Hacker News Reader

Comment by Tomis02 | Hacker News Reader

> The Tesla FSD system on the other side aims to be a generic driving AI. ... can it ever get so reliable, that the driver can completely hand over to car ever?

Using the current machine learning algorithms? Not likely. Imagine releasing a new airplane and telling the pilot - "this airplane flies well in 99% of scenarios, but 1% of the time it refuses to respond to pilot commands and kills everyone on board. We don't know why it happens but we'll keep training the model". That's the state we're currently at with "AI"; as someone put it, it's alchemy. If it works, we don't know why. If it doesn't, we don't know why - but we tune the parameters until it seems to work. We cannot prove it will work for every possible input. ... Not exactly a scientific approach by any means. Such self-driving cars might be accepted in USA but I imagine they'd quickly get banned in Europe after a few random fatalities.

View on HN

In the 4th quarter, we recorded one crash for every 4.31 million miles driven in which drivers were using Autopilot technology (Autosteer and active safety features). For drivers who were not using Autopilot technology (no Autosteer and active safety features), we recorded one crash for every 1.59 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 484,000 miles.

https://www.tesla.com/VehicleSafetyReport

there are a few things to unpack here

1) Autosteer and active safety features are not "autopilot", its adaptive cruise control.

2) if they think something bad is about to happen, it beeps and hands over control, thus its the drivers fault, as the system wasn't engaged at the time of the crash

3) its not their self driving tech.

exactly. Steer assisting and active safety features are standard features for every upper class car you buy today.

1) Autosteer + active safety features is exactly what autopilot is. What do you think it is?

2) Tesla count 5 seconds after a disengagement and count it as a fault of autopilot if a crash occurs.

3) It's part of their self driving tech, but it's different to FSD - yes.

> Autosteer + active safety features is exactly what autopilot is.

If you ask a normal person on the street what "autopilot" means in a car, you'll get a range of answers. But the consistent opinion is that autopilot means that you don't need to concentrate. Most will say that its automatic driving, and will drive for you.

But autosteer requires concentration. Its just sparkling lane assist.

> it's different to FSD - yes.

Which is the point. Its marketing fluff. Autopilot isn't really that, its adaptive cruise control with lane assist. Thats the thing that rankles. All of this stupidity, injury, and noise comes from a marketing decision. One designed to cover up that the CEO over promised and wildly under delivered _yet again_.

Autopilot in a plane is cruise control and flight heading (lane assist). Is that not the same thing?

“FSD” and “Autopilot” are both SAE level 2 systems, which by definition, never take over driving responsibility from the driver.

They may have the worlds best driver assist features, but they have yet to produce a product that drives itself in any capacity.

https://www.motorweek.org/images/SAE_Chart.png

SAE is good for classification but not for evaluating performance. By definition autosteer isn't "driving", but it does effectively drive for you at the level you'd expect, and it can take you 300 miles on a highway without disengagements and the $12k package can even overtake cars slower than you.

You’re only looking at one metric of “performance”: the number of conditions it operates in.

The SAE is looking at a different performance hurdle between levels 2 and 3: reliability.

That's my point. Mercedes might be able to operate unsupervised on a some predefined map of highways while under a set of conditions and at a certain speed, but that doesn't mean it's a good product that solves real problems if it can't take over driving unsupervised on a regular road trip where you're driving 70+mph.

"Tesla's Favorite Autopilot Safety Stat Just Doesn't Hold Up": https://www.wired.com/story/tesla-autopilot-safety-statistic...

There isn't actually any substance to that article. Why don't the numbers hold up?

It's covered in the article.

"NHTSA's Flawed Autopilot Safety Study Unmasked (2019) - The safety regulator's claim that Autopilot reduces crashes by 40% was based on flawed data and analysis, which it attempted to keep secret.": https://www.thedrive.com/tech/26455/nhtsas-flawed-autopilot-...

"Tesla Autopilot Safety Stats Said Imbued With Statistical Fallacies, Interpret Cautiously": https://www.forbes.com/sites/lanceeliot/2019/06/09/tesla-aut...

"In 2017, the feds said Tesla Autopilot cut crashes 40%—that was bogus": https://arstechnica.com/cars/2019/02/in-2017-the-feds-said-t...

These are refuting NHSTA numbers, not Tesla’s numbers.

According to the article it seems to be:

1. NHTSA has reiterated that its data came from Tesla, and has not been verified by an independent party (as it noted in a footnote in the report).

2. Second, it says its investigators did not consider whether the driver was using Autopilot at the time of each crash. (Reminder: Drivers are only supposed to use Autopilot in very specific contexts.)

3. And third, airbag deployments are an inexact proxy for crashes.

... which all sound like really flimsy reasons to conclude "doesn't stack up." A more honest summary based on those would be something like "hasn't been verified by a 3rd party yet."

It's incredible how openly the media makes things up about Tesla, likely just to generate page impressions.

That article is pretty old. More recent analysis is much more specific as to how the data is flawed, see tweet and associated paper:

https://twitter.com/Tweetermeyer/status/1488673180403191808

The primary issue is that the data doesn't adjust for road classification in any way. Highway driving, where autopilot is used, has considerably fewer crashes per mile than city driving. Tesla compared Autopilot's rates against all driving rather than just highway driving which would be the relevant metric.

That adjustment alone almost completely eliminates any safety advantage of Autopilot before you get into any of the other adjustments like age.

So being safer on highways should be completely disregarded? Okay so this Mercedes news should also be completely disregarded

Tesla's data doesn't even show that they are safer on highways, I don't think you understand the original criticism.

It's like this:

Tesla: Look how great we are, our apple is redder than their orange.

Critic: It doesn't make sense to compare the colors of those things in that way. Here's why ...

You: So being a redder apple should now be disregarded?

Tesla’s data shows autopilot is safer (crashes less) than not on autopilot. The only question mark I’ve seen in this discussion is on NHTSAs data regarding Tesla.

Could you explain the flaw in Tesla data? How is crashes/mile not a good proxy for safety?

Tesla count any crash above 12mph. Airbag deployment isn’t required.

That's an apples to oranges comparison, due to differences in car quality and situations driven.

If you only engage Autopilot in the easiest situations to drive in, it might look safer artificially because of that.

Comparing a Tesla to all vehicles means old cars with less safety features enter the picture, skewing the statistics again.

>In the 4th quarter, we recorded one crash for every 4.31 million miles driven in which drivers were using Autopilot

But this is not FSD , is a combo of human + some driver assistance, come back when Elon discloses how many times the human was forced to intervene to prevent a crash.

I am wondering why Tesla fans think this stats are proving that so called FSD is safe , don't you understand stats? did PR tricked you,? this proves that soem assistant like one keeping safe distance between cars is safer, make all the data public and let us compute the real stats please.

Reply on news.ycombinator.com