Basically think of the NN as an expert driver that just has access to 30s of video data. For most situations this will be okay, but sometimes knowing approximately where you are will help you greatly. Often you can figure out where you are, but there will be times when guessing if you are in turn on red or not area will be tricky.
Adding another sensor adds failure modes. But without the GPS available, navigation will be very tricky anyway, so you might aswell assume it will be available but add some data where you remove it to teach it to drive acceptable without it.
Compared to the vision the amount of data is very low, two doubles vs millions of ints. It cost very little to input it to the neural network in extra compute/bandwidth.
I think FSD has a high level map and compares what it sees to the map, sometimes updating the map.
What it sees always take priority over the map... but if the majority of high rated drivers also turn left on red in a particular location, in the absence of any other visual clue, I think the car will perhaps assume that is what it should be doing... When there is no map data, what it sees is the only consideration...
Map data is most useful when vision is reduced...
So driving in a snow storm, with no lead car, and no map data, because no other Tesla has ever driven that route, may be problematic...
There is a difference between map data as a guide, and map data as an essential component,
Often what trusted safe drivers do in response to a particular set of visual inputs is the right thing to do, regardless of the technicalities of road rules. When safe drivers don't obey the letter of the law, the law might be sub-optimal.
In relation to judging V12 progress, I'll reserve judgement until Chuck Cook puts it though a series of tests on unprotected left turns, narrow roads etc, or I get it, and can test it myself. My hunch is Chuck will have it long before I do, and he will quickly find any problems,
If Chuck finds problems (and he probably will) the next questions are how fast FSD improves and if it regresses in some areas. What I expect is Chuck will find problems, but we should see a steady rate of improvement with no regressions... Any timeline predictions I may do, will be after I see Chuck test the next 3 versions....