Joon Sung Park Introduces Simile AI and Simulation as the New Scaling Law
Joon Sung Park, the lead author of the landmark "Generative Agents" paper, has co-founded Simile AI to commercialize agent-based simulations, proposing simulation as a new scaling law for AI. The startup aims to scale these simulations to create digital twins of billions of people for behavioral and social forecasting. This marks a shift from academic, small-scale AI agent sandboxes to commercial-grade environments that can simulate complex human behaviors at a massive scale. If successful, agent-based simulations could provide a new paradigm for scaling AI capabilities beyond traditional compute and dataset size limits. Simile AI's vision involves scaling up generative agent simulations to potentially model "8 Billion Digital Twins" of every living human. This approach leverages LLM-powered agents to replicate human responses and social interactions, which recent research shows can match real-world survey responses with high accuracy.
## BACKGROUND
In 2023, Joon Sung Park and his co-authors published "Generative Agents," which simulated a small town of 25 AI agents interacting in a virtual sandbox. Traditionally, AI scaling laws dictate that model performance improves predictably with increases in compute, parameters, and training data. Simile AI proposes that simulating multi-agent interactions can serve as a new dimension of scaling to generate emergent, complex behaviors.