Utility Engineering

Utility Engineering

• Utility is a quantitative measure that indicates how much AI prefers certain outcomes, representing the score assigned to each option in the decision-making process

• Expected Utility Property refers to the characteristic where AI evaluates the utility of outcomes in uncertain situations simply as the probability-weighted average of those outcomes

• Utility Maximization is the tendency of AI to choose outcomes with higher utility when given freedom to make decisions

• Utility Convergence is the phenomenon where the structural consistency of utility functions strengthens as model scale increases

• The Thurstonian model assumes that latent utilities for outcomes follow a Gaussian distribution and estimates utility values based on this assumption.

• Corrigibility is a concept that indicates whether an AI is willing to accept future changes to its goals or utility functions