Virtual Town of AI Agents: Stanford and Google’s Experiment Now Open-Sourced on
Stanford and Google researchers have developed an experiment in which
virtual agents, represented by instances of ChatGPT, interact in a
fictional town. While these agents, such as “John Lin” and his son
“Eddy”, don’t physically move in a virtual space, they respond to
complex textual prompts about their surroundings, leading to
interactions like conversations between family members. The experiment
provides a fascinating insight into mimicking human behavior using a
non-state preserving API like gpt-3.5-turbo. What stands out is the
meticulous process of prompting each AI agent based on the minutiae of
their immediate circumstances, allowing them to “remember” and “react”
to situations. The result is a system where AIs, despite their
limitations, can carry out “believable simulacra of human behavior”.
This breakthrough, showcasing how generative agents can be developed, is
now open-sourced on a GitHub repository. A key highlight from the
research is the use of text-based prompts to guide the AI agents, making
it work seamlessly with non-state preserving models.https://github.com/joonspk-research/generative_agents
- 0 Replies
Sorry, there were no replies found.
Log in to reply.