Abstract
Artificial intelligence has advanced rapidly through scaling of model parameters, datasets, “
“and computational resources. Scaling laws have demonstrated predictable improvements in model “
“capability; however, capability alone does not fully describe the value of an AI system deployed “
“in real-world environments.
Category: Collaboration
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.
