OpenAI burns 20% as much compute on monitoring as the model itself, SemiAnalysis says

kevinnbass · x · 2026-09-17

SemiAnalysis's Max Kan revealed that after OpenAI enhanced its chain-of-thought monitoring following the Hugging Face incidents, monitoring alone now consumes roughly 20% as much compute as the underlying model itself. He argues alignment and interpretability research will require far more compute, and that people need to internalize how compute-intensive safety work really is. Commenter kevinnbass draws an analogy to humans spending huge "compute" policing group norms, predicting AIs will soon censor and threaten other AIs.

Original post →

More from Models

Models channel →