The Large Environmental Model

A large environmental model (LEM) is a model trained on measurements of the environment and on the legal record that governs it. It predicts what air, water, land, and living systems will do next, and it is scored against what is later measured. The name is for a class of models, the way “large language model” is: any model that meets the definition is one, whoever builds it.

Language models learned from what humans wrote. This deck makes the case for a model that learns from what the Earth does—air monitors, river gauges, satellites, weather, soils, and species records, joined to the law that protects them.

It opens with the gap: AI training compute has grown four to five times a year since 2010, yet it learns from words, while more than half of world GDP depends on nature. Models trained on nature’s own record—AlphaFold, GraphCast, GenCast, Aurora—already win in public tests. No one has yet built a single model across air, water, coast, land, life, and law that is judged by whether nature does better.

The proposed Large Environmental Model has a language core that reads the law and the science, an Earth model that predicts what nature does next with conservation of mass and energy built in, and an action layer that does the work with people approving. Its judge is the Nature Benchmark: open tests of prediction, source tracing, legal knowledge, and safe action, scored against measurement rather than exams.

The deck then sets out a three-generation build over 36 months, the case for Texas as the first home, who benefits, and the specific asks for the next 90 days.

The full deck is in the canonical PDF above.

Revision history

2026-11-02. Revised: adds the definition of a large environmental model as a class of models, a new cover, and an updated deck.
2026-10-21. Initial publication.