"If we get AGI wrong, there's no second chance. We need coordination architecture, not constraint architecture. Constitutional AI proves it works."
Recognized that traditional safety (filtering bad outputs) doesn't scale to AGI. Instead, he built Constitutional AI: a coordination architecture where capability and values grow together. Anthropic proves that alignment is not constraint but coordination.