The distillation and open-weights debate of the past few days reached a point where Anthropic CEO Dario Amodei apparently decided to weigh in himself. On July 27 he published a post with one blunt takeaway: Anthropic has never advocated for a ban on open-weight models.
What the fight is about
The backdrop: proposals are circulating in Washington to ban US companies from using Chinese open-weight models like Kimi K3. In response, a long list of tech companies — led by Nvidia — signed an open letter defending open models. And somewhere in the middle, Anthropic caught the blame: some accused the company of wanting to ban open models to protect its own business.
That’s exactly what Amodei pushes back on. Open models without dangerous capabilities are a public good, he writes — they cost nothing beyond the compute to run them, and they deliver real value to developers, researchers and businesses. A blanket ban, he adds, wouldn’t solve any of his serious security concerns anyway.
What Amodei actually fears
Two scenarios keep him up at night. First: that authoritarian states — chiefly China’s leadership — build more powerful AI than the US and use it for military superiority or deep repression. Whether such a model is open or closed is irrelevant here — the most dangerous model might be one trained in secret and handed only to the military. Second: that powerful models get misused for cyber or biological attacks. Here he concedes that open weights can carry higher risk, because guardrails are hard to bolt on and, once released, weights can’t be pulled back.
The three measures he really stands for
Instead of a ban, Amodei calls for three things: don’t sell powerful chips to China and crack down hard on smuggling; shut down industrial-scale distillation operations; and put all sufficiently capable models — open and closed alike — through mandatory safety testing. On that last point he sees something close to a consensus forming, even across vendors and borders.
My take
This is a strikingly candid piece — Amodei takes the wind out of the “protectionism dressed up as safety” accusation by separating the positions cleanly: he has nothing against competition from open models, but plenty against authoritarian escalation and untested high-risk models. Whether or not you share his biology worry, the split he draws — “open vs. closed” versus “tested vs. untested” — is the smarter framing. My hunch is that’s exactly where this lands.
Sources: