joon sung park

Simile has demonstrated simulation accuracy at 85% of human self-replication using generative agents with memory, planning, and reflection. This crosses a threshold for commercial deployment (Fortune 500 concept testing, earnings call simulation) and opens a path to simulating macroeconomic, political, and climate collective-action problems at scale, potentially solving long-standing social science questions.

Moving beyond single-prompt LLM personas, Simile builds agents that wake up, form relationships, plan parties, and exhibit emergent social phenomena. Multi-agent simulations allow customers to model downstream implications of decisions (e.g., EV launch impact on non-EV perception), a capability no current market research method provides.

CVS and other Fortune 500 customers use Simile to test thousands of concepts across representative synthetic populations, bypassing the bottleneck of recruiting real users and the physics of field experiments. The platform grounds simulations in real human data via Gallup partnership, then extrapolates to answer unlimited follow-up questions at near-zero marginal cost.

Frontier models optimize for rational superintelligence (CPU), but simulating society requires modeling irrational, diverse human preferences (GPU). Simile trains a behavior foundation model on randomized control trial data and life-narrative interviews to translate attitudes into behavioral predictions, creating a new model category distinct from OpenAI/Anthropic.