In the context of neural network interpretability, superposition refers to the phenomenon where a model encodes more independent concepts than it has neurons, by representing each concept as a direction in activation space rather than assigning one neuron per concept. Because high-dimensional spaces contain vastly more near-orthogonal directions than dimensions, the network can store a very large number of weakly interfering features simultaneously in a much smaller set of neurons. This compression is efficient but comes with a cost: it makes features harder to isolate during analysis and contributes to unexpected model behaviors when unintended feature combinations activate together.

Think of it like a radio tower broadcasting hundreds of stations on closely spaced frequencies. Each station can be received clearly with the right tuner, but the tower itself is emitting all of them simultaneously in the same physical space. A listener without the right filter hears noise, just as an interpretability researcher without the right probing method sees a tangle of mixed signals in the neuron activations.

Mechanistic interpretability researchers expose superposition by training sparse autoencoders (SAEs) that decompose a layer's activations into a larger dictionary of monosemantic features: reconstructed = decoder(relu(encoder(activation))). Anthropic's work on SAEs applied to Claude found tens of thousands of interpretable features—far more than the number of neurons in the probed layer—demonstrating superposition at scale.

Superposition is studied alongside Embedding geometry, Latent Space structure, and Neural Network internals; understanding it is foundational to interpreting why LLMs produce Hallucinations and unexpected outputs. You will encounter it in mechanistic interpretability papers, AI safety research, and discussions of why individual neurons in large models are rarely cleanly interpretable.