No PriorsNo Priors Ep. 87 | With Co-CEO of Waymo Dmitri Dolgov
CHAPTERS
- 0:00 – 0:46
Waymo today: scale, cities, and why robo-taxis matter
Sarah and Elad introduce Dmitri Dolgov and frame Waymo’s current footprint: 100,000+ paid rides weekly across multiple cities. They set the agenda around what it takes to deploy autonomy at real scale and what comes next.
- •Waymo’s origin as Google’s Chauffeur Project (2009) and later spin-out
- •Current operational scale: paid rides and weekly miles
- •Topics preview: deployment, scaling, safety, and future roadmap
- 0:46 – 2:09
From DARPA Urban Challenge to Google’s self-driving bet
Dmitri recounts entering the field during the DARPA challenges and how the Urban Challenge made autonomy feel inevitable. He describes the early Google effort starting with a small team and eventually becoming Waymo in 2017.
- •DARPA Grand Challenge vs. Urban Challenge and what they proved
- •Stanford team involvement and the “clicked for me” moment
- •Google project launch in 2009 and company formation in 2017
- 2:09 – 3:11
The early self-driving “lineage” and skepticism before deep learning
The discussion highlights how much of the autonomous driving ecosystem traces back to a few university labs and DARPA teams. Dmitri reflects on how “crazy” the goal looked at the time—before AlexNet and modern deep learning momentum.
- •Field lineage from Stanford/CMU cohorts into industry
- •Pre-AlexNet era constraints: limited tools, models, and compute
- •Widespread skepticism and the psychological commitment required
- 3:11 – 4:49
Milestones before products: exploring the space, then rejecting ADAS
Waymo’s early years focused on prototyping and learning rather than shipping a defined product by a fixed date. Dmitri explains they initially considered an ADAS-like product, but pivoted to full autonomy around 2013 after progress clarified what path mattered.
- •No fixed deadline at the start; emphasis on exploration milestones
- •Early product idea resembled advanced driver assistance
- •2013 pivot decision: commit to full autonomy instead
- 4:49 – 6:52
Generational discontinuities: Firefly to public service to Gen 5 scaling
Dmitri breaks down key ‘step changes’ that drove recent growth rather than smooth progress. He recounts the 2015 ‘zero-to-one’ Firefly milestone, the Gen 4 Pacifica deployment in Arizona (public in 2020), and the Gen 5 platform that enabled expansion to four major cities.
- •2015: first truly driverless rides (Firefly) as the ‘zero-to-one’ moment
- •Gen 4 Pacifica + Arizona deployment: focus on repeatability and maturity
- •Gen 5 platform (JLR-based fleet) as the scaling foundation
- 6:52 – 9:05
Where to deploy vs. where to learn: choosing cities and operating domains
Sarah probes why Arizona came first and how Waymo selects deployment environments. Dmitri explains the split between deployment (medium complexity to go end-to-end) and development (seek hardest conditions), and how they think in terms of operating domains (ODD) rather than ZIP codes.
- •Deployment site criteria: de-risking with “medium complexity” environments
- •Development strategy: target dense, high-speed, harsh-weather scenarios
- •ODD-first framing: map cities to operating-domain coverage
- •Forward-looking lenses: market, technical complexity, regulation
- 9:05 – 10:32
AI breakthroughs behind the newest driver: transformers plus the full machine
Dmitri attributes the biggest leaps primarily to AI rather than hardware alone. He points to ConvNets as an earlier boost and transformers/bigger models as crucial for the latest jump, but emphasizes the surrounding system—data engine, simulation, and evaluation—matters as much as architecture.
- •AI as the main driver of capability; hardware improvements help but aren’t decisive
- •ConvNets (post-2013) vs. transformers and larger-scale training
- •‘Architecture is an enabler’; the data/eval/sim stack is the differentiator
- •Need to build and evaluate in tandem to reach autonomy-grade reliability
- 10:32 – 11:42
How Waymo evaluates readiness: metrics, simulation, and the Safety Framework
Sarah asks how evaluation works internally and how it intersects with regulatory safety cases. Dmitri describes a metrics-driven approach supported by extensive data, open-loop and closed-loop simulation, and an integrated readiness-and-safety framework that gates releases.
- •Two coupled problems: building the driver and evaluating it
- •Hundreds of metrics and the data required to measure them
- •Open-loop evaluation vs. closed-loop simulation needs
- •Readiness and Safety Framework as the aggregate validation methodology
- 11:42 – 13:34
Safety results vs. human baselines: what the data shows
Elad presses for comparative safety levels. Dmitri cites Waymo’s published Safety Hub results based on 22M rider-only miles, claiming better-than-human performance that improves with severity, and references third-party insurance analysis with Swiss Re.
- •Empirical comparison to human benchmarks across collision severities
- •Lower-severity outcomes ~2x better; severe/airbag-type outcomes ~6x better
- •Safety Hub publication and transparency through reported data
- •Swiss Re study: reduced damage claims; strong bodily-injury results (smaller dataset)
- 13:34 – 15:32
Regulators, public trust, and why scaling must be gradual
The conversation shifts to what regulation should look like and why Waymo doesn’t ‘flip a switch’ to scale instantly. Dmitri emphasizes transparency, iterative rollout, and the need to earn trust with communities and regulators as the core pacing factor.
- •Regulatory posture: enable safety mission while requiring responsibility
- •Permits and long-running engagement with regulators and communities
- •Scaling is exponential but bounded by trust-building, not just tech
- •Trust is hard to earn and easy to lose; transparency is central
- 15:32 – 18:39
AI vs. hardware framing, and the push to simplify sensors and reduce cost
Elad asks about the perceived Waymo-vs-Tesla dichotomy (hardware-centric vs software-centric). Dmitri argues it’s fundamentally an AI problem, while hardware still matters because perception quality is foundational; he then explains Gen 6 priorities: simplification, cost down, and better rider UX.
- •Waymo view: ‘it’s all about AI,’ with hardware as an advantage in the physical world
- •Once the core works and evaluation is strong, focus shifts to optimization and speed
- •Gen-to-gen hardware: more capable, simpler, cheaper; economies of scale
- •Gen 6 focus: cost reduction + user experience improvements
- 18:39 – 21:30
Can you remove sensors? Answering with empirical ablations, not ideology
Elad asks whether Waymo can predict which sensing modalities are truly necessary. Dmitri says they can now run data-driven ablations (remove LiDAR/radar, add noise) and measure the safety/performance impact, concluding camera-only can ‘drive’ but doesn’t meet their autonomy safety bar.
- •New capability: empirical testing of sensor removal and degradation scenarios
- •Camera-only may function, but not at required performance for full autonomy at scale
- •Acceptable sensor stack depends on product definition (ADAS vs full autonomy)
- •Operating point choices (resolution, cleaning, modality mix) tie to safety bar
- 21:30 – 23:30
Business focus: ride-hailing first, partnerships always, broader modalities later
Elad asks whether robo-taxis are the near-term plan and how Waymo commercializes beyond CapEx constraints. Dmitri says ride-hailing is the primary focus, but the long-term mission is deploying a generalizable driver across deliveries, trucking, and personally owned vehicles—primarily via partnerships rather than building cars.
- •Ride-hailing as the main near-term product and learning loop
- •Waymo positions itself as building a generalizable ‘trusted driver’
- •Future applications: deliveries, long-haul, personally owned vehicles
- •Go-to-market preference: partner across vehicle and operations ecosystems
- 23:30 – 26:51
How autonomy reshapes car ownership and cities—and why timelines are hard
The hosts explore how autonomy could reduce private car ownership and alter land use (parking, urban design). Dmitri avoids firm timelines but reinforces that many benefits (safety, accessibility, land efficiency) are unlocked only with scale, and argues Waymo is beginning to earn the right to claim real-world safety impact.
- •Expected trend: greater shift to on-demand mobility, especially in dense cores
- •Auxiliary benefits: accessibility and better land use as parking needs shrink
- •Refusal to speculate on exact timelines; emphasis on scale prerequisites
- •Today’s scale enables more ‘empirically unambiguous’ safety benefit claims
- 26:51 – 34:43
OEMs, infrastructure, and designing vehicles around riders (the Gen 6 car)
Sarah asks what happens to traditional automakers when AI becomes a primary value driver. Dmitri argues OEMs remain essential because vehicles and form factors still matter, infrastructure changes can help, and future autonomous ride-hailing vehicles should be designed around passenger experience—spaciousness, easy entry, and privacy for work/calls.
- •Waymo builds the ‘driver’; OEMs build diverse vehicles and form factors
- •Complementarity with traditional OEMs rather than replacement
- •Public-sector/infrastructure role and multimodal integration (transit hubs)
- •Rider-centric vehicle design: flat floors, lower entry, sliding doors; open questions like seat orientation
- •Privacy and new in-car behaviors (meetings, calls) as part of the value proposition
- 34:43 – 40:16
The real difficulty: full autonomy vs. driver assist, and the long tail of ‘nines’
Sarah probes why trucking/delivery aren’t necessarily easier and what lesson the industry should take away. Dmitri argues the core chasm is ADAS vs full autonomy at scale, where rare, high-severity edge cases dominate; modern models can get impressive demos quickly, but not the reliability needed to remove the driver.
- •Key distinction: full autonomy at scale vs. driver assist systems
- •Different domains (freeways, delivery) change contours but not core complexity
- •Long-tail events at low frequency still must be handled safely at high speeds
- •Modern VLM/transformer approaches can produce quick prototypes, but not autonomy-grade safety
- 40:16 – 42:40
Iteration today: data flywheels plus architectural and simulation sophistication
Sarah asks whether Waymo’s improvement loop now resembles other AI companies: find failures, collect data, retrain, deploy. Dmitri says it’s both continuous data/eval flywheels and deeper methodology choices—moving beyond pure imitation learning into synthetic data and closed-loop simulation, potentially at intermediate representations for efficiency.
- •Core loop remains: identify weaknesses, collect/curate data, retrain, validate, deploy
- •Pure end-to-end imitation can plateau (especially for full autonomy)
- •Need synthetic data and closed-loop simulation to ‘go the distance’
- •Simulation at intermediate representations can be more scalable than sensor-level replay
- 42:40 – 44:30
Looking back and forward: existential questions answered, now optimize and scale
Dmitri closes by reflecting on two decades of progress and the shift from uncertainty to momentum. He claims Waymo has demonstrated buildability, evaluability, user demand, and a path to commercial viability—turning the focus to optimization and expanding access responsibly.
- •Past existential questions: can we build it, evaluate it, and make it viable?
- •Current confidence: safety record, evaluation maturity, strong user feedback
- •Next phase: optimization, scaling, and broader geographic reach
- •Mission focus: measurable reductions in road harm through trusted autonomy
