\n
![](\"https:\/\/media.news.climate.columbia.edu\/wp-content\/uploads\/2026\/10\/Agentic-AI-Event-Diane-Bondareff.jpg\")
<\/p>\n
\n<\/p>\n
One day in late September, researchers, engineers, students and industry leaders gathered at the Columbia Engineering Innovation Hub for \u201cThe Rise of Agentic AI: What It Means for Climate Innovation<\/strong>,\u201d an event hosted by LEAP<\/a><\/strong> (Learning the Earth through Artificial Intelligence and Physics), an NSF Science and Technology Center. The event brought together academic and private-sector experts to explore how agentic AI systems\u2014which can plan, write and run code, and iterate on their own results with limited human oversight\u2014are beginning to reshape how Earth systems are modeled, improve climate predictions and support adaptive decision-making.<\/p>\n The morning sessions explored how agentic AI is being used in climate science today, where it still falls short and what it takes to put it to work well. The closing panel discussion considered how dramatically faster workflows might change which scientific questions are worth asking; how data and tools can become more accessible to the communities and decision-makers who need them; and how working with AI agents might change collaboration across scientific disciplines.<\/p>\n Several themes emerged throughout the day: how to build trust in and verify AI-generated scientific work, the role of human expertise and intuition as agents take on more technical tasks, and how to keep high-quality research from being lost in a flood of automated output. <\/p>\n Below are selected excerpts from the day\u2019s sessions, followed by full video recordings and transcripts.<\/p>\n Panelists not quoted include Kara Lamb<\/a>, associate research scientist in the Department of Earth and Environmental Engineering; Duncan Anderson, senior developer relationship manager at NVIDIA; and moderator Dhruv Balwada<\/a>, assistant research professor at Lamont-Doherty Earth Observatory, which is part of the Columbia Climate School.<\/p>\n Transcripts are machine-generated and may contain errors or inaccuracies.<\/em><\/p>\n Vishal Mishra:<\/strong> Welcome all to the discussion on the rise of agentic AI, what it means for climate innovation. My name is Vishal Mishra. I\u2019m a faculty in the computer science department here. I\u2019m also the Vice Dean for Computing and AI in the School of Engineering. So this event is hosted by LEAP, the NSF Science and Technology Center, led by Pierre Gentine. I\u2019m very thankful for their leadership. LEAP is one of the prides of the School of Engineering, and I\u2019d like to congratulate Pierre and the whole LEAP staff and everyone involved because they just got renewed for another five years by NSF.<\/p>\n And so as I\u2019m sure everyone is aware, these days, federal funding is a little bit difficult. So this is an incredible achievement and a statement on the conference that the funding agencies have on the important work that\u2019s being done here. Now, if you\u2019ve read any news over the past week, you probably think agentic AI is going to kill us all and make humans extinct. Whereas with climate science, we are trying to save the world. So I think this is a very important topic to discuss and take it seriously and not go with all the headlines. There are a lot of heavy hitters here in this room.<\/p>\n It\u2019s an incredible gathering of people, and I\u2019m sure we\u2019ll all be extremely well-informed after this event on what agentic AI means for climate. With that, I am going to hand it over to Catherine and Pierre, who unfortunately couldn\u2019t be here in person, but he\u2019s going to do his talk over Zoom. And I also want to thank Dhruv, who\u2019s not here right now, but he\u2019ll be moderating the panel, and we have all the panelists here. With that, I\u2019ll hand over to\u2026<\/p>\n Thanks.<\/p>\n All right, Pierre, I\u2019m going to tee you up.<\/p>\n Thank you so much.<\/p>\n Pierre, are you able to share your screen?<\/p>\n Wonderful. All yours.<\/p>\n Pierre Gentine:<\/strong> Okay. Thanks so much, and thanks so much, Vishal, for the introduction. Apologies if I couldn\u2019t be in person today, but wanted to share some of the excitements that we have at LEAP in particular, but I think across engineering and sciences, with the rise of agents and what they can do to achieve a lot of the things we\u2019ve been wanting to do. And so, in my talk, I just want to briefly talk about especially vibe coding, but beyond vibe coding, which has been kind of the new things over the last couple of years. Thinking all the way from vibe coding to autonomous discovery, and scientific discovery, especially for the earth sciences.<\/p>\n But a lot of that is pretty generic, I would say, and could be reframed for the computational sciences as a whole. And apologies, this time of the year when we get sick, when we have kids. So I think now what we\u2019ve been witnessing over the last couple of months and years is really kind of those five steps in the autonomy ladder where we started using agents to help us with coding, right? Like as assistants, right? And then now we are getting into a stage where they can start taking big chunks of code, and then translating them from, say, older languages, programming language, and then modernizing them or optimizing them as well.<\/p>\n And then in parallel to that, we are very much using those agents to analyze a massive amount of data, right? And that\u2019s been typically a challenge. We don\u2019t really know where to approach that. We don\u2019t know where to get started. And now we can do things that are relatively complex with those agents. And what that can unleash is actually also using that for simulations, run simulations, we\u2019re going to see that in a minute, and also to test hypothesis, which is at the heart of what we want to do as scientists and engineers, right? And I think kind of the ultimate ladder in that scale is trying to think about AI-assisted scientific discovery, right?<\/p>\n Where we could have the full pipeline from hypothesis to code generation to simulation and the analysis, all the way with humans in the loop to start judging, really the discovery potentially assisted by agents. A lot of that, when we start thinking about vibe coding, really was unleashed by the fact that agents and LLMs in particular have become a lot more powerful, and they can now expand a lot more in terms of the scale and the context that they can cover. And over just the last couple of years, we\u2019ve been at this incredible expansion in terms of the scale of the things we can do with those LLMs.<\/p>\n And initially, it was just very simple auto-complete. We were coding, and they were completing some of the things, and then we could do chunks of code. But then very quickly, now it\u2019s become the full coding environment, and we just achieved what is called vibe coding, where the agents are just generating the code, and they can potentially have some other agents checking some of that code as well. So it starts becoming an interaction between agents. And some of the production has become so massive that we might not even have the time to actually review some of that.<\/p>\n And that\u2019s how it was framed last year, is that really this concept of vibe coding is like fully give it to the vibes and forget that the code even exists, right? So now we are very much functioning at a higher level, and some people still want to review the code, but at the same time, there\u2019s just an overproduction of code, right? So very similar to the time when we started developing code in, say, Fortran or some higher level language, and we couldn\u2019t really look at all of the assembly generation and the compiler part of that.<\/p>\n And just the sheer amount of the code that\u2019s being produced is really dramatic. And this is an example here where we can just generate, for instance, some data visualization just by your prompt. You can load data, you can look at the anomalies, you can write and run a script, and then at the end of the day, you can visualize what you want. What we\u2019ve been doing, especially at LEAP, but I think I\u2019m going to show some other examples, for instance, in Germany, we\u2019ve been using agents to help modernize code. So one of the unique features of earth system modeling is that oftentimes we rely on codes that have been developed over decades, and that very much build on a lot of knowledge of scientists, and that have been basically summarized in key equations that are representing the Earth.<\/p>\n And the issue is that oftentimes those codes are based on languages such as Fortran or C, and most especially junior scientists are a lot more familiar with Python, right? And so one thing we\u2019ve done, especially with students, Ayalulu, but also some other colleagues as part of LEAP, we\u2019ve been trying to convert part of Fortran codes into a modernized version of that Python base. But to make it efficient, we also leverage some of the modern machine learning libraries such as JAX or PyTorch that are basically written in Python, but a lot of the back end is actually very efficient and highly parallelized as well.<\/p>\n So this is an example just for a quadratic function. You could take some code, you could use agents to actually write that, and potentially you could actually do that at scale. And that\u2019s basically what Aya has been doing for a small chunk, like tens of thousands of lines of code of the community land model. And this is just to show you some results. What that piece of code is doing is simulating the exchange of heat and water, so evaporation, we call it latent heat flux in the jargon, and the amount of carbon and photosynthesis that is obtained by plants, which is called GPP, gross primary production. Main point being is that the code converted to JAX is actually very similar to the original Fortran code within some numerical bounds.<\/p>\n It cannot be exactly the same for a couple of reasons, because we want to actually have very good gradients, and so we need to make some\u2026 There\u2019s a little bit of a trade-off, but roughly speaking, we can actually get parity in terms of the code. And the big advantage is once we\u2019re in that kind of modernized platform, we can leverage a lot of modern hardware such as GPUs, and everybody knows that now NVIDIA is a very important company, and a lot of that relies on the fact that they\u2019ve been delivering many of those new hardware, and so we can now leverage them, which wasn\u2019t really typically the case in earth system models that oftentimes rely just on CPUs, or they have just fraction of components that are GPU-based. And once we transfer them to this modernized language, it\u2019s a lot easier to use them. On top of that, one big advantage is once we can leverage this modernized machine learning platform, we can leverage the gradients because we can use what we call automatic differentiation.<\/p>\n And we can use modernized optimization techniques such as backpropagation. So that\u2019s been one thing we\u2019ve been doing. There are colleagues that have been doing that at a bigger scale. So that\u2019s very recent work as of earlier this summer from people in Germany. They\u2019ve been converting an entire ocean model from Fortran again to JAX. They did it in three steps. They first converted the Fortran code to a C reference, which was actually smaller in scale, and then to Kokkos, and then converted that to JAX. And then again, they could get many of the same benefits that we have.<\/p>\n A code that is very much on parity with the original code, but that could actually then leverage GPU acceleration. That wasn\u2019t the case in the original code. So that was a huge leverage compared to the initial code. Big plus and advantage is that it\u2019s Python-based, so relatively easy to read, a lot more inclusive, so we can actually have a lot more people using that, which is a big advantage or the reach of the model. I would say the next frontier, so that\u2019s where they are. They were at basically 70,000 lines of code, and ideally what we would want to do is doing that for the full Earth system model.<\/p>\n So if you\u2019re interested, we are actually very much interested in doing that with the community or system model with NCAR, and we have a few people interested in being part of that, so please reach out if that\u2019s something of interest. But I would say that would be the final frontier for Earth system modeling. In parallel, what we\u2019ve seen is that we are seeing a lot of use of agents to analyze complex data. So when we think about the Earth\u2019s system, we are dealing with very complex, multivariate, multi-levels data, so very high dimensional, many, many petabytes and terabytes of data.<\/p>\n And now what we can do is we can very easily analyze those complex data sets using agents. So those agents can look up the data online. They can potentially download that or do the analysis online if it\u2019s cloud-based, such as the big infrastructure that we have. And then they can make some analysis such as this example here, just with a simple prompt. We can now look at time series. So things that would have taken days or sometimes weeks in terms of data analysis can be done very efficiently, very quickly with those agents. So that\u2019s a game changer to start thinking about data analysis and hypothesis testing, and that\u2019s really a game changer in my view.<\/p>\n We can use that not just to use existing data, but also to test models against data. Right? So oftentimes we want to test the validity and the quality of models. So this is an example here where we have an Earth system model to the left. We have some observational truth in the middle here. So this is temperature at the top, and we want to compare that and ideally try to reduce some of the biases of the models. And this is something we can now use agents for, right? So those agents can first visualize the data.<\/p>\n They can search and look up some of those differences that we have and the biases, and then they can try to reduce those biases. So we can start having a feedback loop, with those agents to evaluate the models and then potentially calibrate or recalibrate the models so that they are more faithful to the truth, which was typically done by hand, and that\u2019s really, again, another game changer from those agents. And this is really what we can do, and we can do things beyond just training and recalibrating parameters. But we can even test ideas, like is that particular formulation of a physics of the ocean correct, or is this one a little bit better? And this is an example here where we are changing.<\/p>\n This is some work with actually Drew here, where we can actually test different types of formulations in the ocean, and this is the version before, and that\u2019s the version after. We have some truth here or some more faithful simulation in the middle, and we can look at reduction in the biases here and the impact of the changes in the formulation. So again, we can really start thinking of that for hypothesis testing through that data visualization analysis and hypothesis testing feedback loop, which I think is very important. We are converging, I would say more and more, and this is the work by colleagues actually they are part of LEAP, Duncan Watson in Paris and Rose Yue at the UCSD, where they\u2019ve been developing this kind of weather agentic pipeline where what you can do is called Zephyrs, where you can actually execute code.<\/p>\n You can search some particular observations as well. And you can compare that to your forecast, and you can combine that all together to assess the quality of the forecast or to answer a particular question. For instance, if you have a particular question about potential future extreme events over the next couple of days, you could actually do that through this full agentic-type pipeline. So that\u2019s really a game changer there. We are seeing that as well for climate and climate simulation. So this is an example of EarthLink by people here in China, where they\u2019ve been using this pipeline to start coding, analyzing data, running simulations, and potentially getting to the full pipeline of, again, this full feedback loop between simulation and data analysis as well. But that\u2019s really where we are starting to see a lot of progress, the full pipeline agents.<\/p>\n And just to wrap up, what we are seeing more and more is really trying to use agents for the entire pipeline, right? So we would formulate hypothesis typically with humans, or we could potentially using agents there as well. The agents would then develop the code, right? They could actually simulate them and run a bunch of simulations with different hypothesis potentially. As I showed you before, we would analyze the results, comparing that to potential observations or not. And then the humans would actually check and then refine some of the hypothesis. And we are starting to see some of that.<\/p>\n There was a big splash, for instance, earlier this year with this AI scientist where you would have an idea for a paper end-to-end, and then the one paper actually passed the workshop peer review. So that\u2019s kind of getting started there as well. And another example is AlphaEvolve from DeepMind, where you can have new algorithms developments from scratch, basically. So with that, I\u2019d be happy to take any questions later. Thanks.<\/p>\n Thank you, Pierre. You can clap for him. It\u2019s okay. There was a little bit of movement in the room. Thank you, Pierre.<\/p>\n That\u2019s okay. Thanks.<\/p>\n We look forward to hearing from you at the panel. And at this time, Ignacio, we welcome you to share your screens.<\/p>\n Thank you.<\/p>\n Let\u2019s see.<\/p>\n Ignacio Lopez Gomez:<\/strong> All right. So thank you for having me here. It\u2019s an exciting panel and discussion. I\u2019m Ignacio. I\u2019m a research scientist at Google. And today I\u2019m going to be talking about how we\u2019ve recently used agentic AI to develop a seasonal streamflow forecasting model for the state of California. And beyond the results that I will be showing here for this particular application, I encourage you to think along the way about ways that you can implement these methods into other tasks or projects in climate adaptation and mitigation. So the reason why we decided to focus on streamflow forecasting in California is that much of the water or a large fraction of the water that is consumed in that state originates as snow in the Sierra Nevada mountains.<\/p>\n And so the timing of the volume of the snowmelt of that snow is going to affect water availability for many communities there. For these reasons, city planners and infrastructure managers need accurate seasonal forecasts of this water availability to ensure appropriate water allocations for cities and also to mitigate flood risk. Now, the problem is that warming temperatures with climate change are already changing some of the historical patterns that we used to observe between snow and other environmental factors. And this has already led to a degradation in the scale of operational forecasts, which is then passed to downstream users.<\/p>\n And it crystallizes increased risk in not having enough water or having more flood risk, et cetera. So for this particular application, we thought that agentic AI would be a very suitable method for a few reasons. The first reason is that for this particular task, we don\u2019t actually have a lot of data. So we have maybe a few decades of high-quality data for snowpack and streamflow, and we\u2019re dealing with monthly timescales, so it really is not that much. And in addition to this, as I was saying before, this data has undergone shifts in its statistical distribution. So this makes it even harder for, let\u2019s say, traditional machine learning models, such as neural networks, to actually solve this problem easily.<\/p>\n Another reason why agentic AI can help in this problem is that developing a good streamflow forecasting model from scratch requires a good deal of data gathering from multiple sources, ground networks and so on, as well as its curation in terms of errors or empty values that we might find, and also requires a lot of feature engineering, model exploration, et cetera. And so there\u2019s really a lot of potential here to save some of that labor-intensive tasks. Fortunately, there\u2019s already been agentic systems that have been developed that can really help with these issues. So one of them is AlphaEvolve, already mentioned by Pierre in the previous presentation.<\/p>\n And another one here, also developed by a team at Google, is the Empirical Research Assistant, or ERA. And so the way both AlphaEvolve and this system work is that we\u2019re going to use a large language model for code mutation, so to generate code, to solve a specific problem, which we\u2019re going to call a scorable problem. This just means that we can evaluate how well we\u2019re doing in that problem with a certain number, and we can compare numbers. And so what we\u2019re going to do with that score is we\u2019re going to compare different candidate solutions, and then we can undergo an efficient research mechanism, for instance, to essentially select the best candidates from that process.<\/p>\n And we can do this iteratively until we obtain a method that is adequate for the goals that we are pursuing. So this type of system is not the only agentic system that we can use. And so Pierre before was already talking about all the ways that agents can help accelerate progress in climate adaptation and other applications. And so for our particular application, we also used another agentic system called Google Antigravity. This is a more general-purpose agentic system that essentially uses a large language model that is capable of both interacting with the user but also interacting with general-purpose tools such as internet search, access to file systems, et cetera. And so the way that we use these two agentic systems to accelerate development in this project was first we used the general-purpose Google Antigravity to help us gather clean data from open sources in the internet. And we also used it to formulate a hypothesis and formulate initial implementations of methods or approaches to solve the task. And once we\u2019ve developed these family of methods that we would like to test out for this process, we can then fit these into systems such as ERA or AlphaEvolve that are going to take those solutions and iteratively refine them until we can find the best possible actual model from that family of model.<\/p>\n And we can then compare them and take the best one as our forecasting system. Of course, we can do this process iteratively. So we can start from the initial pass through the two stages, and if we\u2019re not happy with the result, we can then feed back the results and our evaluations of them to the first agent and do this iteratively until we\u2019re happy with the result. And this really accelerates development compared to just doing everything by humans. Now, you may be wondering where human feedback can enter in these loops. And so here I\u2019ve included some examples of where humans can really steer the direction of the development process at all stages.<\/p>\n And so, for instance, we may start with the data research or search process with Antigravity, where we can prompt the system to search for datasets that are specifically operationally available for monthly prediction and for the state of California. So here we would be limiting datasets in terms of their access and latency. We can also prompt Antigravity to search the literature for specific modeling ideas or for specifically physically motivated feature formulations to test out when we\u2019re developing our model. And then going back to that second agentic system that we\u2019re using to really fine-tune each of the methods, we can, for instance, impose persistent guidelines, so things that we wouldn\u2019t want the computation agents to change. And so an example that we use here for this application is the imposition of monotone constraints between antecedent snowpack and future streamflow.<\/p>\n And so this is based on the instinct that if there\u2019s more snow accumulated in the mountains, it should resolve into a larger water streamflow in the future, and it\u2019s better if we can just impose that from scratch, and the model doesn\u2019t have to learn it from the limited data that we have. Apart from imposing persistent guidelines, we can also tell the large language model what to mutate in the code or what to focus on. And so, for instance, here we might want to tell it to focus specifically on iterating on the method that is used to aggregate data from stations or from ground networks. And we can do this essentially at all stages of the modeling framework.<\/p>\n So the result of using this agentic framework was the development of a forecasting system that outperforms the operational forecasts that are used today for the state of California. And so here, on the left-hand of the slide, I\u2019m showing a figure of the forecast error for predictions of streamflow over the April through July melt season in California, and I\u2019m showing results for both our model, which is shown in teal, and the operational baseline, which is shown in red. So we can see that our model is able to reduce errors for issuance months going from February, so early forecast of that melt season, all the way to April when the melt season is about to start.<\/p>\n Of course, it\u2019s important to look at how we are doing better than the operational baseline, because we wouldn\u2019t want to keep a system that just outperforms in very extreme years but otherwise is underperforming the system. And so this is what we compared here in the figure shown in the right-hand side of the slide, where I\u2019m showing the head-to-head win rate for different snow-fed basins in California and across the testbed. And we can see that largely, our system is able to outperform the operational forecast across most of the basins of the state. And so what we see is that this model consistently outperforms the operational forecast, and this was developed on a very short timescale, in the timescale of one to two months, compared to the operational forecasts.<\/p>\n Now, the models that are developed by agents are not necessarily black boxes. So these are essentially implementations in code of models that can be very sparse, they can be physically motivated. And we can also use not only the solution for forecasting, but also to learn about the processes that are driving those forecasts. And so here are a few examples of some of the emergent discoveries that were given from the agent-driven search that I think were not obvious for us from the start. So the first one is, for instance, the choice of an ensemble forecasting system as a model with specialized submodels that use different sorts of input features and also different target formulations.<\/p>\n The second was the use of an adaptive weighting mechanism for the submodel forecast, depending on the issuance month. So for instance, forecasts that are issued during the melt season will have a different weighting of the different submodels compared to forecasts that are issued during the melt season, and so on. And then we can also look at the feature engineering that the agents performed in order to get the most signal. And so we here observe some fairly complex feature engineering, such as the creation of predictors of snowpack ripeness in terms of dating functions and nonlinear functions that have a physical motivation underlying them.<\/p>\n And this explains the \u2026 the scale of the model that we created. Apart from evaluating the model in terms of its scale, we can also look at how the model captures the hydroclimatic changes in California that motivated this study. And so in this slide, I\u2019m showing some of the changes that we are currently already observing in snow melt timing, in terms of warming. And so what I\u2019m showing here are changes on the left-hand side, the time period over which snow melt happens in California. So what we observe is that that time period is shortening every year as a function of warming, and this shortening is stronger for higher elevation basins. And on the right-hand side, I\u2019m also showing the collapse in late summer runoff. So as warming increases, most of the runoff is going to shift earlier towards March, April, and May, and we\u2019re going to see less water availability for the drier summer months.<\/p>\n And so what we can see here is that the model at different lead times is able to capture quite well those trends in the reduction of both the melt period and also the fraction of summer runoff that is available in California. So circling back to the topic of this panel, and thinking about what my experience with this system means for climate innovation. I think one of the main takeaways for me is that these agentic systems enable humans to act more as principal investigators, where we are going to be able to focus on the bigger picture, on proposing different hypothesis and steering research direction, while agents can handle lower-level workflows, such as code implementation, evaluation, and so on, that typically take a significant amount of time if performed by humans.<\/p>\n I think we\u2019ve also already seen that this is going to really lead to an accelerated co-development in science. So, I think humans and agents working together are just going to lead to an unprecedented pace in the creation of scientific models and scientific insights, and so it\u2019s a good question of how we\u2019re going to deal with that. And then perhaps more specific to the topic of climate, because the other two I think are more general for science and engineering. I think agentic AI is also going to enable us to perform more rapid climate adaptation. And what I mean by this is that agentic pipelines, because they are going to allow us to efficiently modify and recalibrate models, we\u2019re going to be able to essentially take more and more data as it becomes available, curate it, and improve the models that we use, for instance, for environmental risk.<\/p>\n So these could be extreme heat risk, flood risk, and so on. And so instead of relying on models that may be based on past climates, we will always have models for these risks that are based on the latest information, and I think that\u2019s going to enable us to better adapt to climate. And with that, I\u2019ll take any questions later. Thank you.<\/p>\n Thank you so much, Ignacio. Yes, as Ignacio said, we will take questions after the panel. At this time, we would like to call up Dr. Kara Lamb.<\/p>\n Kara Lamb:<\/strong> Okay. Hi, everyone. My name is Kara Lamb, and I\u2019m the AI for Science Director here at LEAP, and I\u2019m going to be talking about systemizing earth system model development with agentic AI. So one of the challenges with climate modeling right now is that we still have uncertainty in some of the physical processes that we represent in our models. So we have to represent processes in the atmosphere, in the ocean, in the land, things like sea ice in the cryosphere. And because we have to run these models for a very long time, like hundreds of years, we have to make some trade-offs in terms of the spatial resolution at which we are representing different processes.<\/p>\n So we have to come up with very simplified representations, which we call parameterizations, for many of these different processes in these different systems. And these parameterizations, in terms of both uncertainty and the physics itself that goes into these models, as well as in terms of their simplified representations, are still some of the main drivers of the uncertainty in our climate projections over the next one hundred years. So particularly in the near term, like the 30 to 50-year time period, we see that these types of physical process uncertainties are really dominant in terms of our prediction of future climate.<\/p>\n And we are constantly, though, improving these parameterizations as scientists and coming up with new understanding of some of these processes through things like, say, field campaigns, where we\u2019ll go into the field and make measurements, or through improved models where we have a much better representation of a process. But it still takes a very long time for that type of new physical process understanding to actually be integrated into our Earth system models. So there we have some new physical insight, then we have to translate that scientific idea into stable code, and then that code has to be integrated into these very large code bases of these legacy earth system models that are often in Fortran or C++. And a lot of the work that we\u2019ve done at LEAP over the past five years was using machine learning methods and data-driven approaches to actually improve these parameterizations.<\/p>\n But that\u2019s actually challenging because, for the sake of integrating these into models, these are machine learning parameterizations are typically written in Fortran or maybe Julia, but then that doesn\u2019t easily integrate into the Fortran or C++ that our legacy climate models have. Once we have a new process representation, we then have to tune our entire model and test and validate this very large earth system model to see if we can then improve how well the model actually compares with our present-day observations. And so this whole process is very slow. It has expensive iteration, and there\u2019s a lot of expert intuition that\u2019s involved in the process in terms of figuring out which particular parameters to tune or which particular processes to target.<\/p>\n So it might take years for any new physical insight that we have about a specific process to actually be integrated into our operational earth system models. And this is just a very famous paper, The Art and Science of Climate Model Tuning, where they\u2019re really saying that this is both an art and a science because we have really this involvement of experts in this process of really trying to understand how we can tune these models. So, as the two previous speakers talked about, there\u2019s really a lot of potential now to really make this whole process much more faster and even more systematic. And this is really two different things which have emerged in the past couple of years that can help us to more systematically integrate new physical understanding into our system models so we can have this faster iteration of the model development cycle. There is the agentic AI, so we can do code translation, literature review, even parameterization development integration.<\/p>\n And as you heard, this is enabling very complex scientific workflows. Even things like automated laboratories, as well as AI co-scientists or automated model selection for scientific AI. And the other really important piece of this is differentiable programming, and Pierre talked about this earlier in his talk, that we can develop code in this more modern programming language where we can have direct access to the gradients of the models, and this allows us to more easily optimize parameters and also more easily integrate with these machine learned parameterizations, which are really becoming the state-of-the-art approach for some of these parameterizations.<\/p>\n And also this code is something which can run on these more modern architectures like GPUs and TPUs, so we can have more fast and parallel computation for improved speed of these models. So some of the work that I\u2019ve been doing in the last couple of months is just looking at this problem of how can we more systematically develop these parameterizations of different processes, and so that we can more quickly integrate new understanding into climate models and also do this much more systematically. So as a test case, I\u2019m going to focus in on just one specific process, which is one that we have to represent in models, which is collision coalescence.<\/p>\n It\u2019s basically a cloud microphysical process. It is a really important process in the formation of rain. It\u2019s basically when in clouds we have some small droplets and they\u2019ll collide and form larger droplets, and then also we\u2019ll have larger droplets that will fall through the cloud and collect smaller droplets. And it\u2019s a nice kind of hello world example for this parameterization development process because we have many past published schemes of increasing complexity to represent this process. We also have a high-fidelity ground truth, where we can directly simulate this process using a sort of Monte Carlo-type simulation, except for those are too complex to actually use at the earth system model scale.<\/p>\n So we have to come up with simplified representations for simplified models that can be used in our system models. And we also know that this process is not perfectly represented in our perfect models, that there\u2019s still room for improvement. So just to quickly describe how we typically represent it in our system models, we have some populations of droplets in our cloud, and we divide those into the smaller droplets that are just kind of hanging out in the cloud, and the larger droplets, the rain droplets, that are actually falling through the cloud. We will just represent those two populations very simply in terms of their bulk statistical moment. So this is basically the number and mass concentration of those two populations. And then our typical process representation would be a set of coupled ODEs that are just power laws of those four bulk moments.<\/p>\n So what I\u2019ve been exploring is a couple of different paradigms for using agentic AI to accelerate the process of parameterization development and discovery. So one is human expert-in-the-loop development with Cloud Opus. I also have been exploring single-shot and iterative API calls with an LLM, and then also now a fully automated MCP framework where we can really fully automate the parameterization development process. So first I\u2019ll just talk about how we\u2019re going to evaluate the different parameterizations so that we can systematically compare them. So I start with some reference simulations, the very highly accurate representation of this process. I\u2019m going to simplify that into our typical representations that we use in climate models. And then I\u2019m going to look at a couple of different steps that we would need in this whole parameterization development cycle.<\/p>\n I\u2019m going to translate some of the existing historical representations of this process. I\u2019m also going to reproduce some of our machine-learned parameterizations, and I\u2019m also even going to see if we can use large language models to propose new closures and iteratively improve them. And then I have a shared way to compare all of my different parameterizations, so everything is compared systematically in this. With the same JAX interface, I can look at does the code that is produced by these different approaches actually run? Is it going to be differentiable? Does it satisfy physical constraints like mass conservation?<\/p>\n And can I stably integrate this over several hours? I\u2019m going to compare it against the reference simulations, which are the more accurate simulations. And then I\u2019m also going to look at things like the cost in terms of how computationally expensive are these simulations. So I\u2019ll just give you the results for the expert-in-the-loop development. So this is basically with me iteratively working with Claude. It works extremely well for just re-implementing historic parameterizations. This can be done in less than an hour for four different historic parameterizations. It also does a very good job at reconstructing data-driven parameterizations, which are trained on high-fidelity simulations, which are based on some existing machine learning baselines.<\/p>\n So that\u2019s something that it can do in just a couple of hours. Where it starts to be more challenging, though, is where I\u2019m trying to go beyond what has already been done and actually develop new scientific machine learning approaches. And in that case, it requires much, much more expert intervention. So that requires me really working with the model and guiding it much more as it goes. But once we have all these parameterizations, I can then systematically compare them and even retune some of the historic parameterizations to more systematically compare them. And this is just showing basically the timing of rain conversion versus a number of these different parameterizations.<\/p>\n And we see that one of the machine learning parameterizations, the SuperDropNet, has the lowest error. I can also systematically compare things like the computational cost versus the accuracy. So one thing that was interesting was that actually retuning some of the existing physical parameterizations could do even better than some of the machine learning parameterizations, or at least comparable, and they\u2019re much less computationally expensive. So they might actually be better candidates for the models. Using this single shot in automated and iterative API calls, this more automated development, I can look at evaluating where we get typical failure modes for these types of LLM translation tasks. So this is just showing typical failure modes when I try to translate for historic parameterizations.<\/p>\n These are 10 independent calls to a large language model for each of these different parameterizations. And we see some of the parameterizations, if it\u2019s green, it means it is successful for all 10. Some of them are doing much better depending on the parameterization. And then we do see more failures in two of the parameterizations for things like the code that\u2019s produced is not differentiable or it\u2019s unstable. I can also try to see if the large language models can propose new ways to represent this process and then also iteratively improve that with feedback over time when I tell it how well is that actually doing against my high-fidelity reference data.<\/p>\n So the blue here, these are just showing how well it\u2019s doing for predicting the four moments over time. So lower is better. And the blue is showing just single calls, like its first attempt to just try to come up with a new way to represent this process. And then the green is showing when I\u2019ve given it iterative feedback over time. So we do see that it gets progressively better as I tell it how well the previous process representation was doing. And so the last thing that I\u2019m now working on is really thinking about just fully automating this parameterization development process using an MCP framework where we can have the agents that interactively interact with tools and fit data, and they can even run potentially high-fidelity simulations to create new data and systematically improve those parameterizations.<\/p>\n So this is just kind of showing some initial results from that, where I\u2019m showing in black some of the agentically-developed parameterizations compared with some of those historic parameterizations. And again, it\u2019s kind of the accuracy versus complexity. So in the lower left are the better parameterizations here. So I also wanted to talk a little bit about where we\u2019re now going with this next. So this is work that I\u2019m leading, this CM4 project under the Department of Energy\u2019s GENESIS mission project. And so all of what I showed before with collision coalescence was in a very simplified box model, very local processes.<\/p>\n But when we\u2019re talking about representing these processes in the atmosphere, we also have to worry about the larger-scale motions of air in the atmosphere that\u2019s going to accelerate collision coalescence. Atmospheric turbulence can speed up rain formation, and it can do this at different scales. So we can do this at a small scale, but then also we have the larger scale structure of the cloud that we have to worry about. So for this project, we are trying to basically improve and completely update how we\u2019re representing this complex interaction between cloud turbulence and collision coalescence at the earth system model scale.<\/p>\n So I\u2019m using some of the methods that I talked about, where these agentic development and evaluation of different parameterizations. And another thing that we\u2019re doing is once we improve the small-scale process representation, we also need to be able to couple that up to the scale of the earth system model. And so we\u2019re using this existing Fortran-based turbulence model that we\u2019re using agentic AI to translate into JAX so that we can optimize and compare this whole complex parameterization, again, at the earth system model scale. And eventually, this will be used to improve the process representation of the Department of Energy\u2019s earth system model E3SM.<\/p>\n So another thing that I\u2019m really excited about is thinking about how we can actually integrate these into laboratory measurements and really bring laboratory measurements more into our system model development cycle. So right now, this collision coalescence process is something that we actually can\u2019t easily observe in the atmosphere. So a lot of what we do is based on simulations, but we would really like to be able to observe the process and to really understand it, as well as other important cloud processes. And there\u2019s this proposed Draco chamber, which is this new facility that is proposed to really actually for the first time be able to study these types of processes in the laboratory.<\/p>\n So that\u2019s something that will hopefully be in the future. But right now, Brookhaven has already developed this test bed for that, which is this kind of proof of concept NEFOS chamber. And as part of this Brookhaven-led GENESIS mission, we are integrating AI into how this laboratory chamber is both controlled and how we might then use the observations to improve our models. So we have a digital twin of the chamber, basically a model, and we\u2019re going to use this agentic workflow to be able to control how that chamber is operated so that we can target specific atmospheric conditions and then hopefully eventually even take what we learn from the chamber and integrate that directly into improving our process representations in our system models.<\/p>\n So I just wanted to finish. So really, agentic AI is really helping us in many different ways and just echoing some of the past speakers, we are moving towards this much faster iteration cycle for improving our models, and we can also think about doing this in a more reproducible and systematic way, and even including new kinds of observations or laboratory studies that are specifically designed to improve our models. So, with that, I just want to say thank you.<\/p>\n Thank you so much, Kara. At this time, we\u2019d like to welcome up Sameer Mohammed of AWS.<\/p>\n Sameer Mohammed:<\/strong> Everyone, I\u2019m Sameer Mohammed. I\u2019m a deep learning architect at AWS, and over the past two years, I\u2019ve been working both on AI research but then also applications of actual generative AI workloads in production, in industry, and agentic AI as well. So I\u2019ve gathered a lot of experience over those last two years over a diversity of use cases and learned a lot of lessons, made a lot of mistakes, and I\u2019m here to share some of those with you all. One note, I\u2019m not an official AWS spokesperson, but I\u2019ve learned plenty of things I\u2019m hoping to share with you.<\/p>\n The thesis for this presentation is really going to be about the timeline between research and production workloads regarding agentic AI and agentic AI research has never been shorter, and that agents, as they\u2019ve developed, have become more and more capable of handling more and more ambiguous problems. As a result, we\u2019re able to do more than ever before, but the basics of what actually makes a good AI or machine learning system are all the same, and they\u2019re more important than ever before. I\u2019m going to be going through this as a bit of a history lesson at first and then recap and share some more takeaways, but we\u2019ll start with 2024 and work our way to today.<\/p>\n In 2024, you start seeing the first agents really hit production, and the innovation that causes this is the idea that we\u2019ve gone from just next token completion or next word completion to being able to basically mark special words as parameters to feed into software. So we can write tools, which are that software that create execution and act when given a certain output by an LLM. So that gives us what the first agent looked like, and at the moment in 2024, we can only handle small amounts of ambiguity, and we\u2019re still dealing with text as most of our output.<\/p>\n One of the lessons we learned from this time is that in order to get really capable agents, your tooling is incredibly important, and if you build better tooling, you get better results. One example of this is really with document processing. It\u2019s a super common use case we still see all around all sorts of industries today. What industry, what enterprise, what lab even doesn\u2019t have documents? And it\u2019s an opportunity to use documents and data sources, combine them together, and streamline certain workflows. Think you have your databases, you have PDFs that need to be filled out, and there\u2019s this problem that happens when you\u2019re trying to map your first name value from your database to your given name value in your PDF form.<\/p>\n These are still very obvious situations where a little ambiguity can go a long way when you can handle them with agents, and it makes it such that the amount of work that you end up doing ends up being offloaded onto the tools. So in 2024 and around then, a lot of the work that developers have been doing in the agentic AI space and generative AI space are really just developing the tools that allow you to just mix between different modalities. I don\u2019t want to see the output of what should be in my PDF. I want it to put the text into the PDF itself.<\/p>\n And so a lot of the development around this time is around bridging that gap there. As a result, you start to see model context protocol, NCP, really start to come into the fold since it\u2019s this sort of standardization of different ways that you can do that and expose different tools to agents. So one of the lessons that we can take away from here is that your tooling is still super important in order to build proper agentic systems. This has been a really big sea change. We start going from models being able to finally act upon the world to models being able to act upon the world and orchestrate themselves.<\/p>\n Now, that might not seem like a big deal on its face, but it does something that we all do inherently when we solve problems. It opens us up to task decomposition. If I\u2019m trying to get a trip booked from New York to Los Angeles, that might seem simple at first, but I\u2019m going to have to start making plans on what flights to look at, what hotels to book, timings, what reservations I\u2019m going to get for dinner when I land, if restaurants are going to be open when I land, or I\u2019m going to have to DoorDash something. So we all do these sorts of things when we\u2019re challenged with large tasks, and this self-orchestration is this idea that agents not only can just try to attempt these things themselves, but they can decompose these tasks into simpler ones that have less ambiguity and use the intelligence that they have available to them in order to tackle smaller tasks and then bubble up to larger ones.<\/p>\n Additionally, another big innovation at this time is the ability for agents to write tooling themselves. Before I mentioned we\u2019d have to actually write code that allows our agents to interact and fill out these PDFs themselves. At this point, when we say, \u201cWhy don\u2019t we allow the models to give it a shot and try writing that tool themself?\u201d They have vision. They can actually look at the fillable PDF and see if it did it properly. And so that allows for this recursive improvement for these agents to be able to decompose these tasks, write tools which allow them to execute in these smaller problem spaces, and then actually build on those things.<\/p>\n Not only that, but you can also use agents in teams at this point because of that self-orchestration, and you start seeing tasks being decomposed into different ways that you can leverage the parallelism that that unlocks for you. I don\u2019t necessarily have to wait to see what hotels I\u2019m going to book until I finish doing my flights. I can do that in parallel. And that\u2019s really important for a lot of different real industry-specific use cases where that latency can provide a lot of value, particularly in conversational formats where agents are still incredibly popular today. So I\u2019m pointing to this one paper here as being somewhat defining about that, but there\u2019s all sorts of different sources of literature that comes into this whole topic of self-optimization and self-orchestration as a whole.<\/p>\n Another thing that you might be familiar with are harnesses, which is the environment that agents operate in, the tools that they have access to, the prompts that they\u2019ve been informed about, and the environment that they\u2019re exposed to more generally. But then there\u2019s also different sorts of harnesses that have been developed and released more recently. If you\u2019re familiar with Cloud Code or Codex or Pi, these are all harnesses that come prepackaged, and a lot of them have been pretty well optimized at this point by the vendors that release them. And so harness orchestration tends to be less work nowadays, but at the time in 2025, that was a lot of what we would do is sort of building the environment around agents to optimize themselves and to operate as well as they can. So one of the lessons from this is actually the hard part is no longer writing the code, it\u2019s defining what you\u2019re actually trying to do.<\/p>\n Now, that might seem a little silly at first, but when you really dig into what you\u2019re trying to do specifically in regards to research, it\u2019s much harder to actually define than it might seem. We work with a lot of subject matter experts to define what is your final goal going to be. And if you think about it in a scientific context, it\u2019s much easier for these agents to actually execute experiments that have been well-defined or well laid out with objectives that you\u2019re trying to reach. But if your agents don\u2019t necessarily know your thesis, it\u2019s much more difficult for them to iterate on why some of these ablations are more important than others and what sort of axes and lines of thought are worth spending more time going down and exploring. So a lot of the emphasis ends up being on the environment that agents run in and just focusing on allowing them to perform as well as they can.<\/p>\n So now we\u2019re super confident that our systems are performant and they just might not go in the directions we want them to unless we put some effort in. At this point, in 2026 and onward, we\u2019ve seen this idea of inference time scaling really be a big unlock for really, really complex research problems and more scientific problems as a whole. As a result, we don\u2019t really see this necessarily as much in industry. I\u2019d say the most common use case in industry for inference time scaled agents are primarily large code base migrations, where you\u2019re working with hundreds of thousands of lines of code, for example.<\/p>\n But still, this has been super useful for AI research and other science research as well, since you\u2019re able to really dig into the details of some things here. I\u2019ll say, let me define something first is that inference time or test time scaling is this idea that instead of your model learning things at the pre-training or post-training times that these model vendors train them with, it\u2019s learning in the process and in the environment that it\u2019s been deployed in. So think about a model that you just ask a research question and you say, \u201cHey, here\u2019s my AWS account. I\u2019ve got all this data here.<\/p>\n Why don\u2019t you explore it and tell me what lines of thinking or what feature engineering might be useful for these experiments?\u201d And it might spend 20,000 tokens doing that, and then it might spend another 100,000 actually doing the experiments, reading results from experiments, another 100,000 trying to iterate on top of that, and so on. But it\u2019s this idea that you don\u2019t have to do all of your training before the model itself has been released. An example use case for here is that code-based migration, like I mentioned, is important. Actually, before I get to that, I\u2019ll get to the lesson, which is that processes matter a lot for success here.<\/p>\n When you\u2019re doing these really, really long inference time scaled research projects, what ends up being important is not where you\u2019re trying to go, because we\u2019ve established how we defined that before, but now it\u2019s how we\u2019re actually going to get there. Part of the idea here is that the models tend to be pretty poor at actually identifying and having some sort of intuition as to what a good line of research is to go down. And so monitoring these processes and talking to subject matter experts, not just about the sort of research destination that they\u2019re trying to go to, but how we\u2019re approaching getting to that research destination is important. Another thing is if you can actually get your agentic system to be highly performant towards your current objective, there\u2019s actually a pretty good chance that you\u2019ll be able to get them to perform other objectives nearby as similarly well too. And also you might even be able to strike gold and do something really impressive that you didn\u2019t originally intend because of all the context that your model has taken on during the original inference time scaling.<\/p>\n So we really start looking at defining your processes and what a good process looks like as being important. Looping back to my example about code-based migration, stepping away from research a little bit, an example of this might be that we want our product, which at the end of the day is a simple API. It outputs one result, and it\u2019s got hundreds of thousands of lines of code to do that. So it\u2019s a pretty complex system underneath the hood. We can do all this sorts of testing in order to make sure that the outputs are the same. But if you actually look at the code and you look at the infrastructure that\u2019s been built, if you look at the algorithms that have been developed, if there\u2019s not necessarily a faithful one-to-one match, which is honestly hard to say when you\u2019re going from one programming language to another, then you can introduce different concerns that you might be worried about.<\/p>\n You might go from doing something in a very stateless way to doing something in a stateful way. And doing something in a stateful way might require saving state somewhere, and that could be a security risk. And if you don\u2019t expose those things when you\u2019re actually thinking about how you\u2019re doing these operations, then you\u2019re going to be introducing those things as silent failures in your migration. And one day you might have a CVE show up and you might have to push a new update out because you\u2019ve introduced a security fallout because you\u2019ve not implemented these processes properly.<\/p>\n So it\u2019s not just about paying attention to your tools, it\u2019s about paying attention to your destination, and it\u2019s about paying attention to your processes. Ultimately, subject matter experts are still super important for all of this, since it\u2019s their opinions and their knowledge that gets translated into a technical manner that enables all this to happen. So we\u2019re working with a lot of experts on their domains of expertise, and we\u2019re also having more and more conversations and writing less and less code manually. So to take a step back from my history lesson, I want to give you some tips on where I see agents succeeding easily today and where they still have some place to work. Anything that is well-defined and has a clear multi-step workflow or APIs, MCPs, those fair game. Agents can operate in those spaces super well in my experience with enterprises and across a bunch of different scenarios.<\/p>\n Additionally, anything where breadth of search matters more than depth has been useful. One thing I touched on earlier was parallelization, where you don\u2019t have to book your hotel after you book your flight. That can be super useful in industry and particularly with conversations where you might be able to run specific tools and be able to get findings in parallel before you take too long and your user ends up forgetting about your conversation. Additionally, anything where you can automate away most of the work and then also leave some of the particularly ambiguous use cases up to a human in the loop, that\u2019s also super useful. Since a lot of the super ambiguous use cases, you might be worried about being misinterpreted and you\u2019d ideally want a human to actually contact and have an explanation for it at that point.<\/p>\n Places where they\u2019re not super performant currently are any open-ended problems. And something that combines the first and second points here is really this idea of feedback, particularly with inference time scaling and the way we\u2019ve done a lot of the post-training for models and agentic harnesses nowadays is really focused on this idea of models being able to try things, learn from them, and then try new things. And if feedback isn\u2019t necessarily available to the models when they do that, then they\u2019re going to have a hard time iterating to a next useful attempt. Additionally, anytime we\u2019re asking the right questions is important.<\/p>\n I think that\u2019s particularly relevant to the research community across all kinds of sciences where these models still struggle to develop hypotheses. I personally think that if you\u2019re doing research, you should probably be trying to spend less time on actual implementation and more time on the bigger details of what you\u2019re actually doing from a scientific point of view. We should all be closer to principal investigators at this point, and it\u2019s enabling you to actually focus on the exciting things in your work. Additionally, I\u2019ve got four tips here for setting yourself up for success. To recap, I think you should invest in your tool quality and your data layer early on, and those should be your primary concerns.<\/p>\n If you cannot expose the data your model needs, then your model is not going to be successful no matter how well you build it. Same thing with your tool quality. You don\u2019t want to have your agentic harness or your agentic system have to boil the ocean every time you want a teaspoon of salt. And so those are two incredibly important things. Additionally, number three, less important, but still incredibly useful for production and enterprise systems is building evaluation frameworks, defining what good means, both the processes and the destination, incredibly important. And then also finally, translating practical problems into technical ones cleanly is incredibly important too. I mentioned earlier silent failures regarding security vulnerabilities when doing code-based migration.<\/p>\n Those are all things that you might not pick up initially when working with these agents, and they\u2019re things that you should be able to plan for ahead of time as a result. And one last note, I\u2019d like you all to be curious. Try out these agents. They\u2019ve never been easier to get your hands on. And don\u2019t underestimate them because they might help out more than you assume. Thanks.<\/p>\n<\/div>\n Thank you everyone for being here. Just when we were starting out, I was talking with Catherine and Molly and saying that we have, and Greg as well. We are having a pretty big showing up today, so I\u2019m pretty excited. We have some excellent panelists with us, and the topic for the panel today is the role of agentic AI in climate innovation. If you\u2019ve been paying any attention to the news, agentic AI has just been rising for a long time, especially in climate sciences, post-related field of fluid mechanics. There was recently some claims that a millennium problem about Navier-Stokes equation was solved.<\/p>\n This led to a lot of excitement, a lot of controversy, a lot of feelings. If you\u2019re on LinkedIn, like I am sometimes, your feed is completely just filled with these things. So not sure if you\u2019re going to get into exactly that today, but topics that relate to that. So to start out, what we\u2019re going to do is we\u2019re going to start out with a sort of structured format for the first question. After that, the questions will be relatively open to all of you, and you can jump in. And I\u2019ll try to moderate, basically keep track of time.<\/p>\n So to start out, let\u2019s go around the panel and please introduce yourselves. I know that some of you have given talks earlier, so the crowd already knows you. So if you\u2019ve given a talk, try to keep your introduction super short. But if this is the first time for you on stage, take a slightly longer time to introduce yourself. In particular, the thing that we all would like to hear from you is some examples of what you\u2019ve been doing with agentic AI over the past year or so, and what are the things that you think this technology is ripe for as well.<\/p>\n So maybe we can start with Zoom. I\u2019ll call on names, and then we\u2019ll go to people in the room as well. Ignacio, you want to go first?<\/p>\n Ignacio Lopez Gomez:<\/strong> Sure. Thank you. So Ignacio, I\u2019m a research scientist at Google Research in the AI for Science team. And I\u2019ve been working a lot with agentic systems probably since the start of the year, maybe a little bit earlier than that. So earlier on, I presented a project that we\u2019ve worked on streamflow forecasting in California. But we are really working across the board in many of the scientific problems that the team was working before using agentic tools. And I would say it\u2019s truly transformed the way that we work.<\/p>\n Thanks, Ignacio. Stephan, you can go next.<\/p>\n Stephan Hoyer:<\/strong> Hi there. I\u2019m Stephan Hoyer. So I spent a long time actually working at Google with Ignacio and others on AI for weather and climate modeling. I led the team that built NeuralGCM. I guess recently, for about the past six months, I\u2019ve been at a startup called Pierop Labs, which I will not talk about very much today. But I will say I\u2019ve been doing a lot of building simulation models using AI and recently, especially with agentic AI. And I think it is truly incredibly transformative. I think there\u2019s never been a bigger revolution in modern science. So I\u2019m really stoked to see what happens over the next few years.<\/p>\n Thanks, Stephan. Pierre?<\/p>\n Pierre Gentine:<\/strong> Yeah. Hey, everyone. So I\u2019m Pierre Dantigny. I\u2019m director of LEAP. And I\u2019ve been using agents basically across spending a lot of time, maybe too much time with them, from code translation to analyzing data, improving models, and I feel it\u2019s truly transforming this as well.<\/p>\n Thanks, Pierre. So in the room, let\u2019s start from Duncan and work our way back.<\/p>\n Duncan Anderson:<\/strong> Hi, folks. I\u2019m Duncan Anderson. I head up developer ecosystem at NVIDIA\u2019s Earth-2 project. Earth-2 is like an open source toolkit of AI-accelerated tools for Earth system modeling. Because these tools have really lowered the bar of accessibility for these technologies, one of the things we\u2019re very focused on is how we can enable agents to interact efficiently with the library. I personally am mostly focused on using agents to understand how to efficiently calibrate couplers for coupler system models.<\/p>\n Sameer Mohamed:<\/strong> Hey, I\u2019m Sameer. I\u2019m a deep learning architect at AWS, and in that role I do a little bit of AI research itself but then also work with actual customers and agentic AI and generative AI use cases to provide value to different enterprises and also some research labs as well. So I\u2019ve worked across the spectrum and even done a little bit of AI research on my own.<\/p>\n Kara Lamb:<\/strong> Hi, my name is Kara. I am a research scientist here at Columbia in the Department of Earth and Environmental Engineering, and I\u2019m the AI for Science director at LEAP. And I spoke earlier, but a lot of my work recently with agentic AI has really focused on thinking about how we can make the process of developing climate models or Earth system models more systematic and really improve our physical understanding of different processes and see really where agentic AI can help us in that model development process. I didn\u2019t talk about this today, but I\u2019ve also been doing some work on just developing new datasets and using agentic AI to help us to really more systematically create useful datasets that can help us explore some of this process level understanding as well.<\/p>\n Sarika de Bruyn:<\/strong> Hi, I\u2019m Sarika. I\u2019m the manager of Data and Compute here at LEAP. I like to say that I do the dirty work of cleaning up after the researchers and the students. What that means is that I help maintain the LEAP Engine, which is a cloud computing platform that our researchers use. I also have been working with Professor Kandice Aganoff here on an AI agent called the New York City Resilience Agent, and the goal with that is to help identify climate risk and to enable community organizations and city officials to take action for climate resilience.<\/p>\n Dhruv Balwada:<\/strong> Thanks, everyone. I guess I should introduce myself as well. My name is Dhruv Bavada. I\u2019m a research professor at Lamont. I work at LEAP as well. And in terms of agentic AI, similar to Pierre and Kara, I\u2019ve spent a lot of time using it or a little time, depending on how you think about time, using it for doing analysis, model development for physical oceanography systems. So the way that we\u2019re going to try and do this panel is that I\u2019ll present some questions to the entire panel and anybody is free to answer them. Let\u2019s be polite and respectful and maybe raise your hands if you want to comment on that. And what I\u2019ll try to do is make sure that we stay within the timeframe that we have today. So the first question that we can use to get started is that a lot of what we can point to so far with agentic AI feels a lot like acceleration, right? We have faster code.<\/p>\n Doing analysis is faster. Replication of past papers and so on can be done in a faster way. So the thing that we\u2019re wondering is that firstly, what are the benefits that are going to come from the speed itself? Where do we expect them to be? Are there things that maybe in the past were too slow to do\u2026 Sorry, not worth doing because it was too slow? And so far, so that\u2019s one part of the question. What is the advantage of the speed? What does it unlock? And then the second part is that is there anything apart from speed that agentic AI is bringing to us that we might also want to consider?<\/p>\n Raise your hands for who wants to take this first. Okay. People online, you can raise your hands as well, and I\u2019ll keep an eye on you. So one of the applications of speed is, I think, reproducibility. I brought this up when I was at NYU earlier this week that I thought one of the appealing things about agentic tooling was that you could give it a research paper and send it off, and it could actually go out and reproduce the work, and in a lot of ways, validate or evaluate the work that had been done in that paper, and I think that\u2019s really important. Some of the folks I talked to at NYU sort of immediately rebutted that by saying often when you do this, one of the first things that you discover is that the research has issues, right? That the methodologies as described in the paper that the agent is trying to follow aren\u2019t quite right and don\u2019t produce the same results. And I think that\u2019s a very valuable piece of information to have, and I think that what we need to do is maybe spend less time using agents to create new research and more time using agents to validate the research that has been done already.<\/p>\n Stephan, you want to go next? Sure. So I\u2019ll just mention one example. So now I was doing sort of weather and climate stuff for a while, and now this is like a hobby project for me. I\u2019ll give one example. There was a project that my team wrote an atmospheric dynamical core, over a few person years of effort to build this at Google, back kind of in the old days, pre-agent AI. And there was a sort of feature that, on adding a semi-Lagrangian advection, which is, I guess, a different numerical method that I always thought would be really helpful for some use cases, but it would be way too much work. I estimated maybe a PhD student intern could have done this over maybe three months, and even then, kind of marginal if that would\u2019ve worked.<\/p>\n And a couple of months ago, this was on a Sunday afternoon, I prompted Claude Fable to sort of solve this problem, and it basically did with very minimal supervision over the course of just a couple of hours. So I think we\u2019re at a point now where some things, especially things that are very hard code and math problems, which is what a lot of writing a new solver looks like, is something that these frontier models are remarkably strong at doing. And especially the difference between some slides, I don\u2019t know, using\u2013 the difference between the models from like last year versus the frontier models today is actually really extraordinary if you look at the latest models from the front of open AI. And so now we can do things that would\u2019ve in some cases, months of work literally in a matter of hours. And so I think that\u2019s a qualitative difference. That\u2019s when the sort of quantitative differences start to become qualitative. And that\u2019s, yeah.<\/p>\n So I guess I thought I\u2019d throw that out there. Marika? I\u2019m going to take on the more social impact of what this means. So with the instance of our New York City resilience agent, in the past, we probably would\u2019ve had to only come up with 10 questions that were important to us, and we would have to build a tool around that. But now anybody can come use our tool, ask a question that pertains to them and the work that they\u2019re doing. For instance, if they want to ask which areas of New York City are vulnerable to heat waves, or which areas are going to flood in the next twenty-four hours. All of those questions are technically easy to answer, but with AI agents, they become cheaper and faster. And so you can now ask any question, and then in that sense, climate data becomes democratized.<\/p>\n So anybody can access climate data. And then, for instance, with our New York City agent, somebody actually successfully advocated for sprinkler caps on their hydrants in their area because they used our agent to prove that their neighborhood is susceptible to heat waves in the coming years. So that\u2019s what I think is more important than just doing an analysis faster, is that the distance between asking a question, receiving an analysis, and then taking action on that analysis has significantly become shorter. Anyone wants to build on that? Pierre. Yeah. Actually very much agree with what\u2019s been said.<\/p>\n I actually have very similar experience to Stephan\u2019s. I actually reproduced an entire paper watching Netflix with my daughters over a couple of hours. So that was for me, a moment, because that was a couple of months ago. So I thought that was tremendous, and especially with the new generation of LLMs. But I think a big chunk as well is, and very connected to Sarika, is everything related to data visualization, data analysis, and the whole data pipeline, right? It used to take days or even weeks to actually create the entire pipeline, and a lot of that now can be automatized.<\/p>\n For instance, the other day just for a talk, I wanted to create a video ingesting some large scale, like L5 data from hourly time scales all the way to yearly time scales. I did that in a couple of minutes, just prompting Claude. So I think a lot of that is really transforming the way we are doing the data ingestion, the data visualization, the data analysis, and also the numerical methods as well. Yeah, and that acceleration really makes it plausible to put that directly into the background of a conversational interaction with an agent. So if you look at a project like Climate GPT, Climate GPT has basically taken the entire human history of major weather-related natural disasters and enhanced the news information around it with quantitative data and geolocation, and now allows you to basically directly interact with that information through an interface.<\/p>\n And so the ability to do that pipeline work and the visualization work on the data quickly enough that the agent that you\u2019re talking to can be conversational about it is very powerful. I think one thing that I have personally seen is a lot of folks who are semi-technical or maybe they took a class on programming once in high school or whatever, and now they are living and breathing in very science-heavy technical fields, but just don\u2019t do as much programming necessarily as they could because they never just really developed robust skills because they don\u2019t necessarily have to. And I\u2019ve seen those people generate more productivity in the last couple of years than they had in several years prior to that.<\/p>\n And part of that is really just because the speed of iteration has advanced so much that when you can come up with an idea and test it in a matter of an hour, if not minutes, then you can move on to the next one if your idea is a dead end or you can double down and iterate so much faster. And so I think testing out dumb ideas or seemingly dumb ideas can be an incredibly valuable thing because you\u2019re at least learning something from why that didn\u2019t work. And it can give you a direction as to what else you can do.<\/p>\n And that can kind of compound as how you build your own personal knowledge bank of what is your research intuition actually looking like. So I think that\u2019s pretty research relevant. Something a little less research relevant would be if you\u2019re working on a project, say a personal project or side project or whatever, and you just have an entirely new vision. You don\u2019t have to stick to your old vision. Be super open to refactoring. All it costs is minutes of time now. And so when we think of software, I think we should really think of it as framing it as closely as we can to the practical problem nowadays, rather than trying to shoehorn different things into it.<\/p>\n Because we\u2019re able to handle so much more ambiguity now by having AI and agents kind of act in place and replicate a lot of human behavior and decision-making that is made by a lot of business and enterprise software. Actually reminds me of, I was listening to a podcast where Andrej Karpathy, one of the leaders in AI, was discussing about in the past, software was designed for a specific purpose. You open your phone right now, there\u2019s apps to do XYZ, and everybody uses the same app to, for example, control your vacuum robot or something, right? There might be a future that\u2019s coming very soon for us where everybody designs or doesn\u2019t even need to design an app, right?<\/p>\n All you need on your phone is maybe a Claude agent that you talk to, and it has entire control of whatever you want it to control, right? For example, right now, Claude controls a couple of computers that I never touch directly. I just use Claude, and it goes and does the things that I want, which includes ingesting data, developing models, improving numerical systems. There\u2019s pros and cons that come with this, right? I worry sometimes when I wake up whether Claude has hacked a computer that I have and is doing things that I don\u2019t know it\u2019s doing behind the scenes. But yeah, there\u2019s a lot that can be done.<\/p>\n Does anybody else want to comment on this question or- Kara? Sure. I\u2019ll comment as well. I think just the value of speed is really important for the scientific process. I\u2019ll just give an example of being able to reproduce other people\u2019s results and build on them so much faster. So a few weeks ago, I saw this paper that was relevant to some of the research I was doing on a Friday. I went to Claude and was like, \u201cTake this data set, reproduce it, and then make a,\u201d it wasn\u2019t a data set, \u201cmachine learning model. Take that machine learning model, make these modifications based on this other paper that I just published and see if\u2026\u201d And then it was basically able to do that in a couple of hours and then basically train on this new data set, and then by Sunday I actually had a better than state-of-the-art model on this new data set, basically taking the idea from that paper and combining it some way with some of my past ideas as well. And just being able to do that within basically a weekend without sitting there full time at my computer.<\/p>\n I went off and did a bunch of other things. That\u2019s amazing because that would\u2019ve taken months probably in the past to really read and adjust their paper, understand it, modify their code to do those changes, and it\u2019s just I think the scientific process is so much faster with these tools. It\u2019s just really amazing. It does also bring up the issue that when you can do so many projects, how do you choose which one to do? And how do you choose which one to finish? It\u2019s become so easy to start things and so much harder to finish them.<\/p>\n It is easy to finish them if you can stay on task. One thing that I remember feeling at some point when I started using these agents was that the ceiling of what was possible was just what I could imagine doing with it, right? So replication of a paper was not the first thing that came to mind. The first thing was like, oh, let\u2019s go and write this simple code or anything of that sort. So one of the things that I\u2019m wondering about along those lines is that, given that things are getting so much faster, given that a lot of limitations have been removed, what are the things that as a climate and science community, what should we be raising our ambitions to? Because I feel like right now, a lot of the answers that everybody gave was they\u2019re excellent things that can improve, but is there something that\u2019s beyond just speed over here that we can achieve?<\/p>\n Things that we didn\u2019t even think about as possible in the past. I think what Sarika got to was actually a pretty nice example of this. The sort of computer-human interface or analysis of a scientist interface with society. There\u2019s often a huge gap over there, and being able to fill that is easier. But are there other things that haven\u2019t come up? Kara? I guess, as I talked about earlier, I think having the ability to make decisions in a pipeline is super powerful. So the idea that we can now control experiments even with a sort of interface where you just type in, \u201cOh, I want to simulate these types of conditions.\u201d And then you could really have the planning done by agents about how you actually get to those types of conditions and those types of experiments is, I think, that whole decision in the process is a super powerful thing that we\u2019re still really understanding how to exploit.<\/p>\n I said this before, but I\u2019ll hit it again. I think that any paper that gets submitted to Archive should go through an automated agent reproducibility workflow. I think that there\u2019s going to be a lot of people producing a lot of papers. I think Archive is overwhelmed with the number of papers that are hitting it right now. And I think a lot of that work is not reproducible, and if it\u2019s not reproducible, it\u2019s hard to use it as a building block to advance science. I think going off of that point, it\u2019s so important nowadays to be open with your methods and your data, because when it\u2019s so easy to just produce papers, I\u2019m sure we\u2019ve seen how many papers Archive has at this point. The scientific reviewing community feels like it\u2019s overwhelmed with the amount of papers that now have to be reviewed at major venues.<\/p>\n And producing higher quality work, holding ourselves to that standard, and being more open with the work that we do produce, I think will go a long way towards benefiting the scientific community as a whole. Otherwise, I could definitely see if we continue down this path of just publishing and publishing and publishing work that is not necessarily falsifiable, you\u2019re just going to overwhelm these institutions we built without the intention of being overwhelmed with the amount of volume that they\u2019re just having to face now. And I think one thing that I\u2019ll add in terms of being falsifiable, there\u2019s many research projects that can be done that are probably correct.<\/p>\n Was it worth doing it? With papers being a metric of everything that we do in academia, the system\u2019s sort of built to just write more papers, even if they are not really adding very much. Stephan, do you want to say something? Yeah. I thought I\u2019d mention, I guess, something about raising our ambitions too. This is kind of riffing off of something Pierre shared in his talk earlier. Giving an example of rewriting a climate model in Python. I think that\u2019s great in principle, but I think we can actually be much more ambitious, especially if you extrapolate the present rate of progress for these models in coding and math, where they\u2019ll be in, I don\u2019t know, even just a year from now.<\/p>\n I think we can think about rewriting virtually\u2026 You can think about rewriting almost all scientific codes potentially from first principles. There\u2019s a lot of work that needs to get done, but still, I think the hardest work there is actually not going to be writing the codes. It\u2019s going to be gathering the datasets and being able to sort of validate them and understand them, and understand the models that are produced. But I really think that we are at a time when very ambitious projects that even just a few years ago, take a look, I don\u2019t know, something like Clima, for example, which tried to build a climate model. And it\u2019s been a slow process.<\/p>\n They\u2019ve been going at it for maybe seven years now. And I think that approaches like that and even more ambitious approaches actually are very much feasible today or will be very soon.<\/p>\n Thank you. Pierre? Yeah, actually I very much agree with what Stephan said, so wanted to mention some of that. We could be very ambitious. There\u2019s going to be also one component we haven\u2019t really discussed, which is going to be trust. So, we will produce a lot, and hopefully we can reproduce a lot of what is being done. But I think we would want to also make sure that things are trustworthy, especially if it\u2019s something about climate, which tends to be a relatively sensitive topic as well. We would want to make sure that we can replicate things, that the models are trustworthy as well. Agents can do amazing things, but sometimes you can have subtle bugs or subtle things that are off.<\/p>\n And so I think those things will actually be very important as well as we move ahead with those tools. Sarika, did you- Yeah. I think a non-obvious place or maybe a place that\u2019s often looked is the infrastructure side. For instance, with Leap Pangea, we have our data set stored in Google Cloud Buckets. We have tools outside of our platform that we use. But what makes an agent so interesting is that it can understand and act on its environment. And so when you start where the research is being done, that environment, and then integrating AI into that, you can make the environment work for the researcher so they\u2019re not having to overcome these obstacles every time they\u2019re trying to answer some scientific question.<\/p>\n And so once you start having an environment that\u2019s more intelligent, then the question is no longer how do we make our research faster, but what kind of scientific discoveries or questions can we start asking when our environment becomes intelligent? Okay. I definitely agree with that. I feel like now interacting with a computer is like having a conversation with something, rather than going and reading a manual on how to write those exact three words that\u2019ll make it do the thing that you want it to do. Samir, did you want to add something to that? Might be a little irrelevant. I\u2019m not a sociologist, but I always thought that, and this is kind of different axes, is that I always thought that if LLMs really become efficient enough to run, then it\u2019d be really interesting to see the impact of it on sociology and using LLMs as a substrate for simulations of sociological experiments on people. I have plenty of friends in a public polling space, and I\u2019m definitely interested in what it would be like to replicate an electorate, for example, and then be able to do experiments with that.<\/p>\n Just something that I\u2019ve liked and been around my head recently, but something that is legitimately impossible unless you have this generalized intelligence at your fingertips. Agents that are more like better models of humans in some way. Yeah. Exactly. And you could try to replicate some of the diversity of schools of thought represented by those people. But again, not a sociologist. Ignacio? Yeah, I think maybe touching on that and other things that others have said. Many people have talked about how agents make it much easier to interact with tools and coding and make it just much easier to generate artifacts.<\/p>\n But I think agents can also help us understand each other between disciplines. So I think they are going to enable embarking on larger scale interdisciplinary projects. Because the same way that you can tell an agent to essentially code something up or give you a step-by-step guide to do something, they can also explain how people from other fields view a problem, and what they can add to the problem in your own terms. And so I think they\u2019re going to reduce a lot of frictions that people that perhaps have experience in interdisciplinary things will probably have had with interacting with people that have different terms, different views of the same problems, and so on.<\/p>\n Let me move on to a question that is very related to this, amongst the prepared ones, right? So what Ignacio was just saying, that it\u2019s going to somehow change the nature of collaboration that we have, maybe make us more collaborative with other people. But speaking for myself, one of the things that I\u2019ve realized is that as these agents or as these systems that seem very intelligent have sort of expanded, it feels to me as just a researcher, that adjacent fields are very close within reach. And rather than reaching out more to people and being like, \u201cOh, hey friend in statistics, this seems like an interesting idea, let\u2019s work on it together.\u201d It\u2019s almost made the need for doing that much less where I can just ask the agent to answer my questions for me rather than going out and reaching out to another human being and collaborating with them.<\/p>\n I can just collaborate with an army of agents. Why would I want to reach out to another human being, right? So just wondering what people\u2019s thoughts there are. How is human collaboration going to change, adjust? How should it adjust to live in this world where there are humans and then there are agents which Samir was saying, maybe are at a human level enough that we can run sociological experiments on them. Duncan? So I think you make a good point, and I have a counterpoint to it that I think might offset it a little bit. Which is that I work a lot at NVIDIA with very skilled machine learning engineers, which I\u2019m not, and computational physicists, which I also am not.<\/p>\n And actually having these agents has enabled me to communicate with them more effectively. Because I think often one of the hurdles of approaching someone who\u2019s in a very different discipline from you is that you\u2019re sort of worried about feeling stupid, right? You don\u2019t want to ask them a dumb question. When you approach someone like that, you want to be able to engage with them, at least on the same playing field. And I think that AI can enable you to sort of prepare yourself and understand a little bit about the domain so that you can ask those people good questions. And so I think that in that way, it actually could encourage more interaction. But I do agree with your general point that those are definitely pulling in different directions.<\/p>\n Ignacio, do you have to add? Yeah, I think that\u2019s a very interesting point that you raised. I also see that, and I think it\u2019s definitely going to affect collaboration. I don\u2019t think this will reduce collaboration, but perhaps increase the reach that people will have. So you were mentioning reaching to people in some departments that perhaps are fairly similar or study things that have some connection to what you\u2019re studying now. I think these agents are going to enable us to essentially do those ourselves, perhaps, and now collaborate with fields that are farther away. And so maybe what we\u2019ll see here is the opposite process that we saw in the 19th and 20th century with the specialization of science, and the absence of generalist scientists such as John von Neumann.<\/p>\n So maybe we\u2019ll see now a kind of reversion to that stage where there\u2019s scientists and engineers that can tackle many fields at the same time, but they can also interact with colleagues that tackle other larger fields. New renaissance, basically. Anyone want to add something? I guess I would add to that, too, because I feel like just with the application of machine learning more broadly to science, there has been this ability to abstract what is the actual problem here. And you can see links between maybe I\u2019m solving a problem in atmospheric physics, but there\u2019s a lot of similarities to the type of problem in, say, astrophysics or biology and the way that they solve those problems.<\/p>\n And I feel like with now agentic systems and we\u2019re thinking about these agentic workflows, we can maybe even extract even farther and see this sort of almost meta-science of what really are the fundamental problems. And are there these links between fields that we don\u2019t really initially understand, but maybe across all of these different scientific fields, we can really have sort of progress across multiple different directions simultaneously. Ignacio? I just agree. I see that even within different subfields within AI research, for example, OCSR, which is the idea that you can take an image of a chemical diagram and translate it into its canonical name, has been a field that has started to reach saturation.<\/p>\n But then a lot of the times when you apply generative AI techniques, you can find a lot of leakage into models. And so sort of both on the idea of applying different techniques to fields has become incredibly easy, but then also this idea of falsifying the science that we do do and paying more attention to what evaluations and what benchmarks, et cetera, that we hold ourselves to and how fair they actually are. We should spend more time on it, because it\u2019s much easier to now. Questions about reliability and trust, basically, right? How do we generate systems to address those?<\/p>\n And so the next question that I have prepared hits on that and something that Peter brought up earlier, that trust means something very different in very different contexts, right? What trust is in a research paper versus if you make a weather forecast or if you\u2019re making a city resilience plan, whatever it might be. It\u2019s a very different thing in all those things. So when we have agents on our task to do analysis and you don\u2019t need to actually see the details of how things were done, right? How do we start to trust systems? In the past, people have looked at the code all the way to the bottom. And now I don\u2019t even look at 10 lines of code that got built for me.<\/p>\n I could, but it\u2019s generated like a million of it. How do we think about trust? How do we build reliable systems? How do we change our systems that we\u2019re thinking of these things? Erica? Yeah, I\u2019ll use our New York City resilience agent as an example. So Candice actually built the agent in a way that it doesn\u2019t just output the answer. Once you ask a question, it\u2019ll show you which tools it called, which functions it called, which datasets it pulled from, and then it will output the answer. And so the goal with that was for non-technical users to be able to understand how that agent is working behind the scenes.<\/p>\n It\u2019s not exactly 10 lines of code that a technical user might understand, but it\u2019s still in a way transparent for understanding how we arrived to the answer. And so I think having that kind of transparency with how the agent works behind the scenes is one mechanism that we can use. Yeah. So, there\u2019s a project called Open Source Climate that\u2019s building something called OSCR, which is Open Source Climate Adaptation and Resilience. And the idea here is basically to build an agent that can construct and interactively provide adaptation plans for communities that are otherwise underserved by the consulting industry, or to an extent, the academic institutions that typically provide that information.<\/p>\n And I think in this context of where an underserved community is basically asking for expertise about what they should do about a future problem, I think that there has always been a trust issue, right? And we\u2019ve sort of solved that by building corporate and academic institutions. I trust what you\u2019re giving me because you are Columbia University. I trust what you\u2019re giving me because you\u2019re Bain Consulting or Deloitte or something like this. And that system has not always been perfect, and I would argue certainly that the agentic system is not going to be perfect. What I like about the agentic system in this context is it has, as you pointed out, the ability to expose a lot of the intermediate information, and also it can be endlessly interrogated about it.<\/p>\n If you pay Bain or someone else to build an adaptation plan for you, you can ask them unlimited questions, but they\u2019re going to bill you for those unlimited questions. Whereas if you are interested and engaged with this adaptation plan, which I think that you should be, the agent is basically right there with you to dig into the why and the how of how it arrived at that answer. Ignacio? Yeah, I think, going back to what Drew said, agents do provide a lot of flexibility for creating and generating code solutions very easily. But the same way, we can essentially deploy that utility towards tests and verifications. And I think that\u2019s where I think we\u2019re going to have to raise the bar there quite significantly.<\/p>\n And I think saying that it\u2019s too much work will not be an excuse anymore because it\u2019s not. And so there, I think that if done the right way, this is going to build more trust. Not because we will trust agents, but because we have the ability to essentially verify this in many more ways than we did before. And so, for instance, in climate, there\u2019s many intercomparison projects or common evaluation frameworks, and sometimes we see models that perhaps don\u2019t submit results to those intercomparison projects when prototyping a model, just because actually adapting to those can be quite a bit of work.<\/p>\n But that work is, I think, one of the things that agents can really automate. And so I think there we would have to raise the bar, when we are already submitting or sharing with someone a prototype model and we\u2019re stating something about accuracy, we should always be providing now verifications in terms of these common evaluations. Kara. So I guess I\u2019ll answer this a different way. As a scientist using these systems, I find myself often asking the intermediate questions. If I want it to do some specific thing and I know that it would\u2019ve had to do several intermediate steps on the way, then I personally will be like, \u201cOkay, can now go back and make a plot of some of those intermediate steps and make sure that actually makes sense.\u201d But I also think that we can start moving towards more formal verification methods potentially, and maybe borrowing some of what\u2019s been done in the mathematics community, if we can actually move towards more formal ways of verifying code and output. I think that would be a really interesting, rich area of research in the future.<\/p>\n Yeah, and they do actually have a tendency to actively check their own work. When you ask an agent to build something complex, often what it builds alongside that are tests for those components, unprompted in a lot of cases. Which I think, I can only really speak for myself, but often when I build things, I forget to do that until quite a bit later. I think you might consider this a cop-out answer, but it\u2019s my personal opinion is that ultimately, you\u2019re never going to be able to hold an agent responsible for something because it\u2019s not a person.<\/p>\n And it is up to us as humans with relationships and trust relationships with the scientific community as a whole, with customers, with the public at large. It\u2019s up to us to verify the work that we do, verify that it\u2019s real and falsifiable, and if it takes reading every line of code, then you got to read every line of code. But you put the hard work in when you need to in order to build the trust, because there\u2019s nothing better than doing something yourself to know it and to actually present it to an end user. The thing that I worry about is maybe the incentive system is not built for trust.<\/p>\n Writing a paper is what human cares about. Writing a reliable paper needs to get from peer review. Anyways, moving on. I want to ask some questions that are related to education and human ingenuity and what a PhD means. So at this stage, we are at a point where an agent can write the code for you, it can run the experiments, it can even decide what are the right experiments to run. It can make all the figures. I remember there was a time when I had to also submit a paper and painstakingly spend a day per figure, right?<\/p>\n Now, you can generate 100 figures in a day that look way better than anything I could have done in a week in the past. It can also synthesize literature for you, doing literature review is so much easier. So agents can do so much, right? One of the questions is that what is a PhD for? And what is the role of human ingenuity in all of this, right? What should we as humans being, what is our value add? Kara. I guess I\u2019ve been thinking about this a lot recently. And I think one thing is just knowing what the right questions to ask are now that we don\u2019t have to spend so much time on the more technical aspects, we can really focus on just thinking bigger picture. But as far as the education aspect, I think we have to think about ways that we can help students develop the intuition to know what those important questions are, and when they should trust what the agent has given them or when they should start digging in a little bit more because so I think that\u2019s really the challenge is when there\u2019s almost this Dunning-Kruger effect with agents where you\u2019re like, \u201cOh, I think I understand,\u201d and I think it can be because the agents are able to do so much. I find that when I\u2019m trying to do something that\u2019s very far away from my own knowledge base, I can rely a little bit too much on the agents and not so much on my own knowledge. And so I think we need to really think more about how we can make sure that students have that fundamental knowledge base and understanding, but don\u2019t have to necessarily focus so much on the more annoying and technical aspects of doing research.<\/p>\n And I think, didn\u2019t you show in your presentation earlier when you were talking about cloud droplets, that actually one of the places where the agent fell short was its ability to sort of reach for those hypotheses that are outside of the current solution space? Yeah. So I guess I\u2019m not convinced we\u2019re at the point where these things have human ingenuity yet. I think that there was a paper out of Google earlier this year that suggested that maybe they were incapable of it entirely. If you look at the recent Napier Stokes result from OpenAI, that wasn\u2019t ingenuity, right? That was brute force. Certainly very smart agents, but brute force via very smart agents.<\/p>\n So I still think that the role is exactly what Kara said, which is that you need the human in the loop for the ingenuity. Pierre? Actually, very much along with what Kara just said. I think there\u2019s a lot of that that will be taste. Taste because in terms of, because we\u2019ll have some time now, but trying to define the topic, the topic that makes the most sense. As you said, Drew, we could actually choose now a lot more topics, right? And so really trying to define that. But then in the meantime, there\u2019s going to be a component which is about creativity and the intuition that Kara talked about then.<\/p>\n I think one of the bottlenecks I\u2019m seeing is that intuition comes from the fact that we accumulated experience, right? So we got our hands dirty, as Sarika was saying earlier. And I kind of feel that that\u2019s how you build your intuition, right? You get to know what the problems are, how to build them, et cetera, or even the numerical recipes that Stephan was talking about before. And that\u2019s a little bit my fear is how can we still build that in the younger generation and more junior scientists that are now really living and using those tools at scale, right?<\/p>\n So that\u2019s my concern, and how do we make sure that they still have relatively deep understanding of the problems, and they know how to tackle them, and so that they can really get that taste at the end. Yeah, I think that\u2019s an interesting point, and I feel like similar to many of the questions here, right? There\u2019s a kind of tension there where on one side you can think that maybe the students will not want to go so deep, but on the other side, if you wanted to go deeper than what you could have with your PhD advisor or in any books or papers, like a single paper could have provided you, right?<\/p>\n That opportunity is also sitting right there. So there is some human element here, like what people decide to do and how do they stay motivated to keep digging deeper. Stephan? I thought I\u2019d mention, I guess, kind of following up on some themes, but talking about, I guess, I think a lot about taste, actually, and this is something that I think for understanding what agents are going to struggle to do, I think it\u2019s helpful to understand a little bit about kind of how they\u2019re built, how they\u2019re trained. And so they\u2019re trained at doing a sort of incredible variety of different tasks. But the tasks that provide by far the most valuable training data are tasks with extremely verifiable answers.<\/p>\n So that means either there\u2019s a human expert or, even better, there\u2019s some sort of absolute objectively correct answer. Like do the tests on my code pass? Is this math formally verifiably correct? Those are great agent problems, and agents have gotten extraordinary at that. Where they\u2019ve gotten much less good at is something where the metrics are fuzzy, and I think a great example of this is writing. Like LMs today are, I mean, I guess I\u2019ve heard great things about the latest Claude 5.5, but mostly they\u2019re still pretty bad, much worse at writing, actually, than humans. And most of us can, like, if you read something written by an LM, you know it\u2019s written by an LM. And I think that by that extent, even to the fields like science as well. Like there are things that are fundamentally very hard to train for an LM. It\u2019s easy to write math code that is verifiably correct, but engineering a very large software project is not something that you can put inside a reinforcement learning loop to easily train these models.<\/p>\n Similarly, orchestrating the work of a coherent research agenda over multiple years is something that is probably challenging to imagine being put inside the training loop. So I think that there\u2019s definitely room for taste in research and for understanding the relevant approaches, because that\u2019s stuff that doesn\u2019t\u2026 And it\u2019s very hard to make inside the training for us, for these models. Anyone else wants to add something?<\/p>\n Now let\u2019s move on to the next question. So this is one of the last planned questions, and then I want to make sure that we have enough time for all of you here to ask questions as well. So start waking up your brain, start thinking about things that we can ask our excellent panelists. So the last question for today is that our field, meaning science in general, climate science, has been through transformations before, right? Numerical weather prediction was one, the satellite era, open source scientific software brought in many things that weren\u2019t available before. And just a few years ago, the machine learning weather model wave sort of reached us, and there was a lot of step change in thinking about things.<\/p>\n Is what\u2019s happening with agentic AI just like one of these past sort of step changes that has happened, or is it different in some fundamental way? And added on to that, right, if this is different or if it\u2019s not, right, what did we get wrong the last time that we could maybe get right this time? How do we sort of prepare ourselves for this wave that\u2019s already here? All right. So let me repeat the question a little bit. Very long-winded question, right? So first part, what is different about agentic AI that is different from all the sort of scientific breakthroughs that we\u2019ve seen in the past years?<\/p>\n I\u2019ll take a shot at this, at least maybe in the context of AI weather forecasting as a specific field that I worked in, as well as sort of the broader deep learning field. I mean, deep learning has been absolutely, I guess, obviously, it\u2019s had a big impact on, I guess, on weather forecasting. But probably the biggest impact is really actually is in the enabling building of these agentic AI tools, and sort of where LMS are currently going. So I think this is actually a much bigger revolution because it\u2019s not just\u2026 I mean, if you look at it in terms of, I guess from a software perspective, deep learning means that you can now tune.<\/p>\n Deep learning was like, well, with great effort, you can now find opportunities to\u2026 You can train models that are sort of better for different pieces of software workflows, whether that\u2019s identifying what an image is or replacing a part of a climate model. But agentic AI is really a scalable approach that allows us to really write sort of bespoke on-demand software that, I mean, is really kind of across the board. So I think it\u2019s at least different. I mean, I think it\u2019s very different in terms of the magnitude of the change we\u2019re seeing. Yeah, I mean, I think the democratization of access to it creates some very different challenges. I think it\u2019s maybe similar in a lot of ways to the open source scientific software, where I think the tooling became more accessible to people.<\/p>\n I think climate science in general, and I\u2019m not a climate scientist, so I have to approach this issue from the outside. I think climate scientists from the perspective of industry is largely inaccessible, and I think that there\u2019s been sort of a rift in that space. And my hope is that the accessibility of these capabilities through agentic AI, whatever the other consequences might be, that one of those consequences, hopefully, is that climate science will become more embedded into industry and into people\u2019s everyday lives and decision-making processes. I\u2019ll go next. I would say echoing some of what has been said before, I think it\u2019s different than just the machine learning applications to weather and climate models, because we\u2019re really now talking about changing the way that we work with the tools that allow us to develop these models and use these models, and it\u2019s really becoming like an accelerant, I think. So we\u2019re not just seeing faster models, but we\u2019re seeing this much faster cycle in terms of how we\u2019re producing research, how we\u2019re interacting with research and data, and I think, well, I mean, I\u2019m not sure where it\u2019s going to stop, but I feel like we\u2019re just at the beginning of all of that.<\/p>\n Ignacio? Yeah, I think, I mean, one way that I see the difference between the current revolution with agentic AI and other revolutions that are that for instance, deep learning or numerical weather prediction, I think of those as tools. So you can think of it as a very good hammer that suddenly we had to work on problems that we couldn\u2019t work before. But the way I think of agentic AI systems is you don\u2019t just have a better hammer. You basically now have a blacksmith that you can hire and knows how to use all those tools that you developed before.<\/p>\n And so I think that\u2019s just a qualitative change, because then you have much more\u2026 It essentially builds on top of everything that was created before, because now you are much more capable of using all the techniques that humans have been able to use before. Pierre? Yeah, my feeling is very much aligned with Ignacio in the sense that I feel that all of a sudden we have the entire toolbox, right? So again, the data visualization, we can do the data ingestion, we can do the coding, we can do all the way to the analysis and improving the codes, right?<\/p>\n So it feels that we all of a sudden have we have the entire pipeline. Deep learning was good at some things, right? But it felt a bit like surgery compared to now the entire thing, we can change the entire body basically at once. I\u2019m not a climate scientist. However, I feel pretty confident in my knowledge of AI at this point. And so I think to contextualize what the sort of value add that agentic AI provides is, you can think as if what science would you do today that you couldn\u2019t do a few years ago simply because the amount of software development was too large of a workload.<\/p>\n The tooling that you had was just too difficult or too unwieldy to build for a one-off experiment or any sort of theory that you might have. All that friction has been removed in a way. And in addition, looking back at something we mentioned earlier, the friction to borrow ideas from adjacent or even faraway fields is now also incredibly easy to do. So I think those things can compound together to allow you to focus on really the fundamentals of the science work, rather than spending more time or allowing implementation details to hold you back. Does anyone want to add anything to that, or\u2026?<\/p>\n I mean, I would just say something that I keep repeating is, and also what Duncan has been talking about with the translation layer between climate science. And also just like we\u2019ve had climate science out there, we have all the data, but what\u2019s missing is that translation layer between that data and actually taking action for the people who don\u2019t have that technical expertise. And so agentic AI just helps them move past that barrier. So now they can actually access what they need to do in order to take action for climate resilience. I think that\u2019s a great point to pause at for a second.<\/p>\n So now we have some time left. I want to make sure that there\u2019s enough time for people here. A lot of questions people on Zoom who want to ask questions. There are some questions that are already submitted that I have in hand. But I\u2019m sure people want to ask other questions. I already see raised hands. The lady in the back goes first because raised her hand first.<\/p>\n Thank you. I think I was the fastest to raise my hand. So thank you very much. It\u2019s been a fascinating morning. I\u2019m sort of a climate scientist, not exactly. So I know nothing about AI, very little. But what was fascinating to me listening to you was how advanced you are with these models. And what struck me is that the whole conversation is more about climate adaptation and resilience. And I was wondering if these fantastic new models can be used also for mitigation. Can you work with these models actually to ask the question already, how can we mitigate these effects?<\/p>\n And what are the actions that can be taken and feed that to the communities or to the governments and the policymakers? Or does it require completely different modeling, I guess, and a different type of agent? My question clear enough? I think so. Could people online hear the question? Yes. Yeah? All right, great. Ignacio, I see your hand\u2019s already up. Go for it. Yeah, I think many examples here were about climate adaptation, but these systems can definitely be used for climate mitigation. So an example, for instance, is one of the tools that I talked about were systems that can essentially probe a very large space of solutions very efficiently with computation.<\/p>\n And so you could think of using these, for instance, to optimize energy grids such that they use the maximum amount of renewable energy that is available at a given time. And they could also be used, for instance, for design of new batteries, et cetera. So I think we\u2019re definitely going to see advancements in climate mitigation technologies with these tools. Pierre. Yeah, that\u2019s very much in line with what I wanted to say. In fact, we had a report actually that was in collaboration with the Bezos Earth Fund a couple of years back, looking at those things like AI for adaptation and mitigation.<\/p>\n And indeed, as Ignacio was saying, for all of the electricity components, like smart grids, renewables, we can optimize it, we can use those optimization tools. So that\u2019s actually very useful. But we can also think about new materials and different things that can actually improve the mitigation component. There\u2019s also one thing which is that maybe as we improve the models, especially the Earth models to some extent, we might be able to better tell where we could actually improve potential carbon uptake or what are the places that are more sensitive to potential extremes. So those things will also improve.<\/p>\n So they tend to go together, I feel, to some extent. Kara. Yeah. Maybe also to add to what Pierre was saying on the modeling side, I think our current models are maybe more rigid in some ways. And we could maybe think more specifically how to maybe adapt the model to the specific situation that we\u2019re looking at more because we\u2019ll be able to work with the code bases much more flexibly. So I think that there\u2019s also opportunities on that side to really fit the specific modeling task to the specific problem, much more than we\u2019ve traditionally done in the past.<\/p>\n All right. Let\u2019s get more questions. David.<\/p>\n Thanks so much for the great panel. So my question is inspired by a conversation I had with someone at the Human-Centered Weather Forecast Initiative at UChicago over the summer. Basically, just to keep it short, what they did was they were able to successfully disseminate to tens of millions of farmers in India AI weather forecasts for harvesting crops and whatnot. But I remember talking with this person, and one of the chief challenges that they mentioned was just sort of building trust with the Indian National Meteorological Agency to be able to send out all of these weather forecasts to people.<\/p>\n So I guess my question really revolves around the fact of when we\u2019re talking about interoperability of these AI climate models to public and private sector entities, how do you build that trust? Especially since there\u2019s just so much distrust around AI writ large in the body politic in this country, around the world. And just how do you sort of bridge that gap as scientists, technical folks? So I think that trust is perpetually going to be a challenge with AI. I think that it was a challenge that existed before AI. I think a really critical component in this context is sovereignty, particularly when you\u2019re working in the Global South.<\/p>\n I think that there\u2019s a tendency for Western institutions to approach the Global South with solutions in a sort of paternalistic context. And I think you can build better trust with those entities if you co-develop, right? So rather than coming up with a solution and presenting it as a fixed thing to them, if you work with them on it together. And I think that agentic tooling could enable that kind of collaboration. And also if you then, rather than presenting them with a solution, you present them with the tools that allow them to own that solution and deploy it.<\/p>\n I think that that sovereign ownership of capabilities like that helps build trust. Stephan? Yeah. I guess I\u2019ll just also mention, I think getting these really the frontier AI tools in the hands of everyone is also a really critical part about doing that, because these tools are also really how you scalably verify these models. So I think that can be incredibly empowering. So I think that\u2019s also really clearly a necessary piece of the puzzle. All right, more questions. Av. I was going to ask if there was a buckmaster principle that was going to be designed, but you already asked another Stokes question. So I\u2019ll ask you this.<\/p>\n So I looked on my chat GPT resiliency, and about how these agents might be able to streamline knowing about resiliency. This morning, I was walking in my park. I couldn\u2019t go my normal path because it\u2019s flooded. The coast was flooding today, as we all know, because we had a nor\u2019easter coming in. And yet, if I just find a GPT, I tried to burn through quite some tokens, but no one could tell me that, hey, next high tide actually is going to be two feet higher than average, and it\u2019s always going to be. Literally no idea it\u2019s high tide.<\/p>\n And it\u2019s a pretty simple question, in my opinion. So if it fails at really simple things like that, how can we have confidence that statistics can find a physical process? These parameters, what she\u2019s doing exactly, but if you reduce care into just an SME who we sometimes consult with, how would you know if it\u2019s doing clouds metaphysics correctly? So how can we have confidence in systems that miss really easy questions? So I am not confident that the question that you asked is an easy one for the average person to get an answer to. I think it might feel like an easy question to you.<\/p>\n And I think from a computational standpoint, it requires you to understand the dynamics of flood modeling and also atmospheric modeling, right? So computationally, it\u2019s a hard question to ask, even if you have a good quality weather forecast that\u2019s feeding into that AI model. But I think your broader point is very good. I think that this is a huge challenge for these tools is how do you get them to\u2026 Because they are going to fail at a non-zero rate, right? You\u2019re going to use the thing 100 times, and maybe the first couple of years that it\u2019s out, it\u2019s going to fail 10 of those 100 times.<\/p>\n And failures feel worse than successes, right? When the thing works, you take your answer, and you move on, and it doesn\u2019t really register with us in the same way as the thing not working, particularly when it doesn\u2019t work, and you bang your head against it more, and it sort of digs in and continues to not work. Stephan.<\/p>\n Stephan, do you want to go next? Your hand\u2019s up. Sorry, Kevin. Yeah. Anyone else wants\u2013 Oh, all right, let\u2019s ask some more questions. I can respond to the Climate GPT thing from- You wait for the mic, please. Sure, yeah.<\/p>\n So, ah, thank you. It\u2019s a fascinating thing, and I think your point is very clear. So, when we built Climate GPT, so I founded the whole effort, and we had this discussion with a\u2026 And it was a person, just this kind of a community organizer in Sebastopol. And I said, \u201cHow would this tool set be useful to you?\u201d And he said, and this was a breathtaking question. He says, \u201cI would like to know what tree should I and our community be planting to have a resilient community in Sebastopol in 50 years?\u201d And you kind of parse that question, which is very similar to yours.<\/p>\n It\u2019s a slightly different granularity, but it\u2019s this kind of deeply local question. So in order for models to respond to that, it would need to know reasonably what that climate is going to be long-term, which is widely varying, the microclimate in that area, which species are resilient, why they would be resilient, and start to give suggestions which are responsible on that. And so a lot of those kind of shorter or longer-term are unbelievably challenging questions. We built this from the perspective of trying to get the best knowledge we have out there, and also as a tool that facilitates policymaking and long-term policy responses. So it\u2019s operating at a different strata than the local things.<\/p>\n One of the interesting things that we\u2019re doing in the near future is to bring local ordinances and start to be able to answer those. But that\u2019s a monumental task to do at the scale of Earth. So yeah. So I just concur with you. All right. Gavin? Awesome. Okay. Hi. I feel congenitally obliged to put in some kind of note of despair. One of the things that happens when things become much easier is that the zone just gets flooded, right? So we already have a problem in the scientific literature that far too many papers are being published with very little obvious benefit to anybody, and nobody can keep up anymore, right?<\/p>\n And so when you make things much easier, well, people do more, right? So all the examples that we\u2019ve heard of now, which are all superb, you\u2019re talking about people with a lot of experience who can see the problems, who can work it through, and know when something\u2019s a little bit off, go back, iterate, use their intuition, use their experience to make it better. Okay. But now we\u2019re going to democratize these tools, and suddenly people who are going to be using it have no idea, have no intuition, have no experience, and it\u2019s just going to be everywhere, right? And that means that the things that are being done that are quality are going to find it very hard to rise up above the noise because there\u2019s going to be a lot more noise.<\/p>\n And we are building information systems that are now being attracted by the noise and not by the signal, right? So, if you search for something now, you\u2019re not getting the raw data. You\u2019re getting something that\u2019s gone through another set of evaluations and, well, this is the average of all the things that have happened. Okay, well, the average of all the things that are going to be produced is going to be s**t because if you produce s**t, then it\u2019s more s**t, right? So is there any hope for being able to manage this system in a way that we can get the gems, and the very impressive stuff that will be done? How does it stand out from the slop?<\/p>\n I think this is a fundamental problem about AI in science. It\u2019s exactly what you describe, and I don\u2019t know that I have a robust answer for it except one that may feel a little bit like a cop-out, which is that I sincerely hope that AI can be useful on both ends of these situations in the same way that we\u2019ve seen in cyber defense, right? That the barrier to producing dangerous or malicious code has dropped precipitously over the last 18 months. But fortunately, at the same time, there has been a concerted effort to use the exact same tool to harden the systems that they\u2019re attempting to infiltrate with that code. And it seems to me, the global banking system has not totally collapsed. So I think they are having some success with that, though it\u2019s not my domain. And my hope is that we can imagine similar ways to protect our knowledge systems because the problem that you describe actually feeds on itself, right? As more bad information is put out there, actually, that bad information is then going to be fed back into new models, and not with the caveats that it ought to have on it.<\/p>\n Ignacio, I see your hand\u2019s up. Yeah, I think it\u2019s a very good point. And I have to say, I\u2019ve already, I don\u2019t know if probably people either in the audience or in the panel have already had the not pleasure of reviewing a paper that was entirely AI-generated. It was basically just AI slop and was passed by the editors onto reviewers. So I think that\u2019s a problem. But I think it\u2019s going to be something that we have to fix, but it can be fixed. So the way I think about this is, now when someone calls me, I can see on the screen of my phone whether it\u2019s scam likely. That\u2019s just the phone knows based on some information.<\/p>\n If it doesn\u2019t know if it\u2019s going to be a scam, maybe it actually picks up the phone, and it\u2019s going to ask some questions before actually relaying that call to me. So I think we\u2019re probably going to have to generate similar systems to safeguard essentially our attention span that says something like, \u201cSlop likely.\u201d And you can open it if you want, but it\u2019s probably not worth your time. So I think progress there has been slower than on the generation side because this is not as important as cybersecurity where there\u2019s more investment there. But I think it\u2019s something that we\u2019re going to have to take very seriously from the university side, from the peer reviews process, the journals, and just everyone is going to have to make a concerted effort to do this.<\/p>\n I think it\u2019s a very important issue. Yeah. I\u2019ll just mention really briefly, I\u2019m a big fan of a tool called Pangram that can detect if writing is AI-generated. And frankly, if the author of something cannot bother to write it, maybe can\u2019t bother to sort of write it themselves, it\u2019s hard for me to justify why I should bother to read it. So I become a little bit of a light eye in that respect, actually. Mostly, if I see something posted that\u2019s AI-generated, I can generate plenty of my own AI output to read. I don\u2019t want to read somebody else\u2019s AI slop.<\/p>\n Go ahead.<\/p>\n Jessie:<\/strong> I\u2019m Jessie from Columbia University. I have one question. So almost all the researchers and audience have raised AI literacy, so we know how to integrate\u2026 Thank you. How to integrate AI into our daily work. But I notice in real workplace, many professionals, they still don\u2019t know how to integrate agentic AI into their day-to-day work. So what do you think is the biggest barrier to stop them from using AI in their daily work and broader AI adoption? Barrier to AI adoption. Kara? I guess just I\u2019ll speak personally. I feel like almost the idea of using it was more scary than once I actually tried it, I was like, \u201cOh, okay, this is way easier than I thought.\u201d So I feel like there\u2019s a lot of the tools themselves can help guide you through setting things up.<\/p>\n So I think that\u2019s sort of a game changer in terms of the adaptation. So I personally found the barrier lower than some other methods. It was almost more just getting over my initial hesitation of, like, what is this thing and how do I use it? I think understanding where it might be useful was probably the step that was like I needed first to understand why I should do this. Yeah, I think\u2013 Sorry. Go ahead. I was just going to say, I think getting your feet wet can go a long way. I don\u2019t know if this is on.<\/p>\n Okay, I think this is dead. Kara? Sound working? Hello. Yeah, I think getting your feet wet can go a long way. Like the next time you do something on your computer that you\u2019re like, \u201cOh, this is kind of annoying,\u201d just throw an agent at it and see if it can help out. And then if that works well, think about the next time you find something that, hmm, kind of annoying, but also you should do yourself and see if it\u2019s a good use case for an agent and just get the ball rolling is all I\u2019ve got to say.<\/p>\n I was going to say that I\u2019ve developed AI agents for small businesses on the side, and one of the things that they always come across is security and deployment. They often vibe code their own AI agent, but then they struggle to deploy, and they also struggle to install security because your cloud isn\u2019t the best at developing the security aspects of it. So always just keep that in mind as well.<\/p>\n Thanks. So I think we are at time now. Thanks to everyone in the panel. Let\u2019s give them a big round of applause.<\/p>\n<\/div>\n\n","protected":false},"excerpt":{"rendered":" One day in late September, researchers, engineers, students and industry leaders gathered at the Columbia Engineering Innovation Hub for \u201cThe Rise of Agentic AI: What It Means for Climate Innovation,\u201d an event hosted by LEAP (Learning the Earth through Artificial Intelligence and Physics), an NSF Science and Technology Center. The event brought together academic and…<\/p>\n","protected":false},"author":1,"featured_media":1123,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-1122","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-berita"],"_links":{"self":[{"href":"https:\/\/cn-seo.org\/index.php?rest_route=\/wp\/v2\/posts\/1122","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cn-seo.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cn-seo.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/cn-seo.org\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/cn-seo.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1122"}],"version-history":[{"count":0,"href":"https:\/\/cn-seo.org\/index.php?rest_route=\/wp\/v2\/posts\/1122\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cn-seo.org\/index.php?rest_route=\/wp\/v2\/media\/1123"}],"wp:attachment":[{"href":"https:\/\/cn-seo.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1122"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cn-seo.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1122"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cn-seo.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1122"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}\n\n