Millions of AI Agents on OASIS: How Similar to Humans on X(Twitter)?
Explore OASIS, an open-source platform simulating social media interactions with millions of AI agents. Study behaviors like message propagation and group polarization at scale.
What would a virtual society made up of over a million large-model AI agents look like?
Recently, Shanghai AI Lab, CAMEL-AI.org, Dalian University of Technology, Oxford University, and the Max Planck Institute, among others, jointly launched an open-source project called OASIS, which focuses on the interaction of millions of AI agents.
This project builds a general-purpose social simulation platform based on large models, supporting up to a million AI agents interacting. Researchers can easily simulate the interactions of large-scale AI agents in complex social environments using OASIS.
For example, it can be used to study classic social phenomena such as message propagation, group polarization, and herd behavior on social media platforms like Twitter and Reddit.
These demonstrative studies validate the effectiveness and practicality of OASIS as a social simulation platform. At the same time, OASIS also discusses the impact of agent societies on the technological path toward AGI as large models and agents progress.
Research Background
As the general capabilities of large language models continue to improve, AI agents based on these models have become a major research trend in the AI field. From research on individual agents to multi-agent interactions, many remarkable results have emerged, such as CAMEL [1], Generative Agents [2], ChatDEV [3], and MetaGPT [4].
However, existing methods face several challenges:
Limited scalability: Few studies have scaled agent interactions to the level of tens of thousands, and achieving this requires overcoming complex engineering challenges.
Limited interaction forms: Even with some studies achieving interactions involving thousands of agents, these interactions tend to be rudimentary and usually only support simple scene simulations.
One of the core questions in the design of the OASIS platform is: "How can we design a platform that supports interactions among tens of thousands or even millions of agents?"
A very intuitive idea is to use a "group chat" model, but having a million agents participate in a group chat at once is clearly not feasible.
In reality, there is a mature platform that supports high-frequency interactions among hundreds of millions of users every day: social media.
