The way Autopilot works right now, it takes in days from the 8 cameras and runs neural nets on each camera feed

then combines the predictions of each camera in planning and control, and uses that to move the car if both computers plans match

the way it works after the rewrite is that the hardware composites all the cameras into one “3d” image view

and then you just run the neural nets on that combined view

This lets the labelers have better tools where they can label video in 3D rather than individual frames

and then the chip team is working on dojo, special chips specially designed to hardware accelerate training using the 3d video as input

so the labelers just label the video, and dojo trains on massive amounts of it at ludicrous speed to build all the neural networks AP uses