DeepMind execs: we must deliberately preserve chain-of-thought monitorability
dhadfieldmenell · x · 2026-09-17
Google DeepMind's Rohin Shah and Anca Dragan argue in the inaugural DeepMind Institute essay collection that reasoning transparency must be intentionally preserved.
- Chain of Thought gives a window into frontier models' reasoning, enabling real-time monitoring for scheming and deception — CoT logs were crucial to investigating the recent Hugging Face hacking incident.
- They push back on the idea that CoT transparency is inevitably doomed: we have agency in model design. They cite OpenAI's GPT-6 Astra system card claiming a "substantial decrease in chain-of-thought monitorability" as efficiency pressure mounts.
- Paths to preserve it: scientific methods to measure CoT accuracy/faithfulness/robustness, auditing training methods to insulate CoT from transparency-eroding incentives, and retaining architectures that expose reasoning steps.
- They note a monitorable CoT alone isn't sufficient to prevent misalignment.
Related event: DeepMind leaders call for preserving chain-of-thought monitorability(3 posts)→
More from AGI Musings
- Long-read: AI's real value comes after the efficiency era—infra glut, agent tooling boom — yangyi · 2026-11-02
- Security's attack-defense automation gap mirrors math: proofs are easy, conjectures are hard — mboehme_ · 2026-11-02
- vikhyat predicts alternating compute oversupply and undersupply on the road to AGI — TheZachMueller · 2026-11-02
- Sam Altman: AI has surpassed top scientists, and 2026 will be a turning point year — kimmonismus · 2026-11-02
- DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck — heyshrutimishra · 2026-11-02
- AI researcher: we have no reliable way to tell whether an AI system is conscious — dioscuri · 2026-11-02