Linear representation hypothesis

There is significant empirical evidence suggesting that neural networks have interpretable linear directions in activation space.