James Parr, founder and CEO of Trillium Technologies, a technology company specialising in the application of AI and cybernetics to challenges, such as climate change, resilience and disaster response, looks at the deep obstacles to overcome for artificial intelligence to become a force for good and herald a true ‘intelligence age’.
Article from Responsible Science journal, no.8, April 2026.
Thomas Paine’s The Age of Reason reflected a structural pivot toward empiricism, putting the acquisition of knowledge ahead of belief. This simple reframing changed the world. Similarly, the Space Age broke the bounds of our terrestrial centricity, providing an orbital vantage point and reframing our place in the universe. The Information Age – powered by the humble transistor – has subsequently reframed the way we work, live, shop, communicate, and love.
By these definitions, we’re now experiencing another reframing. The artificial neural net (ANN), like the flint axe, prism, telescope, steam engine, electric motor, rocket, and transistor before it, is opening a door to a new epoch of human experience which we might call the Intelligence Age. (An observation that Stanley Kubrick pre-empted in “the most famous jump cut in cinema” in his masterful 2001: A Space Odyssey.)
Historical epochs aren’t demarcated by a single technological invention, however; they are defined by the profound infrastructural and epistemological shifts those inventions precipitate. New epochs are cultural as much as scientific or technological. In 2026, public discourse is still distracted by the novelty of ANN algorithms passing the Turing test or the emergence of artificial companions. However, focusing solely on these anthropocentric applications obscures a much more profound shift at a socio-technical, geopolitical, and ecological level – which we ignore at our peril. From an operational and infrastructure perspective, the new era emerging is much deeper: the active integration of deep learning and other machine learning methods with pervasive sensor networks, universal autonomy, and distributed cloud computing to create a cyber-physical computational fabric that stretches around the Earth and out into the solar system.
The mechanics of this cyber-physical fabric rely on moving beyond isolated, task-specific computational nodes toward a dynamic, networked intelligence. Compute, data, decisions, and autonomy need to be deployed anywhere they are needed – such are the limitations of the speed of light – effectively erasing traditional geographical bottlenecks. A primary example of this synthesis is the development of networked intelligence for Earth observation, such as Instrument-to-Instrument (ITI) translation and autonomous "tip and cue" models we have built at FDL.ai. Historically, satellite observation was a passive endeavour characterised by significant latency.
As of 2026, a low-resolution multispectral satellite can detect an anomaly, say, a potential methane leak or the sudden thermal signature of a wildland fire, and autonomously cue a trailing, high-resolution hyperspectral satellite to precisely segment the plume or quantify the emissions. By installing machine learning onboard satellites, the time between observation and actionable insight is drastically reduced, fundamentally altering the capacity to respond. For the first time, we can listen to the symptoms of our planet and triage a solution.
This emerging networked intelligence is already transforming applied science. In heliophysics, orchestrated AI models now analyse the magnetic structure on the Sun’s surface and solar wind measurements to predict the geoeffectiveness of coronal mass ejections, providing advance warnings of geomagnetic storms that threaten our technologically dependent society. In disaster response, in-orbit machine learning has enabled the first flight-proven demonstrations of real-time global flood extent segmentation, delivering situational context directly to emergency responders.
These isolated breakthroughs are converging toward a unified framework that we have been calling Earth System Predictability (ESP). Previously, environmental monitoring technologies functioned like scattered components of an engine; individually sophisticated but lacking the integration necessary to drive coherent action. Earth System Predictability is being accelerated by the maturation of Foundation Models.
Just as the technology sector has seen rapid evolution in language and vision models, the scientific community is developing massive, pre-trained geospatial models. Geospatial foundation models such as NASA’s Prithvi, Clay, and ESA’s TerraMind ingest enormous volumes of disparate data to encode deep representations of physical phenomena. Once trained, they require only computationally lightweight adaptors to perform highly specialised downstream tasks, from predicting localised environmental states to discovering complex causal pathways in climate data.
However, as Kranzberg famously observed, "Technology is neither good nor bad; nor is it neutral." The uncritical deployment of AI introduces epistemological risks, no matter how good the intention. We are seeing that commercial AI is structurally incentivised to produce fluent outputs, a characteristic that can fatally mask statistical fragility and hallucination. For Earth System Predictability to be scientifically valid – particularly in high-stakes scenarios regarding flood evacuation, wildfire containment, or grid management – foundation models must trade fluency for epistemic humility. They require architectural mechanisms that systematically process real-world uncertainty and constantly error-check their conclusions.
A model must possess the mathematical capacity to explicitly state when it lacks the data to make a reliable determination. Initiatives pioneering the systematic handling of real-world uncertainty for geospatial foundation models, such as FDL’s SHRUG-FM, ensure systems output probability maps and variance metrics alongside their predictions. We believe these are an absolute prerequisite in creating the vision described above. If a model doesn’t know enough to give an answer, it can cue another satellite to help. Integrating these reliability bounds allows scientists to accurately gauge data and model fidelity, ensuring any downstream decision – whether human or robotic – retains critical and dynamic error-checking.
Another stressor we need to factor is the macroeconomic trajectory and attendant acute geopolitical and environmental distortions. The dominant industry paradigm overwhelmingly favours the development of monolithic, artificial general intelligence (AGI) models containing ultimately quadrillions of parameters (‘Q-models’). Training and operating these massive architectures requires staggering energy resources and immense capital expenditure – where even sand is becoming a contested resource. Consequently, this foundational compute is increasingly centralised within a handful of monopolistic technology conglomerates – antithetical to the vision of a distributed and ubiquitous computational fabric enabling us to seamlessly make the trillions of decisions to better steward our planet.
For the international scientific community, delegating the processing of critical environmental and infrastructural data to these centralised entities is strategically untenable. Put simply, it is a profound loss of intellectual sovereignty. When a nation relies entirely on external infrastructure for basic predictive capabilities, it implicitly accepts the embedded cultural biases, operational priorities, and proprietary constraints of the provider.
Equally critical, the energy footprint of continuously routing vast quantities of telemetry across transoceanic cables to distant server farms contradicts the global imperative for decarbonisation – which AI-enabled audits are showing to be more urgent than ever. Moreover, relying on an AGI monopoly creates a fragile dependency, stifling local innovation and indigenous thought-styles, and homogenising the approach to regional challenges.
Mitigating these risks requires a deliberate structural pivot toward sovereign, hybrid computational ecosystems. Rather than defaulting to centralised nodes, scientific blocs and national governments must cultivate localised compute infrastructure powered by geographically proximate renewable energy. A responsible strategy involves developing moderately sized, highly optimised foundation models – for instance, highly optimised open-weight models in the 7-billion to 70-billion parameter range. When fine-tuned on high-quality regional data, these smaller models routinely match or outperform massive general models on specialised scientific tasks at a fraction of the energy cost, making true localised compute financially and ecologically viable.
Moreover, these models can be tailored specifically to regional data and cultural contexts and to solve problems through a ‘mixture of experts’ – only escalating when the large models are needed. Under this hybrid approach, a domestic model handles local inference, ensuring everyday analytical tasks and the processing of sensitive data remain securely within national borders. The massive, energy-intensive offshore models are relegated to a supplementary role, utilised primarily for sanity-checking complex anomalies or providing baseline global context. This drastically reduces the dependency on external monopolies and limits unnecessary energy consumption.
Maintaining localised compute also preserves scientific independence. By retaining control over their algorithms and proprietary datasets, regions can cultivate unique, highly specialised AI capabilities, such as advanced localised agriculture or specific medical data analysis, which can then be exported to the global community. Through this lens, we can start to get a better definition of the Intelligence Age – when it uplifts the wellbeing of all humankind. By that definition, we aren’t there yet.
Every historical epoch follows a trajectory from technological innovation to profound economic and cultural shifts. The current manifestation of this emerging epoch is not an inevitability that must be passively accepted. To harness the potential of this new promethean change responsibly, the scientific community must advocate for decentralised, sustainably powered cyber-physical infrastructure that prioritises epistemic rigor and regional sovereignty over centralised, opaque monopolies.
When this is achieved, and only then, will we be able to say that we are in the Intelligence Age.
Image credit: Steve A Johnson via Unsplash
