Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The Work

Building AI for the physical world requires transitioning from "move fast and break things" to a rigorous "move fast and ship safely" philosophy, as the cost of error involves human lives rather than digital tokens. Unlike digital AI, physical AI demands high reliability from day one, necessitating a "nines" approach where each order of magnitude in performance requires fundamentally different engineering strategies. Key technical pillars include utilizing redundant sensing modalities—cameras, LiDAR, and radar—to overcome environmental limitations, and adopting a "structure-augmented end-to-end" model that balances learned representations with physical constraints. Success relies on a powerful flywheel comprising the agent, a high-fidelity simulator, and a critic, all guided by quantitative metrics. Ultimately, earning public trust through transparent, evidence-based safety records is the most significant competitive advantage in deploying autonomous systems at scale.