FLOPs and beyond Decoding the AI Act’s Systemic Risk Criteria
summary
Panel maps how the EU AI Act labels GPAI as systemic using a 10^25 FLOPS threshold as a presumption of risk, alongside possible criteria like parameters, data quality, modalities and 10,000 registered business users. Discussion shows compute is a useful trigger but not a full proxy for harms such as hallucinations; demos include a Bing false indictment and a GDPR summary error. Promised mitigations: red teaming, tracking and reporting, watermarking, and a practical fix, zero-shot translation, that cuts hallucination rates.
AI-written summary · about
chapters (8)
- 0:00Welcome
- 1:46AI Act’s systemic risk framework
- 15:05Compute and AI governance
- 20:22AI hallucinations and systemic risks
- 30:42Systemic risk design and mitigation
- 55:46International governance and AI safety
- 1:00:47Q&A
- 1:19:50Closing remarks
related talks
01h 08mTHE GOVERNANCE OF AI: CONVERGENCE OR DIVERGENCE?
35mRisks and Mitigations for a Safe and Responsible AI
01h 09mGOVERNANCE AND REGULATION OF AI FROM THE PERSPECTIVE OF AUTONOMY AND PRIVACY
39mGPAI Code of Practice & Systemic Risk
01h 06mEMPOWERING THE AI ACT: LIMITS AND OPPORTUNITIES
44mAI, Corporate Responsibility & Democratic Legitimacy
35mThe European AI Act: a data scientist’s perspective
53mAI: Law, Policy and Common Sense Suggestions on How to Stay Out of Trouble