At first glance everything looks cute: a small town as a computer simulation. In interactive dollhouses in the style of The Sims, 25 AI agents interact with each other in natural language in a shielded environment (sandbox). An agent wants to host a Valentine’s Day party, and his buddies spend two days spreading the invitations. Everyone coordinates to show up at the party at the right time.
However, no people move the figures, they move themselves and make unsolicited decisions. One almost wants to say “autonomously”. Because an agent structure expands the capabilities of large language models to store a complete record of all experiences of the acting “agent”. All of this happens in natural language, just as we humans use it to communicate and interact with our environment.
Interactive experiment with AI agents
The setting does not originate from a children’s birthday party, but from a scientific experiment on the capabilities of modern artificial intelligence (AI). The research report (“Generative Agents: Interactive Simulacra of Human Behavior”) written about this partly answers the question of the extent to which today’s AI systems are already able to imitate human behavior. The agents carry out activities that are reminiscent of human activities: they wake up, prepare breakfast, go to work. Painting, writing, forming “opinions”, perceiving each other and initiating conversations.