Pace: A Multiplayer Game About the Game Theory of AI Development

Summary

Paradigm has created Pace, a multiplayer web game based on a paper by Drew Fudenberg and Andrew Koh about competition between AI labs. The model treats labs as players that can accelerate or decelerate development: they gain financially by staying close to or ahead of a rival, but face greater catastrophe risk when the leading model moves beyond a shared safety frontier. Whether players race at maximum speed or pace development along that frontier depends on factors including catastrophic-risk probability, discount rates and transparency about progress. The game makes the strategic tension concrete through delayed information, temporary privacy around acceleration, gradual and irreversible decisions, an unpredictable safety frontier, and increasing acceleration that reflects possible recursive self-improvement. The authors note that real-world control is harder because development also depends on long-term commitments to compute, data and labor; labs can keep progress secret longer, the safety frontier is uncertain, and there are more than two competitors. They argue that similar dynamics apply to competition between countries such as the United States and China, making verification, transparency and shared safety standards important to any coordinated slowdown.