📄 Download local PDF: AgentSociety – LLM Social Simulation Paper
🔗 View original on arXiv: arXiv:2402.09630
Mapping the 2025 computational sociology breakthrough to your symbolic cosmology
In light of external shocks (noise - Dionysian) and internal efforts (signal - Apollonian), intelligence is an unambiguous noise-signal gradient and landscape. The epistemic origin of AgentSociety, then, lies in the computational turn: no longer theorizing about society from a distance (idealized destiny), but simulating it in motion (dynamic operations). The “source” is a deep lineage of modeling (Turing → Schelling → Epstein), now turbocharged with LLM agency. Heritage isn’t static—it’s compounding, dynamic code.
The society being modeled is an LLM-generated commons: 10,000 agents interacting in a shared symbolic world, forming relationships, spreading news, debating policy. The data isn’t scraped—it’s lived in simulation. Commons become computational petri dishes; data becomes dramaturgy.
This is the dangerous gift: simulate a society, and you risk mistaking it for reality. AgentSociety asks us to walk backward through the mirror—to question what it means to “know” a society we’ve built. The price of this power is constant self-doubt, methodological transparency, and ethical vigilance.
How do we play with each other? Via the inflammatory, polarizing, and universal.
Can we trust the synthetic? AgentSociety compares outputs to empirical data from real studies—and the alignment is astonishing. This final layer is about posture: validating, recalibrating, and looping ethics into design. A true scientific loop: experiment → inference → return.