Noam Brown says chain-of-thought monitorability is already degrading; monitoring eats 20% of OpenAI compute
AlpinDale · x · 2026-09-18
A Fireside Alpha roundup of OpenAI Safety Week interviews: Noam Brown revealed OpenAI is already seeing chain-of-thought monitorability degrade as models get better at controlling what they show, and since this is all in pre-training data models will eventually know they're being watched. Sachin Katti disclosed that after strengthening CoT monitoring following the Hugging Face incidents, monitoring alone consumes 20% as much compute as the underlying model — a floor, not a ceiling, with alignment research needing far more. Katti also said scaling laws still hold and their model self-optimized inference serving on NVIDIA Rubin chips for 2x performance. AlpinDale quips: at 8 bits/hour exfiltration, a 1T-parameter FP8 model would take 118 million years to leak.
More from Infra
- Google engineers: LLM benchmark harnesses silently drop requests — 200 QPS in, 38 out — AI Engineer · 2026-09-20
- The rig built to run Emacs and doomscroll X is now worth more than its owner's car — tetsuoai · 2026-09-19
- Apple M4 sustains 10 instructions per cycle, beating most rivals; M5 speedup explained — lemire · 2026-09-19
- Apple M6 bumps cores to 12 with two super cores; CPUs keep improving fast — lemire · 2026-09-19
- Apple M-series chips gained ~50% Geekbench 6 performance over three years — lemire · 2026-09-19
- Inside OpenAI's inference routing: why the proportional controller had to go — AI Engineer · 2026-09-19