Dena Temple Raston (2:34)

I'm Dena Temple Roster, and this is Click Here's Mic Drop. Paris Perdicaris has spent the past decade figuring out how AI interprets the physical world. Hurricanes, ocean currents, weather systems.

Paris Perdicaris (2:51)

My background over the last 10 years, I've done a lot of work on AI and deep learning applied specifically to physical systems.

Dena Temple Raston (2:59)

So it made sense that Microsoft would come calling as they were trying to figure out if AI could handle one of the Biggest data problems we have weather. Weather, it turns out, produces a ton of data.

Paris Perdicaris (3:12)

So Starting from the 50s, 60s and 70s, we have been collecting around the clock from satellites, from weather stations, air balloons, aircrafts, ships in the oceans, or buoys.

Dena Temple Raston (3:27)

So it's an information goldmine. Decades of public data gathered by agencies like NOAA or its European equivalent, the ecmfw. Big tech's breakthroughs depend on their treasure chest of data. So Paris built Aurora, an AI system that learns from all that weather data.

Paris Perdicaris (3:47)

The way a sailor predicts the weather is experience based, right? So they're looking back to their prior experience. They've been in the situation many times. They kind of feel the wind and see the clouds, and from that information, they can make a prediction of what's about to happen.

Dena Temple Raston (4:00)

Aurora doesn't understand the atmosphere. It just sees it in terms of patterns.

Paris Perdicaris (4:06)

And so far we have seen that on average, it outperforms all existing operational forecasts for hurricane tracking and can generate tracks, can predict accurately tracks maybe four to five days before those have an impact on land.

Dena Temple Raston (4:25)

Now, that sounds impressive, but there's a caveat. It appears to do well with these broad predictions like hurricanes. But critics say that broad accuracy is one thing and being able to spot the conditions that lead to, say, flash floods and microbursts is another. And those are the storms that seem to show up more often these days, so they've started to matter more. Microsoft isn't the only big tech player trying to use AI to predict weather. Google's been developing AI weather tools. And Huawei in China, it has too. They all feed on the same public data. Paris also added a few other data points to his model.

Paris Perdicaris (5:08)

We tried to go a step beyond that and use a lot more diverse data, include also data from the oceans, you know, air chemistry and air pollution. To build a more general model that goes beyond weather forecasting, he added things.

Dena Temple Raston (5:22)

Like fossil fuel emissions, ocean chemistry, air.

Paris Perdicaris (5:26)

Pollution, for instance, concentration of carbon dioxide in the atmosphere or concentration of nitrogen dioxide.

Dena Temple Raston (5:32)

Feed the model millions of examples, and slowly it begins to connect the dots.

Paris Perdicaris (5:38)

It's a little bit like your brain when you're learning. So the first phase you can think of as a general education phase, where the model is going to see through all these decades of historical data. And now we're talking about millions of hours of, you know, Earth evolution data, everything from temperatures and wind patterns and ocean currents.

Dena Temple Raston (6:05)

It's like giving an AI system a liberal arts education in Earth systems. Then you send it to graduate school to learn about hurricanes, air quality, and.

Paris Perdicaris (6:15)

Ocean waves, for instance, high resolution weather forecasting, or air quality prediction, or ocean wave prediction, or hurricane tracking, or any other interesting prediction task that we might be interested in in order to understand and monitor the Earth system.

Dena Temple Raston (6:33)

But forecasts don't live in a vacuum. They depend on humans and politics. That's when we come back. Stay with us.

Recorded Future Advertisement Narrator (6:55)

In cybersecurity, your greatest fear isn't the threats you see. It's the critical signals lost in the noise. Every day, security teams sort through millions of potential threats. That's why Recorded Future exists, to give you precision intelligence tuned to your needs. Our advanced AI detects patterns humans might miss, while our threat intelligence experts, veterans of military and intelligence services, provide crucial context. With Recorded Future, you gain the confidence to identify critical threats and the precision to act before they become attacks. Learn why 1900 + customers, including 45 + sovereign governments, trust us to detect threats faster and and achieved 350% plus ROI within a year.

Dena Temple Raston (7:54)

Traditional weather prediction takes labor, balloons, equations, supercomputers. What AI can do is shortcut the math.

Paris Perdicaris (8:04)

AI models excel by learning patterns in actual data, rather than trying to solve complicated systems of mathematical or physics equations like the traditional tools are doing.

Dena Temple Raston (8:17)

That makes them faster and cheaper and sometimes more accurate.

Paris Perdicaris (8:21)

So tools like Aurora generate everyday forecasts basically from six hours ahead, maybe up to 10 days ahead.

Dena Temple Raston (8:29)

But here's the garbage in, garbage out.

Paris Perdicaris (8:33)

If I give my model the wrong inputs and then it tells me that it's going to snow in Greece in August, then obviously that is an erroneous prediction that I should filter out.

Dena Temple Raston (8:45)

And that's where humans come in.

Paris Perdicaris (8:47)

There's an additional element here and layer, which is the human expertise of a human meteorologist. So the way official forecasts are issued is through a human that has a specific expertise for the region. That human has synthesized many different forecasts from Aurora and other models and based on their expertise, are issuing the official forecast.

Dena Temple Raston (9:09)

So AI isn't replacing meteorologists, at least not yet. But what if the data itself begins to disappear?

Recorded Future Advertisement Narrator (9:18)

Hundreds of employees at the national oceanic and Atmospheric Administration have been fired, and more layoffs in the agency could be coming today.

Dena Temple Raston (9:26)

In the early days of the second Trump administration, some 1300 members of Noa either resigned or were laid off, and the agency still faces more cuts. And as we reported On Tuesday, Project 2025, the administration's legislative playbook, calls for NOAA to be disbanded and privatized. And that could mean fewer weather balloons, fewer satellites, less data. Since so much of AI is, you know, good information, good Data producing good results. If you have, say, half the number of balloon launches or half the number of satellites actually looking at the weather, is that going to inevitably affect the ability for Aurora to be accurate?

Paris Perdicaris (10:11)

Absolutely, yes. I mean, likely. If those changes impact the quality of the data we have access to and the quality at which we are able to monitor the Earth system, that's going to have a direct impact on the AI models, which are only as good as the data that they've seen.

Dena Temple Raston (10:30)

When Paris heard about the NOAA cuts, he says he immediately worried for Aurora.

Paris Perdicaris (10:35)

Yeah, absolutely. I mean, the funding cuts at NOAA shook up the community and definitely for us at develop AI models, we are relying on noaa. We run relying on their expertise, their data. So of course it's a blow. And you know, if it continues in this direction, it is going to slow down progress.

Dena Temple Raston (10:54)

The first version of Aurora was trained before the cuts. It runs fine, at least for now, but in a few years it'll need new data. And with the climate shifting faster than ever, less or outdated data is the last thing forecasting needs. Though Paris, for his part, remains optimistic.

Paris Perdicaris (11:14)

We're at the beginning and in fact, we have only scratched the surface when it comes to the data that is available out there for building those AI systems.

Dena Temple Raston (11:23)

Optimistic. There's still plenty of untapped data. But the real test of AI won't be the sunny day it gets right. It'll be the storm it never saw coming. From Recorded Future News. This has been Click here's mic drop. It was written and produced by Megan Dietry, Sean Powers, Erica Gaeda, Zach Hirsch, Lucas Riley and me, Dina Temple Rousten. It was edited by Karen Duffin. We'll be back on Tuesday with an all new episode of Click Here. Have a great weekend. SA.

Recorded Future News Announcer (12:41)

Looking for more of the cybersecurity and intelligence coverage you get on Click Here. Then check out our sister publication, the Record. From Recorded Future News. You'll get breaking cyber news from reporters in New York, Washington, London and Kyiv, among others. And you'll see for yourself why it attracts hundreds of thousands of page views every month. Just go to the Record Media.