Thanks.
Weird how people have such different experiences.
Have you never had NOA abort a lane change in the middle, or give you the red hand on wheel alert, or phantom brake, or oversteer into an exit or miss lane exit entirely?
Don't get me wrong. I've had good experiences with NOA too. But I have experienced some issues too.
I have a 2018 Model 3 with AP2.5. Commute 60 miles one way weekly, 90% highway.
I'll use the above questions to frame my experience. I have had lane changes abort, mostly manually commanded changes, and maybe once from NOA commanded change. All for identifiable reasons (missing lane markings or approaching cars). I've never gotten the red hand alert during normal driving on a highway "for no reason". I have gotten it on non-highway roads when something weird happens, usually lost lane markings during a curve or traffic blocking the view or something. Phantom braking used to be much worse. Now it's far better than it once was. Yes I see it but rarely, and more importantly, predictably. It happens at certain points, which I know where those points are and can plan accordingly. I won't call this an "oversteer" but it does aggressively steer into a certain exit lane I use during my commute. Never ever missed an exit.
My general outlook is probably more optimistic than most have stated here. I agree with other comments that have said overall AP has improved over time while some updates have regressions, which are usually temporary. I think we have to keep in mind that AP is still a work in progress - aka not perfect and not finished. The way I approach it is to learn its strengths and weaknesses and use it when it's practical. "Set it up for success" you could say. I know when I can trust it and I know when to be cautious because either I'm unsure if it will succeed, or I know it's likely not to succeed. As an example, we all know it isn't intended to automatically stop at traffic lights and make a left turn through a complex intersection, so I don't "expect" it to and don't get disappointed if I try to make it do that and it fails (although it HAS successfully made a left intersection turn before). Based on that usage pattern, it's not perfect, but I'm happy with it.
Many have said it starts to try to change lanes for an exit roughly 1 mile early. As far as I can tell, it still isn't designed to/doesn't have the capability to intelligently seek out a gap in traffic to perform a lane change if there isn't one directly next to you. In other words, it won't accelerate to get into a gap ahead of you because it knows there isn't a gap behind you. Since I know that, I don't expect it to succeed in situations when that occurs (yet). If that does occur, I manually accelerate or decelerate (using AP speed, not disengagement) to put it into a position that it CAN lane change, and then it does great. It's not FULL self driving yet, it's driver assistance. I like when it changes lanes for me, so I help it get into position and it helps me change lanes.
There are two places on my commute where it hiccups. It's consistent every time. At one exit, the road merges from two lanes to one and it wants to put me in the one that disappears. This is consistent and a very simple fix. I just cancel the lane change and all is fine. This is clearly a lack of map/lane knowledge of that particular exit and I expect is highly likely to improve over time. If I DON'T cancel the lane change, it still works, it just merges back over when the lane ends. While it's not exactly like a human would do it, technically it still "works". Another exit has a slight upward slope to it compared to the highway. It tends to phantom brake here but only for a brief instant and then continues on, only a few mph lost. This did not occur on a previous update, just like the aggressive lane changes into exit lanes. I chock this issue up to the AP code being more broad and capable of detecting more "stuff" than it used to, which is a step toward it being a more general solution than just staying in lanes and taking exits - this is a good thing. My hunch is it thinks it "sees" something that it wasn't previously trained to "see" at this upslope exit, but then it realizes it isn't actually there. While it's a bit buggy in certain cases like this, I believe, overall it is getting more capable on the back end, which we don't directly see, and will improve.
My only real issue with NOA is speed commanded lane changes. I'd say it's about 50/50. There are definitely times when it recommends a change as I approach another car and I'm thinking "why? This car isn't going slower than me." Then 10 seconds later it is decreasing speed because the car ahead of me is in fact going slower than me, and i'm totally impressed! The negative side is when both lanes are relatively slow and the other lane is slightly faster it will recommend a change. I personally don't like continuously changing lanes like this so I just cancel it, but it keeps suggesting it. Because it immediately puts on the blinker when it recommends the change and doesn't allow you any time to cancel before the blinker comes on, I look like an idiot who keeps signaling over and over and never changes lanes. For this reason alone, I disabled unconfirmed lane changes and I've been much happier with it since then. This last case is really not AP's fault. That's when it recommends a lane change, and changes lanes, just for the offending car to then accelerate and me not be able to pass (without accelerating), now blocking the lane. Tesla can't change how OTHER people drive, so good luck to them on fixing that problem.
Last, I'll say this. NOA progress has slowed/stopped for a while. I don't know a ton about code but I know some, and I think it's safe to say the jump from NOA to full self driving is a massive leap. They can't just turn on FSD when it's 80% ready and say "have at it". That's a tall order. Also I think it's worth considering that Tesla has recognized that leap can't happen on AP2.5 hardware, so now they're at this point where they have to split the code between AP2.5 and AP3. If it's MUCH harder to get beyond the current state of NOA with the complaints we have mentioned in this thread, it makes total sense to kindof pause trying to "make it work" in its current form on AP2.5 and focus on improving it on AP3. As of now, we have only seen "FSD Preview" rolled out to the public, which leads me to believe we are basically all still on AP2.5 code, even if you have AP3. Elon has also said that everything so far has been optimized for AP2.5. I expect when AP3 is fully "turned on" with code designed for AP3, that's when the significant improvements will start to happen.