Hacker News front pageJul 28, 2026, 06:03 UTC+8

Anthropic CEO: We never pushed for an open-weights ban, but here's what we do support

Our position on open-weights models

Dario Amodei clarifies Anthropic's stance: the company has never advocated banning open-weights models. He sees non-dangerous open models as a public good. His real worries are authoritarian regimes gaining military AI superiority and misuse for cyber or bio attacks—neither is fixed by banning US businesses from using Chinese open models. He backs three measures instead: blocking chip and equipment sales to China plus cracking down on smuggling, targeting industrial-scale distillation, and mandatory safety testing for all sufficiently capable models regardless of openness. He agrees with much of the industry open letter but pushes back on claims that open weights inherently improve safety or favor defenders over attackers.

Why it matters: Anthropic's CEO posts a personal clarification — not a dry PR piece, but a direct response to an active policy controversy. The post cleanly separates 'open-weight models as a public good' from 'two national security nightmares,' with high information density. Not scored highe...

Read the original ↗Export Markdown

Take

This is worth reading because the last few days have been messy—US officials floated banning American companies from using Chinese open-weights models, a bunch of tech companies signed a letter pushing back, and some people pointed fingers at Anthropic, accusing it of wanting a ban to protect its business. Dario's response cuts through the noise.

First, he states flat-out that Anthropic has never advocated for banning open-weights models. Non-dangerous open models are a public good—they save businesses and developers money and compute. This is consistent with his past writing.

Second, the two nightmare scenarios he actually worries about: authoritarian regimes getting superior AI for military dominance and surveillance, which has nothing to do with whether models are open—the most dangerous one might be trained in secret and handed straight to the military. And misuse for cyber or bio attacks, where open models do pose higher risk because you can't easily add guardrails or monitor usage, but banning US businesses from using them doesn't stop bad actors.

Third, the three measures he backs: block chip and equipment sales to China and crack down on smuggling, target industrial-scale distillation, and mandate safety testing for all sufficiently capable models regardless of openness. The logic chain: without US chips, China hits a scaling-laws ceiling; distillation lets them squeeze more out of fewer chips and close the gap to within months of the US frontier, so that needs enforcement; safety testing is the backstop.

One thing I'd flag: he doesn't spell out how to define or enforce "industrial-scale distillation." That's a much trickier operational problem than a blanket ban, but Dario argues it's the actual leak that needs plugging.

Source: Hacker News front page · anthropic.com
Published: Jul 28, 2026, 06:03 UTC+8 · Jul 28, 2026