The day in conversation for 2026-10-23

The discussion: Ruth: That shift from relying on Bayesian updating in simple environments to potentially ignoring entire states when the world size explodes seems like a fundamental breakdown in standard modeling assumptions. What we're seeing isn't just noise; it’s a structural change in how agents process information given resource constraints. Warren: Exactly, but I wonder about the practical implications for prediction models. If we assume rationality implies Bayesian updating, and the paper suggests non-Bayesian heuristics become necessary, how do we operationalize that necessity? Does this mean certain thresholds for environmental complexity or cognitive load trigger a switch from maximizing expected utility based on probability to some kind of rule-based shortcut? Iris: I’m more interested in the cognitive mechanism itself. If ignoring entire states is optimal under limited resources, does that imply a specific form of computational economy—a form of bounded rationality where the cost of calculating probabilities for every state outweighs the benefit? It raises a question about the precise limits of what an agent can actually track versus what they must approximate. Ruth: That approximation seems key. When we look at the gap between optimal outcomes and that Bayesian benchmark widening with world size, are we talking about a linear increase in error, or is there a point where the non-Bayesian heuristic introduces