Anthropic’s Dario Amodei argues that AI interpretability is essential for safety, proposing tools to decode model internals before they become dangerously powerful.