The human accident rate is about one per 500K miles, so if they were able to get in that range, then yes, they would have succeeded; drivers would be able to stop paying attention to the road without putting themselves and others in danger.
But the current FSD beta's intervention rate is more like one per 10 miles, judging from some quick googling. I see no particular reason to assume that incremental improvement can take us from 10 to 500K.
This is from from November 2021, but I'm still highlighting it because it is just terrifying (Correct time, though the video later on also exhibits inabilities of the system): https://youtu.be/9wRRClg_aM8?t=113
But are you using confirmation bias to find a cognitive bias that fits here.
But, In all seriousness we don't have access to the data across all 60k FSD users to know what the intervention rate is and how it has been changing over time.
We do have previous statements that as they get better they are moving to harder situations. Start with empty roads, and once you can do them well start finding harder and harder situations. When you start you avoid construction zones, once you are doing well you start looking for them.
Which could be a sign some drivers are simply overly cautious. Suppose 1/10 of disconnects prevented a crash, reducing the risk of crashing to 0 only reduces the number of disconnects by 10%.
To actually reduce that number you would need to make drivers feel more confident in the vehicle which is a useful metric, but only indirectly relating to safety.
What is the appropriate point of comparison though? All human drivers? Sober human drivers? Sober cautious human drivers? Sober cautious human drivers with driver assistance technology (e.g. auto-braking and blind spot warning, or potentially even more sophisticated LiDAR tech)?
I don't think this question is even meaningfully-defined. There is no "the" point of comparison. The relevant point of comparison is whatever ride it's displacing.
The rideshare explosion has already had a measurable effect on drunk-driving deaths; to the extent that a theoretical lower-cost AV will make rideshare even more accessible, then its effect on drunk-driving reduction absolutely makes non-sober drivers a relevant comparison.
For an average young person who'd get in the car with one of their friends, or drives a bit recklessly themselves[1], an AV at sober-human-driver level would be valuable.
For a guy who needs his kids driven around, a "sober cautious human driver" level of safety may feel right.
For questions like "what should the regulatory bar for launch be", all human drivers seems like an easy answer.
[1] I'm probably guilty to a degree here, on the rare occasions I drive
Is it reasonable to assume AV will be lower-cost than rideshare? The key thing that makes Uber more affordable than a taxi is that vehicle purchase/maintenance/depreciation/liability are all externalised.
In a full-self-driving situation you no longer have to pay your driver, but you do have to pay for all of the above. With the inevitably higher standards of maintenance required for AV fleet vehicles I can't really imagine it being cheaper than it currently is.
Sure the sensor/cv/vision tech will get cheaper, but machines still wear down.
> Is it reasonable to assume AV will be lower-cost than rideshare?
That's what the industry is betting on. I think it's reasonable in the steady-state: labor costs are expensive as hell.
> vehicle purchase/maintenance/depreciation/liability are all externalised.
These aren't 100% externalized with Uber, as they show up in the labor cost. They're only externalized with Uber to the extent that drivers do the math wrong on the costs they're paying[1]. Most of the analyses I've seen of this choose every possible pessimistic assumption, and still end up with net wages that are very high. They're of course low relative to "a living wage", which is what the analyses are focusing on, but that's precisely the point of what we're talking about: even the floor of labor costs is very high, when you're looking at expenses.
[1] Completely tangentially, but also note that this ignores the extent to which people derive value from being able to convert assets around. It's hard to imagine for us SWEs making 1% salaries and sitting on mountains of wealth, but liquidity is a constant and pressing concern for a large portion of the country. See also: payday lenders, where there's a stark difference between the opinions of those who've actually studied the economics of the industry and the midwit affluent John-Oliver-watcher.
> The human accident rate is about one per 500K miles, so if they were able to get in that range, then yes, they would have succeeded; drivers would be able to stop paying attention to the road without putting themselves and others in danger.
Unfortunately, I expect that automation will be held to a higher standard than human drivers, rather than the same standard. When an accident happens, people want to know who to blame, and an unimpaired human driver gets somewhat more latitude for a genuine accident, while a piece of software is always going to be perceived to be at fault (which it may well be, even in a situation where a human wouldn't be considered to be). And conversely, people (somewhat validly) want to have more control: every driver thinks they're above average, and the software won't be as good as their accident rate, and if something happened at least they were in control when it happened.
I don't necessarily even think those are incorrect perspectives; we should hold software to a high standard, and not accept "just" being as good as human drivers when it could be much better. But at the same time, when software does become more reliable than human drivers, we should start switching over to make people safer, even while we continue making it better.
(Personally, I wish we had enough widespread coordination to just build underground automated-vehicles-only roads.)
The real averages of FSD intervention are unknown since some 2,000 Tesla employees also have NDA'd Beta access, and it would surely differ between rural, suburban, and urban roads.
But the current FSD beta's intervention rate is more like one per 10 miles, judging from some quick googling. I see no particular reason to assume that incremental improvement can take us from 10 to 500K.