A regulatory classification triggered by the compute used in training, imposing reinforced obligations on the most capable models.
The European artificial intelligence regulation deals separately with general-purpose models, which serve no assigned purpose and underpin applications their provider does not know about. Among them it singles out those presenting systemic risk and imposes reinforced obligations, covering model evaluation, adversarial testing, serious incident reporting and cybersecurity protection. The trigger is remarkable for being purely quantitative: the cumulative compute used in training, measured in floating-point operations, above a threshold set in the text. The choice is openly a second best, since capability cannot be measured directly, and it carries its flaws with it, a fixed threshold aging as method efficiency improves, and a highly capable model trained efficiently potentially falling below it. The text therefore provides for revising the threshold, and leaves the Commission power to designate a model on other criteria, restoring judgment where the number fails.
Regulation (EU) 2024/1689 presumes that a general-purpose AI model presents systemic risk where the cumulative compute used for its training, measured in floating-point operations, exceeds ten to the power of twenty-five.
systemic risk GPAI, modèle à usage général, GPAI, seuil de calcul