process on how we develop batteries that are going into these cars. So again, first, second, third, fourth order effects kind of across these industries. 

0:47:42 - Andy B

I need to disagree with something Joe said earlier. I should have done it earlier, because I fully agree with you guys. But it's not that AI is only doing things that people can't do before. It's that AI is helping us go back to what humans have been great at doing for a long time, like when everybody was a farmer. Everyone had a personal and close relationship with their food and could treat their food well and treat their local soil well. Now we're scaling that. So AI is opening new frontiers of what we can do, but it's also mirroring what we have to learn from history and from you know, our ancestors around the world, who have already done things well, who basically were doing gene editing via crop selection for tens of thousands of years before we ever called the gene editing. AI helps us learn from them and apply that at larger scale. 

0:48:41 - Joe C

I think the yeah, the term scale there is really important. Things that we may be able to do in the past if we were like monitoring our own little field, our own, the family's food. But now you know, our systems are responsible for feeding a global population and we need extra help with that. Humans can't do it alone and this computing power gives us the power to do it. 

0:49:06 - Andy B

And I personally really like the fact that, instead of having to farm year round, I like get to play video games sometimes. I think that's great. It's a great thing that we did for ourselves. Well done humanity. 

0:49:18 - Kiran V

Yeah, I don't know if I would cut it as a farmer. 

0:49:22 - Andy B

Sidebar. This will get cut out, I'm sure. But at one point during COVID I guess the UN was trying to figure out the some committee was trying to figure out the ideal vaccine delivery thing and they were really worried about food system collapse because so many farms -  there are about 350,000 farms on earth that feed everyone Well, which is a smaller number I think then most people would expect, and almost all of them have just one or two matriarchs or patriarchs, who's the only person that knows what the fuck is going on? And this committee recommended that the very first people to receive these vaccines should have been like special emissaries, with a shot going to every farm and giving injections to those people, because those 350,000-ish people are the one standing between us and complete collapse. Isn't that crazy? 

0:50:19 - Kiran V

Yeah, it's crazy. Very fragile. 

0:50:23 - Andy B

Shout out to farmers. Okay, so I have been the AI hype girl for this entire episode. We've been talking a lot about how machine learning AI is being used in multiple steps of the supply chains; the second, third, fourth order effects for good. Let's be very real. They're also being used by people that you could say, diplomatically, are bad actors. I'm going to go ahead and call them evil people, selfish people.

So the same open source computer vision model that can be used to predict flooding, probably also is to predict veins of precious oil running underground and mine them deeper and more hungrily. So when we've been talking about, AI is a tool and what matters is the hand that wields it, because it's such a force multiplier, it lets one group or person escalate what they're doing at such scale, their decision making power. It's very important that the person with the decision making power is making good choices. There are heavy investments of an AI machine learning in oil, natural gas mining, all kinds of problematic heavy scale resource extraction projects from earth. Anybody can use satellite imagery and computer vision to figure out where our forests are growing, somebody can use it to decide where should I plant more trees and somebody can use it to decide where should I go, cut trees. 

0:52:13 - Joe C

It's just one more reminder that sort of everything that we're talking about fits within the confines or has to play with capitalism, and we, you know we must follow the money to determine how these things are used. Yeah. 

0:52:31 - Andy B

And don't, for a second, trust any of the content you see from anyone who stands to make a buck from the planet telling you they're using AI for efficiency, unless they can provide you the sort of data that Blue River did, where they're saying like, look, we've reduced this thing by this much. If they're not that transparent and that clear there's there's some bullshit happening. I get so much clear data and facts from the people who are proud of their work. The people who know they're being shady, they're not transparent. 

So we don't have all the numbers because it's so hard to measure things. We don't have big numbers to share of like how much more efficient like fracking is with AI. Just know there's somebody who's using AI to make fracking more efficient. 

0:53:38 - Joe C

In the examples I looked at. I think it was really good to step back and ask, like, what's in it for them? Beyond the goodwill of helping the planet? You know, with the textile company I mentioned, they have a service that they can provide to people who make textiles and can profit off of reusing textiles, and so to me that makes sense, that they have a place in the supply chain where they can both make profits and do some good. I would be skeptical of any company, you know, doing something green just out of the goodness of their heart. 

0:54:19 - Kiran V

Yeah, and I think with like any tool or technology, right when it comes to capitalism, it's very much a double-edged sword and you know you can use that tool or technology to do some really good things and have great impacts. And you know there's a lot of nonprofits around the world that you know try and capitalize on those impacts. But on the flip side right, there are going to be companies like Exxon. That's like oh wait, a second. You mean I can mine oil more cheaply so I can get more oil out of the ground for spending the same amount. They're absolutely going to take advantage of that because it's a business and they're trying to make money. So how these AI tools are applied to these different challenges is going to be really important in the long-term outcomes and impacts on you know, something like climate change. 

0:55:20 - Joe C

And I hope that all these AI innovations will lead some of those companies who've maybe been in the tradition of harmful behavior to realize that doing something a little different is suddenly more profitable. We hear a lot about how, you know, the cost of solar power has gone down thanks to regulation and, as a byproduct, the industry is being built up to actually get that into place. For some of these companies that you know wanted to make money off of providing power in the past, maybe oil was the only way to do it, but now there's options and they can look at the books and see what's actually going to come out cheaper and realize that maybe there's a different way. And they don't even need to consider if it's good for people or not. They just need to know it's good for profits. 

0:56:11 - Andy B

Something that I struggle with is that the company I work for, Verta AI. We are doing a lot of things that can help models be green, so we are offering model distillation services. That basically takes big bottle and makes it smaller so it runs faster and cheaper, and we're really good at doing the serving in a way that's really fast and performant. But the thing is, green doesn't sell in all circles, so I can't do product marketing. That's like look at how you could reduce your carbon footprint if you were serving with Verta or using our services. Instead, I have to position it as look at how much money you can save. Drives me crazy. 

0:57:00 - Kiran V

I almost feel like there needs to be this baseline of splitting the difference of profits with mother nature. So it's like if I had something that I was doing and suddenly AI helps me do it 50% more effectively, we could say, all right, you keep 25% of those profits Great, you did a bunch of work, but 25% of that needs to go to Mother Nature, whatever that means. But I think that would be so cool if humans could just say like all right, we're all going to do this for the earth. We'll still get more profits as individuals or organizations, but the earth is also actively benefiting and there's a clear bottom line for nature in a lot of these efforts. 

0:57:50 - Joe C

So there is, on that topic, a little bit happening here with carbon credits, and I don't know the entire story here, but you can get carbon credits for essentially purchasing and maintaining green spaces or pieces of the earth, let's say, that contribute to a healthy planet. So I read an article in Wired Magazine not long ago that someone was working to preserve large swaths of seaweed fields in the Caribbean. The field itself, underwater, keeps the planet very healthy. It's an important part of the ecosystem. So by maintaining it, a government will pay or give credit to that person to help them offset maybe some other shady thing that they're doing that does produce carbon. It is a win-win because that field gets to thrive and the planet gets to thrive. The person who has a business to run gets a break on the repercussions of their carbon output. What else? There's a lot of benefits to that, but it does sort of rest on the idea that a government needs to be in place to enforce this and provide the regulation and sometimes funds to make it possible. 

0:59:28 - Andy B

I don't trust governments, but I do trust people. That means I hope, when you listen to this, that you feel like you know what to ask for from your representatives and from your companies and your service providers. Like I know, you could use AI to do this better. Why are you not? Like get informed and then get sassy everybody. You are entitled to ask for your planet to be treated with kindness and respect. 

0:59:59 - Joe C

All right, cool. So we're going to start to wrap up a little bit here, but just before that, I want to talk about the future and kind of where a lot of these ideas are heading. One interview that I came across was with Sims Witherspoon, who is the head of DeepMind's Climate Action Group, and this is again Google's AI division, Deepmind, and they were asked what's needed, what's needed to take this further, and they had two great points, the first being that a lot more data is needed (https://www.wired.com/story/wired-impact-deepmind-ai-climate-change/). 

As we talked about quite a bit, data is really what drives AI and makes it possible and gives us understanding into, you know, what is possible with the AI, and there's a lot of gaps in that. When it comes to climate related data sources, they mentioned the Climate Change AI group. 

CCAI is a nonprofit that really believes that machine learning can fight climate change, and one thing that they've maintained is a dataset wish list that outlines some of those gaps, and so, basically, you can consult this wishlist and they outline like, hey, what do we still need to look at? And it covers a lot of agriculture things, but also urban planning, how even EVs and batteries operate, data on building and construction, all these different things that are sort of climate related. So Christmas is coming up, go look at their data wishlist and get out there and find that data for them (https://www.climatechange.ai/dataset-wishlist.pdf). 

And then the second thing that they mentioned that is really needed is domain experts. So because climate touches so many different areas, we need experts on this. So an electrical engineer, for example, might need to be consulted because they're a player in this. A farmer might need to be consulted. You know, groups that make vehicles need to be consulted, and that requires domain expertise, because no one's going to know the business or the implications like they do. 

1:02:06 - Kiran V

This is actually something we discussed in one of our other episodes - How AI is Made, and these are the subject matter experts, or SMEs, who are often involved in developing AI, because it is really the knowledge in their head that we're trying to put into that model. So you need the people that know how to do these things really well to feed that data and information into the model so that the machine can now effectively take over what that human was doing.

Cool, well, thanks everyone again for tuning in to today's episode of AI FYI on climate change. As you heard, this is an enormous topic and we can have many, many more hours of discussions, so we would love to hear from you to you know, figure out, what are you guys interested in, what type of topics do you want us to go into? 

So send us an email at aifyipod@gmail.com. Check us out online at aifyipod.com and like, subscribe and share on your favorite podcast listening platform. This is Andy, Joe and Kiran signing off.