Key points
- Waymo serves 500,000 trips a week and 4m autonomous miles weekly across 15 US cities.
- Waymo's 220m autonomous miles show its Driver is 17 times better than humans at serious-injury crashes, preventing one every eight days.
- Dolgov says a demo is 1% of the work, with each additional nine of reliability costing ten times more effort.
Waymo co-CEO on shipping physical AI without breaking things
Dmitri Dolgov argues that physical AI is arriving now, and Waymo is the most mature example. The hard part is turning a demo into a product when mistakes happen in atoms and cannot be undone. Waymo's bet is that a unified foundation model and a high-fidelity simulator can make physical AI safe enough to scale, provided evaluation is built before the technology.
Takeaways
- Physical AI demands moving fast and shipping safely. Dmitri Dolgov says mistakes in atoms have no undo, unlike digital AI. A car at freeway speed covers about 100 feet per second, so inference must run onboard. The four gaps are cost of error, latency, data and validation.
- A working demo is one percent of work. Waymo's project began in 2009 with a dozen engineers. By demo standards driving was solved in 2010, yet a service took about ten more years and five more to reach 500,000 weekly trips. Each nine of reliability costs roughly ten times more effort.
- Sensor fusion sees what cameras cannot. Waymo fuses cameras with LiDAR and also uses radar. In a Phoenix dust storm and at night, LiDAR detected pedestrians and children the camera missed. Dolgov warns against anchoring to today's component prices, since hardware commoditizes and gets cheaper.
- Waymo rebuilt around each AI wave. It used ConvNets in 2013, Transformers from 2017 and now VLMs and frontier models. The Waymo Foundation Model is a multimodal world-action model with think-fast and think-slow paths. Its generative decoder predicts other actors and plans decisions across fifth- and sixth-generation hardware.
- Structure must channel scale, not fight it. Dolgov cites Richard Sutton's bitter lesson that methods scaling with compute and data win. Waymo's structure-augmented end-to-end approach enables real-time validation and cheaper training. It also gives verifiable feedback for reinforcement learning. Structure that fights scale always loses.
- Simulation is itself a big AI model. Closed-loop evaluation requires a high-fidelity simulator, built on Waymo's behavioural and sensing world models and Google DeepMind's Genie 3. It lets Waymo train on synthetic rarities such as a plane landing on a freeway or an elephant in an intersection.
- Safety data backs Waymo's superhuman claim. Across more than 220m fully autonomous miles, the Waymo Driver is about 17 times better than human drivers at serious-injury crashes. Someone dies on the world's roads every 26 seconds, and Waymo says it prevents a serious injury every eight days.
What comes next for physical AI
- Dolgov predicts the next decade of AI will happen in the physical world, as the last decade happened in the digital world.