Simulating Humans at Scale: Simile's Joon Sung Park

Summary

Joon Sung Park describes Simile as an applied AI lab focused on simulating human behavior and societies with higher fidelity than general-purpose frontier models. He traces the origin story to Stanford work like Smallville and Generative Agents, where LLMs plus memory, planning, and reflection produced believable 25-agent town dynamics. The company’s thesis is that today’s models are good at objective problems, but human society requires models that encode diversity, irrationality, preferences, and long-tail life history. Simile collects behavioral data through interviews, surveys, Gallup-style panels, RCT repositories, and customer-owned data, then evaluates outputs with population-level metrics like total variation distance. Park frames the long-term ambition as a “CERN of human society” for studying everything from product decisions to bank runs, climate coordination, and democracy collapse.

Notes

• Smallville was the Stanford prototype that demonstrated generative agents with memory, planning, and reflection in a 25-agent town.
• Park’s research path runs through Stanford’s early foundation-model work, including Opportunities and Risks of Foundation Models.
• Simile is positioned as an applied AI lab building human-behavior simulations from real data, not just product wrappers.
• The company’s thesis is that foundation models learn attitudinal web text, but human society requires richer behavioral modeling.
• Simile collects long-tail life-story data, interviews, and surveys to close the attitude-versus-behavior gap.
• Simile says it is building a foundation model of human behavior, literally trained on RCTs and customer data.
• Frontier models are used to coordinate research, but Simile builds its own model where the gap is human diversity rather than objective reasoning.
• The product starts with concept testing, then expands to product testing, simulated product experience, and earnings-call simulation.
• Simulation is valuable because it scales beyond ad tests and can represent populations that online experiments miss.
• Population-level evaluation uses total variation distance, with under 0.5 treated as strong evidence for some decision-making cases.
• Park splits simulation problems into convergent and divergent classes, which changes how confidence should be estimated.
• He says simulation needs standards analogous to p < 0.05 in inferential statistics.
• Park sees major societal applications in macroeconomics, bank runs, climate action, democracy collapse, and the monetary system.
• Park says science fiction’s mature societies usually have both AGI and simulation, and he thinks simulation is now ready for its first real attempt.