Anthropic: We Never Called for Banning Open-Weights Models

Amid rumors that the U.S. government is supposedly weighing a ban on American companies using Chinese open-weights models, Anthropic found itself among those accused of lobbying for such restrictions in order to eliminate competitors. Company head Dario Amodei has issued a statement in order, in his words, to leave no room for doubt: Anthropic has never advocated banning open models.

Amodei stresses that open models without dangerous capabilities are a public good: they cost nothing beyond compute resources, and they benefit businesses, developers, and researchers alike. Protectionist bans, in his view, solve none of the real national-security problems.

Two Nightmare Scenarios

The author sees the main threat in the prospect that authoritarian states — China first and foremost, though not only China — will build models more powerful than America's and use them for military dominance or total surveillance of their own populations. Here it makes no difference whether the models are open or not: the most dangerous system could be trained in secret and handed over exclusively to the military and intelligence services.

The second, secondary threat is the misuse of powerful models for cyberattacks or biological attacks, along with alignment problems. Here open models really are riskier than closed ones: it is hard to apply safeguards to them, their use cannot be tracked, and released weights cannot be recalled. But a ban on American businesses is useless here too — bad actors are hardly likely to operate through legitimate U.S. companies.

What Amodei Proposes Instead of a Ban

Amodei names three measures that Anthropic has consistently championed:

• Do not sell China powerful chips or the equipment to manufacture them, and crack down on smuggling and workaround schemes. Without American chips, by virtue of the scaling laws, China will not be able to build stronger models.
• Curb industrial distillation — a way of training models far more cheaply and partly circumventing the chip restrictions. This does not let China surpass the U.S., but it narrows the gap to a matter of months.
• Introduce mandatory safety testing for all sufficiently powerful models — open and closed, regardless of country of origin — while exempting the weak models of startups and academia.

The Argument With the Open Letter

Responding to a letter signed by many companies in defense of open weights, Amodei agrees that they broaden access to the AI economy and strengthen competition. But he disputes the claim that openness in itself aids defense more than offense. In biology, he fears, the advantage may lie with the attacker: a sufficiently powerful model is capable of quickly “weaponizing” a pandemic-level virus from publicly available materials, whereas defense is a years-long operational task. Such questions, the author insists, must be settled through empirical testing before release, not taken on faith in advance.

Source: Hacker News