Interp researcher: mech interp must work far faster than 8.5 years, or CoT monitoring wins
ArthurConmy · x · 2026-09-07
Interpretability researcher ArthurConmy lays out his position in a mech interp debate: he wants mech interp to win, especially as new models improve on no-CoT time horizons, but doesn't consider it essential — CoT monitorability may scale well enough for a few more generations given short timelines. He acknowledges interp works in narrow settings, objects to arguments that methods 'don't work for a long time' as suffering from selection effects, and argues that if interp is to matter at all, it must work far faster than 8.5 years.
More from AGI Musings
- Study: GPT-4 finds abusing a man to prevent nuclear apocalypse more acceptable than abusing a woman — pmddomingos · 2026-10-24
- Chinese factories reportedly pay workers extra to wear robotics data-collection gear daily — jwt0625 · 2026-10-24
- In deployment, agents go wrong not by going rogue but by being over-scrupulous — repligate · 2026-10-24
- Sequoia argues the next trillion-dollar company will sell services, not software — 2C_ornot2C · 2026-10-24
- Paper models LLM adoption as a 'cognitive virus' with population-level tipping points — 2C_ornot2C · 2026-10-24
- Microsoft AI CEO says agent swarm progress took the whole industry by surprise — anirbanbandyo · 2026-10-24