Descriptions:
Waymo Co-CEO Dmitri Dolgov delivers a structured seven-lesson talk at Y Combinator’s Startup School, drawing on nearly two decades building autonomous driving technology. He opens with current operational scale: over 4 million fully autonomous miles driven every week across 15 U.S. cities — equivalent to more than 300 years of average American driving per week — accomplished with a superhuman safety record.
Dolgov frames the core challenge of physical AI around four gaps absent in digital AI: the cost-of-error gap (mistakes cost lives, not retries), the latency gap (a car at freeway speed moves 100 feet per second, demanding millisecond onboard inference), the data gap (no internet-scale labeled dataset exists for the physical world), and the validation gap (unlike a chatbot, you cannot ship a minimum viable autonomous vehicle and iterate from user-reported edge cases). He advises startups against anchoring business models to current hardware prices, noting Waymo is on its sixth-generation hardware suite with dramatically reduced costs at each generation.
A substantial portion covers how Waymo has rebuilt its stack through successive AI waves — CNNs around 2013 for perception, transformers from 2017 for behavior prediction and planning, and today’s VLMs and frontier world models — each time carrying bleeding-edge research into a safety-critical production system without regressions. Dolgov argues the hardest organizational muscle to build is not doing applied research, but shipping that research to production repeatedly across multiple technology generations while reducing system complexity rather than adding to it.
📺 Source: Y Combinator · Published August 03, 2026
🏷️ Format: Keynote Launch
