Anthropic has published a 10,000-plus word paper arguing that the real AI danger isn't just job losses, the warning CEO Dario Amodei has been repeating all year. It's recursive self-improvement: AI systems capable of designing and training their own successors.

The paper reveals Claude now writes over 80% of Anthropic's code, and floats a coordinated pause on frontier AI development if rival labs agree to verifiably do the same.