Saving the world one algorithm at a time | The Age of A.I.

1.43M views7284 WordsCopy TextShare

YouTube Originals

Many say that human beings have destroyed our planet. Because of this these people are endeavoring t...

Video Transcript:

Kurt Vonnegut said, "Science is magic that works. " I buy it. Makes perfect sense.

. . but it wasn't long ago that we couldn't understand what caused entire species to become extinct, why the ground shook, or crops dried out.

One of the promises of A. I. is that it'll enable us to use machine learning for prediction and conservation, anything from protecting wildlife to anticipating earthquakes.

Seeing the future may not prevent disasters, but I think we can all agree we need an equipment upgrade. What the heck. Let's give it a shot.

Is a sixth mass extinction on the horizon? "Reply hazy. Try again.

" [ranger] Are you ready? We are looking at the one with the collar. [Eric Dinerstein] There are about 250,000 parks in the world.

. . [low growling] .

. . and about 80,000 of them are under threat.

[Marc Goss] So this is a trail that's both being used by elephants and by poachers. We think so, yeah. [Dinerstein] We've come all the way over here to Africa to stop poachers before they kill by using artificial intelligence.

Fingers crossed that this works. [chirps] [elephant bellows] [ranger 1 speaking] [ranger 2 speaking] [Downey] Poachers kill roughly 35,000 African elephants every year. A single pound of ivory tusk can sell for $1,500, and tusks can weigh 250 pounds each.

Elephants are about a decade away from extinction, but it's not just about protecting one kind of animal or keystone species. There are about one million other species which are in danger of being wiped out. Animals affect vegetation, biodiversity shapes ecosystems, all of which, in turn, impact people.

It's all connected. So, it's not a stretch to say that protecting African elephants protects humanity and our future. [ranger 1 speaking] -One degrees?

-[ranger 2] One degrees. -Ten? -[ranger 1] Ten?

[ranger 1 speaking] The Mara Elephant Project is a protection organization that includes active anti-poaching work on the ground, and the research and monitoring of elephants. [Downey] Here in the grassy plains of southwestern Kenya, a team of just 50 rangers patrols an area of more than 3,000 square miles. [people singing in other language] To keep tabs on the larger bulls.

. . .

. . aerial teams tranquilize them, fit them with GPS collars, -and track them via satellite.

-[GPS beeping] It also allows us to look at that historical data and say, these are corridors. These are conflict areas. These are areas where elephants are at risk.

[Downey] In the hopes of identifying poachers, rangers have installed trap cameras throughout the park. A traditional camera trap system takes a photo of anything that moves. .

. [camera shutter clicking] . .

. which then results in all kinds of false positives. [Goss] The man hours of having to look through thousands of photos.

. . maybe one got a poacher, and it was three days ago.

[insects buzzing] That's too late for those animals. They're already dead. [Downey] Clearly, traditional camera systems are not enough.

[Anna Bethke] So these are all different. . .

[Goss] Snares. You have everything from barbed wire to cables, and then when an animal comes through, the snare closes onto it. -These are for killing, for bushmeat.

. . -[Bethke] Okay.

. . .

but you find often elephants also go through, and you see they've cut the trunks off, um, which is a really sad one. If we are able to get those poachers before they've killed the wildlife, red-handed with their snares as they're coming in, we're really now getting ahead of the game. Every one of these, if we hadn't removed it, costs an animal its life.

[Bethke] Jeez. I am the head of A. I.

for Social Good at Intel. It's the conjunction of AI technology with socially impactful projects. [Goss] These are drop spears, specifically for elephants.

[Bethke] Oh, my gosh. [Goss] So what they do is they put a hoist on it, and they hoist it right up in the top of a tree, really, really high up, and then there's a trip wire, and then when the elephant walks underneath. .

. It's nasty stuff. It's a problem that's not going away.

[Bethke] What I do is talk to people and organizations that have a mission, and I connect them with the technology that they need to see their missions through. [Dinerstein] How's it going? [researcher] Hey, Eric.

Good, good. [Downey] Intel partnered with Resolve, an N. G.

O. that uses technical innovation to solve some of the planet's more pressing environmental problems. [Dinerstein] Resolve is a small non-profit.

We look at trying to save the world's large mammals and prevent the sixth mass extinction from getting any worse. [Downey] They began developing an A. I.

-powered anti-poaching device they call "TrailGuard. " [Bethke] TrailGuard at its core is a motion-capture camera designed to help prevent poaching. Thanks to the amazing invention of a new chip, it brings artificial intelligence to the edge, to the chip itself.

[Bethke] So what we have done is input this very powerful computer chip called a Vision Processing Unit. All the pictures that the camera trap is taking will go through this VPU chip, and it'll figure out if there is a person present or not. .

. Okay. And only send you the pictures there where there's a person.

So it's gonna reduce, like, 95% of that noise -that you're getting. -Wow. [Bethke] It has an A.

I. algorithm in it which looks at every single picture, and sees if this is a picture that the park rangers are interested in, and right now, they really just want to know if there is a person going past. [Downey] The algorithm is fed thousands of images of both humans and animals.

It analyzes body shapes, facial geometry, movement, and other features until it learns how to distinguish one from the other, from any angle, in any light. [Bethke] It takes a picture whenever any motion happens in front of it, but the A. I.

is going to send only images of people going past. Image recognition is looking at an image and understanding what's in it. It's the perfect example of something that we take for granted.

Human beings can just, you know, look at an image and understand what's there without even thinking about it, but it's actually an incredibly hard problem. Amazingly, we actually have computers now who, at least within certain parameters, can do this as well as human beings can. [Dinerstein] The beauty of this system is, here in the Mara, the amount of time it takes from when the poacher walks in front of the camera to when it says, "That's a human," and sends it to the headquarters.

Under two minutes. We would be able to then deploy a ranger team to that site, and hopefully make an arrest before they've even had a chance to come in and kill those, those animals. [researcher] We need to lower the detection probability.

. . Right.

. . .

so we don't miss anything. [Dinerstein] We don't mind if we get false positives, but we don't want to get false negatives, that is, a human going by, and the A. I.

did not detect it. [Bethke] What I think I'm worried about is that at night, it's gonna be even harder for it to pick it up. One of the hardest things with deploying technology is you're never exactly sure how it's going to work in the field.

There are a million different things that could go wrong. [Downey] Anna, Eric, and the rangers head out to the Mara Reserve to set up the cameras for their first ever real-world test. [Dinerstein] The parks in Africa are really large because the big mammals that live there need really big spaces in order to survive.

That really complicates the problem of trying to protect them. . .

but the reality is that poaching incursions are not random. In fact, poachers actually try to follow trails, because they want to get in and out as fast as they can. [Bethke] This is beautiful up here.

[Dinerstein] There's about ten major access routes that account for 80% of all the poacher traffic. [ranger Wilson] I think this is where the elephants normally come. [Dinerstein] What might be more inconspicuous?

-What do you think? -[ranger Wilson] Mmm. .

. I want to hide it behind this branch. .

. [Dinerstein] Yeah. [ranger Wilson speaking] [Dinerstein] Excellent.

[Bethke] This looks awesome. So, you have the view. Poachers would not even suspect that there is a ranger or a camera here.

This is a hidden ranger. [chuckles] -[Bethke] A hidden ranger. -[ranger Wilson] Yeah.

[ranger] So tonight, my friends and I are going to play a little bit of a cat-and-mouse game with the cameras they put out. We're going to mimic the actual poachers. Let's see if the cameras catch us.

[Dinerstein] Does the artificial intelligence work at night? Which is the critical thing, because a lot of poaching happens at night. [Downey] Night is also more difficult for image recognition.

Details and features are harder to see, so for this test, the A. I. will have to rely on a narrow data set.

[Dinerstein] We knew that it works in the lab, in the United States, but that's not the same. [Goss] So this is a trail that's both being used by elephants and by poachers. We think so, yeah.

[Downey] 12 miles away, at MEP headquarters, the team waits for a trigger from TrailGuard. [suspenseful music playing] All right, well, here I see four images have come in. [Downey] The A.

I. 's computer vision has detected something. Person, animal, or other?

So, let's check the first one. Fingers crossed that this works. -[Goss cheering] Hey!

-[Bethke] Yes! There it is. Well done, guys, well done.

-[Bethke] Awesome. -[Wall] We've got our poacher. [clapping and celebrating] I am, like, over the moon.

I'm so excited that the demo worked. -[Dinerstein] Oh, great! -[all cheering] -[Goss laughs] -[Wall] Two for two.

Look at that! [laughs] [Bethke] Yes. .

. Oh, it's so good. [Wall] Nice, four for four.

That's pretty good. Catching poachers has really been a sort of cat-and-mouse behavior, if you like. The TrailGuard, really, is giving us this new edge.

We will get the poachers now before they've even had a chance to come in and kill the wildlife. [Dinerstein] This is a global problem. There are many more bad guys out there than we have the good guys defending these parks.

Right now, we can help tilt that balance back to the good guys to save the world's large mammals and prevent the sixth mass extinction. [Downey] The mass extinction is when most of the Earth's animals and plants die out, wiping out biodiversity. Probably the best-known one was 66 million years ago, when a giant asteroid hit Earth, decimating everything.

[howling] Now, we're arguably in the midst of a sixth mass extinction. Poachers are part of the problem, but so is climate change -and what we eat. -[cows mooing] All of us.

Can we rewire our destructive behavior? [man speaking] [Matias Muchnick] Sergio Barroso, he is one of the most talented chefs in the world. He has taste buds that no one else has.

[speaking Spanish] [Downey] Okay, how about this one, what do a Chilean chef, his insanely sensitive taste buds, and a robot named Giuseppe have to do with saving the environment? Good question. [speaking Spanish] [Downey] This meal might just hold the key to saving our planet.

It's the ultimate test. [upbeat music playing] [Muchnick] We have very ambitious dreams. We do want to change the world, but we're not dreamers.

. . We're more of doers.

So let's add 90 grams. [Muchnick] Animal-based protein food has been our main source of nutrition for the last, you know, thousand years. I actually grew up eating steak, but the way we're farming it.

. . [cows mooing] it's not the right way.

The food industry has become the common denominator to every major environmental ill known to humankind. . .

deforestation, water scarcity, world hunger! Unbelievable, right? [Downey] Eating meat causes climate change?

Sure. Cows and other animals emit methane, a harmful greenhouse gas. One third of the world's farmable land is used to feed livestock.

Look at it this way, eating one burger has about the same environmental impact as driving a gas car for ten miles. So, what do we do? [Muchnick] There is a new way, a better way of making food, and that's plant-based.

[Downey] Matias partnered with two PhD pals to co-found NotCo, an A. I. start-up with a humble mission.

. . Save the planet by reducing meat consumption.

[Muchnick] So we can see here that most of the ingredients make a lot of sense. Cabbages, rice, pumpkin, sunflower oil, pea protein. [Downey] Their challenge is not creating a plant-based alternative to popular animal proteins.

Plenty of companies already do that. So here many formulas are being generated and seems that they make sense, right? [Downey] NotCo is trying to crack something a little more elusive.

. . [Karim Pichara] Seems that he's using three different oils.

[Downey] . . .

taste and perception. Using an A. I.

algorithm they call "Giuseppe," they're trying to make people think they're eating steak, or eggs, or milk, when they're actually. . .

not. [Pichara] He was trying to find the omega-3 from a mixture of three different oils. [Pichara speaking] .

. . a technology that is able to tell us how to reproduce an animal-based food just by using plants.

[Muchnick] The algorithm is able to understand that there are clear connections between the molecular components in food and the human perception of taste and texture. [Downey] The magic behind Giuseppe is all about chemistry. The A.

I. looks at the molecular makeup of foods, like, say, milk. It then creates a list of ingredients from its most basic building blocks.

Finally, using machine learning and a massive database, Giuseppe then recombines select elements from plant-based foods to recreate the taste, and texture of the original. [Lav Varshney] Humans are good at reasoning about two ingredients or maybe three ingredients at a time, but after that, it becomes very difficult for us to think about it, but the machine can start thinking about five ingredients, ten ingredients, and how they all go together, and what the flavor profiles will be, and that's really the great power of the machine. [Pablo Zamora] What are you doing?

You're working on the fibers? There are a lot of similarities between plants and animals, because we share part of the chemical nature. All of them have DNA, all of them have RNA, all of them have proteins, lipids, and carbohydrates.

An almond and a walnut share, like, 97% of the molecules are the same molecules. Just the 3% give the identity to your brain that the walnut is a walnut, and an almond is an almond, so we need to really identify the molecular features that give the identity to food to pick up the specific type of plants and the specific ingredient from plants to rebuild. Yeah.

Smells like sushi. Not fresh sushi. [chuckling] [Zamora] I mean, that is kind of the magic behind NotCo.

We are building an ecosystem that really will see food in a unique way. . .

because Giuseppe suggests really. . .

crazy ingredients sometimes. [blender whirring] [Downey] The first food they tried to recreate is one of Chile's favorite condiments. .

. mayonnaise. [Zamora] We identified that cabbage, in a specific environment, will release a molecule that is really similar to lactose, and to your brain, it's kind of the same thing.

[Muchnick] We tasted the emulsion, and we said, "It tastes exactly like mayo. . .

" but it's red. The algorithm still didn't understand that one of the characteristics that we value in the sensorial experience of a product is color. [Downey] So they solved taste, and texture, but kept tweaking their algorithm until Giuseppe came up with a better formula for color.

[Muchnick] We are the third biggest mayo suppliers in Chile, and the best-selling mayo that comes in squeeze bottles. [Downey] The success of Not Mayo led to the creation of Not Milk, Not Ice Cream, and Not Meat. Now, they're working on Not Tuna.

We have devastated oceans. [seagulls crying] Generating a replacement for tuna is something that will move the needle in the world, so fish makes sense. Victor, if you can go to the berries, and extract the color, please.

We need to make a blue tuna. It's the. .

. the normal tuna. We have a cabbage that is prepared like kimchi.

[Muchnick] It's the looks, it's the taste, the aftertaste, and other dimensions on, uh, on sensorial experience. That is key, man. [Pesse] We can add this to make an emulsion.

[Muchnick] If we don't do things that are as good or better, things won't change, and we won't move the needle. So, this looks like tuna color. -Yeah, really good color.

-It's really concentrated. . .

Yeah, I think that is good. It's quite similar. One of the things with large-scale behavior changes, it's often not driven by nutrition or sustainability considerations, it's driven by flavor.

I think everyone has this experience of wanting to try the new chip flavor, and if we can have that same property in healthy and sustainable foods, that can be really powerful. -[Giuseppe] Hey. -[Pesse] Hello.

So these are the first formulas of the Giuseppe Tuna. [Downey] Today, the NotCo team is taste-testing two of Giuseppe's tuna recipes. What do you think?

It smells pretty fishy. You know, it's like, it's like, um. .

. It's like this kind of like ceviche smell, like and some algae. .

. It's really the very first time that we ever tried a fish on Giuseppe. The color is from berries and the kimchi that we used.

Wow! The texture is good. Mm-hmm.

The taste is a little bit lacking. [Muchnick] The algorithm is not ready yet to understand the complexity of a muscle. How do we explain that to an algorithm?

Can we try that one? Mm. I like one this one better.

Like, the flavor of this is better. -The flavor is way better. -Taste.

. . [Zamora] But the texture for tuna, it's this one, but that one is better in taste.

Wow, I'm impressed, for being the first shot. When Giuseppe suggests something, it's the first iteration, right? We need to keep running this.

I mean, there is where the scientists take control of the process, and they try to really dig in, into the molecular, structural component of the food. In terms of cans, I think that we're nearby, you know, we are really close. Not Tuna is the new Not Chicken of the Not Sea.

92% of our consumers are non-vegetarian. They don't care about sustainability. What they care about is eating tasty food.

[Muchnick] Everything that Giuseppe suggests, all the recipes, all of the products end up here, and this is basically our R&D, product development facility. [Muchnick] Sergio Barroso, he is one of the most talented chefs in the world. [Downey] That's not an understatement.

Barroso cut his teeth at El Bulli, in Spain, arguably the best restaurant in the world, and his own joint in Chile is top 50. He's a culinary innovator, can taste flavors invisible to the average human tongue, and carries around his own personal spatula, gunslinger-style. [Pesse] Oh, you have your spatula.

Yes, always, always. You never know when you have to taste something. [laughing] [Downey] Suffice to say, NotCo couldn't have picked a more discerning palate to put their latest A.

I. foods to the ultimate test. The most important part for me is the flavor, because without flavor, we don't care the other part.

The acidity and the oiliness is very, very balanced. You would never think that this is doing without eggs, you know? [Pesse] Okay, now, Sergio, you will taste the Not Milk.

[Sergio chuckles] Okay. Cheers. Salud.

First, the texture is like milk. Yeah, is very tasty. Is very, very, very tasty, and I think is more flavor than the milk that you can buy at the supermarket.

So the Not Milk is more milk-- -Yes. --than the milk. More milky!

[Muchnick] We started with a plant-based burger, because we thought, a burger is not as complex as, as a steak. The texture is the same, the same as when is cooked. [Barroso] When I see something like this, one part of my mind is thinking "What.

. . what can I do with this," no?

So I want to try to, to cook with these, with the milk, with the meat, with the mayonnaise, with everything, no? I think I can prepare something very, very special. [Muchnick] Fantastic.

Challenge accepted. [laughs] [Yann Yavin] Salud! -Sante!

-Sante. . .

[Barroso speaking English] This was planned one day prior to this, so Sergio is with his team, preparing an eight-course dinner for the very selected group of people. Expectations are very, very high. Anxious, nervous, you know, getting, uh, mixed feelings of excitement.

[speaking in Spanish] [Barroso] We are going to eat everything in the same way that we prepare it in a normal recipe in the restaurant. We would put one liter of milk, we are going to put one liter of Not Milk. This is the ultimate challenge for us.

[Downey] At the end of the day, if NotCo's gonna take off, their food just has to taste good for everybody. . .

A. I. or no A.

I. I like to start the menu like this-- [slaps palm] [conversing in Spanish] [Barroso] We want to say in the first tapa that this is something serious. [De Vicente speaking Spanish] Oh, my god.

Oh. . .

I couldn't believe that it was not mayo. [people conversing in Spanish] [Muchnick] He's using Not Meat for dumplings. [Barroso] Normally with the real meat, you need three minutes, but you don't know, because this is the first time you cooked with a Not Meat.

[Muchnick] We're so nervous because this is going to prove that plant-based food can actually replace animal-based food. [man] Senoritas. .

. [people conversing in Spanish] My brain told me that I wasn't eating meat, but my soul told me that it was meat. We are going to prepare the, the last dessert with the Not Ice Cream.

We are doing something similar like a sweet taco with the cotton candy. [Barroso speaking Spanish] [delighted laughter] [speaking Spanish] [Barroso] They weren't thinking all the time that they are eating something different with "Not" products, and they were happy with the flavors, with everything, no? -[speaking in Spanish] -[applauding] It makes me feel fantastic.

Sante. Gracias a todos. Gracias.

-Muy bueno. Muy bueno. -Gracias.

[clinking glasses] [Muchnick] We do want to change the world. The dream is big, the dream is there. [Downey] The dream of NotCo is that it uses A.

I. to preserve our planet from the damage humans are doing. .

. but does it cut both ways? Can A.

I. protect people from destructive forces of nature? [Domingos] The laws of nature don't change.

Human beings change in response to A. I. .

. but nature doesn't. [rumbling] If you take earthquakes, if you can predict when earthquakes will happen.

. . this can save a lot of lives.

[Harold Tobin] It's been 319 years since the last big one here. Seismically, it's the quietest zone in the world right now, so the possibility that it's quiet because it's locked and loaded and ready to go, it certainly has freaked a lot of people out. [rumbling] [glass smashing, panicked screaming] [police dispatch] 25-238, we have a major injury accident.

They're advising major injuries. [radio messages overlap] We have a 50-foot section which has collapsed. [Tobin] If a magnitude-9 earthquake took place on the Cascadia fault, we'd be expecting tens of thousands of casualties.

Right now, we can forecast that big subduction zones or major faults will have an earthquake at some point in the future, but that's not what people want to know. They want to know is the earthquake coming imminently? In six hours?

In one day? In 30 seconds? [Downey] The Cascadia fault line, more than 600 miles long from Vancouver Island to northern California, could cause the biggest natural disaster in North American history.

It's generally accepted that The Big One is coming, maybe within 50 years, but experts can't be certain, or more specific. [Doug Gibbons speaking English] [Tobin] The PNSN is a network of seismic sensors, and the idea is that we are continuously monitoring for earthquake activity of all scales, all the times. [Gibbons] This is one of the 400 seismic stations in Washington and Oregon.

This is a strong-motion accelerometer, so it is looking for large earthquakes. [Tobin] Earthquake Early Warning system is not predicting an earthquake. [Gibbons] We're doing a stomp test.

I stomp on the ground, the instrument detects that shaking. [Tobin] It's identifying those first waves that are coming in. [beeping] [Gibbons] All 400 of our instruments are sending a constant stream of data back to our data center.

[Paul Bodin] In each of those 400 sensors, there's a lot of signals which aren't earthquakes. [Tobin] So we're generating this incredible volume of data. This is probably a truck going by.

[Gibbons] Where the A. I. comes in would be filtering out what we call cultural noise.

Trains, trucks, people. [Downey] Vibration info from all 400 sensors in the region is fed into a machine-learning algorithm, which is trained to differentiate earthquake tremors from, say, construction or buses. Using machine learning and a huge database of known sounds, the A.

I. can quickly sort through the noise of the natural world. [Bodin] Oh, there's some interesting stuff going on here, guys.

[Bodin] Machine learning enables us to quickly find the signal of earthquakes. We can do it much faster and better. [emergency dispatcher radio] [Tobin] We're working hard on an earthquake early warning system, and it's called Shake Alert.

The emergency management center here would be one of our key immediate users. -[alerts blare] -Earthquake! Earthquake!

If it's a big earthquake, then within a second or two. . .

[dispatcher] Unit clearing on five, identify. [Tobin] . .

. our computer algorithms determine what area it's going to affect. .

. [telephone operator] Damage assessment. .

. [Tobin] . .

. and then create a warning for that. Depending on how close you are to the source, the warnings can be very short, like less than a second.

. . [sirens wailing] .

. . to maybe as much as three minutes.

[deep rumbling] [Downey] . . .

but when a massive earthquake struck northern Japan in 2011. . .

[Downey] The residents had warning. . .

between 20 to 90 seconds, depending on how far they were from the quake's epicenter. . .

. . .

and yet, almost 16,000 lives were still lost. Is there any way A. I.

could buy us even more warning time by predicting the next big one? [Chris Marone] Earthquake prediction is a really hard problem, because we don't know where the earthquakes are gonna happen. Not really, right?

We don't know when they're gonna happen, and we don't really know how big that zone is that the fault reaches before it all of a sudden snaps. Well, what do you think, bud? [Sharan] Yeah, we can go ahead and actually make an offset right now.

[Marone] Yeah, let's do it. [Downey] Chris and his team have come up with a way to create earthquakes in the lab. You're at, uh, 2.

3 right now. [Marone] We're simulating the kind of thing that happens in the upper few miles of the Earth's crust. [Sharan] So I'm just gonna, I guess, set this up before we get started.

[Marone] The goal today is to try to figure out if we can create laboratory earthquakes under a range of conditions, and use machine learning to predict that whole range. [Downey] On a small scale, this test mimics what happens on a huge fault like the Cascadia. It's basically two massive tectonic plates, one under the coastal northwest, and one under the Pacific.

One slides under the other, and sometimes snags, causing friction and pressure. [deep rumbling] When it's too much, they violently slip, causing an earthquake. So that's one side, and then I basically repeat the same thing for the other side block.

. . [Downey] That's what Chris and his team are trying to replicate with mini granite blocks in the lab.

[Marone] We've got lab seismometers that are located right next to the fault zone. [Sharan] . .

. and I try to align it with the lasers. [Marone] Giant press, homebuilt, driven by hydraulics.

[Sharan] Now this is ready to go. I'm just turning it on right now. [hydraulics whirring] [Marone] In the lab, we're controlling the stresses on the fault, we're controlling how fast the stresses build up to some critical point where it's going to break.

There are groans and creaks on the fault. Micro earthquakes, a lot of little mini-failure events. Yeah, let's see.

Let's look at the acoustic side of this. [Marone] We're listening to everything that happens in the faults, and we're using machine learning to find the patterns in it. Is there something happening that we can use to say, "Ah, this is about to go big.

" -Yeah. -There's another one? So now we are actually recording.

We're hearing all that little chatter you hear. [rock cracking] [cracking] [Marone] Let's zoom back in, because we're gonna get a nice big one here right now. Here it comes.

[rock grinding and cracking] [thuds] -Bang. -There we go. Yeah.

That was big. [Marone] That yellow line is the measure of the shear stress on the fault. We're watching it build, and the whole time it's increasing, there are micro earthquakes that we're listening to, and what we've realized is that we can use those micro earthquakes along with machine learning to predict the time of the next event.

[Downey] These micro earthquakes that they're listening to are too faint for the human ear, but not for machine learning. It uses faint signals to predict earthquakes we can hear and feel. [Marone] That was a nice, big high-frequency event.

We're probably gonna get another one-- [rock cracks] -Ooh, that was a big one there. -That was big. Listening to the small events to teach us about when the big events are gonna occur.

No, that's fabulous. That's, that's so cool to see. Another one is coming up.

We're predicting the time of the next earthquake, and we're predicting the duration of it. How many cycles do we have already? [Marone] Something like 20, 30 events.

It's a game changer, using artificial intelligence, because now we can use machine learning to ask questions about why. . .

why is that happening? Is there some geometric structure that's building inside the fault zone, that is somehow seen by artificial intelligence? Machine learning prediction is here.

The red line is the experimental data. Every time it drops, that's a lab earthquake, and then the blue line is a model prediction based on machine learning. It's impressive how well it works, right?

Yeah, that you see the same thing in different cycles, exactly the same cluster appears just right before failure. [Marone] Yeah, you can see, you know, it's not perfect, but it's damn close to perfect. We've shown beyond a shadow of a doubt that A.

I. , machine learning, can predict laboratory earthquakes. We're probably going to get another one.

[rock cracks] Yeah, there it is right there. That's awesome. [laughs] Yeah, right?

The challenge is how do we scale that from what's being measured in the laboratory to what we measure with seismic sensors over an area of hundreds or thousands of square kilometers. [Varshney] As artificial intelligence goes from the lab out into the world, the world is much more complicated, and so there's a lot of factors that might not have been modeled. If we can transfer out, if we can generalize, that would be really, really amazing.

[Marone] How many years is it gonna be before we can predict or forecast something about Cascadia? I do think that's within our lifetime. [Downey] It's a bold statement, but one that no longer seems out of reach, especially as machine learning gets better at identifying patterns to better forecast earthquakes or other natural disasters.

. . and if A.

I. can predict calamity, could it go one step further, and prevent it? Machine learning is being used by corporations and governments to help solve large-scale conflicts and catastrophes.

One area that causes a lot of problems is food, or lack of it. [Mark Johnson] Agriculture is just a critical component of national security and the health of the world. Food shortages sometimes lead to famine.

Famine leads to political unrest. The way I see A. I.

is it's not a scary force to be feared, but something that's going to help us address these big problems. [Downey] Mark is the co-founder of a company called Descartes Labs. [Krisna] This is a max composite of the what the detector picked up from last winter.

This year, the whole place lights up. [Johnson] We're a young company. It's me and my merry band of physicists.

One of the coolest things we do at this company is to try to look at the bleeding edge of science, and try to figure out how that can affect our lives. [Mike Warren] There's a tremendous amount of satellite imagery out there. That's one of the biggest datasets that humanity has collected.

[Johnson] The most astonishing thing to me about satellites is that we have had excellent scientific information coming down since the '70s, and yet it's really hard to use that information. [Downey] There are thousands of satellites photographing Earth, so Descartes built a supercomputer in the cloud that uses machine learning to analyze these images and make models from the information. They're trying to predict when disease, disaster, or even war might strike.

[Johnson] Artificial intelligence has progressed so much in the past decade, and now you can start to build models on top of this data instead of just having a bunch of pictures that are hard to deal with. We've been trying to find all the solar panels in the U. S.

using this satellite imagery. [Johnson] So the first task, we built what we call a similarity search engine. [Downey] The search engine uses object recognition, a type of A.

I. that learns to identify and differentiate specific things within images. Instead of scanning for poachers in Africa, they're looking for everything from solar panels to riverbeds.

The computer is just seeing numbers. It does not see the image like we see it. It just sees raw numbers, so we teach it over time that's what a river looks like.

That's a street. That's a building, and the algorithm is able to, over time, if it sees enough examples of this, you know, it says, "Yeah, I get it. " [Johnson] So we started looking around for a problem that we could solve with this, and what we settled on was agriculture.

Plants are really neat because they're like these little factories, right? Just by watching how the light bounces off of these plants, you can tell a lot about the production of this factory. [Downey] For innovators, it's not totally unusual to find an answer, and then reverse-engineer a question.

The question Mark asked his A. I. was about corn.

How many cornfields are there, where, and what are their growing patterns. . .

and then comparing satellite images over time, can we predict how much corn the country would produce next year? At the beginning of the project, that was kind of a pipedream to see if that was possible. Could you use sensors flying around hundreds of miles above those cornfields, never having seen an ear of corn, and get a really good, accurate prediction?

We combined two things, so the satellite data will give you a sense of the health of the plants today. Then we looked at the weather data, which will give you a sense of how healthy the plants will be in a week or two. [Downey] In 2017, the U.

S. produced over two billion bushels of corn. Descartes' estimate was within one percent.

. . .

and I think what really shocked the industry was this wasn't a bunch of agronomists, these weren't people who were experts in corn. We were a bunch of physicists just using the principles of physics and light. This really woke up the market that data can really change traditional forecasting methods.

[Martin Ford] If you have lots of data, and you've got an algorithm that can analyze that data, it can learn things that no human being would be able to perceive, and it will be able to make predictions that no human being would ever have been able to make. [Downey] . .

. but it's not just corn. You can do this with almost anything that's photographed from space.

Water, forests, factories, roads. I'm looking at mapping rice paddies in Asia. The brighter yellow is a higher probability of being a rice paddy, and anything darker purple or blue is a low probability of being rice.

Can you tell the health of a crop at all? One of the ways we understand the health of crops is to look in the infrared bands. So this is beyond red in the electromagnetic spectrum, which tells you a lot about the health of crops, just, humans can't even do that.

This is in Iraq, where they grow rice. Most of the water in this area comes from snow melt. There's a lot of places where the food supply is highly dependent on the amount of snow that falls in the winter.

[Krisna] The blue here is like looking at snow from 2014, and 2015 was a drought year, and there was a lot less snow, and it correlates to how much rice there was in this region. 64% less rice. [Johnson] So if we know how much snow falls in the winter, the farmers can better plan their crops.

Maybe they plant drought resistant crops. This should give you some indication of later in the summer when the snow is melting, how much rice that area can handle. [Downey] Their A.

I. -based forecasts turned heads, and caught the attention of DARPA, the defense research arm of the U. S.

Military. DARPA is the Defense Advanced Research Projects. They look at the hardest problems that are gonna affect the military of the United States.

[Steve] We've been pretty heads-down on the science, but I know you work really closely with the defense department. Do you think this is going to influence their thinking about conflicts that may come, or. .

. Oh, absolutely. I think that, uh, having advanced warning of problem areas is.

. . is very critical.

[crowd chanting] You know, the Arab Spring was touched off by, you know, wheat shortages that led to shortages of, of bread. No food, no anything! Hosni Mubarak.

[crowds shouting] The problem that they challenged us with is looking at food production in the Middle East and North Africa, and the goal is to understand where there might be food shortages. [Steve Truitt] When you don't have water, when you don't have food, when you don't have shelter, you fight, right? You will do anything in order to survive.

[gunfire and shouting] [Evans] Really hope to, uh, get an update on current status. Okay, so we've been looking at this current year growing season. We're seeing a lot of healthy wheat fields in Syria this year.

It looks like they're, we're about 20%, um, above last year's production -in Syria. -This is a wonderful leap forward in the ability to get, um, you know, objective measures of what's really happening in agriculture in these troubled areas. [Johnson] It's really important to know where there's crop failure so we can send aid there before it gets really bad and turns into famine.

The fact that we can look across the world, and find where famine might happen four months from now is. . .

it's mind-blowing. [Johnson] It's pretty astonishing to me that you can look at the health of crops with satellites that are flying hundreds of miles above the Earth. [Downey] The idea that corporations and governments have technology that predicts, and maybe prevents, large-scale human catastrophes, like war and famine, is mind-blowing.

[Johnson] Computers are sorta like three-year olds right now, and we're training them to be our helpers, to help us make better decisions, to help us be better humans. I believe deeply that science is going to help us save the planet and save ourselves. [Downey] In the old days, people used to think disasters were caused by God or magic.

Now we know better. If a sixth mass extinction happens, it's probably gonna be on us. [Dinerstein] We have this technology that we can use to save life on Earth.

[Downey] A. I. might not prevent disasters, but new scientific tools, like machine learning, image recognition, and predictive modeling, might help us at least get ahead of them.

[Ford] Artificial intelligence, machine learning is gonna be a very, very powerful tool for making predictions at a precision that previously has been impossible. [Varshney] There's been a lot of projects in sustainability that are deploying artificial intelligence to address problems that the world faces. [Johnson] If we can have a better insight into what the future looks like for food, then we can really save a lot of people.

[Downey] Conserving the planet or preserving our species doesn't really have anything to do with magic. . .

but it would be divine.