Waymo Co-CEO Dmitri Dolgov: "Move Fast And Ship Safely"

Dolgov opens by distinguishing between AI in the digital world and AI in the physical world, noting that Waymo currently operates over 500 autonomous trips per week across 15 US cities with over 4 million fully autonomous miles driven weekly—equivalent to 300 years of average American driving per week.

The first major lesson addresses the enormous gap between a working demo and a production product. Waymo achieved a 90% capability demonstration in 18 months (2009-2010), completing 100,000 autonomous miles across 10 routes before any major AI breakthroughs like ConvNets or Transformers. However, scaling to a reliable service took 15 additional years. Dolgov explains that reliability and performance follow an exponential curve of "nines"—each additional nine of reliability (99%, 99.9%, etc.) requires roughly 10 times more engineering effort. Demos might need one nine of reliability, while fully autonomous vehicles in the public require multiple nines. With millions of miles driven weekly, rare events become daily problems.

The second lesson establishes that required reliability levels fundamentally dictate technical architecture choices. Different sensing approaches have performance-versus-effort curves that flatten at different points. Waymo chose a multi-modal sensing approach using cameras, LiDAR, and radar rather than relying on cameras alone (as humans do) because superhuman autonomous driving requires complementary sensors. Cameras fail in darkness and glare; LiDAR provides 3D structure; radar penetrates weather and measures velocity. These sensors are fused into a unified world view rather than serving as backups.

Lesson three emphasizes the need to repeatedly ride waves of technological innovation. Waymo has rebuilt its driver around major breakthroughs: ConvNets (2013), Transformers (2017), and VLMs/world models (recently). The critical challenge is not prototyping new technology but integrating bleeding-edge research into production without regressions, reducing complexity simultaneously, and doing this repeatedly across multiple innovation waves. Dolgov advises setting the launch bar to demand both breakthrough performance and radical simplification.

The fourth lesson applies the "bitter lesson"—that general methods leveraging massive compute and data outperform handcrafted engineered solutions. Waymo bets on high-capacity foundation models trained on vast data, then distills into efficient models running on vehicles. However, this must be balanced with intentional structure. Dolgov introduces "structure-augmented end-to-end" learning, which augments learned embeddings with materialized structured representations (encoding physical laws, road rules, predictable object behavior). This structure must channel scale rather than fight it. Structure validation enables safety checks at inference time, improves efficiency in training and evaluation, and provides strong feedback signals for reinforcement learning.

Lesson five establishes that high-fidelity closed-loop simulation is essential. Open-loop evaluation (observing input-output pairs) is insufficient for safety-critical agents. Closed-loop simulation requires understanding physics, semantics, traffic, and weather. Waymo built behavioral world models (before the term became popular) and now builds sensing world models that produce realistic sensor outputs. This enables training and evaluation on previously unseen synthetic rare scenarios (planes landing on freeways, elephants in intersections, etc.) before real deployment.

The sixth lesson describes building an ecosystem of three interconnected AI systems: the agent (the driver), the simulator (virtual learning playground), and the critic (rigorous evaluator). All three share fundamental reasoning and generative capabilities through a shared foundation model. This creates a powerful flywheel: real-world deployments generate data that grounds the simulator; the simulator generates harder edge cases; the agent learns and improves; improved agents deployed in the world generate more data. Metrics must guide this flywheel toward desired outcomes.

The final lesson elevates evaluation and metrics to strategic importance. Model architectures and algorithms spread quickly, but without good metrics, development is flying blind. For physical AI, evaluation must go deeper than model-level performance—validating every component from physical layers to on-board behavior to off-board systems and operational processes. Waymo calls this the "safety and readiness framework." Trust is earned gradually through relentless field proof, not through technical cleverness or demos. Waymo's practice of publicly publishing safety data and research creates competitive advantage that's difficult to replicate.

Dolgov concludes that these lessons compound together: setting the right reliability bar determines technical approach; intentional structure boosts scaling; riding tech waves reaches required capability levels; the AI ecosystem with agent, simulator, and critic guided by metrics creates the flywheel enabling superhuman safety performance. Current data shows the Waymo driver is approximately 17 times safer than human drivers regarding serious injury crashes across its operating areas. He frames physical AI as the next decade's frontier—analogous to where digital AI was years ago—with all necessary ingredients now available.