Governance

Compute Threshold

Threshold of computational resources (FLOPS) used to classify AI models for additional regulatory requirements.

Definition

Compute Threshold is a quantitative metric used in AI regulation to classify AI models based on the computational resources used in their training, triggering enhanced regulatory requirements.

EU AI Act Threshold (Article 51):

  • 10^25 FLOPS: A general-purpose AI model is presumed to have high impact capabilities (and thus systemic risk) when the cumulative amount of compute used for its training is greater than 10^25 floating point operations
  • Measured in: Floating point operations (FLOPS)
  • Adjustable: The Commission can update this threshold through delegated acts based on technological developments

Rationale for Compute Thresholds:

  • Proxy for Capability: Computational scale correlates with model capabilities
  • Measurable: More objective than capability-based assessments
  • Predictable: Allows providers to know classification in advance
  • Scalable: Can be adjusted as technology evolves

Current Context (2024):

  • GPT-4 estimated at ~10^25 FLOPS (near threshold)
  • Newer frontier models likely exceed this threshold
  • Threshold roughly corresponds to current "frontier" models

Obligations When Exceeding Threshold:

  • Enhanced model evaluation and testing
  • Systemic risk assessment and mitigation
  • Serious incident tracking and reporting
  • Adequate cybersecurity protections
  • Cooperation with authorities

Limitations:

  • Does not account for training efficiency improvements
  • Same compute can produce different capability levels
  • Does not address fine-tuned or open-weight models
  • May become outdated as technology advances

Sources

  • EU AI Act Article 51
  • Executive Order 14110