Anthropic put GPU kernel optimization in the same Safeguards section as hacking and biology. WTF?

The author expresses strong disappointment that Anthropic categorized GPU kernel optimization and large language model architecture improvements under safeguards concerning hacking and biology during model training. The author argues that limiting these efficiency-focused skills restricts open weights accessibility and prevents models from running efficiently on older hardware. They criticize the restriction as an opaque safety measure implemented before user interaction. Open article →