“The Superposition Hypothesis posits that dense neural networks are projections or 'shadows' of much larger, sparser neural networks, allowing them to represent more features than they have neurons.”