You can't base an FSD purchase decision based on the current state of the software, because the software is in active development.
When Tesla introduced AP2 in late 2016, it took them over a year to get the software to a level where it provided value beyond TACC. And now, it works pretty well at lane keeping and lane changing (though it is a little too aggressive in wanting to move out of the "passing" lane).
NOAP is relatively new, and overall works pretty well in most circumstances. The decision on when to lane change needs improvement - but that's something that Tesla will continue to fix through OTA, based on experience across the entire fleet of AP2/3 vehicles (something the other manufacturers don't have).
Using NOAP on roads other than limited access highways will likely require purchasing FSD. Even without regulatory approval to operate without a driver, FSD should begin providing value on non-limited access highways in the next 12 months - allowing the vehicle to operate under AP on more roads, even in driver assistance mode. And that will begin to provide value in owning FSD.
Of course there's uncertainty - and purchasing FSD now is a bet on Tesla eventually getting the software to provide value to justify the purchase cost. Will FSD provide something of value long enough to provide sufficient value over the remaining time you'll have the vehicle (even in driver assistance mode)?
If you're only planning to keep the Model 3 for 3 years, then it probably doesn't make sense to purchase FSD - but if you're planning to keep the vehicle for 6-8 years, then it may make more sense - especially since Tesla will likely charge more in the future to activate FSD after purchase.
When we purchased our 2017 S and 2018 X, we decided the risk was low enough to justify purchasing FSD for both vehicles, and based on the progress Tesla has made since those vehicles were purchased, and our plans to keep both vehicles for 8 years or 100K miles, we're comfortable with our decision. Though proportionately, FSD costs much less for an S/X than it would for a Model 3 - which might have had a different decision if we'd purchased a Model 3 last year instead of an X...