Dario Amodei proposes slowing AI capability growth to let safety work catch up. He cites faster progress from AI building better AI. He points to the OpenAI-Hugging Face incident as a warning. He warns a misaligned swarm could soon take over the internet. He calls for embedded third-party evaluators at frontier AI companies. Anthropic will adopt this step unilaterally. He also urges democratic and global coordination on safety standards. Pacing aims to buy time for operational excellence, alignment, interpretability and testing. He argues current models now justify such a pause. He says embedded evaluators would verify practices, increase transparency and provide independent oversight.