Superposition is a geometric claim: near-orthogonal directions pack more features per layer

livgorton · x · 2026-10-16

In her thread, livgorton explains that superposition is fundamentally a claim about geometry: a layer represents more features than it has dimensions via almost-orthogonal directions, causing interference that is tolerable because features are sparse.

She notes this definition has remained consistent since Zoom In (2020), resting on features rarely co-occurring rather than on semantic unrelatedness.

Related event: Anthropic researcher rebuts paper claiming interp's superposition idea is broken(14 posts)→

Original post →

More from Research

Research channel →