you forgot that auto pilot likely is only used or only works under certain conditions. So that 1.2B miles is selective. You can't extrapolate that to the 10-20 deaths, which I assume you calculated from ALL miles driven
But are the parts that are easy for machines also what is easy for humans? Human attention decreases when things get more simpler and more monotonous while an autopilot can deliver constant attention.
The problem with systems like autopilot is that the human still needs to pay attention. The driver needs to take over in a matter of seconds when the system encounters a situation in can't handle. Autopilot makes things simpler and more monotonous while still requiring perfect attention from the human.
In theory that's a huge problem. In practice we have 1.X billion miles of data and it does not seem to be such a huge deal. Either the systems are already fairly good, or people mostly pay attention.
Granted, I am approaching this from the perspective of a more relaxing driving experience not necessarily from a pure safety standpoint. People spend 20,000+ hours driving in a lifetime making that less stressful is a huge benefit even if they are still stuck in their cars and can't get work done.
We need better number to do this comparisons, similar car models, similar roads conditions, similar driver population, since Tesla keeps the data hidden from the public I incline to believe the data is not in Tesla favor.