Thermodynamic, Information-Theoretic, and Epistemic Constraints on Artificial Intelligence Systems

Abstract
For a fixed hardware generation and a fixed task-success threshold, broadening the admissible task and input scope is
hypothesized to increase energy per verified successful output and the cost of achieving specified verification coverage.