PowerBot Talks S01E08 - Casper Kaae Sønderby - Co-Founder at Twig Energy

Hello and welcome to Powerful Talks, the show where we talk to the brightest minds in energy markets. My name is Helmut Spindler, I'm your host today and my guest today is Casper Sonderby, CEO and co-founder of Tweak Energy, Halakasper. Hello, it's a pleasure to be here, Casper. I think it would be best if you introduce us to Tweak Energy and you roll a Tweak Energy before we go back in time. Can you do a bit of a short introduction of what Tweak does? Yes, of course. So I'm just engineered by training now, CEO, so I do more things that are like non-technical now, but Tweak Energy, I mean, we are a power trader and an aggregator, so we trade power in Europe and in the US and we optimize assets on both the intraday and the ad markets and on the initial services. So that's us in like 30 seconds, probably the hard facts, the country area you come from. When did you start? So we founded in 2021 in Copenhagen by me and my brother, so we are the Choukov honors. Now we are 18 people, on the trading side we are covering most of Europe, pretty much was epics spot colors and are pretty big on the intraday part trading there and I think what sets us a little bit of parties that we run, we fully automated the trading setup. So we have no people with background in energy trading and so it's much more a computer science company or machine learning company in that regard. And then on the asset side, we are active in the Nordic right now, have solar batteries and load banks online in Denmark and Sweden, basically expanding that tool to the Nordics and like other countries in primarily, I mean, just down Europe right, I mean Germany is obviously interesting because it's big, there's also many other Eastern European countries that are very interesting and an element tells you, so we are looking at that as well. Super fascinating, quite some growth due to all bites, it's young agents, it's also interesting that you're going into assets, where did you start? How did you get it to the energy industry? How did it happen? Yeah, how did it happen? I mean, that's a good question, I mean, I mean, I can do the long story and then maybe it makes sense why I ended up in energy because I think we are podcasts because we can do the long story. We can do the long story, yes, but yes, so as I just briefly mentioned, I'm an engineer by training actually by medical engineering, so I did that in Denmark, then kind of drifted more and more towards machine learning statistics and then eventually I did a PhD in machine learning where I was on the called probabilistic deep learning, so it's the kind of early deep learning wave, at least in Europe, we were some of the early ones there, right here, so that's back in 2000 and something like that. So I mean, you know, you publish a bunch of papers and do a lot of math and they're super fun, I think it's really a nice time to be in that field because just the amount of ideas and growth was just staggering, right, so we were initially visiting these conferences with a few under people, right, and then when I left or like finished my PhD, the main conferences were 5,000, 10,000 people and you have all these major companies fishing for talent there. So just, I mean, it's just pure luck that you're joining on that kind of wave development. Where did this wave start, the wave didn't really start a lot earlier, right? So it's probably more like in 12, I would say, so there's an image net moment it's called when there's a guy in Jeffrey Hinton, he had a PhD student and then you just would keep an eye on Alex Krishinsky, I don't pronounce his name correctly, but they basically have this, like when this competition in image recognition, where they just blow all these traditional methods out of the, just demolished them completely, much much better than everyone was at least looking at that. What I actually worked on might be taking, this was like the taking, right, to catch the talks, yeah, exactly, there's like this data set, right, that has about a million images called image net and there's a hundred categories, a thousand maybe actually, then you have to classify them. So by today's standard is super trivial and also my PhD, right, I mean, I worked on some like generative AI back then, so mostly on images and the images regenerated back then was like 128 by 128 pixels right off of like photos of people or something like that. So that was considered a high resolution and I'm amazed by the development in the last five, 10 years and this was also in 2015, this was also 2015, like 15, 16, yeah, something like that, right? Um, I didn't knew that Jenny, I, um, image generation was already a thing there, that's super interesting. Oh, but I have been going on for a long time, I mean, for, for like decades, it has just been, you know, actually, I had this kind of, you know, in my PhD, I did a study where we studied like grass textures, we tried to generate grass, like kind of close up photos of grass because they had kind of interesting structure and like high frequency structure. But I mean, that's, that's just where the field was back then, right? And, and now we're talking, you know, these super high resolution, almost like interesting sort of from, from real images, that's just amazing, I think, but did you stumble upon that field, how did you get it to deep learning? Yeah, I don't know from how did I, so basically, I mean, so me and my brother, we have this kind of little bit on off professional career trajectory. So I think he, he actually started in another place in Denmark and then moved to Copenhagen and we're just like, you know, ping ponging a little bit on, on, on, on, on machine learning. And I actually worked in a, after my master's, I worked at the vaccine development, actually a amount of vaccines for, from what to resist and bacteria infections, but then I wanted to get a bit more deeper into the method. And then I mean, we got just stumbled on some of these tables on, on, on, on deep learning. And really, there was no people doing it in, in, in Denmark back then at least. So we showed up at a, like, there's all a little bit of that turned out to be my PhDs who were like, just, you know, talked to him about, is this was interesting. I, I think I can say that he was a maybe a little bit burnt by neural networks because he's turned on that, you know, anticipating in the AI, when talking to, in the 90s. So he was, I think he was a little bit like, you know, it's good. You have this, like, go, go look at it, but, you know, we, we have, we have failed here once before, but, but, um, eventually he, you know, he, he, he was a very, very, very enthusiastic as well. And then he, he stumbled on funding and, and, and actually he turned out to be, I think, one of them, most important, deep learning, like, like, neural network, um, who have really built that community in the Nordics, I would say, um, and I think that was just, just, yeah, that was really a starting point. So, you know, a lot of kind of small events, was it a release of a library that's kicked things off globally? Do you think? Or was it just, you know, the ability of laying around within really graphic cuts? Yes. That's a good question. I think, I mean, it's like, compute, right, um, a lot of ideas have been like, build up over decades, uh, and, and small, smaller X on how you can optimize non-linear functions. Uh, and then, then of course, like, GPUs were starting to be fast enough back then, you know, it was a lot harder than this today, uh, two, two, many things on, on, on, on accelerators you have to, I mean, you, you had to work a little bit harder for your, for your kind of, uh, fast algorithms. So things like combination, right, of those, those, like, compute data. So imaging, it was definitely super important, someone collected a million images and, and actually labeled them, um, that's just like beyond somewhat odd mechanical Turk probably. Yeah. I don't actually know how they did it. Uh, it's actually, uh, I can't make any click. Yeah. It's, it's an American professor like Lee, Lee Fifa, she's called, and, and actually also the guy who ran a Tesla AI pilot, I was, what is it named, um, for us, I mean, no, sorry. Sorry for the, but, but I mean, it's just like really interesting time to be in that field, I think one more question, because I think this like history of, um, machine learning is super fascinating. Like, when this team, this university team came along and totally won the image net competition, just like a random innovation. Could someone in 2008 have taken it via GPUs and done the same? So was it just time for someone to discover that? Or was it really like where Nvidia GPUs in 2008, not fast enough, but it took until 2013? Yeah, so that's a good question. I think a lot of things happened, right? So I mean, Nvidia released something called CUDA, which is the programming language, like, yeah, see like programming, right? So you can actually write to some, at least in the early versions, there was still pretty hard to write to a GPU. So this is like early version, right? So you need to be a pretty good programmer, right? To like, first of all, just use a GPU, signal also paralyzed to move the GPUs. So super hard. So you need like skills on that. And then I think the other thing, well, I think actually this kind of AI winter was still a little bit going on in the sense that, that neural networks were considered a little bit in, like, you know, this bastard machine learning technique, right? That like, it had no guarantees on performance. So people like, like, basically, you're optimizing a super non-linear function with something called gradient descent, stochastic gradient descent. And that just, I mean, it turns out to work really well, but now, like, the formal guarantees of the performance are not, you get no guarantees. So hope it can be absolutely bad. And people have kind of, around that time, people are looking into these, like, probabilistic methods, like Gaussian processes or, like, niche is passing, like, graphical models. And those ones, I mean, they have, and patient inference as well, right? Where you, you have these, you do these kind of large graphical models, and you try to infer, like, probabilistic distributions or predictions. And some of them, Gaussian processes, for example, I have super nice properties that they guarantee, they give all kinds of guarantees. Turns out, then in practice, the guarantee is much worse than what you can get with a new network. So, so, I mean, it cannot be absolutely bad, right? But it's also not as good performing. And I think a lot of people were kind of a little bit, let's say, stuck, but at least focusing a lot on these, like, methods that had more kind of, like, rigorous guarantees on how well they could perform. And then someone just, I mean, you cannot argue with kind of performance in a competition, right? Well, there's no, you cannot cheat or, like, overfeed or something like that. So, so there was really, I think, what brought people around. So, I mean, many small things. And then I think actually, also, a lot of credits, you go to these groups that worked on neural network for a long time without having, like, really hard time publishing anything. And so, that's only maybe 15 years ago, right? It was super hard to publish anything from, like, with a, with a neural network machine learning method, because, was that not fancy? A little bit like today, right? You can only publish something that's neural networks almost, but someone comes with another method. It's super hard. Yeah. So, that, that's the, I think, you know, many, many small things, but also human factor, right, involved on what's cool and what's not so cool. But did they release Kuda to you know that? Actually, don't know that, but I think it's like early, early 2000s, if I were to guess, I think initially, I think it was mostly for scientific programming, actually, but turns out, like, you just need to multiply a lot of matrices right to do machine learning. So, so that's, I don't really know how the triangles at the matrices, how they are connected, but seems to work. I mean, a GPU is just like, it can just, like, do a lot of stuff in parallel, right? So, that's really it, right? And graphics is also a generally, basically, matrix multiplication. Okay, so you were there at the first conferences, on, yeah, people are, yeah, yeah, what happened then? What did you do? So, you, you, you've written your thesis around that. How did you move forward? Yeah, so then actually, as I said, me and, and my brother, we have this kind of an on-off relationship, then we actually started a company together with a few other people. So, we thought, right, we were in a very good spot when we finished our PhD, because there was like a lot of people wanting to basically hire and people that experience in deep learning and, and all that. So, we started a company where we wanted to do augmented hearing, basically, like, augmented hearing. So, you can, yeah, so that's basically, like, imagine you, you have instead of having a very expensive pair of hearing aids, then you take, like, our idea was to take like AirPods, and then do, basically have one microphone, and then we want to separate out a speech from background noise. That's a problem. You cannot solve perfectly, but you can solve it with machine learning, for example, by learning what's the, what's the noise, and then try to reconstruct the missing parts. And we, our idea was to sell it to hearing aid companies that for, you start the reason there's a lot of them around Copenhagen. And I think we actually made something work on a technical level. I think we just did a really poor job at explaining to the hearing aid companies why this was super groundbreaking, and the next thing they should, they should look at. So, they basically, like, I'll not call it turned us down, but at least they, they did not want to pay like anything close to what could, could make this into a successful company. There was a business. Yeah, there was super focused on this kind of latency thing, right? They said, oh, you cannot do anything with hearing aids where, you know, latency is more than a few milliseconds. And we wanted to run like the algorithm on a cell phone right, and then stream it to there. And then you have maybe 10, 20 milliseconds latency. It's a little bit annoying, but I think if the, if the kind of alternative is to not be able to hear anything, then I think it's a great trade off that you can slightly delay the very crisp voice, but at least we did not manage to sell them that idea. So, so eventually I, well, we, I took up a job in the job offer in in Google Brain in Amsterdam. So, move there with my family, and that's this kind of long, long-term research division of Google, and there's like really good good research, research out there, and so a really nice environment to join. And it's just kind of back then, I mean, it was really fortunate time for anyone like me, someone like me, because it was this, you know, there was a massive amount of calling like FOMO from the big tech companies. So, they were just like, they were really deploying a lot of resources in this. So, they were running this giant university lab essentially, where you, you know, you, you come and you, I mean, you're welcome, it's a, you get a lot of compute, and you just, you know, they tell you what you think is important for Google in three years, five years. And that's it. No restrictions, okay. No restrictions, and a little bit weird, like on-boarding experience, right? Because you're, I hope you're great. What do I do? And so, I did a little bit, I continued a little bit on doing some things on these like probabilistic modeling. Eventually, then started working on, on, on well forecasting. So, just me initially, and, and then a few other people, and, and we wanted to essentially speed up well forecasting. So, you could get something else close to real-time forecasting in, in Google Maps, for example. So, imagine that you can, you can predict precipitation, for example, with a minute resolution, and maybe a hundred meter, like, made it in time resolution, and a hundred meters or something like that, resolution ratio dimension, then you can, you can essentially route life cyclists and pedestrians around the, the showers. So, that was really the idea for us. And we started working on that, and also a little bit early days on, on well forecasting. So, we, I mean, just built that team, and, and eventually grew to like, 10, 15 people when I left. And, at that point, we were, like, testing the algorithms in, in, in Google products, right? But I, there was like three years, I think, working on that, we published a string of papers on that, and, and then also worked on course, of course, on, on internal development. Yeah. So, so that was my, my time at Google. I also, then the reason I ended up in the energy, actually, is because, you know, we looked at, Google, of course, procures a lot of energy, and they run these massive data centers, and they have a lot of, by a lot of power from, from, for example, wind turbines in the Midwest of the US. So, we also looked at, basically, like, instead of trying to predict, like, rain, you know, you plug in another thing on your weather model, and then try to predict the power production, and eventually, also the prices. So, we, we did that, and we, I mean, had some, I think, like, back then, we thought it was really good results on predicting power prices in the US. I think it turns out, there was a, a few things that we did not understand about how the power might work. So, it was, I think, was still better, but it was not, like, as ground-breakingly better than what we thought. But, essentially, you know, we could, we could, in the US, right, do have this, do you want to sell, at their head, or do you want to sell, and build time market? That's the question, as to ask yourself, if you're a generator, and we could just do that better, by, by basically, doing machine learning. So, that was my idea to deploy this. Did it deploy? We did not actually deploy that, not, not the energy part. I mean, it was not a research thing. Yeah, we had some kind of ideas about, right, that there, suppliers, like Google service providers had some issues, essentially, on performance. It was the starting point. But we, I don't know if it's deployed now. Of course, I don't know what's going on now. But back then, I mean, back then it was much more research. And also, it was like these heydays of our call, these industry labs, right? They were all completely open. We should just publish everything we did. And so a little bit the best of all worlds situation. So your thesis is that today they would not publish everything. They would keep things that you can see. I don't know about it. Yeah, I think definitely all the labs are not publishing anything anymore. So I mean, like Google and all the other tech companies right are getting a little bit more into what I would call a competitive mode or one mode or whatever you call that, where they compete with each other. So back then, it was more about growing the market, I think. You know, make sure that these people don't go to a garage and start something that will just rot us. So we want to inhouse all of that. Keep that busy. I don't know if you want to just prevent them from doing it. Or like at least make sure it happens in your own bag out. And not how did this Google labs concept work? Was this like a university? But we felt any obligations to teach. And you could do whatever you want. And it's just research. Or if you require to work on certain products, because you also said you worked on like, or the default was to integrate this into Google Maps. So how free are you at this time to decide what you want to do? I was a back then. That was like pretty much complete freedom. I mean, you're just, it was very sprawling. Google brain renders kind of very broad, like like non-hierarchical system, where basically, you know, you take. You get like, I don't think we were like a few hundred people, maybe 500 or something, across the US and Europe. And they just, you know, the goal is just to do like fundamental research in machine learning statistics, something that can help Google both short and long term. And there's very few kind of restrictions on what you could do. I mean, I think I were probably one of those who were a little bit more, I don't call it product minded, but maybe thinking about what can I do that would be useful for like a company like Google. That was something else. That was also why I thought whether it was interesting, because essentially Google, the mission was to, you know, make all information available. And I think one thing that was not available is choose spatial data. That's also for custom. That is just not readily available outside of the, like, very technical groups. So that was, but I mean, general, total freedom. Actually, the interesting thing is Google also has another research department, just kind of the sister or twin sister, what do you want to call it? Called the deep mind, actually interned at deep mind. So I know how it works there as well. But they run them on, like, kind of a radical structure where they, you know, they have these grand ambitions. And then they really, like, try to have these big projects and for a lot of people at them. So, so Google ran also those type of labs. OK. So this is where deep blue and all these stuff also came out of, I assume. Yeah, it's actually deep blue, actually, called it is alpha-fort. I say, yeah, yeah, yeah, yeah. Deep blue was there, yeah, right? Yeah, exactly. It was like '97, I think. But yeah, I mean, alpha zero, alpha fold, all of these things came out of deep mind. And I mean, I think both methods have the merits right. Google brain is definitely the more chaotic one. But they, I mean, Google brain invented the transformer right, fusion models that are now a huge thing right in image modeling were, like, some of the fundamental contributions just like happened at a few people sitting at the same kind of aisle of disks as I had, right? So, so, you know, it's just like a few people writing and pondering ideas. And then, like, you know, they're super smart, also, them. I would say much smarter than I am. Most of them are right, so, or maybe all of them. But, you know, then you, you know, you have these research ideas coming out right. And, you know, then they just, some of them just work. And they just, like, explode in popularity and turn more into an engineering problem. Where did they create the transformer? Yeah, so that's in good. So, this is like, this is like the, like, the origin of all the Chennai stuff that we're seeing today. What do you know? And also, it's just correct. Yeah, exactly. And also, a lot of, like, basically, anything in machine learning is on, like, images have transformed, image recognizes have transformers. And then, generic, generic, generative AI are pretty much transformers. Like, language models are transformers. So, all these GPT models. So, GPT is actually a general free-trained transformer. Like, abbreviation for that. There was the one and two ones, which really did not take off so much. But the third one, I mean, then they just scaled it to something big, right? But the kind of core idea was developed in Google. It's actually not called anything with transformers. It's called, I think, attention is all you need to the paper. And I think it has many, as many interesting things about that paper, but I think what really, what they did was they created something that scaled really well or fit really neatly onto the accelerators we have available. So, essentially, a transformer uses a lot of flops, meaning you need a lot of compute. And we have a lot of that on a GPU. We have a little bit less of data bandwidth and all stuff, so they're really good at utilizing really like a perfect fit for how GPU works. I think that's, I mean, out of the reason why they are working so well now. Take us back to weathercast, but they are like this global weather models. And I assume that the research you did also was based on this global weather models. How did you break this down to, like, smaller region forecasting? How does that work? How do you do that? Yeah, that's a good question. I mean, we just started thinking about what's useful for Google, so we started looking at the US. That's the biggest market for Google. And then our idea was to take like a weather model is actually an interesting, like how weather forecast works today is that you have a physical model of the atmosphere basically. And then you try to, it's simply to write up the like a partial differential equations that describe how will the atmosphere evolve. And that means that you chuck up the atmosphere into these kind of cubes. And then you need to know the pressure, the temperature of the humidity, and a few other things. Less than 10 fundamental variables. And then you can, is it right up the equations of how everything will evolve? The problem with, and that's just physics, it's first principles. The problem is that to make this computationally feasible, the cubes are something like a kilometer by a kilometer by a few hundred meters big. And a lot of the stuff you care about, for example, cloud formation, that happens at a much smaller scale. So that means that your physics doesn't really resolve the weather at the kind of right granularity. And then you start adding all these small approximations. It's called space-like machine learning models in a weather model. They call it something else, but really it's a function approximation to a physics problem that you cannot solve exactly because your scales are too big. So we thought that they're already doing this. So we're just going to rip out all of these physics and put in a machine learning model instead. And then just take the current state, all these values in these cubes, and then predict into the future. What will the future stage be? So that's really how we approach it. And I think that essentially is what people are doing today as well. And we had a lot of pushback on that, I think, back then, from methodologists who had all kinds of-- well, I think we just did not talk the same language on how you should-- what is the useful weather forecast? So they were used to looking at these physical models, right where you can see, like, layouts, and all these things kind of look real in the sense that it looks like we'll weather their forecast. So it's essentially on a physics simulation of how could the weather evolve into the future. And then they wiggle their initial parameters and rather another one. Then they do it 50 times, and they have 50 different kind of evolutions of the weather. And then they somehow created and picked the forecast. And we were more-- I said, I did my piece in probabilistic modeling, right? So we were interested in the probability of rain, for example, and not, like, some sample of, is there a cloud here? Just more-- what's the chance of rain at this area, at this particular time? And turns out, then, initially, it's very crisp. And then, as you move forward in time, like, your direction horizon increases. Then the-- like, it gets increasingly blurry, right? Because you are in uncertainty increases. So you get these more and more uncertain. Like, there's some chance of rain here, but it's not particularly high, not particularly low, because you really don't know once you are beyond to see this. And we discussed a lot with methodologists. They always complain. This does not look like cloud or rain pattern. And we-- We're always trying to explain that this is not a probability. It's not a cloud that ranged. We had a lot of interesting discussions on that coming back to energy markets now slowly. How do you see to weather forecasting? When do I see there are a lot of companies out there that provide weather forecasting. There is a bunch of players that are coming by the thing we're doing. Are I based forecasting and improving things? How do you see the scene? How will this move forward to expect the consolidation here or is this a market that has room for a lot of players? How do you see that? Back then we already, I mean, I was already back then convinced that this will work and we just saw these scaling properties right back. You know, we get more data, we get a bigger model and it just works better. My manager, he always had the thing right, you need a new compute data and conviction than your success is guaranteed. And I think this is just a really good example of that. I'm still super convinced there will work and we've seen the last maybe 12 months. There have been these like a lot of models that are now starting to work significantly better than the physics-based models. I think what they have actually not solved yet and I think that's the kind of missing pieces. So you have the initial condition right and the weather forecast takes about 6 hours on a global scale to make. What we have solved now is basically taking the initial condition then put in a machine learning model and predict the future. But the initial conditions are actually still derived from the physical model or something called the data ingestion. And that's a take takes a long time that takes about four hours maybe. So actually the bulk of the work still happens in a physical model in most. So we still and the problem with that is that these pipelines taking a lot of data and it happens at at NOAA in the U.S. and ECMWs and and they taking all these weird data from like planes that are kind of flying through the atmosphere and they measure the pressure and like the looms and ships and satellites all kinds of stuff. And they basically have this problem right that you have these kind of past measurements in the atmosphere. And now you have to somehow propagate that to get an initial conditions for all these more cubes in the atmosphere. That's just a hard problem. So they go back and forth between what does my data tell you what did my previous model run tell me what what did the model I ran maybe 6 hours ago tell me about what the state should be now. And then they have this kind of rubber band in between where they go back and forth and that takes about takes a long time and it's called like data assimilation if someone's interested. I think that's that's the that's the frontier now. How do you go directly from the raw data into into like a neural network model essentially and then predict like both the current initial conditions and the future like state of the atmosphere. But then on the consolidation part I think there's a lot of companies like this to wear in Europe, right? That's someone called Celuria Celurian sorry I can open in the West also a few other ones to climb it. It takes the tries to build these fundamental models essentially like a like jet GPT for for for with them. I think I'm sure we'll work from taking upon a few I'm maybe a little bit more uncertain on how they want to monetize it how I was the business model around this that's that's for me an open question why because I think you know you have this little bit like this problem where you know you are cut you're maybe a little bit better than then the current models maybe you're like let's say if like 20% better than the physical models are but it's also not like that the the the state one agencies they will not also do this. So Eastern dog you have already have their own AI model in production so maybe they will not for in the same manner resource arrived there probably being I maybe a little bit like a one generation behind maybe so you I think you can always be better right but how do you sell something where you are a little bit better I think it's called the mouse trap fallacy right where you know how do you know you have something that works but is it fundamentally better what you sell. And I think that's an open question maybe maybe it is super valuable or maybe if not it's just anything that's the open question for me that's a very interesting interesting point because I think like imbalance price for us with this castes with dexter energy you know to mirroring you have this slippage effects anywhere for limited market for a weather forecast my assumption would be if my forecast is just a little bit better then I'm the winner in that market and all the people become to me for the forecast and the rest would go out of business I would consider this a winner takes it all market yeah but then you have the problem right in in energy for example right you have this opposite problem right that once you forecast is the best one then your errors are the most expensive ones because then everyone will I can imagine like in Germany right is is what are they called the energy meteor is that what the name of the for if they make a mistake right then half of the traders will will make that mistake and then it's super expensive so I think that's actually an argument for in-housing it that you actually don't want best forecast you just want one that's decorated okay so it's a portfolio management question in the end put forward and risk management in it you are you don't actually care about the errors per se you care about how expensive your errors are and of course there's a at some point there's a regulatory issue right if you if you are really bad at forecasting then of course you you can create problems for the grid right but I don't think we are talking more like in in some percentest points better right so I'm for me that's that's I think for in energy that's the that's the interesting question for for these companies right how do they like either should monetize it themselves right and then they're not a weather provider anymore or or are they make it like they maybe make the value I think your previous guess from optimering I think he mentioned also right that you have the strawberry that you know once you are big you have market impact essentially you to cleat the value of your own forecast for me it's just a question if this is like a race to the top where we'd see a few companies exceeding it the rest will not be there anymore which would mean that they would have really need to invest yeah so I think that's that not sure I would if you want to be a fundamental like a physics simulator of the of the world essentially or at least out of out of shell of the earth then you need a lot of computers and you need a lot of in video data center graphic cards yeah you can buy whatever you are favorite accelerator is right but you need something that's a you need a lot of them right and that's that's a lot of money I think so this is a long-term investment game super perfect customer take us into twig now yeah so so actually I I left them a Google brain after three years great time there Google is a great company to work for but I had a few small kids so I moved back to Copenhagen more for for family reasons to be closer to grandparents and you know the drill for having small kids and then you know I you know I had the luxury of being able to take a little bit like sabbatical and think about what I should work on next and so I had three things I had like self-driving cars robotics and and something related to climate that was my my three list and actually very close to a joining a self-driving car company called Wave that is now turned out to be pretty successful in in the UK unfortunately I don't know my wife did not the last minute a second the the break on on moving to London with some some small kids you have this didn't happen are you happy that this didn't happen because self-driving cars I think that this is like the the next winter that we are seeing the self-driving car winter or do you know I disagree agree in the sense that I think most companies have the wrong approach but I think actually the wave for example that I was in talks with I think they had the right vision that they wanted to solve this as and like no high definition maps like insulin machine learning no mistakes just you know you have a huge essentially like a new network that goes directly from steering and and like something that resembles Google maps goes from like you know you want to go here on this map and then it has like a bunch of cameras and and then you're you're you're going directly to steering and breaking an acceleration and it works it's amazing to see that drive around in London that thing right there's no kind of hard-coded logic on it like avoids to pedestrians bicycles buses all kinds of that's really amazing piece of work they did that and I think they the good thing about that is it's very scalable right you know you don't have to you don't have these millions and millions of lines of weird rules you just have a huge like the machine learning model in there you need to feed more data and really understand when it fails and then you you generate data for the failure cases failure most and I think they compliance perspective if I would have feared I can understand the beauty of it. Yeah, there's of course like compliance right and I think I also think maybe compliance for machine learning is a little bit sometimes you know people have this idea that you cannot that it's completed black box right which I think is a little unfair in the sense that you know a human being if you if you could kind of like look into a human brain in the moment it makes a decision most likely they not always are super conscious about why they do it you can always ask them afterwards right why did you do x and they will have a perfect explanation retrospectively right so I think if you could kind of ask them just in there right yeah but there's some regulatory issues of course they are actually driving around in London and if you as it is so so I think they somehow have solved that are they doing better than Waymo mean that's probably a hard question to answer I think Waymo have a lot of cash right so they are a Google company I think Waymo is more advanced from a like or maybe more mature I think I think someone like like fully machine learning driven is just fundamentally the only thing that will work because it's impossible to write up the rules on how you drive a car that's I don't think you can do that that's the not possible I think I lost my belief in self-driving cars within the next 10 years but let's see let's see okay yeah join the self-driving car company probably a big maze probably not we don't know no you can spend all your time thinking about that right but but I then I eventually started like holding on on climate and how can I essentially make an impact here and you know I have a problem in the sense that I am a computer scientist and I think we have an infrastructure problem that was the fundamental question I in front of me and thinking about how can I help with converting my my knowledge on computer science and machine learning into impact on on infrastructure that will make our world a little bit a green carbon neutral or whatever we want to call that so I mean I knew that we I had some good knowledge on machine learning right and and also on weather forecasting actually then what happened is that my brother he was living in California Cupertino worked for Apple their special projects team on something very secret and then they he moved back to Copenhagen and then we you know started chatting and about how can you how can you do something interesting here and and I started then in Denmark you know there's a lot of energy traders so I started visiting them and just kind of you know trying to figure out how does this energy world even work and and you know where does the money go and how do you actually operate it so that I did not actually figure that out because you know they're super secretive they don't want to tell you what they're doing so that was for me intriguing I think you know I'm the type of person that then thinks that you know if they don't want to tell what they're doing then it's because it's not super hard it's the the one you figure out what they're doing you can also do it and the other thing was that you know my background is is is in computers right so I think why why do they have all these people like have these like iron eyes of people sitting in front of desks and the and you know trading power 24/7 why can it not be automated might as it have to work that way so that was the what started a trade energy eventually so so we eventually had basically like you know a three-step plan to how very naively enter this market and and have an impact on on the grid so so the first one was that we can automate power trading that was just you know point one I mean that was you know easier maybe easier said than done but but I think we can now say that we have we have solved that and the the next step was was to integrate then with with basically infrastructure so meaning basically we plug them into our our platform and start operating them and selling the power and and controlling the operation and and the third point for us is is to make sure that we deploy projects fast enough at a large scale so that was our three-step plan for for business plan that we then went out and tried to sell we got a few investors abroad and then started programming and and you know getting market access and eventually I think around maybe 21 just before end of 2021 we got licenses to operate on some Tzos and then ran like a more rudimentary version of an automated trading system back then did ran me and my brother and and that was very stressful right because we had a 24/7 operation two people that was well not not the best few months of my life but then we hired in a lot of great people a little bit later after we got proof of concept and then I think just been growing from there basically adding more markets making the platform more advanced more robust all the all all those things right yeah and I can maybe I can how many markets did you start with just we started out just with one Germany empire that was it and then I mean we just added more markets right since then we are pretty much at all European markets and there's a lot of things right that you don't know you have to do you have to sketch your right you have to we actually build our own scheduling system and integrating on this email client with the S was very painful so it's a great off that was hard eventually we found out you could buy that solution but you know when you don't even know what it's called then it's hard to buy yeah but I think I think the interesting thing right was that I think we joined it I mean we were kind of lucky in the sense rather we joined just before like Russia invaded Ukraine right and all the a couple of things that happened there right but but on the power market you had this kind of massively volatile environment then for for a year something and we I mean it was not hard to make anything work in those times you know there's you know honestly if you had a reasonable prediction about anything then you could earn some money on power back then so we were just like fortunate to start at that time and scale up the operations back then so that was that that was really the starting point and I mean we still don't have any I mean we have no human traders back then we still don't as I said I think we don't have anyone with power trading background right so we run it as a engineering shop which I and also we're locating companies actually Denmark has a lot of power trading companies but they are actually all located in another part of the country it's called August they have these denser commodities the the OG of energy trading in Europe I think and they basically you know there's been like it like this kind of a small small droplets of new companies dripping from a denser commodities the last 15 years or something so there's a lot of really big community of power trading you know but I mean we're actually not really connected to that for just different company we tap more into the computer science and machine learning now so that's just interesting to kind of you know see two different approaches I think they both have merits right yeah just don't understand how you can sit in front of the screen eight hours and being super concentrated and execute I think it's interesting that you went in and said okay we're going to build a prop shop on then we're doing assets relatively so yeah I think that's the I mean for me one thing I learned from my previous startup is that you need to find somewhere making money right and you can either get that from investors or you can you can you can try to rhythm yourself and that's I mean I was a little bit tired of making a product that is hard to sell so show you know we just wanted to make our own money essentially but and and I think we we for me I think I think the power might go well like short term power will be fully automated from that's that's I'm not any doubt about that it's more matter of how what's the time horizon on where it will be automated that has always been my belief but from back then and I still think that's true human traders have a hard and harder time on keeping up with the information and the the speed that especially as you get hooked to the delivery right so but then I think I think that the not the problem but at least that one issue with having a prop shop right is like all these prop shops they they have this you know we help stabilize the grid or whatever right and move the power to wherever it's needed I think that's a little bit like too shallow and ambition I think they can do more and I think they should do more and and so I think you can you can utilize a lot of that knowledge that you learn by by operating on on the power market to make the asset operation much more efficient for me I think like long term I think that's that's for interesting goal for us right that now we I think we are a large participant on the intraday markets and I think make those market more efficient, more liquid and all those good things. But I think that also only scales to some, like at some point right, those markets are efficient, right? And then the marginal gain of being more efficient and not so high, then I think the impact then for us is more direct in the sense that we want to participate in making sure that assets actually get deployed on the grid. And if we do well on the trading ride, we also, we can support that impact that up with money right and know how. So you put strap yourself with some prop trading to now be able to trade assets. What kind of assets do you not trade? So we trade solar and batteries and like load banks, flexible load. Just the three major components, then there's some variations, but really that's what makes up the bulk of it. Yeah, I mean, you know, it's very interesting with those as you also wear the power markets are relatively homogeneous across Europe, like antular service markets are also in principle relatively in harmonized. But the entrance or like the actual market exit is not so easy. So there's a lot of work on APIs and stuff like that. And also there's these markets are pretty shallow. So they quickly fill up and then they're really high prices for a while and then you just completely tank just really sort and so it's interesting experience to be in those markets. So for example, in Northern Europe, right, you have a market called FCD that's like an antular service market, like it's like FCD containment research. But basically that went from like hundreds of euros per hour per megawatt to like two in two months, and that's that's painful right, if you heard that shock primary source of income. Yeah. That's a risk. That's a risk. Exactly. That's telling all the people that are going that are betting horses on insulin markets. It's not a market that's getting bigger. No, it's interesting right because it's kept, it's like it's, I mean, regulatory, I mean, we need X amount and we'll buy that pretty much whatever it will cost. But yeah, I think I think it will always be a part of the revenue from flexible assets. It's just more, you keep, I mean, also in the Nordics right, you have these insane revenues you earn maybe somewhere between seven, eight hundred, nine hundred thousand per megawatt installed battery capacity for a year. It's like a payback time of 12 months of your project, which is also like, I mean, that's just not sustainable right, humans can can, can sell that and keep them running fast. Yeah, but there are also people out there that are saying, yeah, we are also deploying new renewables, the insulin markets will big, we'll get bigger. This is what people are saying, totally, totally heard the contrasting opinion. They're trading away the day in balance, this will not get bigger. If you are relying too much on the cellular markets, and we're missing out on short-term power wholesale trading, then you're making a big mistake. Yeah, I think that's true. And also, you need to, you need to stack out your assets differently right for participating in energy trading compared to, like a lot of the NGO service markets are like power rated instead of energy rate, and I mean, energy trading is matters like megawatt hours and not so much on the megawatt, so that's just fundamentally different specs of the system you want. And also, I think on the scaling, the, the ventilator service markets are scaled based on an in minus one event, so meaning your largest asset will drop out, and now you have to survive that. How much do you need to procure? I mean, that's just not going to 10x, you know, you're not going to build a nuclear land of 10 gigawatts that will drop out. That's not possible. You, so the big, the scaling, like the limiting one in the Nordics right now are this finished nuclear reactor called OU, OU Kula 3. Sorry for my pronunciation of that, but that's the interesting right, you add a lot of firm power and then so nuclear as firm right and and base load, and then you are actually a service requirements increases because now that's the biggest asset online in the grid, but yeah, but I totally agree with you, it will, I mean, it's a cap market, I always expect this to be an equilibrium between a salary and that's what happened in Germany right. So now actually a lot of the people who are in FCR, they are now back in in the intraday and like these quarter auctions, because there's more lucrative compared. So then actually FCR prizes have increased in Germany for a while and just because you make more on being in the energy market, so I told you guys, I think that's true, you that's the stable point you'll reach. Where do you find the assets now? What is yours to add the cheaper going forward with twig? That's a good question, I think that's actually very hard and one of the things we're thinking about, like where's this going long term right, because I think actually I listen now to a few of your podcasts and I think you said right that the aggregator role is, it might be a tough one in the future and I to some degree would agree with that. I think there's a lot of like the pure like aggregator role I think is not a race to the bottom, but maybe that will be competitive, I think, something we learned. So a lot of the like where you find your assets is more on the building partnerships. So again, like we have a lot of knowledge on like how you should take out the assets like what are the market development, all this getting on early with like strong partnership with developers or people who invest. So you inform them about how they should invest and then part of that is also that you operate the assets. I think that's a way forward for us. I think other companies are maybe following a little bit more like a commission, like having a very efficient sales organization, I think that's a said that what you need if you want to hunt assets on conferences and stuff like that, that's just not what we are doing right now, we are more on the hardening with fewer but closer and I don't know what's the right way, but it's just more also, I think what we are good at. We are very tech heavy organizations or so like we need to talk to people who know like technical details as well. We cannot manage like many small customers, we can manage to bigger ones. That's how I worked. I think there's also a huge difference between battery trading and the classic renewables. It's more about how do I distribute risk, how should the contract look like? So this is a lot of origination stuff, doing special types of contracts while flexibility trading is, yeah, sales force on the ground, being active in a lot of markets, having the ancillary connections, so they are different ways to succeed and our like long term, I mean I think what we really, I mean for a long time in the Nordic for example, like it was pretty easy to be an aggregator because you just need to fit into a city and if you fast markets they're called and then you're pretty much doing optimal. Now those markets disappeared and we have always told our partners that you need to be ready for the energy markets as well. And that's where I think we really good is on the trading part, right? And that goes straight from our experience, from the like the intraday markets in general that we know how they work and we can, yeah, we can we can operate those efficiently. So that's what I think is our competitive edge here on the asset operation. Of course, it also means that you know, intraday and energy has to be a big component of the revenue before that vision or whatever one call it is true, right? I think we're seeing some signers of that now, especially in Germany, also in the UK, and then our belief is that I've got to have them in the rest of Europe as well. What's next for Twig Energy? Yeah, so what's next? I mean, so, you know, these ones are pretty much the, you know, 0.1 and 2 on my business plan list, so also, I mean, automated power trading. I think we have solved that in a way, it's more an hour scaling issue. And on the intro getting with Ashes, we also have an alright sized portfolio now of asset in operation. So the last one is how do we enable projects being deployed? And that's what we are pondering with now, how vertically integrated should we be? So one thing we did was that we invested in a battery project or developed it with a developer on Denmark, software has been operational now since, like this summer, Mr. Biggest project in Denmark, so we owned that in a joint venture with the developer. And I think, I mean, that's an interesting way forward. I think as how can you to enable the projects, and this is a very direct way, right? You know, we have money, we have no other hand, and but we are basically figuring out how can we make sure that we build that scale? So that means that we are not so interested in building, you know, these one, two, five megawatt projects. How do we build like 50, 100 or 200 megawatt projects that actually will have an impact? And currently, I think it's lithium ion projects, right, for someone like us that are interesting down the route, like we know is what's the next technology, but it's probably plenty coming along. So that's what we are currently looking at. How do we, do we have a role to play in that field? That's the question we are asking ourselves right now. Yeah, I think that's, does that answer your question? Absolutely, absolutely. As we're nearing towards the end of the episode, Kasper, I do have a few questions on topics that are regularly discussed that I ask every guest. The first one is, how do you see the hydrogen future? Do we, we'd even hydrogen, better you see applications? Well, I think like burning hydrogen as a fuel is, what's it called, a crime against thermodynamics. It's, I think that's just not efficient, you know, I think maybe to me, I'm a little sure, so airplane? Yeah, but I don't think, yeah, I don't even think hydrogen there. I imagine a pressure vessel on a plane, right? Your plane has to look very different. Like you cannot put pressure vessels in the wings of a plane. That's just not gonna work, I think. But I think maybe it's the answer a little bit, or maybe you have asked the question of what kind of burn instead of fossil fuels, and then your answer is hydrogen. And I think maybe that's just the wrong question, right? You should have asked yourself, how can I, maybe solve my problem? And then you would have already electricity, right? With, with many of those applications. I think for, I'm not an expert, but I think like to me, it feels like biofuels are just like perfect for planes and ships maybe also, and we should just stop burning biome as in like, you know, bio or what are they called? Like burning woods in a, in a kind of the, for heating is just dumb, almost I would call it like nothing. We should use those, but like that bio material should be used from very, like those really hard to obey applications. Okay, second one, what about nuclear power? Nuclear, so also, you know, next controversial one. I think, I mean, this is why I'm asking the question. I mean, I mean, the easy answer is right. You need to, like, don't, don't, don't shut down nuclear. That's really not smart. So keep wanting those, I think, I think you need everything. 50 years of eight. Yes, I think we have put ourselves in such a bad place now that we need to do everything we can. And that's include like running all the nucleus. We have also building new ones. I just more think the timeline is like in zero pride, we're talking 10 years maybe for 20 years on, and we can do a lot in the meantime. So I think like we just need to build all renewables and build all the new free began. I'm pro any kind of power addition to the grid that we can. And in Europe, we have kind of lost ability to produce new power almost, right? We could do it, but very expensively. Yeah, that's true. I mean, that just feels like, I mean, we just like all decided that now, you know, maybe it's mostly an offense to the Germans, but if the German, the German, you could say, we are happy with using a French design, and we all use that, and maybe that would be cheap. But I mean, I don't know. We need to get projects going. I think that's true. Medium size reactors just make more sense to like do one big copy that. Yeah, I think it's not a, it's not a technical problem. It's all kinds of problems, like political problems, permitting, zoning, all that stuff, right? Now do you see the future of US power markets? Or generally? That's also a figure of US power markets. Yes, I mean, we also actually active in the US power market. So I think just, you know, a very different system, right? It's not surprising. I think create power markets, which we can say that, but I think what's very interesting thing at least for like, for example, in Irkord, right in Texas, I always find it amusing that they, you know, they have very good alignment between their ideology of the free market and the way they designed the power system. So, you know, they have an energy-only market where you basically have a very limited capacity markets so that this is where you pay people for being backed up right in Europe that is very small. In Texas, instead, they have this system where the energy price will go extremely high sometimes when there's like too little power, too little right. And so, and then they just say, you know, free market, you know, people can expect that and it's just kind of a nice that they align in their way. The other thing I was there about, you know, in Europe, right, we always, you know, I think taunting the US a little bit about reliability of their grid and its old, and all that stuff right, and it's broken and blah, blah, blah. I think that's probably true. But on the other hand, they have extremely cheap power compared to Europe. And they actually, in Irkord, Irkord is, you know, like Texas, they add a whole lot of renewables at batteries. It's by far the ones that add the most in the US as a state. And that's just, you know, they have this rule that you can, you can get connected to the grid, but we might curtail you if there's no capacity. But we'll connect it fast. And in Europe, we have the kind of opposite way that the guarantee that it will not curtail you, but it will take you seven years or something, whatever, some pretty long time. And I think that's just, for me, interesting. I think, you know, it almost came to the point in Europe, right? We don't have a, we cannot add new data to it, but like power generation in Europe, right? So we have no data send us, for example, almost in Europe because it's not because of course, it's just because there's no power. And I think that's kind of a, that's sad. Yeah, I mean, sure, if you look at the US Henry hub gas prices that we're really low. So if we would have this in Europe, then we would also have like electricity prices in the same rate. Yeah, that's probably true. We had that, we had that like gas prices, like I can remember TTF gas being traded 10 euros per megawatt hour. And yeah, that's true. I mean, now we lost the Russian gas, right? There was, and I think now we're, I mean, LNG is much higher price, right? So that's what we rely on now. I agree it's not so simple as I put it right. I just more think the kind of, I think we have a problem with like electrifying fast enough because we do not add enough power generation to the grid. And I mean, that just means we don't have any data send us. We don't like we de-industrializing all that stuff. In Europe, I think that's, well, that's not a good trajectory, I think. Just because you said great market design, great market design in the US, which role is it that you have when you say great market design? Is it as a pure trader or what do you generally think of the market design compared to the European one? You're very deep in both. Yeah, I mean, I think both are great. I think like, I think the US as, you know, they basically have traded like spatial granularity. So they have a lot of different prices at each substation. And then they have low time resolution instead. And in Europe, we have this have less, I kind of, spatial granularity, but we have this very fine granularity of the intraday market in time, right? So I think, I don't think even to both, but maybe you can kind of go in between those two. I don't know if they're good, but also the US have something called the central dispatch system, where the ISO is called the controls who are dispatched in Europe. We have this where, you know, it's actually the balancing responsible that dispatched on units. So I don't actually always want to like best, I think both works and I marriage, yeah. I think the European one feels a little bit more harmonized. And, and, and, and kind of make me mature on some ways, but I don't know. I would not know which one is best or most effective or anything like that. I just think they're interesting, different design choices for solving a problem. That's, you said to me, how do you fries power in a grid? Super last question. Can you name a rising star in the energy industry? It cannot be, it cannot be power, but yes, no, I would not name myself. I mean, so I'm fortunate enough to, we have a very nice office and some very nice office mace from rail and it's a real energy. And they do essentially, like they, they essentially make PPAs available for small and medium-sized companies, and I think it's great so you, you can purchase a fraction of the PPAs into the, and then they will act as your power, but I think it's super nice. and in a very direct way. of contributing to the addition of new power to the grid. And then I have another one. And I'm just being patriotic for the Sega of being Danish. But there's not a lot of interlaps. A lot of great data companies out there. Sorry. So data labs. Sorry. Interlaps. Interlaps. It's essentially also an entity space, but they do entity optimization for buildings. So imagine that you can then we have a lot of smart meters and you can hook your smart meters for power and for heat and up to that platform. And then they will try to figure out if you like given your building layout and your like how it looks and what just what do you like a building of similar size normally used. Do you then have some times when you use too much or do you have a lead in water for example or a new heating system. And they basically sell that and this kind of you know a little bit. And you know optimization is always a little bit great I think compared to development of new stuff. But I think it's just very direct impact on and they can they can save a lot of the energy on many buildings like. And any 30% ride that's pretty significant given a lot of our energy goes into heating alright. That's it. Thank you very much. Casper for coming to our show. Goodbye to all the listeners out there. See you in the next episode. Bye bye.