构建物理 AI：Waymo 自动驾驶的经验教训

Building Physical AI: Lessons from Waymo's Autonomous Driving

本期速览 · Overview

Waymo CEO 分享在物理世界中安全部署 AI 的七条技术经验，强调与数字 AI 的差异。

Waymo's CEO shares seven technical lessons for safely deploying AI in the physical world, emphasizing the differences from digital AI.

演示只占工作的 1%，真正的挑战在于达到所需的可靠性等级。
A demo is only 1% of the work; the real challenge is achieving the required nines of reliability.
闭环模拟对于物理 AI 代理的训练和评估至关重要。
Closed-loop simulation is essential for training and evaluating physical AI agents.
评估和指标是战略护城河；在技术之前构建它们。
Eval and metrics are the strategic moat; build them before the technology.

可工作的演示与真正的产品之间的差距巨大；演示只是工作的1%。
The gap between a working demo and a real product is vast; the demo is only 1% of the work.
可靠性和性能需要指数级的“九个九”；每增加一个九需要十倍的努力。
Reliability and performance require an exponential ladder of nines; each additional nine takes ten times more effort.
对于物理AI，闭环模拟对于训练和评估至关重要。
For physical AI, closed-loop simulation is essential for training and evaluation.
构建物理AI智能体需要智能体、模拟器和评判者的生态系统，并由飞轮驱动。
Building a physical AI agent requires an ecosystem of agent, simulator, and critic, powered by a flywheel.
评估和指标是战略护城河；在构建技术之前先构建它们。
Eval and metrics are the strategic moat; build them before building the technology.

最好的AI时刻看起来像什么都没发生；安全是隐形的。
The best AI moments will look like nothing happened; safety is invisible.
在物理AI中，你必须快速行动并安全交付，而不是快速行动并打破东西。
In physical AI, you must move fast and ship safely, not move fast and break things.
弱感知导致安全曲线过早平坦；多种模态对于超人性能是必要的。
Weak sensing leads to a safety curve that flattens too early; multiple modalities are necessary for superhuman performance.
对抗规模的结构总是失败；引导规模的结构总是胜利。
Structure that fights scale always loses; structure that channels scale always wins.
你的模型可能被泄露，但数亿英里的安全数据更难复制。
Your models can be leaked, but hundreds of millions of miles of safety data are much harder to replicate.