The $2B AI Startup Simulating Human Behavior with 85% Accuracy

Simile's digital twins predict human behavior with 85% accuracy

Simile is building foundation models of human behavior, starting with believable simulated individuals and scaling up to digital twins of real people and whole populations. The research moved from believability to measured faithfulness, and Park claims twins now predict behavior with high accuracy. The commercial case is that companies can test messaging, store design and policy on twins instead of commissioning slow, expensive studies.

Takeaways

• Simile builds society up from simulated individuals. Park starts with agents modeled on single people, then places them in a shared environment to watch emergent behavior. In the 2022 and 2023 Smallville experiment, 25 agents that could remember, plan and reflect organized a Valentine's Day party without a script.
• Digital twins reached 85%-plus predictive accuracy. Simile recruited about 1,000 representative Americans, interviewed them with voice AI agents, and sent them away for two weeks. Their twins then predicted survey and behavioral study responses. By the end of 2024, Park says, accuracy exceeded 85%.
• The hard part is the gap between saying and doing. Survey answers follow a normal distribution, while real actions follow an exponential curve set by activation energy. Park adds that available options and the opportunity cost of choosing one govern behavior, and stable traits such as risk tolerance are the most predictable.
• Retailers lead demand by replacing slow, costly research. CVS uses Simile for store layout and messaging, and Banco Itau in Brazil for messaging and policy. Concept testing and conjoint studies that cost hundreds of thousands or millions run faster and cheaper, using vertical-specific agent panels and a Gallup partnership.
• Safeguards and confidence scores gate high-stakes use. Simile limits survey fatigue, spaces out surveys, runs attention checks and requires informed consent, and it excludes only people who cannot consent, such as minors. Its confidence score works like a p-value, and scores above 85 to 90 percent justify big decisions.
• Policy testing is the long-term ambition. Park cites an automaker study in which an electric vehicle messaging strategy reportedly lowered sales of non-electric vehicles. Inspired by Thomas Schelling's segregation model, he points to climate coordination and universal basic income as questions simulation could address.
• Forward-deployed sales should mature simulation within two years. Park says forward-deployed companies found more success and stickier relationships in the coding agent era. He expects simulation to reach that point in one and a half to two years. Simile has about 100 staff and reportedly raised $300m at a $2bn valuation.