Silicon Valley-Driven Hype for Self-Driving Cars | Hacker News Reader

Silicon Valley-Driven Hype for Self-Driving Cars

nytimes.com

73 pointsdbcooper52 comments

It's not Silicon Valley hype that's the problem. It's Tesla hype. Tesla's "autopilot" is way under-sensored. One TV camera, a bumper-level radar, and some ultrasonic range sensors. That's just not enough. (Tesla is supposedly adding more cameras in later models. They still need windshield-height radar.) It's also not clear how much redundancy Tesla has in the control system? Backup processors? Redundant steering motors? Redundant brake actuators. Google has all of those. Does Tesla?

The Tesla crash and rollover on the Pennsylvania Turnpike is troublesome.[1] That's the situation Tesla's "autopilot" is supposed to handle. Here's the crash location - Penna Turnpike mile marker 160, eastbound.[2] Concrete center divider, metal guard rail on the right, clear pavement markings, summer driving conditions. There's no excuse for a crash in that situation. That may have been an outright system failure. NHTSA and the Pennsylvania State Police are investigating.

Autopilot bug reports are written in customers blood. Musk seems to be in denial about that.

Denial isn't going to work. Look how much trouble General Motors got into over ignition switches with weak detent springs.[3] Keyrings with too many keys could under some circumstances move the switch and turn off the vehicle. (It wasn't just loss of propulsion; power steering, power brakes, and airbag power were also lost.) GM has paid out $3 billion so far, recalled 800,000 cars, and fired 15 people.

[1] http://www.nytimes.com/2016/07/07/business/us-safety-agency-... [2] https://goo.gl/maps/TfcuZR67oXr [3] https://en.wikipedia.org/wiki/General_Motors_ignition_switch...

Tesla has now reviewed the logs and says that there’s “no data to suggest that Autopilot was engaged at the time of the incident."

http://electrek.co/2016/07/06/tesla-model-x-rollover-acciden...

Tesla's response is evasive and designed to be misquoted as above. Here's their official statement:

“We received an automated alert from this vehicle on July 1 indicating airbag deployment, but logs containing detailed information on the state of the vehicle controls at the time of the collision were never received. This is consistent with damage of the severity reported in the press, which can cause the antenna to fail. As we do with all crash events, we immediately reached out to the customer to confirm they were ok and offer support but were unable to reach him. We have since attempted to contact the customer three times by phone without success. Based on the information we have now, we have no reason to believe that Autopilot had anything to do with this accident.”

Tesla has no idea if their autopilot was engaged. They lost communications during the crash, probably when the vehicle rolled over.

The customer told the Penn. state trooper involved that they were on autopilot at the time.

> Tesla has now reviewed the logs

This is an extremely misleading interpretation of the article. The article says Tesla has only reviewed the basic crash logs, which don't contain information about autopilot state. The detailed crash logs, which have this information, have not been received because the antenna was damaged in the crash.

So Tesla has only reviewed cursory information that says nothing about autopilot state; therefore, they can state they have "no data [suggesting] autopilot was engaged". But you cannot conclude autopilot was not engaged, because there is critical data missing.

Very misleading.

> Tesla has now reviewed the logs and says that there’s “no data to suggest that Autopilot was engaged at the time of the incident."

is a quoted directly from the article. It is also the title of the story.

Quoted from the part of the article the author crossed out and rewrote because it's inaccurate. And the title is a quote taken out of context.

I had not seen that edit. Agreed that the title is inaccurate.

The trend on these Telsa / EV blogs has been to take Tesla's logs as gospel and discount human observation.

There have been other reports of users claiming autopilot being engaged when it was not and reports of not noticing that summon was enabled even though one can't really miss that (that has subsequently been redesigned).

Human factors for these systems is incredibly challenging since such a wide audience of users useses them.

For the record, the crossed out and updated section was added two days before you posted your comment. At that time, the author added a link right at the top to a newer, more accurate article. The author also added the word "[Updated]" to the title so people would notice this change.

What I find interesting is that the tech for the autopilot largely comes from Bosch and has been in kther cars for a long time. It's branded "assist systems" in German cars and they are super anal about the driver oeeping the hand on the wheel at all times. I have a passat eith lots of the assist systems and I honestly think that in the current form it's also a safety risk of you start to depend on it too much. It encourages bad behavior.

Audio, BMW, Mercedes etc don't sell the driver assist as being an auto pilot. The driver is made aware that these are assisting tools and nothing else.

Without disputing your points about Tesla, there are a couple issues with your comments on the Pennsylvania crash which cannot be overlooked.

The Google Maps image date is September 2014 (fine print at bottom of page) and we must assume it looked the same on the accident date. It's entirely possible it looked very different on the accident date. Paint spills, fire or previous accident damage to road surface, debris on roadway, etc.

Also missing from all the reports are weather conditions. Did I miss them? I don't see it mentioned at all in the Times or Detroit Free Press articles. Are we assuming it was a sunny and clear day? That view is going to look very different during a thunderstorm.

Pennsylvania seems to be way ahead of other states in their willingness to research safety issues and make changes to their roads and laws. It was the first state I know of that made it a law that you have to have your headlights on if you are using your rain wipers. Makes sense now but they did it first and it definitely helps others see you. They also paint big white circles on the interstate in areas where rear-end crashes are a problem, so you have a very clear visual and know to keep at least three 'dots' behind the car ahead. There are plenty more examples but I feel like this is the perfect state to have something unfortunate like this happen.

> It's also not clear how much redundancy Tesla has in the control system

There need to be enforced standards. Prove to us that your car can handle itself (no pun intended), then we'll let your "autopilot" operate unrestricted. Not before.

We also desperately need auditing of the code running their systems. This is a situation where an if statement could kill somebody. These types of problem is why the NHTSA exists.

>> Tesla's "autopilot" is way under-sensored.

That actually makes it easier to develop the AI. More sensors means more noise and more work to get rid of it, it means more data and more expensive computations to crunch through it.

This is not to defend Tesla (I could care less) but to point out that "more sensors" doesn't solve the fundamental problem, that of creating a piece of software that can autonomously navigate an arbitrary environment.

We're nowhere neare developing that kind of software and, like people say in academia, it's really not clear how we're going to get to the point where we do.

The thing that strikes me as more than a little off-key about the hive mind-like push for fully autonomous cars is that most drivers don't want them: http://www.bizjournals.com/sanfrancisco/blog/techflash/2016/...

"...only 15 percent of respondents would want a fully self-driving vehicle as their next purchase. Meanwhile, 39 percent said they would prefer a vehicle that has some self-driving features. Nearly all respondents (95 percent) said they would want access to a steering wheel, gas and pedal controls even if a vehicle were self-driving, suggesting a wariness toward the technology. Brandon Schoettle, the study’s author, noted that public attitudes toward self-driving vehicles haven’t shifted much over the past couple of years, despite well-publicized developments in the space."

This sort of reminds me of Google Glass- the tech early adopters were praising it highly and worked the hype up to a thick lather, but "normal" people wouldn't be caught dead using it and the product belly flopped badly. The entire exercise seems like the self driving technology is being built for its own sake, as though the coolness of the engineering challenge will stimulate sufficient demand to justify the endeavor.

That really means nothing. If you asked people "do you want a machine that dries your clothes" or "do you want a computer you carry in your pocket" a long time ago you probably would have received tepid interest.

That's completely untrue.

It most certainly doesn't mean nothing. It means that that there are significant trust issues that must be overcome before autonomous driving technology can get widespread acceptance. This is not a pure engineering problem that must be solved. Technologists can endlessly repeat "autonomous driving is inevitable" but that alone won't make it so.

How can you trust a thing that doesn't exist? It's certainly understandable why people might hedge their bets. It's a lot to ask to get commitment to something sight-unseen. I have faith that the trust will arrive soon after the technology does.

In the article, they state that Google thinks it will be 30 years before cars will be fully capable without steering wheels. If I were a survey respondent, I wouldn't want to wait that long to get my next car either.

I think most people have driven a car with cruise control. Some cars have the "stay in lane" feature. People have certainly been riding in cars driven by others. I don't think it is a huge leap for them to imagine what it would be like riding in a passenger seat but car just drives itself.

As most political science people would know, it depends how the question was asked. It can be asked to insinuate that there would be a Windows 7 computer running their car and it would be taking the person 80 mph on the highway or through the crowded 5 lane bumper to bumper traffic to a job interview, or some important appointment they might say "no". They can immediately imagine horror scenarios happening.

I only found the abstract to couldn't see how questions were formulated:

http://www.umich.edu/~umtriswt/PDF/SWT-2016-8_Abstract_Engli...

Neither of those is an immediate evolution of their predecessor. Autonomous cars literally have the exact same structure of their predecessor (the manually driven car), they just simply automate the manual labor required to operate the machinery.

> If you asked people "do you want a machine that dries your clothes" [] a long time ago you probably would have received tepid interest.

You can ask this question of people today and still receive at best tepid interest. Dryers are unpopular in China.

And Europe*

*Washers with the built in dryer that never really works don't count.

Even if there is no consumer interest, the value to the transport industry as a whole will be enormous.

Sure, there is already significant interest in some specialized industries (long haul trucking, transportation around large scale mines, and others), and that interest will grow incrementally as the technology matures. But the article is more about the general consumer case. I think consumers will gradually warm up to autonomous driving as it is proven out on a large scale but that it will take much longer than early adopters predict.

The "as their next purchase" part of that option seems like it would seriously distort the results. I'm sure plenty of those people were thinking "I don't think I'd trust a self-driving car that came out today, but it's be pretty cool to get one later once I know they work well".

Instead of asking whether they want to buy a self-driving car, ask if they want a 75% discount on Uber rides. I think we all know the answer to that.

We don't know the answer to that, because it would be determined by how much the public trusts technology that doesn't exist yet (at least not in a form ready to be deployed en masse by Uber). And if they did trust it, there'd be no reason for Uber to give a 75% discount for using it.

Spend some time out of a city. Or at least trying to consider what life is like outside of a city.

I grew up on a farm in the country. Perhaps you should try to consider that:

1. 80% of the US population lives in urban areas.

2. The cheaper Uber gets the farther out from population centers it can expand.

What makes you think that owning, maintaining and operating a fleet of self-driving cars will be that much cheaper than just hiring people who own and maintain their own cars.

Reminds me of my parents who for every car they've ever bought got 0 add on features. I wouldn't be surprised if autonomous capability was seen as an excessive car feature and thus shied away from by everyone used to running their cars into the ground before replacing them with the bare minimum model replacement.

All they have to do to flip that poll on its head is to sell autonomous cars as a low end product rather then a high end product. As is Tesla is leading the pack with "cool car that cost $80k that you'll never own!"

Polls aside, there's a chance having autonomous driving could end up giving you a reduction in insurance premiums, resulting in it actually adding value that way.

Google Glass was a vitamin. Autonomous driving is a pain killer.

Something that annoys me about all this self-driving focus is that they are not targeting the obvious low-hanging fruit:

1) Focus on perfecting low speed bumper-to-bumper traffic fully autonomous driving on freeways.

2) Focus on passing legislation to allow zero driver liability and awareness when using the low speed bumper-to-bumper traffic self-driving mode.

The speeds are low and so the risk factor is reduced by orders of magnitude. Even if an accident happens, it won't be fatal because of the low speeds. And that should also help with the legislation issues. Also, this feature would be very popular. Imagine a commercial with typical bumper-to-bumper traffic. Then the camera zooms in and the driver is sleeping. When the car detects that bumper-to-bumper traffic is going to end soon, it triggers a loud alert which wakes the driver.

And yet all the car companies are ignoring this. I remember reading about early lane keeping and automatic speed cruise control and it would not activate under slow speeds. I was incredulous when I read it. They made this not work for exactly the most useful scenario? I don't care about self-driving for road trips, they don't happen often. And door-to-door self driving is the ultimate goal, but it's going to take a long time to get there. But helping to alleviate the horrible bumper-to-bumper driving by allowing drivers to do something else can be done with current technology. The whole self-driving effort feels like no one has done any market research to see how to make it help with real world situations.

Volvo XC90 has this feature. It's basically lane keeping with automatic cruise control that works between 0 mph and 30. It also requires you to keep hands on the wheel (that's not a problem, imho).

The only drawback to me is that if the car stops completely, the autopilot shuts down and has to be manually restarted (click button) to resume.

    > It also requires you to keep
    > hands on the wheel (that's
    > not a problem, imho).

That stops you from reclaiming the time, so is probably nowhere near as valuable/desirable as one that allows you to repost cat gifs.

This could make traffic worse. The only thing keeping bad traffic from getting worse is that past a certain degree of badness, some people start choosing to cancel or defer driving trips because they are not willing to drive with that degree of delay and/or annoyance. Making traffic easier to put up with, will have the same effect as reducing the price of any good would have -- increase the amount "consumed"

Self driving cars will lead to worse traffic unless tolls go up enough to make up for it

There was a recent story about a self-driving long haul trucking caravan across Europe touting the success of such. It was done in the spring and lots of high-fives around it...The route was in Sweden, Denmark, and Germany...aside from perpetual autobahn construction...each of those have very good roads and it was done in the spring.

"Self driving cars" still can't distinguish all driving hazards. The recent Tesla fatality, many stories made it a point to point out the driver was watching a movie.

So, let's move self driving cars to California roads. They will deal well with traffic in places like 880, etc. But, can they deal with poor road conditions, the overall shitty level of driver competence here, etc? I think that is many years off and not yet proven.

Another area to consider - uptake in new cars. I just bought a 2016 Tacoma in the last year. These trucks are known for lasting decades. I don't commute. I don't see replacing it anytime soon, let alone, ever having a self driving car for taking me places in the back country/forrest roads/etc.

It is hype and maybe it will happen for some fleet situations, but I don't see autonomous cars being the mainstay for private individuals in my lifetime.

Uptake in new cars will be a non-issue when Uber and Lyft make self-driving taxis.

The movers in industries are never the end consumers, but the large corporations.

Cities will replace HOV lanes with Autonomously Driven Vehicle lanes, and that will include the buses.

I've seen lots of stories about "cities" and their plans.

In the bay area, Morgan Hill to Mill Valley, keeping things to 101, are part of the same suburban gridlock. Do I see that area changing by "autonomously driven vehicle lanes"? No. And you are an idiot if you think it will happen. Why?

HOV lanes aren't only used by long distance commuters. You aren't going to get HOV/Autonomous lanes through SF. Sorry, After the 89 quake they tore down freeway infrastructure. 880 -> Richmond-San Rafael, yeah no.

Please provide me with a valid scenario where I can get from any of the bay area communities a commute over 30 miles where an autonomous vehicle would work. This is the norm. If you assume anything about future plans, you lose.

Hwy 85 was supposed to help south bay commutes, 2 months after opening, it was still faster to go 17 -> 280 -> Page Mill to get to Stanford from Santa Cruz. Oh yeah, that was still in the 90s.

I have no idea where you live, but, at least in the bay area, maybe you need a commute history lesson.

And as someone who lived near what is now Almaden and 85 before 85 went in (amd my parents were well aware of road plans, 85 was on the book for decades) the freeway turned upper middle class neighborhoods into problem areas when it went in.

"Autonomous lanes" much like light rail or Bart lines, have a community effect beyond the "making it easier to move people" effect.

"Aware of the conventional wisdom that robotic cars are about to cause an epochal “disruption,” automakers are eager to demonstrate that they are fully engaged. A result has been a drumbeat of announcements auguring the imminent arrival of robotic cars, almost as though they were the next generation of iPhones. The breathless statements are especially beguiling for members of the public without the engineering background required to understand the challenges that remain. In other words, most people."

I'm glad we have a "former technology reporter" to tell us this.

This is strangely similar to "flying cars". Those have been "just around the corner" for, I don't know, the last 30 years probably, if not more.

Related thread on the front page right now: "Simple questions to help reduce AI hype (smerity.com)"

https://news.ycombinator.com/item?id=12062294

Excerpt:

> ...good performance on a dataset also doesn't necessarily imply good performance in the real world. I'd wager that even trivial self-driving car models, given reasonable data, would be able to score in the high 90s if we decided on a reasonable notion of accuracy. The real stickler is the last few percent. When we make a mistake at high speed on a road, potentially under questionable conditions, the results can be disastrous. If the cost of a mistake is high, accuracy really doesn't mean anything, especially over a dataset that might not be representative in the first place.

    > When we make a mistake at
    > high speed on a road,
    > potentially under 
    > questionable conditions, the
    > results can be disastrous

Given that that's no less true for humans, why do we require 100% accuracy from cars, rather than just "on average better than humans"?

I found this article to be on the mark.

It's interesting that Google's project director estimates full autopilot will arrive between 3 and 30 years incrementally. Up to this point, most news and my friends seem to think it'll be here within 5 years.

When Google says "incrementally" they mean where it will arrive, not level of autonomy. They'll be driving in sunny California long before snowy Minnesota.

Right. We're on the same page.

The idea of having a "summer car" and. "winter car" is pretty common in Minnesota. Summer car being a nice one, perhaps fun to drive, that gets stored in the winter so that "road cancer" doesn't eat the body. A "winter car" has a good heater, a good battery, and good tires.

There is also the problem of analyzing why certain AI actions where taken when investigating accidents: https://www.technologyreview.com/s/601860/if-a-driverless-ca...

What surprises me is that last-mile and local transportation of non-living cargo hasn't gotten more attention. I can't wait to get cheap delivery from practically anywhere in town and nearly on demand.

I don't think Musk really understands the AI tech that he's managing. He thought Open AI would be the path to true AI, and he can't accept that a current flaw in Tesla's autopilot may be material to the company.

Musk has done a ton of great stuff. To be better, he could use someone who he trusts and is an expert in the field, like LeCunn, to simmer Musk's expectations of AI. Right now he seems brimming with confidence about AI to the point he (1) is claiming we're already living in a simulated world, (2) has started his own true AI venture, and (3) is ignoring the relevance of his own product's failures. I can only imagine this is because he is surrounded by Yes men.

I could be 100% wrong.

Reply on news.ycombinator.com