Hacker News Reader

Hacker News

Comment by wilg | Hacker News Reader

Full thread

wilg

I would like the NHTSA to gather and publish robust data on:

• Crashes per mile on each manufacturer's "self-driving" systems

• Crashes per mile on each manufacturer's LKAS (lane keeping assistant) systems

• Crashes per mile on each manufacturer's TACC (traffic aware cruise control)

• Crashes per mile on each manufacturer's regular cruise control

• Crashes per mile under manual control

Break each out by highway, suburban, city, and car model year. Also include whether the crash resulted in injuries and deaths, and whether those deaths were occupants or not.

Then we can actually determine where the problem is, and the scope of it.

View on HN

Seriously. Tesla loyalists will say there’s net fewer accidents because of Autopilot, and maybe there are, but it’s still Tesla’s responsibility to prevent the system from being misused or entering dangerous situations, whether through negligence or ignorance. Hopefully the recent recall was a wake up call.

I'm not sure what this has to do with what I said. I want to know if there actually are fewer accidents because of Autopilot versus other situations. People are suspect of Tesla's own published data, but we can just easily answer it. For some reason, nobody advocates for finding out what's happening, just complaining one way or the other.

More numbers would be great they always parrot Teslas statistics comparing the number of miles per accident that people felt comfortable enabling AP (easiest highway miles, more comparable to when other vehicles have cruise control enabled) vs all miles.

> Then we can actually determine where the problem is, and the scope of it.

Unpopular opinion: there probably isn't a problem? None of these system is going to be perfectly safe, but there's quite a bit of evidence at this point that they're all safer, as actually used (which, yes, includes some misuse) than a human driver is at doing the same task.

No one wants to hear that though, so instead we take "where the problem is" as a prior and continue arguments like this ad nauseum.

Full disclosure: my car is going to drive me to Idaho in the morning.

Then this data will prove that, and eliminate the problem of people worrying about it.

There's plenty of data already though (of which a very clear LACK of signal should be a pretty obvious one). People, likely you included, discount that stuff because it doesn't confirm their priors. And so they demand new data that they believe will. And if it doesn't? Goalposts are movable for a reason.

At the end of the day you either believe or you don't. But if there was an actual problem, we'd know. There are millions of these things on the street now.

It's not a question of belief, it's a question of collecting the data. The data I mentioned is reasonable, easy to collect, and currently unavailable.

You're elsewhere in this subthread dismissing existing evidence (insurance rates) because it doesn't confirm your priors. I think we both know what would happen if this study magically appeared. "Demanding" data is a cheap evasion. The requirement for rational discourse is accepting the data you have.

lmao, nobody presented evidence based on insurance rates. I just suggested there are many other compounding factors that make insurance rates difficult to use for this purpose.

It's also absolutely hilarious to me you're assuming I think Autopilot is unsafe.

You could look at liability insurance rates. If the same driver has a lower insurance rate for a car with a certain automated system, then that system probably makes driving safer.

Too many confounding variables. People want to know if these systems are safe. Should be trivial for a regulatory agency to make this happen.

I suspect age correlates with vehicle choice and know it correlates with accident rates, so I'm not sure those figures would tell you what you think they're telling you.

OK, include driver age. Have some people put together a robust research plan that accounts for various biases. Should not be particularly difficult.

I see you claiming it's not difficult, but I don't see you making it happen. If it's so simple, take your developed plan to the authorities and make it happen. Show them what data they need to collect and how they can force manufacturers to provide it.

> • Crashes per mile on each manufacturer's "self-driving" systems

...counting miles where self-driving was activated including the 5 minutes afterwards — https://www.motortrend.com/news/nhtsa-tesla-autopilot-invest...

(Honi soit qui mal y pense.)

This is a silly article. The numbers Tesla releases themselves say "we count any crash in which Autopilot was deactivated within 5 seconds before impact", so Autopilot aborting less than one second before a crash would still be counted as an Autopilot crash. (I think this happens because AEB takes over.) But sure, this should obviously be accounted for in some well-run government study.

Reply on news.ycombinator.com