Redwood Researchers Propose Counting Nearly All AI Company Work as Capabilities
Summary
Alek Westover and Nathan Sheffield of Redwood Research propose defining capabilities activities as all AI company work except idle compute, narrow product development, customer serving, and six listed safety tasks. They say false positives beat false negatives when pacing compute. They urge developers to let third-party auditors measure and publicly report the split.
Key Points
- Westover and Sheffield count a GPU as idle only when the whole card sits unused for a non-negligible stretch.
- Their narrow safety whitelist covers five areas: interpretability, CoT informativeness, model persona, model organisms, and theoretical alignment work.
- Anthropic has published a share-of-AI-R&D safety figure without disclosing its definition or auditor verification, the authors note.