David Wood, JBA Risk Management & Jochen Papenbrock, NVIDIA: Showing the world how to revolutionise modelling (388)

Hello, welcome all welcome back to the insect podcast, Zoya here, and this week we're joined by JBA Risk Management and NVIDIA. Two organisations working together to revolutionise catastrophe modelling with high performance computing and AI. Joining Matthew in the studio this time are David Wood, managing director at JBA and Jochen Papenbrock, head of financial technology at NVIDIA. To get it, they explore how GPU powered flood simulation and physics informed AI are enabling faster, more accurate and more scalable flood models, both in the UK and around the world. If you're curious about how AI accelerated catastrophe modelling is reshaping, how insurers understand and manage flood risk, this episode is for you. Now let's dive into the conversation. Jochen, Dave, it's great not just to see you but to see you in person, we're back in the studio for a face-to-face interview, managed to get both of you in London at the same time, brilliant to see you. Jochen, welcome back to London. Thank you, thanks for having me. And Dave, coming down from Yorkshire. That's right, yeah, coming down for the Oasis insights conference, but I'm going to head down afterwards. So a good opportunity to come into the studio. Great conference, lots of interesting people there. And Jochen, Dave won't be a fan of when I say that it's always wonderful, we talk about insurance but everybody I think will have heard of in the video and it's always great to have people coming in from outside of insurance. So thank you for carving some time out of your busy day to join us too. Sure, yeah, I mean, Nvidia works in every industry including insurance, banking. That's where I work. Well, we'll hear more about that in a minute. So just a bit of an introduction. So JBA, you found it back in 2012, or the modelling part was, you're one of the leading providers of flood risk data in the UK and also now you've got a global catastrophe model for flood. Dave, you've been JBA since 2016. We'll hear a bit about your background in a minute. Head of product development for five years and now you are MD. Jochen Papenbroker, you are head of financial technology and I'm reading this from LinkedIn, lead DevRel Banking, Amir at Nvidia. So welcome. Jochen, I'm going to have to come straight back to you on lead DevRel Banking. What does that mean in practice? Yeah, so DevRel is short for developer relations manager and Nvidia is a developer first company. So we engage with developer ecosystems and my role basically is to connect into this ecosystem to help the ecosystem leverage our compute platform, our software frameworks and so we have a lot of partners and video as a partner organization, cloud service providers, computer makers, consultants, software vendors and we connect them and then work with the banks in my case because that's my focus industry and also the insurance companies. So that's why it's called a developer relations manager. Now I say it's a fascinating concept. We'll talk a bit more about this. It's a developer first one, particularly as it explains a little bit, I think about how Nvidia is scaled. And of course we talk a bit about how Nvidia is the heart of a lot of what we're doing or what people will know with generative AI. But Dave, just come back to you a bit. So the pleasure of doing this podcast is that by definition everyone we talk to is ended up in insurance or I guess in your case working insurance. But you came at this from pharmaceuticals and then that was all a little bit not exciting enough so you find yourself an insurance but just tell us a bit more about your background. Yeah, sure, studied chemistry originally and pretty early on realized I was never going to make a lab chemist. So I moved into computing something that was called cheminformatics and that was all about data modeling and data science and its application to pharmaceutical research. After a PhD for about ten years I worked in the pharmaceutical industry and several big multinational companies and also a startup as well. And the focus there was really about developing data modeling platforms that could accelerate drug research. Now the pharmaceutical industry had got a little bit heady after the 90s days when it started making billions on individual drugs and there was a bit of a calm down from that. So decided it was time to look elsewhere and discover JBA. But yeah, the surprising thing for me I think is really how much of a crossover that it is. There's a lot of data science and statistics and software development. There's a lot that relates to. And a lot of uncertainty as well which is kind of why you need the computer which I guess brings us back to you. So I mentioned people will be familiar with the video as powering the Genevieve I today. Of course the history of the company goes back and I think it's quite an interesting parallel between this back around about 2012 with JBA and what happened in the video. So it'd be great just to hear a little bit more about how didn't the video get to where it is today. I mean more than 30 years ago we invented the GPU which is graphics processing unit. So it's a piece of technology that would help accelerate graphics for computer games for example. But we very early on realized that you would be able to program these things and build a programming language around it called CUDA. And around 2012 people were experimenting a lot with AI models, deep neural network models which is multilayered deep neural network models. And there were competitions around like who has the best model to differentiate between for example cats and dogs, a typical example of a predictive AI model. But there wasn't that much progress until this AlexNet competition where some researchers were able to leverage GPUs for training models much faster. So they were able to expose the model to a lot more data and examples. And this really improved the accuracy of these models a lot. And that was sort of a breakthrough for us because we had already developed software around that. But then we invested more in building CUDA and other libraries based on CUDA that would specialize in certain mathematical operations like matrix multiplication which you need for deep learning, the training of deep learning models. So that was sort of our ignition into the deep learning space. So we are really a very deep AI company these days. And sort of implicit in what you said, but you mentioned earlier that before you started that shift over to GPUs and the scaling factor is pretty extraordinary. I mean, I think you're saying like a sort of 1% improvement in terms of that they'd be able to recognize the cat versus a dog. But then that changed dramatically didn't it once you had the sort of the more powerful computing power? Yeah. So by accelerating the compute you can expose more material to the model. You can make this training much faster, much more efficient, safe energy, safe cost. And this is what we do. And later on we did the same with generative AI and transformer models. Yeah. So chat GPT wouldn't exist without NVIDIA GPU technology and all the other models that we see physics AI, gen AI, agente AI. So really driving that actually we are in the middle of an industry revolution that is powered by accelerated computing. And generative AI, AI models is just one example, but you can also apply this, accelerate computing to, for example, simulation, Monte Carlo simulation, numerical models, for example, in weather and so forth. So there are many workloads that you can accelerate with accelerated computing, including data processing, data preparation, which is important. Obviously, before you train your nice AI models, you need to clean up your data and curate the data sets. So we have built sort of end to end accelerated computing stacks and SDKs. So people can benefit from accelerated computing power in every step of their AI journey. We like to pick our acronyms. So SDK software developer kit, I think, if I've got my acronyms correctly. Yeah. Exactly. Because I mean, developing hardware and massively parallel compute with GPUs is one thing. But you need to make it accessible and available to the ecosystem by providing software frameworks, development kits, blueprints, and microservices such that developer communities around the globe can get easy access to those. Ideally, leveraging existing open source frameworks and accelerating them with a GPU so that developers don't really have to change their original code and can easily adopt GPU technologies. What was it that took you to the video? How did you discover the company? How did they discover you? Yeah, I ran my own AI startup working in financial services, building some models and products around that. And then at some point met some people at Nvidia. They had this program called Nvidia Inception, which is targeted to startups. And at first, I didn't really take notice, but like a few days later I had more time to read into the material. And I was like, whoa, this company is really ahead of the curve in AI. I was really surprised. And yeah, that was my way in, basically. So Dave, that brings us back to JBA. So back in 2012 around, and you also, JBA had a recognition of the power of GPUs. Yeah. In fact, it was a little bit before then still. So 2012 was when the JBA risk management form, but the origin story goes back a little bit before then. And so it was early 2000s that Simon Waller with his team started developing our flood simulation software, which is called J flow. And actually, it was about 2007, York and I think, or they're about the Rob Lamb, who's managing director of our JBA trust now. He had a really early insight that these flood simulations could be accelerated with the GPU computing. So, you know, that predates CUDA and some of the other libraries that made it easier to do these things. But, you know, it was a couple of years in the making, but what that allowed the team to do was to accelerate these simulations by a factor of over a thousand, right? And all of a sudden, that enabled the possibility of doing high resolution flood simulations at a national scale, you know, the team produced five meter resolution flood maps for the UK. And these flood maps using this technique became a gold standard for assessing flood risk in the UK insurance market. And then after that, it led to flood maps for the whole world and catastrophe models as well. So, yeah, it was quite early days for GPU computing, but it was that innovation that really led to the formation of JBA risk management in 2012. Yeah, now, I mean, I, at the same time, different company, but it is recognizing, well, not personally, but I was recognizing the power of GPUs. I can share some money into the video at the time. But let me, I guess, blood also is so complicated modeling it. You really couldn't build a credible flood model unless you had the power of the GPUs to be able to do it. Unlike, unlike, I mean, more traditional early models for hurricanes or earthquake, you just, they're really critical to get the, I guess, both the resolution and take kind of all the variances you get in and building a flood model. Yeah, run the full simulation, how the water flows over the terrain models, that's right. And so now, what do we, uh, 13 years on, or we don't, you know, in 2012, but so what's happening now with the relationship with, with the video? Yeah. Okay. So, you know, there's obviously a little bit of history there because it was an innovation based on Nvidia's chips on their GPUs that led to the formation of the company. So there's history and affinity there. But we're also using AI extensively through all our development processes and a range of applications, things like observations or, you know, finding ways to accelerate our flood simulations further, identifying quality issues and we're using language models as well. So, AI has kind of become part of the machinery at the company. But our interests recently were in the emerging AI weather models and we wondered whether these would have applications in catastrophe modeling. You know, would they allow us to produce richer events sets, uh, for instance? So, yeah, with the event sets, I think the principal challenge for us there was that you're working with the observed record. It's a few decades long. It's a limited source of data and we've got to build up our risk model from that. And, you know, whenever a significant weather event happens, people tend to say that it's unlike anything seen before. So that was the problem that we were looking into trying to address with AI weather models. We could use numerical methods. We could run simulations to produce new unseen weather patterns, but these are really computationally expensive. And so, again, the advantage with the AI approaches is the kind of acceleration you get, you know, a thousand times faster again. So, yeah, around that time that we'd started looking into these techniques to see whether they had applications in events sets that you can got in contact and showed us some of the latest developments with Earth 2 and that led into the project that we've worked on with them. Yeah. And separating the high from noise. But you're just on that one. I mean, again, top to it for people that if you're trying to learn and separate all the noise and high marketing from what the reality is. What's your technique for getting to truth in a busy, busy life? Yeah, I think it's important to make sure to be connected to the right ecosystems and to learn what's the state of the art and where I do people gather to discuss future developments in AI, right? So, GTC conference is one of those places where I got to learn about Nvidia more and about all the other AI companies that work with us. So, it's a huge ecosystem around that and seeing what these people are building, what they are working on is very important also for a startup person because as a startup person, you think you are the genius. You have the new product that nobody else have and just attending one of these conferences is a reality check because you will find companies that are doing what you had in mind for almost like three years or whatever. So, you get informed, you get benchmark, you get state of the art and art of the possible information, then you add your own ideas to that and connect to that ecosystem and actually build something. And the GTC conference, that's what the conference in the video is, one of the leaders in running that in California. And I know you, I think JBI, were talking that at the last conference you ran. Exactly. GTC stands for GPU technology conference and it's power by Nvidia, but in effect, it's the entire computer and AI industry gathering, meeting and presenting and there are also a lot of education and training workshops around that and we invite our entire ecosystem to present the latest and greatest and also to educate people how to leverage GPU technology. And of course, JBI has got a presence in the US and you got a presentation that I think is now available online from the presentation from GTC. Yeah, that's right. Yeah, it's available on the GTC website for downloading and listening. It relates to the work that we've been doing with Nvidia on Earth2 and AI weather models. Excellent. So, working with JBA and insurance, you mentioned before Physics AI, which is kind of one of the other, I guess, sort of branches of, of generative AI we've been learning about. I'm sure other insurance companies are talking to working with the other, any examples you can give of people you're doing work with? Yeah, absolutely. So, at that GTC conference, there were around, I don't know, 25,000 visitors and hundreds of companies presenting and there was JBA obviously and Dave on stage and also our engineers, but also there was the insurance company AXA presenting their take on Earth2 and on weather models and the hurricane tracking how to leverage Physics AI models there. And for those ants, Millie, with Earth2, can you spend a little bit about what that project is? Yeah, this is one of the go-to-market strategies of Nvidia. Just putting the GPUs is not enough. It's a very important piece of hardware, but you need to make it accessible to people. You need to build these frameworks that we talked about, right? So, and we saw that AI will also revolutionize the way we do atmospheric models and the way we get inside about extreme weather and climate. So, at Nvidia, we were thinking about how could we build a platform for people to leverage the power of supercomputing in this area? Yeah, so we built a certain framework, certain platforms. We built something called Earth Studio, which gives you the ability to build inference pipelines and to leverage open source models and leverage Python libraries. Yeah, so make it really easy for developers to build something in their domain. So, they could keep concentrating on where they create the value, where their expertise is, they know how, but at the same time, put the easy button to the accelerated computing part, because that is quite complex, but we really simplified as much as possible with huge investments actually into the software and software development kits and frameworks. So, people could really concentrate on what they are good at, but at the same time benefit from accelerated computing. And David also mentioned it's you can create more simulations and you can accelerate everything here. And Nvidia is really obsessed with acceleration and compute performance for good reason, because to create value in the industry, be it the insurance or the banking or whatever industry, leveraging accelerated computing, you can make things faster, easier, cheaper to run, you save energy compared to not using accelerated computing. So, it's also the ability of certain companies to leverage that compute to create a competitive advantage or to create value versus not using this. So, we are always very keen on benchmarking, using accelerated computing versus not using it. An access to STF people who want to use that is a subscription model or the people can they get access to it for free or how does that work? Yeah, it's several elements of earth to platform, there's physics, Nemo, which is an open source framework for training these models of two studios like an open source Python library to build custom inference pipelines, including pre-trained models, data pipes, and you can build those huge ensemble pipelines to create more scenarios, let's say, of future weather developments in just in no time and no cost compared to doing everything numerically. Yeah, so the acceleration benefit, it comes from the GPU power from the hardware side, but also it comes from so-called surrogate models. This is this physics AI models, it's like physics informed deep neural networks that would approximate other models, and by this you can further accelerate generating, for example, large ensembles. So the entire acceleration benefits comes both from GPU power, but also leveraging the right AI models and training them to simulate the data. Yeah, and I think what's really interesting is somebody listening, thinks they're not technical enough, is it actually the concept or not that difficult to understand technically? By offering more scenarios, you do what ensembles, which is basic actions, are different scenarios. Essentially, you've got the power to just do more simulations, we'll talk a little bit about how this works for JBA. To me, that is the incredible sort of, it's like the step up, if you like, this is a dramatic shift, is that ability to run many more scenarios. Before I come to you, David, you mentioned it in passing about the Nvidia as recessed by. for those of us who've been watching the video and just seen the extraordinary growth of the company, maybe regret not investing earlier. You also mentioned the company is quite small, relatively speaking, to sort of the position in the world, and it'd be great just to understand a little bit more about that concept, about how that business has grown, and the sort of philosophy about, you mentioned a little bit about developer first, but about enabling your, I think, enabling your clients, rather than necessarily sort of building your own infrastructure and sort of absorb the cost yourself or your building thinks to enable your clients in turn creates more uses in the video. That's the simplification of what you're doing. Yeah, because we just focus on what we are really good at, and that is building excellent hardware, building the software frameworks to leverage that hardware, put the easy button to it, and then approach a huge network and ecosystem of partners and collaborators that work with us. There are millions of developers around the globe, leveraging CUDA, and I mentioned in the beginning of our partner network of cloud service providers, computer makers, and also software vendors. So part of our go-to-market in general is with collaborators and partners. So we are in search for who is the leader in a certain industry, how does this ecosystem look like, and then work with those software vendors, for example, that are specialists in what they are doing, and they leverage our full stack platform, which is the hardware software layers. I talked about it like IRF2 Studio and these open source libraries, and JBA was an ideal candidate to approach, because they were the leader in the field, they are the leader in the field, and they were very open to innovation, trying and testing those systems, and that's why I think we have such an important collaboration, and that's why you were on stage, because you were really showing to the world at GTC this largest AI industry conference on the globe, showing people how to leverage accelerated computing and to revolutionize climate and weather modeling really. So that was a pretty good tagline, you might even use it for the podcast, sharing the world how to revolutionize modeling. So we've talked a little bit about it, but can you just give some more examples about what does this mean for your JBA or catastrophe modeling in general now, you've got access to IRF2 and the video. Yeah, sure, I'll talk a little bit about the proof of concept study that we did within video, and that was what we talked about at GTC as well. And so with the higher throughput that we get from these models, we were able to, well, there were two things that were presented to us. One was long rollout, so long simulations. So we were able to run seasonal simulations. And the other thing was the huge ensembles, because of the throughput of these approaches, we were able to run ensembles of a thousand members, which really gets you into the realm of understanding the probabilities and the probability distributions. So for the proof of concept study, we focused on last winter, 2324. We ran the global models and ran global simulations, but we focused on the Elb Valley in Germany for the purpose of our analysis. And so we built up these thousand seasonal simulations and used those to derive an event set. And what we saw from that is a few things. Firstly, it worked well. We were getting sensible levels of rainfall, both on an individual daily level, and also looking at the full season as well. And the distribution that we were getting in terms of the extreme events was actually really good. You know, we were seeing more extreme events that occurred than in the actual season, which was pretty wet. And we were able to run those into the catastrophe model and build up a picture of how the population could have been affected. And the answer is that the most extreme, it was sort of three times more of the population was affected than in the actual season. So yeah, we've kind of reassured ourselves that there's some real value there. And we can really start to enhance the events that's to go into the catastrophe models. So in terms of some of the advantages that we can get out of these approaches, well firstly, we've got a global simulation. So these things run at 25 kilometer grid pretty much. And what we're doing is we're simulating weather parameters. And from that, we extracted precipitation, but there's no reason it could just be precipitation. We can extract from those global weather simulations of range of different perils. It could be anything from wind to hail to potentially even things like drought as well. But on top of that, with some of the ecosystem that's provided by Earth too, we've got the opportunity of using downscaling models so that you can get down to real local scales. And whilst some of these techniques are still being developed, they're still being prototyped. What we're seeing is considerable progress over time. So what's not quite possible yet, give it another year and I expect it will be too. So what we've got is essentially this global weather engine that allows us to pull out all sorts of different perils, different variables that are of interest, catastrophe modeling. We've got the ability to look at local scale as well. And we can also use this approach because it's so fast in real time as well. So we can do counterfactual analyses of events that are playing out and understand the range of possibilities that might happen and feed those into our models as well. The last one is interesting. I mean, there was all interesting, sorry, but so you're saying that if you've got an event evolving, you can start to give people a sense as to how that might continue what the sort of increased flooding could be from a particular event. It's sort of almost not quite real time, but sort of. Well, that's it. These are kind of future ideas. It's not things that we're working on just yet, but there's certainly the kind of things that you can see this moving into in the future. So as the weather changes, in principle, we can feed the current weather conditions into these models, run large ensembles, understand what can happen at the tails as well as what's most likely to happen and that can give you real insights into how events might play out. Well, just as we kind of get towards the end, it's a day for one for you and some of the questions you're going to come back to you on, but so something that is interesting and learning more about how they can get access to what you're offering, what's the best way to learn more about JBA or contact you directly to get some examples. Well, in terms of understanding the kind of work we're doing, I'll point people to the blogs that we've got on the website. We've published a blog on our website. There's also one on MV as website as well and it was written by a couple of staff members John Ashcroft and Alison Paulson, and that's a really nice summary of the work that we've been doing. So people with a general interest to dig into a bit more of the details can look there. But in terms of getting in contact with us, if you go to the website, there's a help at JBA risk email address that people can use and they'll be sent through to the relevant people. So that's a good way of getting in contact. But we also try and make sure we've got a good presence at the various conferences. So driving is at the conferences that we attend is a good way of getting in contact too. Good, including us. Thank you for your support for that as well. You're confused for people that now want to get access to the video, whether they're insurance or somewhere else, what's the best way to do that? Yeah, we have dedicated websites for Earth2 Studio and the platform itself. We have on our website build.nvd.com, we have some of these open source models in place, inference, microservice platforms there. Because it's, as I mentioned before, it's not only the training, but also the inferencing. So running these models very efficiently, they're generating thousands of examples. So we demonstrate how that works. We also have the recordings of the JBA talks there. And I think that's really a very good educational piece because we really focus on creating additional business value in innovative areas where nobody has been before. So JBA and NVIDIA, we are pioneers in revolutionizing this atmospheric modeling industry. And it's very important when you do this to really show the business value. And that's what we did in the extensive studies together, like our experts sitting with together with the JBA experts for a couple of weeks, even months, to really show what's the quantifiable business value and our material really shows how we did that. Because acceleration per se, that's nice to have, but how does it translate into value? For example, you see or you can generate more scenarios on samples that you haven't observed before. So you can better prepare for the future floods, for example, yeah, or other parades, like better prepare for hurricane seasons. And also this downscaling is important. So you can have those models, but at a much better granularity because then it becomes really relevant for the insurance and banking industry who are sort of invested and they provide a protection for assets, like, for example, in an airport or an AI data center, which is very expensive. So leveraging those models, you have a better idea of measuring that risk. And that's about to hit so that the insurance market, the insurance market, and also the banking sector, for example, marked to market as a back securities, as backed by buildings, for example, and where you need to sort of recalculate the physical risk like every day, potentially. And I'm not sure if we currently really doing that properly in the industry. So with these kind of models, we will see and improved risk modeling capabilities. And that's where we sort of were keen on documenting this, how we work with JBA, and that we could prove the quantifiable business value and demonstrate to people how to leverage those modern know-how and plus data and leverage the compute to really come to new levels of an accuracy of modeling. There'll be a very willing audience amongst the risk managers responsible for these large asset portfolios, more as the actual building owners as opposed to the investors, who are also interested, but there's an increasing expansion of those organizations wanting to understand their own exposure to risk, which I can see that makes complete sense. We'll come to the future in a minute, but before we go there, just a question for both of you is anything that we haven't spoken about that you think is important to add. So Dave, anything we haven't covered that you would want to? Yeah, I'm a very strong believer in being sort of client and user-led as we approach these new development. It's easy to get excited about what's possible with these new technologies, but we really need to make sure that these are sort of focused and designed on the actual needs of the people using the models. So I guess there's a general invitation and request for anyone who's interested in this kind of science and technology to get in touch, because we really value the input in shaping how this works its way into our production catastrophe models. I know in your new role, because your previous role was more in looking, you're now, you're sort of, if we call it a roadshow, meet Dave, you're open to anybody who wants to meet you face to face. Yeah, that's it. I'm getting myself out there and visiting various clients at the moment, so just get in touch and I'm happy to come and visit people. And Jochen, for you, anything we haven't covered that you think is important, people to know about in a video? No, I think we pretty much explained everything why we work together, how we work together, and I just want to use the time here to express gratitude and say thank you to the JBA team and to Dave and people who put trust into our innovative technology and we're willing to test it, to explore it. It was quite a journey and now we are sort of at the forefront together and that was quite an interesting time. It took us some months to get there and I'm really happy that we are here and can discuss these wonderful applications and with the quantifiable business value, yeah, so it was an amazing journey. To echo that as well, I think the support that we've had from the NVIDIA team in terms of taking what you've got and turning it into the products has been fantastic, it's been a good working exercise. Well, I mean, I'm quite selective when I use the word exciting, but it's sitting here listening to the two of you and hearing the story and having followed a bit of it. It is incredibly exciting. As I mentioned, we've heard in a video in different contexts, but actually I think for anybody involved in risk assessment and insurance, to know this is a kind of next, you know, beyond the next frontier, actually, of what's possible is really, really encouraging and you know, sudden extent, you know, because it's we're all collaborative and you know, allows other people to come in and look at other perils and things that previously would have been too difficult or too expensive to model as well. Well, this is absolutely fascinating and really enjoyed that. It's kind of also slightly relevant that we're sitting here in the studio in just near to the Silicon Roundabout and Elf Street, but we've all got jobs to do, conferences to go to, so I'll let you both go. But again, thank you very much for joining us. Thanks very much for having us. Thank you. Well, if you've made it this far, then I'm pretty sure you found that as interesting as I did. The Instech Podcast comes out every Sunday morning, where we spot like the latest news, leasing voices and freshest updates across insurance that you need to know about. If you would like to take part in these conversations, head to www.instech.co to find out how you can join our network and be a part of the insurance intelligence for the curious.