Dmitri Dolgov on why a working demo is 1% of the product
Waymo's co-CEO turns two decades of self-driving into a seven-lesson playbook for physical AI, where a working demo is 1% of the work, every added nine of reliability costs roughly 10x, and the real moat is the evals, simulation, and hundreds of millions of autonomous miles backed by publicly audited safety proof.
Human Lives, Not Tokens
Building AI for the physical world means clearing four gaps that set it apart from digital AI — the cost of an error, the latency budget, the missing data, and the day-one validation bar — because you are moving atoms, not bits.
the cost of a mistake can be measured in human lives, not tokens. There's simply not an undo and a retry button.
A Working Demo Is 1% of the Job
Waymo's demo drove itself in 18 months, but the shippable product took 15 more years, because reliability climbs an exponential ladder where every added nine of performance costs roughly ten times the last.
at scale, the long tail is the problem space. It's your entire problem statement. When you drive millions of miles per week, a rare event that might happen once in a million miles, that just becomes your daily reality.
Your Nines Pick Your Architecture
Most technologies ride a steep early curve that tends to flatten, so the recurring trap is picking the tech with the fastest ramp only to hit a plateau below the reliability your product actually needs.
common failure mode is picking the tech that gives you the fastest early ramp, riding that steep curve, feeling like you're winning
Add Capability, Subtract Complexity
Waymo rebuilt its driver around every AI wave — CNNs, then transformers, now VLMs and world models — but the rare muscle is not adopting new tech, it is forcing each wave to also simplify and unify the stack instead of fragmenting it.
set your launch bar to demand both breakthrough performance and at the same time radical simplification and unification.
Structure That Channels Scale, Not Fights It
Sutton's bitter lesson says methods that scale with compute and data beat handcrafted knowledge, and Dolgov's refinement is that structure which fights scale loses while structure that channels it wins.
bet on a system that's maximally learned and minimally constrained and leverage structure intentionally to boost performance and scaling laws both in training and in evaluation.
Your Simulator Is as Hard as Your Agent
Trying a risky action and watching its consequence is rarely something you can do safely on real roads, so closed-loop simulation is essential for physical AI — and a good simulator is itself a giant world model, as hard to build as the driver it trains.
the problem of building a good realistic simulator is just as hard as building the agent itself.
You Don't Build One AI, You Build Three
Shipping a physical agent takes three AIs sharing one foundation — the agent that acts, the simulator it practices in, and the critic that judges it — wired into a flywheel where real-world deployment keeps making all three smarter.
once you have these three pillars you can create an incredibly powerful flywheel to accelerate your progress.
Eval Is the Moat, Not the Model
Model architectures are well known and spread fast, so the durable advantage is not the model but the evals, metrics, and hundreds of millions of autonomous miles of publicly audited proof that are far harder to replicate — turning trust into the real moat.
If you can't quantitatively define what good enough means, you're not really building a product. You're just iterating on your demo.