Waymo Co-CEO Dmitri Dolgov: "Move Fast And Ship Safely"

Why does it matter? Because the gap between a working demo and a real product just swallowed 15 years.

Waymo declared autonomous driving solved in 2010 after 18 months of work. The actual product — half a million trips per week, 15 cities, no safety driver — arrived 15 years later. Dmitri Dolgov, Waymo's co-CEO, uses that arc to trace exactly why every AI hype cycle keeps producing spectacular prototypes and almost no real products — and why the pattern is structural, not accidental.

• Every additional nine of reliability costs 10x more effort than the last — most founders don't count their nines before counting demo views • Physical AI faces four gaps digital AI never had: lives-not-tokens error cost, millisecond latency, no physical internet, and a validation wall you can't iterate past when errors are fatal • Eval and metrics are your strategic moat — not your model architecture, which competitors can replicate in months • Building physical AI means building three AIs: the agent, a high-fidelity simulator, and a critic — and the flywheel they create is what enables safe scaling

Waymo declared autonomous driving 'solved' in 2010. The product arrived 15 years later.

In 2010, a team of roughly a dozen engineers drove 100,000 autonomous miles and completed 10 pre-specified 100-mile routes — each end-to-end, without human intervention — in about 18 months. Day, night, traffic, pedestrians, cyclists, construction zones. Capability complete. They felt like they were on top of the world.

Then the engineering reality arrived.

Every additional nine of performance costs roughly 10 times more effort than the one before it. A demo might need one nine. A copilot needs a few. A fully autonomous vehicle operating around children at freeway speeds with no safety driver needs a whole stack — and reaching each new nine requires a fundamentally different approach, not more of the same.

"The demo took 18 months. The product took about 15 years."

At scale, the math compounds into something stranger: driving 4 million autonomous miles per week means an event that happens once in a million miles is now your daily reality. The long tail isn't an edge case. It's the entire product surface.

What makes this cycle so persistent is that AI breakthroughs accelerate the wrong part. Every wave — ConvNets, Transformers, LLMs — makes it roughly 100x easier to build a working demo. The tail, where the actual nines live, moves too, but the effect is muted by comparison. So each breakthrough produces a fresh cohort of founders who mistake a spectacular prototype for a product.

"The recurring mistake of every cycle is spending on the demo what you should be saving for the nines."

Count your nines before you count your demo views.

In the physical world, a mistake costs lives, not tokens — and there is no retry button

Four gaps. That is how Dolgov frames the chasm between digital and physical AI, and each one demands rebuilding your architectural intuitions from scratch.

First: cost of error. A language model that hallucinates costs you a retry. A robot that fails at the wrong moment costs someone their life. "There's simply not an undo and a retry button." That asymmetry shapes everything downstream — you cannot ship at 80% confidence and iterate your way to safety.

Second: latency. A digital assistant can take seconds to respond. A car at freeway speed covers 100 feet every second. All inference, every decision, must run on a computer that fits in a trunk, in milliseconds.

Third: data. Digital AI got the internet — a vast, pre-labeled cache of human knowledge assembled over decades. Physical AI got nothing equivalent. The physical world was never digitized at scale.

Fourth, and most corrosive for anyone porting a SaaS playbook: the validation gap. In digital AI you ship something good enough, let users surface the edge cases, and iterate from there. That feedback loop — cheap, fast, public — is simply unavailable when your edge cases can be fatal. You need a very high level of safety and confidence before your first deployment.

But the opposite is equally true: there is no substitute for real-world field data. You cannot perfect a physical AI agent in the lab and release it overnight. The escape from this bind is a rigorous, formal deployment framework — not a shipping philosophy borrowed from consumer software.

Your model can be leaked. Two hundred million miles of validated autonomous operation cannot.

Model architectures proliferate fast. The best ones are well known, and new ideas spread across the industry in months. Data matters enormously, but without strong metrics you are flying blind — you cannot prioritize which data to collect or evaluate the ROI of any change.

What competitors cannot replicate is proof.

"Build your eval before you build your technology. Build your eval and your metrics before you build your product. If you can't quantitatively define what good enough means, you're not really building a product. You're just iterating on your demo."

For physical AI, model-level evaluation is not enough. Waymo's safety and readiness framework — built and refined over years — evaluates every layer: physical hardware, on-board behavioral models, off-board systems, and all operational processes wrapping them. Dolgov considers it one of the company's most important assets.

That is also why trust compounds into a durable business advantage. Based on over 220 million fully autonomous miles, Waymo's latest safety data shows the system drives about 17 times better than human drivers in the areas where it operates, measured by crashes causing serious injury. Every 26 seconds, somewhere in the world, someone dies in a road crash. At Waymo's current scale, that data translates to a serious injury prevented every eight days.

Those numbers cannot be claimed. They have to be earned, mile by mile, in public. "Your models can be leaked, algorithms can be replicated, but hundreds of millions of miles of fully autonomous operations in the real world, backed by evidence-grade evaluation and publicly audited proof — that is much, much more difficult to replicate."

Eval is infrastructure. Treat it that way from day one.

You are not building one AI — you are building three, and their flywheel is the only safe path to scale

The agent is not the product. The agent is one-third of the system.

Dolgov's framing is direct: you need the agent, a high-fidelity simulator, and a critic — and all three are roughly as difficult to build as each other. "The problem of building a good realistic simulator is just as hard as building the agent itself."

The simulator is not lightweight tooling bolted onto your stack. It is a big AI in its own right — one that must understand physics, semantics, traffic dynamics, and weather, and produce synthetic scenarios realistic enough that agents trained inside it transfer safely to the real world. Waymo's behavioral world model, built on Google DeepMind's Genie 3, generates controllable high-fidelity scenarios: cars stopped on freeways, dust storms in Phoenix, complete darkness with no headlights, even a plane landing on a highway or an elephant walking through an intersection.

The flywheel these three create is what actually enables safe scaling at speed. Real-world deployment generates data. That data grounds the simulator, making it more realistic. The simulator generates harder edge cases for the critic to score. The critic tells the agent how to improve. The agent gets smarter, gets deployed, generates more data. Each turn of the cycle compounds.

Closed-loop simulation — where the agent takes actions, observes consequences, and updates — is absolutely required for evaluation and extremely valuable for training. Open-loop scoring is necessary but insufficient for a safety-critical system; you cannot evaluate counterfactuals without letting the agent act.

Build the simulator and critic as co-equals to the agent from day one. They are not supporting tools added later. They are the mechanism by which you earn the right to scale.

Structure that fights scale always loses — the key is injecting it as an amplifier, not a constraint

Richard Sutton's bitter lesson — that methods which scale with data and compute always beat handcrafted engineering — has held through every major AI wave. Dolgov does not dispute it. He extends it.

The nuance is that structure is not monolithic. Structure that constrains your solution space will lose to scale every time. Structure that channels scale — injected intentionally, as an amplifier — consistently wins.

Waymo's approach: structure-augmented end-to-end. Pure vanilla end-to-end models have a genuine advantage — the gradient back-propagates from the actual task all the way through the model, letting the encoder learn exactly the representations the decoder needs. That produces rapid early progress. But for superhuman performance in a safety-critical autonomous vehicle, vanilla end-to-end hits a ceiling.

"If you need to reach superhuman levels of performance in a fully autonomous agent in a safety-critical environment, just doing kind of that basic vanilla end-to-end is not enough."

The fix: augment learned embeddings with materialized structure representations — physics, rules of the road, object behaviors — that do not constrain what the model can learn but give it better scaffolding for training and evaluation. This unlocks real-time validation at inference time, lets you mix and match training and evaluation at different levels of abstraction, and delivers stronger feedback signals for reinforcement learning.

The Go analogy is clean: a camera-to-actuation end-to-end system could theoretically learn Go, but a 19x19 board as an intermediate representation does not limit the model — it just makes scaling far more efficient. Physical AI has analogous structure waiting to be used.

"Bet on a system that's maximally learned and minimally constrained."

Camera-only perception hits its safety ceiling far below what full autonomy actually requires

Humans drive with just eyes. That is a proof of existence — but only for a human-level reliability target.

If the goal is full autonomy at superhuman safety, weak sensing leads to a safety curve that flattens out way too early. Waymo uses cameras, lidar, and radar — not as backups to each other but as complementary physical sensing modalities fused into a single, unified view of the world.

Cameras deliver high resolution and color but degrade in darkness and glare. Lidar provides direct 3D structure measurement and performs identically in pitch darkness or a blinding sunset. Radar punches through fog, rain, and snow, and directly measures velocity using Doppler. Each modality has an encoder; all three are fused into a richer representation than any single sensor could produce.

The difference in practice is concrete. In a Phoenix dust storm, the camera shows approximately what a human would see — nearly nothing. The lidar, in the same frame, clearly identifies a pedestrian at the roadside. In complete darkness with two people about to jump over a construction barrier onto the road, the camera captures very little; the lidar catches them with time to react.

Hardware economics follow a different curve than most founders assume. Waymo is on its sixth-generation sensor suite. Each generation delivered meaningfully better capability and meaningfully lower cost simultaneously. "Do not anchor to today's component prices. Betting your company, betting your approach on today's hardware prices is just betting your company on a number that has a fairly short shelf life."

Design the sensor stack for the reliability ceiling your product actually needs — not for today's cost baseline.

Every AI breakthrough makes the demo 100x easier — and barely moves the reliability tail that actually matters

The macro pattern becomes visible once you have watched it repeat. Each wave of AI progress — ConvNets, Transformers, LLMs — compresses the time to a working demo and extends the life of the hype cycle that follows. The tail, where the nines live, moves too, but asymmetrically.

"Every AI breakthrough makes it that much easier to get started. Your demos, your prototypes, they get a hundred times easier. But the tail — that's where the hard problems are — that moves much less. And that's why every hype cycle produces a wave of absolutely spectacular demos and very few real products."

There is a second trap embedded in how companies adopt breakthroughs. When a new architecture or training paradigm arrives, the instinct is to spin up a tiger team. That is fine — you should. But you need to know in advance what success looks like for your whole product stack, not just the prototype. A failure mode Dolgov has watched repeat: a technically impressive project succeeds and then hits a dead end because no one designed the integration path. The success becomes deflating rather than accelerating.

The bar Waymo applies when adopting new technology: it must deliver breakthrough performance and radical simplification of the stack simultaneously. Capability gains alone do not clear the bar. Fragmentation is a regression, even when disguised as innovation.

"Set your launch bar to demand both breakthrough performance and, at the same time, radical simplification and unification."

If the new breakthrough only helps the demo and leaves the tail untouched, you have a better science project, not a better product.

Physical AI is where digital AI was a few years ago — and the same traps are already being set

Every ingredient is now in place: generative world models, proven architectures, affordable compute and sensing, demonstrated scaling laws, and a real product operating at scale in public. The infrastructure cost that made autonomous vehicles exclusive to well-capitalized incumbents will not hold that shape for long.

What that means is the reliability gap — the 15-year chasm between demo and product — will not disappear. It will replicate across every new physical AI domain. The founders who navigate it are the ones who build their eval framework before their model, resource the simulator and critic as co-equals to the agent, and resist spending on the demo what they should be saving for the nines.

The next decade of AI will happen in the physical world. The companies that build it right will not just have better technology. They will have proof.

---

Topics: autonomous vehicles, physical AI, robotics, AI safety, startup building, product development, sensor fusion, foundation models, simulation, reliability engineering, Waymo, AI hype cycles