← NLA experimentsNatural Language Autoencoders · follow-up series
S3·b
Matched SAE features through the AV
GemmaScope-2 (L40, 16k, l0=60) decoder directions nearest each diff-of-means axis, run through the same two cells as the dom vectors.
- sae
- gemma-scope-2-27b-it · resid_post L40 16k medium
- matching
- cos(dom, W_dec rows), top-3 per axis
- features
- 0
Status —Feature matching runs right after the vector battery in the GPU phase — nothing here yet. Refresh once the run starts.
§2
Decodes
Feature directions get the full Cell B mult sweep; Cell C runs at 3.5× and 5× the feature's max corpus activation (the 015 dose-response window). Compare against the same axis on the s3-vectors page.