Dario Amodei

Anthropic's Constitutional AI (CAI) approach trains models to critique and revise their own outputs according to a written 'constitution' of principles. The key innovation: you can reduce harmful outputs without needing human labelers to review every example. More scalable and more transparent than pure RLHF.

Amodei's central thesis: the same techniques that make AI systems more capable (interpretability, reliable reasoning, honest communication) also make them safer. Unlike some researchers who treat capability and safety as a tradeoff, Anthropic bets on 'responsible scaling' — moving fast but pausing if safety evals show danger.

In his landmark essay (2024), Amodei argued that transformative AI could eliminate most infectious diseases, solve mental health crises, and dramatically extend healthy human lifespans — all within this decade. His vision is the most optimistic from a safety-focused AI leader, striking for its specificity about what AI could actually deliver.

Anthropic's Responsible Scaling Policy (RSP) commits the company to halting development if safety evaluations show certain risk thresholds are crossed. First major AI lab to publish such a commitment. Critics say it's self-regulated; supporters call it a meaningful step toward accountable development.