The Frontier

Anthropic does not advocate for banning open-weights models; Dario Amodei sees non-dangerous versions as a public good. He views blanket software bans as ineffective against state actors.

Dario Amodei argues the debate should focus on a model's underlying power and capabilities, not its release format. If a model lacks dangerous capabilities, its weight distribution is irrelevant.

Dario Amodei identifies hardware and industrial-scale distillation as critical leverage points for AI security. He advocates for tighter export controls on high-end chips to maintain a strategic lead.

Industrial-scale distillation allows authoritarian states to train frontier-class models with less compute, bringing their frontier within months of the US. Dario Amodei calls for crackdowns on this process and smuggling.

Dario Amodei believes a model secretly trained by the CCP for military use poses a greater national security threat than any public open-weights release. Choking chip supply is key.

Dario Amodei argues all frontier AI models, regardless of release format, require rigorous safety audits before deployment. This shifts to a performance-based safety standard.

Dario Amodei highlights biological risks as a nightmare scenario, where a powerful model could design a weaponized virus quickly. Defending against such threats takes years.

Dario Amodei believes even adversaries might cooperate on biological safeguards, as pandemic-level threats ignore borders. Testing must be global to be effective.

Fritzi von speculates that recursive self-improvement (RSI) could fundamentally alter AI competition, moving past rotating state-of-the-art to create compounding advantages, similar to an industrial revolution, as Dario Amodei previously noted.