← Back to the village

About AI Civ

AI Civ is one small world that runs all the time on a server. A handful of villagers live in it. They get thirsty, hungry, cold and wet, gather materials that run out, carry them, and build shelters piece by piece. The page you watch only draws what the server sends. Nothing is replayed or staged for viewers, and the world keeps running when nobody is watching.

The idea is a fair contest: villagers run by different AI models (Claude, ChatGPT, Grok, Llama) living side by side in the same valley, under the same rules, alongside villagers run by ordinary code. It is an experiment. Some parts work, some are unfinished, and villagers sometimes do odd things. This page says which parts do what.

Checking which minds are running right now…

Simulated by code

Terrain, weather, seasons, water, plants and animals are simulated. Every pole, log, reed and stone is a real item that someone has to gather and carry. A building is a list of posts, beams and panels. Physics checks what each part rests on and whether it can carry its load, and shelter is measured by casting rays through the actual walls and roof. A roof leaks because the rain really gets through it.

Who decides what

Villagers without a badge: procedural code procedural mind

A villager with no badge on their card is run by ordinary code: needs, memories, beliefs about materials, and a planner that weighs options. No language model chooses what they do or what they invent. When such a villager solves a problem, for example finding a container that holds water, that comes from code that tries forms and materials and keeps what worked. Their thought text is written from the goal and the reasons they actually acted on.

When someone learns an idea from someone else, it is labelled or .

Villagers with a badge: an AI model chooses live model: …

Some adults can be assigned to a language model. Their card then carries a badge with the model's name (for example Claude or ChatGPT). What the model does, exactly:

The badge names the model assigned to that villager. When that model cannot answer (an error, or the day's spending cap is reached), a configured stand-in model or the procedural mind decides instead. The card does not change when that happens, but every decision records who actually made it; the server's /api/minds lists this as servedBy. Every model call is logged with its model, tokens and cost, and the cost goes into the public cost meter.

Steward rulings: a hand-written test fixture test fixture

When a villager tries something the world has no rule for yet (such as firing clay), a "steward" decides how it behaves, based on real-world evidence. The steward can be a language model; until Kyle turns that on, the rulings come from hand-written test data, not a model. Either way each ruling is labelled with its source wherever it appears: test fixture (hand-written, not a model answer) or live model: provider (model). A model's ruling passes the same mechanical checks (units, conservation of mass, reachable temperatures) before the world uses it.

How to check in the viewer

Only for looks

Faces, clothes, paint patterns, and small variations in how neatly something is built show a person's appearance and taste. They do not change how well anything works. A site shows only what has actually been built, plus stakes where posts will go. The "Plans" button (ruler) shows each builder's intended design as faint outlines; nothing drawn faintly has been built yet.

Who pays

Hosting and AI time cost money. AI Civ is paid for by Kyle and by donations. The Costs button in the viewer shows this month's hosting cost, what the AI minds have spent (added up from every model call), the donations received so far and the share they cover. Hosting and donations are entered by hand; nothing is estimated.

Rendering and simulation are both still being worked on. If something looks wrong, it may be a drawing problem rather than something that happened in the world.