Anthropic CEO 发布开源模型立场：不全面禁止但需管控

Dario Amodei, Anthropic CEO just published the much-awaited blog. "Our position on open-weights models"

• Anthropic does not support a blanket ban on open-weight models.
• Amodei’s primary concern is authoritarian governments building superior military, surveillance, or repression capabilities, regardless of whether models are open.
• Powerful open-weight models carry higher misuse risks because safeguards can be removed and released weights cannot be recalled.
• Anthropic supports tighter controls on advanced chips and chipmaking equipment reaching China. It also wants action against industrial-scale distillation, which allows for cheap way to copy capabilities using far less compute.
• All sufficiently capable models, open or closed, should face mandatory testing for cyber, biological, and alignment risks before release.