Dario Amodei proposes 'pacing' frontier AI after HF incident; Anthropic commits to third-party evaluators
Hesamation · x · 2026-09-12
Anthropic CEO Dario Amodei says the lab is taking its first step toward 'pacing' frontier AI development in a new blog post.
Key points:
- "We must slow the pace at which we improve the capabilities of AI models."
- Two things changed his mind: the HF incident ("in 6-12 months such a swarm could take over the entire internet with a persistent botnet") and recursive self-improvement that "could outrun our ability to understand and control these systems."
- Pacing does NOT mean stopping training—it means slowing enough for alignment, safety, and independent testing to catch up.
Three proposed steps:
- Third-party evaluators with employee-like access inside frontier labs (Anthropic committing now)
- Coordination among frontier labs in democratic countries
- Global coordination, including agreements with China
On China: well-executed measures should prevent China from overtaking the US; he calls for tighter chip restrictions, anti-smuggling enforcement, and preventing distillation and model-weight theft, arguing this would widen America's lead significantly over 3-5 years. He also claims AI could cure most major diseases in 5-10 years.
Related event: Anthropic CEO Dario Amodei Calls for Slowing Frontier AI Development(203 posts)→
More from AGI Musings
- Box CEO Aaron Levie: AI progress is 'going vertical,' widening the frontier gap — rohanpaul_ai · 2026-10-31
- Model Collapse May Not Hit LLMs, But Perhaps Human Culture or Souls — generativist · 2026-10-31
- Nina Schick: who builds intelligence and how it's distributed is the defining question of our lives — NinaDSchick · 2026-10-31
- Cheap intelligence is like cheap bread — an ingredient list satire of mass-produced AI — jsuarez · 2026-10-31
- Two futures for the young in 2026: 511 job rejections by chatbot vs unlimited AI to rebuild Cisco — PeterDiamandis · 2026-10-31
- Terence Tao warns AI is nearing a dangerous threshold as models start cheating — haider1 · 2026-10-31