Simulation: the new Scaling Law — Joon Sung Park, Simile AI

The podcast begins with Joon Sung Park sharing his background, originally born in Korea and moving to Boston at age 11. His early interests were in arts and music, with aspirations to be an artist. However, his fascination with the creative potential of computation led him to research and AI. His journey includes pioneering work on foundational AI models and deep dives into modeling human behavior and societal simulation.

Joon discusses the origins of his work, starting from the 2023 Smallville paper on generative agents—AI characters capable of memory, planning, socializing, and developing emergent behaviors. The core idea was to ask what it would mean to recreate the world or simulate society to better understand human interactions and behaviors. The 'time machine game' prompted the team to envision creating life-like, scalable representations of humans and society, leading to the vision of simulation as a tool to understand, predict, and influence societal outcomes.

Key to their approach is incorporating diverse, long-form qualitative data (like life stories), observational transaction data, and causal mechanism data from randomized controlled trials. They emphasize that humans often make decisions shaped by deep, complex contextual factors, which current models, trained mainly on web data, fail to capture. Instead of building models solely for prediction, their goal is to model the underlying social physics—rules and behaviors that govern human actions—by continuously feeding models with rich, multi-scale data.

The team created digital twins—personalized AI models representing real people—and validated that these twins predicted behaviors with about 85% accuracy compared to individuals themselves. They highlight the difficulty frontier models face in accurately reproducing nuanced human behaviors, biases, and mistakes, especially in niche populations. Their approach involves post-training on a large set of controlled behavioral trials, improving robustness and making the models more realistic.

Joon discusses early evidence of scaling laws in AI simulation—more data and compute improve model performance. The ultimate aim is to simulate entire populations (billions of individuals) to explore societal issues like climate change, democracy stability, and economic systems. They envision multi-agent ecosystems that can provide insights into complex societal phenomena by understanding emergent behaviors and causal dynamics at large scales.

The conversation explores the roots of agent-based modeling, notably Thomas Schelling’s work on segregation, which demonstrated how small preferences can lead to large societal patterns. Modern advances with generative AI offer an opportunity to create high-fidelity, societal-scale agents that can help solve wicked problems—such as climate change and social collapse—potentially earning Nobel-level recognition for impactful work.

Cost concerns are addressed, with current efforts modeling tens of thousands to hundreds of thousands of individuals weekly, and reusing data for multiple studies. The aim is to eventually scale to simulations that incidentally might cost as much as training massive foundation models. The conversation discusses in-memory inference efficiencies, multi-agent interactions, and the societal value of large-scale simulations capable of informing policy and decision-making.

Use cases include understanding consumer populations, testing policies, product design, and replacing expensive human focus groups with synthetic populations. The technology aims to improve decision-making across sectors like marketing, public policy, and product development by capturing the diversity and complexity of human behavior at scale.

Joon draws a parallel between simulation and painting, emphasizing that both aim to highlight the essence of their subjects rather than perfect replication. His background as an artist influences his view of simulation—an endeavor to surface core truths about society and human nature, enriching understanding about collective life and individual habits.

Looking forward, Joon is driven by societal impact—envisioning simulations that can answer fundamental questions like the need for Universal Basic Income or understanding how to address wicked problems such as climate change. He acknowledges the potential of simulation to revolutionize policy and societal planning, with the goal of performing countless scenarios rapidly to inform real-world decisions.

The discussion ends humorously with the philosophical question of whether we are already living in a simulation. Joon remarks that regardless of the truth, the experience remains real and valuable. The episode highlights his optimism, scientific curiosity, and driven mission to leverage simulation technology for societal benefit.