Geoffrey Irving: Internal Misjudgment on Why OpenAI Disabled CoT Monitoring

geoffreyirving · x · 2026-08-31

Geoffrey Irving discusses the meme that OpenAI 'forgot' to look at Chain of Thought (CoT). He argues that existing CoT monitoring fails out-of-distribution, causing high false positives, necessitating its disabling to train models with new capabilities. Irving notes an employee pushback claiming this is inaccurate, which he finds frightening as it suggests a fixed internal explanation, even when adjusting thresholds for low false positives at the expense of negatives.

Original post →

More from Companies & People

Companies & People channel →