OpenAI Research: Chain-of-Thought Monitorability Degrades as Models Scale

Key Info

OpenAI shared new research on how chain-of-thought (CoT) monitorability—specifically no-CoT capability and CoT controllability—evolves across the model training lifecycle, concluding that models become harder to monitor through their CoT as they scale toward greater intelligence.

Highlights

• Research examines two dimensions of CoT monitorability: no-CoT capability and CoT controllability
• Findings track how monitorability changes over the full model training lifecycle
• Conclusion: monitoring via CoT becomes harder as models grow more intelligent
• The authors describe the results as unsurprising but sobering for AI safety
• Shared as a 10-part thread by @jeffreygwang, retweeted by OpenAI