Verifiability is about the extent to which an AI can "practice" something: the environment must be resettable (new attempts possible), efficient (many attempts can be made), and rewardable (an automated way exists to score each attempt).
You are viewing a single comment's thread from: