bibkey: “till2024truefeatures” authors: “Demian Till” year: 2024 title: “Do sparse autoencoders find true features?” doi: null claim: “Frequently co-occurring features may be replaced by a composite feature because the sparsity saving outweighs reconstruction costs elsewhere.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://www.lesswrong.com/posts/QoR8noAB3Mp2KBA4B/do-sparse-autoencoders-find-true-features”

Do sparse autoencoders find true features?

Verified locator

LessWrong, 22 February 2024. Qualitative predecessor of the composite-atom argument; the fiber-law volume gives the exact unit-atom l1 form of this argument.

Declared identifiers: https://www.lesswrong.com/posts/QoR8noAB3Mp2KBA4B/do-sparse-autoencoders-find-true-features.