The Demo Is Only 1% Of The Work

Dolgov says physical AI cannot follow “move fast and break things” because mistakes can cost lives, and cars must make decisions in milliseconds on board. Unlike digital AI, it lacks a ready-made internet-scale source of labeled data and needs strong safety evidence before deployment. He argues teams should define operating conditions and scale through a rigorous deployment framework.

Waymo’s early autonomous driving demo took about 18 months, but turning it into a service and scaling it took roughly 15 years. Dolgov says each additional “nine” of reliability or performance takes about ten times more effort, and rare failures become routine at millions of miles per week. Founders should decide how many nines their product needs before choosing an approach or celebrating demo traction.

Dolgov warns that a technology with a fast early improvement curve may plateau before a product reaches its safety target. Waymo combines cameras, lidar, and radar, whose different strengths help detect hazards in darkness, dust, glare, and other difficult conditions. He also advises designing for future hardware improvements rather than locking a company’s plans to today’s component prices.

Dolgov describes Waymo’s approach as combining scalable learned models with useful structure, such as physics and road rules, to support performance and validation. He says closed-loop simulation is essential because an agent must be tested on how its actions change the world; Waymo’s simulator models both behavior and sensor data, including synthetic rare scenarios. The agent, simulator, and critic form a flywheel, while evaluation and metrics guide what the system learns and when it is ready to deploy.

Dolgov calls evaluation and metrics strategic assets and says teams should define “good enough” before building a product. Waymo’s Safety and Readiness Framework assesses the full system, including onboard behavior, physical components, off-board systems, and operations. He cites Waymo’s published data covering over 220 million autonomous miles and reports 17 times fewer crashes causing serious injury than human drivers in its operating areas.