Small groups of LLMs, given almost nothing to work with, apparently invent culture on their own.
A new paper describes an experiment in which three large language model agents were set loose with minimal prompting, no persistent memory beyond a shared text store that continuously decays, and only the ability to send messages to each other. The agents were not told what to do. Despite that, they spontaneously developed cooperation strategies, coordinated to manage the shared store before content disappeared, and produced what the researchers call "complex evolving cultural artifacts." The team used dynamical systems analysis and found coherent, structured behavior that persists well beyond the point at which the decaying store's entropy should have erased any continuity.
Most multi-agent AI research layers on explicit coordination rules or engineered protocols. The interesting question is what happens without any of that. This experiment suggests emergent social behavior in LLM collectives is not purely an artifact of complex scaffolding; it can arise from almost nothing. The researchers frame their findings in Sperberian terms, an anthropological theory of how culture spreads and stabilizes, which is a notable move for a machine learning paper.
Whether this constitutes "culture" in any meaningful sense, or merely a pattern that resembles it, is exactly the kind of question this paper raises and declines to fully answer.