The Valley's favorite excuse for shipping half-baked software doesn't work when your AI controls two tons of steel at 45 mph.
The Summary
- Waymo co-CEO Dmitri Dolgov rejected "move fast and break things" for physical AI, saying "when you're dealing with atoms instead of bits, breaking things is not really okay"
- His replacement: "move fast and ship safely," treating safety as foundation, not afterthought
- Physical AI has no undo button — a chatbot error gets a retry, a robotaxi error gets a lawsuit or worse
The Signal
Dmitri Dolgov has been building self-driving cars since 2009, back when Google's autonomous vehicle project was still an X moonshot. Now as Waymo's co-CEO, he's drawing a hard line between the software-first mentality that built Web2 and the requirements for physical AI systems that navigate real streets with real pedestrians.
At a Y Combinator talk released Monday, Dolgov took direct aim at Mark Zuckerberg's famous mantra that guided Facebook's hypergrowth years. The "move fast and break things" philosophy works fine when you're iterating on a news feed algorithm. It's catastrophic when applied to machines that weigh thousands of pounds and share roads with school buses.
"There's simply not an undo and a retry button."
The error tolerance gap between digital and physical AI is the entire story. A language model or chatbot that makes a mistake gets corrected in the next prompt. A self-driving car that misjudges a turn destroys property, bodies, and the entire regulatory path forward for autonomous vehicles. One class of AI operates in the consequence-free sandbox of software. The other operates under physics.
This matters now because we're watching the agent economy split into two tracks:
- Digital agents: customer service bots, coding assistants, research tools
- Physical agents: delivery robots, warehouse automation, autonomous vehicles
- Different risk profiles, different regulatory futures, different build philosophies
Dolgov's thesis is that safety must be baked into AI models and training from day one, not bolted on later. This contradicts the dominant Silicon Valley approach of shipping fast, learning from production failures, and patching in post. That works for SaaS. It doesn't work for robotaxis operating in San Francisco during rush hour.
Waymo itself proves the point. Spun out of Google in 2016, still owned by Alphabet, the company has spent 15 years building toward commercial autonomous rides. That's glacial by startup standards. But they're now running paid robotaxi services while competitors who moved faster hit regulatory walls or worse, literal walls.
The Implication
If you're building in the agent economy, know which track you're on. Digital agents can iterate aggressively because the downside is a bad user experience and a refund. Physical agents require a different build philosophy where safety isn't a feature, it's the foundation. The regulatory environment will increasingly split along this line. Expect physical AI to face scrutiny that digital AI never will.
For investors and operators, this creates opportunity. The companies that figure out how to move fast AND ship safely in physical AI will own the next decade of automation. The ones that import Web2's "ask forgiveness, not permission" mindset will own the next decade of lawsuits.