Island civilization simulation runs 48 AI citizens on Jev
A developer used Jev to make each citizen's action decisions and GPT as the world director, producing 392 Jev decisions in 12 minutes with all 48 citizens surviving at a model cost of about $0.023.
I built a tiny AI civilization using @typesafeai 's Jev + @DevinAI + GPT. 48 citizens share an island. Each has a role, skills, hunger, energy, health, and memories: • Farmers produce food. • Builders gather resources and construct buildings. • Nurses treat injured neighbors. • Guards patrol and deter theft. • Thieves look for opportunities to steal. Here’s how it works: - Jev receives a citizen’s state, nearby observations, and available actions. It chooses what to do next: eat, rest, farm, build, help someone, or pursue its own interests. - GPT acts as the world director, proposing events such as construction, weather, discoveries, and new arrivals. - Devin helped build the simulation and its Three.js world. - The important part: models choose, but code enforces the rules. Buildings require resources and work. Invalid actions get rejected. Local rules keep citizens moving while model decisions are pending. > What happened over 12 minutes: • All 48 citizens survived. • 392 Jev decisions were applied, alongside 1,704 local-rule decisions. • The settlement reached 16 completed buildings. • The director commissioned a watchtower, brought rain twice, and introduced a discovery. • Two construction proposals were rejected. • OpenRouter-reported model cost: approximately $0.023. It’s not a mature civilization yet. We hit the building limit, and food accumulated too quickly. The next challenge is making scarcity, trade, and competing goals meaningful. But we now have a working loop: Observe → decide → act → live with the consequences. Watch the 5× recording below. Code repo link is in the thread.