Generative agents: living artificial lives in Smallville

In Generative Agents: Interactive Simulacra of Human Behavior (April, 2023), the authors discuss the use of the ChatGPT LLM to generate the behaviour of 25 agents that live and work together in a simulated village called Smallville.

The agents wake up, have breakfast, open up their shops, get coffee, organise parties and gossip with each other. Their activities are surprisingly… human:

These generative agents produce believable individual and emergent social behaviors: for example, starting with only a single user-specified notion that one agent wants to throw a Valentine’s Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time.

What I found most horizon-expanding about this work was the demonstration of the results of using the LLM in a wider setting. Rather than the end result being the text the LLM outputs, instead the social interactions of the agents are the output. The LLM’s text output is instead being used for its encapsulation of behaviour in its language generation, and is fully hidden within the system’s runtime.

The LLM is put to use in a great variety of ways within the system. A fascinating example was the way the researchers structured each agent’s memories and used the LLM to decide which events an agent experiences would be more important to that agent. The LLM’s ratings of interestingness are fed to a wider system, which uses the ratings to decide which memories to use during another use of the LLM: asking it for the actions an agent should take. The system is full of ways prior LLM result texts influence future LLM result texts in ways designed to support this world simulation.

Overall, a very eye-opening exploration of the possibilities within the space.