Simulated worlds have always been designed: their rules were written by hand, and their inhabitants were scripted. Language models change this, because intelligence can now serve as the substrate a world is made of, not just a tool used inside one. Simulation is becoming something generated rather than something assembled.
Terrarium1 is a planet-scale simulation engine for this paradigm, running persistent worlds populated by autonomous agents. Each agent in the terrarium is an independent model with its own memory and thought process, and no central model directs its behavior or scripts its choices. This means that when two agents interact, two independent models are reasoning about each other in real time.
Our key innovation is a unified physics engine that acts as the bedrock for individual agent actions and conflict resolution in the world. Every action in the terrarium is staged, and the engine decides what becomes true, and it resolves every action against the complete state of the world. Due to this intelligent, model based world engine, agents can attempt any action they can describe, including actions we never anticipated, and the world remains consistent.
Because the world is generated by models rather than assembled from fixed systems, Terrarium1 is compute-bound rather than design-bound, so new scenarios require only new descriptions, and simulations can scale infinitely across scenarios and durations while maintaining absolute agent freedom and realism.
We have developed the first world engine of its kind, built on a general intelligence substrate. Every increase in the substrate’s capability gives us more capable agents and a more capable world engine, unlocking vast possibilities.