Glossary · Term

superposition

← all terms

Definition

Plain language

The idea that a model packs far more concepts than it has neurons by encoding each one as a blend across many neurons at once.

As stated in the literature

The superposition hypothesis: networks represent more features than dimensions by encoding concepts as directions (sparse combinations) rather than dedicating one neuron per concept, which produces polysemantic neurons and motivates dictionary learning to recover the underlying features.

Also called: superposition hypothesis

Why it matters: It explains how models cram so much into limited space and why their neurons are messy, which is the puzzle interpretability tools try to crack.

For example, a model with a thousand neurons can juggle far more than a thousand concepts by spreading each one as a faint pattern across many neurons.

Heard on the show

“That packing has a name — superposition.”
Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent

Mentioned in 3 episodes

  1. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
  2. 098
    Finding Millions of Readable Concepts Inside a Real, Deployed AI Model
  3. 002
    An AI Ran a Real Optics Lab for 21 Hours and Found a Transformer-Shaped Pattern in Light

Related concepts

Related terms