Been thinking about this a long time...my AP1 car has had some scary moments, but is approaching "silky" in most scenarios . But my suggestion to everyone with AP cars is to leave it off and drive your very best, diligently. Before you get fired up about my opinion, read why I think this is best course of action:
As I understand AP now, it's uploading data to AI in the cloud, parsing the data, sending data to the cars to improve their accuracy and behavior. Given that (at least for AP2 cars), which is the better way for us to behave? Scenario A: If we allow EAP to do it's thing (it makes its occasional misjudgments and we jerk the wheel back, accelerate or brake frantically), it only learns that it's done something wrong but it didn't see how to pass through that area correctly. Scenario B: However, if we drive the car unassisted (but EAP is still watching) and pass through a "dangerous" spots like narrow bridges, over oddly painted crosswalks, under bridge overpasses, etc, without incident, the car will learn that behavior and thus behave better in those "trouble" spots for the current MX/MS and 1000s of M3 drivers soon to pass that spot.
So, unless I misunderstand the way Tesla EAP/AI system is functioning, I think we should be teaching the system ourselves, as good experienced drivers, and not allowing EAP in its nascent stages so be figuring it out by itself. We often note that EAP is "like driving with a teenager"...very few of us would set the cruise at 80 for a teenage driver in dense urban freeway traffic and then look down at our phones, would we? In fact, we'd have had our pre-teens watching our driving behavior for years before they get behind the wheel. Then we'd start them slowly, with some off-street training, and THEN ease into public driving, and THEN more complex scenarios. The magical thing about EAP is that it's leapfrogging the human model, but we shouldn't expect it to parse an incredibly complex environment immediately without proper training. I'm guilty too...my de facto behavior has been to set EAP near the speed that I would drive myself, based on how quickly I want to get somewhere, but would I do that with a teenage driver?? No...I'd advise them to go slower, watch them like a hawk, ready steer at any time.
So I think we've been approaching EAP the wrong way, and my suggestion has a broader Tesla community benefit: stop using EAP, but let it watch you drive. This TMC (and TMC community should consider harnessing the wisdom of a crowd of good Tesla drivers to more rapidly improve EAP, so we can all enjoy a rock-solid EAP (and FSD) sooner, and possibly prevent catastrophic incidents). Some will say "I paid for it so I demand that it work now". I paid for it too, understanding fully the caveats that came with it, but now I have an opportunity to help improve the technology just by doing what I was already doing before?!? AFAIK, no other brand owners can say their car gets better as they drive it.
And of note to the OP, my AP1 car has improved substantially since I bought it, and further still since Musk's Silky Smooth comment. I had expected his comment would only apply to AP2 cars, but I'm impressed with the smoothness of the system now. But I don't look away for more then a moment and my hands are always ready to catch that "teenage swerve" FNAR.
There has been absolutely no acknowledgement from anyone at Tesla or from my Service Center that any fleet to mothership learning has or is taking place. I have driven the exact same route on my daily commute since March 2017 with my EAP HW2 car and the exact same errors happen with either Auto Steer / TACC or both engaged. There are several areas that I HAVE to cancel AS / TACC if there is anyone anywhere near me due to the dangerous maneuvers EAP makes.
I am sorry to say I really don't believe that anything you wrote about possible learning is actually happening.