Elon Musk: Tesla Autopilot | Lex Fridman Podcast #18

Musk argues that non-autonomous cars will soon be as niche as owning a horse, and claims an autonomous car could be worth five to ten times more than a non-autonomous one over the coming years.

With hundreds of thousands of cars continuously collecting multi-sensor data, Tesla believes it holds the vast majority of relevant driving data, enabling rapid neural network improvement and edge-case learning.

Tesla’s in-house FSD chip provides an order-of-magnitude more processing than prior NVIDIA hardware, with dual redundant systems, so future capability gains can largely come from over-the-air software updates rather than hardware swaps.

Any time a human takes over from Autopilot, Tesla analyzes whether it was for convenience or a system shortcoming, using these events to refine trajectories (e.g., optimal paths through intersections) and improve behavior.

He predicts that fairly soon Autopilot will be so much safer than humans that requiring manual supervision may actually reduce safety, analogizing to how elevator operators became an unnecessary and dangerous layer over automation.