Hacker News Reader

Comment by pcblues | Hacker News Reader

I dabble in image recognition. I could understand the signage on the back of the truck looked like an overhead sign much further away. This is the sort of whoops, sorry accidents ML will make. I'd still like to see a deaths by people driving normal cars/normal cars on the road vs. deaths in automatic cars/automatic cars on the road, by the hour on the road. If it's less than real drivers then I say charge on.

View on HN

> deaths by people driving normal cars/normal cars on the road vs. deaths in automatic cars/automatic cars on the road, by the hour on the road

These are not qualitatively the same. Drivers have a lot of control over their individual accident rate, by driving more dangerously or more safely. I want to know if the automatic system is safer than me, not just safer than the average.

And to state the obvious: the appropriate regulatory reaction to a death caused by AI is not "charge on" but investigation and improvement.

Since you may be on the reciving end of a self-driving-system gone awry, having them simply be better than average would be a good thing. Especially if we require them to always be better than average as they begin to proliferate.

> I'd still like to see a deaths by people driving normal cars/normal cars on the road vs. deaths in automatic cars/automatic cars on the road, by the hour on the road

Since "deaths using autopilot" is a statistical bucket unique to Tesla no such comparison can be made.

We have no equivalent death numbers for "normal cars" operated either 100% manually under autopilot-appropriate conditions, or primitive cruise-control. If you just make the comparison against automotive fatalities in general, it's a bullshit comparison, since it would include all driving conditions.

>I'd still like to see a deaths by people driving normal cars/normal cars on the road vs. deaths in automatic cars/automatic cars on the road, by the hour on the road

This is not enough, you will need to also make sure you consider:

- type of car , what safety systems have (maybe compared with a car with intelligent safety system the stats will look completly different, maybe even worse for Tesla)

- age of the car, probably will correlate a lot with the above,

- the road , we know Tesla will fail to drive on hard road section so would be fair to make comparisons on the same easy roads sections.

- IMPORTANT: you need to count the cases where a Tesla human driver intervene and saved the day, this incidents won't appear as crashes but they would have been crashes if the human was not there, you would need to ask Elon to be transparent with the publish and publish this, from what we see in YouTube videos the human driver had to intervene a few times in an hour to save the situation from critical situations

- optionally produce a statistic where you exclude drunk drivers, people on phone, people that got asleep , this cases could be fixed with cheaper tech so I would be curious if the cheaper tech would be safer but is just not sexy enough (or in fact as I suspect self driving is not about more safety and just having a private driver, the safety is just an inconvenient , without it Elon would have removed the Beta label from the FSD

> This is the sort of whoops, sorry accidents ML will make.

Vision based ML. Is this not one of the many cases where LiDAR would have prevented this from happening?

Or really any kind of radar where you fallback to, if there's something detected we should probably stop. A lot of Tesla folk will go on about radar causing brake checking, but I have a vision only Y and get it just as much as I did on my radar equipped 3.

More sensor input, especially at different frequencies which are going to have different characteristics and so different capabilities and limitations regarding what they detect, is certainly going to provide more data. Since LIDAR is on the order of $1000 per vehicle now, it shouldn't really be a cost concern, but the engineering to incorporate and properly merge that additional sensor input is not insignificant. Tesla has certainly been aggressive in its "FSD" market push, and the quantity of accidents absolutely should be a concern. If the NHTSA determines that the system as-is is sufficient (that is, at least better than the average), then they should say so - but the reverse holds true too.

This article gets into it in a lot of detail —https://www.forbes.com/sites/bradtempleton/2020/10/28/new-te...

It’s per distance and not per time, and number of crashes and not number of deaths, and the roads where autopilot can be enabled are usually safer and easier, but it is indeed safer: https://www.tesla.com/VehicleSafetyReport

Here we can see again that monocular depth estimation is an ill posed problem.

Maybe we should not be relying on ML to do this until it will no longer make a "whoopsie" and decapitate the occupants on the reg?

But instead we'll be lab rats so some SV manchildren can play with robots

The issue with comparing these numbers is that the distribution of quality and safety of roads is not the the same for data collected by Tesla vs. "human drivers". It is easy to claim that in relatively safe environment of a interstates and highways Teslas did better than human drivers everywhere else per 1M km.

TL;DR; even if this is truly your only metric to judge these systems, you can not simply compare these numbers.

Is this satire?

Why would it be? You can’t let a small handful of deaths grind progress to a halt especially when the alternative is the hundreds of thousands of deaths we see on the roads today. Take precautions, learn from mistakes but don’t shut everything down.

Except, Tesla removed radar from their cars, a relatively inexpensive and simple mechanism that significantly reduce their fatality rate.

It’s a false dichotomy here it’s frustrating, especially when they had implemented a check to their software and then removed it to save money on a $50,000 car

> Except, Tesla removed radar from their cars

If I had to guess and be mean, even though I like Elon Musk, I'd say that was one of his directives; I liked watching Elon talk about rocket technology, but one day I stumbled about his ideas about software development at PayPall and it made me cringe.

Yeah I think that’s a huge mistake. I would have hoped Elon had enough foresight to see that volume production would make LIDAR cheaper. He has painted himself into a corner. However, he seems like he could play off changing his mind on that decision pretty easily so I hope they go that route at some point. Only problem is it would make a lot of the Tesla’s sold so far incapable of doing FSD so I don’t see that happening anytime soon.

So many reasons!

Like the "I dabble in image processing" lead in (which like saying 'I've been known to brush my teeth' then launching into a discussion of removing the thyroid via arthroscopic surgery through the lip)

Then the tone of the rest of the comment, so light and cheery with "whoops sorry!"

And then the implication that a system where a classifier that confuses a truck for an overhead sign killing people is ok* as long as cherry picked stats check out (*for those of us in the space without arbitrary restrictions on our sensor suits because a shyster tweeted "we only need cameras) that wouldn't

I work in the AV space and despite staking our livelihoods on this problem space yet haven't managed to pick up such a cavalier attitude.

Human lives are not fungible. If I gave you a button that saved 8 random lives and killed 4 others would you press it?

If that button were 100,000 lives to 4 I would.

I mean if you want to dodge a realistic question with an unrealistic rebuttal you're free to...

AP has been finally implicated in 15 deaths since inception.

Do you think AP has prevented 375,000 deaths... or essentially wiped out traffic deaths in the US for 9 years?

Hint: more than 0 people died in car crashes last year

I was sure it was, but now I’m not so sure having read a few the sibling threads

Reply on news.ycombinator.com