Chinese researchers develop Light Society to simulate 1 billion AI agents

Researchers from elite Chinese institutions, including the University of Science and Technology of China, Tsinghua University, and Fudan University, have developed a simulation framework called Light Society. This system is capable of operating over one billion AI agents, each possessing human-like characteristics such as personality, memory, beliefs, and decision-making abilities supported by large language models (LLMs).

Light Society represents a significant technical leap, overcoming previous computational bottlenecks that limited simulations to approximately 10 million agents. The framework achieves this scale through two primary innovations: a mixture-of-models engine that distributes cognitive workloads and the use of knowledge-distilled surrogate models, which utilize smaller, faster models to mimic more complex ones.

To ensure realism, agents are instantiated with demographic profiles and values drawn from the World Values Survey. The researchers have demonstrated the system's utility through experiments involving trust games to observe cooperation patterns and opinion diffusion to model how ideas propagate through massive social networks. The framework is intended to serve as a tool for predicting complex social phenomena and testing methods of influencing public opinion at a planetary scale.