What's this about?
After a week of speculation and NVIDIA's pro-open-weights letter, Anthropic published where it actually stands. The gist: it says it has never advocated banning open-weight models and calls harmless ones "a public good," but it wants restrictions — notably on distillation — and frames its central worry as authoritarian governments (the CCP most of all) building frontier models to achieve lasting military superiority. It landed the same day Kimi K3 proved an open Chinese model could reach the frontier, so nobody missed the timing.
What exactly is "distillation"?
Distillation is training a smaller or rival model to imitate a stronger one's outputs — an apprentice copying the master's finished dishes without the master's training. Anthropic wants limits on others distilling from its models; critics note Anthropic itself trained on vast scraped data, so a rule that "our outputs are off-limits but everyone else's inputs were fair game" reads as one-sided.
What's the community saying?
- "Ladder-pulling": the top-voted read is that this is regulatory capture dressed as safety — "a whole lot of words to say we don't support open weights."
- The distillation double-standard: "we can digest all information legally or not because we're the good guys; you can't digest from us."
- Who decides "dangerous"? Several note that "harmless open models are fine" is empty when Anthropic (via lobbying) helps define harmless.
- The China framing itself — some call it standard geopolitical playbook, others find the military-superiority risk genuinely reasonable.
- A minority defends it as a fair, honestly-stated position even if self-serving.