Elon Musk Rolled Out Autopilot Despite Engineers' Safety Concerns, Says Report (theverge.com)

An anonymous reader quotes a report from The Verge: When Elon Musk announced last fall that all of Tesla's cars would be capable of "full autonomy," engineers who were working on the suite of self-driving features, known as Autopilot, did not believe the system was ready to safely control a car, according to the Wall Street Journal. The WSJ report sheds more light on the tension that exists between the Autopilot team and Musk. CNN previously reported in July that Musk "brushed aside certain concerns as negligible compared to Autopilot's overall lifesaving potential," and that employees who worked on Autopilot "struggled" to make the same reconciliation.

A major cause of this conflict has apparently been the way Musk chose to market Autopilot. The decision to refer to Autopilot as a "full self-driving" solution -- language that makes multiple appearances on the company's website, especially during the process of ordering a car -- was the spark for multiple departures, including Sterling Anderson, who was in charge of the Autopilot team during last year's announcement. Anderson left the company two months later, and was hit with a lawsuit from Tesla that alleged breach of contract, employee poaching, and theft of data related to Autopilot, though the suit was eventually settled. A year before that, a lead engineer warned the company that Autopilot wasn't ready to be released shortly before the original rollout. Evan Nakano, the senior system design and architecture engineer at the time, wrote that development of Autopilot was based on "reckless decision making that has potentially put customer lives at risk," according to documents obtained by the WSJ.

Comments

Re:There's always one or two voices....

[JohnFen]

anyone who's tried to execute a change or deliver an outcome will always find one or two dissenting voices in any organisation of scale.

Absolutely true. And it's equally true that it's foolish to not take those dissenting opinions very seriously (even if, after careful consideration, they don't change your plans).

In any organization, there is a strong "rah-rah" tendency, and people tend to suppress their own doubts. Nobody wants to be a wet blanket or potentially risk their career by not seeming to be a "team player". So the voices of those who point out problems need to be listened to much more carefully than the voices of those who say "everything's great".

Someone always has to make the tough call

[RhettLivingston]

I've been in engineering organizations releasing new products that had life saving or threatening potential. It is always an agonizing, scary hard call as to when you've passed the threshold of risk.

There is a bell curve with a peak. You rarely hit the peak. If you make the call too late, you cost the lives of those you might have saved - too soon, you cost lives of those who might have saved themselves.

Even if you hit the peak perfectly, you'll always be able to truthfully argue that some people are being saved who would have died and some are dying who would have lived. The peak is a point of balance between the two - not a perfect elimination.

I can remember many times hating my bosses when they released a product that I didn't feel was ready. As an engineer, I have to be over-focused on the problems and stand no chance of seeing when I am perfectly perched on that probability peak. They had to pry the projects from my hands to get them out the door. I actually begged in tears once. But, in retrospect, I can't think of any case where my bosses weren't right in releasing the product that I was concerned about releasing.

What we need to force progress is for attorneys to get smart and start figuring out how to file more effective suits for lack of progress toward autonomy. How many are dying today because we don't have it? We need to focus hard on that.

Re:Someone always has to make the tough call

[RhettLivingston]

Having the license in this country is often career-ending, much like having a PhD. It can make it very difficult to get a job. I've been in corporations that had thousands of engineers and never met anyone I knew to have it. I think they tend to be in certain structural and mechanical, and most civil and architectural engineering areas. The electrical, aeronautical, and computer engineering professions have much less of this.

Regardless, there is no such thing as a vehicle on the road today that does not make some safety compromise. Not one single vehicle uses the best-known safety mechanism for every single aspect of the car. No one could buy it if they did, and it wouldn't meet other necessary criteria if every compromise was made in the safety direction. Our government often has to force the matter by making regulations like the ones coming down the pipe soon to require all vehicles to have automatic braking technology. This is tech that has been available for a while, but many engineers must be signing off on vehicles that are killing people, otherwise, the government wouldn't have to be stepping in.

Everything engineered makes these compromises. For example, every building might be designed to handle a 500-year quake, but what happens if a 5,000-year quake comes along?

Airbags are an interesting example. Even the best airbag systems kill some people who would not have died without airbags. But they save many more that would have. So, you accept the compromise. Many years ago, seatbelts did the same and still do. Yet, we have them, and are even required by law in most places to wear them.

With the autonomous vehicle question, it is ready to deploy when it will save more people than it will kill when measured versus human drivers (all of them, not just the competent ones). To wait any longer would be killing those people that it might have saved. Of course, determining when that point is is a near impossibility. The hard call will either be made or the vehicles will never be made because the engineers will never be able to say with any product that it is not flawed in some situation - often in which it is being misused by the consumer.

Realistically, we do wait longer than the point of net balance because the public does not understand statistically-based decisions very well. When it is your family member that died because the tech failed you want to blame the tech without looking at the whole picture. We often don't even know when our family member died because the tech that could have saved them was held back because it was being over-engineered.

Often, these hard decisions are the reason for regulation - not to protect the public but to allow the companies protection in deploying something that a big picture organization like the government has determined will be a net benefit to the public while being a detriment to some individuals. The engineers then have the excuse of having met the regulation. It seems to work better with our minds.

Absent specific regulations and tests to target (which is the ideal situation in a free society), the business leaders are usually the ones who make the tough calls.

Re:Not surprised

[Rei]

It's a lot simpler than that:

This article is bullshit.

Sorry to be so blunt, but it's journalistic malpractice. The author is confusing Enhanced Autopilot (EAP) with Full Self-Driving (FSD). To be clear:

  * Some safety features related to autopilot, such as automatic braking and the like, are available to everyone for free.

  * EAP is an optional add-on available today ($5k on the Model 3 if purchased at the time of buying the vehicle, $6k as an over-the-air upgrade) which provides lane-following (requires hands on the wheel and driver attention) and driverless summon features (very low speed, "back out of / into a tight space / drive down the parking lot" stuff without a driver in the car). More to the point, there are two entirely different versions that have existed over the year. AP 1.0 was used on earlier vehicles, based on software and hardware from Mobile Eye. Tesla and Mobile Eye split in a contract dispute. Mobile Eye claims that Tesla wasn't using their hardware right. Tesla says that Mobile Eye found out that Tesla was working on an in-house Autopilot system and demanded that they stop as a precondition to get to continue to use their hardware. Mobile Eye says they knew about Tesla's internal work but didn't feel threatened by it. Regardless, Tesla was forced to switch to their internal version, AP 2.0, which was a step backward. AP 2 is just now catching up to the features of AP1.

  * FSD is Tesla's current goal, where the vehicle can drive itself without you having to have your hands on the wheel or paying constant attention. You cannot use FSD, even if you buy it. It costs $3k on the Model 3 ($4k as an over-the-air upgrade later). The article is talking about FSD being rolled out before engineers think it's ready. To reiterate: you can only buy FSD right now, you can't use it until it's ready. Tesla apparently tried to clarify this for the author:

When reached for comment, a Tesla spokesperson referred back to the company website, where a disclaimer for Autopilot reads that “self-driving functionality is dependent upon extensive software validation and regulatory approval.”

The author apparently nonetheless still failed to understand what that means. You Cannot Use FSD. Period. If engineers are complaining about FSD being rolled out too soon, they're complaining about Tesla selling something that its drivers aren't going to be able to use for too long of a period of time. And you know what, I fully agree with the engineers in that regard - I think it's wrong of Tesla to sell something that there's a big question as to whether they'll be able to get it working reliably enough or pass the serious regulatory barriers in its way.

But if engineers are complaining about FSD, then it's not complaints about EAP. Because the two are very distinct things. EAP isn't perfect, don't get me wrong - and the 1.0 / 2.0 switch was a big setback (they still don't use all of the cameras on the vehicle). But it also pesters drivers enough if they show signs of not paying attention to the road (e.g. not holding onto the wheel) in order to overcome its imperfections (the level of pestering was significantly increased after AP1's fatal accident, in which the driver was apparently watching movies during most of his trip).

Re:Deep neural nets will never give us full autono

[BasilBrush]

Neural Nets are very specifically NOT rule based. They are trained.
GOFAI was pretty much a phrase invented to label stuff that IS NOT the neural net approach.
Autonomous vehicles do not need AGI. It's very much a single domain system. You don't need your autonomous car to be able to diagnose diseases for example.

Re:Not surprised

[Rei]

They implied that Tesla is currently having people drive something that its engineers deem unsafe. This is simply not the case at all. If the engineers were complaining about selling FSD, they're not complaining about anything that consumers are actually driving.

Everyone who buys FSD does so on their assessment of how likely they think it is that Tesla will actually deliver. There is zero confusion among anyone who buys it about the fact that they can't use it right away; the option always includes the "you can't use this until it's finished and legally approved" disclaimer next to it. It all comes down to how optimistic or pessimistic you are about the technology. I'm a pessimist, and will not be buying it. Some of Tesla's engineers working on it are apparently also pessimists. I'm not surprised. It's a crazy-hard task, and very different from human-supervised autosteer / EAP.