Part 1: A World From a Seed
A handful of procedural rules can grow a whole living world. First a planet from a seed, then life that adapts to it, procedurally, and with an LLM. This is how games can teach our agents.
You don't need a mountain of data to build a world. A few rules plus a seed grow one, oceans, mountains, climate, and even life that fits the world. That emergence, rules in, world out, is exactly how a game can teach an agent. Play the world below.
In the intro I told the story: regulated work is starved for data, learned world models need data we don’t have, and games solve the exact same problem a different way. They don’t store worlds. They grow them from rules.
Let me show you, not tell you.
A whole world from one number
Here is a real, living world grown from a single seed. A handful of rules turn that one number into terrain: where the oceans sit, how high the mountains climb, how cold it runs, how wet it is. Nothing about the world is stored. The rules are tiny; the variety is endless. Drag to spin it; scroll or pinch to dive from orbit down to the surface.
That is the whole point of a world model: author the rules once, and you get an unlimited supply of worlds to work in. No data set required.
And it is alive
A world is just scenery until something lives in it. So we asked a simple question: given this exact seed, its temperature, its rainfall, its biomes, what would live here? Below, each world grows two creatures side by side. One is procedural, its body and behavior fall straight out of fixed rules read from the terrain. The other is LLM-designed: a model reads that specific world and invents a creature to fit it, a furry grazer for a frozen seed, a water-hoarder for an arid one. Drag the seed and watch both adapt.
Here is the quiet part said out loud. This is a sandbox for watching three things at once: procedural generation hands us an endless supply of real, different worlds; an LLM reads each one and adapts life to it; and underneath, each creature runs a small, readable policy, a goal and a handful of rules it follows on its own, with nothing hand-scripted. That is the whole secret. A world you can actually see is a world where you can actually measure what your AI does. Later we point the exact same machine at a lending world and let agents adapt inside it.
Why this matters for regulated AI
A game grows a believable world from rules and lets things adapt inside it. That is exactly what we need for a domain like lending, where we can’t practice on real customers but we do have the rules: the regulations. Author the world from the known laws, drop agents in, and let them learn where a mistake is free.
The planet was the easy half. The hard, interesting half is the life: how a creature decides, on its own, when to drink, when to eat, and when to simply wander, and why that same tiny, readable policy is how we teach an agent to handle a rule it has never seen before.