Jim talks with George Hotz about running Comma, an open-source driving assistance company. They discuss breaking the carrier lock on the iPhone at seventeen, Google's Project Zero, zero days, Mobileye & proprietary perception algorithms, cameras vs lidar, 6 levels of self-driving automation, the reliability of human driving, self-driving cars as "demo complete," why corner cases aren't the issue, integrated world models, the challenge of defining lane lines, recognizing the right part of the road, behavioral cloning, the hugging test, Comma's data set, the small offset simulator, how to install Comma in a car, what it does, why high-precision maps aren't useful, problems with Waymo's approach, "trackless monorails," why current systems still use remote-control driving, hyper-fragile centralized systems, Tesla's approach, against magical inflection points, self-driving as a stepping stone to artificial life, why Comma doesn't do marketing, the regulatory environment, eyes off vs hands off, why self-driving cars are easier than general robotics, liability, functional safety, the Tinygrad machine learning framework, who's using it, and much more.
Episode Transcript
Comma
Tinygrad
George Hotz is the founder of comma.ai and the tiny corp. He is working on self driving, robotics, and ML infrastructure with the goal of creating an operating system for silicon-stack life.
