Joon Sung Park, creator of the Stanford Generative Agents paper (Smallville) and founder of Simile, explains how AI can now simulate human behavior at 85% accuracy—meaning the simulation predicts what someone would do almost as well as they could predict their own behavior.
The Smallville origin: 25 agents in a small town, each with a persona, routine, and relationships. One agent (Isabella, cafe owner) spontaneously decided to throw a Valentine's Day party, invited people, and they actually showed up—including Klaus who asked his crush out. This wasn't scripted; it emerged from the agent architecture.
The product today: Simile partners with companies like CVS. You define the population you care about, Simile collects real data through interviews (using their own RL-trained interviewer that asks 'tell me the story of your life') and surveys via their Gallup partnership, then creates agent simulations that can answer any question about that population.
The say-do gap: LLMs are trained on what people say online (attitudinal data), but there's a real gap between what people say and what they do. Simile's behavioral models (trained on RCT repositories) close this gap.
CPU vs GPU analogy: today's frontier models are like the CPU of intelligence—rational, objective, superhuman at math. Simile is building something closer to a GPU—models that represent the diversity of human values, preferences, and irrationality. You need both.
Convergence vs divergence: some simulations converge (network structure always forms hubs, like PageRank). Others diverge (elections, wars). For convergent questions, errors compound but still converge. For divergent questions, run 100 times and show the distribution of outcomes.
The North Star: simulation could be the Hubble telescope for social science. Where Hubble changed our understanding of the universe, simulation could change our understanding of humanity—macroeconomics, segregation, climate collective action, democratic stability.
I am somebody who is quite inspired by science fiction. And when you read science fiction that covers societies that have progressed far enough in its technological maturity, you always see two pillars. You have some version of AGI, and you have some version of simulations that really help guide the society.
Smallville was basically a game town of 25 agents living in it. Individual agents had a description of persona, but they would actually wake up the morning, do their routines, go to work, actually have relationships, sort of like people would, and they would actually have emergent phenomena, like having parties and so forth.
We demonstrated that using our architecture and the models, we can actually predict people's behaviors 85% as accurately as people replicate their own.
Turns out, people are irrational. We have a lot of subjective values, preferences and tastes. So you actually start to see divergence in model size going up and the performance in its ability to predict and simulate human behavior.
Simulation can be that for human society. The thing that does excite me, there's a lot of focus on natural sciences, but how can simulation really unlock our understanding of humanity and social sciences?
AL
