superposition hypothesis

Models seem to know far more distinct concepts than they have neurons to store them in. The superposition hypothesis explains how: a layer packs many features into the same set of neurons by representing each feature as its own direction in activation space, and overlapping those directions. Picture a small room where you fit thousands of arrows by letting them point every which way — none gets a private axis, but as long as few are active at once, you can still tell them apart.

This works because features are sparse: in any given context only a handful of concepts are present, so the interference between overlapping directions stays small enough to tolerate. The network trades a little noise for a huge gain in capacity, storing more features than dimensions. The cost is that individual neurons become polysemantic — one neuron fires for several unrelated things — which is exactly why reading a model neuron-by-neuron is so confusing.

Superposition is the central obstacle that motivates sparse autoencoders. If concepts were each given a clean neuron, interpretability would be easy. Because they are folded together instead, much of the field is about unfolding them: recovering the hidden, near-orthogonal feature directions that the network compressed into a smaller space.