I paid for FSD three years ago when it was on sale, mostly so I could get HW3, as I figured that would give me the best possible on-highway EAP experience, which I do value. Since actually installing FSD software about a year ago, I've used it a handful of times on surface streets, and as I expected, it's never performed well enough for me to want to use it for any length of time, particularly in the small historic downtown where I live, with narrow streets, lots of pedestrians, somewhat unexpected lane shifts and so on.
With every new release, I do try it out on the interstate once or twice - my commute to work is over 90% interstate - and while it can be smoother than EAP, it makes too many poor decisions, particularly with regard to lane changes, to be useful to me. Most of the way, I drive in the HOV lane, and there are 2 or 3 HOV-only exits to the left on my route; none of them are ever actually part of my itinerary, but FSD always seems to drift into them, forcing me to override. EAP handles these exits much better, though I do still have to be vigilant when passing them on EAP, as it's not perfect, either.
FSD also insists on making lane changes at inappropriate times, often seemingly for no reason, and with very little warning; I suspect that in some cases this is due to Tesla's navigation engine, which is actually pretty good apart from the fact that it fails to take into account time savings from HOV lanes when building a route, even with HOV lanes enabled in the options. So it will often want me to go around the city (where there are no HOV lanes), as that route might be faster than the route through the city without HOV lanes, but in reality, the same route through the city WITH the HOV lane is the fastest. Waze knows this, but Tesla's system does not, and so while on FSD, the car wants, for instance, to leave the HOV lane on I-75/85 in downtown Atlanta and move across 6 lanes of traffic so it can take the exit for I-20 west and eventually I-285 north, which eventually meets up with I-75 again on the northwest side of town. However, I know (and Waze confirms) that sticking with I-75's HOV lane all the way through town is likely to be faster, and far less stressful, as it requires zero lane changes and there are almost no semis, unlike 285, which is littered with them.
On the flip side, when lane changes ARE necessary to maintain the proper route, FSD will often wait much too long to initiate them. Just to see what it would do, I have let it get as close as 3/4 of a mile to my exit with no action on its part, while four or five lane changes to the right were necessary. Of course I took over manually and knew that I wouldn't make my exit without driving dangerously, so I went on past my exit and made a detour. Normally in that situation, I'd begin moving over at least two or three miles in advance, making each lane change when it's easiest and safest to do so, rather than trying to do it at the last minute.
Ideally, I would like FSD on-highway behavior, as I do believe it's smoother than EAP and like the fact that it'll nudge over for large trucks and such, but with me still in charge of all navigation and lane-change decisions - basically, EAP with the FSD driving engine. I know I can kind of approximate that by not inputting a route into the navigation system, but I do find it useful to have my route on the big screen, for ETA and the very good visualization of traffic along my route. Hopefully that's where we are eventually heading, with one Autopilot engine for all cars, regardless of whether they have FSD enabled or not.