As we say in aviation - what's your mission, if I may ask? And how long do you plan to keep the car? Personally I think if you look at the progress of AP2 in year one it has been spectacularly fast and should encourage you not discourage you.
Stop focusing on Elon's imaginary timeframes. Stop listening to the hater brigade for a minute and start paying attention to people like
@verygreen who bring real knowledge, and look at the actual progress rate of AP2/2.5. Do some reading about neural nets and reinforcement learning and what is happening on the cutting edge. Remember that learning involves lots of struggle with no visible progress and then breakthroughs in performance happen. Remember that the brain of AP 2/2.5 is a swappable board. Finally, remember that Tesla is still, one year later, betting the company on vision and selling the FSD option. They are ramping up the Model 3 with a vision based autonomy system. This says something. Also remember a lot of progress is likely happening in traffic logic simulation that we know nothing about.
The simpletons here on TMC believe that "No me car turn by itself! Been one year! Me angry! Me car no workie! Me angry! Elon dumb! This impossibowl! See Waymo! See big spinning dome! Dat key! Me want spinning dome too! Why Elon dum?"
For me, my mission is driving long distances on highways for business. AP 2 suits this mission perfectly in current 0.42 form.
To believe the software will NOT improve drastically from here you have to buy into several ridiculous claims:
1 - The autopilot team suddenly stops making progress.
2 - Tesla mysteriously becomes the only self driving player unable to teach traffic logic to its cars, despite having a world class AI team and having the largest fleet of vehicles in the world by orders of magnitude, gathering image data for learning.
3 - Despite the fact that vision plus radar is already reliable for lane keeping and cruise control, for no apparent reason Tesla is unable to make the cars turn left and right without hitting things.
At this point vision is great and getting better. The real problem is traffic logic - that's where the hard part is and that has nothing to do with sensor tech. That's stuff like "If I cut one lane over what will the guy two lanes over do?"
George Hotz gives a great talk on this recently where hen points out that sensing is the easy part - the hard part is decision making and that has zero to do with sensors.
Also remember that reinforcement learning is bleeding edge tech and Karpathy is an expert at it.
For other stunning reinforcement achievements see Alpha Go Zero's recent crushing of the original Alpha Go.
@verygreen gives some good comments elsewhere here recently on the challenges of reinforcement learning for driving.
But if you want guarantees that the current setup will take us to Level 5 - I think you're asking too much. Very robust L3/4 I think is the likely outcome. Lack of rear radar or a camera high off the ground is my concern.
However even during a recent rainstorm I was able to discern incoming vehicles from behind on the rear camera despite it being covered in vision distorting water drops.
I can see us getting to a system that alerts you to take over if it loses confidence in its rear vision. So a Level 4 that requires frequent cleaning of the rear camera (maybe a quick wipe before each departure and then a touch up wipe every few hrs).