Tapes Together Strong: The Co-evolution of Computation and Cooperation

Details

Classical evolutionary game theory studies cooperation by isolating social interactions from physical costs of behavior. Artificial life models study self-replication without formalizing the social dilemma of resource acquisition versus shared energy preservation. Autopoietic Game Theory (AGT) puts both together. In AGT, agents are randomly initialized Z80 machine code programs that compete for shared energy needed to replicate. The social dilemma is embedded directly into the physics of computation: stealing (defection) destroys shared energy, slows execution, and can prevent reliable replication. The authors show empirically and through a simplified theoretical model that this coupling can make defection self-limiting even in well-mixed populations — without requiring kin selection, spatial structure, or reputation mechanisms. Evolved programs in several Z80 environments suppressed stealing behavior. Spatial assortment (allowing related programs to cluster) further supported structural complexity and task performance. The framework can incorporate exogenous pressures: math tasks structured as sequential social dilemmas, with rewards tied to computation budgets, were also studied. Authors include Blaise Agüera y Arcas (Google) and researchers from MIT and McGill. The paper's primary contribution is theoretical — it proposes a new computational substrate for studying how cooperation could have emerged in early life or could be engineered into AI systems where shared resources constrain individual optimization.