A preprint links targeted expert changes in open-source AI models to much longer outputs, but reports no live hardware bit-flip demonstration.