Providers of General Purpose AI (GPAI) models face extensive obligations under the EU AI Act
– covering the entire lifecycle, from model development and training to system integration, release, and ongoing updates.
Highly compute-intensive models trained with more than 10²⁵ FLOPS are classified as GPAI with systemic risk. These models are subject to additional requirements. In particular, providers must carry out a continuous and comprehensive cyber risk assessment throughout the entire lifecycle.
This assessment must consider not only technical factors but also ethical and societal risks.
More details can be found in the guest Heise article by Mirko Ross here:
https://www.heise.de/hintergrund/KI-Modelle-nach-den-Vorgaben-des-AI-Acts-entwerfen-11076636.html
