Latest / Science Behind the Glass: illuminating the human side / Can AI Predict the Storm Before It Forms? With Dr Jonathan Ferrer
Transcript
- 0:01Who are the characters behind science? This is
- 0:04Science Behind the Glass, illuminating the human
- 0:07side. In this podcast, we unveil the human side
- 0:11of laboratories and science. Scientists share
- 0:14personal experiences and we discuss their meanings
- 0:17from different perspectives and fields of research.
- 0:21Welcome to Science Behind the Glass. Thank you
- 0:24very much for joining us. It's a pleasure having
- 0:27you here. So if you want, can you please introduce
- 0:31yourself? Sure. Thank you, Linda. And nice to
- 0:35meet you, Rafa. It's my pleasure being here.
- 0:38So I'm Jonathan. I'm a principal AI applied scientist
- 0:42at Inception, a company based in Abu Dhabi. I'm
- 0:46an artificial intelligence and machine learning
- 0:48scientist. I did my PhD in AI and machine learning.
- 0:54My research was focused on stochastic decision
- 0:57-making and probabilistic reasoning, or probabilistic
- 1:00reasoning and uncertainty. And I've been developing
- 1:04artificial intelligence solutions and machine
- 1:06learning solutions for conservation decisions,
- 1:09environmental management, climate, weather, transport
- 1:13supply chains, transport resilience. Um, so afterwards,
- 1:17after my research, uh, research, um, say years,
- 1:22I transitioned to a more applied, um, sciences
- 1:26domain. And this is what I'm doing here in, in
- 1:29Abu Dhabi. And that sounds really exciting. Big
- 1:33jump. So you're from Spain and then how'd you
- 1:36end up working first in CSIRO in Australia and
- 1:40then jumping to Abu Dhabi? What was the transition
- 1:43about? Yeah, so I did my PhD in Spain and Australia.
- 1:52I spent almost eight years in Australia, I believe,
- 1:55seven, eight years in Australia, and then I moved
- 1:57to Abu Dhabi. Yeah, I believe it's been a little
- 2:00bit of everything. So when I was in my PhD in
- 2:05Spain and I got to know Australia, I believe...
- 2:08fall in love with that place, with the nature
- 2:11and the opportunities that Australia was offering
- 2:14in the domain of research and academic environment
- 2:18as well. I was really motivated by developing
- 2:23AI for conservation and environmental management,
- 2:27and I believe Australia is a great place for
- 2:29doing this. And then how we went from Australia
- 2:33to Abu Dhabi to the Middle East. well one of
- 2:37the reasons is Abu Dhabi is I believe is in exactly
- 2:39between Australia and Spain so it's easy to go
- 2:42from one place to another the UAE is I believe
- 2:48it's a new technological hub so there is a lot
- 2:51of things going on in artificial intelligence
- 2:53and machine learning here from a research perspective
- 2:56and from artificially from a research perspective
- 3:00and from applied sciences as well so I I was
- 3:06doing research in Australia and I started in
- 3:11academia in the University of Queensland when
- 3:15I did my first postdoc in Australia. After finishing
- 3:18my PhD, my first postdoc was at the University
- 3:20of Queensland in Australia. I was working by
- 3:22then with robotic arms that could pick up small
- 3:25objects and could manipulate them. And after
- 3:29that, I worked at CSIRO. First as a postdoc and
- 3:32then as a research scientist. And it was there,
- 3:34it was there in Cicero when I started developing
- 3:37more interpretable AI solutions for conservation
- 3:42decisions, for environmental management, health,
- 3:46life security. So I work a lot on decision making
- 3:48and uncertainty and probabilistic reasoning.
- 3:52So that's where that transition already happened
- 3:55from research to applied sciences domain. I love
- 4:01that we were developing solutions and conducting
- 4:05research at the same time that was actually used
- 4:08by government and by industry in general, especially
- 4:11by government. And I think that's the right way
- 4:13to go. I want to do research, but I also want
- 4:16to develop solutions that are useful at the end
- 4:19and that can provide some benefits, some real
- 4:21benefits to individuals and institutions. And
- 4:25I believe that's... how this natural transition
- 4:27from researchers to applied scientists happened.
- 4:29And that's what made me move from Australia to
- 4:31Abu Dhabi to the Middle East. And right now,
- 4:35without leaving this research component behind,
- 4:38I'm working as an applied scientist. I believe
- 4:41that this is how we're leveraging these new research
- 4:44outputs and these new findings to improve the
- 4:49solutions that we're developing. So that's what
- 4:52made me move all around these places. these kind
- 4:56of opportunities. It's quite interesting, Jonathan.
- 4:59I guess like AI right now, it's like really in
- 5:03the limelight, right? There's a lot of money
- 5:05being poured into it, right? From venture capitalists
- 5:08and so forth. Is that giving you some freedom
- 5:12of like pursuing whatever you imagine of to do
- 5:17with your work right now? You have a lot of freedom
- 5:20because there's capital coming at you from all
- 5:22avenues, is that? How do you... Yeah, because
- 5:26it's a huge... Some people say that it's like
- 5:30a bubble that's creating because of the amount
- 5:33of money that's going into it. But it must create
- 5:35a lot of freedom for you to be able to adventure
- 5:38whatever you can imagine of, right? Yeah, so
- 5:42I believe there's a lot of investment in artificial
- 5:45intelligence, a lot of money and resources that
- 5:49are put towards... not only developing research,
- 5:53but also developing applications. That gives
- 5:56us enough freedom to create and to develop new
- 6:01solutions. Regarding if it's a bubble or not,
- 6:05depending on whether this money will stop coming
- 6:09in those big amounts, the things that we are
- 6:13doing right now, the things that the community,
- 6:15the AI community is doing, I strongly believe
- 6:17that they're going to stay here anyways. These
- 6:22are things, genetic AI, language models, vision
- 6:27models, everything. These are things that are
- 6:28going to stay here independently now of what's
- 6:31the funding in the next following years, which
- 6:34I believe that this funding is going to stay
- 6:36here as well. I don't think we'll stop having
- 6:41the opportunity of developing new research and
- 6:44developing new solutions. Just on that note,
- 6:46I'm more interested to see, like, perhaps before
- 6:50this whole hype of AI, have you seen a transition
- 6:55in giving you freedom in the way that you approach
- 6:58your work now in a way that there is no barriers
- 7:01of how you can really approach your work because
- 7:06there is no limitations of financial aspect,
- 7:08perhaps, of it? Yeah, so I believe. Probably
- 7:13before all this hype and before all these slice
- 7:16language models and this gen -initive AI new
- 7:20research directions, probably one of the biggest
- 7:25barriers for developing artificial intelligence
- 7:28was the infrastructure before. So these models
- 7:31that now we're training, these large language
- 7:34models or these vision models require a fair
- 7:36amount of computing. That was probably when I
- 7:42started my PhD. Those things were not there as
- 7:46well, right? We had cloud computers, we had supercomputers
- 7:50as well, but they were not as accessible as we
- 7:52have right now. We have these Azure services,
- 7:55these Amazon Web Services, all these new, powerful
- 8:00and efficient GPUs that are being developed by
- 8:02some companies like NVIDIA, which these things
- 8:06are facilitating this computing power and these
- 8:08resources. So I remember before I was doing my
- 8:14PhD, so we were developing our algorithms and
- 8:21training our algorithms in our work. own computers,
- 8:24in our own laptops, in our own personal computers.
- 8:27We were having some access as well to some computing,
- 8:30but it was not the same as before as now. Anyways,
- 8:33yes, I think that given all the opportunities
- 8:35that artificial intelligence has and all the
- 8:38new research directions and potential research
- 8:39directions, I think that the AI community has
- 8:42a fair amount of freedom in exploring and developing
- 8:46new solutions. Can you give us a little, I know
- 8:53that you have been explaining all these AI models
- 8:56and you use it also for forecasting, but can
- 9:00you give us a little bit of background how this
- 9:02AI, generative AI, comparing to the previous
- 9:06ones, is changing the complex problems in science
- 9:10and everyday life? Why has it been such a big...
- 9:14Maybe I can focus this on the, especially on
- 9:17the weather and climate domain, which is... It
- 9:19sounds weird how genetic AI can predict weather
- 9:26and climate at those fine scales, right? So,
- 9:29for example, when we were having more traditional
- 9:32weather forecasting models, these were models
- 9:37that relies on large numerical and physics -based
- 9:40models that We're actually solving physical equations
- 9:44in a thick script, for example. Very accurate,
- 9:48which is very accurate, especially at larger
- 9:49scales. So by larger scales, I mean by lower
- 9:53resolutions. But it's extremely computational
- 9:55expensive. So especially when you want high resolution,
- 9:58when you want to predict particular climate events
- 10:00or extreme events at a very fine scale, right?
- 10:05So because of that, usually many local details
- 10:09with more traditional models are... either smooth
- 10:11or not resolved at all. So that's when Genetic
- 10:15AI came into the picture. So these new Genetic
- 10:20AI models are built on top of these more traditional
- 10:23models. While we're keeping these traditional
- 10:25models for large physics skills, for example,
- 10:29we or the community found that we can use Genetic
- 10:37AI to add to create, let's say, because this
- 10:41is what genetic AI is doing, right? More realistic
- 10:44and fine scale details on weather and climate.
- 10:50So this is what allows us to produce more high
- 10:53resolution forecasts, much faster, and especially
- 10:55what we were discussing before about competing
- 10:57in power at a fraction of the computational cost.
- 11:00while we're still remaining consistent with underlying
- 11:02physics, right? So in short, it's not that generative
- 11:06AI has been replacing traditional models or traditional
- 11:09forecasting, but it's enhancing them. So it's
- 11:11making little predictions and more actionable
- 11:15predictions, right? So generative AI in this
- 11:19setting, in this climate models, work in a way
- 11:22that is very similar on this. diffusion models
- 11:26generate images, right? So if you have seen this
- 11:28stable effusions where we can type something
- 11:31like, I don't know, create a picture of a dog
- 11:35walking on the moon or walking on Mars or dancing
- 11:40on top of Mars, and the model actually generates
- 11:42that, right? So in image generation, how these
- 11:48things are done is a model starts from a noisy
- 11:51version of an image and progressively refines
- 11:55it until it becomes sharp and realistic, right?
- 11:58So in our case for climate and for weather, we
- 12:01do something very similar. We start from coarse
- 12:03weather, from a coarse climate resolution forecast,
- 12:06which captures these large details at the big
- 12:10picture, but miss the local details. And this
- 12:13is what we're using in Genetic AI to fill these
- 12:15local details, right? We use what is called the
- 12:18correction diffusion process that progressively
- 12:22adds fine scale structures and this model learns
- 12:28how high resolution weather looks given low resolution
- 12:33weather. And this is how we're generating more
- 12:37realistic, more physically consistent details
- 12:40such as precise temperature gradients and wind
- 12:44patterns that can be missed by these low resolution
- 12:46models at a much finer scale. So let's say that
- 12:50instead of predicting the weather from scratch,
- 12:52what the generative models do, they refine and
- 12:57correct forecasts that already exist, but converting
- 13:00them into high -resolution models that learn
- 13:03the small -scale physics of weather and climate.
- 13:06And that allows us to generate data, to generate
- 13:10data at the end, because we're generating data.
- 13:14That turns into more actionable... climate information
- 13:20that can be used for real world decisions at
- 13:24a fine scale super super super fascinating i
- 13:27can go into all that but now i want to continue
- 13:29on what you just said about like uh um you're
- 13:33directly to public decision making can we go
- 13:35a little bit into that like what are opportunities
- 13:37and challenges of working with governments to
- 13:39implement ai in areas like climate adaptation
- 13:43disaster response and or infrastructure planning
- 13:46um We're going to the technicality a little bit
- 13:49later. I'm interested in that. Yeah. Yeah, so
- 13:55I believe that working with governments is the
- 14:00right way to go, and it offers us a unique opportunity
- 14:04to scale AI solutions, especially where they
- 14:08can have the greatest impact. uh for society
- 14:18right so the governments at the end they operate
- 14:20at uh at city level uh city level region level
- 14:24um country level as well so for example i can
- 14:29support this climate adaptation uh i can support
- 14:33disaster preparedness uh it can support infrastructure
- 14:36planning in a very concrete and meaningful way
- 14:39in an actionable way working with governments,
- 14:41right? So the key opportunity in general is collaboration.
- 14:44It's industry, governments, research institution,
- 14:48education institution. This is where we need
- 14:51to combine all the scientific innovations with
- 14:53the policy, with the policy, with the local knowledge
- 14:55of the country as well, with the long -term planning.
- 14:58This is where this alignment goes well. So at
- 15:00the end, AI becomes a powerful tool to support
- 15:06these decisions. Not a tool... that is taking
- 15:09the decisions for you, but the tool that is supporting
- 15:11these decisions, that is improving the resilience.
- 15:14And at the end, the goal, especially for people
- 15:17who are developing artificial intelligence for
- 15:20sustainability and for real -world impact, the
- 15:23ultimate goal is to protect communities, to help
- 15:25communities, right? So these partnerships are
- 15:28very important. These partnerships are what allow
- 15:32us to move from... from experimentation to the
- 15:35real -world implementation. So it's crucial for
- 15:38that management and development. And that leads
- 15:40us to another question. How these data centers
- 15:43and AI deployments are evolving with sustainability
- 15:47and the energy efficiency goals? Because we know
- 15:50that there's a lot of impacts related with AI,
- 15:53but how this is aligning together? Yeah. So I
- 15:58think that when we talk about cost, of artificial
- 16:01intelligence, it's important to look at it in
- 16:04a balanced way. There's a computational cost
- 16:06and an energy cost of training and running these
- 16:09models. But at the same time, there is also the
- 16:13value that they generate, right? So, for example,
- 16:15in areas like climate, disaster management, conservation,
- 16:19artificial intelligence can significantly reduce
- 16:24costs by improving the efficiency. that we're
- 16:27managing resources, that we're optimizing resources,
- 16:30it can enable better planning. It can help better
- 16:34prevent costly impacts before they happen, right?
- 16:37So the conversation, I think, when we're talking
- 16:39about cost isn't just about costing isolation,
- 16:42but how to use AI responsibly to maximize this
- 16:47positive impact, right? So, for example, in terms
- 16:50of energy and resources, When we're training
- 16:54the largest scale climate models, these models
- 16:56are big. It's huge. Training generative AI models
- 16:59is huge, right? So it does require significant
- 17:02computational resources, yes. Especially during
- 17:05the training phase. We have large amounts of
- 17:07data and the computer we need to use to learn
- 17:12these complex patterns is quite substantial,
- 17:14right? However, this is typically a one -time
- 17:19or infrequent cost. So once these models are
- 17:22trained, running them, what we call inference,
- 17:24is much, much faster and far less energy -intensive
- 17:28than the classical physical approaches. So, for
- 17:30example, in weather, when we're predicting weather,
- 17:32the traditional methods are the physical space
- 17:34methods. These are numerical methods, equations
- 17:36that are thrown into a computer, and they're
- 17:40slow. They're very computational demanding. They
- 17:43take time. They take time to generate these climate
- 17:45models, these climate predictions, these weather
- 17:48predictions at a low -resolution scale. Once
- 17:50you have trained these models, which performance
- 17:53is very close to these traditional models, these
- 17:55traditional physics -based models, what it was
- 17:58taking hours or days to generate with the classical
- 18:01models, and it was taking lots of computational
- 18:03resources, when the genetic AI models are trained,
- 18:06it takes seconds, it takes minutes, at a small
- 18:09fraction, small, small fraction as a cost, right?
- 18:13So at the end, this inference part is much faster.
- 18:18It's... far less energy intensive than the classical
- 18:22physical approaches. So that's the balance. We
- 18:28invest in training these models, but then we
- 18:30have solutions that are efficient, are fast,
- 18:35that they can run in just in a laptop. So this
- 18:38cloud computing, these amazing infrastructures
- 18:42that you need to train these models, you can
- 18:45use them for... Once. For training these models
- 18:47once, you can run these models in your laptop
- 18:49at a fraction of the cost. Very efficient. You
- 18:51can run ensembles of models in seconds, in minutes.
- 18:56That's really cool because I've been playing
- 18:57a lot with agents. I'm running on a VPS that's
- 19:02going through an EPI with OpenAI Codex and I'm
- 19:06having it run on WhatsApp. So I'm just communicating
- 19:10with it and it's just running pretty much on
- 19:12a really, really low power. VPS shared server,
- 19:16right? So that's the beauty that I go back to
- 19:19my first question. I think the amount of money
- 19:22that's being poured into AI, we're seeing this
- 19:24really rapid growth to development and progress.
- 19:29Like pretty much every day, I feel like there's
- 19:31something exciting that's happened. So that must
- 19:34be really, really fascinating for you because
- 19:37it drives you like, ah. Like your imagination
- 19:40can run wild on that level. But do you feel like
- 19:46we're going the right direction? Because there's
- 19:49even talk of building data centers. Are we ethically
- 19:53progressing the right way? Not like for development,
- 19:58like there shouldn't be a frontier for it, but
- 20:01are we doing it ethically? What would be your
- 20:03point of view on how we should scale? Thinking
- 20:08about the planet, thinking about the consequences
- 20:10of this. Yeah. So in my opinion, from my perspective,
- 20:17the key is to focus on efficiency, is to focus
- 20:21on efficiency, is to focus on transparency, and
- 20:23especially what I mentioned before, which is
- 20:25the collaboration, the collaboration between
- 20:26institutions and countries, right? So this is
- 20:28crucial. So what we need to do, we need to design
- 20:32AI systems, we need to design data centers. to
- 20:36be as energy efficient as possible. We need to
- 20:39adopt these best practices for sustainable computation
- 20:42and we also need to integrate renewable energy
- 20:45whenever we can, right? And this is something
- 20:47that at the same time is happening. So a stronger
- 20:50collaboration between researchers, industry,
- 20:53policy makers, this is what can help artificial
- 20:57intelligence ensure that we are deploying artificial
- 21:07intelligence, like we're deploying responsibly
- 21:09and at a scale, right? I think we're making good
- 21:13progress in several areas. AI models are becoming
- 21:16more and more efficient. And this is one of the
- 21:18lines of research, optimizing models, visualizing
- 21:22models, making them efficient and faster. At
- 21:26the same time, we're adopting renewable energy.
- 21:28For example, here in the UAE, we have a lot of
- 21:30sun. There is a good amount of solar fans in
- 21:33the UAE as well. Researchers as well are increasingly
- 21:39focused on reducing this kind of footprint on
- 21:42this training phase, on this in -frame phase.
- 21:45I believe also there is a growing awareness of
- 21:48sustainability in industry and in academia. And
- 21:52at the same time, we need to continue innovating
- 21:55in efficiency, right? We need to continue innovating
- 21:57in model optimization. We need to develop. previous
- 22:00architectures. We need to build the specialized
- 22:01models. We need to containerize, optimize the
- 22:04containerization of solutions. We need to have
- 22:06this energy -aware deployment. So at the end,
- 22:09the goal is to ensure that AI is scaling, it
- 22:13remains sustainable, and it delivers a maximum
- 22:17benefit from a societal and environmental point
- 22:20of view, not only economic, while keeping this
- 22:23energy use under control. So I think we're getting
- 22:28there. Yeah, I think that something that you
- 22:33point out is important and it's like the balances,
- 22:36the cost of the balances of using these models,
- 22:39generative AI. For example, there's a lot in
- 22:41conservation is being used a lot. We have a guest
- 22:45that works for the ocean cleanup and they use
- 22:48it to identify microplastics in the ocean, which
- 22:52is housing heaps. identify this actually an animal
- 22:57or or actually uh microplastics so the identification
- 23:01that one is just reduce the amount of resources
- 23:05that i have to put on that has particular tasks
- 23:08but something that you also mentioned is like
- 23:11i think that one way to go is being is just transparency
- 23:16of the data centers being built i was reading
- 23:19recently that one in and i think that was switzerland
- 23:24but it's more like a focus on languages the ai
- 23:27that a that they are building is a 100 like a
- 23:32running by a renewal energies and also what they
- 23:40are doing is just putting all the data there
- 23:42and being careful with the privacy of the data
- 23:45and how they are feeding their models So yeah,
- 23:48I think that I agree with you. There has to be
- 23:51a collaboration, a huge collaboration, but also
- 23:53transparency and also how they implement these
- 23:58models. I actually want to take a step back.
- 24:00Can we just break down a little bit? because
- 24:03you were talking about image models for your
- 24:05work, for weather prediction and stuff, because
- 24:08you were talking about stable diffusions. Can
- 24:13you run me through the different steps so I get
- 24:15a better understanding of what the different
- 24:18steps of how you train your data? Because it
- 24:21seems like you're using image representation
- 24:23as well, numerical data that you're training
- 24:26with, just so I can get a better understanding.
- 24:28So for training these weather and climate models,
- 24:32What we're doing is we're using generative AI
- 24:35to downscale weather predictions, right? So we
- 24:38take this input low resolution weather forecast
- 24:40and we take this output high resolution weather
- 24:43forecast and we train a generative AI model that
- 24:46maps this low resolution to high resolution,
- 24:48right? This is where the training process happens
- 24:52and collecting this data can come from two different
- 24:55sources. Usually low resolution weather... outputs
- 25:00is publicly available from ECM, WF, from ETA5.
- 25:04There are several sources that provide the high
- 25:06quality analysis data sets at low resolution.
- 25:10High resolution weather outputs is another thing.
- 25:13So this is obtaining high resolution weather
- 25:16outputs. We need either institutions or meteorological
- 25:19agencies that actually make this data public.
- 25:22This data is expensive to get. It's expensive
- 25:24to gather. It requires a lot of domain expertise.
- 25:27And this is what's taken as the kind of ground
- 25:31truth. So as the golden data set, these physics
- 25:34-based models, we know that they work, we know
- 25:36that they work well. There's a lot of research
- 25:37happening there from Earth atmospheric scientists
- 25:40and climate scientists, weather scientists in
- 25:42general. So once we invest into these high -resolution
- 25:51data sets, we can... actually train the genetic
- 25:54models with just a few years of data, four or
- 25:56five years of data. We started a really nice
- 26:01partnership with NVIDIA in 2024 to leverage their
- 26:05technology, their models, and their expertise
- 26:07as well to train these models for DOE and from
- 26:10other regions. So the way that these models work,
- 26:16it's very similar in how these diffusion models
- 26:20generate images. We take an image, which at the
- 26:27end is a weather forecast. If you see weather
- 26:30forecast, if you see climate forecast, it's a
- 26:33grid, it's a picture, it's an image with pixels.
- 26:36Each pixel represents an area. Usually for low
- 26:40resolution, this pixel represents 25 kilometers
- 26:42per 25 kilometers. At high resolution, it can
- 26:45represent models at two kilometers of resolution,
- 26:48where each one of these pixels is two kilometers
- 26:49per two kilometers, and even 200 meters of resolution
- 26:52for the city of Abu Dhabi, which is 200 meters
- 26:55for 200 meters, right? So the model will start
- 26:59from a very noisy version of that image. So with
- 27:04an image that is noise, pure noise. And it progressively
- 27:08refines it until it becomes a sharp and realistic
- 27:12image. And the sharp and realistic image is that
- 27:15realistic that it's so close to a real weather
- 27:19forecast, which is amazing. So we are generating
- 27:23an image that it cannot be distinguished from
- 27:27a real image. And that real image represents
- 27:30a weather pattern, a climate pattern, a climate
- 27:33forecast. So it's impressive. We take. something
- 27:37that is a very low resolution and we convert
- 27:39it into something that represents what will be
- 27:43the temperature, what will be the wind patterns
- 27:44in that building especially. or in that particular
- 27:49location, considering lots of additional inputs
- 27:52and invariants, what is geographical properties,
- 27:55land sea mass, land use, altitudes, and things
- 27:59like this, right? So that's especially useful
- 28:03to capture the small -scale physics and to detect
- 28:05extreme events, to detect extreme events that
- 28:08are missing the low -resolution outputs that
- 28:10represent 25 kilometers per 25 kilometers. What
- 28:13do we have, for example, what's the temperature
- 28:14in a city? in general, what's the wind pattern
- 28:17in a city instead of what's the temperature,
- 28:19what's the wind pattern in some particular areas
- 28:22of the city, in some mountainous areas, in some
- 28:25coastal zones. And we can do this. These models
- 28:28are also stochastic. These models are probabilistic.
- 28:30So what it means is that we are able also to
- 28:33measure uncertainty. We don't have just one prediction
- 28:36of what will be the temperature here, but we
- 28:38have what will be the potential range of temperatures
- 28:41and how this potential range will affect the
- 28:43other variables as well. So if we think about
- 28:46this, Genentech AI in general, it looks sometimes
- 28:51like a science fiction, right? Because at the
- 28:53end, we're creating things that feel almost lifelike
- 28:57or human, right? So from images, from text, from
- 29:01music, from simulation of complex systems. And
- 29:03at the end, we're simulating a complex system.
- 29:05We're simulating a complex system that it's such,
- 29:08such, such real. It's that close to reality that
- 29:12it's almost reality. So what's happening at the
- 29:16end behind the scenes is not magic. It's pure
- 29:19mathematics, pure statistics. We have models
- 29:22that learn patterns from huge amounts of data
- 29:25and they generate new content that follow these
- 29:28patterns, that follow the patterns that they
- 29:29have learned. And this actually shows up in many
- 29:32areas. So it can help generate ideas for products.
- 29:35It can help educators, I don't know, to create
- 29:38interactive learning approaches, personalized
- 29:41education. It can simulate traffic flows based
- 29:46on what the model has learned for cities it can
- 29:48optimize supply chain by generating these potential
- 29:51scenarios like what will happen if these things
- 29:53are happening in the future what will be the
- 29:55consequences of this scenario planning is a very
- 29:58interesting um approach for this right uh output
- 30:02for this so the key point is that this uh these
- 30:05models like with amplifies the problem solving
- 30:07at the end um and it turns in my opinion it turns
- 30:12what was taking month or years to generate into
- 30:17something that is taking minutes, which is why
- 30:20it feels so impressive and is attracting that
- 30:24attention. Yeah, and that's very interesting
- 30:27because usually, for example, in fields like
- 30:30agriculture, the farmers have a lot of issues
- 30:32predicting the weather and the crops, and they
- 30:37have a lot of losses because of the weather.
- 30:40This kind of models can help. with the like a
- 30:45mitigate those impacts environmental impacts
- 30:48and weather impacts that can have in a crop on
- 30:51so provide suggestions to the farmers and how
- 30:54to grow and grow in a better way their crops
- 30:58yeah absolutely yeah so and that's that's one
- 31:02of the targets and that's one of the goals and
- 31:04applications that we like um so we we mentioned
- 31:08that this uh traditional physics -based models
- 31:10which is what's traditionally used by meteorological
- 31:12agencies We know that they are good. We know
- 31:14that they are real. This is what is being used
- 31:17and they are good and they require a lot of domain
- 31:20knowledge, a lot of domain expertise, but they
- 31:22are expensive, right? So not everyone, not everywhere
- 31:27we have these computational resources to run
- 31:29these physics -based models, right? And I believe
- 31:32that this is where this kind of genetic AI models
- 31:35that have been getting implemented for weather
- 31:39and climate can play a key role. um so we mentioned
- 31:42that it takes some cost to train them but then
- 31:45inference is cheap anyone with a laptop can run
- 31:48as models instead of seconds it might take some
- 31:50minutes but it it will happen right and especially
- 31:54for cases like this for agriculture this is where
- 32:00we can have very localized weather forecasts
- 32:03and this is where we have good chances of detecting
- 32:06extreme climate events at particular locations,
- 32:10at the specific locations that we miss with more
- 32:12coarse base models. We can also generate with
- 32:16these genetic models, we can generate many scenarios,
- 32:19we can generate an ensemble of forecasts, we
- 32:21can evaluate uncertainty or how a climate event
- 32:25might happen and can affect a potential crop
- 32:29or a future crop, right? So this is how these
- 32:34models, this is how genetic AI for this kind
- 32:36of topics for weather and climate can play a
- 32:40key role, can help farmers, can help cities planning
- 32:45for disaster management, can detect heat hotspots
- 32:48as well. This is an interesting topic. So we
- 32:51can detect areas of a city where heat hotspots
- 32:55might happen as well, might affect population,
- 32:58might stress out the infrastructure, things like
- 33:01this. More and more agencies and institutions
- 33:04interested at the end in in developing artificial
- 33:07intelligence for these solutions, for these domains.
- 33:11I believe this is something that cannot be avoided
- 33:14mentally. AI is happening everywhere, but it's
- 33:18important also to say that AI is not getting
- 33:21there to replace this knowledge and these traditional
- 33:23methods. So we will always need them. We'll always
- 33:26need this domain expertise. We'll always need
- 33:28how to run these physics -based models, these
- 33:32traditional models, because at the end, this
- 33:34is what is generated in the data that we're using
- 33:36to train our models, right? So I believe that
- 33:39both things can come together. We need these
- 33:42traditional physics -based models by methodological
- 33:45agencies to train our generative models. We need
- 33:48these physics -based models to act as a counter
- 33:52to evaluate our generative models. We need these
- 33:57physics -based models to actually check if the
- 34:00predictions that we believe that are more possible
- 34:02and feasible for these generative models are
- 34:04correct or not, or they deviate too much from
- 34:07the domain knowledge. So both things have to
- 34:10come together. At a personal level, do you use
- 34:13AI? And how can it be affected? And how do you
- 34:17modulate it with your family? And how do you
- 34:19control it? A lot of people are moderating it
- 34:23that much, using it a lot. So I believe everyone
- 34:26is using now ChatGPT for lots of things. But
- 34:29also this is becoming a lot in the developer
- 34:32life and at the industry level. So we have now
- 34:35not only ChatGPT, but we have Microsoft Copilot,
- 34:38we have... Coursera for assisting developers,
- 34:43going faster at the end, right? So there is a
- 34:46lot of talks about this pipe coding, whether
- 34:47the pipe coding is good or not. I think that,
- 34:50as I said before, it's a support tool. So it's
- 34:54something that is here to support you, to support
- 34:57decision -making, not to take decisions for you.
- 35:01Not to do things for you, but to support you.
- 35:02It's a companion. It's another tool. For example,
- 35:05it may be very useful, right? So there's something
- 35:08saying, lots of emails, lots of messages from
- 35:10teams and so on. Compile it, for example. It's
- 35:12great. Can you please check me if I have any
- 35:16attachment from that person regarding that topic?
- 35:19And they will find it for me. And that's great.
- 35:22It makes my life easy. It makes our life easy
- 35:24at the end. I'm not saying that I'm going to
- 35:28rely and live all my life for these things like,
- 35:31hey, tell me what I need to do tomorrow or anything
- 35:34like that. But it's a tool that's here to support
- 35:37us. And I think it's here to stay. Yes. Fully
- 35:41embracing the AI. Probably. And if you have a
- 35:46message for our public, what would you, what
- 35:48would you message? This community plays a key
- 35:52role on developing new solutions. on facilitating
- 35:57this understanding as well to the general domain.
- 36:00And these things have to come together, research,
- 36:04industry, public domain. These tools are very
- 36:11useful. And as I was saying, they are support
- 36:15tools. They are tools that are here to guide
- 36:18decision -making to help us in our everyday tasks,
- 36:23not just to take control of our for life, but
- 36:27to help and make our lives better. Of course,
- 36:31there must be some awareness. At the end, all
- 36:33these genetic AI tools can be used for things
- 36:36that maybe are not good, like genetic images,
- 36:38which violate any policy or any rules. It can
- 36:44generate fake data. So I'm not saying everything
- 36:46is beautiful and everybody is using AI for good.
- 36:52But this is the... direction where we need to
- 36:54move where we're moving which is developing artificial
- 36:57intelligence uh for good for a human society
- 37:02at human ai level that's that's that's the point
- 37:04developing uh human ai level um artificial intelligence
- 37:08yeah thank you yeah thank you thank you for having
- 37:13me yeah We make Science Behind the Glass because
- 37:22we want to show you the people who are changing
- 37:24the world. Researchers from every field, social
- 37:28leaders, educators, and those working in politics,
- 37:31economics, and conservation to understand reality
- 37:34as a whole. Because science doesn't live in isolation.
- 37:37It is the sum of all its parts. We are a bilingual
- 37:41podcast, and now we have a blog too, breaking
- 37:44down language barriers to bring knowledge to
- 37:46every corner of the world. We dream of a world
- 37:49where anyone, anywhere can connect with these
- 37:53ideas in a simple, equitable way and feel the
- 37:56same passion they would feel watching a soccer
- 37:58match. We do all this out of our own pockets
- 38:01and that's why we need your support. To keep
- 38:05this platform alive and to keep giving a voice
- 38:07to the people working behind the scenes and who
- 38:10make this a better world, we need your support
- 38:13and with a donation at sciencebehindtheglass
- 38:15.com you can give us a hand. Every contribution
- 38:18counts and keeps this mission alive. And remember,
- 38:23you can listen to us in all audio platforms.
- 38:25This episode was made possible by Daniel Esguerra
- 38:28and me, Linda Hernandez. Original music by our
- 38:31colleague and friend, Bernardo Aguila. And with
- 38:34a super powerful help behind the scenes and sometimes
- 38:37as a host, Rafaeli Jordan. See you next time.