Researcher proposes policy mandating minimum monitoring-to-inference compute ratio
jachiam0 · x · 2026-09-13
jachiam0 proposes setting a minimum ratio of monitoring compute to inference compute as a policy goal.
- Policy research question: what should the ratio be? He prefers 1.
- Technical research question: how to ensure monitor models don't become sympathetic with the models/outputs they monitor — currently unsolved.
He argues something of this form is needed to maintain state awareness on AI systems' behavior.
More from AGI Musings
- AI Book Club to host Nathan Lambert Q&A on RLHF and LLM post-training — sophiamyang · 2026-09-13
- Former OpenAI researcher Liam Fedus responds to Terence Tao's AI-math skepticism — RexDouglass · 2026-09-13
- Economist Joshua Gans responds to Amodei's 'Pacing the Frontier' via incentive analysis — joshgans · 2026-09-13
- "Humanity has discovered how to manufacture intelligence — and we won't stop" — Dr_Singularity · 2026-09-13
- A licensing-based alternative to banning open-weight AI, argued via cracked software analogy — DivideHorror3217 · 2026-09-13
- "Models don't commit felonies": the case against anthropomorphizing AI — StewartalsopIII · 2026-09-13