Before I begin I'm going to acknowledge that I'm unable to take Elon seriously when it comes to anything relating to Autopilot or FSD. He lost all credibility in those specific topics. I do take him seriously when he talks about SpaceX though.
It's a strange dichotomy to say the least.
The way I see it is EAP is essentially a piecemeal approach so I don't find it that hard to believe FSD will be anything different. Especially if you remember the old days where they gradually raised the speed at which you could use EAP. Oh, and how they initially started with only one camera activated, and then two cameras.
It's also worth pointing out that there is precedent in the market place of having ADAS systems work at only traffic speed. Like you can drive a 2019 BMW X5 with its enhanced traffic assist package at up to 37mph without hands on the steering wheel. It has a driver facing camera (it's only a visual camera and not the fancy camera like the Supercruise system uses).
The Audi A8 in Germany has L3 capability at up to 37mph on the white listed roads (only divided highways). You can read a book, watch the scenery, etc at these low speeds. Audi hopes to reach 80mph in the near or midterm future.
Will Tesla go that route with FSD? I dunno, but they can.
There is also the question of at what point do EAP features end, and FSD features begin?
The easiest example is stop light detection. I certainly believe Tesla will have stop light detection in HW3 within 3-6 months after release. But, it likely won't be given control over the brakes. They'll probably go through a 3-6 month validation period to make sure the detection rate is accurate enough to enable it.
Tesla likes having lots of time to "validate" things. Like right now we're supposedly validating NoA lane change suggestions so that Tesla can unlock unconfirmed lane changes.
I find it really amazing that people who argue that EAP is almost feature completely so easily glance over two of the hardest things about the EAP. Those two things being Smart Summons, and Unconfirmed lane changes. In both situations the accuracy of the camera detection is paramount.
Even when EAP is feature complete I'd argue that it's still far away from what's really wanted due to weaknesses in the system.
NoA with EAP is essentially L3 ready if it works REALLY, REALLY well. The only difference at that point is transferring liability to the car from the driver for an approved stretch of freeway. You can't sleep as you have to be ready to take over. With L3 you're given a certain amount of time to take over, but it's not enough to take over from sleeping.
So what Tesla is going to do is add features to HW3 vehicles that HW2/HW2.5 vehicles don't get. These features won't specifically target EAP, but will profoundly impact how the vehicles operates under EAP that they might as well.
Things like debris detection, and possibly pothole detection (dunno how well HW3 can detect potholes using the forward facing cameras).
So while you might not see FSD features I will say that you will see some dramatic improvements in some areas.
When will FSD actually be a reality? It will likely start right around the time of EAP on HW3 is so good that it can easily be approved by whoever these regulator people are. In Germany it should be straight forwards as by then the Audi A8 will already be driving on L3 in Germany. In the states the situation is more difficult because it has to be approved state by state. California is the most important state to follow when it comes to Tesla, and I don't know how that will play out. California is such a nanny state that they have ridiculous cancer warnings on everything under the sun. They'd probably try to put a label on the sun if they could. So if L3 can't be enabled in California then I doubt any other state will have it. If it's allowed in California then I expect it to quickly spread.
I don't expect us end users seeing much in the way of surface street FSD except for things like roundabouts, etc. They'll probably sneak in roundabout capability into HW3 EAP, and not say anything about it. Not the crazy roundabouts, but the simple ones.
With L3 on freeways I think we'll hit the limits of what HW3 and the current sensor suite can do.
Usually hardware can only be in front of SW for a generation or two.
The software gets to a point where weaknesses are seen in the hardware, and then the whole pattern repeats. The next car will have rear corner radars, and an inexpensive Lidar. Tesla will simply buy a small Lidar company that has a cool product.