Multi-agent social simulation · Controlled experiments
AI Town: The Model or the Environment?
Change the rules. Watch the town.
Let residents work, trade and talk, then compare simple rules and changed environments to ask which collective behaviors require model judgment.
Research question
It is tempting to credit AI for the order in a town that talks and lives. Rent, rewards and action rules also shape every choice. This study uses repeatable comparisons to separate model judgment from the structure of the environment.
Approach
- Record work, trade, social interaction and resource changes alongside dialogue.
- Compare model judgment, simple rules and judgment-removed variants while keeping other conditions comparable.
- Change rent and social conditions, repeat the runs and assess differences against criteria set in advance.
Where it stands
A separate living-town iteration now has 20 residents who have walked, worked, produced, spent and rested according to model plans. It is currently paused; continuous autonomous life and full memory reflection remain in development. Earlier controlled results remain unchanged: the model and judgment-removed groups did not meet the preset separation criterion, and the main measures did not separate the model from simple rules. The living-town iteration is not a new result for that experiment.
Research extension of a16z’s open-source AI Town. Credit for the original project, artwork and interface remains with their creators.