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

Current Operational Scale

• The Waymo fleet serves approximately 500 trips per week and drives over 4 million fully autonomous miles weekly.
• Operations span 15 cities across the United States.
• Cumulative total exceeds 200 million fully autonomous miles and 20 million rider-only trips.
• Safety record is 17 times better than human drivers regarding crashes causing serious injury.
• The fleet prevents a serious injury roughly every eight days based on current operating scale.

Fundamental Gaps in Physical AI vs. Digital AI

• Cost of Error Gap: Physical mistakes result in potential loss of life, unlike digital errors which typically require a simple retry.
• Latency Gap: Milliseconds matter in physical environments (e.g., a car traveling at freeway speeds covers 100 feet in one second) compared to seconds in digital applications.
• Data Gap: Physical AI lacks a pre-labeled, digitized "internet" equivalent for the real world, requiring the collection of real-world data.
• Validation Gap: High safety confidence and rigorous validation are required on day one before deployment, as the cost of error prevents an "iterate later" approach.

Lesson 1: The Demo-to-Product Gap

• A working demo represents only about 1% of the total engineering work required for a scalable product.
• Performance gains follow an exponential curve where each additional "nine" of reliability requires roughly 10 times the effort of the previous one.
• Waymo achieved a functional demo in 2010 but took 15 years to reach commercial scale due to the "long tail" of rare edge cases.
• Scaling accelerated significantly in recent years, going from 100 million miles to the next 100 million in just seven months.
• Founders must count required reliability "nines" before building demos to avoid underestimating the engineering burden.

Lesson 2: Sensing Architecture and Hardware

• Human-level performance is insufficient for full autonomy; the goal requires "superhuman" safety margins.
• Waymo utilizes a multi-modal sensor suite (cameras, LiDAR, and radar) rather than relying on a single modality.
• Camera: Provides high-resolution color but degrades in darkness and glare.
• LiDAR: Measures 3D structure directly, effective in darkness and against blinding sunlight.
• Radar: Penetrates environmental obstructions like fog, rain, and snow, and measures velocity via Doppler.
• Sensor data is fused at the encoder level into a single, unified world view rather than acting as independent backups.
• Hardware strategy prioritizes future-proofing; betting on current component costs is discouraged as prices typically drop and complexity reduces over generations.

Lesson 3: Managing Technology Waves

• Successful companies must repeatedly ride waves of technical breakthroughs (e.g., ConvNets, Transformers, VLMs) while maintaining system stability.
• New technology adoption requires a strategy for unification and simplification, not just capability gains.
• The "Waymo Foundation Model" is a multimodal, world, action, and language-aligned model powering the driver across multiple hardware generations and vehicle platforms.
• The model utilizes a "System 1 / System 2" architecture:
  • Fast Path: Processes raw sensor data for split-second, safety-critical geometric reactions (instincts).
  • Slow Path: Handles complex semantic understanding and scene context (reasoning), such as identifying a car on fire.

Lesson 4: The "Bitter Lesson" and Structural Augmentation

• General methods leveraging massive compute and data outperform handcrafted engineering approaches (Richard Sutton's "Bitter Lesson").
• Structure that channels scale wins, whereas structure that fights scale fails.
• Waymo employs "structure-augmented end-to-end" learning, combining learned embeddings with materialized structural representations (physics, road rules).
• Benefits of structural augmentation:
  • Enables real-time inference-time validation and safety checks.
  • Improves training efficiency by allowing evaluation in structured representation spaces.
  • Provides verifiable feedback signals for reinforcement learning and loss function design.

Lesson 5: High-Fidelity Simulation

• Real-world closed-loop simulation is mandatory for training and evaluating safety-critical agents, as open-loop evaluation is insufficient.
• Simulators must act as "generative world models" that understand physics, semantics, and sensing realism.
• Waymo utilizes behavioral world models (trained on real data) coupled with sensing world models (e.g., leveraging Google DeepMind's Genie 3).
• This allows the training of agents in purely synthetic, rare scenarios never encountered in the real world (e.g., planes on highways, elephants in intersections).

Lesson 6: The AI Ecosystem Flywheel

• Building a physical agent requires a three-pillar ecosystem: the Agent (driver), the Simulator (virtual playground), and the Critic (evaluation engine).
• All three pillars are based on the same foundation model architecture to ensure shared reasoning and generative capabilities.
• The Flywheel Mechanism:
  • Real-world agent deployment generates data.
  • Data grounds the simulator to increase realism and generate harder edge cases.
  • The Critic evaluates performance on these cases to provide feedback.
  • Improved agent deployment generates more data, accelerating progress.

Lesson 7: Evaluation and Metrics as Strategy

• Evaluation frameworks and metrics are more strategic than model architecture; they define product readiness and safety.
• Evaluation must extend beyond the model to cover the entire stack: physical layers, onboard behavior, off-board components, and operational processes.
• Waymo's "Safety and Readiness Framework" guides all development and deployment decisions.
• Publicly audited safety data and evidence-grade evaluation are critical for earning trust from customers, regulators, and communities.
• Earning trust through operational evidence creates a defensible business moat that cannot be replicated by algorithmic leaks alone.

Future Outlook

• Physical AI represents the next decade of the AI industry, mirroring the digital AI revolution of the previous decade.
• Ingredients for success include generative world models, proven architectures, affordable compute/sensing, and scaling laws.
• The ultimate goal is to move beyond tech as a "science project" to delivering tangible safety and quality of life improvements.
• Waymo expects to expand the Foundation Model to new applications including trucking and personally owned vehicles.