Artificial general intelligence (AGI) is often imagined as humanoid robots orchestrating complex tasks. Yet a deeper and more impactful trajectory is emerging: AGI-like capabilities acting as powerful tools to solve society’s most persistent challenges. The discussion at CERN Sparks! Serendipity Forum in September 2021 highlighted this shift. The talk underscored that AI is already driving transformative progress in understanding and managing the natural world, and that the path toward AGI could accelerate our ability to address climate change and other global issues. Rather than focusing solely on autonomous machines, the emphasis is on intelligent systems that can learn, reason, and act across diverse domains with broad applicability. The core idea is that the urgency and multifaceted nature of climate challenges demand algorithms that can generalize, adapt, and cooperate with human experts across disciplines. In this view, AI is not merely a set of specialist tools but a catalyst for a more capable scientific enterprise. The forum emphasized that progress in AI will be most meaningful when it is deeply integrated with real-world applications that require both predictive insight and effective control.
Understanding the AI Landscape: Prediction, Control, and the AGI Horizon
AI research today is often organized around two broad categories of problems: prediction and control. Prediction models are designed to learn patterns in a domain—such as weather systems, environmental processes, or epidemiological trends—and to forecast how those patterns might evolve over time. These models excel at discovering structure in vast datasets, identifying causal relationships, and generating probabilistic scenarios that inform decision-makers. Control models, by contrast, focus on guidance and action. They aim to induce desired behaviors in agents operating within an environment, adjusting strategies in real time to respond to changing conditions, constraints, and goals. This division reflects a practical reality: understanding what might happen is not enough if we cannot steer outcomes in a desirable direction. The dual emphasis on prediction and control is central to building more capable AI systems that can assist humans in both understanding complex phenomena and shaping their trajectory.
Developing an effective path toward AGI requires advances in both domains, with careful attention to how these areas interact in real-world settings. Environmental and social systems are inherently dynamic and uncertain, often featuring nonlinear feedback loops, emergent properties, and rapid shifts. These characteristics challenge even the most sophisticated predictive models and demand robust control strategies that can remain stable under variability. The pursuit of AGI-inspired capabilities thus involves designing algorithms that can reason about uncertainty, adapt to new contexts, and transfer knowledge across tasks. It also requires an emphasis on collaboration across disciplines, ensuring that models are informed by domain expertise and validated against real-world data and objectives. The two domains described here—weather forecasting and plasma control for fusion—illustrate how this integrated approach can yield tangible benefits while advancing the science of general intelligence itself.
Two real-world domains where DeepMind has actively contributed to climate-related challenges, while simultaneously developing what is required to progress toward AGI, are weather prediction and plasma control for fusion energy. In weather prediction, the inherent variability of atmospheric processes makes precise modeling extraordinarily difficult. Yet, by leveraging large historical datasets, researchers can infer the causes and effects that shape weather events. The core idea is to borrow generative modeling techniques, previously used in image and video generation, and apply them to weather data. This approach enables the creation of multiple plausible scenarios from limited observations, providing a more nuanced understanding of potential outcomes and their associated uncertainties. In collaboration with the Met Office, the UK’s national meteorological service, these models have demonstrated the ability to take as little as 20 minutes of weather data and generate several hypotheses for radar maps, enabling more accurate predictions of heavy rainfall within a 90-minute horizon. This level of short-term forecasting is crucial for timely decision-making in emergency services, energy infrastructure management, and flood warning activation, all of which rely on rapid and reliable information to protect communities and assets.
The long-term implications of improved weather forecasting extend beyond weather agencies. When meteorologists have access to more accurate and timely forecasts, decision-makers across sectors can plan with greater confidence. For emergency responders, better predictions mean faster mobilization and more effective deployment of resources during extreme weather events. For energy managers, accurate forecasts of precipitation and temperature can influence power grid operations, hydropower availability, and demand response strategies. For flood risk management, improved lead times and spatial resolution can guide the pre-emptive closure of critical infrastructure and the dissemination of warnings to vulnerable populations. The overarching narrative is that AI-driven weather prediction exemplifies how advanced modeling techniques can translate into practical, life-saving applications. It also demonstrates how generalized AI technology can extend its reach into core societal systems, reinforcing the argument that progress toward AGI can yield meaningful, broad-scale benefits when aligned with essential public needs.
Beyond the immediate benefits, the weather domain offers an instructive blueprint for how AI research can scale across related domains. The same family of generative models, trained on comprehensive historical data, can be adapted to learn patterns in other complex, stochastic systems. The insights gained from weather modeling—such as handling incomplete data, quantifying uncertainty, and producing interpretable forecasts—provide a foundation for tackling additional challenges where reliable foresight is critical. As the models mature, they will become more capable of integrating disparate data streams, incorporating physical constraints, and adjusting to evolving climate regimes. The result is a more resilient forecasting capability that supports proactive planning and risk mitigation across industries, governments, and communities. This example embodies the broader thesis: AI research that advances our ability to predict environmental processes can also contribute to the development of generalizable reasoning and learning mechanisms essential for AGI.
In parallel with weather forecasting, researchers are exploring plasma control for fusion energy, seeking to create reliable, clean, and abundant power. Fusion energy holds exceptional promise as a source of energy that is practically limitless, carbon-free, and self-sustaining. However, achieving stable fusion reactions requires managing highly complex, dynamic plasma behavior in ways that are both precise and robust. This domain presents one of the most challenging engineering puzzles: how to maintain novel plasma shapes that maximize energy output and sustain stability for as long as possible. The work in this area reflects a broader strategy for AGI development, where solving a highly challenging, real-world problem can illuminate general-purpose control principles and reinforcement learning techniques that generalize well beyond a single task.
In practice, this research involves close collaboration with premier institutions to push beyond the limitations of handcrafted models. By partnering with the Swiss Plasma Center at EPFL (École Polytechnique Fédérale de Lausanne), experts apply deep reinforcement learning algorithms originally developed for robotics to the domain of plasma control. The aim is to create controllers that can adeptly manipulate different plasma shapes and configurations, achieving responsive, real-time adjustments at a high temporal cadence. The reported achievement—an effective controller operating at 10,000 interactions per second—illustrates the potential for AI systems to handle fast, high-dimensional control problems in high-stakes environments. This kind of rapid interaction loop is essential for maintaining stability and performance in fusion devices, where errors or delays can have significant consequences for safety and energy yield.
The broader message from these efforts is clear: progress in AI is most meaningful when it is anchored in real-world domains that demand both predictive insight and agile control. The fusion and weather projects illustrate how domain-specific challenges can drive the development of generalizable AI capabilities. They also highlight the indispensable role of cross-disciplinary collaboration. Complex scientific questions do not lie neatly within a single field; they require the synthesis of machine learning, physics, engineering, meteorology, and domain-specific expertise. This interdisciplinary approach is not simply beneficial—it is essential for moving from narrow, task-specific AI to more general, adaptable, and robust systems.
The Power of Interdisciplinary Collaboration: Why Partnerships Are Essential
The path toward robust, real-world AI that can meaningfully contribute to climate solutions is inseparable from collaboration across disciplines. The progress described in weather forecasting and plasma control could not occur without the synergy created by partnerships among AI researchers, meteorologists, plasma physicists, engineers, and data scientists. The right collaborations bring multiple perspectives to bear on a common problem, enabling researchers to identify the most relevant questions, validate models against real-world data, and interpret results in a way that is actionable for practitioners and policymakers. In these endeavors, AI researchers bring methodological innovations—such as advanced generative modeling and reinforcement learning—while domain experts provide critical context, validation benchmarks, and practical constraints. The result is a more rigorous, credible, and impactful body of work that can be translated into tangible benefits for society.
Moreover, the forum stressed that forming a shared vision of what AGI could enable is a collective act. It is not merely about single demonstrations of technical prowess but about building a broader scientific culture that welcomes diverse perspectives from both natural and social sciences. This inclusive approach helps ensure that the development of AI technologies is guided by ethical considerations, societal needs, and long-term implications. It also fosters a sense of shared responsibility for steering the direction of AI research in ways that maximize human welfare while minimizing risk. When scientists dream together with a broad community of experts—from climate scientists to sociologists, from policymakers to engineers—the potential for designing AI systems that genuinely serve humanity grows substantially.
This collaborative ethos is grounded in a common scientific approach to navigating complex questions at the heart of the world’s most urgent challenges. The integration of AI into climate science, energy systems, and policy-relevant decision-making requires more than technical breakthroughs; it requires a framework for experimentation, validation, and knowledge transfer. Interdisciplinary teams can design experiments that simulate named scenarios, test robustness under a range of conditions, and assess the practical viability of AI-driven interventions. They can also develop best practices for interpreting model outputs, communicating uncertainty to decision-makers, and deploying AI solutions in ways that align with real-world workflows. Such practices are essential for ensuring that AI-driven insights translate into safer, more effective actions that protect people and natural resources.
In this context, the idea of “dreaming together” gains new meaning. Imagining what a world with AGI could look like is not a speculative exercise; it is a strategic process that helps identify the kinds of problems such a system should be prepared to address. It invites input from a spectrum of disciplines and sectors, enriching the potential applications and guiding research toward areas with the greatest potential for social good. The talk at CERN Sparks! Serendipity Forum underscored that the most impactful AI advances will be those that emerge from collaborative, cross-disciplinary efforts rather than isolated, siloed work. The message for researchers and institutions is clear: cultivate partnerships, share knowledge openly, and engage with stakeholders from government, industry, and civil society to shape the direction and governance of AI deployment in climate-relevant domains.
From Prediction to Action: How AI Moves Climate Science Forward
As AGI research advances, addressing climate change moves from a theoretical objective to an actionable, scientifically grounded program. The process requires not only developing more powerful predictive models but also creating reliable control mechanisms that can translate forecast information into timely, effective interventions. The fusion and weather projects illustrate a dual pathway: better forecasts reduce risk and enable better preparedness, while advanced control strategies — guided by AI — can improve the stability and efficiency of energy and industrial systems. This integrated approach recognizes that climate resilience hinges on both anticipating changes and tailoring responses to evolving conditions.
In weather forecasting, the practical impact of AI manifests in two primary dimensions: accuracy and timeliness. Large-scale models can infer complex atmospheric dynamics from historical data, enabling meteorologists to generate multiple plausible weather scenarios. The ability to produce a set of competing hypotheses for radar maps empowers decision-makers to weigh risks and deploy resources accordingly. This is particularly valuable for emergency services, where rapid, decisive action can save lives. It is also crucial for energy management, where demand and supply dynamics are sensitive to weather fluctuations. By anticipating heavy rainfall events up to 90 minutes ahead, authorities can implement protective actions, adjust grid operations, and activate flood warning systems with greater confidence.
Beyond short-term weather events, the broader implication is that advanced AI systems can help communities adapt to climate variability more effectively. When predictive models capture the range of possible futures, planners can design more resilient infrastructure, plan for extreme scenarios, and communicate risk with clarity. These capabilities are not merely computational feats; they translate into tangible protection for people, properties, and ecosystems. The weather work thus serves as a concrete demonstration of how AI research can progress from theoretical insight to real-world utility, validating the premise that AI-driven innovation can contribute meaningfully to climate adaptation and disaster risk reduction.
In plasma control for fusion, the translation from model to manipulation is equally consequential. Fusion energy is poised to offer a clean, abundant source of power if the remaining scientific and engineering challenges can be overcome. The application of deep reinforcement learning to plasma control represents a strategic step toward turning theoretical models into practical control policies that can operate within the demanding environment of a fusion reactor. The ability to control plasma shapes and configurations at high speed strengthens the prospects for stable energy production and longer pulse durations, both of which are essential for achieving net energy gain. While the current work focuses on specific control objectives and experimental settings, the underlying principles—learning from interaction, adapting to changing plasma states, and optimizing control actions—embody the broader capability targets for AGI: robust, generalizable reasoning and decision-making in complex, dynamic systems.
The convergence of these research threads—predictive modeling with actionable control, domain-specific problem-solving, and disciplined collaboration—illustrates a compelling pathway for AI to contribute to climate solutions in a tangible way. It highlights how optimizing both foreseeing and guiding processes can yield a more resilient, responsive society in the face of climate variability and energy challenges. Importantly, the approach maintains a clear link between scientific advancement and real-world outcomes. By aligning AI development with urgent public goods, researchers can ensure that progress in AGI is not an abstract milestone but a driver of practical enhancements to safety, sustainability, and prosperity.
The Road Ahead: Causality, Efficient Learning, and Transfer in AGI
The journey toward AGI is not solely about scaling numbers or increasing computational power. It also involves addressing core cognitive and learning challenges that remain central to general intelligence. Among these, establishing reliable causal understanding stands out as a foundational capability. Causality enables models to reason about what would happen under different interventions, which is crucial for both prediction and control. In climate-related domains, causal reasoning supports more robust inferences about how interventions—such as emission reductions, energy system changes, or adaptation measures—will influence outcomes. Building AI systems that can discern causal relationships in complex, noisy environments is a step toward more trustworthy decision-support tools that can guide policy and practice with greater confidence.
Another essential area is learning efficiency. Real-world deployment demands models that can learn effectively from limited data, adapt to changing conditions, and generalize beyond their training environments. Transfer learning, meta-learning, and continual learning are all central to achieving this level of flexibility. As AI systems encounter new climate-related tasks or novel experimental setups—such as different fusion devices or alternative energy configurations—they must leverage prior knowledge while remaining responsive to new information. Progress in these learning paradigms will enable AI to contribute across multiple domains without requiring exhaustive retraining for every new scenario. This capability is a key aspect of the broader AGI objective: developing flexible, general-purpose learning systems that can scale across tasks and contexts.
The relationship between specialization and generalization is at the heart of AGI development. Specialized AI tools have delivered impressive performance in narrow tasks, but their effectiveness is limited by their narrow scope. The hope is that a continuum emerges in which specialized components can be composed and coordinated by higher-level reasoning to address increasingly complex problems. This modular and hierarchical approach to AI design aligns with the need for robust, adaptable, and scalable systems that can operate in the diverse, dynamic environments characteristic of climate science and energy research. By exploring how these components interact—how predictive models feed control policies, how feedback from the environment refines forecasts, and how domain-specific knowledge informs general strategies—researchers aim to cultivate a more coherent and capable form of AI.
The science of AGI, as illustrated by these interdisciplinary efforts, is not a single breakthrough but an iterative process of learning, experimentation, and collaboration. It involves continually refining models, testing them in real-world contexts, and sharing insights across fields to accelerate progress. The fusion of predictive accuracy, reliable control, causal reasoning, and efficient learning forms the backbone of a more versatile AI system. As the field advances, more opportunities will surface to apply these capabilities to climate resilience, sustainable energy, and global health, among others. The overarching aspiration is to move toward systems that can reason about cause-and-effect, learn rapidly from experience, and transfer knowledge across different but related domains. In doing so, AGI research can become a catalyst for solving the interconnected challenges that define our era.
Broader Implications: Ethics, Governance, and Responsible Innovation
As AI systems grow in capability and begin to touch critical aspects of climate action, energy, and public safety, ethical considerations and governance become increasingly important. The deployment of sophisticated AI for weather prediction, disaster readiness, and fusion control must be guided by principles that protect people, ensure transparency where appropriate, and foster accountability for the outcomes of AI-enabled decisions. Responsible innovation involves not only achieving technical milestones but also building trust with users, communities, and policymakers who rely on AI to make informed choices. This includes clear communication about uncertainties, limitations, and the potential risks associated with AI-driven interventions.
The collaborative nature of the work described here underscores another governance imperative: the inclusion of diverse voices in setting research priorities and evaluating societal impact. Involving natural and social scientists, engineers, policy experts, and community stakeholders helps ensure that AI tools address real needs and align with public values. It also supports efforts to anticipate unintended consequences and to design safeguards against misuse. The responsibility to steward powerful AI technologies extends beyond researchers; it encompasses institutions, funders, and regulators who shape how AI is developed and applied.
In the context of climate and energy, governance also means ensuring equitable access to the benefits of AI-enabled innovations. The goal is to prevent disparities in the distribution of improvements in forecasting accuracy, disaster preparedness, and energy reliability. By embedding equity considerations into research agendas and implementation plans, the AI community can work toward outcomes that are more just and inclusive. This orientation helps guarantee that the advancements in AGI and its applications contribute to reducing vulnerability and increasing resilience for communities that are often most at risk from climate impacts.
Toward a World Where AI Solves Broad Societal Problems
The central claim of the CERN Sparks! Serendipity Forum talk is both aspirational and pragmatic. It asserts that solving climate-related challenges—and doing so with AI that generalizes across contexts—is a powerful driver for scientific progress in AGI itself. The path forward involves continued investment in predictive modeling, control systems, and cross-disciplinary collaboration, as well as a commitment to responsible innovation and inclusive governance. By maintaining a focus on real-world impact, researchers can ensure that AI not only advances theoretical understanding but also yields concrete benefits for people and ecosystems worldwide.
The journey toward AGI is not a single leap but a long arc of discovery, testing, and refinement. As climate science, energy research, and related fields provide rich, real-world testbeds, AGI-oriented AI systems can learn to reason more deeply about causality, improve their learning efficiency, and develop transferable knowledge across tasks. The work on weather prediction and plasma control demonstrates the value of confronting complexity with a combination of data-driven insight, physical understanding, and human expertise. It also reinforces the idea that the most transformative AI breakthroughs will emerge when diverse scientists collaborate to tackle questions at the intersection of technology and society.
In summary, the talk emphasized that AI is already delivering meaningful advances for climate-related challenges, and that AGI-level capabilities hold the promise of accelerating both scientific discovery and practical solutions. The emphasis on prediction and control, the concrete achievements in weather forecasting and fusion plasma management, and the central role of interdisciplinary collaboration together outline a compelling roadmap. This roadmap points toward a future where intelligent systems help society anticipate, prepare for, and adapt to the realities of a changing climate, while also expanding the boundaries of what is possible in energy, science, and technology.
Conclusion
In the evolving conversation about artificial general intelligence and its role in addressing climate and energy challenges, the key takeaway is that real-world impact arises from a synergy of predictive power, robust control, and cross-disciplinary collaboration. Weather prediction and plasma control for fusion energy stand as emblematic examples of how AI can move from theoretical capability to practical, society-serving outcomes. By harnessing large-scale data, leveraging advanced modeling, and deploying agile control strategies, AI research can help communities prepare for extreme events, optimize energy systems, and push the boundaries of what is achievable in fundamental science. The path to AGI, when guided by diverse expertise and shared purpose, has the potential to transform both our understanding of complex phenomena and our ability to respond to the urgent needs of the planet. As this work continues, it will be essential to nurture partnerships, uphold ethical standards, and maintain a clear focus on the public good, ensuring that AI-driven innovations contribute to a safer, more sustainable, and more resilient world.