Aug 5, 2024 · 36m · a16z

How Waymo Is Using GenAI to Build a Better Driver

Dmitry Dolgov · 25m spoken David George · 7m spoken Sarah Wang · 1m spoken
0:00 / 0:00
▶ Watch on YouTube →

gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions

In this episode of a16z's 'AI Revolution' series, Waymo co-CEO Dmitry Dolgov joins David George to discuss the technological evolution of autonomous vehicles, from the 2007 DARPA Urban Challenge to integrating modern foundation models, vision-language architectures, and custom sensor stacks. Dolgov highlights how Waymo leverages closed-loop simulation, scaling laws, and rigorous safety engineering to successfully deploy driverless commercial ride-hailing services.

How this conversation actually went

Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →

The host as informed peer 3.7 Guest teaching 3.5 Guest disagreement 0.4 The host pushing back 0.8
05100:0010:0020:0030:001:15–4:52 · The host as informed peer 2/10 Revolution Series Title Graphic David George asks Dmitry to reflect on his early DARPA challenge days and career evolution. Dmitry warmly explains the 2007 Urban Challenge setup and how it sparked his work leading to Google Self-Driving Car Project and Waymo.4:52–9:06 · The host as informed peer 3/10 Integration of ConvNets, Transformers, and GenAI in Autonomous Driving David asks about layering generative AI onto traditional machine learning, prompting Dmitry to gently reframe the scope by detailing the chronological integration of ConvNets, Transformers, and VLMs.9:06–13:14 · The host as informed peer 4/10 Simulation, Synthetic Data, and Closed-Loop Testing David highlights the debate around synthetic data utility, and Dmitry details how closed-loop simulation generates long-tail edge cases and statistical realism for training.13:14–18:36 · The host as informed peer 5/10 Applying Scaling Laws to Autonomous Systems and Onboard Compute Constraints David presents the debate between rules-based edge-case engineering and pure end-to-end AI, prompting Dmitry to reject the dichotomy and explain why end-to-end models alone cannot solve full autonomy.18:36–21:52 · The host as informed peer 5/10 Real-World Deployment, Weather Conditions, and Pick-Up/Drop-Off Challenges Dmitry discusses pick-up and drop-off challenges, and David demonstrates solid domain understanding by listing specific urban friction points like garage doors and driveway blocking.21:52–25:28 · The host as informed peer 5/10 Safety Record, Swiss Re Study, and Safety Validation Framework David cites recent 15 million mile safety data, while Dmitry enriches the discussion with joint Swiss Re study findings and Waymo's attentive-human benchmark model.25:28–27:37 · The host as informed peer 3/10 Generalizable Driver Vision, Go-To-Market Strategies, and Partnerships David asks about future commercial dynamics, prompting Dmitry to detail Waymo's generalizable driver vision across ride-hailing, trucking, and Uber app integrations.27:37–30:00 · The host as informed peer 3/10 Hardware Modalities: LiDAR, Radar, and Cameras David asks whether LiDAR will remain necessary, leading Dmitry to explain the complementary physics of cameras, LiDAR, and imaging radar for full autonomy redundancy.30:00–32:26 · The host as informed peer 5/10 Why Driverless Autonomy is Uniquely Hard Compared to LLMs David contrasts the rapid commoditization of digital LLMs with the consolidation in AVs, which Dmitry attributes to physical world noise, safety stakes, and real-time millisecond latency constraints.32:26–36:24 · The host as informed peer 2/10 Humbling Early Memories and Career Advice for Builders David prompts Dmitry for early memories, leading to an amusing 2009 anecdote about an AV navigating around debris falling from a dump truck, followed by career advice for young builders.1:15–4:52 · Guest teaching 3/10 Revolution Series Title Graphic David George asks Dmitry to reflect on his early DARPA challenge days and career evolution. Dmitry warmly explains the 2007 Urban Challenge setup and how it sparked his work leading to Google Self-Driving Car Project and Waymo.4:52–9:06 · Guest teaching 4/10 Integration of ConvNets, Transformers, and GenAI in Autonomous Driving David asks about layering generative AI onto traditional machine learning, prompting Dmitry to gently reframe the scope by detailing the chronological integration of ConvNets, Transformers, and VLMs.9:06–13:14 · Guest teaching 4/10 Simulation, Synthetic Data, and Closed-Loop Testing David highlights the debate around synthetic data utility, and Dmitry details how closed-loop simulation generates long-tail edge cases and statistical realism for training.13:14–18:36 · Guest teaching 5/10 Applying Scaling Laws to Autonomous Systems and Onboard Compute Constraints David presents the debate between rules-based edge-case engineering and pure end-to-end AI, prompting Dmitry to reject the dichotomy and explain why end-to-end models alone cannot solve full autonomy.18:36–21:52 · Guest teaching 2/10 Real-World Deployment, Weather Conditions, and Pick-Up/Drop-Off Challenges Dmitry discusses pick-up and drop-off challenges, and David demonstrates solid domain understanding by listing specific urban friction points like garage doors and driveway blocking.21:52–25:28 · Guest teaching 4/10 Safety Record, Swiss Re Study, and Safety Validation Framework David cites recent 15 million mile safety data, while Dmitry enriches the discussion with joint Swiss Re study findings and Waymo's attentive-human benchmark model.25:28–27:37 · Guest teaching 3/10 Generalizable Driver Vision, Go-To-Market Strategies, and Partnerships David asks about future commercial dynamics, prompting Dmitry to detail Waymo's generalizable driver vision across ride-hailing, trucking, and Uber app integrations.27:37–30:00 · Guest teaching 5/10 Hardware Modalities: LiDAR, Radar, and Cameras David asks whether LiDAR will remain necessary, leading Dmitry to explain the complementary physics of cameras, LiDAR, and imaging radar for full autonomy redundancy.30:00–32:26 · Guest teaching 4/10 Why Driverless Autonomy is Uniquely Hard Compared to LLMs David contrasts the rapid commoditization of digital LLMs with the consolidation in AVs, which Dmitry attributes to physical world noise, safety stakes, and real-time millisecond latency constraints.32:26–36:24 · Guest teaching 1/10 Humbling Early Memories and Career Advice for Builders David prompts Dmitry for early memories, leading to an amusing 2009 anecdote about an AV navigating around debris falling from a dump truck, followed by career advice for young builders.1:15–4:52 · Guest disagreement 0/10 Revolution Series Title Graphic David George asks Dmitry to reflect on his early DARPA challenge days and career evolution. Dmitry warmly explains the 2007 Urban Challenge setup and how it sparked his work leading to Google Self-Driving Car Project and Waymo.4:52–9:06 · Guest disagreement 1/10 Integration of ConvNets, Transformers, and GenAI in Autonomous Driving David asks about layering generative AI onto traditional machine learning, prompting Dmitry to gently reframe the scope by detailing the chronological integration of ConvNets, Transformers, and VLMs.9:06–13:14 · Guest disagreement 0/10 Simulation, Synthetic Data, and Closed-Loop Testing David highlights the debate around synthetic data utility, and Dmitry details how closed-loop simulation generates long-tail edge cases and statistical realism for training.13:14–18:36 · Guest disagreement 3/10 Applying Scaling Laws to Autonomous Systems and Onboard Compute Constraints David presents the debate between rules-based edge-case engineering and pure end-to-end AI, prompting Dmitry to reject the dichotomy and explain why end-to-end models alone cannot solve full autonomy.18:36–21:52 · Guest disagreement 0/10 Real-World Deployment, Weather Conditions, and Pick-Up/Drop-Off Challenges Dmitry discusses pick-up and drop-off challenges, and David demonstrates solid domain understanding by listing specific urban friction points like garage doors and driveway blocking.21:52–25:28 · Guest disagreement 0/10 Safety Record, Swiss Re Study, and Safety Validation Framework David cites recent 15 million mile safety data, while Dmitry enriches the discussion with joint Swiss Re study findings and Waymo's attentive-human benchmark model.25:28–27:37 · Guest disagreement 0/10 Generalizable Driver Vision, Go-To-Market Strategies, and Partnerships David asks about future commercial dynamics, prompting Dmitry to detail Waymo's generalizable driver vision across ride-hailing, trucking, and Uber app integrations.27:37–30:00 · Guest disagreement 0/10 Hardware Modalities: LiDAR, Radar, and Cameras David asks whether LiDAR will remain necessary, leading Dmitry to explain the complementary physics of cameras, LiDAR, and imaging radar for full autonomy redundancy.30:00–32:26 · Guest disagreement 0/10 Why Driverless Autonomy is Uniquely Hard Compared to LLMs David contrasts the rapid commoditization of digital LLMs with the consolidation in AVs, which Dmitry attributes to physical world noise, safety stakes, and real-time millisecond latency constraints.32:26–36:24 · Guest disagreement 0/10 Humbling Early Memories and Career Advice for Builders David prompts Dmitry for early memories, leading to an amusing 2009 anecdote about an AV navigating around debris falling from a dump truck, followed by career advice for young builders.1:15–4:52 · The host pushing back 0/10 Revolution Series Title Graphic David George asks Dmitry to reflect on his early DARPA challenge days and career evolution. Dmitry warmly explains the 2007 Urban Challenge setup and how it sparked his work leading to Google Self-Driving Car Project and Waymo.4:52–9:06 · The host pushing back 1/10 Integration of ConvNets, Transformers, and GenAI in Autonomous Driving David asks about layering generative AI onto traditional machine learning, prompting Dmitry to gently reframe the scope by detailing the chronological integration of ConvNets, Transformers, and VLMs.9:06–13:14 · The host pushing back 1/10 Simulation, Synthetic Data, and Closed-Loop Testing David highlights the debate around synthetic data utility, and Dmitry details how closed-loop simulation generates long-tail edge cases and statistical realism for training.13:14–18:36 · The host pushing back 3/10 Applying Scaling Laws to Autonomous Systems and Onboard Compute Constraints David presents the debate between rules-based edge-case engineering and pure end-to-end AI, prompting Dmitry to reject the dichotomy and explain why end-to-end models alone cannot solve full autonomy.18:36–21:52 · The host pushing back 1/10 Real-World Deployment, Weather Conditions, and Pick-Up/Drop-Off Challenges Dmitry discusses pick-up and drop-off challenges, and David demonstrates solid domain understanding by listing specific urban friction points like garage doors and driveway blocking.21:52–25:28 · The host pushing back 1/10 Safety Record, Swiss Re Study, and Safety Validation Framework David cites recent 15 million mile safety data, while Dmitry enriches the discussion with joint Swiss Re study findings and Waymo's attentive-human benchmark model.25:28–27:37 · The host pushing back 0/10 Generalizable Driver Vision, Go-To-Market Strategies, and Partnerships David asks about future commercial dynamics, prompting Dmitry to detail Waymo's generalizable driver vision across ride-hailing, trucking, and Uber app integrations.27:37–30:00 · The host pushing back 0/10 Hardware Modalities: LiDAR, Radar, and Cameras David asks whether LiDAR will remain necessary, leading Dmitry to explain the complementary physics of cameras, LiDAR, and imaging radar for full autonomy redundancy.30:00–32:26 · The host pushing back 1/10 Why Driverless Autonomy is Uniquely Hard Compared to LLMs David contrasts the rapid commoditization of digital LLMs with the consolidation in AVs, which Dmitry attributes to physical world noise, safety stakes, and real-time millisecond latency constraints.32:26–36:24 · The host pushing back 0/10 Humbling Early Memories and Career Advice for Builders David prompts Dmitry for early memories, leading to an amusing 2009 anecdote about an AV navigating around debris falling from a dump truck, followed by career advice for young builders.

speaking balance: gold is the host, purple is the guest (3 minute bins)

0:00 · the host 0% · guest 100%0:00 · the host 0% · guest 100%3:00 · the host 0% · guest 100%3:00 · the host 0% · guest 100%6:00 · the host 0% · guest 100%6:00 · the host 0% · guest 100%9:00 · the host 0% · guest 100%9:00 · the host 0% · guest 100%12:00 · the host 0% · guest 100%12:00 · the host 0% · guest 100%15:00 · the host 0% · guest 100%15:00 · the host 0% · guest 100%18:00 · the host 0% · guest 100%18:00 · the host 0% · guest 100%21:00 · the host 0% · guest 100%21:00 · the host 0% · guest 100%24:00 · the host 0% · guest 100%24:00 · the host 0% · guest 100%27:00 · the host 0% · guest 100%27:00 · the host 0% · guest 100%30:00 · the host 0% · guest 100%30:00 · the host 0% · guest 100%33:00 · the host 0% · guest 100%33:00 · the host 0% · guest 100%36:00 · the host 0% · guest 100%36:00 · the host 0% · guest 100%
Sharpest disagreement ▶ 15:36 Rejecting end-to-end dichotomy

Dmitry directly pushes back on David's premise of end-to-end AI replacing legacy approaches, labeling it a weird dichotomy and insisting that full autonomy requires deep learning plus extensive safety engineering.

Hardest push from the host ▶ 14:42 Challenging end-to-end taglines

David explicitly pushes Dmitry with an intentionally simplifying premise about end-to-end driving taglines that he acknowledges might drive Dmitry crazy, challenging him to defend Waymo's hybrid architecture.

Biggest teaching moment ▶ 28:06 Sensor physics breakdown

Dmitry educates David on why LiDAR cannot be omitted for full autonomy, breaking down the distinct physical properties and redundancy advantages of active LiDAR, imaging radar Doppler velocity, and high-res cameras.

The host holds their own ▶ 20:26 Urban pick-up edge cases

David demonstrates strong domain knowledge by offering specific, accurate edge cases about dense urban pick-ups—such as opening garage doors and driveway blocking—prompting Dmitry to confirm he hit the exact subtleties of the problem.

the scores for every segment, with the reasoning behind each
ChapterTopicThe host as informed peerGuest teachingGuest disagreementThe host pushing backWhy
Revolution Series Title Graphic 2300 David George asks Dmitry to reflect on his early DARPA challenge days and career evolution. Dmitry warmly explains the 2007 Urban Challenge setup and how it sparked his work leading to Google Self-Driving Car Project and Waymo.
Integration of ConvNets, Transformers, and GenAI in Autonomous Driving 3411 David asks about layering generative AI onto traditional machine learning, prompting Dmitry to gently reframe the scope by detailing the chronological integration of ConvNets, Transformers, and VLMs.
Simulation, Synthetic Data, and Closed-Loop Testing 4401 David highlights the debate around synthetic data utility, and Dmitry details how closed-loop simulation generates long-tail edge cases and statistical realism for training.
Applying Scaling Laws to Autonomous Systems and Onboard Compute Constraints 5533 David presents the debate between rules-based edge-case engineering and pure end-to-end AI, prompting Dmitry to reject the dichotomy and explain why end-to-end models alone cannot solve full autonomy.
Real-World Deployment, Weather Conditions, and Pick-Up/Drop-Off Challenges 5201 Dmitry discusses pick-up and drop-off challenges, and David demonstrates solid domain understanding by listing specific urban friction points like garage doors and driveway blocking.
Safety Record, Swiss Re Study, and Safety Validation Framework 5401 David cites recent 15 million mile safety data, while Dmitry enriches the discussion with joint Swiss Re study findings and Waymo's attentive-human benchmark model.
Generalizable Driver Vision, Go-To-Market Strategies, and Partnerships 3300 David asks about future commercial dynamics, prompting Dmitry to detail Waymo's generalizable driver vision across ride-hailing, trucking, and Uber app integrations.
Hardware Modalities: LiDAR, Radar, and Cameras 3500 David asks whether LiDAR will remain necessary, leading Dmitry to explain the complementary physics of cameras, LiDAR, and imaging radar for full autonomy redundancy.
Why Driverless Autonomy is Uniquely Hard Compared to LLMs 5401 David contrasts the rapid commoditization of digital LLMs with the consolidation in AVs, which Dmitry attributes to physical world noise, safety stakes, and real-time millisecond latency constraints.
Humbling Early Memories and Career Advice for Builders 2100 David prompts Dmitry for early memories, leading to an amusing 2009 anecdote about an AV navigating around debris falling from a dump truck, followed by career advice for young builders.

Statements from this episode (12)

Disclosure
Dolgov: Waymo is combining autonomous driving AI with vision-language models
“So that, that's what we've been very focused on that Waymo most recently is taking kind of the AI backbone and all of the AI, the Waymo AI that is over the years we've built up that is really proficient at this task of autonomous driving and combining it with …”
Dmitry Dolgov Aug 5, 2024 ▶ 8:44
Insight
Dolgov: Building AV drivers requires iteratively modeling simulation environments and actors
“To build a good driver, you need to have a very good simulator, but to have a good simulator, you actually have to build models of, like, realistic pedestrians and cyclists and drivers, right? So it's, you know, you kind of do that iteratively.”
Dmitry Dolgov Aug 5, 2024 ▶ 12:18
Disclosure
Waymo has completed over 15 million fully autonomous rider-only miles
“I mean, we've driven, you know, tens of millions of miles in the physical world, and at this point we've driven more than fifteen million miles in full autonomy, what we call, you know, writer-only mode but we've driven, you know, tens of billions of miles of …”
Dmitry Dolgov Aug 5, 2024 ▶ 12:59
Insight
Dolgov: Autonomous vehicle AI scaling requires rare-case data over raw mileage
“It has to be, you know, the right kind of data that, you know, teaches the models or trains the models to be, you know, good at the rare cases that, that you care about.”
Dmitry Dolgov Aug 5, 2024 ▶ 14:09
Insight
Dolgov: Distilling huge AI models beats training small models directly
“Where you're much better off training a huge model and then distilling it into a smaller model than just training small models.”
Dmitry Dolgov Aug 5, 2024 ▶ 14:37
Insight
Dolgov: End-to-end AI models alone cannot achieve full autonomous driving
“You can take, you know, an end-to-end model, go from sensor to, you know, trajectories or actuation. You know, typically you don't build them in one stage, you build them in stages, but, you know, you can do, like, backprop through the whole thing, so, you kno…”
Dmitry Dolgov Aug 5, 2024 ▶ 16:58
Assertion Partly supported
Waymo operates 24/7 across San Francisco, Phoenix, LA, and Austin
“In terms of where we are today, you know, we are up, you know, driving in all kinds of conditions. We're driving, you know 24 seven in San Francisco, in Phoenix, you know, a little bit those are the most mature markets, but also in L.A. And in Austin, and, you…”
Dmitry Dolgov Aug 5, 2024 ▶ 19:19
Assertion Supported
Dolgov: Waymo driverless cars cut injury crashes by 3.5x
“And I think 3.5 acts as the reduction in injury. And that's about two X reduction in the police reportable kind of lower severity incidents.”
Dmitry Dolgov Aug 5, 2024 ▶ 22:18
Assertion Partly supported
Dolgov: Swiss Re study showed zero bodily injury claims for Waymo
“We published a joint study with a Swiss RE. Which is, I think, the largest global reinsurer in the world, and the way they look at it is, you know, who contributed to an event, and there we saw, ah, like the same theme, but the numbers were, ah, very strong, t…”
Dmitry Dolgov Aug 5, 2024 ▶ 23:37
Prediction Open · timeframe Aug 2027
Dolgov: Waymo Driver will expand into deliveries, trucking, and personal vehicles
“We envision a future where our, the Waymo driver will deploy, be deployed in other commercial applications, right? There's deliveries, there's trucking, there's personally owned vehicles.”
Dmitry Dolgov Aug 5, 2024 ▶ 26:01
Disclosure
Dolgov: Waymo builds its own custom imaging radar in-house
“So if you build an imaging radar, which we do ourselves you know, it allows us to, you know, give you an additional redundancy layer, and it has benefits also in active sensor.”
Dmitry Dolgov Aug 5, 2024 ▶ 28:44
Prediction Held up
Dolgov: Waymo will continue using LiDAR, radar, and cameras together
“So yeah, I think the trend, you know, for us that we'll, you know, using all three modalities just, you know, makes a lot of sense.”
Dmitry Dolgov Aug 5, 2024 ▶ 29:46
Made with StarZero

Turn any episode into a week of clips.

This entire site, over 1,000 episodes transcribed, diarized, checked and made playable, runs on the StarZero media pipeline. Drop in your own episode and the podcast clipper finds the moments worth sharing, cuts them, captions them, and reframes them for every feed.