100,000 Rides a Week, Six Times Fewer Airbag Crashes: Waymo’s Co-CEO on Eighteen Years of Full Autonomy

Sarah Guo introduces Dmitri Dolgov, co-CEO of Waymo, on *No Priors*. The company began as the Chauffeur project inside Google in 2009 and spun out at the beginning of 2017. It now runs more than 100,000 paid rides a week across San Francisco, Los Angeles, Austin, and Phoenix. Guo takes the cars herself and has been campaigning, half-jokingly, for better South Bay coverage. Dolgov has been in the field about eighteen years.

He dates his start to 2006 and the DARPA Grand Challenges. The first desert race asked teams to drive about a hundred miles autonomously; nobody finished, though a lot of progress was made. A repeat of the challenge produced finishers. DARPA then staged an Urban Challenge in a mock city meant to imitate public roads. Dolgov was on the Stanford team. That, he says, was the moment it clicked: he saw the future and the benefits, and he has not looked back.

The Google project started in 2009 as a small group and grew into Waymo. Elad Gil notes that almost the entire industry traces to a handful of labs — Sebastian Thrun’s Stanford group, Carnegie Mellon, and the people who walked from those teams into Google. Dolgov’s reply is short: it is definitely a small world. There was no ImageNet, no conv nets, no Transformers, and onboard computers that look tiny next to a phone. Eighteen months after they set the first milestone, they hit it anyway.

Gil reminds him that this was considered a crazy thing to do. AlexNet did not exist yet. People said it would never work. Dolgov is honest about the internal version of that doubt: the team was not sure whether they were a little crazy or completely insane. In the university DARPA days it had been a research project. At Google the belief was that they could make it work, and that if they could, the mission justified it. They had almost no data on how long it would take or how hard it would be.

They did not treat the first years as a product plan with an end date. For a couple of years there was no product in mind and no target date. The first order of business was to explore the space. They set milestones whose only job was prototyping, learning, and understanding the boundary of the problem. After those two years they decided there was something there, and the first product they thought might be viable was what people now call an advanced driver-assist system, with a small number of years attached to the forecast.

They worked on that path, made more progress on the core technology, and decided it was the wrong path. They wanted full autonomy. The pivot was around 2013. Assist and a driverless car, in Dolgov’s framing, are not nearby points on one curve. They are different orders of magnitude. The call came roughly a decade before the rest of the industry talked that way in public, and it is the decision everything later in the conversation sits on.

Gil points at the present: something like 100,000 rides a week, five million a year annualized, more than a million miles a week and growing exponentially. What was the inflection? Dolgov does not describe a smooth curve. He describes discontinuous generational steps from the 2013 decision to today. It took about three months to go from 50,000 to 100,000 — he says miles in one pass and the hosts have been talking trips, but the point is the doubling was recent and fast.

2015 was the zero-to-one moment: the third-generation hardware suite, sensors and computer, on a custom vehicle they called Firefly. They took a few rides with nobody behind the wheel. The next skip was the fourth-generation driver on Chrysler Pacifica minivans, deployed fully autonomously in Chandler, Arizona, and opened to the public in 2020. The goal of that generation was not to capture a market. It was to do the thing repeatedly: mature the driver, evaluate it, ship on a regular cadence, put it in front of real customers, hear the feedback, and iterate.

Then they jumped to the fifth generation, on Jaguar I-PACE vehicles, which is the fleet in Phoenix, San Francisco, Los Angeles, and Austin — the platform they believe they can take to scale. City choice depends on the time horizon. For development you want the hardest problems: densest streets, highest speeds, harshest weather. New York in winter is the extreme he names. For a deployment, the fourth generation needed a medium-complexity area so they could run the whole loop and de-risk it, which is why Chandler came first. San Francisco is good for learning and a real commercial market; downtown Phoenix adds more high-speed roads. They think in operating domain, not zip codes, and future cities get three lenses: commercial market, technical complexity, and the regulatory environment.

Asked what broke through to a scalable fifth-generation driver, Dolgov is blunt: it is all about AI, full stop. Each hardware generation gets more capable and simpler, which is a boost, but the last jump was Transformers, bigger models, and compute. Conv nets after AlexNet, around 2013, had given them a lift and then plateaued in the wrong place. Architecture, he keeps saying, is an enabler. What you actually need at the level they care about is everything around it: the data engine, the flywheel of training and evaluation, and a simulator. You have to think about evaluating the driver in tandem with building it.

Internally that starts with deciding which metrics matter, then bringing data to support them. They have hundreds. Some can be scored open-loop on logged data. Some need closed-loop simulation, which means a realistic, scalable simulator. Those metrics guide training of the Waymo Driver and funnel into validation. The aggregate of the methodology is what they call the readiness and safety framework. Building and evaluating are two problems that have to move as one. That is the part he is glad the hosts asked about, because it is insanely difficult and easy to skip in a demo.

Later in the conversation Gil describes the graduate-student version of the stack: take an off-the-shelf VLM, images or video in, fine-tune so the decoder emits trajectories instead of words, and drive tens or hundreds of miles that will blow your mind. Dolgov agrees. Transformers and VLMs are mind-blowing; any good computer-science student would build that today. Then he asks the question that separates a prototype from a product: is it enough to remove the driver, drive millions of miles, and put up a safety record that is demonstrably better than humans? No.

Dolgov says they can now talk about beating human benchmarks with empirical data. The latest public point, on what Waymo calls its safety hub, is based on 22 million fully autonomous rider-only miles. Against human baselines by collision severity, they are about a factor of two better at lower-severity outcomes, and the gap widens as severity rises. For airbag-deployment collisions they are about a factor of six better. That comparison includes no notion of causality or fault. Bring fault in, and most of the remaining events are ones he describes as unavoidable.

A harder third-party check came from Swiss Re, which he calls the biggest global reinsurer. Waymo shared data; Swiss Re ran the analysis. On a smaller set, a little under four million miles, they found about a 4x reduction in damage claims versus the human baseline and a 100 percent reduction in bodily-injury claims. From Swiss Re’s point of view the result was statistically significant. That is the first time Dolgov sounds willing to talk about the mission as something happening on the road, not a slide: fewer people hurt.

So why not go bigger, faster? The United States drives just over three trillion vehicle miles a year. Waymo’s million-plus a week is rounding error. They are scaling exponentially — three months from 50,000 to 100,000 — and they have the permits. The constraint he names is process. The technology and the product are new and different. It has to be a transparent, iterative, gradual conversation with regulators, communities, and riders, earning trust rather than turning on a switch. People who have ridden use the word “magic.” People who have not are anxious. You cannot cross that in one step. Gil’s elevator-operator analogy is the same adaptation, one generation of infrastructure later.

Andrej Karpathy had told the same show that Tesla looked software-driven and Waymo hardware-centric. Dolgov does not really accept the split. It is all AI, full stop, plus the fact that you have to operate in the physical world, so seeing well is an advantage. A human can drive with one eye closed or without corrected vision; they will not be a great driver. Hardware still matters, and every generation has increased capability while simplifying and bringing cost down. That was true from four to five, and it is the primary focus of generation six, along with designing the car around the passenger: more space, flat floors, a lower step-in, sliding doors. Radar used to be expensive; put it on enough cars and it rides the usual hardware curve.

The qualitative change he cares about is not a new sensor. For years the industry was climbing, hitting what looked like a summit, and finding another range — the old mountain analogy. They had not cracked the nut, so every breakthrough was a hypothesis about yield. Now they have a driver that works and machinery that can evaluate it, which means they can answer sensor questions with data instead of theory. Ablate LiDAR and radar, drive cameras only: can it still drive? Of course. Is the performance good enough for full autonomy, at scale, against their readiness-and-safety bar? No. Change the product to driver assist with a human on the loop, and the operating point changes. Assist and driverless are not the same answer with fewer sensors.

Near-term business is ride-hailing. They are a technology company building a generalizable driver whose mission is to be the world’s most trusted, and the TAM is more than three trillion miles in the United States and more than ten trillion worldwide — delivery, long-haul trucking, personally owned vehicles. They have never been in the business of building vehicles. They partner with OEMs, tier-ones, operations companies, and network companies, because the opportunity is too large to go alone. In dense cities, younger people already want fewer cars; autonomy continues that trend. Land now wasted on parking and cars that sit idle 90 percent of the time can be used better, but those benefits only arrive at scale. Safety is the primary one they feel they have finally earned the right to talk about.

The driver, he says, can still sit in different form factors: a city hailer, a truck, a family car for a longer trip. That is complementary to what OEMs already make. They are also subsidizing Waymo trips to public-transit hubs in the cities they operate, treating robotaxis and buses as complements. They did not copy Zoox’s face-to-face seating; motion sickness is a real constraint, and facing forward has benefits. What they do see, daily, is privacy. Riders take work meetings, blast music, call in from the back seat. Without another human in the cabin, that hypothesis is showing up in feedback. The bottleneck to a billion miles a week is not, in his telling, a missing chip. It is earning trust one step at a time in a product people are not used to.

Guo asks why teams that share Waymo’s lineage went after trucking or local delivery, which looked easier, and why more autonomous miles are still happening in Waymo passenger vehicles. Dolgov’s first-order split is not car versus truck. It is full autonomy at scale versus driver assist. Slow deliveries and freeway trucking change the contours of the long tail, but they do not let you skip the core: a generalizable, evaluable driver, a noisy physical world, people doing weird things, and decisions where seconds matter. There are no silver bullets.

Freeways feel more structured on the nominal case. The rare events are still there, at higher speed and higher severity: construction, grills and mattresses falling off the vehicle in front, cars spinning out, reckless driving, motorcycles, pedestrians jaywalking. If something happens once per million miles, a prototype never sees it, which manufactures early optimism. That is why highway autonomy is not the easy mode people imagine. Go back to 2009: about a dozen people, ten routes of a hundred miles each — freeways, downtown San Francisco, around Lake Tahoe — with no intervention from start to finish. It took maybe eighteen months, with no ImageNet, no conv nets, no Transformers, tiny computers. Getting started has always been easy. Every technical wave makes the first part of the curve steeper. The complexity lives in the many nines of the long tail, which a prototype and a driver-assist system never have to pay.

Is the iteration cycle now just find bad cases, add data, retrain, like other AI companies, or are they still changing the architecture? Both, he says. The flywheel is the heart of it. The simplest version is an end-to-end model trained only to imitate human drivers: pixels in, behavior out, rebalance the data, maybe without a simulator at all. You will make progress. You might plateau in the right place for assist. You will plateau in the wrong place for a fully autonomous system. To go the distance you still loop on new data, but what sits in the loop has to change: synthetic data, closed-loop simulation, then simulation in intermediate representations, because piping everything through sensor-level simulation at scale is not practical.

After close to two decades he is more excited than ever. The vision was always there. The existential questions were whether they could build the thing, whether they could define good enough and evaluate it, whether people would use it, and whether it could be commercially viable. Operating at this scale, they have demonstrated they can build it, they are proud of the safety record, riders are using it, and it can be done in a cost-efficient way. What is left is optimization, scaling, and bringing the driver to more people and more places. Perception, safety-critical real-time systems, and a realistic simulator will spill into other robotics, but they are staying laser-focused on driving, because that is where the trillions of miles are.