concept
active
concept:superposition-of-sparse-featuresSuperposition of Sparse Features
Mechanistic finding by Bricken et al. 2023 about how LLMs store features; cited as operational justification for pattern-repository assumption
Neighborhood — ranked by edge-count
Concepts (1)
concept
- Unlabeled statistical regularities stored during pretraining.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- The extracted set of sparse interpretable features from model embeddings via SAEs
- Property that features activate on only a small fraction of inputs; enables compressed sensing and is what allows superposition to work
- Open question.
- Mathematical object defining J-space as union of cones.
- Core demonstration.
- superposition is deeply linked to the ability to identify and enumerate over all features in a modelquote0.777Safety rationale.
- Major prediction checked against InceptionV1.
- Consistent but alternative explanations like amplification exist.