TL;DR

Dario Amodei has published a direct rebuttal to claims that Anthropic wants open-weights models banned, saying the company “has never advocated” for such a ban. He backs three alternatives instead: chip export controls, a crackdown on industrial-scale distillation, and mandatory safety testing for all sufficiently capable models. The post lands days after Anthropic conspicuously declined to sign an industry letter opposing restrictions on Chinese open models.

Anthropic was the notable absentee from last week’s open letter backing open-weight AI, which by Monday carried more than 130 signatures including OpenAI, Nvidia, Microsoft and Google. That absence drew accusations that the company was pushing for restrictions to shield its own premium-priced models from cheaper competition.

Amodei’s answer is that models without dangerous capabilities are “a public good” costing nothing beyond the compute to run them. His stated worries are elsewhere: authoritarian states building more capable models than the US, and powerful models being misused for cyber or biological attacks. Neither, he argues, is fixed by stopping American businesses from downloading Chinese weights — “bad actors are unlikely to be legitimate US businesses.”

Where he does break with the letter’s signatories is on the security case for openness. The letter argues broad access helps defenders more than attackers. Amodei thinks the reverse is at least as plausible, citing biology as a field where weaponisation could outrun defence by years.

Politico, which first reported the open letter, noted the reaction from former White House AI czar David Sacks, who posted that Anthropic “won’t stop until they kneecap open source.”

For UK readers, the substantive point is the third proposal. Amodei wants pre-release testing applied to open and closed models alike, with startup and academic models exempted — a design close to what Britain’s AI Security Institute already does voluntarily. AISI’s own recent work is cited in the post’s footnotes, and its assessment that open-weight release creates “a persistent and irreversible risk of misuse” for dangerous-capability models is quoted directly.

Looking Forward

The disagreement now on the table is narrower than “ban or not”. It is whether capability testing should become a legal precondition for release, and who runs it. Amodei concedes such a regime only works if it is global, which means Chinese participation. UK firms building on open weights should watch whether testing obligations attach to model publishers alone, or reach downstream to deployers too.