Dan Robinson joins Theo Jaffee and Sofia Puccini to demonstrate a game about competition between AI labs. Players balance cash flow and technical progress against a moving safety frontier. Dan Robinson explains how individually rational choices can create a prisoner's dilemma: both players may prefer a slower collective pace while fearing the consequences of falling behind.
Dan Robinson distinguishes circumstances that favor racing, voluntary restraint or multiple possible equilibria. The demonstration explores uncertain safety progress, delayed information and the limits of coordination without reliable communication. Dan Robinson argues that greater transparency can help competitors avoid harmful races, while emphasizing that the game simplifies real institutions and strategic choices.
Dan Robinson also discusses recursive self-improvement and the bottlenecks that remain when research labor is automated. Compute, data and other scarce inputs can limit acceleration, while tasks that resist automation may become increasingly important. These are lessons from simplified economic models and exploratory games, not precise forecasts of what AI labs will do.
Watch on YouTube
