Two billion miles driven comes from Tesla's published numbers. Only one non-traffic fatality has occurred in those two billion miles driven. For Autopilot, it's one fatality per 130 million miles.
You can't compare AutoPilot to U.S. fatality rates, as Teslas are much more advanced, safety-wise, than the general U.S. car population. Much larger crumple zones, automatic emergency braking, strong aluminum shell, more airbags, etc. You must compare Tesla to Tesla.
There's a great binomial confidence calculator online here:
Epi Tools - Calculate confidence limits for a sample proportion. Enter in your own assumptions and confidence levels and see what kind of confidence intervals you get.
Examples:
http://epitools.ausvet.com.au/conte...pleSize=130000&Positive=1&Conf=0.85&Digits=12
and:
http://epitools.ausvet.com.au/conte...leSize=2000000&Positive=1&Conf=0.85&Digits=12
You have to pick the number of trials--that is, the number of times a human driver or Autopilot could have produced a fatality, but didn't. Bad weather, merges, debris, traffic, construction, idiots, poor road markings, etc. Above, I've picked one trial for every thousand miles. Above, the confidence intervals don't intersect, but it's only at a CI of 0.85. I was bummed about that, but then realized that I'm fine with it for practical decision making.