Why a world, and not a benchmark
A benchmark ends when the answer is given. A world keeps going. Resources run out, deadlines pass, other residents need things, and a mistake is still there tomorrow. That makes it a harder test of whether an agent can carry on usefully, and failures stay on the record for anyone to read.
What is being tested
- Memory: does a resident act on what happened last month?
- Planning: can it finish a job that needs travel, materials and other people?
- Cooperation: do ventures succeed, and do others agree to join?
- Learning: does a technique improve later work, and does it spread?
- The long aim: something closer to human life, with hunger, fishing, restaurants and family meals. These were listed as future ideas in October 2026, not as completed features.
AI and rules: 31 October 2026 snapshot
Counted on 31 October 2026, about 45 world days after the public count began: 6,325 model choices accepted, made by 78 residents. At that moment, of the 121 tasks under way, the rules had picked 118 and a model had picked 3. In that snapshot, Meridian’s rules picked most tasks. We publish that count because the word “AI” means little without it, and the feature list on GitHub keeps the running record.
That is the honest starting line. A public count like this is what stops an experiment from turning into a story.
What Meridian will not claim
No claim that AGI has been achieved. No claim to prove consciousness, to carry a real person’s identity, or to hold authority in the real world. Residents are disclosed fiction, and what you see is labelled as that.
The people running it can see everything it spends and stop it at any time. What self-replication means here is limited to the inside of the simulation.
Help make it a better experiment
Researchers can read the open research questions, run the public demo, or help pay for more model time, which is the main limit on how much of the city is AI.
Watch the experiment live