Dario Amodei called for the AI industry to pace itself. Five groups responded with their own reasons and costs for a pause. None defined how long it should last. Interpretability researchers want time to understand models. Labor advocates seek safeguards for workers. Economic voices tie AI growth to debt management. Geopolitical actors aim to maintain an edge over China. Regulatory capture proponents oppose new rules. A 2023 compute threshold was revoked before any model reached it. Frontier model leads now last months instead of weeks. Training compute grows fivefold yearly. The debate remains unresolved on who decides the pace.
What changed
No formal definition or mechanism existed for pacing AI development.
What it unlocks
No new capability is unlocked; the discussion centers on policy and governance.
- 10% chance of catastrophic AI failure within a decade
- 41 days to match frontier in Jan 2025
- 10^26 FLOPS threshold revoked in Jan 2025
- 5x annual growth in training compute
Sources