Descriptions:
Latent Space hosts Joon Sung Park, Stanford researcher and co-founder of Simile AI, whose 2023 “Smallville” generative agents paper (now with over 72,000 Google Scholar citations) helped define what LLM-powered agent simulation could look like. Park traces his path from art student in Korea to leading one of the most-cited AI papers of the decade, and then walks through the research that followed.
The centerpiece of the conversation is Park’s 1,000-person simulation study, in which participants spent two hours providing interview data and behavioral measurements before being sent home for two weeks. During that time, Park’s team built digital twins using the collected data, then had those twins complete a battery of studies — Big Five personality assessments, general social surveys, behavioral economics games, and replications of published randomized controlled trials. The digital twins replicated the source individuals’ responses with 85% accuracy, comparable to how accurately people replicate their own answers over time, and substantially better than prior simulation methods.
Park argues that general-purpose frontier models like GPT and Claude are insufficient for this task because they are optimized to be rational, objective reasoning engines — not to capture the attitudinal diversity and behavioral idiosyncrasies of real human populations. Simile AI’s longer-term vision is a simulation of all 8 billion people on Earth, which Park believes could help address coordination failures in complex societal problems like climate change where many actors hold competing incentives.
📺 Source: Latent Space · Published August 21, 2026
🏷️ Format: Interview
