Hacker News Reader

Comment by majormunky | Hacker News Reader

It feels a bit odd for them to say, autopilot could not have been used in that area. "In a reconstruction using a similar Tesla, the agency found that Autosteer was not usable on that part of the road, meaning the Autopilot system couldn't have worked." Whose to say that Tesla didn't update something between the crash and when the NTSB tested autopilot in that area? I don't think they did do that, but are there any safeguards in place to detect that happening?

Its one thing to say that a car cannot do something because of a physical issue on the car, vs some software that can be updated or have bugs that haven't been caught yet.

View on HN

This point feels a little strained. There is a question: "Was autopilot in use?", and an experiment was performed to answer that question, which produced an informative result, which was reported. You seem to be attacking the linguistic structure of the statement by interpreting "coudn't" to mean some kind of existentially inviolate truth and not just... the result of the experiment.

I mean, come on. This is what people were saying within hours after the crash: autopilot as shipped simply doesn't have behavior consistent with this accident. It won't command the accelerations that would have been required on that tiny street and (as measured by the NTSB) won't even engage in that particular environment.

My car has never allowed AP to be engaged on roads without lines painted on it. There was no update that took that functionality away.

Yes this is true.

Basically every Tesla owner that saw the story knew right away that it was being sensationalized.

I was very surprised to see the LA times jump on it when they certainly have numerous employees that drive these vehicles.

It is sensationalized, but it could have a bug and not a designed behavior. This could have been a software bug or a rare combination with hardware or environment conditions.

Great, so the bug would need to be proven. A newspaper can't decide what happened the day of the crash with no evidence.

> My car has never allowed

That is an anecdotal result

> There was no update that took that functionality away

Maybe not explicitly, but we all know how software works

> There is a question: "Was autopilot in use?", and an experiment was performed to answer that question

Why an experiment? Isn't there a Black Box, like on aeroplanes, with logs to definitively answer the question?

From the article:

The report said the car's lithium-ion battery case was damaged, and in the ensuing fire the infotainment system's onboard storage device was destroyed.

The very next sentence of the article:

However, the "restraint control module," which stores data such as whether seatbelts were in use, how fast the vehicle was going, acceleration information, and airbag deployment, although damaged by fire was recovered and turned over to the NTSB's laboratory to evaluate.

The point is that Tesla can do over the air updates, which means (in theory!) that the experiment could be gamed because it would use a different version of the software.

My wife's truck has a lane guidance system. It will gladly keep you in a well marked lane. When there are no lines or bad lines it makes a loud ding and shows a big orange cancelation message on the dash display. This is a current year F150. I can only imagine Tesla has at least as good a system as Ford. As for updates, the infotainment system has a record of the last check-in and the last update, as well as the current systems version, similar to your phone or computer. With the number of Tesla hackers out there, it seems nearly impossible for the conspiracy you're suggesting to be a reality.

On the other hand wouldn’t Tesla be irresponsible if it didn’t manually review sections of road involved in a crash and perhaps override the settings if automated systems had made bad calls?

Can you imagine a second wreck on that same stretch of road and the accusations against Tesla for not doing the manual review envisioned above?

Have Tesla said definitively that they cannot override capabilities without updating the version number?

I think manual review might actually be the worst way to solve this problem, because you're relying on human labor to fix every single hiccup, which ruins the point of an automated system anyways. I think for sure they run the route through a bunch of simulations to tune the ML mode, but I seriously doubt they would go as low-level as using telemetry to disable certain features by hand. The best solution would be updating the model so that it makes the right calls on the right type of road - so that when it encounters a similar road, you don't need to wait for another fatal accident before kicking the review process into gear and making the necessary changes again.

tl;dr it doesn't scale and would be a logistical nightmare to maintain individual configurations for every road the system struggles on

Of course it would not solve the issue in a generic way. To me, the reasonable fix would be to immediately disable the feature in the specific area manually and then retrain the model or whatever it is they do.

>it seems nearly impossible for the conspiracy you're suggesting to be a reality.

What conspiracy? majormunky is saying Tesla might push out updates every so often, and that might effect whether autopilot can be used, and the investigators might not have realized that.

> Whose to say that Tesla didn't update something between the crash and when the NTSB tested autopilot in that area? I don't think they did do that, but are there any safeguards in place to detect that happening?

That one. To do so for a very specific location with no visible lane markers (implied by the article) to disable the lane guidance system seems very much like a stretch. "I don't think they would, but are there safeguards" implies the author of the comment is describing intent. I don't think it's a stretch to say that's a statement of conspiracy.

I guess you're right. But the same basic thing could happen without a conspiracy.

The article says that it couldn't have been in autopilot because autopilot requires lane markings for it to be activated and that area has no lane markings. It's nothing to do with software versions.

Is there any possibility for bugs caused by errors in the software or hardware?

Not a statistically meaningful possibility. If someone suffocates you don’t assume all the oxygen in the room just happened to randomly end up away from their face.

Do you have a source for that statistics? I have seen bugs in my Tesla but never had the oxygen issue happen. This is just anecdotal of course, but I suspect your comparison is off.

Sure, but the existence of individual bugs is very different that picking a very specific bug.

It’s the same as assuming that because someone regularly wins the lottery a specific ticket must have good odds. The space of all possible bugs is so vast, if the odds of every possible bug where as high as a lottery ticket then the software would be unusable. Some bugs are of course more likely than others, but assuming this one is more likely than average requires justification.

Elon also said that the vehicle didn’t even purchase full self driving capabilities.

Autopilot also includes the less-capable levels, not just FSD. For example, Enhanced Autopilot uses Autosteer and TACC which are both mentioned in the article.

> Elon also said

A man who uses his celebrity status to pump cryptocurrencies in order to enrich himself and the companies he owns has no moral standards and no credibility.

Even if he was pumping crypto, that doesn't mean he has no morals or that he would lie about technical details, particularly where he can be easily found out.

Hah, the more evil version of billionaire dodgy car salesman would have a database script to modify the purchase order to retroactively remove that item.

"Self-drive? Our records say he never bought it!".

To be clear, are you referring to the guy in the company selling cars claiming that the current models sold today will be fully autonomous with only software patches?

Might be worth a read on https://en.wikipedia.org/wiki/Ad_hominem

Doesn't matter what you think of the character, it has nothing to do with this instance.

It has everything to do with the character, since he provides no evidence for what he says, we have to trust him. Therefore, the first question we must ask is: is this person trustworthy?

They used the same version of software.

Local software != map data, necessarily.

Tesla has version of map data, if they did their job they used the same version for both.

Also, I don't think this feature actually uses map data.

It was reported that the versions are the same.

TL;DR let’s ignore the regulatory agency report, that confirms what literally everybody was saying, they are also being fooled. Tesla evil hurr durr.

also just because the driver was in the driver seat at the time of being on camera doesn't preclude some sort of stupid behavior like leaving the driver seat while on "autopilot"

All facts aside, people are still required to be in control of their cars and need to be legally able to intervene.

"But I used autopilot" is probably the most used excuse to avert fault away from the driver. Especially in the most broken insurance system on the planet where an accident like this leads to such a change in life that the driver is scared of the outcome.

my understanding is that whether autopilot is in use or not has no bearing on your insurance. as far as the insurance company is concerned, it's an accident while you're responsible for the operation of the car.

I’ve heard that the presence of Autopilot is actually viewed negatively by insurers.

This goes against intuition-wouldn’t a machine that behaves very reliably be better than a human driver - but my dad used to say that the insurance company’s calculation was only ever about how likely they thought they could win if they sued the other party. I don’t know if that’s true, but something rings true about it.

From an actuarial perspective, all forms of driver assistance contain the potential for moral hazard.

My insurance rates are the "same" as a similarly priced Audi or such.

As Teslas are safer over time, one would expect rates to gradually go down. Some repairs are a bit more expensive, so could be a bit of a balancing act.

> As Teslas are safer over time, one would expect rates to gradually go down.

In Germany it's currently the opposite due to the overly high amount of Teslas that are declared as total damage because the auditors that work for insurers usually don't know that their frames can be partially replaced.

Additionally there's no availability to repair a Tesla in practice at that point, but the given software lockout policy of Tesla is a different story.

At the same time, “he/she used autopilot while it is only drive assist” is a similarly overused excuse made by Tesla all around. If so many do so, maybe they did buy that car anticipating that feature, which makes it at least misleading advertisement.

The ironic result of this could be that insuring an Autopilot car gets more expensive than driving a normal car due to the high cost in investigating an accident after it happened.

Also, they shouldn't need to reconstruct the issue, there should be a black box recording all the events.

According to the article, the main storage aka black box was destroyed in the fire. They are trying to pull data from a secondary controller that was less damaged.

Reply on news.ycombinator.com