Anthropic CEO Dario Amodei is sticking to his view that capable open-weight models carry significant risks, but he's not against open weights in principle.
In a blog post, he stresses that "Anthropic has never advocated for a ban on open-weights models," adding that models "that don't have dangerous capabilities are a public good."
But Amodeiy says he has two "nightmare scenarios": authoritarian states, especially China, could build more powerful models than the US and use them to gain military superiority or surveillance capabilities. And powerful open models could be misused for cyber or bio attacks. AI labs know from their usage data that such attempts are already happening with closed models as well.
Open weights are riskier, he argues, because they lack safeguards and a released model can't be recalled. While open models also empower defenders, Amodei sees a strong imbalance on biological threats, where sufficiently capable models could weaponize viruses quickly while building defenses would take years even in the best case.
A study by the UK AI Security Institute found that open-weight models trail frontier models by a few months on cybersecurity. Amodei has warned about the capabilities of his own models before, warnings that helped prompt a US ban, so he's at least consistent in his arguments.
