The TED AI Show: Can AI predict (and control) the weather? w/ Dion Harris and Tapio Schneider

[00:00:00] [SPEAKER_04]: Audio Collective.

[00:00:35] [SPEAKER_00]: You can play around, you can drill in, you can understand what was happening in 1972 on this

[00:00:41] [SPEAKER_00]: specific day.

[00:00:42] [SPEAKER_00]: What was really exciting is to see people who aren't climate researchers go and literally zoom in

[00:00:47] [SPEAKER_00]: on their neighborhood and see what was happening or see what the foliage looked like at

[00:00:52] [SPEAKER_00]: that time and understand what's really happening.

[00:00:54] [SPEAKER_06]: This is Dion Harris from Nvidia.

[00:00:57] [SPEAKER_06]: The company that's been creating a complete digital twin of the planet, appropriately called

[00:01:02] [SPEAKER_06]: Earth II.

[00:01:04] [SPEAKER_06]: So, I can use Earth II to travel back to my current neighborhood in Austin, Texas.

[00:01:09] [SPEAKER_06]: In 1972, I can go all the way down a street level and see the oak trees lining the

[00:01:14] [SPEAKER_06]: blocks, the green expanse of Zilker Park.

[00:01:18] [SPEAKER_06]: But you can also see the specifics of what the climate looked like, the air quality,

[00:01:23] [SPEAKER_06]: the weather systems, the monsoon patterns, and how they all interact.

[00:01:28] [SPEAKER_06]: But, even more importantly, we can use these complex climate models to travel into the future

[00:01:35] [SPEAKER_06]: and visualize multiple outcomes and extremely high resolution and do it all really,

[00:01:41] [SPEAKER_06]: really fast.

[00:01:42] [SPEAKER_06]: Something that traditional climate models were nowhere close to achieving.

[00:01:46] [SPEAKER_00]: Roughly about 45,000 times faster in terms of creating an actual forecast.

[00:01:52] [SPEAKER_00]: It's not just one forecast that makes it interesting.

[00:01:55] [SPEAKER_00]: It's being able to do thousands of forecasts that can give you a much better representation

[00:02:01] [SPEAKER_00]: of what the likely outcomes come out.

[00:02:03] [SPEAKER_00]: So, just to give an example, we are working with a central weather administration of Taiwan

[00:02:07] [SPEAKER_00]: and they're often hit by type phones that are coming inland.

[00:02:13] [SPEAKER_00]: And so, what's interesting about Taiwan is it's an island.

[00:02:15] [SPEAKER_00]: And so, you have to be able to understand how and when you can relocate but their relocation

[00:02:20] [SPEAKER_00]: options are somewhat limited.

[00:02:21] [SPEAKER_00]: And so, by giving them more granular understanding of where and when type phones will

[00:02:28] [SPEAKER_00]: hit landfall for example, they can then model and quickly build re-evaluation or relocation

[00:02:35] [SPEAKER_00]: programs based on that data.

[00:02:37] [SPEAKER_00]: So having more resolution gives them more information quicker about specifically where and

[00:02:43] [SPEAKER_00]: when air is going to be impacted.

[00:02:47] [SPEAKER_06]: So it's safe to say that AI technology is ushering in a new era of powerful tools to predict

[00:02:53] [SPEAKER_06]: and respond to climate change.

[00:02:56] [SPEAKER_06]: But it goes beyond that.

[00:02:58] [SPEAKER_06]: Geoengineering is giving us the ability to intervene and control the weather.

[00:03:04] [SPEAKER_06]: Does this mean that AI could actually solve climate change?

[00:03:10] [SPEAKER_06]: I'm below the will to do, and this is the Tedi Aisha where we figure out how to live and

[00:03:14] [SPEAKER_06]: thrive in a world where AI is changing everything.

[00:03:22] [SPEAKER_05]: Imagine this.

[00:03:24] [SPEAKER_05]: In 2030, the CFO of a Fortune 100 company is a bot.

[00:03:28] [SPEAKER_05]: I'm Paul Michaelman and on Imagine This will be exploring possible futures and the implications

[00:03:33] [SPEAKER_05]: they hold for organizations.

[00:03:35] [SPEAKER_05]: Joining me will be BCG's top experts as well as my co-host gene, BCG's conversational

[00:03:41] [SPEAKER_04]: Gen AI agent.

[00:03:42] [SPEAKER_04]: Blending human creativity with AI innovation, this podcast promises an unmasched listening

[00:03:48] [SPEAKER_04]: journey.

[00:03:49] [SPEAKER_04]: Join us on Imagine This from BCG.

[00:03:55] [SPEAKER_02]: Sometimes things in the world of technology are complicated and need careful

[00:03:59] [SPEAKER_02]: explaining.

[00:04:01] [SPEAKER_02]: Sometimes they just need a little hard truth.

[00:04:03] [SPEAKER_01]: I don't think anyone is going to buy up an honor with crypto at any point in the foreseeable

[00:04:08] [SPEAKER_01]: future.

[00:04:10] [SPEAKER_02]: I'm Lizio Liery, the host of Slates What Next TVD, your clear-eyed guide to technology,

[00:04:14] [SPEAKER_02]: power and the future, Friday and Sunday wherever you get your podcasts.

[00:04:25] [SPEAKER_06]: Now, I don't know if I buy that AI is a panacea for the climate crisis, but these

[00:04:30] [SPEAKER_06]: models are an amazing example of how AI will be transformative in the climate space.

[00:04:36] [SPEAKER_06]: So I'm feeling very hopeful and excited.

[00:04:38] [SPEAKER_06]: This is where we ought to be putting our AI technology to work, but with every technology

[00:04:43] [SPEAKER_06]: there comes great challenges and unintended consequences.

[00:04:47] [SPEAKER_06]: And because it's earth we're talking about, we have to tread carefully.

[00:04:51] [SPEAKER_06]: The stakes here are massive.

[00:04:54] [SPEAKER_06]: Our guest today is climate physicist, topio Schneider, who's going to talk us through

[00:04:59] [SPEAKER_06]: the promises and the parals of this new era of climate modeling.

[00:05:06] [SPEAKER_06]: So topio, you've been working in climate science for a solid 30 years now, but I want

[00:05:12] [SPEAKER_06]: to go all the way back to the beginning.

[00:05:14] [SPEAKER_06]: What attracted you to the climate space in the first place?

[00:05:18] [SPEAKER_03]: As a physics student, I was always interested in physics of everyday life.

[00:05:22] [SPEAKER_03]: How does it refrigerate over work and how does it consist of over work?

[00:05:26] [SPEAKER_03]: That kind of physics I found absolutely fascinating.

[00:05:29] [SPEAKER_03]: And as I progressed as a physics student, I realized that the physics I was doing and learning

[00:05:35] [SPEAKER_03]: was increasingly a bit further removed from daily life.

[00:05:38] [SPEAKER_03]: And I decided I want to work on physics at the energy of sunlight by definition.

[00:05:44] [SPEAKER_03]: That's sort of the daily life physics.

[00:05:46] [SPEAKER_03]: And that's how I got to climb it as a part of physics.

[00:05:50] [SPEAKER_03]: I must say it did play a role in my decision making that it matters to people.

[00:05:55] [SPEAKER_03]: It was clear already then 30 years ago that global warming is happening and will impact

[00:05:59] [SPEAKER_03]: all our lives.

[00:06:01] [SPEAKER_03]: That was a factor as well, but the primary motivation was wanting to understand this incredibly

[00:06:06] [SPEAKER_03]: complex system.

[00:06:07] [SPEAKER_06]: And what a complex system it is indeed.

[00:06:10] [SPEAKER_06]: I'm curious, what has been the trajectory of AI into the work that you've done?

[00:06:15] [SPEAKER_06]: Right obviously statistical analysis has been a thing we've had good old fashion AI as

[00:06:19] [SPEAKER_06]: well.

[00:06:20] [SPEAKER_06]: And now of course we've got this generative AI deep learning wave that's happening.

[00:06:23] [SPEAKER_06]: I'm curious what's been the trajectory of AI?

[00:06:26] [SPEAKER_03]: I started out in working in a biophysics group in the first wave of neural networks

[00:06:31] [SPEAKER_03]: in the 1990s.

[00:06:33] [SPEAKER_03]: So I had some early exposure to the early days of machine learning as it was then on.

[00:06:39] [SPEAKER_03]: But what really fundamentally changed for me was when I decided to work on more smaller

[00:06:46] [SPEAKER_03]: scale processes, which are the processes that are most uncertain in climate models,

[00:06:51] [SPEAKER_03]: and wanting to use data much more extensively than they have been used before.

[00:06:56] [SPEAKER_03]: I think that's when I really started to think about how we can use data,

[00:07:00] [SPEAKER_03]: started collaborating with one of my colleagues at Caltech and Restored trying to form

[00:07:04] [SPEAKER_03]: light ways we can use data well for climate purposes, which is quite different from

[00:07:11] [SPEAKER_03]: many other applications of machine learning.

[00:07:15] [SPEAKER_06]: Can you talk a little bit more about these data sources in how those have been evolving?

[00:07:18] [SPEAKER_06]: What's the kind of data that's most useful in your applications?

[00:07:22] [SPEAKER_03]: So the true age of satellite observations of Earth's atmosphere,

[00:07:26] [SPEAKER_03]: ocean started around 1980.

[00:07:28] [SPEAKER_03]: And since then data volume has kept increasing exponentially.

[00:07:33] [SPEAKER_03]: Right now we're receiving from NASA alone about 50 terabytes of data every day from space alone.

[00:07:39] [SPEAKER_03]: And in addition we have sensors, autonomous vehicles and oceans and the like and it's truly,

[00:07:44] [SPEAKER_03]: it has become a very data rich field, the climate sciences in ways that they were not

[00:07:50] [SPEAKER_03]: 30, 40 years ago. So the way the data that we have are most commonly used right now is for weather

[00:07:57] [SPEAKER_03]: forecasting. So whenever you get a weather forecast, what has happened before you get it is that

[00:08:03] [SPEAKER_03]: weather forecasting center has is simulated all the data we have, we use that as initial condition

[00:08:08] [SPEAKER_03]: for forecast. This has led that really to a big jump in the quality of weather forecast that

[00:08:13] [SPEAKER_03]: many people don't quite appreciate. It's about 25 years ago, something called 4D variational

[00:08:18] [SPEAKER_03]: data simulation. It led to a big jump in the quality of weather forecast we have. What's interesting

[00:08:23] [SPEAKER_03]: is that in the climate space, so when you think about, now let's run the model, let's a bit like

[00:08:28] [SPEAKER_03]: a weather forecasting model but run it for decades or centuries and say what will happen decades from now.

[00:08:35] [SPEAKER_03]: Their data have been used much much less extensively, primarily to evaluate models to say this

[00:08:40] [SPEAKER_03]: is good or bad after the fact, but not directly to inform the model and that's the piece that we

[00:08:46] [SPEAKER_03]: want to change in this climate modeling alliance, the Klima project I'm leading is use the data

[00:08:51] [SPEAKER_03]: directly to inform the model to achieve higher quality of predictions and projections for the future.

[00:08:59] [SPEAKER_06]: Okay, to make this a little bit clearer, basically Topio is advocating for using all this observational

[00:09:05] [SPEAKER_06]: data that's just sitting around not merely to evaluate the models, based on the predictions

[00:09:11] [SPEAKER_06]: that they produce but to train the models themselves so they keep in mind all that historical

[00:09:16] [SPEAKER_06]: observational data to do better climate modeling. So I'm curious Topio, has the impact of these recent

[00:09:24] [SPEAKER_06]: AI developments altered your expectations for what's going to be possible with these models

[00:09:30] [SPEAKER_06]: that are specifically focusing on climate predictions on these longer time horizons?

[00:09:35] [SPEAKER_03]: Yeah, definitely. It's maybe useful to talk a little bit about how these models are being developed

[00:09:40] [SPEAKER_03]: and what's sure, what a day and my life used to look like. These models are essentially

[00:09:45] [SPEAKER_03]: solving Newton's laws and the laws of thermodynamics and other scale and a global grid. The problem

[00:09:51] [SPEAKER_03]: is that these the meshes of this grid have a size of some of between 10 and 10 kilometers as typical

[00:09:58] [SPEAKER_03]: today and there's a lot of stuff that are much smaller in scale than the meshes of climate model.

[00:10:06] [SPEAKER_06]: Okay, so what Topio is saying here is important. So let me break it down. Imagine the entire earth

[00:10:12] [SPEAKER_06]: as a giant puzzle. Climate scientists are trying to solve this puzzle but they only have very

[00:10:17] [SPEAKER_06]: large pieces to work with. Each piece representing an area of 10 to 100 kilometers. They use these

[00:10:24] [SPEAKER_06]: pieces to build a picture of how these climate systems work, fitting them together and seeing

[00:10:29] [SPEAKER_06]: how they interact. The problem is there are many tiny but important details like clouds

[00:10:35] [SPEAKER_06]: that are way smaller than a single puzzle piece. It's kind of like trying to see a butterfly

[00:10:40] [SPEAKER_06]: in a puzzle where each piece is the size of a car. In other words, the resolution of traditional

[00:10:46] [SPEAKER_06]: climate models is way too fuzzy to be able to discern these tiny important details in clarity

[00:10:53] [SPEAKER_06]: thus creating a lot of uncertainty and as you'll hear from Topio, AI is able to mitigate this problem.

[00:11:00] [SPEAKER_03]: And so what you have to do is find some empirical way of representing what clouds do

[00:11:06] [SPEAKER_03]: given what you know on larger scales, on the mesh size of the climate model. And that was

[00:11:16] [SPEAKER_03]: pretty tedious process and it has been reasonably successful but in this process and the

[00:11:28] [SPEAKER_03]: use in climate predictions. And what machine learning tools change is that for these small

[00:11:34] [SPEAKER_03]: scales now we can learn what they do from data. A lot of companies that are delving into machine

[00:11:40] [SPEAKER_06]: learning are experiencing this issue of they're not really grounded in the physics of the real

[00:11:45] [SPEAKER_06]: world right? I mean just to give you an orthogonal example of video generation, you know persons

[00:11:50] [SPEAKER_06]: running backwards on a treadmill or a glass is breaking and it's behaving like plastic because

[00:11:56] [SPEAKER_06]: you know this model doesn't understand sort of the cause and effect relationships and sort of the

[00:12:01] [SPEAKER_06]: the rules of physics that govern that environment, what's the type of work that y'all are doing to

[00:12:06] [SPEAKER_06]: anchor these predictions that you do into the physics of the real world? Yeah that's a good

[00:12:12] [SPEAKER_03]: analogy actually. I mean what you don't want is a climate model that hallucinates physics right?

[00:12:17] [SPEAKER_03]: And for a video or even for weather forecast if there's something wrong whether it looks funny

[00:12:23] [SPEAKER_03]: and you correct it right? For a weather forecast if it's wrong to often be don't trust it and go

[00:12:27] [SPEAKER_03]: to the different source. The additional challenge of course for climate is that you do not have

[00:12:33] [SPEAKER_03]: an easy validation case, you do not immediately know when something goes wrong if it takes years

[00:12:39] [SPEAKER_03]: or even decades for these changes to become manifest. So what do we do to deal with this problem?

[00:12:46] [SPEAKER_03]: What we do is we use the laws of physics that we know and embed machine-running tools

[00:12:53] [SPEAKER_03]: inside the laws, inside conservation laws and that gives us a few-wish an insurance policy

[00:13:00] [SPEAKER_03]: that what we produce is reasonable and even more to the point, we want to predict there's

[00:13:07] [SPEAKER_03]: something we have no data for. We don't have data for the future and the future can be entirely

[00:13:12] [SPEAKER_03]: different from what they currently seeing. So it's known as the Out of Distribution Challenge

[00:13:16] [SPEAKER_03]: machine-running. So by using physics as far as we can, it helps with this Out of Distribution

[00:13:25] [SPEAKER_06]: Challenge as well. That's a really good point. Obviously the future is unknown but this is like

[00:13:32] [SPEAKER_06]: as close as we can get to a real crystal ball. So with advances in you know we're obviously

[00:13:39] [SPEAKER_06]: sensing the world in greater fidelity and greater frequency. We've got these, we've got this

[00:13:44] [SPEAKER_06]: beautiful tool that is machine learning and you're anchoring it in the laws of the real physics

[00:13:49] [SPEAKER_06]: that govern the real world. I'm curious what are your hopes for what these models will be able to do

[00:13:55] [SPEAKER_03]: in the very near future? Yeah so maybe let's start from what we need and then say how we get there.

[00:14:01] [SPEAKER_03]: I think what we need is assessments of risks extreme weather, extreme climate risks for the next

[00:14:08] [SPEAKER_03]: two decades. Climate is changing, we need to mitigate as much climate changes we can but some

[00:14:14] [SPEAKER_03]: climate changes unavoidable. We need to get ready for what is coming and build our infrastructure

[00:14:18] [SPEAKER_03]: so that they're right sized and cost effective for the world we'll all in and have it 10-20

[00:14:24] [SPEAKER_03]: years from now. So you need two things. You need to reduce model errors and you want to quantify

[00:14:32] [SPEAKER_03]: uncertainties errors so that we can take them into account and planning decisions if you're an

[00:14:36] [SPEAKER_03]: engineer building stormwater management system you don't want to just know the mean rainfall or

[00:14:44] [SPEAKER_03]: any kind of expected values but you want uncertainty ranges you want risks. The biggest uncertainties

[00:14:49] [SPEAKER_03]: in climate modeling come from these small skill processes. For example the clouds dominate

[00:14:53] [SPEAKER_03]: the uncertainties in climate predictions. We don't quite know what clouds will do under global

[00:14:57] [SPEAKER_03]: warming that dominates uncertainties and how climate will change. And the machinery tools

[00:15:05] [SPEAKER_03]: we're talking about I think if huge potential there one concrete example that turns out to be

[00:15:10] [SPEAKER_03]: important is how does a cloud exchange error with its surroundings through turbulence? Turns out

[00:15:17] [SPEAKER_03]: to be it's what are the key controlling factors for how clouds behave in a global warming

[00:15:22] [SPEAKER_03]: that processes hard to measure. It's even hard to simulate precisely and infer from simulations

[00:15:27] [SPEAKER_03]: but they're machinery tools that allow you to indirectly infer what that process looks like

[00:15:34] [SPEAKER_03]: that you can then use an climate model and achieve large error reductions or colleagues at MIT.

[00:15:41] [SPEAKER_03]: If developed in ocean model similar story there is small scale turbulence and in ocean you can

[00:15:46] [SPEAKER_03]: similarly learn from data how to model that in the context of this large physical model

[00:15:53] [SPEAKER_03]: and then we are still talking about perhaps order 10 kilometer scale resolution you still need

[00:15:58] [SPEAKER_03]: to get to this kilometer scale and here's another really good use of AI tools it's for

[00:16:04] [SPEAKER_03]: scaling or super resolution if you wish that you fill in the details that a climate model does not

[00:16:09] [SPEAKER_03]: produce using the AWS for the present climate climate projections for future climate to produce

[00:16:16] [SPEAKER_03]: the localized climate risk assessments that we ultimately need. So you brought up really two points

[00:16:22] [SPEAKER_06]: one is rather than getting this like macro scale picture we need to give decision makers

[00:16:27] [SPEAKER_06]: and local authorities this micro scale picture so they can it can be a lot more actionable

[00:16:31] [SPEAKER_06]: and then as we reduce the error the quality of predictions the accuracy of the predictions

[00:16:37] [SPEAKER_06]: will obviously go up will there be a feedback loop there where basically machine learning will

[00:16:42] [SPEAKER_06]: allow us to get better at predicting the future as we collect actual data or will that window

[00:16:48] [SPEAKER_06]: keep moving out and the future will always be unpredictable. No people get better and better at

[00:16:54] [SPEAKER_03]: predicting what will happen at least say for the next 20 to 50 years or so I think that's a good

[00:16:59] [SPEAKER_03]: time horizon to focus on we should be able to provide good predictions. The reason I choose

[00:17:04] [SPEAKER_03]: this time scale is that if you think longer term uncertainties in what we assume and do will

[00:17:12] [SPEAKER_03]: start to dominate how much you will be emit and that's obviously not something you can model

[00:17:17] [SPEAKER_03]: from first principles so that becomes a conditional for providing scenarios but for the next few

[00:17:23] [SPEAKER_03]: decades the uncertainties we have are dominated by model uncertainties and then by just

[00:17:27] [SPEAKER_03]: the natural chaotic variability of the atmosphere and oceans. So the model uncertainties we should

[00:17:33] [SPEAKER_03]: be able to reduce dramatically and then it's just no way to do just locally alone they

[00:17:38] [SPEAKER_03]: globe it's all interconnected but to get the local information then you need these downscaling

[00:17:44] [SPEAKER_03]: two worlds risk assessment two worlds so you need to build a valued chain of models that are interlinked

[00:17:51] [SPEAKER_03]: and in the end you want people to use all that information so that's one thing to say I have a

[00:17:56] [SPEAKER_03]: fantastic diffusion model for downscaling but it's quite a different story now to give this an

[00:18:01] [SPEAKER_03]: answer of people who need this information is small-town planar so you need to build good user

[00:18:06] [SPEAKER_03]: interfaces that make it very easy for stakeholders to access this information in their decision

[00:18:12] [SPEAKER_06]: workflow. I love your point about good user interfaces and it actually begs the question

[00:18:21] [SPEAKER_06]: how do we translate these insights that you're getting from these models into climate action

[00:18:26] [SPEAKER_06]: at that local level? Can you paint a picture of how those local decision makers would engage with

[00:18:33] [SPEAKER_03]: this type of data? So climate action I think we need to distinguish two pieces there is mitigation

[00:18:38] [SPEAKER_03]: so reduce emissions and there's adaptation adapt to whatever is coming. I think for the mitigation

[00:18:43] [SPEAKER_03]: part that's largely policy problem, mental technological progress problem and somewhere as we know

[00:18:49] [SPEAKER_03]: enough about the climate system additional information is probably not going to change the

[00:18:56] [SPEAKER_03]: public private sector organization that makes any decision of a future of a few decades

[00:19:02] [SPEAKER_03]: will have to adapt the climate change. In municipal planar we'll want to plan store some

[00:19:07] [SPEAKER_03]: water management infrastructure that's one type of information they need information on

[00:19:12] [SPEAKER_03]: precipitation extremes that get from now an architect building designer we'll want to build

[00:19:17] [SPEAKER_03]: a building in which it's still comfortable to be inside a future decades from now so they want to

[00:19:23] [SPEAKER_03]: temperature, probability of temperature extremes for long temperature extremes for some time

[00:19:28] [SPEAKER_03]: at the downstream end of this value chain what you need is an ecosystem of tools that

[00:19:34] [SPEAKER_03]: cater to different sectors specifically and meets people in their decision making process.

[00:19:42] [SPEAKER_03]: I think once you have those tools available that will trigger action right flood protection

[00:19:46] [SPEAKER_03]: of course another good example rising sea level increase the risk of storm surges with rising

[00:19:53] [SPEAKER_03]: sea level you want to know what those risks are and then for actively design your levies

[00:20:00] [SPEAKER_03]: your built infrastructure accordingly. I think that's that's a kind of climate action

[00:20:05] [SPEAKER_03]: that will be triggered with better signs and better information on the risks again there's

[00:20:12] [SPEAKER_03]: a whole mitigation site which is hugely important of course these scientific information

[00:20:20] [SPEAKER_03]: you know it's important but in some ways we have what we need to be no we need to reduce emissions.

[00:20:26] [SPEAKER_06]: Exactly I mean it's almost like we know what action we have to take to prevent climate

[00:20:30] [SPEAKER_06]: catastrophe and we have to do it now but we still have a hard time sort of visualizing the impact

[00:20:37] [SPEAKER_06]: of that action or really even the lack of that action in a very concrete way and so climate

[00:20:43] [SPEAKER_06]: change obviously has become really politicized or polarizing as a topic but maybe when you have

[00:20:49] [SPEAKER_06]: that model in front of you it might become more clear cuts so do you see a world in which these

[00:20:55] [SPEAKER_06]: models and applications built on top of them would make these risks seem more tangible and more

[00:21:05] [SPEAKER_03]: these risks to be tangible and easily accessible to people having say apps for consumers

[00:21:12] [SPEAKER_03]: for people can contextualize what's currently happening and put it in context of future risks

[00:21:18] [SPEAKER_03]: I think it would be tremendously helpful in raising awareness for what is happening

[00:21:25] [SPEAKER_03]: you mentioned polarization I have to say as far as the climate change questions are

[00:21:32] [SPEAKER_03]: still clearly polarized but I would say polarization is decreasing there's just a reality

[00:21:39] [SPEAKER_03]: that climate is changing at this point if you're a business and you make any decision that has

[00:21:44] [SPEAKER_03]: a reach of a few years to decades you would just lose money if you don't think about climate change

[00:21:50] [SPEAKER_03]: and that becomes the lowest common denominator that people can agree upon that they have to

[00:21:54] [SPEAKER_03]: worry about it I expect the polarization to for other decrease on this issue is simply because

[00:22:00] [SPEAKER_03]: climate change is happening and that's just a question what do we do about it

[00:22:04] [SPEAKER_06]: it's amazing that economic incentives are aligned here right and like to your point 10 and 20

[00:22:09] [SPEAKER_06]: years are a time scale which impacts all of us it's not this nebulous hundreds of years from now

[00:22:15] [SPEAKER_06]: thing that we're planning for which makes a huge difference I think before we get into mitigation

[00:22:22] [SPEAKER_06]: I do want to talk very quickly about like AI is like the panacea for everything these is like

[00:22:27] [SPEAKER_06]: well how do we solve this problem well AI of course and then AGI will magically come along

[00:22:32] [SPEAKER_06]: and solve everything for all of us in your mind what can AI not help us do from a climate perspective

[00:22:39] [SPEAKER_03]: I think the potential for modeling into like a huge for anything involving softer

[00:22:44] [SPEAKER_03]: but it's important to keep in mind that we live in a very material world match it in fertilizers

[00:22:49] [SPEAKER_03]: are produced by primarily by the Harvard Bush method which relies on natural gas and large source

[00:22:55] [SPEAKER_03]: of emissions steel production cement production right these are all enormous industries putting out

[00:23:01] [SPEAKER_03]: massive amounts of material those things are not so easy to change quickly this of course

[00:23:09] [SPEAKER_03]: aviation is also a good bit harder aviation is 2% of global emissions it's it's important but

[00:23:15] [SPEAKER_03]: if you can deal with everything else that's already pretty good but everything else is still a number

[00:23:19] [SPEAKER_03]: of fairly recalcitrum problems there so I think AI may not be the magic bullet for all of that

[00:23:24] [SPEAKER_03]: but that being said it can play a role in finding solutions now will the next large language model

[00:23:31] [SPEAKER_03]: find us a way of producing nitrogen fertilizers that doesn't involve using natural gas and

[00:23:37] [SPEAKER_03]: lead sub greenhouse gas emissions will be definitely not there yet there are folks out there that are

[00:23:42] [SPEAKER_06]: advocating for you know more it's like stronger intervention right and so there's been

[00:23:47] [SPEAKER_06]: recent popularity in cloud seeding in fact I'm aware of you know like 23 year old kids who are

[00:23:55] [SPEAKER_06]: doing startups that are going gung ho about this in El Sagundo and so I'm kind of curious you

[00:24:00] [SPEAKER_06]: outline the complexity of this problem of modeling this phenomenon weather phenomena at this like

[00:24:07] [SPEAKER_06]: it's very hard to model some a system this complex but does that end up having some like ripple

[00:24:12] [SPEAKER_06]: effect in some other part of the world like do we understand that I'm kind of curious what you think

[00:24:17] [SPEAKER_06]: about these other forms of climate intervention there's climate interventions on different

[00:24:22] [SPEAKER_03]: levels from speculating to things that can actually be done seeding clouds to make it rain in a given

[00:24:28] [SPEAKER_03]: place that has been tried since the advent of weather forecasting effect that was the original

[00:24:33] [SPEAKER_03]: hope when John for Neumann and other story by the forecasting programs in the wake of world

[00:24:37] [SPEAKER_03]: or two that it would lead to weather modification by on large these programs have not been successful

[00:24:43] [SPEAKER_03]: the other way in which you can't see clouds is under the setting of geoengineering very you

[00:24:49] [SPEAKER_03]: just changed the cloud cover of earth especially low clouds of the oceans that will reflect

[00:24:55] [SPEAKER_03]: more sunlight and that will offset some of the warming that comes from increasing greenhouse gases

[00:25:02] [SPEAKER_03]: there various ways of of setting warming from increasing greenhouse gases

[00:25:08] [SPEAKER_03]: seeding clouds it's fairly speculative it may work what almost certainly is possible is put air

[00:25:15] [SPEAKER_03]: sold so little particles could be a sulphate or silicon particles into the stratosphere as rockets

[00:25:23] [SPEAKER_03]: they would reflect sunlight it would lead to cooling offsetting some warming that seems

[00:25:28] [SPEAKER_03]: technologically feasible it's relatively clear that that it would be doable to do this

[00:25:35] [SPEAKER_03]: you would you could offset the warming what you could not offset easily are precipitation changes

[00:25:43] [SPEAKER_03]: that go along with global warming so the real challenge with these kinds of scenarios becomes

[00:25:50] [SPEAKER_03]: with a surge of controlling climate globally you know some warming and some parts of the world

[00:25:56] [SPEAKER_03]: may be quite desirable for agriculture and the and the architects say so people may not want

[00:26:04] [SPEAKER_03]: that warming not to happen of course by enlarge warming is not desirable it has all sorts of

[00:26:11] [SPEAKER_03]: attendant risks but then if you do geoengineering you might change the monsoon rainfall in India

[00:26:17] [SPEAKER_03]: and that will have severe implications that people there obviously wouldn't like so there is a governance

[00:26:23] [SPEAKER_03]: problem and an ethical problem was in charge there's the obvious moral hazard problem so supposed

[00:26:30] [SPEAKER_03]: we find ways of offsetting some warming does it give us a cut long to keep emitting then the risk

[00:26:37] [SPEAKER_03]: if you do this for a while any kind of geoengineering is that you're offsetting the warming say you

[00:26:43] [SPEAKER_03]: do it for 30 years you offset all the warming that would have happened in 30 years pretty something

[00:26:47] [SPEAKER_03]: at your point 60 degrees centigrade or so but for one reason or another you miss the fix of injecting

[00:26:54] [SPEAKER_03]: aerosols into the stratosphere or seeding the clouds and then you'll get all the warming that you

[00:26:59] [SPEAKER_03]: set over a timescale of a few months in one bag with ends which is what's called determination shock

[00:27:06] [SPEAKER_06]: so for those of you who haven't heard the term before or read the book termination shock is

[00:27:12] [SPEAKER_06]: kind of like putting the planet on a climate change pain killer if we suddenly stop the treatment

[00:27:18] [SPEAKER_06]: all that warming that we've been masking hits us all at once kind of like withdrawal but for the whole

[00:27:24] [SPEAKER_06]: planet it's the shock of terminating our quick fix hence the name so there is another

[00:27:33] [SPEAKER_03]: governance ethics hazard in that as well I think it's good to have that discussion on a society level

[00:27:42] [SPEAKER_03]: cloud seeding as a research program I support because it's really one of the big uncertainties

[00:27:48] [SPEAKER_03]: and climate protections is what will happen to clouds or how does pollution affect clouds for broadly

[00:27:55] [SPEAKER_03]: and at the very least these programs will lose a day at that question whether we should do

[00:27:59] [SPEAKER_03]: as a global scale my personal take us at least to be extremely cautious right we're messing with

[00:28:05] [SPEAKER_03]: the system we don't fully understand whatever we do will have global implications we don't have

[00:28:10] [SPEAKER_03]: effective governments mechanisms that seem very difficult you could find a globally optimal solution

[00:28:15] [SPEAKER_03]: if you can globally agree on the loss function to minimize all in terms of it then I think

[00:28:21] [SPEAKER_03]: it would be possible but that's a problem right I mean we won't globally agree on the loss function

[00:28:25] [SPEAKER_03]: the objectives for different countries different stakeholders will be very different

[00:28:30] [SPEAKER_06]: case and point China and India are already saber rattling over China's weather modification

[00:28:36] [SPEAKER_06]: program in these worried it's messing with their monsoons and the rivers that they share across

[00:28:42] [SPEAKER_06]: their borders which no surprise could have an impact on agriculture and food security

[00:28:48] [SPEAKER_06]: well I certainly hope that we come up with a globally optimal solution and doesn't take an

[00:28:53] [SPEAKER_06]: eccentric billionaire to just go you know have carte blanche and just start doing this

[00:29:00] [SPEAKER_06]: without the consent of this like planet that we inhabit together I'd love to change gears a little

[00:29:05] [SPEAKER_06]: bit just on the impact of AI systems right like there is on in terms of power consumption how do

[00:29:13] [SPEAKER_06]: think about this race towards more data more compute equals more intelligence we must do this

[00:29:20] [SPEAKER_06]: and the climate impact that this sort of race is happening having yeah climate science has

[00:29:28] [SPEAKER_03]: been big users who put computer sense exists so if your electricity users are by implication

[00:29:33] [SPEAKER_03]: and sometimes people joke we should just take the the attendance you two emissions

[00:29:39] [SPEAKER_03]: that combat climate simulations into a count strike in this emulation it's a joke it's

[00:29:44] [SPEAKER_03]: tongue and cheek I mean it's not that big in a global scale right but of course it's a serious concern

[00:29:50] [SPEAKER_03]: I mean the the civil lining here perhaps is that we're getting ever more compute for

[00:29:55] [SPEAKER_03]: the same amount of electricity energy use so that's the good news of course training AI models

[00:30:01] [SPEAKER_03]: right now is incredibly expensive given that it is so expensive you can simply scale it with

[00:30:10] [SPEAKER_03]: bigger computers indefinitely so we probably have to find more energy efficient more data efficient

[00:30:17] [SPEAKER_03]: ways of achieving what today's big models achieve and I think that's the new frontier built

[00:30:24] [SPEAKER_03]: models that are more energy efficient smaller and just as good and I think that's where we'll see

[00:30:31] [SPEAKER_03]: a we will see a lot of progress in it coming years that would be my expectation now to get two

[00:30:36] [SPEAKER_06]: science fictiony here but like you know data centers are already very resource intensive we're

[00:30:41] [SPEAKER_06]: going to keep producing more data to as a society that's certainly not going away how do you see

[00:30:46] [SPEAKER_06]: sort of like this this industrialization the next evolution of industrialization technology

[00:30:52] [SPEAKER_06]: getting built into the type of climate models that you're building like are you accounting for

[00:30:56] [SPEAKER_06]: this already like or um you know how does that how does that work when we do climate simulations

[00:31:02] [SPEAKER_03]: we take emissions scenarios as given so there is some economists social scientists scientists

[00:31:09] [SPEAKER_03]: and to like getting together and mapping out several plausible futures for emissions so

[00:31:17] [SPEAKER_03]: technological progress is expected and alike and what we take from that are these scenarios

[00:31:22] [SPEAKER_03]: and then we make climate projections conditional on those scenarios the unambiguously good news

[00:31:30] [SPEAKER_03]: is the rapid decrease of the cost of electricity that is renewably produced the cost of solar power

[00:31:40] [SPEAKER_03]: has decreased by almost a factor of 10 and roughly were last 15 years or so so so

[00:31:45] [SPEAKER_03]: what we need is renewably produced electricity and that's increasingly feasible

[00:31:51] [SPEAKER_03]: and that then can power large data centers so I think it is a science fiction scenario

[00:31:57] [SPEAKER_03]: who would have to worry that that data centers are kind of eating our climate future I think

[00:32:02] [SPEAKER_03]: I'm more optimistic there that's in fact there'll be renewably power it getting more efficient

[00:32:07] [SPEAKER_03]: and we can sustainably compute for whatever we need in the next few decades

[00:32:13] [SPEAKER_06]: what is your vision for the future right um you know on one hand a lot of people view AI itself

[00:32:18] [SPEAKER_06]: as a very polarizing subject some people are extremely optimistic about it to the point where

[00:32:23] [SPEAKER_06]: people are like we must keep accelerating accelerator die on the other hand people are like oh

[00:32:27] [SPEAKER_06]: no we got to pump the brakes are we deploying and proliferating this technology far too quickly

[00:32:32] [SPEAKER_06]: how do you think about this conundrum and on the spectrum I think my own vericus it's certainly

[00:32:39] [SPEAKER_03]: not representative but I think it is indicative of how AI can be hugely beneficial I think in

[00:32:45] [SPEAKER_03]: the in the long run for many for us it really takes a lot of the dreadree out of the day to

[00:32:50] [SPEAKER_03]: day work we can learn functions from data that before we had to guess by hand and it was tedious

[00:32:55] [SPEAKER_03]: it's increasingly coming a very efficient tool so I think I'm extremely optimistic and I think

[00:33:02] [SPEAKER_03]: in the long run this will increase productivity and I see some of that in my daily life now

[00:33:08] [SPEAKER_03]: and I'm very optimistic about that part of course it doesn't mean you know some jobs will

[00:33:15] [SPEAKER_03]: become less important others more important there will be winners and losers just as they were

[00:33:20] [SPEAKER_03]: at the beginning of the industrial evolution that will happen again I think I see huge potential

[00:33:25] [SPEAKER_03]: I think I see huge risk not using that potential in setting us back I'm less worried about

[00:33:36] [SPEAKER_03]: saying computers taking over those kind of scenarios they're not a large concern for me right now

[00:33:42] [SPEAKER_03]: what are you looking forward to? I mean for me personally of course as I want an amazing climate

[00:33:47] [SPEAKER_03]: model that reduces uncertainties quantifies uncertainties and then build verticals on top of it

[00:33:53] [SPEAKER_03]: that goes to local scale information that goes to to apps that consumers can use to to assess

[00:34:04] [SPEAKER_03]: compliance and to like I want climate information to be permeating economic decisions in a

[00:34:12] [SPEAKER_03]: rational and effective way and I think it's achievable one easy one is anyone who purchases

[00:34:18] [SPEAKER_03]: property right you'd like to know what the risk of flooding and wildfires in that area are

[00:34:24] [SPEAKER_03]: and like to know this granular accuracy and in a way that you can trust and I think that's achievable

[00:34:33] [SPEAKER_03]: and just simply you know when you talk with your friends about the weather today it's an unusually

[00:34:39] [SPEAKER_03]: hot day have contextual information how unusually hot is it in the past what does it's going to be like

[00:34:45] [SPEAKER_03]: 10 20 years from now just inform daily discourse with with that type of information I would find that

[00:34:51] [SPEAKER_06]: incredibly helpful too. Tapio thank you so much for your time it was a pleasure talking to you and

[00:34:55] [SPEAKER_06]: I'm really really really really enjoy this conversation and me too thank you thank you bill well

[00:35:06] [SPEAKER_06]: so after my conversation with Dion and Tapio there's a couple things I just want to stop and

[00:35:11] [SPEAKER_06]: marvel at number one our planet is covered in sensors let's appreciate that for a second it's

[00:35:18] [SPEAKER_06]: not just satellites and space and underwater drones it's light are measuring aerosols and clouds

[00:35:24] [SPEAKER_06]: it's a radar measuring ice sheet thickness the boat load of sensor data out there constantly

[00:35:30] [SPEAKER_06]: monitoring this planet below my mind and only now are we starting to fuse it all together

[00:35:36] [SPEAKER_06]: and number two we're building some predictive climate models using the same AI technology

[00:35:42] [SPEAKER_06]: we use for cellular whimsical things like AI art generators or 3D video games that's just cool

[00:35:49] [SPEAKER_06]: and a big reminder that it actually matters what we do with technology speaking of technology

[00:35:55] [SPEAKER_06]: if these models are accurately predicting climate disasters we may be tempted to use geol engineering

[00:36:01] [SPEAKER_06]: to modify the weather but whole boy does something like this require global coordination in lock

[00:36:07] [SPEAKER_06]: step otherwise I see a new kind of geopolitical crisis brewing in the future literally the weaponization

[00:36:14] [SPEAKER_06]: of weather instead we should think of this crystal ball as a sandbox for scenario planning generating

[00:36:21] [SPEAKER_06]: a model of what the earth could look like in the future if we make a set of decisions and put them

[00:36:26] [SPEAKER_06]: in motion we can use it like a canvas for global coordination so if we continue with the current

[00:36:32] [SPEAKER_06]: path of excessive emissions and energy blindness the models will map out just how dire the

[00:36:38] [SPEAKER_06]: consequences will be but hey the future isn't fixed and these models can also show what other

[00:36:45] [SPEAKER_06]: futures are possible if we take intensive coordinated action to mitigate some of the harm caused by

[00:36:51] [SPEAKER_06]: climate crisis that is if we shift away from fossil fuel towards clean energy sources if we

[00:36:57] [SPEAKER_06]: work to reduce emissions and preserve biodiversity these next generation models allows to travel

[00:37:04] [SPEAKER_06]: forward in time and visualize the impact of our actions to make a more tangible vision of the world

[00:37:11] [SPEAKER_06]: we actually want to live in the today I show is a part of the TED audio collective and is produced

[00:37:24] [SPEAKER_06]: by TED with cosmic standard our producers are Elefetta and Sarah McCray our editors are Ben Ben

[00:37:31] [SPEAKER_06]: Cheng and Elehanzar Sallazar our showrunner is Ivana Tucker and our associate producer is Ben

[00:37:37] [SPEAKER_06]: Montoya our engineer is Asia-Plaar Simpson our technical director is Jacob Winning and our executive

[00:37:44] [SPEAKER_06]: producer is Eliza Smith this episode was fact check by Dana Colashi and I'm your host

[00:37:50] [SPEAKER_06]: Belawville Sedu see y'all in the next one