Key Takeaways
- Thomas Schelling won the Nobel Prize by showing that mild individual preferences, not just overt racism, cause total neighborhood segregation.
- Traditional agent-based models from the 1980s and 1990s stalled because simple red and blue dots lacked the richness of real human behavior.
- Joon Sung Park and Simile AI build digital twins using interviews, transaction data, and trials to simulate social decisions at 85% fidelity.
- The real goal of large-scale behavioral simulation is finding the specific path today that prevents catastrophic future failure modes.
The Red Dot Trap in Behavioral Modeling
In the 1970s, economist Thomas Schelling set out to study why American cities segregated so sharply. Most people assumed complete segregation required overt racism. Schelling proved them wrong with a checkerboard and two colors of coins.
He gave each coin a mild preference: stay put as long as at least one-third of your immediate neighbors look like you. If that threshold dropped below one-third, the coin moved to the nearest open space. The result was stark. Even with broad tolerance for integration, the system quickly segregated into solid monochromatic blocks.
Park points to Schelling to show why human intuition fails at scale: “One of the striking finding of this paper or this Asian-based model was for the longest time people thought the segregation within society was caused by explicit and overth racism. But if you look at this model, people's preference towards living with people of the same color, that preference can be very minute, but the very small difference actually causes the society to segregate completely over time.”
Schelling won a Nobel Prize for that insight, but the field hit a ceiling soon after. As Park explains, “agent-based models for the longest had impact in the in the 1980s, '90s to some extent early 2000s, but it has now sort of gotten forgotten by the community a little bit because as you can imagine, red dots and blue dots is not really a rich description of people.”
Psychohistory Without the Fiction
When Park designed the generative agents architecture behind the Stanford Smallville paper, he treated LLMs as a fix for Schelling's red dot problem. Instead of rigid rules, agents gained memory streams, reflection loops, and language.
That shifts simulation from a toy to what Isaac Asimov called psychohistory: using mass behavioral models to guide policy decisions before disasters strike.
Park frames the ambition through the opening plot of the Foundation series: “if you read foundation series literally the first act is there's a group of scientists who have found out that oh our galactic empire is going to collapse and we're going to have 30,000 years of unrest. And they basically run psycho history, the simulator that tries to teach them, okay, how can we keep this unrest to,000 years?”
Most teams treat simulations like survey engines. They ask what users might say on a form. Park argues that misses the entire point of the technology: "So really what simulation allows you to do in its highest form is you give it not a problem or question like what would people answer to the survey. That's not what we do. What we tell it is here is a goal that we have in the context of foundation. We want to keep the unrest to a thousand years. what is the path that we need to take now to get to that particular future and that's what simulation allows you to do."
At Simile AI, Park feeds agents observational transaction logs and randomized controlled trial data to hit 85% fidelity against real human cohorts. That allows builders to stress-test tokenomics, product launches, or fee increases in software before exposing real customers to bad edge cases.
What to Do With This
Stop surveying your users about features they have not seen yet. Write down the top three failure states for your next rollout, build five distinct customer personas as system prompts with transaction histories, and run fifty simulated interaction loops to spot emergent churn patterns before you push the code to production.