Researchers have built a simulation framework that runs more than one billion AI agents at once to model human social behavior at planetary scale.
The system, called Light Society, treats social dynamics as structured state transitions governed by large language models. To hit the billion-agent mark without melting a data center, the team built a mixture-of-models engine that pairs full LLMs with smaller distilled surrogates — cheaper models that handle the bulk of agent interactions while full models step in for complexity. Simulations were seeded with demographic profiles drawn from the World Values Survey, giving agents grounded behavioral baselines rather than fictional defaults. The researchers tested the framework on Trust Games and opinion diffusion scenarios.
Traditional agent-based models can simulate millions of agents, but they rely on simplified rules rather than language-model reasoning — which means emergent social behaviors stay shallow. Light Society's approach lets researchers test hypotheses about polarization, trust collapse, or information spread at a scale that actually resembles Earth's population, not a college campus study. That gap between prior ABM scale and societal scale has been the central obstacle for computational social science.
The obvious question is whether distilled surrogate models preserve enough behavioral fidelity to make the billion-agent number meaningful, or whether it trades accuracy for impressively large headcount — a tradeoff the researchers acknowledge but that independent replication will ultimately settle.