Simile co-founder and CEO Joon Sung Park is extending ideas from the 2023 Generative Agents project into models of human behavior. In a Latent Space podcast, he describes a goal of simulating people and populations so organizations can explore decisions before applying them in the real world.
The approach uses long-form interviews, observational and transaction data, and randomized controlled trials. Park argues that frontier language models are often optimized to sound rational and helpful, which can make them poor representations of people who are inconsistent, biased, or acting on incomplete information. Simile therefore post-trains models to capture some of the causal mechanisms behind human choices rather than relying only on prompting.
Park’s earlier research created digital twins of 1,000 people from two-hour biographical interviews. The article says their responses reproduced behavior and attitudes 85% as accurately as the original participants reproduced their own answers two weeks later. That result is presented as evidence for the method, not proof that broad human behavior can already be simulated reliably.
Potential uses include product research, policy testing, and studying how individual actions combine into population-level outcomes. Park also discusses representative synthetic populations and models that may eventually require data-center-scale computing. The project’s largest ambition—simulating all eight billion people—remains a long-term objective rather than a demonstrated capability.