“Language models have more superposition than vision models because their features are less sparse and their feature importance curve is flatter.”