When conditions are difficult, if it's being stubborn, you might fly for a long time, or you might fly for a short time, or something like that; it's like a different quality, or the day has a stubborn quality. At some point, it dawned on me, hey, you have to put that into a different scale, a different dimension, and then the "not flyable" actually fits in better there. From a certain level, you can just say, okay, I'm not going now if the forecast says it's stubborn, or I'll wait until tomorrow, it's calmer then.
Podz-Glidz, the Lu-Glidz podcast.
Stories from the cosmos of paragliding. Today with Peter Waldner, and I am Lucian Haas. The AI revolution is now making its way into meteorology as well. There are AI models that can predict global weather developments up to ten days in advance at least as well as traditional physical models. In several areas, they even outperform these with only a thousandth of the computing power. Of course, something like that fuels the hope that with the support of AI, we will soon be able to assess flight weather for paragliders more accurately. But, and this is a big but, the field of precise thermal forecasts is not exactly the focus where the weather services are putting their AI efforts.
They are more interested in better extreme weather forecasts. So, conditions where we typically shouldn't be in the air at all. But there is hope. What if AI fans from directly within the paragliding scene start pumping up thermals forecasts with AI? In this process, artificial intelligence learns from previous forecast data and actual flights on those respective days to bring meteorology theory and expected flight reality closer together.
The Swiss Peter Waldner has started exactly this kind of project. The 60-year-old from Zurich programmed an app called Best-Air, which can be found at the website bestair.ae. He talks about what role artificial intelligence plays in it, why Peter doesn't want to leave the assessment of a day's flyability to the AI yet, and much more in this episode 185 of Podz-Glidz.
By the way, I can only present these kinds of stimulating conversations in Podz-Glidz because there are supporters who financially support my work on the podcast and blog. Join in. You can find out how on the Lu-Glidz Gleitschirm-Blog, specifically on the Support page.
Peter, are you actually a fan of artificial intelligence? I mean, in that sense, do you use something like that in your everyday life?
Of course, I mean, I'm a total fan. I find artificial intelligence so incredibly helpful. I'm just surprised by everything you can do with it. I've already used it in all sorts of different life situations,
I'd say, for things like medical or psychological or technical stuff, or I don't know, or with authorities, of course with software, with computers, I've gotten help from AI. I'm just amazed by what it can all do. And I think that's probably just the beginning. It's developed so incredibly in the last few years. I mean, I'm just blown away, I have to say, yeah. What do you use? Sure, the classic chatbots, OpenAI or? I use various things now. I use Gemini a lot, ChatGPT a bit less. For programming lately, a lot of Claude from Anthropic. That's very smart when it comes to programming, yeah.
I assume you probably had your first contacts with ChatGPT or some other AI chatbot too. Did you try back then to say, I'll ask this chatbot, how well can it predict for me, for example, for paragliding, say for today, look at the weather, where can I go flying today? Did you ever try something like that? Really? I never did that. I would have thought you would, if you're so fascinated by AI. But that question, no,
I never asked that. I have, you know, while traveling or something, I was in Bassano recently and I asked what you can do there in the evening, and stuff like that. But for flying directly, I have to—I'll try that right after this. But I suspect there's still a bit of work to do until you get a text forecast like, go to the Beppi launch site near Bassano today and take off at one o'clock. I don't think that works yet, maybe in three or four years.
Yeah, well, I've tried that every now and then over the past year because they always say the AI is even better now. It's gotten better, and then there are these OpenAI versions 4, 5, and whatever. And I noticed, I always did this here for a launch site—I live in Bonn, so for the Bonn region—and at the beginning, it really hallucinated and spat out some nonsense, saying, "Yeah, yeah, of course you can go flying today, the sun is shining and the wind is weak" or something like that. And by now, AI learns, it adapts to you a bit and all that, and it probably knows by now that I'm a paraglider and stuff like that. And now it actually says, "Look, with these wind conditions, you could go to Finkenberg today, that's a south-facing slope near us where you can fly."
And you should be able to fly thermals there. However, there's so much cloud cover that it could be a bit difficult to find thermals. And sometimes there are already terms in there, if you say, okay, someone who doesn't really know about the weather, not that you could believe them, but it sounds like that. That
sounds like
well, I'd say. And in a lot of what I'm seeing now, it actually fits so well with what's already there that you could say, if it's already like this, it probably won't take long until you can say, at least this basic assessment, where could I, I'm visiting Switzerland now or the region somehow, where can I go flying here? And that it could eventually spit out, yeah, yeah, there are the XYZ launch sites and they fit the wind directions. And regarding the thermals forecasts, I also retrieved those somewhere from MeteoSchweiz or something, the glider weather report says,
it works
today, so you can just give it a try. Plus, the wind speeds are still suitable for paragliders. That doesn't mean you have a great weather report yet, but I assume that will eventually work relatively well. I believe that immediately.
I think AI will unfortunately make a few things redundant as well, maybe even my systems at some point, but that's just part of it. I still think it's a great development, and what you're describing—well, I'm going to try it out after our interview. The
I asked about the introduction because you've programmed a Meteor app now that also works with AI. Not exclusively, but where AI plays a role. How did you come to do that? Yes,
so on one hand, it's because of my fascination with AI, because I just think that's where things are heading.
In the
...time that a lot of things are being supported by AI today that used to be done conventionally. And on the other hand, because I always thought,
Actually, the existing forecasts are good, but what's missing for me a bit is the validation. By validation, I mean someone says tomorrow will be good, and then you go flying—and who hasn't experienced that? You're standing at the launch site. Then you take off and think, yeah, today is going to be awesome, someone said so, right? And then you suddenly notice in the air, no, today actually isn't going to be awesome. It's flying, okay, maybe I'll get ten kilometers out, but then I'm on the ground. And that's what I mean by validation, right? So I have my suspicions about the previous forecasts, that they might simply be unable to really check if their predictions actually work and are truly accurate, perhaps due to a lack of manpower and such.
I think, you know, large forecasting systems like MeteoSwiss or the German Weather Service probably have people who do exactly that, but I don't think our small paragliding division has that, and I thought to myself, hey, you could do that and get help with AI. You can't do that by hand. You need, you need help there, and that was actually one of the first thoughts, that I really want to validate this a bit now and develop a learning system that gets better and better.
Your app is called Best Air, so the best air. Yes, the AI basically tells you where it should work best in practice in the end. I think that's probably the idea behind it. Now, this app mainly offers—the core of it is thermals, regional thermals for all of Europe, but your focus is also on Switzerland because you also use this Swiss CA1 model as a basis for it. How does the AI come into play there? I assume you first take the basic stuff that these Meteor models have calculated and create a thermals forecast based on that, however you do that—you can explain that right away as well.
And what does the AI do at that point? How does it learn, and how does it ultimately influence the results?
I can briefly explain that. I calculate around 700 thermal regions twice a day. For that, I take meteorological data from the Deutscher Wetterdienst, specifically the ICON EU, and from those, I take the, let's say, most promising data. These are, for example, data about wind, of course, about temperatures, about solar radiation, about thermal activity or potential. And from those, in a first step using what I call a heuristic model, I actually calculate a classic forecast. So I get a score from 0 to 4, I get some kind of number.
For every region. And every two
hours. A quick side question, because I might have transcribed that wrong. Does that mean you also use this ICON EU model for the thermals in Switzerland? Yes. That means the others are just... you only show, you also show wind maps and stuff. That's where the ICON CH1 from Meteo Schweiz comes into play. Exactly.
I use the high resolution for the wind, precipitation, and cloud data. For the thermals forecasts, I use the slightly lower resolution so that I have the same data basis across all of Europe. Ah, okay.
That was just a brief interjection. Sorry for the interruption. Just keep going, how do you proceed with the ICON EU data? Exactly. Then I calculate that...
It's just an algorithm I've developed where I first create a conventional—I call it a heuristic—forecast, and then I contrast it or supplement it with flight data. And flight data means real flights that I analyze. Specifically, I analyze them in terms of climb, distance, and the maximum altitude reached. It's a pretty involved process. I also break them down by region and time windows. And then I always look at a section—well, my computer does—and basically calculate the real performance that the air allows on that day.
And I think I've come up with a pretty cool algorithm. You have to imagine, for example, if you fly a kilometer and sink 100 meters, that's not really any performance. You're just gliding, it was calm air and it doesn't really bring much. But you could also circle at the same spot for, say, half an hour. And that means you've got some decent—well, not decent, but a bit of thermals. And if you're, say, 500 meters higher at the end at the same spot, then you've obviously proven that there was something there, right? And I take those numbers. Specifically, I deliberately don't take the worst ones, because those might include beginners.
Or if there were beginners involved who don't even turn when they fly into the thermals. Or I don't take the best ones either, because I don't want something like, if Kregelmurer starts, he just always finds something. That's somehow magic or something, maybe there's nothing there at all, but he finds something anyway. No, it's probably not like that. But I take the 90th percentile. That means I take a very good flight and it's then representative for me of what actually happened in this region, in this hour.
Aha, so that means you don't—so I assume you pull these flights from XContest, but you're not saying that on that day, for example, in Fiesch in Valais, 100 flights were made and you don't take those 100 flights and analyze them individually, but rather you say you take the 90th percentile of them, so you take out the 10 best ones and that is the reference flight for you for that day.
So I do look at all of them, but then I basically discard all the others. In the end, you have to narrow it down somehow; for me, the 90th percentile is what matters, and from that, I generate a number again, like earlier with the heuristic model, or specifically from 1 to 4 for the potential, and then I correlate that with a forecast using an AI model. From that, I derive a correction factor that, hopefully—and I'm convinced of this—will keep getting better, because I also have to tell you about a weakness right now: I still don't have nearly enough data. I started collecting data in January, and as you can imagine, there isn't particularly much in January.
You can't find many long flights there.
Flight data, and then it slowly became interesting. I think February was still pretty bad, but from March onwards it was pretty good. From that perspective, unfortunately, I still don't have enough data, and above all, I don't have any summer thermals yet. Summer thermals are, as we all know, quite different from spring or autumn thermals. Even in terms of the values, you have completely different solar radiation, more stability, and so on. So the model still has to learn. That's also quite exciting. These models are really clever because, if you calculate it, they don't just show you the results; they also show you how wrong or how right they are.
And there's a mean error that's spat out by the model. And you can see, yeah, it's not that good yet. I'm really still missing data there. That's why the forecast is still quite heavily, specifically 80 percent, based on this heuristic model, and only a small part, one-fifth, is still influenced by the AI model. I hope that I can eventually throw out the heuristic model completely and only output the AI forecast.
can. From a conceptual standpoint, you now have a lot of regions in Europe.
Yes. Well,
But maybe in February, there was some flying, maybe some thermals were already being flown in the Alps, but not in the lowlands for a long time. There might not have been any thermals at all in February. Yes. Do you basically calculate a small individual AI model for each of your thermal regions, which ultimately says, I optimize the AI model for each region individually? Or is it something where the optimization basically runs fundamentally and, ultimately, what you determine in Fiesch, you apply just as well to the Lake Constance region and ultimately maybe also for the Black Forest and the North German lowlands? So, are the correction factors the same across Europe? Yes, that's the
The nice thing about it is that they aren't factors; rather, the model is a bit cleverer. It's just one model to answer your question, but the model naturally has a lot of information. It also has the latitude and longitude of the region. Above all, it has the region's ID, meaning it knows that the data from Fiesch, for example, has ID 125 and the data from Farnas has ID 126. It knows that these are different regions and can do something with that, in the sense that they are qualitatively different. The model can, for example, take some
...temperature gradient for Fiesch calculated differently than for Farnas or Bonn. That's the exciting thing about these models. They are quite complex. They're also referred to as tree models. It's not one decision; it's hundreds of decisions made based on the data you feed into it. These decisions also get smarter the more data you have available. Because of that, it doesn't actually need 700 models. That would also blow my capacity a bit. It's also the case that for individual regions, you'd probably never get a decent model.
There are no flights there. You've got one every three weeks that happens to pass through on a cross-country flight. Exactly.
You'd never get a decent forecast there. That would be a bit of a problem from that perspective too. But the problem is that the complexity of these models is very great. That
means the model knows for itself—for example, in Northern Germany, I have regions where I have hardly any data, and I don't really correct it there. Yeah, exactly.
That's maybe a bit of the downside. People always say AI is a black box. That's true. What exactly is being calculated in this model—I've googled a lot and had Gemini explain these models to me—but of course, I don't understand them down to the last detail. To some extent, I have to trust, as you probably do when you use ChatGPT, that they'll provide you with something meaningful. But that's not always the case. I experienced a classic hallucination again yesterday. It was Gemini. Then I said, what you told me simply isn't true.
Then Gemini said, "Oh sorry, I hallucinated that."
These machines are honest enough nowadays to admit that. That's also a problem, of course. I might never know exactly what is being calculated, for example, in the lowlands of France. But I hope there's something else, something sensible. I've also thought about how sensible it even is to make forecasts for regions that are hardly flown in, because you might never get the chance to actually validate them. But that remains a bit of a project for the future. Maybe I'll strike them out at some point.
a few regions. What exactly is being corrected there? In your model, in the output, you first say that for every thermals region, you basically provide four categories, options. A day with only glides, a day for local flying, a day for cross-country flying, or it's a hammer day. Those are the four categories you've defined in Best Air. Does the AI now decide which category is output? Or does the AI also decide to say, for example, the cloud base is at 1,500 meters? The weather models said 1,700, but I just round it down because you won't reach them or whatever. Where does the AI's intervention end?
Which output value is actually corrected or co-determined by this in the end?
Right now, it's only two values, namely the potential and not as a category. You just said there are four categories. It's a continuous number. For example, it might spit out 2.7 for a region. I then convert that. It's located in this sector. 2.7 is "streckenfliegen" [cross-country flying]. So it would be green now. But actually, it's a continuous scale. It spits out the potential. And the second thing it spits out is the reachable altitude. That's pretty easy to measure with what the pilots actually flew. And I have the model calculate that as well.
Because I want to know that in advance, not afterwards. And it's simply in meters over more. Those are the two values that the model spits out at the moment. And I then output those on my map. Or rather, I'm currently correcting the heuristic model by 20% with this AI output. And I hope that I'll be able to do it only with that eventually.
Question about the display: in the model—so what you see on the website or in the app—you get a representation over the day with the height winds and, for example, base heights. There are small little clouds drawn there. Is this base height shown there the one the model output? The real cloud base, calculated using classical temps and stuff like that, which you can determine with? Or is it the reachable, i.e., flyable height, which your AI has basically already determined? Yes, exactly, it's still a mix.
It's 80% what the Deutscher Wetterdienst says. They provide a measure for where the thermals stop. There's a value they calculate. And for the other 20%, I'm currently correcting that with my own calculations. So, for now, the focus is still on the so-called heuristic values—meaning it's not a real calculation at all, because that value is actually provided by the Deutscher Wetterdienst, or convective, I don't remember the abbreviation, convection goes up to there and there. The Deutscher Wetterdienst calculates that for me at the moment. And I correct it by one-fifth.
Because even if it's from the German Weather Service, it's not 100% accurate, is it? I'm sure you've experienced that too, right? Like it says the cloud base is at 2,000 meters, but it was actually 500 meters lower, right? So even the German Weather Service is sometimes pretty off. But yeah, that's how forecasts are. I hope I'll get it increasingly better with the actual flight altitude that the pilots then
reach. That means, what happens eventually, when Best Air is—not completely, but well-trained, let's say, over the course of the whole year, also over the whole summer, the data is collected and you do a new evaluation for every day and include the flights and say, okay, now I'll take the best 90th percentile flight from that and calculate it accordingly. That's what it's about again. Then, at the end, you'd really have something where you say, if everything goes well, let's say, then Best Air would give me for my specific region what I can probably actually fly in terms of altitude on such days. Yes, exactly, exactly. That's the vision and that's where it's headed. That means I won't be shown the cloud base height. It could also happen that I can't reach a cloud base at all because the thermals are already so weak beforehand.
Yeah. Instead, if it turns out that way, then Best Air would also show that, no, the flyable thermal height is there and there. And I've corrected this DWD model accordingly so that I provide you with the actual thermals or reachable thermal height based on my experience as an AI. Exactly,
Exactly. Yes, that's exactly how it is.
It hasn't been running for that long yet. You say that March is actually the first month where more or more interesting data starts coming together, because more flying happens then and the thermals also kick in accordingly and are comparable. Are there already things where, from the little bit we have now, you can see, "Ah, the AI is really learning something," where you can say I can see a trend, a tendency, where it could actually get better than the classic thermal forecasts—like, for example, if you take thermal regions, Burn Air or an XC-Therm, which also have such thermal regions accordingly, where you would say the AI could ultimately spit out more realistic values?
Is something like that already noticeable?
Well, it's obviously difficult, you know, if you ask me. For example, are you already better than XC-Therm forecasts or something? That's a tough question. I can just tell you, for instance, I was flying in Zug yesterday, and Burn Air said it was a killer day, but my app said it was cross-country flying, and I was on the ground after half an hour.
Then maybe you're not a reference. Yeah, maybe I'm not the 90th.
Percentile. Derek Kriegel, he flew like a beast, and the 90th percentile would have utilized the cross-country thermalling potential. No, I
Don't let me fool you into thinking I'm at the 90th percentile, but it still didn't turn out to be a disaster... It happened to others too, so maybe I can take myself out of the spotlight a bit. It happened to others, and there were probably some good pilots among them. No, jokes aside, saying "I noticed it like this yesterday at Zugerberg" is obviously not a scientific approach. The question you're asking me is effectively difficult because you'd really have to compare different models with each other. Even here, I think in the gliding world, that's a bit of a pipe dream, but maybe not entirely unrealistic. Someone would have to go and compare XC-Therm, Best-Therm, and Burn-Therm with each other and then perform measurements with a real sample that actually deserves those names.
That is very difficult for me. The advantage I have is actually that every time I run my calculations—I don't do this daily, I do it about once a week—I cross-reference the flights with the DWD forecast data, and the model spits out these error figures every time, and interestingly, they are always getting smaller. This means it's best represented at altitude, because that's something tangible. For example, it shows that for the reachable altitude, we currently have an error rate of 300 meters. That means if it spits out 2000, it could be 1700 or it could be 2300.
This error, of course, decreases with a larger amount of data.
Although, I think a certain amount of error will always have to remain, because there's naturally always so much variability in the weather that even if you have the best model capable of calculating it, you're still going to have it—whether it's exactly that 200, 300-meter fluctuation range—is the cloud now, I mean, how much moisture was in that small valley he just flew over, that it might have been a bit lower or a bit higher at the corresponding spot, it's very difficult to track that at all.
Yeah, and also, you know, I've also thought about the whole concept of regions. Take any region, let's take Pretigau again, okay? You start at Farnars or you start in Madrisa in the back, right? It could be the same region, the same forecast, right? It could be that Madrisa would be a much better choice, right? So there could be different ways to go about it. You could always subdivide these regions into finer ones, right? You could do that. You could even actually do away with regions entirely and make a forecast for every point, for example, that the DWD provides. I stuck with these forecasts...
...sort of roughly assign them to regions, because I think that's also practical, right? Because you don't just fly... well, a long-haul pilot might do 500 to 1,000 kilometers a day, and a forecast that's too granular just isn't realistic. And you often don't even know where you're flying to, right? It has to be manageable somehow. That's why I stuck with it, and a certain level of coarseness and certain errors will also remain with the AI.
That's very clear. For your region, if you're calculating the thermals now, do you take one point from that region where you say, "that's my reference point," or have you defined five reference points for each region that you then average out in the end to get my thermals weather for this region? How do you do that?
Yes, that's an interesting question. Effectively, I do take a reference point. I've hand-picked them for the important flight areas because I looked at where people actually fly. You probably know those Skyways maps where you can see every XC pilot makes a line and over time a line forms. For example, at Matrisa or somewhere like that, there's a red spot because so many people have flown there. And I've naturally tried to place these reference points where people actually fly. And I also include the neighboring points. That means I take a certain amount of blurriness or a certain number of points, but I don't actually take the whole region.
You could also average out the entire region, but there are points that seem less meaningful to me because, for example, nobody ever flies there, or there's a lake, or I don't know what.
Your Best-Air app also has a special feature that I haven't seen with any other aviation weather provider. You don't just list these four categories. For a region, there's gliding, local flying, cross-country flying, and massive day potential, but you also categorize it further and say what kind of demand this day places on the pilot. And there, you've defined three categories: gentle, sporty, and intense. How do you arrive at these categories? And does the AI have any influence on that, or is that something where you say, no, I define that myself? If so, what do you define it by?
And why did you choose exactly this classification?
I actually think this new feature is pretty cool because, well, I'm always looking at what's useful for people, right? And I mean, we've talked about Krigel Moore, who also flies when there's some kind of 50 km/h wind, or even with a Föhn tendency or something. And other people don't do that, and maybe they really don't do it for a good reason, because they just don't have the skills for it. And that's what I mean by this—I've thought a lot about the existing scales, or something, like for example, the lowest color was basically red, or not recommended, and then it went, the next one up was glider, and then somehow regional flying, thermals, cross-country, and then hammer day.
I thought about this scale for a long time and eventually reached the point that it just doesn't work, right? It's not a one-dimensional thing, and especially the "not recommended" part isn't really a good category, because what's not recommended for one person might still be recommended for another, right? It actually has nothing to do with performance, so it doesn't really belong in this potential story; rather, in difficult conditions, when it's gusty, you might be able to fly for a long time, or you might only be able to fly for a short time, or something like that—that's a different quality, right? The day might have a gusty quality. And then at some point it dawned on me, hey, you have to put that into a different scale, a different dimension, and that's how I came up with it.
And then the non-flyable stuff actually fits in better there, right? Like the red category, right? It might not be the same for everyone, but from a certain level, you can just say, okay, I'm a beginner, I'm not going if the forecast says it's gusty, right? I'll wait until tomorrow, it'll be calmer. But
Based on your definition, there can be a "hammer day" that's right at the upper limit of being "choppy," where people say, yeah, a hammer day sounds great, I could go flying, but where the beginner would have to say, no, upper limit choppy, that's probably just too much for me, or I just have to take off at 10 in the morning and land again at 11 and then maybe it works. But
That's exactly the reality, isn't it? I mean, look at it from the outside, right? The local flight school says, "Hey, people are flying in the sunshine," no, in Fiesch, people are flying around in the summer at noon who actually shouldn't be here because it's way too wild. So yeah, and I actually find it really interesting that a pro can still get a lot out of it, while a beginner might prefer to go on a different day, right?
How do you define these three categories—gentle, sporty, and spicy? I mean, what decides when it's sporty, when it's spicy, or when it's just gentle? And what makes a gentle "hammer day," for example? What would have to come together? That might be something the average pilot would wish for, a gentle hammer day.
I suspect that's unfortunately rarely the case because, specifically, or what I factor in, is of course the wind. Or if it exceeds—I don't know the thresholds off the top of my head right now, it's actually continuous—but above a certain wind speed in the relevant altitudes, of course, for example above ground or 500 meters above ground, the demand increases, right? Because lee areas are created and so on, right? If you no longer have forward motion and so on, then it just becomes unpleasant. Then there's wind shear, right? If I see, for example, south wind at the ground and north wind above it, I assume it's going to be turbulent somewhere in between. Then there's the thermals themselves, right?
Because thermals are something beautiful, but they can also get pretty exhausting if they're very hard. Take Fiesch, for example; I factor in a deterioration starting from a certain size, especially regarding this convective potential provided by the Deutscher Wetterdienst. Are you also expecting thunderstorms there, or from a certain level of instability? I factor in a deterioration there. And what else is there?
Föhn, just experimentally, because that's the Föhn—I saw you wrote something about it. Multiple times, yeah. Yes, multiple times, yeah, yeah. I read something from you about the shallow Föhn. That's such a shimmering thing that I don't claim I can explain it definitively with my model. But to some extent, I calculate it, then it flies in. Exactly. Was that it? Yes, I think also rain and thunderstorms of course, but then it becomes very quickly simply unflyable, right? That
is that where you say, so if you, there's also a graphical representation on your app where you can always see, there are these three pillars: gentle, sporty, crisp—you've divided those into three boxes again, and then you always put a red dot in where it's roughly located within that pillar. Exactly.
You can
...pull up and then it's actually unflyable. That's like a red cap on top, the unflyability cap, and it can then be adjusted accordingly based on all sorts of things.
be used. Exactly. And I have to say, for now, I'll probably stick to not calculating the unflyability with AI afterwards, because that's exactly the problem we were just talking about: some pilots might still manage well with certain wind or thermal conditions, while others can't anymore. And that's why the reality, or what the pilots actually do, might not be the measure you want to have, so I'll probably stick with this heuristic model. But the analysis of the flight data, I find that very fascinating. You can get quite accurate information from a pilot's track,
so the computer can read that pretty accurately, whether it was "punchy" or not. And it's surprisingly simple. You can just look at what his climb rates were and, above all, the sink rates.
Then you have a ratio for that. So if you climb quickly and then descend just as fast, or even twice as fast, then it's intense.
So, I think every pilot knows that.
That's an old rule: if it goes up, it also goes down. If you have climb rates over 4 and sink rates over 4 in there, then you can say pretty confidently that it wasn't a smooth flight. And the same goes for ground speed. If you suddenly have 0 meters of forward speed or are even flying backwards or something, then that definitely wasn't a particularly relaxed flight. At least not for me. Maybe there are people who still feel comfortable with that,
but I don't. How did you come up with these three categories—gentle, sporty, and intense? How did you arrive at your classification yourself? Did you just do it by rule of thumb, like I do when I experience it and say, I'll set the limit for gentle flying. There shouldn't be any wind shear over 5 km/h at 500 meters or anything like that. So, what did you take as the transition limits? Well,
That came about a bit through trial and error. In the background, there's also a number that goes from 0 to 4, and anything over 3 is basically in the unflyable range. So, a decimal point—with decimal places, so 2.7 for example—is already pretty intense. But I just thought that for the display, to grasp it quickly, these three-way divisions are actually great. So that you can also say, today it's gentle, or today it's sporty, or today it's intense. But it's actually a purely visual subdivision into these categories. In the background, it's actually a continuous scale.
And what kind of feedback are you getting? Or are you already getting feedback from users saying, "Hey, that fit perfectly every time"? Or do you get feedback like, "Your AI is nice, but it's still hallucinating quite a bit"? Yes, I get
I've actually already received some feedback. Of course, some of it is critical, right? So one person said, I don't know what these colors mean and so on. You have to explain that. I'm doing that now. I promised. I'm doing it next. But I'm also getting very good feedback. People find it exciting. They think it's really something new. A new contribution to these forecasts. And they also like the forecasts. I hope I get even more feedback after this podcast. So I'm very excited when people write to me. I like that. You know, when you're deep into programming like that, you eventually lose sight of the forest for the trees.
It's really interesting for me to know how the people standing out there at the launch site experience these forecasts and the apps, and whether they are useful for them.
Well, you're a paraglider pilot too. I think you've been flying for ten years or so now. Yes. Also relatively intensively, and so you're already quite experienced, I'd say. Now, with your own app, of course, you can talk yourself up a bit, praise your app or whatever. But if you were to take it as it is, would you say, I'd rely solely on the data my app provides, that will be enough for my meteo preparation for a day? Or would you say, no, that's actually just an add-on? You should still look at classic flight weather, meaning look at all the classic weather maps and other forecasts, and then just say, and then I'll take a look at Best Air and see how they and the AI might have assessed it, so that I have a bit of extra help to say, ah, more sporty, more edgy, more gentle, or something like that.
Yeah, so as far as wind and thermals and weather are concerned, I do go out sometimes, even now just with my own app. But actually, I have to say, and I have to give Bernie Herz a huge compliment here. I mean, I'm nowhere near the amount of information he has in his system, am I? When you click into that, for example, you have valley winds, you have landing zones, you have cables, obstacles, you have special notes for wildlife protection, you have hiking trails for hike & flyers and so on. That's just gigantic, isn't it, what he's all worked out there. And I just can't
keep up with, right? Don't you want to? Or could you imagine building Best Air into a competitor of Burn Air at some point? Or is that not actually the intention? You just want to say, in this small niche as a flight decision aid? I think that
I can't manage that, you know? I mean, for now, I'm still a one-man show and I just don't have the resources. I'll probably expand one thing or another as I keep working on it, and I'm very motivated to do that. But I don't know if I'll ever manage to get that level of detail. I can't tell you that right now. I doubt it, actually, right? I have a lot of respect for the work Burn Air has done. And that's true for the other apps too, isn't it? Like XC Therm, for example—I don't know, you probably know it—it's maybe a tad better regarding the thermals forecast, and they also do the wind forecasts very well.
But you don't have launch sites there, for example; you have many other things, but you just don't have those. And it's the same with Paraglideable and the like. All these models cover, like, a specific niche. And that's why the question is, are we competition? I don't think so.
You're a complement, where people can take an extra look and say, "Ah, I'll let the AI give me an assessment of how the day might turn out, roughly."
Yeah, yeah, yeah. So it would be fun for me if I could eventually show that I can offer the best thermals forecasts with these AI predictions. That would be cool, of course.
Best in the sense that it might not be the most accurate specifically for the altitude, but with the addition of "gentle," "sporty," and all that, so that people say, "Ah, that really helped me assess the day accordingly." Exactly, yeah. Now, an app like that—you say you're a one-man show and such—it's also quite... I mean, you probably have to put a lot of time into it, but at the same time, it's also quite data-intensive, I imagine. You always have to read out all these meteorological models, you have to save them somewhere. Do you have your own server, or do you run all of that in the cloud somewhere? How do you do that?
Yeah, well, I don't have a server at home. No, no, no, I rent one, like a virtual server. But that's still manageable, but you're hitting on a point there. I mean, the meteo data is insane; it's an incredible amount of numbers. You think about what the DWD puts out every night or every three hours. I don't know if you've ever looked at it, but there are terabytes of data that can be retrieved. And yeah, of course you can... Does your program do that?
that too? So it keeps downloading, like a terabyte, and says, "Okay, now I'm going to calculate that."
No, I don't try to stay in the terabyte range. These are selected data points. I don't need everything. It's selected data, and I try to keep it lean so I don't end up with huge costs. And no, I'd probably still add a few things, like Sahara dust, to my to-do list. But I don't need many things and probably will never touch them. Still, this whole meteo business is data-intensive. And yes, it involves a lot... I mean, every night and every... when do I do that? By 12 p.m., a few hundred megabytes are already being moved around.
and calculated. So, it's in a form that isn't usable yet. You also have to... convert it and scale it and so on. A lot is already running in the background, yeah. That means you calculate
these models once a day? Twice. Twice? Yes. Once at noon, once at night at... Yes. The midnight run and the noon run? Or how do you do it? Yes, exactly.
So, I get the midnight run at 4 a.m., I think—I don't stick to the exact time—and then the midday run at around 2, I think, or something like that. Well,
You said you rent space in the cloud and have to shovel a lot of data back and forth. That costs money too. And now you're offering best-air completely for free. Even if you set aside your own working time, there are still quite a few costs involved. What motivates you to basically give something like that away to the paraglider scene? Well, yeah, I mean...
I try to keep the fixed costs very low; I believe that's still possible today.
...again, for example, through AI. I'm convinced that in the past, I would have had to hire people, like developers in Pakistan or something, to get to where I am now in a reasonable amount of time. Modern technology also makes things cheaper, and I find that fascinating too. I mean, this virtual server doesn't cost me a fortune. And fortunately, the weather data is also open source, so that means I don't have to pay for it. The garden data is also open source, which wasn't the case a few years ago. So today it's possible, right? A freak like me can do something like this with low costs if he enjoys it.
And I have that. At the moment, I'm not dependent on earnings in this area, which is nice. Maybe I'll charge for some pro features in a year or so. That could be good, but it's also a huge amount of work. You know, you have to set up a payment portal and then calculate taxes with Germany, France, Italy. It's a nightmare. At the moment, it's just that it's good
works. I could have an AI take over your tax returns then. Would you have to program that too, and other things? But that's also fine, I can relate. So, basically, I'm trying with Lu-Glidz to ensure that this funding model I have is also based on the fact that I don't want to deal with too much billing and all that stuff that makes subscription models complicated—you have to be able to cancel them and all that—because it's so much work that I'd actually lose interest in the rest of it, because I'd only be dealing with back-office tasks. See, it's exactly the same for me. That wouldn't be any fun at all.
Yeah, okay, that's interesting. Yeah.
So you don't just have Best Air as your own app now, you've also programmed another app before that's about the live tracking and live ranking of Hike-and-Fly competitions. Exactly. How did that happen?
Yeah, that's a funny story.
Before that, I actually organized my own hike-and-fly race in the Engelberg valley, the Engelberg Köp. And back then, there just wasn't a proper solution for hike-and-fly that was actually reliable during the first exchange. There was one provider—I won't name names now—but they were so unreliable that I just said, no, I don't want that as an organizer. I'm paying a lot of money and then having a lot of trouble because of it. That's not right. So I thought about what it would be like if I made something myself for the Engelberg Köp. And in the first year, it was still pretty primitive, it was only for Android. And somehow it just happened, and then people liked this app and thought, yeah, this is user-friendly and all that.
And lately, there's been an increasing demand for it at more Hike-and-Fly races. And I really enjoy that too, because I get to stay involved with these races a bit and I'm sometimes even there myself. And it's small, a bit of money, but also rather modest. I think it's a great development.
When you say you're involved, do you mean as technical support, or do you actually participate in these races yourself and say, "Wonderful, now I have my own app in my pocket and can see what place I finish in"? Yes, well...
Usually, it's not that I actually participate in the race, like actually running off at the starting gun and all that. But my system is so good now that I might still need to check for about 10 minutes to make sure everything is running smoothly. And then I can jump in at the end and go flying with the group too. So I'm basically participating a bit outside of the competition. Sometimes these are also races. For example, I have the honor of being allowed to judge the Swiss Championship in Hike-and-Fly now. That's also not... well, I told you, I'm not quite in the 90th percentile, it's not quite my level.
But I hope that I'll also be there—it's happening the weekend after next in Jura—that I'll be there to fly
come, yeah. I mean, you probably observe this scene as well because you're an organizer yourself. I mean the scene, the Hike-and-Fly scene, especially in Switzerland, there are now, I believe, almost two dozen competitions across the country throughout the year that are also in relatively good demand. How do you explain this—I don't know if you can call it a hype, but this upswing of Hike-and-Fly, also in terms of wanting to measure oneself? Yeah, I mean, it
is simply one of the most fascinating things. For me, it's the idea, you know, that you really do everything with your own muscles and the thermals, right? And that you get from Salzburg to Monaco with that. It's just something that's simply fascinating. I mean, it fascinates me, right? There's no Bentley involved, no car, no engine, nothing. And yeah, it's maybe a primal dream that fascinates many people. And I mean, Switzerland is ideal for that, right? I always find myself getting all excited about it, right? You can take off from somewhere, climb a mountain, take off, and then glide down into some valley, and then you check the SBB app to see where the next postal bus is.
and it brings you back to Zurich. So it's actually so suitable, right, for this sport. And Switzerland does a lot too. The Swiss Hang Gliding Association also promotes these competitions, supports them. And another important point I find is the equipment, which has always become lighter. There are so many, so many good gliders now that weigh three kilos, right? You didn't have that before. And that just makes it incredibly fun.
Isn't it a bit contradictory? I always imagine hike and fly as, well, me and nature, and I go hiking. And maybe people say, ah, to the exposed but lonely launch site. And that's such a special experience and a special sense of freedom you have there. If you now slap a competition on top of that, your freedom is basically restricted again, because you either have to chase certain waypoints, so to speak, or you're not just, not this, I'm just hiking freely in nature, but you also always have this thought of competition in there and have to be faster than the other person or something like that. Isn't that a bit contradictory for the actual idea of hike and fly?
That
I see it completely your way. Those are two totally different things. I mean, I really enjoy doing both, right? I like going out completely alone or with my dog, and then I'm really out there, for example, somewhere in the Engelberg Valley alone, and then I fly on or down by myself. For me, that feeling of freedom and connection to nature is front and center. And then I actually don't really care if it's just a glide down; then the day was good. The Hike and Fly race is actually a completely different story. Other things come into play there. Personally, I'm not on the podium,
but when I participate, I always find it fascinating to do it together, and it's also fascinating that, I mean, often Kegelmur or other people who just fly top-notch are there, and being out there with them is also incredibly exciting. And in the evening, I think one of the few downsides of paragliding is that you land somewhere and then have to walk home alone, right? And that in the evening you're somehow at the finish line together with your competitors and can have a beer and talk about the day. Like, at this turnpoint, I felt this way and that way.
I always think that's a great thing. For me, this shared experience is more central. People who are very ambitious, they have different goals, right? I can't really say that's my thing, but I can understand it well, right? They measure themselves against each other, like, yes, I can cover 800 meters of altitude in an hour with my heavy backpack. That has its own appeal, doesn't it? But it's not like I'm out there on the front lines. How
how many meters of elevation gain do you usually manage?
Yeah, I'm pretty fit for my age. I can probably do 500, 600 meters of elevation gain if I really get going. How
How old are you? I'm 60. Oh, wow. Well, the others might not see it, but I see it here. I wouldn't have guessed you were 60. Yeah, thanks a lot. So, hats off for that too. With these hike-and-fly competitions, I always wonder—when you're out alone, you come across some difficult ridges, launch sites, or whatever, and you don't have that competitive situation, so you probably, I imagine, assess things much more cautiously: can I launch now or not, do I dare do this or not. And in a competition, of course, the risk simply increases enormously because of this competition situation.
Yes. And it increases enormously in two ways: on one hand, because of the competition, and on the other hand, these competitions often lead into very exposed terrain. Yes. Yes, where you really—I mean, if you take an Eiger tour, where you have to land in pretty rugged terrain next to a mountain hut to then reach this point-to-point and then take off again from there. And these are maybe things where you'd say, in my normal flight, I wouldn't necessarily do that, trying to somehow pump myself down in this rocky terrain. But it is done in such competitions.
Isn't that... well, or put another way, couldn't we find models for how to run these competitions where the risk is lower? Or is this risk part of the thrill, that people enjoy participating in something so intense?
Yeah, I mean, I think you're really touching on a problem there, right? I was raving a bit earlier about the competitions and everything you can experience there. I fully stand by that. But it has this downside, doesn't it? Other sports have them too, like ski racing or motorcycle sports or something. When you push to the limit, it just becomes dangerous, right? And the tendency in the heat of competition to push your own limits for what you think you can still, uh, what you think you can still handle, that's definitely there, isn't it? So, for example, I can tell you, in the Engelberg Cup, we always address that in the briefing, saying, hey guys, it's not about much, we don't have prizes and there's no infinite glory for the winner, or something like that.
It's not about much, don't take risks, right? And keep in mind that the racing fever can carry you away a bit, sweep you up and all that. Luckily, we haven't had any significant accidents in the Engelberg Cup yet. But I know there have been serious accidents, even in other races. That's the sharp side of paragliding; of course, there are serious accidents even without racing. I think as an organizer, you can steer the racing task a bit so that it's perhaps a bit more demanding. You mentioned the Eiger Tour, that's certainly one of the more demanding races, or ones where maybe there's not much margin for error, or where you don't have enough skills, or something like that.
We in the Engelberg Valley are perhaps a bit more gentle, but this effect of overestimating yourself—doing more than you perhaps should—is also a bit of a dilemma sometimes, even in simple terrain. That's a downside that you kind of have to accept. I mean, you're a nurse yourself, you accept that. It's just part of it.
Have you ever fallen into that racing fever trap yourself, to the point where you might have hurt yourself a bit or something?
Yes, unfortunately, I have.
During a three-day race, I made a start in an area that was just a bit in the lee. I actually knew that, and I thought the lee wouldn't be strong or anything. And then it washed me down a bit, and luckily I wasn't seriously injured, but my glider was
destroyed. The race was over and you weren't in the 90th percentile at the end.
No, no, no, no. No, fortunately, not much happened to me, but I've experienced it in my own body and how it's a factor in that.
plays. To go back to the beginning. For Hike & Fly, would you say this Best Air model, or rather the Best Air app, can also be used well for Hike & Fly, maybe to do a risk assessment with this gentle, sporty, intense—so, which tours am I capable of, or something like that? Yes,
So Best Air isn't specifically a Hike & Fly app, of course, but this breakdown based on the pilot's skill level can help. You have to admit that accidents do happen, yes, typically during takeoff and landing, and Hike & Fly is more demanding because you don't have the standard takeoff sites, and sometimes not the standard landing sites either.
An app can't really help that much, can it? Because if you look at this slope and ask yourself, "Can I get off of that?" and there are still some rocks lying around or a few trees and stuff. You have to decide that yourself. Or whether it's still doable and if the wind is right now and everything. So that's just a matter of judgment, and AI doesn't help with that. And I believe that maybe you could eventually send an AI a picture or something and ask, "How should I take off here?" But I don't really think so.
Well, I don't know if I'd want to trust an AI that then says, "Well, you can go ahead and start." Especially because AI tends to agree with you rather than contradict you, since all the models are trained to be perceived as positively as possible. And they'll just say, "Yeah, yeah, just fly, it looks good. I think you can do it." Yeah,
yeah, exactly.
What's the future of Best Air? Where do you see Best Air in a year?
Oh, so in a year, I'll definitely have very good data, over a year's worth. I also want to make the app more user-friendly. That's a big priority for me. I'm currently in the process of programming a little onboarding so that people can actually figure it out. That's also a requirement of mine. I love software and I love good software. And I think there's unfortunately a lot of software in the world that drives people crazy. And I don't want to be part of that. I want to present the complex information that's inherent to Pleitschirmen in the simplest and fastest way possible, so that with just a few clicks, people really know, "Aha, so that's probably how it will be today and tomorrow."
That's my goal. I want to refine it a bit more.
A beautiful closing statement. Keep it as simple as possible for the pilots and keep refining it with AI so that they can really make good use of it. Peter, thank you for these very interesting insights into your AI development regarding weather and thermals forecasts. Thank you, Lucien. I'm going to critically compare the entire project and the results, especially in the region where I'm currently flying, to see what the different thermals forecasts say and if I can eventually recognize whether AI actually makes a difference in the end or if it's just a nice label where you ultimately have to say, well, the error remains so large that regardless of how you look at it, it's a nice representation, so you can use it. But whether it actually brings any advantages or not, I'm curious to see for myself.
How the whole thing develops. In any case, I assume that AI will increasingly take hold somewhere in the paragliding sector in the future. Therefore, you are currently at the forefront of this, and I'm curious to see how it goes from here. I
thanks so much, Lucian Haas, for having me on the show.
Podz-Glidz is the podcast of the paraglider blog Lu-Glidz. Behind it, there is no large editorial team or advertising department, just me, Lucian Haas, as a passionate paraglider and freelance journalist. In the podcast and blog, I combine both to keep the paragliding scene up to date on all the possible facets of this fascinating sport and hobby. And that's completely ad-free and without any paywall. From the feedback of my listeners and readers, I know that many value Lu-Glidz and Podz-Glidz as important and independent information channels. Maybe you feel the same way. For me, both are not just a passion, but also part of my livelihood. Unfortunately, I haven't won the lottery yet or have a fat inheritance in my account.
That's why I have to make sure that the time I put into this is rewarded with more than just intrinsic fun and kind words. And that's where you come in. If you regularly listen to Podz-Glidz and read Lu-Glidz and benefit from it, then please give something back. How much you send me as a supporter via PayPal or bank transfer is entirely up to you. You can find all the necessary information and links on the site www.Lu-Glidz.blogspot.com under the "Support" section. That's it for today. The next episode of Podz-Glidz should appear in two weeks as usual. Until then, go get back up in the air yourself. Take off early, land late, fly far. But above all, have fun.
See you soon, your Lucian.