My estimate / guestimate is three years. What is your estimate / guestimate for no accidents on average every 150K miles?
It's almost pointless trying to estimate/guess such a thing. Tesla themselves probably dont know, since this is new ground for everyone. However, software/AI systems like this tend to increase in reliability asymptotically, with rapid early progress that tails off as major problems are solved leaving only the edge-cases that may be much harder to solve. Right now I would say we are at the start of that curve and will make fast progress, if the beta videos are anything to go by.
The basic question, then, is how far along that curve do we need to be to meet and then beat the average human driver? My guess, given the poor quality of the "average" driver, is not very far. If that is the case, we may (repeat, *may*) see FSD exceeding average human performance in as little as 6-12 months. Technically, the issue has been: can the AI system create an accurate enough representation of the situation around the car for it to make safe and effective driving decisions? Many of the bets/posts on this forum (and others) have been skeptical (and I confess to some of that myself), but Tesla seem very close to cracking the vision issues necessary.
Sure, the car will probably remain stumped by some things, like complex road diversions, or missing road markings. But if it disengages and/or stops safely then I think FSD will be shown to be safe sooner rather than later.
Remember that agencies like the NHTSA are primarily interested in
safety, not accuracy. They dont really care if the car goes round and round in circles, takes wrong turns, or just sulks in a parking spot .. it just has to do these things
safely. It should also be remembered that
right now, the car is probably better than humans at some things. I bet it's reaction time to stop for a pedestrian that runs into the road is far better than a humans.
This means that, while the car may lag behind humans in some intricate situations, its ability to just drive reliably mile after mile may allow it to beat out a human on average sooner rather than later. This is why so many posts here that focus on "the car wont be able to do X special case" miss the point .. the point is,
can the car do better at things humans are bad at? And we are bad at stuff like running red lights, or forgetting to look both ways at a stop, or not seeing a car in our blind spots. Precisely the things that cause the majority of accidents, and
precisely the things the car is going to be very very good at very soon now.