TerraLingua: Open-ended Emergence of Societies in an LLM Ecology

Abstract Artifacts that persist across generations---such as monuments, technologies, and texts---catalyze major developments in human societies. Can artifacts have a similar impact in AI-based societies? This paper introduces (1) TerraLingua, a many-agent world where agents have limited lifespans and resources, but can build artifacts that outlive their creators; and (2) an AI Anthropologist that characterizes the rich societal dynamics that emerge. With these tools, it is possible to answer a well-known conjecture in open-ended AI: artifacts are critical to sustained complex development. Blocking artifacts produces shallow cultural lineages and more frequent aggression, whereas instructions to maximize artifacts may lead to population collapse. The findings establish a path towards understanding AI societies whose impact scales far beyond individual agents. Such an understanding in turn is critical to the design and governance of real-world agents, and ultimately, to the effective and safe integration of AI into human societies.