CoT monitorability not abandoned yet, but new techniques risk a race to the bottom
DavidSKrueger · x · 2026-09-03
AI safety researcher David Krueger acknowledges sloppiness in his earlier claim that CoT monitorability had been abandoned, crediting Leon Lang for the correction. Lang's point: chain-of-thought monitorability hasn't actually been given up yet, though he worries new techniques could trigger a race to the bottom that makes monitoring models' reasoning increasingly difficult.
More from AGI Musings
- Gary Marcus: Generative AI Will Likely Be Seen as a Tragic, Costly Mistake by 2030 — GaryMarcus · 2026-09-03
- AI doom is plausible, which is why the Pause crew are controlled opposition, argues researcher — georgejrjrjr · 2026-09-03
- Engineer: anthropomorphic AI descriptions on podcasts are misleading and non-predictive — jamesdouma · 2026-09-03
- Sam Altman: One California Almond Uses More Water Than 38,000 ChatGPT Queries — Polymarket · 2026-09-03
- kalomaze: compute multipliers don't matter if objective specification is unsolved — kalomaze · 2026-09-03
- World Bank Nigeria RCT: GPT-4 tutoring boosted learning by 0.31 SD, worth 2 years of schooling — alexvoica · 2026-09-03