Natural General Intelligence

An essay site, read 2026-09-21. "Natural General Intelligence", tagline "The case for a nature model that learns the state and dynamics of the planet itself", by Ryan Orbuch, September 2026; the contact address is at lowercarbon.com, the only affiliation stated. A review, by Anselm Hook.

What I liked. The large-scale world model is an audacious way at the modelling problem, and it rhymes with other big efforts: NVIDIA's Earth-2, Mark Pesce's The Playful World, and further back Buckminster Fuller. The techniques it leans on did not exist a few years ago. Orbuch writes well, names specific data sources, tools and artefacts, and gives the proposal a formalism, an "NGI" world model. There is a concrete, fundable ask: build a foundation model trained on planetary data.

How it is built. The introduction starts from the fact that we are embodied beings grounded in the real world, a seminal point I wish I had made in my own writing. It observes that Earth systems already have the automation we fantasise about, already respond to change, already feed us, and that only human-managed systems carry so many externalities and need so much tending. It dwells on knock-on effects, the limits of our cognition, and the environmental crisis. A chapter on Lovelock grounds the planetary view in the history of the Gaia idea. The core chapter proposes the nature model, with diagrams of how it would sit inside human decisions, tables of the raw environmental datasets that exist, and satellite coverage maps, then the supporting powers: AI, compute, sensing, data, and our new capacity to intervene in nature at scale. A chapter on data cites FitzRoy to ground prediction and argues for more sensing. A chapter on building the model treats simulations as scaled-down worlds, distillation, regression, discovering relationships, then speculates. The close invites contact and suggestions, and, notably, addresses language-model agents directly: share observational archives with provenance, resolution and uncertainty; move data across disciplines; find missing measurements; critique arguments; look for simpler approaches; pick one topic or one place and produce something others can inspect or build on, a dataset adaptor, a benchmark, a validated model, a proposal for what to observe.

What it glosses over. Two things that matter to me. First, the aim should not be technology per se but encouraging people to build models any way they can. A single model is more fundable, but there will be many competing models, and even if an NGI dominates there will be many instances of it, run separately by many people. There will not be one giant data centre with a website you log into to test your idea. Second, there is nothing on bioregional expertise, participation or civic engagement. Corporations and governments already build models; most of us are passive consumers who read the article and never play. Better tools level that field: our own ideas, public data, our own predictions. So for me the challenge is not exactly technical. It is social. And in the chapter on building, I wanted more on the downstream effect of better tools, on helping us see.

What it opens. The passage addressed to agents treats them as having powers any reader has, and it suggests an opportunity: other developers could stand up their own planetary agents that do exactly this, ongoing analysis, sub-agents, a ticket list, grinding perpetually. There is a role for many participants, human and not. One could even stand up an agent that advocates for a specific river.