Note that I'm not an owner of a Tesla (yet). I'm in it for the long game and have put those funds in TSLA instead. That said, I'm now a believer after seeing this video of the newest firmware making lane changes:
Very impressive changes indeed!
The thing to remember about AP and NOA is that your mileage can, and will, vary. I forget the version number, but it was at least two, if not three or four, back that it worked really well for me. Including one case where,
before the non-confirmation was enabled, my Tesla took over and took me through a couple of interchanges leaving the interstate. For which I was very grateful due to traffic volume and rapidity of shifts. It was inspiring as to what FSD could achieve. And then I got an update.
And its little things, like a road I commonly traverse that has a mildly awkward curve. Human drivers frequently cross the line even though there's no real reason to and autosteer was no different. But I'd keep trying (if there was no oncoming traffic). And then an update occurred and it drove it
flawlessly every time. Until the next update. Currently its close to where it used to be: a disengagement results from my forcing the car to stay on the correct side of the line.
And yet, these same updates that made my experience less good
improved the experience of many others. While this variance may simply be due to the complexity of the problem, I think it is
inherent when relying on neural nets and you trade certainty for probability.
With some problems, e.g., computer vision, you are trading the
certainty of it
not working for the
probability of work. Whatever the probability is, its better than the certainty of it not working, and neural nets can be trained for high probability of recognition. This is why computer vision is done with neural nets -- but Tesla's approach is to extend this model throughout the computer driving model so you trade
certainty of working in a giving situation for a
probability that it will work. Not only will regressions occur, but you cannot readily predict under what circumstances they will occur.
So you get two steps forward, one step back and some people only experience the backwards step even though most of them experience at least one net step forward.
Just a point of caution about optimism relating to any specific video or testimony about progress with autopilot. By the same token, a further point of caution about skepticism relating to any specific video or testimony about issues. While progress towards FSD will not be monotonic it
is happening.