This is an iHeart Podcast.
Guaranteed Human.
If I were to go back I don't know 30 years in Kenya.
What's the difference between then and now in terms of tree cover?
I'm talking to Philip Thigo, Special Technology Envoy to the Kenyan President.
No, it's big.
So if you think about, we are now 11% to 12%.
Previously, we were more than 20%.
So we are cutting trees more than we are planting them.
In 30 years, Kenya lost half its tree cover.
Half.
And here's where that matters.
Kenya is a mountainous country.
Dotted throughout the highlands are dozens of what Kenyans call water towers, natural reservoirs, densely forested areas capable of absorbing the enormous amount of water that falls on the country during the rainy seasons.
The tree roots and undergrowth secure and capture moisture, then slowly release it into the rivers that flow down into the country's low-lying coastal areas.
But in recent years the water towers have depleted.
Settlements have encroached on them.
Trees have been chopped down.
Thousands of acres cleared.
The natural reservoirs cease to hold nearly as much water.
So now, Kenya is prone to extremes.
Too much water flowing down from the highlands in the rainy season and too little water left during the dry season.
So you have a couple of hours of water, then you have a A couple of hours with no water.
The taps have to be dry by the city authority.
So that's the significance of the water towers we have when they cannot hold water.
Kenya desperately needed to restore its water towers by planting as many trees as humanly possible.
So in the fall of 2023, the Kenyan government took action.
It started a national holiday National Tree Growing Day a day to allow the citizens of Kenya to go out into the forests that dominate the Kenyan countryside and plant as many trees as they can.
And the government decided on a number.
The president's real focus right around how to ensure that we do not lose more forests was, in this very ambitious campaign, around 15 billion trees.
That's right, 15 billion, with a B.
So imagine that number will tell you the ambition, but also it tells you the deficit.
It has to be 15 billion in the next eight years.
15 billion trees over eight years averages out to more than 5 million trees per day.
That's a lot of trees.
But with such a massive goal, how can you track your progress?
How do you know where to plant those trees so they'll have the most impact?
How do you monitor where older trees are still being cut down?
Well, the answer to those questions came from IBM and a little space agency called NASA.
That's right, folks.
Smart Talks is going to space.
My name is Malcolm Gladwell.
You're listening to the latest episode of Smart Talks with IBM, where we offer our listeners a glimpse behind the curtain of the world of technology.
In this season, IBM has gone inside elementary school classrooms, toured formulation labs at L'Oreal and spoken with the fan development team at Scuderia Ferrari HP.
In this episode how IBM is partnering with NASA to build geospatial models using data from satellites to better understand our Earth and solar system.
5, 4, 3, 2, 1, 0.
All engine running.
Liftoff.
We have a liftoff.
32 minutes past the hour.
Liftoff on Apollo 11.
IBM has worked on space-related projects since before I was even born.
A team of 4000 IBM engineers helped create the Saturn V rocket that took Neil Armstrong to the moon.
And when I think of NASA, I tend to picture the moon landing, or the team of people back in Houston guiding the Apollo mission, or the Hubble telescope, or astronauts aboard the International Space Station.
What I didn't think about until now are NASA's geographers.
In order to go places, you need to map things.
This is Kevin Murphy, Chief Science Data Officer at NASA's Science Mission Directorate.
But I think that there's, you know, an assumption that NASA's all about rockets and astronauts.
And certainly that's a really large part, an important part of NASA.
NASA sends people to space and looks out at the stars.
But NASA also looks down at the Earth.
The agency has about 150 satellites that use radar LIDAR Landsat, Aqua Terra CloudSat Aura, low-Earth orbit, medium-Earth orbit, geostationary orbit on and on.
In one sense, NASA makes hardware.
They build rockets and spacecraft and all those satellites that circle the Earth.
But fundamentally, NASA also collects data.
Its scientists and engineers.
People like Kevin want to make the best use possible of all the information gathered by all those many dozens of instruments.
Right now, we gather around 25 petabytes of new observational data per year.
In the next couple of months, we're about to launch a high-resolution global radar.
When that launches, we'll double how much we collect every year, to about 50 petabytes of information.
Actually, since we recorded this conversation, NASA launched that global radar, what they call NISAR.
So NASA is already generating new data at the rate of 50 petabytes each year.
To put that in perspective, a single petabyte could hold about 500 billion pages of standard printed text.
Now, can anyone sort of apply to use this data?
You don't even have to apply.
It's free and open data.
It advances how we understand data.
What we do on Earth and how we see ourselves within the universe.
People can take it for so many different downstream applications.
So you can go to our websites today.
You can search through our tools and you can download information from the Mars rovers.
You can download information from the Lunar Reconnaissance Orbiter or any of the Earth science data satellites.
And give me an example of a really cool application, a really cool use that someone I don't know, an academic, whatever has used your data for.
Okay, so one of the really kind of cool but unexpected observations that we had is that we launched a pair of satellites in the early 2000s called GRACE.
And these satellites orbit the Earth.
And they can measure very precisely the distance that they're away from each other as they orbit the Earth.
And as you go into gravity wells you can actually see a satellite accelerate and the other one accelerate after it right,
And using that information, we were trying to map kind of the gravity fields of Earth.
What they found is that they can actually map below kind of the mass of Earth to where water storage is, for instance.
So aquifers right.
So you can monitor, through gravity, how much water is being depleted or added to an aquifer, or the density of glaciers.
So just to back up for a moment, The presence and density of water deposits below the Earth's surface have an effect on gravitational fields that are being measured in space.
Yeah.
And so does that tell you.
Presumably you learn things like where there's an aquifer, where you didn't think there was an aquifer.
Or if it's being depleted faster.
Yeah, yeah.
So who's using that kind of data?
All sorts of different organizations, whether they're, you know, NGOs or government agencies or people that are planning large agricultural products.
Was that an intentional discovery?
It wasn't.
It was accidental.
It was accidental.
NASA has assembled a historically unprecedented mountain of data about the physical world, free and open to anyone.
And the possibilities for how that information can be used are so vast that even NASA is still uncovering them.
When I was a kid, I loved Legos.
I had a huge bin full of them.
At the time, Legos were really just colored bricks of various sizes.
They weren't as complicated as they are today.
And what I realized even then was that there were more possibilities in a box of Legos than I could ever imagine on my own.
I'd play with my brother and he would show me something that hadn't occurred to me.
And I'd go to my friend Bruce's and see that he was off on some Legos tangent that I'd never even thought of, like a cool bridge or a castle or a truck.
I used Legos one way.
Bruce used his Legos in a completely different way.
NASA's Data Treasure Trove is like a very, very big box of Legos.
And here's the question.
With so much data containing so many possible connections, could IBM, and specifically IBM's artificial intelligence, help NASA scientists uncover patterns and connect systems in a way they've never done before?
Everything started with a question, right?
I'm talking to Juan Bernabe Moreno, director of IBM Research in Europe.
As we advance AI, we have new tools to understand the surroundings, understand the world, understand the language and understand our planet.
And the question that we were asking ourselves was all these new advances that we see in language.
It was a post-GPT moment.
Could we apply the same idea and the same architecture and technology to data about our planet?
The advent of AI created a new opportunity.
What if all of NASA's mountain of data could be organized analyzed, understood by artificial intelligence?
The original idea was to create a geospatial foundation model for the Earth and from there create additional specialized models for other scientific priorities of NASA.
And finally create an AI system that can understand all the data across those specialized models in order to uncover hidden insights and relationships.
Together, these models could unlock an infinite number of potential applications.
I asked Kevin Murphy at NASA about the beginning of these Earth models.
I have some colleagues and we were investigating a number of different avenues of using AI with our data, but also kind of the management and stewardship of the data.
So not only like the observations, but how we make it available to people, make it discoverable.
And they said, Hey, we see these transformer architectures.
We think that they can be applicable to some of the sequential observations that we make.
We'd really like to work with IBM on that.
And I was like, I'm really skeptical.
Why were you skeptical?
Because I hadn't seen those types of tools really produce results that were commensurate with the amount of effort you put into them.
So we were getting some really good results in deep learning approaches, but they took a lot of effort.
But Kevin came around quickly.
When we typically develop a new data product or an algorithm, it takes anywhere from 12 months, 18 months, 24 months to go from data and hypothesis to result which is validated.
We were able to get approximately the same precision for some well-known types of benchmarks Within.
I think it was about four months of starting the work.
Yeah, yeah.
So it happened faster than you thought.
Much faster.
In 2023, IBM and NASA launched a foundation model trained on NASA's harmonized Landsat Sentinel-2 satellite data across the continental United States.
They named the model PRITHVI, the Sanskrit word for Earth.
The first version of Prithvi used only Earth observation images.
And just that was enough to totally change Kevin's idea of what foundation models could do.
But they didn't stop there.
IBM and NASA were encouraged at how well Prithvi worked for Earth observation tasks.
So they decided to create a more complex version of Prithvi that could understand weather and climate data.
They hope this new version of Prithvi would allow researchers to answer new questions about the Earth, from short-term weather forecasting to longer-term climate effects.
Imagine you have a map of all the different temperatures pressures clouds, rainfall and more from around the globe.
With this map, IBM and NASA could implement advanced tasks.
They could track the formation of El Nino or predict how the path of a hurricane would change if the ocean temperature went up by half a degree.
I will always remember this moment.
It was when we created the weather and climate foundational model.
The senior meteorologist of NASA was like.
I cannot believe that has changed the way I think about AI.
And ever since, he's been kind of preaching with this example.
Juan and his team then took the model and decided to test it, really test it.
They took away 99% of the data points and ran the experiment again.
What they were trying to figure out is if the model had learned enough about the basic principles of the Earth, the underlying physics of the way the planet works, to fill in the blanks on its own.
With just 1% of the original data, would it still be accurate in its predictions?
What happened?
The model crushed it.
So it was able to extrapolate on the basis of 1% of the data what the entire picture looked like.
Yes.
Wow.
Because pre-learned everything, right?
Yeah.
It pre-learned the kind of principles of... Exactly.
Yeah.
Oh, wow.
That's very, very impressive.
At that moment when you realized you could do that.
I'm just curious about your emotional.
I mean, did you jump up and down?
What did you do?
It was a very emotional meeting because, you know, having this person say, now I'm convinced, right?
It was kind of a quite special moment.
These moments make your life as a researcher.
IBM and NASA launched Prithvi for weather and climate in 2024.
And while IBM and NASA scientists could use Prithvi to run interesting experiments, they were even more excited about how Prithvi could help people in the real world.
So let's go back to Kenya, Ambassador Philip Thigo and the country's great tree planting project.
So on those initial months there was a massive effort, including a couple of national holidays for tree planting.
Yeah.
Yes.
Where the entire cabinet was sent...
Did you plant trees?
Yes, I did.
Oh my God, I said the entire cabinet plus some of us, we have to be seen.
Are you good at that?
I planted two weeks ago.
Well, it's very easy.
Dig a hole, put a tree in the ground, cover it with a shovel, pour water.
Planting a tree is easy.
But remember, it has to happen 15 billion times.
IBM Research has been operating in Nairobi since 2013.
And what Kenya wanted, at least in the beginning, was straightforward.
The Prithvi model that IBM and NASA built could be used to essentially make the world's greatest map.
And Kenya, with IBM's help, could use that model to make the world's greatest map of Kenya.
The first step was to lay a grid across the topography of the country, break the forest into manageable, bite-sized pieces, each of which could be analyzed separately.
So because a forest is massive when you look at it in terms of green, right?
But when you layer it, you're able to break it into pieces, like into boxes.
And for us that was important because then it's easy to tackle it when it's in a grid system than just as a massive forest.
So that was also what the model was able to do.
Then the model painstakingly sorted through each of those boxes and looked for what Philip calls hotspots.
So you can see, for example, very quickly which other areas are being eroded very fast and that you need to quickly protect.
And that's where you want to target, right?
I mean, it's not possible to do everything at the same time.
Do you have a definition of a hotspot and how many hotspots are there according to that definition?
Oh, there are a lot.
So we have more than 40 water towers.
And I'll tell you, all of them have hotspots.
And the hotspots in my definition are areas that are being degraded faster and in a very unusual way.
You can literally see how human activity is seriously degrading that particular area.
That if you do not have a direct intervention, we'll lose the entire forest.
So that's the hotspot for us.
Because think about cutting 100 trees a day and cutting a million trees a day.
So that's a hotspot.
You want to look at places where there's just unusually high activity of deforestation.
In a hotspot, the size of each box in the grid was 10 by 10 meters, about half a tennis court.
That's how closely they were examining the forest.
So very crudely.
The model ingests all of this satellite data and helps you answer some very specific questions like where should we prioritize our tree planting efforts?
Which areas down to an extraordinary level of specificity are eroding most quickly?
You know, all those kinds of practical questions about how to direct your strategy.
So if you think about a smart forest right and that's really for us we're calling it smart fencing, smart forest.
Everything now is smart because of AI.
So if you think about your usual, what you can see with your eyes, and then the satellite layer which just zooms in, and you see green.
So what the model has been able to do is to create a smart layer, right?
And in that smart layer you can actually see many things, from analytics to the grids, to a dashboard, to an alert.
So we're able to layer to those blocks.
You can quantify degradation by blocks.
You can match interventions.
You can match reforestation.
I asked Philip to imagine what it would have been like to attempt the tree planting project in an era before AI.
His answer was, plant 15 billion trees?
Restore the water towers?
Impossible.
With Prithvi on Kenya's side, though, it's really happening.
What should be clear by now is how versatile Prithvi can be.
Want to know how to combat deforestation?
Prithvi can model that.
Want to know when the best time in the year to plant your crops is?
Prithvi can help predict that too.
Last year, six months after IBM started helping Kenya with reforestation, Kenya needed Prithvi's help on something else, and it was an emergency.
So something was happening in the world that we sort of had these floods that we didn't expect.
In the spring of 2024, Kenya was hit with thunderstorms and torrential rain.
Days and days of it.
And so I got a call from the Red Cross.
They're one of my friends.
And they're like Ambassador.
We need a little bit of help on how we deal with response, because what we're seeing is unusual, right?
Because normally you'd only have one area.
All of a sudden we had an entire country flooding in April.
We had about 3,800 kilometers squared, kind of total land flooded, which is unusual for Kenya.
And so when I got this call, we were like, okay, there's some work we did with IBM.
We only did one function for the trees.
It was actually a climate model.
And we said, can we use this to help us better respond to floods?
And so that was how we started having this discussion with IBM in terms of repurposing the model to help us deal with this new challenge around floods.
Again, Prithvi is versatile.
Prithvi could use everything it knew about the land, the forests and infrastructure to analyze how and where and when floods would occur.
The Kenyan government could then use the model to help the Red Cross organize its response.
Show areas that needed to be evacuated or safe places where the Red Cross could set up camps.
That information was invaluable.
Historically, what has happened is that they would set up camp based on population congregation where people assemble is where they set up a camp.
Not based on any data.
Simply because people are there, they will. come there to provide services and emergency response.
What we realized is that that model doesn't work.
So what we've been able to do with IBM is be able to sort of give Red Cross very specific locations or options where to set up camps.
So if people come here, just tell them, no move here.
That's the safe place. you really want to go.
So I think for me, that was really amazing.
So we are calling them, a very funny word for it, flood assembly points.
We always have fire assembly points, but now we can say we have literally flood assembly points that are safe for citizens.
That's fascinating.
So the model has ingested this incredibly granular picture of the topography and weather patterns of Kenya.
It's just giving you a set of useful predictions about how you should shape your response.
Yes, and what we did remember is that, as I said, it was a full multi-stakeholder capability.
What IBM gave us was a base map.
We didn't have that before.
And a base model.
So you kind of have these layers upon layers upon layers to be able to make intelligent decisions.
Throughout my reporting on this episode, I've been really impressed by what Prithvi can do.
But it doesn't stop at floods and reforestation.
Prithvi has also been used to look at wildfires and floods in the UK.
And Kevin told me that researchers in Africa have even used Prithvi to identify locust breeding grounds, which could help them prevent swarms that destroy crops.
But all these are issues on land.
I mean, I always say to people, 70% of our landmass is ocean.
Kate Royce is the director of the Heart Tree Centre, which focuses on adopting AI into UK's public and private sectors.
And one of those sectors is the blue economy.
Oceans, fish, shellfish.
But oceans are huge, and getting data from oceans is difficult.
So you're dealing with something where there's not a lot of people walking around collecting data.
So the real difficulty is understanding that collecting enough data to make anything make sense.
And oceans are very complex in terms of their interaction with our climate and how they interact with the climate.
So understanding the physics-based models is pretty challenging too.
Once again, enter IBM.
IBM created a new geospatial model to help us better understand our oceans.
Hartree and IBM, along with the Plymouth Marine Laboratory, the UK Science and Technology Facilities Council and the University of Exeter have all partnered to focus the model's power on the waters around the United Kingdom, which ultimately will help the UK's blue economy.
You get these major blooms in algae.
So the ocean goes green and you might see it in lakes as well.
Now, if you are shell fishing and that's what you're harvesting you can't harvest cockles mussels, to be very colloquial when you have algae blooms, because they're poisonous.
So there are certain times of the year where you can harvest and there are certain times of the year you can't.
If you keep having the algal blooms, just to put it on an economic terms, that's a problem.
So if we look at it that way, that's an issue.
So we really do need to try and understand where these algal blooms will happen, when they will happen and how to limit them.
Because obviously, if you're shellfishing as your livelihood, that's going to really impact you.
Kate told me that understanding these algal blooms how they form, why they form and how they move would allow people to better manage them.
What is it you're putting in the water?
Are you putting fertilizers in the water in the near shore environment that is causing those algal blooms?
Is it because we are heating up the oceans, and particularly our near shore environments, that is causing that?
I don't know.
I'm not a specialist, but that's what you're trying to figure out.
Is there something we are doing that is creating those environments, that is causing those algal blooms, or is it natural?
And natural is always a difficult one because I would say we live in a very managed environment.
Particularly in the UK, very few landscapes are natural.
Most of it is managed in some way.
Are we managing it in an appropriate way?
Is there changes in how we behave that could make things better?
Not that I needed more examples to sell me on how useful the Prithvi models are, but Kate gave me a few more use cases that reinforced just how exciting foundation models are for our oceans.
These big brown seaweeds can really help with carbon sequestration.
Imagine if we could improve the environment enough so that we could have more of that, so that we could sequester more carbon.
The other thing is wind power in the UK.
We have a lot of offshore wind farms and we're doing really well with our renewable energy resources.
So where do we put that and how does that impact sand movements?
So these sandbars and things aren't static, they move.
So understanding that is really important for where you're going to put your sub-oceanic infrastructure.
So you've got cables going across the oceans.
If we're going to use our oceans more, we need to understand what that environmental impact is going to be long-term.
The ocean model launched at the end of September 2025.
The research is only beginning.
When I sat down with Kevin Murphy at NASA, I wanted to understand where all of this impressive work was going.
And one of the signature aspects of this work is that it's not just for IBM and NASA researchers.
Anyone can use these models.
So before, if you were a researcher or, let's say, you were a farmer or maybe a technology-informed person that was interested in something like this, you would have to learn about how to do remote sensing, how to calibrate the imagery, how to stitch it together, because you know they come in kind of postage stamps that you have to squash together.
And then you would have to learn about the algorithms necessary to do all the processing, right?
So a lot of work.
And then you could actually do the mapping that you were interested in.
Today, what you can do is you can go to Hugging Face, which is where this model exists in the open, using kind of our open science principles.
And you can apply it to future or historical observations without having all of that background information.
And with the partnership between NASA and IBM, these foundation models are multiplying.
The new version of Prithvi I mentioned launched in September 2024.
Then, in August 2025, NASA and IBM launched another foundation model called Surya, based on data from the sun.
Surya can help predict solar flares which can disrupt communications and increase radiation for high-altitude flights.
And then there's the ocean model I talked about with Kate Royce.
So what does the future look like for all these foundation models built from NASA data?
If I wanted to look five or 10 years out to understand erosion patterns in a coastal town you could give me Eventually, I think we'll get there.
We've really only been doing this for the past few years.
There's a lot of, I think, capabilities to still discover and uncover with how we use these models for especially long-term predictions like you're talking about.
What do you think you can't do and that you'd really love to do?
What's the great white whale problem?
We can't do this today, but I'd like to be able to do it in the future, which is really the linking of the models together.
Right now we have these isolated areas where you know, we have the harmonized landsat sentinel or prithvi geospatial model.
We have the weather model um, which can look at short-term predictions.
We're building out um the heliophysics model to look at the sun dynamics.
But there are probably going to have to be additional models built so that we can understand how they interact with one another.
And that is kind of towards a digital twin of kind of the solar system or earth systems, which I think is a big hairy problem.
But if we understand it, we might be able to address some of the questions that you just asked about prediction.
So if you linked all of those models together, basically what you're saying is you say a digital twin, you're essentially replicating holistically how our world works.
Yep.
And do you think that is achievable?
I don't think it's immediately achievable.
Yeah.
But based on kind of the progress that we've seen in the last three or four years, I think it's more achievable today than it was then.
Do you think you'll see it in your... I sure am hopeful.
Yeah.
I sure am hopeful.
And I've got a couple of years left.
Smart Talks with IBM is produced by Matt Romano, Amy Gaines McQuaid, Trina Menino, and Jake Harper.
We're edited by Lacey Roberts.
Engineering by Nina Bird Lawrence.
Mastering by Sarah Bruguere.
Music by Gramascope.
Strategy by Tatiana Lieberman, Cassidy Meyer, and Sophia Derlin.
Special thanks to the team at NASA's Science Mission Directorate.
Smart Talks with IBM is a production of Pushkin Industries and Ruby Studio at iHeart Media.
To find more Pushkin podcasts, listen on the iHeart Radio app, Apple Podcasts or wherever you listen to podcasts.
I'm Malcolm Gladwell.
This is a paid advertisement from IBM.
The conversations on this podcast don't necessarily represent IBM's positions, strategies or opinions.
Since we recorded this episode, IBM and NASA released Syria, their solar weather model.
In early testing, it showed a 16% improvement in solar flare prediction accuracy.
This is the kind of improvement that helps protect our satellites, our power grids and our GPS systems from the sun's unpredictable nature.
And the next step in this partnership, another model coming in 2026, looking beyond the Earth and the Sun.
The universe of possibilities just keeps expanding.
