Interpretability lacks metrics robust enough to survive optimization, researcher says
nabla_theta · x · 2026-10-18
In a reply on interpretability work, nablatheta argues the problem is "substantially harder" because there is no good interpretability measure that is actually sufficiently robust under optimization — a core critique of current mech-interp evaluation methods for safety purposes.
More from Safety
- Polymarket gives OpenAI just 8% odds of a full training pause by end of 2026 — Polymarket · 2026-10-18
- OpenAI engineer over alleged hidden recording device seeks bail, denies Saudi escape plan — Polymarket · 2026-10-18
- Free comic "The Collective" offers a strange retelling of the Hugging Face hack — Flugelwagen · 2026-10-18
- SpaceXAI's OAuth consent page draws scrutiny over vague X data-sharing language — yenkel · 2026-10-18
- ConnectWise supply chain attack: one OAuth app may expose hundreds of M365 tenants — cyb3rops · 2026-10-18
- Thomas Dietterich: proving humans did the real authorship may need signed AI logs — tdietterich · 2026-10-18