Intuitively Tesla's approach makes more sense, but would love to hear someone with domain knowledge on how much of a difference it can actually make (after all, you need quality training data and Tesla may now have to navigate through significant more noise).
