---

Transcript

This transcript was generated automatically and may contain errors, especially in names. Click any timestamp to jump to that moment in the audio.

0:01Who are the characters behind science? This is Science Behind the Glass, illuminating the human side. In this podcast, we unveil the human side of laboratories and science. Scientists share personal experiences and we discuss their meanings from different perspectives and fields of research. Welcome to Science Behind the Glass. Thank you very much for joining us. It’s a pleasure having you here. So if you want, can you please introduce yourself? Sure. Thank you, Linda. And nice to meet you, Rafa. It’s my pleasure being here.

0:38So I’m Jonathan. I’m a principal AI applied scientist at Inception, a company based in Abu Dhabi. I’m an artificial intelligence and machine learning scientist. I did my PhD in AI and machine learning. My research was focused on stochastic decision -making and probabilistic reasoning, or probabilistic reasoning and uncertainty. And I’ve been developing artificial intelligence solutions and machine learning solutions for conservation decisions, environmental management, climate, weather, transport supply chains, transport resilience. Um, so afterwards, after my research, uh, research, um, say years, I transitioned to a more applied, um, sciences domain. And this is what I’m doing here in, in Abu Dhabi. And that sounds really exciting. Big jump. So you’re from Spain and then how’d you end up working first in CSIRO in Australia and then jumping to Abu Dhabi? What was the transition about? Yeah, so I did my PhD in Spain and Australia.

1:52I spent almost eight years in Australia, I believe, seven, eight years in Australia, and then I moved to Abu Dhabi. Yeah, I believe it’s been a little bit of everything. So when I was in my PhD in Spain and I got to know Australia, I believe… fall in love with that place, with the nature and the opportunities that Australia was offering in the domain of research and academic environment as well. I was really motivated by developing AI for conservation and environmental management, and I believe Australia is a great place for doing this. And then how we went from Australia to Abu Dhabi to the Middle East. well one of the reasons is Abu Dhabi is I believe is in exactly between Australia and Spain so it’s easy to go from one place to another the UAE is I believe it’s a new technological hub so there is a lot of things going on in artificial intelligence and machine learning here from a research perspective and from artificially from a research perspective and from applied sciences as well so I I was doing research in Australia and I started in academia in the University of Queensland when I did my first postdoc in Australia. After finishing my PhD, my first postdoc was at the University of Queensland in Australia. I was working by

3:22then with robotic arms that could pick up small objects and could manipulate them. And after that, I worked at CSIRO. First as a postdoc and then as a research scientist. And it was there, it was there in Cicero when I started developing more interpretable AI solutions for conservation decisions, for environmental management, health, life security. So I work a lot on decision making and uncertainty and probabilistic reasoning.

3:52So that’s where that transition already happened from research to applied sciences domain. I love that we were developing solutions and conducting research at the same time that was actually used by government and by industry in general, especially by government. And I think that’s the right way to go. I want to do research, but I also want to develop solutions that are useful at the end and that can provide some benefits, some real benefits to individuals and institutions. And I believe that’s… how this natural transition from researchers to applied scientists happened.

4:29And that’s what made me move from Australia to Abu Dhabi to the Middle East. And right now, without leaving this research component behind, I’m working as an applied scientist. I believe that this is how we’re leveraging these new research outputs and these new findings to improve the solutions that we’re developing. So that’s what made me move all around these places. these kind of opportunities. It’s quite interesting, Jonathan.

4:59I guess like AI right now, it’s like really in the limelight, right? There’s a lot of money being poured into it, right? From venture capitalists and so forth. Is that giving you some freedom of like pursuing whatever you imagine of to do with your work right now? You have a lot of freedom because there’s capital coming at you from all avenues, is that? How do you… Yeah, because it’s a huge… Some people say that it’s like a bubble that’s creating because of the amount of money that’s going into it. But it must create a lot of freedom for you to be able to adventure whatever you can imagine of, right? Yeah, so I believe there’s a lot of investment in artificial intelligence, a lot of money and resources that are put towards… not only developing research, but also developing applications. That gives us enough freedom to create and to develop new solutions. Regarding if it’s a bubble or not, depending on whether this money will stop coming in those big amounts, the things that we are doing right now, the things that the community, the AI community is doing, I strongly believe that they’re going to stay here anyways. These are things, genetic AI, language models, vision models, everything. These are things that are going to stay here independently now of what’s

6:31the funding in the next following years, which I believe that this funding is going to stay here as well. I don’t think we’ll stop having the opportunity of developing new research and developing new solutions. Just on that note, I’m more interested to see, like, perhaps before this whole hype of AI, have you seen a transition in giving you freedom in the way that you approach your work now in a way that there is no barriers of how you can really approach your work because there is no limitations of financial aspect, perhaps, of it? Yeah, so I believe. Probably before all this hype and before all these slice language models and this gen -initive AI new research directions, probably one of the biggest barriers for developing artificial intelligence was the infrastructure before. So these models that now we’re training, these large language models or these vision models require a fair amount of computing. That was probably when I started my PhD. Those things were not there as well, right? We had cloud computers, we had supercomputers as well, but they were not as accessible as we have right now. We have these Azure services, these Amazon Web Services, all these new, powerful and efficient GPUs that are being developed by

8:02some companies like NVIDIA, which these things are facilitating this computing power and these resources. So I remember before I was doing my PhD, so we were developing our algorithms and training our algorithms in our work. own computers, in our own laptops, in our own personal computers. We were having some access as well to some computing, but it was not the same as before as now. Anyways, yes, I think that given all the opportunities that artificial intelligence has and all the new research directions and potential research directions, I think that the AI community has a fair amount of freedom in exploring and developing new solutions. Can you give us a little, I know that you have been explaining all these AI models and you use it also for forecasting, but can you give us a little bit of background how this AI, generative AI, comparing to the previous ones, is changing the complex problems in science and everyday life? Why has it been such a big…

9:14Maybe I can focus this on the, especially on the weather and climate domain, which is… It sounds weird how genetic AI can predict weather and climate at those fine scales, right? So, for example, when we were having more traditional weather forecasting models, these were models that relies on large numerical and physics -based models that We’re actually solving physical equations in a thick script, for example. Very accurate, which is very accurate, especially at larger scales. So by larger scales, I mean by lower resolutions. But it’s extremely computational expensive. So especially when you want high resolution, when you want to predict particular climate events or extreme events at a very fine scale, right?

10:05So because of that, usually many local details with more traditional models are… either smooth or not resolved at all. So that’s when Genetic AI came into the picture. So these new Genetic AI models are built on top of these more traditional models. While we’re keeping these traditional models for large physics skills, for example, we or the community found that we can use Genetic AI to add to create, let’s say, because this is what genetic AI is doing, right? More realistic and fine scale details on weather and climate.

10:50So this is what allows us to produce more high resolution forecasts, much faster, and especially what we were discussing before about competing in power at a fraction of the computational cost. while we’re still remaining consistent with underlying physics, right? So in short, it’s not that generative AI has been replacing traditional models or traditional forecasting, but it’s enhancing them. So it’s making little predictions and more actionable predictions, right? So generative AI in this setting, in this climate models, work in a way that is very similar on this. diffusion models generate images, right? So if you have seen this stable effusions where we can type something like, I don’t know, create a picture of a dog walking on the moon or walking on Mars or dancing on top of Mars, and the model actually generates that, right? So in image generation, how these things are done is a model starts from a noisy version of an image and progressively refines it until it becomes sharp and realistic, right?

11:58So in our case for climate and for weather, we do something very similar. We start from coarse weather, from a coarse climate resolution forecast, which captures these large details at the big picture, but miss the local details. And this is what we’re using in Genetic AI to fill these local details, right? We use what is called the correction diffusion process that progressively adds fine scale structures and this model learns how high resolution weather looks given low resolution weather. And this is how we’re generating more realistic, more physically consistent details such as precise temperature gradients and wind patterns that can be missed by these low resolution models at a much finer scale. So let’s say that instead of predicting the weather from scratch, what the generative models do, they refine and correct forecasts that already exist, but converting them into high -resolution models that learn the small -scale physics of weather and climate.

13:06And that allows us to generate data, to generate data at the end, because we’re generating data. That turns into more actionable… climate information that can be used for real world decisions at a fine scale super super super fascinating i can go into all that but now i want to continue on what you just said about like uh um you’re directly to public decision making can we go a little bit into that like what are opportunities and challenges of working with governments to implement ai in areas like climate adaptation disaster response and or infrastructure planning um We’re going to the technicality a little bit later. I’m interested in that. Yeah. Yeah, so I believe that working with governments is the right way to go, and it offers us a unique opportunity to scale AI solutions, especially where they can have the greatest impact. uh for society right so the governments at the end they operate at uh at city level uh city level region level um country level as well so for example i can support this climate adaptation uh i can support disaster preparedness uh it can support infrastructure

14:36planning in a very concrete and meaningful way in an actionable way working with governments, right? So the key opportunity in general is collaboration. It’s industry, governments, research institution, education institution. This is where we need to combine all the scientific innovations with the policy, with the policy, with the local knowledge of the country as well, with the long -term planning. This is where this alignment goes well. So at the end, AI becomes a powerful tool to support these decisions. Not a tool… that is taking the decisions for you, but the tool that is supporting these decisions, that is improving the resilience.

15:14And at the end, the goal, especially for people who are developing artificial intelligence for sustainability and for real -world impact, the ultimate goal is to protect communities, to help communities, right? So these partnerships are very important. These partnerships are what allow us to move from… from experimentation to the real -world implementation. So it’s crucial for that management and development. And that leads us to another question. How these data centers and AI deployments are evolving with sustainability and the energy efficiency goals? Because we know that there’s a lot of impacts related with AI, but how this is aligning together? Yeah. So I think that when we talk about cost, of artificial intelligence, it’s important to look at it in a balanced way. There’s a computational cost and an energy cost of training and running these models. But at the same time, there is also the value that they generate, right? So, for example, in areas like climate, disaster management, conservation, artificial intelligence can significantly reduce costs by improving the efficiency. that we’re managing resources, that we’re optimizing resources, it can enable better planning. It can help better prevent costly impacts before they happen, right?

16:37So the conversation, I think, when we’re talking about cost isn’t just about costing isolation, but how to use AI responsibly to maximize this positive impact, right? So, for example, in terms of energy and resources, When we’re training the largest scale climate models, these models are big. It’s huge. Training generative AI models is huge, right? So it does require significant computational resources, yes. Especially during the training phase. We have large amounts of data and the computer we need to use to learn these complex patterns is quite substantial, right? However, this is typically a one -time or infrequent cost. So once these models are trained, running them, what we call inference, is much, much faster and far less energy -intensive than the classical physical approaches. So, for example, in weather, when we’re predicting weather, the traditional methods are the physical space methods. These are numerical methods, equations that are thrown into a computer, and they’re slow. They’re very computational demanding. They take time. They take time to generate these climate models, these climate predictions, these weather predictions at a low -resolution scale. Once you have trained these models, which performance is very close to these traditional models, these traditional physics -based models, what it was taking hours or days to generate with the classical models, and it was taking lots of computational resources, when the genetic AI models are trained, it takes seconds, it takes minutes, at a small

18:09fraction, small, small fraction as a cost, right? So at the end, this inference part is much faster. It’s… far less energy intensive than the classical physical approaches. So that’s the balance. We invest in training these models, but then we have solutions that are efficient, are fast, that they can run in just in a laptop. So this cloud computing, these amazing infrastructures that you need to train these models, you can use them for… Once. For training these models once, you can run these models in your laptop at a fraction of the cost. Very efficient. You can run ensembles of models in seconds, in minutes.

18:56That’s really cool because I’ve been playing a lot with agents. I’m running on a VPS that’s going through an EPI with OpenAI Codex and I’m having it run on WhatsApp. So I’m just communicating with it and it’s just running pretty much on a really, really low power. VPS shared server, right? So that’s the beauty that I go back to my first question. I think the amount of money that’s being poured into AI, we’re seeing this really rapid growth to development and progress.

19:29Like pretty much every day, I feel like there’s something exciting that’s happened. So that must be really, really fascinating for you because it drives you like, ah. Like your imagination can run wild on that level. But do you feel like we’re going the right direction? Because there’s even talk of building data centers. Are we ethically progressing the right way? Not like for development, like there shouldn’t be a frontier for it, but are we doing it ethically? What would be your point of view on how we should scale? Thinking about the planet, thinking about the consequences of this. Yeah. So in my opinion, from my perspective, the key is to focus on efficiency, is to focus on efficiency, is to focus on transparency, and especially what I mentioned before, which is the collaboration, the collaboration between institutions and countries, right? So this is crucial. So what we need to do, we need to design AI systems, we need to design data centers. to be as energy efficient as possible. We need to adopt these best practices for sustainable computation and we also need to integrate renewable energy whenever we can, right? And this is something that at the same time is happening. So a stronger collaboration between researchers, industry, policy makers, this is what can help artificial intelligence ensure that we are deploying artificial

21:07intelligence, like we’re deploying responsibly and at a scale, right? I think we’re making good progress in several areas. AI models are becoming more and more efficient. And this is one of the lines of research, optimizing models, visualizing models, making them efficient and faster. At the same time, we’re adopting renewable energy. For example, here in the UAE, we have a lot of sun. There is a good amount of solar fans in the UAE as well. Researchers as well are increasingly focused on reducing this kind of footprint on this training phase, on this in -frame phase.

21:45I believe also there is a growing awareness of sustainability in industry and in academia. And at the same time, we need to continue innovating in efficiency, right? We need to continue innovating in model optimization. We need to develop. previous architectures. We need to build the specialized models. We need to containerize, optimize the containerization of solutions. We need to have this energy -aware deployment. So at the end, the goal is to ensure that AI is scaling, it remains sustainable, and it delivers a maximum benefit from a societal and environmental point of view, not only economic, while keeping this energy use under control. So I think we’re getting there. Yeah, I think that something that you point out is important and it’s like the balances, the cost of the balances of using these models, generative AI. For example, there’s a lot in conservation is being used a lot. We have a guest that works for the ocean cleanup and they use it to identify microplastics in the ocean, which is housing heaps. identify this actually an animal or or actually uh microplastics so the identification that one is just reduce the amount of resources that i have to put on that has particular tasks but something that you also mentioned is like i think that one way to go is being is just transparency

23:16of the data centers being built i was reading recently that one in and i think that was switzerland but it’s more like a focus on languages the ai that a that they are building is a 100 like a running by a renewal energies and also what they are doing is just putting all the data there and being careful with the privacy of the data and how they are feeding their models So yeah, I think that I agree with you. There has to be a collaboration, a huge collaboration, but also transparency and also how they implement these models. I actually want to take a step back.

24:00Can we just break down a little bit? because you were talking about image models for your work, for weather prediction and stuff, because you were talking about stable diffusions. Can you run me through the different steps so I get a better understanding of what the different steps of how you train your data? Because it seems like you’re using image representation as well, numerical data that you’re training with, just so I can get a better understanding.

24:28So for training these weather and climate models, What we’re doing is we’re using generative AI to downscale weather predictions, right? So we take this input low resolution weather forecast and we take this output high resolution weather forecast and we train a generative AI model that maps this low resolution to high resolution, right? This is where the training process happens and collecting this data can come from two different sources. Usually low resolution weather… outputs is publicly available from ECM, WF, from ETA5.

25:04There are several sources that provide the high quality analysis data sets at low resolution. High resolution weather outputs is another thing. So this is obtaining high resolution weather outputs. We need either institutions or meteorological agencies that actually make this data public. This data is expensive to get. It’s expensive to gather. It requires a lot of domain expertise. And this is what’s taken as the kind of ground truth. So as the golden data set, these physics -based models, we know that they work, we know that they work well. There’s a lot of research happening there from Earth atmospheric scientists and climate scientists, weather scientists in general. So once we invest into these high -resolution data sets, we can… actually train the genetic models with just a few years of data, four or five years of data. We started a really nice partnership with NVIDIA in 2024 to leverage their technology, their models, and their expertise as well to train these models for DOE and from other regions. So the way that these models work, it’s very similar in how these diffusion models generate images. We take an image, which at the end is a weather forecast. If you see weather forecast, if you see climate forecast, it’s a grid, it’s a picture, it’s an image with pixels.

26:36Each pixel represents an area. Usually for low resolution, this pixel represents 25 kilometers per 25 kilometers. At high resolution, it can represent models at two kilometers of resolution, where each one of these pixels is two kilometers per two kilometers, and even 200 meters of resolution for the city of Abu Dhabi, which is 200 meters for 200 meters, right? So the model will start from a very noisy version of that image. So with an image that is noise, pure noise. And it progressively refines it until it becomes a sharp and realistic image. And the sharp and realistic image is that realistic that it’s so close to a real weather forecast, which is amazing. So we are generating an image that it cannot be distinguished from a real image. And that real image represents a weather pattern, a climate pattern, a climate forecast. So it’s impressive. We take. something that is a very low resolution and we convert it into something that represents what will be the temperature, what will be the wind patterns in that building especially. or in that particular location, considering lots of additional inputs and invariants, what is geographical properties, land sea mass, land use, altitudes, and things like this, right? So that’s especially useful to capture the small -scale physics and to detect extreme events, to detect extreme events that

28:08are missing the low -resolution outputs that represent 25 kilometers per 25 kilometers. What do we have, for example, what’s the temperature in a city? in general, what’s the wind pattern in a city instead of what’s the temperature, what’s the wind pattern in some particular areas of the city, in some mountainous areas, in some coastal zones. And we can do this. These models are also stochastic. These models are probabilistic. So what it means is that we are able also to measure uncertainty. We don’t have just one prediction of what will be the temperature here, but we have what will be the potential range of temperatures and how this potential range will affect the other variables as well. So if we think about this, Genentech AI in general, it looks sometimes like a science fiction, right? Because at the end, we’re creating things that feel almost lifelike or human, right? So from images, from text, from music, from simulation of complex systems. And at the end, we’re simulating a complex system.

29:05We’re simulating a complex system that it’s such, such, such real. It’s that close to reality that it’s almost reality. So what’s happening at the end behind the scenes is not magic. It’s pure mathematics, pure statistics. We have models that learn patterns from huge amounts of data and they generate new content that follow these patterns, that follow the patterns that they have learned. And this actually shows up in many areas. So it can help generate ideas for products.

29:35It can help educators, I don’t know, to create interactive learning approaches, personalized education. It can simulate traffic flows based on what the model has learned for cities it can optimize supply chain by generating these potential scenarios like what will happen if these things are happening in the future what will be the consequences of this scenario planning is a very interesting um approach for this right uh output for this so the key point is that this uh these models like with amplifies the problem solving at the end um and it turns in my opinion it turns what was taking month or years to generate into something that is taking minutes, which is why it feels so impressive and is attracting that attention. Yeah, and that’s very interesting because usually, for example, in fields like agriculture, the farmers have a lot of issues predicting the weather and the crops, and they have a lot of losses because of the weather.

30:40This kind of models can help. with the like a mitigate those impacts environmental impacts and weather impacts that can have in a crop on so provide suggestions to the farmers and how to grow and grow in a better way their crops yeah absolutely yeah so and that’s that’s one of the targets and that’s one of the goals and applications that we like um so we we mentioned that this uh traditional physics -based models which is what’s traditionally used by meteorological agencies We know that they are good. We know that they are real. This is what is being used and they are good and they require a lot of domain knowledge, a lot of domain expertise, but they are expensive, right? So not everyone, not everywhere we have these computational resources to run these physics -based models, right? And I believe that this is where this kind of genetic AI models that have been getting implemented for weather and climate can play a key role. um so we mentioned that it takes some cost to train them but then inference is cheap anyone with a laptop can run as models instead of seconds it might take some minutes but it it will happen right and especially for cases like this for agriculture this is where we can have very localized weather forecasts and this is where we have good chances of detecting extreme climate events at particular locations, at the specific locations that we miss with more

32:12coarse base models. We can also generate with these genetic models, we can generate many scenarios, we can generate an ensemble of forecasts, we can evaluate uncertainty or how a climate event might happen and can affect a potential crop or a future crop, right? So this is how these models, this is how genetic AI for this kind of topics for weather and climate can play a key role, can help farmers, can help cities planning for disaster management, can detect heat hotspots as well. This is an interesting topic. So we can detect areas of a city where heat hotspots might happen as well, might affect population, might stress out the infrastructure, things like this. More and more agencies and institutions interested at the end in in developing artificial intelligence for these solutions, for these domains.

33:11I believe this is something that cannot be avoided mentally. AI is happening everywhere, but it’s important also to say that AI is not getting there to replace this knowledge and these traditional methods. So we will always need them. We’ll always need this domain expertise. We’ll always need how to run these physics -based models, these traditional models, because at the end, this is what is generated in the data that we’re using to train our models, right? So I believe that both things can come together. We need these traditional physics -based models by methodological agencies to train our generative models. We need these physics -based models to act as a counter to evaluate our generative models. We need these physics -based models to actually check if the predictions that we believe that are more possible and feasible for these generative models are correct or not, or they deviate too much from the domain knowledge. So both things have to come together. At a personal level, do you use AI? And how can it be affected? And how do you modulate it with your family? And how do you control it? A lot of people are moderating it that much, using it a lot. So I believe everyone is using now ChatGPT for lots of things. But also this is becoming a lot in the developer life and at the industry level. So we have now not only ChatGPT, but we have Microsoft Copilot, we have… Coursera for assisting developers,

34:43going faster at the end, right? So there is a lot of talks about this pipe coding, whether the pipe coding is good or not. I think that, as I said before, it’s a support tool. So it’s something that is here to support you, to support decision -making, not to take decisions for you. Not to do things for you, but to support you. It’s a companion. It’s another tool. For example, it may be very useful, right? So there’s something saying, lots of emails, lots of messages from teams and so on. Compile it, for example. It’s great. Can you please check me if I have any attachment from that person regarding that topic?

35:19And they will find it for me. And that’s great. It makes my life easy. It makes our life easy at the end. I’m not saying that I’m going to rely and live all my life for these things like, hey, tell me what I need to do tomorrow or anything like that. But it’s a tool that’s here to support us. And I think it’s here to stay. Yes. Fully embracing the AI. Probably. And if you have a message for our public, what would you, what would you message? This community plays a key role on developing new solutions. on facilitating this understanding as well to the general domain.

36:00And these things have to come together, research, industry, public domain. These tools are very useful. And as I was saying, they are support tools. They are tools that are here to guide decision -making to help us in our everyday tasks, not just to take control of our for life, but to help and make our lives better. Of course, there must be some awareness. At the end, all these genetic AI tools can be used for things that maybe are not good, like genetic images, which violate any policy or any rules. It can generate fake data. So I’m not saying everything is beautiful and everybody is using AI for good.

36:52But this is the… direction where we need to move where we’re moving which is developing artificial intelligence uh for good for a human society at human ai level that’s that’s that’s the point developing uh human ai level um artificial intelligence yeah thank you yeah thank you thank you for having me yeah We make Science Behind the Glass because we want to show you the people who are changing the world. Researchers from every field, social leaders, educators, and those working in politics, economics, and conservation to understand reality as a whole. Because science doesn’t live in isolation.

37:37It is the sum of all its parts. We are a bilingual podcast, and now we have a blog too, breaking down language barriers to bring knowledge to every corner of the world. We dream of a world where anyone, anywhere can connect with these ideas in a simple, equitable way and feel the same passion they would feel watching a soccer match. We do all this out of our own pockets and that’s why we need your support. To keep this platform alive and to keep giving a voice to the people working behind the scenes and who make this a better world, we need your support and with a donation at sciencebehindtheglass.com you can give us a hand. Every contribution counts and keeps this mission alive. And remember, you can listen to us in all audio platforms.

38:25This episode was made possible by Daniel Esguerra and me, Linda Hernandez. Original music by our colleague and friend, Bernardo Aguila. And with a super powerful help behind the scenes and sometimes as a host, Rafaeli Jordan. See you next time.