“Achieving monosemantic features, where each feature has only one meaning, is the key to being able to reason about a neural network's components independently.”