Comment by frgtpsswrdlame | Hacker News Reader

Comment by frgtpsswrdlame | Hacker News Reader

Parent Full thread

And if you'd like a critical take on that report:

But exactly how NHTSA reached that conclusion has been a mystery. It’s also the subject of an ongoing lawsuit between the agency and a group of researchers aiming to obtain NHTSA’s data so it can try to replicate the findings.

That may be an impossible task.

On Friday, in response to a Freedom of Information Act request filed in March by Jalopnik, NHTSA said it was unable to provide most of the materials and data used in the investigation, citing a “confidentiality determination” it granted Tesla in July. (It’s not uncommon for companies to submit requests to NHTSA for certain information to be treated as confidential. From there, NHTSA then chooses whether or not to grant the request.)

In particular, the agency said, NHTSA is withholding in full an “excel spreadsheet Tesla submitted that contains production data from disclosure because the spreadsheet contains information related to trade secrets and commercial or financial information.”

[...]

Autosteer, however, is relatively unique to Tesla. That’s what makes singling out Autosteer as the source of a 40 percent drop so curious. Forward collision warning and automatic emergency braking were introduced just months before the introduction of Autosteer in October 2015. A previous IIHS study shows that both the collision warning and auto emergency braking can deliver a similar reduction in crashes.

https://jalopnik.com/feds-cant-say-why-they-claim-teslas-aut...

Let's take the most conservative interpretation of that result and say that Tesla's collision warning and automatic emergency braking are responsible for the entire 40% reduction in accidents since that is similar for that feature set in other cars. That means that the introduction of autosteering at worst had no effect on the accident rate. That is a pretty big difference than "This "kind of autonomous driving" is absolutely deadly, by definition."

EDIT: So let me add that yes, this report probably isn't as pro-autopilot as some people are claiming, but I would still consider it positive for Tesla considering it is the only publicly available report on the subject and it shows that autopilot is at least no worse than a human driver.

What bothers me is that Tesla has this data. Their cars are instrumented out the wazoo and uploading everything to the mothership. It would be trivial for them to discount cars that encountered AEB events, didn't use Autosteer etc from the dataset. The fact that even with that data they still feed us BS statistics about stuff like autopilot highway miles vs overall miles really worries me

>Let's take the most conservative interpretation...

I disagree that your interpretation is the most conservative interpretation. The most conservative interpretation is that we know basically nothing about the safety of Tesla's autopilot. There are so many unaccounted variables here - the presence of AEB and collision detection, total utilization of TACC vs TACC + Autosteer, utilization of Autosteer over time (for example: are people conservative with it at first and then gradually increase their usage into riskier driving situations?), non-airbag accidents, the constantly updating nature of Autopilot systems (for example what if an update actually makes the vehicle less safe?), whether people have more accidents initially driving any EV due to being unfamiliar with the instant torque of EVs and so on and so on. Mix in proprietary data which no one is willing to share and I think the only thing we can say is that we don't know.

Also:

>it shows that autopilot is at least no worse than a human driver.

It does not show that!

There are human driven miles in both buckets, for all we know the second bucket could have more human driven miles than the first bucket which is why the accident rate is lower. Tesla has never proven that it's Autopilot system is safer than or as safe as a human driver.

> Let's... say that Tesla's collision warning and automatic emergency braking are responsible for the entire 40% reduction in accidents... That means that the introduction of autosteering at worst had no effect on the accident rate.

That doesn't follow logically at all. Collision avoidance and autosteer were introduced within months of one another. One could be reducing accidents while the other increases them. The most conservative interpretation is that the study may not have properly isolated these variables and is misattributing the safety gain of collision avoidance to autosteering.

You threw an ellipsis over a qualifying phrase in my statement and then responded as if I didn't say it. Your second ellipsis removed "since that is similar for that feature set in other cars". My conclusion rests on the assumption that those features on a Tesla are similar to those features on other cars. The only way for autopilot to be less safe than a human (taking the numbers from that study as true) is for those safety features to be substantially more effective on a Tesla. That is certainly possible, but it seems likes a weird argument to make since it rests on Tesla's technology being simultaneously industry leading and yet not nearly as effective as they claim.

Nah, I was clear you were characterizing that as the most conservative interpretation of the study. Your assumption is logically flawed. To wit:

> The only way for autopilot to be less safe than a human (taking the numbers from that study as true) is for those safety features to be substantially more effective on a Tesla.

You lumped the two types of features together. There's your logical flaw. The most conservative interpretation is that, yes, Tesla collision avoidance could be great while autosteering is "not nearly as effective as they claim."

These features are all interconnected and work together. At the most basic level they are about identifying obstacles and avoiding them. For an accident to happen, you usually need a failure of more than one of these systems.

For example the most recent fatality was directly caused by an autosteering error directing the car towards a median. However the accident could have been prevented if the automatic emergency breaking triggered before the collision. Considering it didn't, I have a hard time imagining that the automatic breaking has surpassed the rest of the industry to a large enough extent that could make up for the autosteering being as dangerous as you are suggesting.

It is certainly possible, it just seems unlikely that Tesla threads the needle to be extremely competent at one feature while also being incompetent enough to bungle a very related feature. That scenario relies on a lot more assumptions than the scenario I proposed which is why I described my original scenario as being the most conservative.

You're just doubling down on very poor reasoning. Collision avoidance can fail 50% of the time and still result in a massive 50% reduction in accidents. Whereas if autosteer fails 50% of the time while seducing drivers into paying less attention, it's going to increase the accident rate. It is not the underlying technologies that matters here, it's the fundamentally different use cases and compares.

> I have a hard time imagining that the automatic breaking has surpassed the rest of the industry to a large enough extent that could make up for the autosteering being as dangerous as you are suggesting.

This is another comprehension error. You claim was "autosteering at worst had no effect on the accident rate." Thus any increase in the accident rate caused by autosteering, not merely a "large" one, would invalidate your claim.

Reply on news.ycombinator.com