WEBVTT - Building Boundary-Breaking Balloons 0:00:15.356 --> 0:00:22.996 Pushkin. Every day around the world, more than a thousand 0:00:23.036 --> 0:00:26.356 weather balloons are launched into the sky. The balloons float 0:00:26.436 --> 0:00:30.636 high up into the atmosphere, sending back information you know, temperature, 0:00:30.996 --> 0:00:34.716 wind speed, air pressure, et cetera, and then a couple 0:00:34.796 --> 0:00:39.156 hours later, the balloons pop. This is basically the way 0:00:39.156 --> 0:00:42.316 weather balloons have worked for decades, and the information the 0:00:42.316 --> 0:00:45.876 balloons send back is really useful for weather forecasts. But 0:00:45.996 --> 0:00:49.196 the information the balloon send back is also pretty limited 0:00:49.556 --> 0:00:51.516 because the balloons only stay in the sky for a 0:00:51.516 --> 0:00:54.916 couple hours, they don't fly very far. If we could 0:00:54.956 --> 0:00:57.916 figure out how to make balloons stay up for longer, 0:00:58.356 --> 0:01:01.516 they could blow in the wind and travel thousands of miles. 0:01:01.796 --> 0:01:05.356 They could travel across continents and across oceans and send 0:01:05.396 --> 0:01:07.756 us back a lot more data and give us a 0:01:07.836 --> 0:01:10.796 much clearer picture of what weather is coming our way. 0:01:16.516 --> 0:01:18.756 I'm Jacob Goldstein and this is What's Your Problem, the 0:01:18.796 --> 0:01:20.716 show where I talk to people who are trying to 0:01:20.756 --> 0:01:25.396 make technological progress. My guest today is Kai Marshland. He's 0:01:25.476 --> 0:01:28.836 the co founder and chief product officer of Windborne Systems. 0:01:29.396 --> 0:01:33.116 Kai's problem is this, can you build weather balloons that 0:01:33.196 --> 0:01:36.076 stay in the air for weeks or months instead of hours, 0:01:36.556 --> 0:01:39.116 And can you pair the data from those balloons with 0:01:39.196 --> 0:01:42.196 AI to make weather forecasts that are way better than 0:01:42.236 --> 0:01:45.876 anything we have today. Kai got interested in balloons back 0:01:45.876 --> 0:01:49.156 in twenty fifteen. He was a freshman at Stanford and 0:01:49.196 --> 0:01:52.876 he joined the Student Space Initiative, which sounds kind of 0:01:52.916 --> 0:01:55.796 fancy but in fact was basically a bunch of college 0:01:55.876 --> 0:01:58.316 kids trying to build their own weather balloons. 0:02:02.156 --> 0:02:06.516 The first half dozen flights of our balloons were complete 0:02:06.556 --> 0:02:08.036 and total failures. 0:02:08.516 --> 0:02:09.756 When do things start working? 0:02:10.636 --> 0:02:14.876 That would be back in June of twenty sixteen. This 0:02:15.076 --> 0:02:18.716 is the end of my freshman year, and I think 0:02:18.756 --> 0:02:22.476 our longest flight that year had been fourteen minutes, so 0:02:23.076 --> 0:02:26.876 shorter than a normal weather balloon. 0:02:27.236 --> 0:02:28.756 Buy a lot, Buy a lot. 0:02:29.156 --> 0:02:35.596 And also the person who had started this project, who 0:02:35.916 --> 0:02:39.876 then became my co founder, had just graduated and was 0:02:39.916 --> 0:02:42.756 about to leave what we thought never to be seen again. 0:02:44.436 --> 0:02:48.076 And school has ended. We can't even get into the 0:02:48.116 --> 0:02:51.916 building because our key cards have stopped working, so we 0:02:52.036 --> 0:02:54.836 have to tailgate someone in to get our balloon launch 0:02:54.836 --> 0:02:58.876 equipment out. Go out to the middle of nowhere in 0:02:58.956 --> 0:03:05.236 the central Valley of California, and if this launch didn't succeed, 0:03:05.676 --> 0:03:06.996 windborne would not exist. 0:03:07.796 --> 0:03:11.516 Okay, so I can guess what happened next based on 0:03:11.556 --> 0:03:12.956 that fact. But tell me what happened. 0:03:13.036 --> 0:03:17.516 Yeah. So it got to be about four am, and 0:03:19.516 --> 0:03:22.676 we're out in the middle of this park, really cold, 0:03:22.916 --> 0:03:27.556 and the sprinklers turn on. They start slowly swiveling around 0:03:27.916 --> 0:03:33.076 towards our delicate electronics, and so I have to hurl 0:03:33.156 --> 0:03:37.636 myself onto the sprinklers, getting soaked, just to save our electronics. 0:03:37.796 --> 0:03:39.876 It's like the nerd version of throwing yourself on a 0:03:39.876 --> 0:03:40.916 grenade exactly. 0:03:40.996 --> 0:03:43.436 I really felt like I was in an action movie sequence. 0:03:43.596 --> 0:03:50.116 Ah good, But that sacrifice really paid off because with 0:03:50.276 --> 0:03:53.916 that flight, we managed to set the world record for 0:03:54.436 --> 0:03:55.636 weather balloon endurance. 0:03:56.316 --> 0:03:57.396 How long did it stay up? 0:03:57.476 --> 0:04:00.396 It stayed up for seventy six hours? 0:04:00.436 --> 0:04:02.676 Wow, and your previous best. 0:04:02.476 --> 0:04:04.276 Was fourteen minutes exactly. 0:04:04.436 --> 0:04:06.716 So you're in college, how do you build weather balloons 0:04:06.756 --> 0:04:10.956 that are better than whatever the federal government? Noah's launching 0:04:10.956 --> 0:04:12.756 our you know, state of the art weather balloon. 0:04:12.756 --> 0:04:17.756 People. So the first big thing is that some new 0:04:17.796 --> 0:04:23.236 consumer electronics have come out, in particular, things like micro 0:04:23.316 --> 0:04:30.876 controllers and lightweight, low power satellite communications, and we realize 0:04:30.956 --> 0:04:35.156 that we can use these innovations to make a balloon 0:04:35.476 --> 0:04:37.716 smarter and to control its altitude. 0:04:38.236 --> 0:04:39.396 What's a micro controller. 0:04:39.596 --> 0:04:45.636 Micro controller is a super small computer that you can 0:04:45.676 --> 0:04:48.116 put in any device to give it a brain. 0:04:49.516 --> 0:04:50.156 And they're cheap. 0:04:50.956 --> 0:04:54.156 They're dirt cheap. Anytime you hear the word Internet of things, 0:04:54.236 --> 0:04:55.556 that means micro controlers. 0:04:56.596 --> 0:04:58.996 And nobody else was doing this at the time, Like 0:04:59.036 --> 0:05:02.756 all the big, you know, well funded weather agencies weren't 0:05:02.916 --> 0:05:05.836 sticking my micro controllers on their on their weather balloons. 0:05:06.276 --> 0:05:12.476 Well, with these big agencies, they are really incentivized to 0:05:12.676 --> 0:05:16.556 keep things stable because weather is this really important thing 0:05:16.676 --> 0:05:20.356 that billions of people depend on, and so they're not 0:05:20.516 --> 0:05:24.356 going to go and say, hey, what crazy experiments can 0:05:24.396 --> 0:05:24.796 we run. 0:05:24.956 --> 0:05:27.556 Their incentive is just like, don't screw it up exactly 0:05:27.556 --> 0:05:29.716 the way we did it yesterday. Don't screw it up 0:05:29.716 --> 0:05:32.996 and you won't get fired exactly. And so specifically, what 0:05:33.236 --> 0:05:35.756 was it that you were doing that no one had 0:05:35.796 --> 0:05:36.396 done before. 0:05:37.196 --> 0:05:41.316 Yeah, So it's a very simple concept that is really 0:05:41.356 --> 0:05:45.636 hard to make work and practice. The way we control 0:05:45.636 --> 0:05:49.476 our altitude is when it gets too high up, we 0:05:49.756 --> 0:05:52.916 vent some of our gas so it has less lift. 0:05:53.156 --> 0:05:56.356 When we fall down too far, we drop some ballast 0:05:56.636 --> 0:06:00.356 so it stops falling. But the hard part about this 0:06:00.676 --> 0:06:04.396 is it has to function at negative seventy degrees celsius. 0:06:04.476 --> 0:06:07.716 That's the temperature of dry ice. You have to have 0:06:08.116 --> 0:06:11.596 a lot of on board software. That's where the micro 0:06:11.636 --> 0:06:15.876 controllers come in to do things like understand is this 0:06:16.116 --> 0:06:19.076 just turbulence that you're hitting or is it actually a 0:06:19.116 --> 0:06:21.636 real change in lift that we need to drop ballast 0:06:21.716 --> 0:06:25.156 or vent gas to account for. And in fact, we 0:06:25.236 --> 0:06:28.716 stayed up twenty four to seven. Every single time it 0:06:28.836 --> 0:06:32.196 talked to us. Every five minutes, we had it play 0:06:32.236 --> 0:06:35.716 an airhorn sound because we were, of course falling asleep 0:06:35.756 --> 0:06:38.036 in the middle of them to see what it was doing. 0:06:39.276 --> 0:06:41.756 So, okay, so you set the record, you stay up 0:06:41.836 --> 0:06:45.196 for what three days watching it? How far did it go? 0:06:46.276 --> 0:06:50.156 It flew all the way across the country and landed 0:06:50.796 --> 0:06:55.916 out in the Atlantic Ocean. But the real reason why 0:06:55.956 --> 0:06:59.076 we decided to form a company around this was because 0:06:59.676 --> 0:07:03.636 we realized the impact this could have. We got into 0:07:03.636 --> 0:07:06.956 this from the engineering side. The lands are fun, but 0:07:07.956 --> 0:07:10.836 we realized, wait a secd eighty five percent of the 0:07:10.876 --> 0:07:15.556 Earth's atmosphere is invisible to humanity. Weather is this crazy, 0:07:15.676 --> 0:07:20.636 unpredictable thing, and we can solve that. And weather is 0:07:20.716 --> 0:07:23.156 so much more than just do you bring an umbrella 0:07:23.276 --> 0:07:29.516 to work tomorrow is the single most immediately destructive aspect 0:07:29.876 --> 0:07:34.316 of climate change, and improving the weather forecast it of 0:07:34.316 --> 0:07:37.996 course helps with adapting to climate change, things like predicting 0:07:38.276 --> 0:07:41.756 where a hurricane will make landfall, but it can also 0:07:41.876 --> 0:07:45.196 help with preventing climate change in the first place. If 0:07:45.236 --> 0:07:49.756 you have a better weather forecast, you can better route 0:07:49.876 --> 0:07:52.876 ships and planes to save a huge amount of fuel. 0:07:53.636 --> 0:07:56.636 You can accelerate the transition to renewables because you know 0:07:56.676 --> 0:07:59.996 when the wind will be blowing or the sun will 0:08:00.036 --> 0:08:03.316 be shining. And we looked at this and said, no 0:08:03.356 --> 0:08:06.716 one else has our balloon technology. If we don't do this, 0:08:06.916 --> 0:08:10.596 no one will. And if we want away from this opportunity, 0:08:10.636 --> 0:08:11.956 we can't live with ourselves. 0:08:13.476 --> 0:08:15.596 I mean, I buy that the stakes are high. I 0:08:15.676 --> 0:08:19.396 buy that it could be very helpful, but like, why 0:08:19.436 --> 0:08:20.876 wouldn't anybody else do it? 0:08:22.036 --> 0:08:26.876 Yeah, a lot of people are trying to improve the 0:08:26.916 --> 0:08:32.756 weather forecast. But what's unique about Windborne is the combination 0:08:33.156 --> 0:08:38.116 of our balloon technology with our AI weather forecasts, because 0:08:38.236 --> 0:08:41.156 really those are the two levers to pull to increase 0:08:41.196 --> 0:08:42.876 the accuracy of a weather forecast. 0:08:44.156 --> 0:08:48.516 You mentioned that eighty five percent of the world lacks 0:08:48.516 --> 0:08:52.276 good weather data, which was surprising to me the first 0:08:52.356 --> 0:08:54.796 time I heard it. But I would think, like, get 0:08:54.956 --> 0:08:57.876 satellites do it. That's kind of I guess my naive 0:08:58.076 --> 0:08:59.956 ish thought. It's like, I know there's a lot of 0:08:59.996 --> 0:09:01.916 satellites looking down at the Earth all the time. 0:09:02.636 --> 0:09:04.876 I'm actually a big fan of the space industry. I 0:09:04.956 --> 0:09:07.596 used to work at SpaceX. But the problem is the 0:09:07.676 --> 0:09:12.036 laws of physics fundamental limit what satellites can measure. Take 0:09:12.076 --> 0:09:16.796 for example, pressure. It turns out pressure is extremely important 0:09:17.476 --> 0:09:20.836 for predicting the weather because it determines where the winds 0:09:20.836 --> 0:09:24.076 are blowing, how the weather systems are moving. But a 0:09:24.156 --> 0:09:29.316 satellite fundamentally cannot measure pressure from space because it's not 0:09:29.476 --> 0:09:31.756 in the atmosphere. It can't see what's going on. 0:09:32.316 --> 0:09:36.596 So you start the company you're good at making weather 0:09:36.636 --> 0:09:38.716 balloons that go up and stay up for a long time, 0:09:38.756 --> 0:09:41.676 and that you can steer. I also read that you 0:09:41.836 --> 0:09:45.596 launched thinking that you could collect data for a tenth 0:09:45.796 --> 0:09:50.676 the cost of existing alternatives, which sounds compelling, but as 0:09:50.716 --> 0:09:53.916 I understand was not compelling enough. Is that right? 0:09:54.596 --> 0:09:57.556 It is. We're now at one hundred and fifty times 0:09:57.556 --> 0:09:59.636 more data per dollar than alternatives. 0:09:59.996 --> 0:10:02.276 Okay, was there some reason one tenth the cost was 0:10:02.356 --> 0:10:03.276 not cheap enough. 0:10:04.196 --> 0:10:07.596 Yeah. When we started the company, we thought it would 0:10:07.636 --> 0:10:11.356 be cheap enough. But it turns out when you really 0:10:11.436 --> 0:10:14.876 look at the scale of weather data collection, you need 0:10:14.956 --> 0:10:20.996 to really understand the atmosphere. You want ten thousand balloons 0:10:21.156 --> 0:10:25.636 aloft concurrently, that's a balloon every sixty miles in the atmosphere. 0:10:27.356 --> 0:10:32.276 And in order to get to that level of scale, 0:10:32.556 --> 0:10:36.156 it's the difference between your company spending one hundred million 0:10:36.236 --> 0:10:39.356 dollars a year and spending a billion dollars a year. 0:10:39.796 --> 0:10:41.596 One was a lot more in reach. 0:10:41.916 --> 0:10:43.796 How do you drive down the cost to one one 0:10:43.916 --> 0:10:45.636 hundred and fiftieth the cost? I mean, I'm sure it's 0:10:45.876 --> 0:10:50.516 many many, many incremental you know, efficiency gains. But like, 0:10:50.556 --> 0:10:51.796 what's an example of one. 0:10:51.996 --> 0:10:56.436 Yeah, Well, one of the big pieces is improving the 0:10:56.636 --> 0:11:00.316 software that flies on it so that the balloon can 0:11:00.356 --> 0:11:03.636 fly for longer. I talked about in the student group 0:11:03.716 --> 0:11:07.316 that first flight lasting seventy six hours. Now our longest 0:11:07.316 --> 0:11:09.756 flight can fly for over four days. 0:11:10.476 --> 0:11:10.956 Wow. 0:11:11.076 --> 0:11:15.036 And the longer you fly, well, the hardware costs stays 0:11:15.076 --> 0:11:18.436 the same, so that means you're collecting a lot more data. 0:11:18.916 --> 0:11:21.436 Yeah. Yeah. So basically, make the balloon stamp for longer 0:11:21.596 --> 0:11:24.916 is the fundamental way that you make it cheaper exactly, 0:11:25.476 --> 0:11:27.436 And how do you make it fly longer? 0:11:27.956 --> 0:11:31.716 Yeah, Well, one of the big things is improving the 0:11:31.756 --> 0:11:35.116 software to better decide when do you vent gas, when 0:11:35.156 --> 0:11:38.396 do you drop ballast or really one of the other 0:11:38.436 --> 0:11:43.116 big things is just making everything smaller and lighter, because 0:11:43.596 --> 0:11:46.596 the smaller it is, the lighter it is, the less 0:11:46.716 --> 0:11:50.236 chance of a leak, the less ballast you have to use. 0:11:50.676 --> 0:11:55.916 And things have just shrunk down so far. It really 0:11:56.316 --> 0:12:00.556 surprises me sometimes when I'm like, wait, that tiny thing 0:12:00.636 --> 0:12:03.436 the size of a dime replaces the thing that was 0:12:03.436 --> 0:12:05.716 the size of a dinner plate before. 0:12:06.676 --> 0:12:10.076 Yeah, that's amazing. So okay, so let's talk about where 0:12:10.116 --> 0:12:12.436 you are today. Let's talk about sort of how it works. 0:12:13.716 --> 0:12:14.756 How big are the balloons? 0:12:15.476 --> 0:12:20.116 So the balloon is two pieces. There's the envelope that's 0:12:20.236 --> 0:12:24.076 the bag that holds all parts the balloon part of 0:12:24.116 --> 0:12:30.356 the balloon, and then there's the main unit, which is 0:12:30.396 --> 0:12:32.996 the electronics and the ballast. 0:12:33.076 --> 0:12:36.836 The stuff exactly, the balloon and the stuff the balloon part. 0:12:36.716 --> 0:12:39.116 Of the balloon. The balloon part of the balloon is 0:12:39.436 --> 0:12:43.356 five and a half meters tall. So that's two and 0:12:43.436 --> 0:12:45.796 a half three times the height of person, depending on 0:12:45.836 --> 0:12:46.636 how tall you are. 0:12:47.156 --> 0:12:51.996 Yeah, so tall. Yeah, it's a big balloon. And how 0:12:52.236 --> 0:12:54.876 how big is the stuff? How big is are the 0:12:54.916 --> 0:12:56.156 sensors in the part that's not the. 0:12:56.316 --> 0:13:01.836 Yeah, that stuff is. I guess this is an audio thing, 0:13:01.916 --> 0:13:04.516 so I would say about this big. 0:13:06.116 --> 0:13:07.476 The size of a basketball. 0:13:07.556 --> 0:13:12.156 Yeah, it's a roughly the size of a basketball. It 0:13:12.236 --> 0:13:16.156 weighs just over four pounds, so about the weight of 0:13:16.196 --> 0:13:22.396 a large duck. Okay, and is kind of long and skinny. 0:13:22.436 --> 0:13:25.516 It kind of looks like a fish. I think of 0:13:25.556 --> 0:13:28.716 it as a trout attached to a giant bag. 0:13:30.156 --> 0:13:33.636 How many of your balloons are in the sky right now? About? 0:13:34.396 --> 0:13:37.076 Yeah, there are a few dozen aloft right now. 0:13:37.556 --> 0:13:37.836 Okay. 0:13:37.876 --> 0:13:40.796 We launch around one hundred a month and are quickly 0:13:40.876 --> 0:13:41.636 ramping that up. 0:13:41.676 --> 0:13:43.076 So where do you launch your balloons from? 0:13:43.796 --> 0:13:47.476 We launch them every day from three continents, South Korea, 0:13:47.676 --> 0:13:50.076 paloelto New York and Cabo Verde. 0:13:50.396 --> 0:13:53.156 Cabo Verde is just off off the west coast of Africa. 0:13:53.236 --> 0:13:53.676 Is that right? 0:13:53.796 --> 0:13:55.916 You got it. One of the reasons why we have 0:13:56.036 --> 0:13:59.356 launch site in Cabo Verde right now. That's right where 0:13:59.516 --> 0:14:02.316 a lot of hurricanes are forming, and so the fact 0:14:02.356 --> 0:14:06.116 that we're collecting data around there is going to be 0:14:06.316 --> 0:14:10.716 really impactful for better predicting the path of these hurricanes. Huh. 0:14:10.756 --> 0:14:13.876 So they start there and then they essentially travel across 0:14:13.916 --> 0:14:16.396 the Atlantic and into the Americas, and if you can 0:14:16.516 --> 0:14:19.876 understand what's going on there, ideally you can sort of 0:14:19.996 --> 0:14:22.716 understand the hurricane where it's going to go in the 0:14:22.796 --> 0:14:26.276 dream scenario, even just as it's becoming a tropical storm. 0:14:26.356 --> 0:14:30.476 Exactly. Yeah, Noah did an analysis of the Impactor data 0:14:30.516 --> 0:14:35.756 had on the twenty twenty two hurricane season, and it 0:14:36.436 --> 0:14:40.836 made the forecasts for Hurricane Fiona about twenty percent better. 0:14:43.876 --> 0:14:46.316 Now it's time for a few ads. After the ads, 0:14:46.556 --> 0:14:50.716 Kai and I will talk about AI. Of course, we'll 0:14:50.716 --> 0:14:54.156 also talk about Winborne's business, what they're selling, and we'll 0:14:54.156 --> 0:14:56.876 talk about the company's quest to build balloons that can 0:14:56.916 --> 0:14:59.476 stay in the air for months at a time and 0:14:59.476 --> 0:15:16.916 make multiple trips around the world. What is your business 0:15:16.956 --> 0:15:18.636 right now, like, what are you selling and who you're 0:15:18.636 --> 0:15:19.036 selling it. 0:15:18.996 --> 0:15:22.236 To right now. Our business has two pieces to it. 0:15:22.596 --> 0:15:26.276 We are selling the observations of the atmosphere we collect 0:15:26.516 --> 0:15:30.356 to the government right now, so that those governments can 0:15:31.436 --> 0:15:36.236 use this data to improve your weather forecast. And just 0:15:36.276 --> 0:15:40.116 a couple of weeks ago, we announced our AI based 0:15:40.156 --> 0:15:43.836 weather model, which is the world's most accurate global weather 0:15:43.916 --> 0:15:44.996 model bar none. 0:15:45.036 --> 0:15:46.836 I feel like you need a little TM when you 0:15:46.876 --> 0:15:50.116 say the world's most accurate global weather model bar none. 0:15:50.276 --> 0:15:51.916 We really do, we really do. 0:15:52.076 --> 0:15:54.596 So let's talk about that. Is that claim validated? I mean, 0:15:54.836 --> 0:15:57.196 I know there was news about that just this year, 0:15:57.236 --> 0:16:01.476 and congratulations, but also, how do I know that's true? Respectfully? 0:16:01.636 --> 0:16:06.116 Yeah, I love that question because anytime a weather company 0:16:06.156 --> 0:16:10.636 makes a claim, you should look at it closely. So, 0:16:10.996 --> 0:16:15.116 first off, we've published our results on our website and 0:16:15.196 --> 0:16:18.876 so you can dig into some of the raw numbers there. Second, 0:16:18.996 --> 0:16:23.076 we're in the process of submitting to weather bench which 0:16:23.156 --> 0:16:28.836 is these benchmarks run by Google, which were the previous 0:16:29.116 --> 0:16:33.916 holders of this record. So we're actively talking with them 0:16:33.996 --> 0:16:38.956 right now about submitting our results getting them validated into 0:16:38.996 --> 0:16:41.956 live up there. It's just a process that takes a 0:16:41.996 --> 0:16:42.996 little while, so. 0:16:43.036 --> 0:16:45.676 Let's talk about the model you built. So I mean 0:16:46.676 --> 0:16:49.236 generally when I talk to people who are working on AI, 0:16:49.356 --> 0:16:51.516 they talk a lot about data, right, And it seems 0:16:51.596 --> 0:16:57.156 like to some significant degree the models are kind of commoditized, 0:16:57.196 --> 0:17:00.316 not quite, but like models are pretty similar, it seems 0:17:00.356 --> 0:17:02.676 in many settings at least one to the other, and 0:17:02.756 --> 0:17:06.476 the differentiator often ends up being the data. Is that 0:17:06.516 --> 0:17:08.076 the case in this instance? I mean, do you think 0:17:08.116 --> 0:17:10.596 you're the best. If you're the best because you have 0:17:10.836 --> 0:17:13.596 all this data from all your balloons. 0:17:14.316 --> 0:17:16.556 I think that you're spot on. Data is the real 0:17:16.596 --> 0:17:20.436 mote here, because what really sets us apart is the 0:17:20.476 --> 0:17:24.356 ability to have all of this data that no one 0:17:24.356 --> 0:17:28.716 else has and to use that both for training the 0:17:28.756 --> 0:17:32.236 model but then also for running the model in real time. 0:17:32.476 --> 0:17:35.956 Data isn't just about training. It's about essentially giving your 0:17:35.956 --> 0:17:40.436 model the prompt of what's going to happen next, and 0:17:40.676 --> 0:17:43.076 you need new data every single day for that. 0:17:44.396 --> 0:17:48.676 Just one thing to clarify. So, if you're selling the 0:17:48.796 --> 0:17:53.556 data to governments, are they not sharing it widely? Like 0:17:53.596 --> 0:17:56.076 they're paying for the data but they're not. But not 0:17:56.156 --> 0:17:58.156 everybody can get at the data? Is that why it's 0:17:58.156 --> 0:17:58.756 a mode for you? 0:17:59.396 --> 0:18:03.236 Yeah, so what they're using the data we sell them 0:18:03.436 --> 0:18:09.236 for is putting it into conventional physics based weather model 0:18:09.596 --> 0:18:14.796 that they then release. So they can't redistribute the data itself, 0:18:15.756 --> 0:18:20.116 but the general public can benefit from its use in 0:18:20.156 --> 0:18:21.116 these weather models. 0:18:21.436 --> 0:18:24.356 And so just to distinguish and this is a distinction 0:18:24.516 --> 0:18:27.636 that's not particular to your company, but you mentioned the 0:18:27.676 --> 0:18:30.676 sort of traditional physics based weather models. I mean, I 0:18:30.676 --> 0:18:33.516 think it's worth spending just a moment here to distinguish, right, 0:18:33.596 --> 0:18:39.476 like between the classic weather model, you know, the sort 0:18:39.516 --> 0:18:42.436 of pre AI weather model, and the AI weather model. Like, 0:18:43.316 --> 0:18:44.996 what's the basic difference between those two? 0:18:45.436 --> 0:18:49.516 Yeah, So a conventional physics based weather model, it takes 0:18:49.516 --> 0:18:52.956 the initial state of the atmosphere and then runs a 0:18:53.156 --> 0:18:59.116 bunch of fluid dynamics on it to simulate what's going 0:18:59.156 --> 0:19:01.196 to happen to all of these different fluids. 0:19:01.676 --> 0:19:04.036 There's a sort of kind of classic sort of feels 0:19:04.036 --> 0:19:08.036 like kind of nineteenth century deterministic. Give me the initial 0:19:08.796 --> 0:19:10.876 condition and I'll tell you what's going to happen in 0:19:10.876 --> 0:19:13.516 the future. It's that right, it's sort of rules based. 0:19:13.836 --> 0:19:21.076 Exactly exactly, and that is nice because you can see 0:19:21.116 --> 0:19:25.236 the exact physics that is being used, but you need 0:19:26.316 --> 0:19:29.996 compute clusters. They cost hundreds of millions of dollars in 0:19:30.076 --> 0:19:33.076 order to run this because the atmosphere is so big. 0:19:34.396 --> 0:19:39.076 And by contrast, a AI based weather model can run 0:19:39.236 --> 0:19:41.396 on just a gaming laptop. 0:19:42.316 --> 0:19:42.516 Huh. 0:19:43.356 --> 0:19:48.956 And what it does, by contrast, is effectively picking out statistics. 0:19:49.116 --> 0:19:51.196 I mean, it's just it's machine learning. I presume it's 0:19:51.236 --> 0:19:52.596 pattern matching exactly. 0:19:53.316 --> 0:19:57.276 It's our model is a transformer based architecture, the same 0:19:57.316 --> 0:19:58.676 thing that powers chat GPT. 0:19:59.556 --> 0:20:02.516 Our A models in general better at this point than 0:20:02.556 --> 0:20:04.316 the physics based traditional models. 0:20:04.916 --> 0:20:07.276 It depends on the use case. But if you're talking 0:20:07.276 --> 0:20:12.876 about hurricanes, yeah, we ran a case study on Hurricane 0:20:12.916 --> 0:20:19.476 Ian and our model would have predicted landfall by two 0:20:19.596 --> 0:20:21.556 hundred kilometers closer. 0:20:21.876 --> 0:20:24.956 I mean, that's another one where like I'm intrigued by 0:20:24.956 --> 0:20:27.676 what you said, but I'm very eager for I'm not 0:20:27.796 --> 0:20:31.356 eager for this hurricane season to come, but like it 0:20:31.396 --> 0:20:34.676 will resonate more with me when you do that prospectively, right, 0:20:34.756 --> 0:20:39.156 of course, I mean, will you this year be predicting 0:20:39.196 --> 0:20:40.676 where hurricanes are going to make landfall? 0:20:41.276 --> 0:20:45.236 We will be predicting that for public safety reasons. We 0:20:45.436 --> 0:20:47.956 aren't going to be you won't tell me. 0:20:48.116 --> 0:20:49.756 Yeah you will, but you won't tell me we. 0:20:49.796 --> 0:20:55.236 Will know a hurricane. Okay, but fair, you of course 0:20:55.276 --> 0:20:58.956 don't want to say, hey, don't listen to Noah about 0:20:58.956 --> 0:21:00.396 these evacuation orders. 0:21:01.076 --> 0:21:05.236 No, no, fair that that is very responsible. So basically, 0:21:05.396 --> 0:21:09.916 you're selling data. You are likely soon to be selling forecasts. 0:21:10.116 --> 0:21:14.956 That's sort of the business. On the technical side, what 0:21:15.636 --> 0:21:17.596 is the frontier? What are you trying to figure out 0:21:17.596 --> 0:21:18.876 that you haven't figured out yet? 0:21:19.636 --> 0:21:23.316 Yeah, well, one of the big areas for innovation is 0:21:23.996 --> 0:21:28.596 on the AI modeling side. We were kind of surprised 0:21:28.716 --> 0:21:31.716 that our models did as well as they did because 0:21:32.356 --> 0:21:35.836 there are a lot of quite frankly obvious things that 0:21:36.116 --> 0:21:40.596 we haven't done, things like just increasing the amount of 0:21:40.676 --> 0:21:44.676 compute we're using to train these models. We use something 0:21:44.756 --> 0:21:49.396 like a fifteenth as much as Google did. So what 0:21:49.516 --> 0:21:51.676 happens when you train it for longer? 0:21:51.956 --> 0:21:54.076 I mean, I'll say, on the AI side, it's not 0:21:54.396 --> 0:21:57.556 obvious that it's optimal for you to be doing the 0:21:57.596 --> 0:22:00.316 AI as well. Right, there's a universe where you are 0:22:00.756 --> 0:22:04.236 optimizing the balloons and the data, and then someone else 0:22:05.276 --> 0:22:08.156 like Google, who has all the computers and all the 0:22:08.196 --> 0:22:12.276 AI knows what to do with doing the AI side. 0:22:12.036 --> 0:22:17.116 Of it, right, Yes, and no, I think that where 0:22:17.156 --> 0:22:21.396 that falls apart is coupling our data with the models 0:22:21.516 --> 0:22:24.676 much more closely. In that there's so much to be 0:22:24.836 --> 0:22:28.156 done in terms of figuring out how to better take 0:22:28.196 --> 0:22:33.516 advantage of our data into particular and also things like saying, 0:22:34.556 --> 0:22:37.316 based on this weather model, where should we be flying 0:22:37.316 --> 0:22:41.236 our balloons to improve the forecast? And we're already using 0:22:41.436 --> 0:22:45.356 our own AI weather forecasts to do that flight plan 0:22:45.476 --> 0:22:48.876 optimization to figure out where our balloons are going to fly. 0:22:49.276 --> 0:22:52.796 So it's really this beneficial effect where the better our 0:22:52.876 --> 0:22:56.116 weather forecast, the better we can fly our balloons. And 0:22:56.836 --> 0:23:00.076 we can then target our balloons to fly to the 0:23:00.076 --> 0:23:02.716 places that will most improve the weather forecast. 0:23:03.396 --> 0:23:04.516 That's a good feedback loop. 0:23:04.676 --> 0:23:06.716 There's also a lot to do on the balloon side 0:23:06.716 --> 0:23:10.796 of things. I wish I could tell you about our 0:23:11.476 --> 0:23:16.436 project that will increase flight time by another factor. 0:23:16.076 --> 0:23:18.876 Of ten, another factor of ten. 0:23:19.076 --> 0:23:21.676 Yeah, so what what are you at now? 0:23:21.716 --> 0:23:23.716 What's the like median flight time? 0:23:23.996 --> 0:23:29.076 Medium flight time is seven to ten days depending on 0:23:29.356 --> 0:23:30.236 how we're targeting it. 0:23:30.396 --> 0:23:33.996 And so you think you're going to get to ninety days. Yep, 0:23:34.276 --> 0:23:38.956 we around the world in eighty days. Yeah, you say 0:23:38.956 --> 0:23:40.356 you wish you could tell me like you feel like 0:23:40.476 --> 0:23:43.076 you're about to do it. Yeah, tell me, tell me 0:23:43.116 --> 0:23:44.716 without telling me what you can't tell me. 0:23:44.996 --> 0:23:49.756 Okay, I'll do my best. So we have had some 0:23:50.356 --> 0:23:56.836 very successful flight tests where we have demonstrated the ability 0:23:57.076 --> 0:24:03.116 to fly without using any finite resources. 0:24:03.356 --> 0:24:06.636 So what what is the key finite resource? The the 0:24:06.676 --> 0:24:08.876 lift gas, the what is it helium or hydrogen? 0:24:08.916 --> 0:24:13.236 What you used to keep Our balloons are compatible with 0:24:13.276 --> 0:24:16.556 both helium and hydrogen, So which we use is dependent 0:24:16.636 --> 0:24:17.276 on the country. 0:24:17.676 --> 0:24:20.316