Reversible Agents Need Robots Waste Bits to See, Talk, and Achieve?

A computer's task is often taken to be that of starting with some input, grinding for a while, and eventually returning an output. Remarkably, all such tasks can be accomplished "reversibly", with an arbitrarily low intrinsic entropy cost, and in reasonable space and time relative to irreversible approaches. Artificial intelligence computation, however, does not fit well into an input-grind-output format. "Reversible agents" should run indefinitely, observing their world, acting to achieve goals, and talking with other autonomous agents. Do these additional requirements introduce more intrinsic entropy costs for such agents? This preliminary survey suggests that they do. Goal states can unavoidably have lower entropy than initial states. Sensing costs, when features of the world relevant to an agent's analysis change unpredictably before the agent has time to reverse that analysis. On the other hand, talk between nearby agents can be cheap.