Why Frontier LLMs Fail at Human Simulation: Park on Irrationality
Joon Sung Park explains why rational LLMs fail at human simulation, dropping to 20% accuracy compared to Simile AI's 85% fidelity digital twins.
Simulating Humanity: from Generative Agents to 8 Billion Digital Twins — Joon Sung Park, Simile AI
With swyx, Alessio Fanelli, Joon Sung Park · Sunday, August 23, 2026
Joon Sung Park, co-founder and CEO of Simile AI and lead author of the landmark Generative Agents ('Smallville') paper, details his journey from artist to Stanford PhD researcher building foundation models of human behavior. He explains how Simile creates digital twins using deep interviews, observational transaction data, and randomized controlled trials to simulate human decision-making and social physics at 85% fidelity. The conversation covers the difference between predicting and shaping the future, why frontier LLMs fail at behavioral nuance, and the long-term vision of simulating all 8 billion people to address complex societal coordination challenges like climate change.
Key takeaways
- Frontier models like GPT-4 drop as low as 20% to 30% accuracy when simulating real human behavior in niche sub-populations because they are optimized for hyper-rational reasoning. Read more →
- Joon Sung Park and co-founder Percy Liang, who coined the term foundation model at Stanford, argue that public web data cannot accurately simulate real human behavior. Read more →
- Static persona prompts fail because frontier models collapse into stereotypes rather than reflecting actual human choices. Read more →
- Simile AI is mapping early scaling laws in agent simulation: pouring in more behavioral data and compute produces predictable, repeatable accuracy gains. Read more →
- Thomas Schelling won the Nobel Prize by showing that mild individual preferences, not just overt racism, cause total neighborhood segregation. Read more →
5 articles from this episode
Why Prompting Cannot Simulate Human Decision-Making
Joon Sung Park on why frontier LLMs fail at social physics, how Simile hits 85% fidelity, and why Smallville ran on markdown files.
Why Simile AI Replaces Persona Prompting with Digital Twins
Joon Sung Park explains how Simile AI builds digital twins from real human data to test UI mockups and predict behavior.
Joon Sung Park on Scaling Simulations to 8 Billion People
Joon Sung Park reveals how Simile AI builds 85% fidelity digital twins to solve planetary coordination problems with simulation scaling laws.
Why Tiny Individual Choices Trigger Mass Emergence
Joon Sung Park explains how generative agent simulations turn Asimov's psychohistory into tools for testing economic and product rollouts.
More Latent Space episodes
- Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna RoboticsOct 18 · 4 articles
- High Taste Testers: The AI Tools We Keep Coming Back ToOct 18 · 5 articles
- From Wet Lab to RL Env: 10,000 Drug-Discovery Campaigns in a Weekend — Ian Quigley, Leash BioOct 18 · 6 articles
- How to kill your SaaS: what we learned replacing a $43K/yr tool with AI agentsOct 18 · 4 articles