Anthropic declined to sign an industry coalition letter on open-weights models. Instead it published its own position, stating it doesn't seek a ban on the practice while arguing for chip export controls, measures against model distillation, and required safety testing.
Distillation, here, means training a smaller model using a larger one's outputs, approximating the expensive model's capability without paying for its development. It's the technique that makes competitive advantage hard to hold even with a closed model, because access to the output already gives away part of the value.
The position describes precisely where the company thinks control is possible. Published weights don't come back and can't be traced. A chip, by contrast, is a physical object with an identifiable supply chain, and safety testing is an auditable process.
Refusing a collective signature and publishing separately is also market signalling. It differentiates the company from competitors on a subject where public opinion and regulation are still forming.
What the discussion doesn't resolve is what's already distributed. A model with published weights stays available regardless of any later agreement between companies or governments.