Summary

Waymo says it has given over 20 million fully autonomous rides, with 10 million occurring in the last seven months. It took eight years from starting fully autonomous operations to providing public rides in four cities; earlier this year, Waymo launched four cities in one day. Dolgov calls the shift a “phase transition” into “rapid, parallel global commercialization.”
The technical stack is broader than an end-to-end driving model. Waymo’s foundation model powers a “driver, simulator, and critic” and combines multimodal sensing, physical-world dynamics, agent behavior, and language alignment. Dolgov’s distinction is “end-to-end, and then what else?” Structured intermediate representations enable runtime validation, closed-loop training and evaluation, and richer reinforcement-learning rewards.
Waymo reports safety performance across its growing mileage. Across more than 170 million fully autonomous miles, Dolgov says the Driver is more than 13 times safer than a human driver for serious-injury-causing collisions in the cities where it operates. At the current scale, he says that means “preventing a serious injury every eight days.”
Sixth-generation hardware is being optimized simultaneously for performance, lower cost, simplification, and volume production. The sixth-generation Driver powers the OHI vehicle platform; fully autonomous operations began earlier this year and are currently employee-only, with access coming to all riders later in the year. Dolgov describes the interior as feeling “like a living room.”
Dolgov argues autonomous driving repeatedly attracts premature optimism because breakthroughs accelerate the easy opening stretch without removing the deployment long tail. AV is “very easy to get started” but extremely difficult to carry through to full autonomy and superhuman performance; convolutional nets, transformers, and large language models reshape “the early part of the curve” but do not change “the long tail of it.”
The near-term business agenda is geographic replication and execution. Waymo was operating fully autonomously in 11 cities, with further US expansion and plans to offer service in London and Tokyo this year. New-city work still requires data collection, environmental characterization, validation, operations, and community trust, although Dolgov says that, more often than not, the Driver is now generalizing “incredibly well.”

Deep dive

1. A 2005 challenge turned autonomy into a 20-year mission

Dolgov spent a year in Japan, attended high school in the United States, then returned to Moscow Institute of Physics and Technology—the school his parents had attended—to study math and physics before earning a PhD in AI. He says those years gave him a technical foundation and, more importantly, the ability to learn and explore independently.
Dolgov calls the DARPA Urban Challenge a “light-switch” moment: autonomy combined technically compelling technology, a powerful mission, and a real product he could personally experience. “Nothing else came close,” and he “never looked back.”
Starting in 2009 as Google’s self-driving car project, roughly 12 people set two improbable goals: accumulate 100,000 fully autonomous miles and complete 10 difficult, 100-mile Bay Area routes without intervention. Working across hardware, calibration, algorithms, tools, and in-car UX, they finished both in about 18 months and concluded that full autonomy was worth pursuing as a product.

2. Breakthroughs shorten the beginning, not autonomy’s long tail

Dolgov’s explanation for AV hype cycles: convolutional nets, transformers, and large language models can produce rapid progress in the early part of the problem, encouraging the belief that “now the problem is going to be” solved. Yet driving remains “very easy to get started” and “very difficult to take it all the way” to a real, fully autonomous product with superhuman performance.
Each breakthrough “reshapes the early part of the curve” but “doesn’t change the long tail of it.” Persistence therefore came from understanding the problem’s true difficulty, not searching for “easy wins, quick solutions, or silver bullets.”
Dolgov says the mission supplied the stamina: worldwide, somebody loses their life to a crash on the roads every 26 seconds. The combination of knowing the mission is important and understanding what the team is up against gives Waymo the stamina to go the distance.

3. Waymo’s foundation model spans driver, simulator, and critic

At the center of Waymo’s AI ecosystem is a foundation model powering three related but distinct pillars: “the driver, the simulator, and the critic.” It must understand physical dynamics, good driving behavior, and how the Driver’s actions affect cars, pedestrians, cyclists, and other agents.
Dolgov characterizes it as a “multimodal world-action-language model.” Images and video are joined by lidar and radar; precise 3D geometry and physics meet behavioral prediction; and language alignment imports general world knowledge useful for driving’s semantics and “deep social aspects.”
Buhler’s end-to-end question draws an important qualification: the foundation model runs from sensors to decisions or actions and learns rich representations between system components rather than relying on an engineered perception-planning interface. But Dolgov rejects the binary framing—“it’s end-to-end, and then what else?” There is “a massive difference between using end-to-end and purely relying on it.”
Waymo augments learned representations with structured, materialized intermediate representations. Dolgov says those layers are critical for runtime validation, closed-loop evaluation and training, and richer reinforcement-learning rewards—requirements that a prototype or driver-assistance system might not need, but a fully autonomous system deployed at scale does. Buhler also says that human feedback from support and drivers is essential to this type of architecture; Dolgov agrees.

4. Hardware simplification and city replication drive the scale-up

The sixth-generation Waymo Driver is its most advanced hardware and sensor suite, while also emphasizing simplification, “drastic cost reduction,” and high-volume production. It powers the OHI vehicle platform. Fully autonomous operations began earlier this year and are currently open only to employees, with access coming to all riders later in the year. Dolgov says the interior feels “like a living room.”
Buhler frames the acceleration as roughly 16 years to 100 million miles and about six months to 200 million. Dolgov adds the commercial milestones: eight years from starting fully autonomous operations to public rides in four cities, then four city launches in one day; more than 20 million rides overall, including 10 million in seven months. “That’s what exponential scaling looks like.”
Entering a city still means collecting data, characterizing its environment, validating the Driver, handling operational components, and earning community trust. More often than not, Dolgov says, the Driver is now generalizing “incredibly well,” leaving high-fidelity evaluation and rigorous validation as the central pre-deployment work.
Dolgov says Waymo is how he gets around, including a freeway ride from Palo Alto to San Francisco, and that his family uses it. His three children are annoyed when a human has to drive; their driving call-outs now focus on “doggies and Waymos.”

5. Safety evidence accompanies the push into global commercialization

Dolgov calls safety the “nonnegotiable foundation” that must shape model architecture, training, evaluation, and team culture from day one. Reaching an initial 90% capability is fundamentally different from achieving the subsequent “next nines” required for autonomous deployment.
At more than four million fully autonomous miles per week, Waymo had accumulated over 170 million such miles. Dolgov says its Driver was more than 13 times safer than a human driver for serious-injury-causing collisions in operating cities—an improvement he says means preventing a serious injury every eight days at the current scale.
He also described a case in which a person—he thinks it was a young woman—lost control of an electric scooter and fell in front of a Waymo. The Driver swerved and braked with what Dolgov called superhuman accuracy and reaction time, and everyone walked away.
His other memorable example involved a pedestrian hidden behind a bus. The Driver could not see through the vehicle, but a sparse lidar return produced by the person’s moving feet beneath the bus let the AI detect the pedestrian, predict the emergence, and react defensively. Dolgov says the capability “blew my mind.”
Asked about the next five to 10 years, Dolgov said Waymo is in “heads-down execution mode,” having moved from “intentional, sequential de-risking” to “rapid, parallel global commercialization.” From 11 fully autonomous cities, the plan is deeper coverage, additional US geographies, and international service in London and Tokyo this year.

FundaAI's Zhou Mo on the AI Industry After Big Four Tech Earnings Digital Freedom, AI Regulation, and the Fight for the Western Internet | The a16z Show