Hacker News Reader

> How is having tons and tons of random camera footage going to help?

One way to think of this is that that the footage is implicitly labelled: we have the benefit of hindsight: we know what the state/location of the vehicle was going into the future. That benefit of hindsight also can serve as implicit labels by knowledge that the vehicle did not crash or collide with something immediately after the footage.

Fairly simple, they have footage from every time a person had to correct AP's driving. Take that footage label it. Train some of it, save some for test cases. Finally don't release an update until it drives better than the current system.

For a lot of problems, clean data labeled data is no longer the bottle neck, or in some cases, there are ways around it. The bigger issue is dealing with the "long-tail" of unknown new scenarios. This is a currently unsolved challenge.