Published on

Stanford/Google researchers create their own virtual world with ChatGPT


This paper presents software agents that mimic realistic human actions, which can be used in various interactive applications like immersive environments, communication practice spaces, and prototyping tools. These agents behave like humans, doing everyday tasks, forming opinions, and interacting with others. They remember their experiences and use them to plan future actions.

The paper explains a system that uses a large language model to store agents' experiences in natural language, combine memories into higher-level thoughts, and recall them for planning. The agents are placed in a virtual environment similar to The Sims, where users can interact with them using natural language.

In tests, the agents show believable individual and social behaviors. For example, they can autonomously spread invitations for a Valentine's Day party, make new friends, ask each other out, and attend the party together. The study proves that the different parts of the agent system—observation, planning, and reflection—are all crucial for creating believable behavior.

This research combines large language models with interactive software agents to develop realistic simulations of human behavior and offers new ways to build and interact with such agents.