Amodei: We never asked to ban open weights — just test them

Anthropic was the conspicuous no-show on Jensen Huang's "Open Weights and American AI Leadership" letter — the one Huang used for his first-ever X post on July 24. The letter started with 25 signatures (Nvidia, Meta, Microsoft, Hugging Face, Mistral) and doubled to 50 within days as Google and OpenAI reversed course and signed. Anthropic and Amazon stayed out. Dario Amodei has now answered the obvious question: "Anthropic has never advocated for a ban on open-weights models."

He concedes Huang's core points — open weights expand access, sharpen competition, hand customers control. Where he splits is the claim that open weights structurally favor defenders over attackers. His named worry is biology: sufficiently capable models, he argues, could compress the path to weaponizing pandemic-class viruses, and a downloaded model can't be rolled back.

Instead of a ban, three asks: tighter export controls on advanced chips and fab equipment to authoritarian governments plus a smuggling crackdown; enforcement against industrial-scale distillation — training on the outputs of frontier systems; and mandatory safety testing applied equally to open and closed models. That last one is the real ask, and it's the one the letter's signatories will fight hardest. The stated target isn't Meta or Mistral — it's "companies backed by an authoritarian state seeking to overtake the US at the frontier."

Why it matters: the open-vs-closed fight just became a fight over who has to pass a test — and Washington is drafting the rules right now.