Stuart Russell explains assistance games as an alternative to fixed-objective optimization and imitation learningImitation learning trains an agent to perform a task by learning from demonstrations of behavior rather than relying only on an explicitly written reward function.. In discussion with Roman Yampolskiy, he describes systems that remain uncertain about human preferencesAlignment is the effort to make an AI system pursue intended goals and follow relevant human values, instructions, and constraints., accept correction and have incentives to permit shutdown, while addressing the difficulty of aggregating different people's interests.
The interview develops a licensing approach based on specific unacceptable behaviors and evidence that risk stays below an agreed thresholdEvaluation measures how well an AI system performs against defined tasks, criteria and failure conditions using repeatable evidence.. Yampolskiy challenges the transfer of mathematical guarantees into the physical world and the possibility of manipulating human feedback. Later discussion considers international cooperation, multi-agent capabilityA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks. and possible economic scenarios, with forecasts and unresolved questions clearly distinguished from proven outcomes.
Watch on YouTube
