AI Act Uses Training Compute as Regulatory Threshold, Expert Calls It Imperfect
MartinSignoux · x · 2026-09-16
A discussion on AI regulatory thresholds: the standard currently used, including in the EU AI Act, is the amount of compute used for training. The expert argues this is imperfect, hence the idea of working on alternative standards — which was the original intent of the national AI Safety Institutes (AISIs).
The replied-to tweet takes a pragmatic angle, questioning a narrative that seems too consensus-driven when three or four competitors simultaneously discover a 10-year, 5% bill while competing against sovereign debt in a liquidity-dry market.
Related event: Debate Over AI Regulation Thresholds: Compute Caps Called Imperfect(3 posts)→
More from Safety
- ITIF: safer AI without slowing progress — the 10% extinction figure has no empirical basis — castrotech · 2026-09-17
- Dan Selsam's AI risk statement called essential reading amid safety funding debate — JacquesThibs · 2026-09-17
- David Sacks tells Amodei and Altman to slow down voluntarily, not via regulation — mikeflache · 2026-09-17
- Bloomberg: The World's Most Reliable AI Detector Has a Human Problem — bloomberg · 2026-09-17
- FT: AI bosses' safety push sparks rift inside OpenAI and Anthropic — financialtimes · 2026-09-17
- AI safety veteran tells Congress to stay out of AI regulation: "You'll make it worse" — GlenBradley · 2026-09-17