Humans, as well. 38,000 dead and 4,400,000 injured every year in the US due to car accidents.
It would be cool to decide on a framework where FSD is tested/approved to be not perfect but better than the average human
I'd love to see all cars being on FSD and communicating with each other around, dropping deaths (and even minor accidents) to virtually zero though.
Outside of (to a certain degree) drugs (and tobacco, which is regulated and known to be harmful) there are very few consumer-facing products which used and maintained properly just randomly kill people sometimes--and we shrug our shoulders and say "stuff happens."
That sounds like a world that'll be great as long as you're in a car, and terrible if you're a pedestrian or cyclist. Kind of like the world we've got now, except with even more victim blaming.
Cars communicating with one another means that they will prioritise QoS for one another, and pedestrian safety will be, as ever it has been, an afterthought—at least as long as there's more penalty for damaging a person's car than for injuring the person themselves.
FSD exists only because we built our entire society around cars.
Even in countries with good train service and lots of cycling, this certainly isn't the case today. Good luck locking everyone up into urban areas.
Long distance journey can be facilitated by interurban rails and railroads.
I'm bullish on self-driving in general, but Tesla's misrepresentation of their self-driving capability has always been just shy of "fraud".
Even average drivers die about once every 100 million miles. AP is so far from that it isn't fair to try and compare. And that's just average drivers, which includes drunks, tired people, 16 year olds, weather-related accidents, etc. Put AP or FSD up against a normal, sober driver in decent weather and there's probably five or six orders of magnitude difference in competency.
to quote wikipedias statistic from 2012:
- the United States had 3.38 road fatalities per 1 billion vehicle-km on its Interstate-type highways, often called freeways.[28]
To match a human, it needs to go half a million miles without an accident, and 100 million miles without a fatality.
Only it wasn't in a tesla, it was in a reasonably recent ICE toyota with adaptive cruise control and lane keeping.
So what was your point about FSD again?
Just watch the YouTube videos from beta drivers (which are cherrypicked to be most flattering to Tesla). The number of disengagements in 20 minutes of continuous driving is disheartening.
Yes. A society that valued both human life and evidence would ban cars entirely.
... hopefully have the sense to evaluate things based on both costs and benefits, and maybe even beyond just the first-order ones.
Even if this was true, there is still the problem that the failure mode of a ML is even less predictable than that of a average human. I.e. only the most intoxicated humans are likely to drive their cars into white walls, but that may happen any day with computer vision on the most well-maintained car, and you may never know what confused it. Or a traffic sign can be defaced in a way that is obviously detectable for a human (up to the point maintenance does not even consider to clean it), but confuses a car in a deadly way.
I.e. even if humans fail, at least they fail in ways that are relatable to us humans. The ways "FSD" fail are incomprehensible and exasperating.