Hacker News Reader

Comment by justinpombrio | Hacker News Reader

> From everything I've heard of Tesla self driving is that per-mile it is more safe than human control (consequentialism argument for self driving).

I think this might not be true! I looked at the numbers a few years ago, and Uber had a much higher fatality rate per mile driven then human US drivers. Waymo was better; I forget if it was better than human or not. I couldn't find comparable Tesla numbers --- only PR-style numbers that could be based on all sorts of weird definitions --- but I don't get the impression that they're at the head of the pack.

EDIT: The number to beat is around ~15 deaths per billion miles driven, on the road in the US.

https://en.wikipedia.org/wiki/Transportation_safety_in_the_U...

View on HN

Tesla reports aggregated accident statistics publicly (for its Autopilot product, not this current round of FSD Beta testing obviously): https://www.tesla.com/VehicleSafetyReport

And indeed, they're better than human drivers. People will nitpick a little around the edges (e.g. cherry picking drivers of cars of the same age or in the same price range, pointing out the lack of specific numbers in various categories, arguing about the definition of "on autopilot"...), but the general truth is that yes: the "better than human" bar has already been cleared. We are not seeing the rate of accidents from this system that we would see if it was unsafe. Period.

> People will nitpick a little around the edges

Sorry what? That is a comparison of highway miles with Autopilot (driver felt comfortable activating autopilot in that stretch) vs all miles in any situation. That's not a nitpick, it's a completely apples and oranges comparison.

And that's autopilot. If you watch a video with FSD beta you can't honestly say it's better than any sober humans who have driven for longer than a week.

FSD beta has had no known accidents at all, though. The argument above was presuming that measurement and statistics can show whether a system is "safe" or not. If you refuse to accept that as an axiom I don't know how to reply.

The question in the linked article we're discussing isn't whether Tesla FSD is "finished" as a product, or perfect, or whatever. It's simply whether or not we should allow it to be tested in a wide beta on public roads.

And where's the evidence that we shouldn't?

> FSD beta has had no known accidents at all, though.

Do you have a source on Tesla FSD? I don't know much about it. The basic info we'd need is number of fatalities, number of miles driven, and to check that these numbers mean roughly the right thing.

You want a source on the absence of evidence? I don't have that. No one has reported an accident to a FSD beta car anywhere I am aware of. The source is me.

You are a fine source! So fatalities = 0. Any idea about miles driven, by Tesla FSD beta cars? Or is there no public info on that?

If it was human drivers (in the US), you'd expect a fatality by around the 70 million mile mark. And fatalities caused by self-driving cars are very newsworthy, so you'd hear about it.

> nitpick a little around the edges

It is not nitpicking around the edges when the foundation is shaky and opaque.

That's "nitpicking" though. Unless you want to claim the existence of a very large population of "autopilot accidents" not present in these reports, it's just not possible to construct a scenario where autopilot is significantly more dangerous than a human driver.

You're attempting to use uncertainty or controversy in the measurement mechanism to stand in for an argument for the contrary point, and that doesn't work logically.

It's not killing more people, basically. That's what "safe" means.

No, this is not correct. Tesla persists in comparing autopilot miles (primarily highway driving, only good weather, recent model cars) to all miles driven by all road users under all weather conditions.

Tesla cannot claim autopilot is safer than a human driver based on this type of comparison.

Still, they fooled you and many others in this thread, so I guess their marketing works.

Very much this is the reality. And moreover, people have been pointing this very, very obvious problem with the Tesla numbers out for years, and yet people continue to take Tesla's statements at face value. It is incredible to watch and the most solid demonstration of how weak the critical thinking skills are in the technology field.

For historical reasons, I have lots of friends who do what is now called "data science" professionally and literally every one of them has asked, unprompted, about whether the numbers Tesla reports are based on non-comparable sets, which they are, early in any conversation on Tesla FSD. It is totally obvious.

Next time people complain about “why do i need math class” here’s a great example. Unfortunately a huge majority of the reasons are for reasoning through advertising BS.

In my experience, the tesla crowd does not care, they'll just cite it in the next Tesla thread.

Do you have a source that constructs a story around existing knowns that does show AP to be unsafe, then? Because again, the deaths and accident counts simply aren't there. There's not enough of them to be "unsafe", no matter how you do the accounting.

That is, take the numbers we have and show that if you select out the right data set from public numbers about other kinds of driving, that the Tesla AP values are actually more dangerous. This surely isn't easy but it's absolutely doable (and, let's be clear, such an analysis would be worth a lot of money to Tesla's competitors -- there's a built-in incentive to do this). But no one has numbers like that anywhere I've seen. Likely because, as seems like the obvious hypothesis, there aren't enough events to make a story for a safety problem via any analysis.

So you are arguing about methodology and not results. That's pretty much the definition of "nitpicking". And it will never end. You'll insist that Tesla AP is somehow hiding phantom deaths for years, because it's your prior. But it's not supported, it's just not.

Tesla has this info but chooses not to share it.

However, here’s an analysis from a few years ago that should at least give you pause in claiming that Tesla’s are definitively safer:

https://medium.com/@MidwesternHedgi/teslas-driver-fatality-r...

Wow, that was a good read. Thanks for sharing

Reply on news.ycombinator.com