My guess is Autopilot is a system that can work all on its own, but it is assisted by metadata based on GPS and just uses automated machine learning. Stuff that requires a firmware update is how the system uses and interprets and reacts to the metadata.
The metadata can be simple values that help give the autopilot weight on how to make its decisions. These values are created by when a majority of Teslas make the same decision in a hold or disengage situation. For example, when 45 of 50 Tesla trips on a highway exit don't take the exit, the metadata tells the autopilot at this GPS location that the left lane marker should be the highest priority over following a car or using a right lane marker. I don't think Tesla requires any manual review whatsoever for this situation--they probably manual review things that Autopilot can never get right, for example turn curve speeds (which needs an update to fix).
This metadata wouldn't be much different than a navigation map having street numbers loaded in, business, names, etc.--even if Tesla has few dozen variables for every section of street, this data would only need kb worth of data to stream while on the move.
I'm guessing it's sort of like a video game replay file, where you can watch a 1.5 hour long 4k hd "video" with a file measuring just a few dozen mb in size. The file just contains what commands were issued during the game, and the program just recreates what happened exactly without actually recording a video.
My opinion anyways!