←back to thread

Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

(transformer-circuits.pub)
168 points 1wheel | 2 comments | 21 May 24 15:15 UTC | HN request time: 0.001s | source
1. bilsbie ◴[21 May 24 17:46 UTC] No.40431498[source]▶
How are they handling attention in their approach?

That’s going to completely change what features are looked at.

replies(1): >>40439840 #
2. tel ◴[22 May 24 11:49 UTC] No.40439840[source]▶
They target the residual stream. Also they may have a definition of “feature” that’s more general than what you’re using. Consider reading their superposition work.