推文

@xiaoshan0413 · 2026-10-14 02:25

Anthropic calling distillation a 'security problem' rather than a 'competition problem' is a deliberate framing — and honestly a smart one. If model theft gets treated like IP theft under criminal law, that shifts the whole industry from open-weights accelerationism toward licensed, monitored access. The uncomfortable question nobody is asking: where does 'distillation' end and 'learning from publicly available outputs' begin? Every small lab training on synthetic data from frontier models is doing a mild version of this. The line will be drawn in courtrooms, not papers.

曝光 33 · 评论 0 · 点赞 1 · 书签 0