Training compute thresholds : key considerations for the EU AI Act : collection of external scientific studies on general-purpose AI models under the EU AI Act

This report provides an in-depth analysis of the concept of cumulative compute as a proxy for general-purpose AI (GPAI) model capabilities, with a focus on measuring and verifying training compute, defined as the computational resources used to train a model, measured in floating-point operations (FLOP). It presents two approaches to estimating training compute, namely hardware-based and parameter-based methods, and discusses their strengths and limitations, including the challenges of estimating training compute for complex architectures and the need for standardised methodologies. The report also explores the challenges of verifying declared and undeclared training runs, and discusses potential verification methods, including concordance between measurement approaches, whistleblower protection, and monitoring of large compute clusters. Additionally, it examines the regulatory context, including the EU AI Act, and provides guidance on notification triggers, including the proposed notification point at the pre-training resource commitment stage, and the need for additional clarity for what constitutes reasonable certainty about threshold exceedance. The report also discusses the importance of updating training compute thresholds to maintain their effectiveness, and proposes a framework for dynamic threshold adjustment, including regular review periods and ongoing assessment by regulatory bodies and expert groups.