RC RANDOM CHAOS

Anthropic's Amodei: We never called for banning open-weights models

· via Hacker News

Original source

Our position on open-weights models

Hacker News →

Anthropic CEO Dario Amodei is pushing back on claims that his company wants US regulators to ban Chinese open-weights models. He argues that open models without dangerous capabilities are a public good and that protectionist bans miss the actual national-security threats. Blocking US businesses from using Chinese models, he notes, does nothing to stop bad actors—who aren’t legitimate US businesses in the first place—and merely shields American AI firms from competition, which he says has never been his aim.

Amodei frames two core worries: authoritarian states building frontier models that hand them lasting military or surveillance advantages, and powerful models being misused for cyber or biological attacks. Open weights are relevant to the second concern because guardrails can be stripped and released weights can never be recalled, but he insists the origin or openness of a model isn’t the deciding factor—the most dangerous system might be one trained secretly for the People’s Liberation Army. Instead of category bans, he backs three measures: keeping advanced chips and chipmaking gear out of China (including cracking down on smuggling), deterring industrial-scale model distillation that lets Chinese labs stay months behind the US frontier on fewer chips, and mandatory pre-release safety testing for all sufficiently capable models, open or closed.

On the industry open letter defending open weights, Amodei agrees with much of it—open models widen access, boost competition, and give customers control—but rejects the assumption that openness inherently helps defenders more than attackers or makes safeguards easier. He points to biology’s potential attacker-defender asymmetry, where a capable model might weaponize a pandemic-class virus far faster than defenses can be mounted. Whether open models raise risk, he argues, should be settled empirically through rigorous testing rather than assumed in advance, and effective testing would ultimately need to be global—requiring even China’s cooperation.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.