Dissertation Title: Population AI: Modeling Societies of Autonomous Agents

Abstract:

Many of the most consequential challenges facing modern society emerge from how millions of decisions interact. Pandemics spread through contact networks. Supply chains fail through cascading dependencies. Labor markets evolve as workers and firms continuously respond to one another. These are population-scale phenomena, yet AI has largely been built to reason about individuals in isolation. This thesis is founded on a simple but consequential insight: intelligence arises from how agents interact.

Large Population Models (LPMs) are the computational foundation of this approach. LPMs represent societies as populations of autonomous agents, each carrying a behavioral model that captures its decision process and local context. The framework models how agent behavior propagates through networks of interaction, allowing population-level outcomes to emerge from many local decisions. To make this practical at scale, we built AgentTorch, an open-source framework that runs LPMs from a single machine simulating a digital city of millions to national supercomputing infrastructure modeling systems as large as the entire US workforce. AgentTorch has been adopted by governments, national laboratories, and organizations globally.

Validation has come on two fronts. On human populations, LPMs have been deployed to monitor health security, food security, and labor market disruption across million-scale populations, revealing risks that do not appear in individual-level analysis. On AI populations, modeling millions of interacting agents enables the discovery of coordination protocols that outperform state-of-the-art AI systems and expert human teams on resource management tasks that have stumped human teams for over thirty years. In both cases, the gains arise from how agents interact rather than from increasing individual capability.

The goal is to build digital societies at the scale of nations: systems with millions of interacting agents that expand human capacity to understand and shape the world.

Ramesh Raskar
Associate Professor of Media Arts and Sciences
MIT Media Lab

Milind Tambe
Gordon McKay Professor of Computer Science
Harvard University

Joel Leibo
Senior Staff Research Scientist
Google DeepMind