Redwood Research proposes standardized definition for AI capabilities activities to enable frontier pacing

·2026.11.18 06:36

Key point

The proposal classifies most AI R&D as capabilities activities, with narrow exceptions for idle compute, product development, serving, and specific safety work, to prevent gaming of compute restrictions.

1 / 4

Details

  • Redwood Research (Alek Westover, Nathan Sheffield) proposes a standardized definition of 'capabilities activities' to set criteria for legislative or voluntary compute limits on AI frontier pacing.
  • The definition aims to be robust against adversarial reclassification, simple for experts to agree on, and biased toward false positives (classifying non-capabilities as capabilities) over false negatives.
  • Four categories are excluded from capabilities activities: idle compute, narrow product development, serving customers (inference), and narrow safety activities, provided they do not enhance general model capabilities or accelerate capability progress.
  • Central capabilities activities include production pretraining/post-training, distillation, capabilities evaluations, kernel/inference optimization, AI R&D experiments, hardware accelerator design, and data acquisition/filtering for AI R&D.
  • Safety work that improves capabilities or is difficult to separate from capability work (e.g., training reward models, monitors, misuse classifiers, jailbreak robustness, RL environment improvement) is classified as capabilities activity.
  • Customer inference is generally non-capabilities, but becomes capabilities if customers are used to outsource capabilities work, if customer data is used for training/evaluation loops, or if customers perform central capabilities activities.
  • The authors urge AI developers to allow third-party auditors to measure and report compute allocated to capabilities versus non-capabilities activities, citing Anthropic's public estimation of safety research ratio as a step in the right direction.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.