No PriorsNo Priors Ep. 87 | With Co-CEO of Waymo Dmitri Dolgov
Sarah Guo and Dmitri Dolgov on waymo’s Co-CEO Explains Cracking Full Autonomy And Scaling Robo‑taxis.
In this episode of No Priors, featuring Sarah Guo and Elad Gil, No Priors Ep. 87 | With Co-CEO of Waymo Dmitri Dolgov explores waymo’s Co-CEO Explains Cracking Full Autonomy And Scaling Robo‑taxis Dmitri Dolgov, co-CEO of Waymo, traces the company’s evolution from Google’s 2009 Chauffeur project to today’s large-scale robo‑taxi service delivering over 100,000 paid rides per week.
At a glance
WHAT IT’S REALLY ABOUT
Waymo’s Co-CEO Explains Cracking Full Autonomy And Scaling Robo‑taxis
- Dmitri Dolgov, co-CEO of Waymo, traces the company’s evolution from Google’s 2009 Chauffeur project to today’s large-scale robo‑taxi service delivering over 100,000 paid rides per week.
- He explains how successive “generations” of the Waymo Driver, driven largely by AI breakthroughs (ConvNets, transformers, larger models, VLMs) plus a robust simulation and evaluation stack, enabled a decisive jump to reliable full autonomy.
- Waymo now claims safety performance significantly better than human drivers, and Dolgov stresses that building and *evaluating* the driver—along with a cautious, trust‑earning rollout—are as critical as the core AI itself.
- Looking forward, Waymo’s primary near‑term business is ride hailing, but its generalized driver is intended to extend to deliveries, trucking, and other vehicle platforms in partnership with OEMs.
IDEAS WORTH REMEMBERING
7 ideasFull autonomy required a pivot away from driver assistance toward solving the entire driving task.
Waymo initially considered an advanced driver‑assist product but decided around 2013 that incremental assistance wouldn’t deliver the safety and impact they wanted; the company refocused on removing the human from the loop entirely.
AI architectural leaps plus a strong evaluation machine enabled the ‘nut‑cracking’ moment.
ConvNets (post‑AlexNet) gave an early boost but plateaued; transformers, larger models, more compute, and a rigorous data/simulation‑driven evaluation pipeline together produced the performance jump that made broad deployment viable.
Evaluating an autonomous driver is as hard and important as building it.
Waymo relies on hundreds of safety and performance metrics, large‑scale simulation (open‑loop and closed‑loop), and a formal readiness and safety framework; this system underpins their confidence in deployment decisions and safety claims.
Waymo’s autonomous service now empirically outperforms human drivers on safety.
Based on 22 million+ fully driverless miles, Waymo reports roughly 2x fewer low‑severity collisions and 6x fewer airbag‑deployment crashes than human benchmarks, and a Swiss Re study found 4x fewer damage claims and zero bodily injury claims in its dataset.
Scaling safely is constrained more by trust and process than raw technology.
Despite exponential mileage growth, Dolgov emphasizes gradual rollouts, transparency with regulators and communities, and the fragility of public trust as the primary gating factors to faster deployment, not just capital or engineering capacity.
Redundancy in sensing is currently necessary to meet a very high safety bar.
While in theory a camera‑only system can ‘drive,’ Waymo’s empirical tests show that removing LiDAR or radar degrades performance below their standards for full autonomy at scale; simplification is pursued carefully while maintaining required safety ‘nines.’
Ride hailing is the near‑term focus, but the Waymo Driver is designed to be generalizable.
Waymo aims to be a ‘driver’ company that partners with OEMs and ecosystem players; after ride hailing, the same core driver could power delivery, long‑haul trucking, and eventually personally owned autonomous vehicles.
WORDS WORTH SAVING
5 quotesThe complexity is in the long tail of the many, many nines.
— Dmitri Dolgov
It’s all about AI. Full stop.
— Dmitri Dolgov
We are starting to actually earn the right to talk about realizing the mission of making roads safer.
— Dmitri Dolgov
Trust is this thing that’s hard to earn, but very easy to lose.
— Dmitri Dolgov
We think of what we’re doing as building the driver. You put the driver in, but you still need the car.
— Dmitri Dolgov
QUESTIONS ANSWERED IN THIS EPISODE
5 questionsIf the main bottleneck is public trust, what concrete steps could further accelerate societal comfort with fully driverless vehicles?
Dmitri Dolgov, co-CEO of Waymo, traces the company’s evolution from Google’s 2009 Chauffeur project to today’s large-scale robo‑taxi service delivering over 100,000 paid rides per week.
How might future AI breakthroughs—beyond transformers and current VLMs—change Waymo’s sensor requirements or safety margins?
He explains how successive “generations” of the Waymo Driver, driven largely by AI breakthroughs (ConvNets, transformers, larger models, VLMs) plus a robust simulation and evaluation stack, enabled a decisive jump to reliable full autonomy.
At what point would Waymo feel comfortable significantly simplifying the sensor suite without compromising its current safety advantage over humans?
Waymo now claims safety performance significantly better than human drivers, and Dolgov stresses that building and *evaluating* the driver—along with a cautious, trust‑earning rollout—are as critical as the core AI itself.
How will the economics of robo‑taxis compare to private car ownership once Waymo reaches much larger scale in multiple cities?
Looking forward, Waymo’s primary near‑term business is ride hailing, but its generalized driver is intended to extend to deliveries, trucking, and other vehicle platforms in partnership with OEMs.
What governance or regulatory frameworks would best balance innovation, competition, and safety in a world where AI, not humans, is doing most of the driving?
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome