Hacker News Reader

Comment by cs702 | Hacker News Reader

Many of the comments here seem a bit... unfair to me, considering that these clips were handpicked.

I watched (and fast-forwarded) through a few of the original, full-length videos from which these clips were taken. The full-length videos show long drives (in some cases, hours long) almost entirely without human intervention, under conditions explicitly meant to be difficult for self-driving vehicles.

One thing I really liked seeing in the full-length videos is that FSD 9 is much better than previous efforts at requesting human intervention in advance, with plenty of time to react, when the software is confused by upcoming road situations. The handpicked clips are exceptions.

For BETA software, FSD 9 is doing remarkably well, in my view. I mean, it's clearly NOT yet ready for wide release, but it's much closer than all previous versions of Tesla FSD I've seen before, and acceptable for a closed-to-the-public Beta program.

View on HN

The fact that it 'only' requires human intervention so rarely is still incredibly dangerous. You can't ask a human to have complete focus for hours on end when they're not making any inputs, and then require them to intervene at a moment's notice. That's not how humans work.

Also, the fact that they're distributing safety critical software to the public as a 'beta' is just insanity. How many more people need to die as a result of Autopilot?

> You can't ask a human to have complete focus for hours on end when they're not making any inputs, and then require them to intervene at a moment's notice. That's not how humans work.

I agree. Everyone agrees. That's why FSD Beta 9 is closed to the public. My understanding is that only a few thousand approved drivers can use it.

> Also, the fact that they're distributing safety critical software to the public as a 'beta' is just insanity. How many more people need to die as a result of Autopilot?

FSD 9 isn't being "distributed to the public." It's a closed Beta. Please don't attack a straw-man.

What are the qualifications of the people selected to participate in the Beta? Do they have any particular experience or receive special training that would set them apart from the general public?

Yes, even if it's a limited release unless these people are professional drivers or Tesla employees these are just people from the public.

What do the people sign that share the roads with those people? Or are they just colateral?

A closed beta needs to happen on private roads. If it happens on public roads it is not a closed beta.

> What do the people sign that share the roads with those people?

Some of the videos show drivers having to agree to take full responsibility for all driving.

I mean even if that’s true so-called beta software shouldn’t be on public roads.

That disn't answer my question. A private company should test their product on private roads.

Nobody signed a waiver to drive next to Joe Blow hauling a mattress in his 82 VW Golf, yet he can do it every day. I don't see how FSD is any different than driving a car with aftermarket modifications. It's not like any of that stuff has to go through NHTSA certification to be road legal.

It’s a public beta, as far as I can tell it’s available to anyone with a Tesla with the FSD package.

This is simply untrue.

Let's clarify. It's public in the sense that a limited number of customers are driving with this on public roads.

Also, Elon keeps promising "the button" which would allow any Tesla FSD "owner" to get the software.

A number of people shelled out $10k for this software that they do not yet have. So there is a real possibility that it may soon be made available to everyone.

So am I wrong in reading that if I purchase a Tesla with FSD hardware, apply for the beta, I can get it?

It literally doesn't matter how well it does in 90% of situations when the other 10% can injure or kill people in relatively basic scenarios like the Tweets presented. I mean the car almost ran into a concrete pillar like it wasn't even there.

> For BETA software, FSD 9 is doing remarkably well

If this was a React website, that'd be great. But it's a production $40,000 multi ton automobile.

People will die today because a driver was drunk/distracted/suicidal/rode raged or had a medical problem.

Are you OK with that or do you think we should attempt to fix that with software? If you do think you should attempt to fix that, do you understand that software engineering is a iterative process? It gets safer overtime.

> People will die today because a driver was drunk/distracted/suicidal/rode raged or had a medical problem.

That's a straw man argument. Mandatory breathalyzer tests before every engine start and yearly medical and psychological check-ups would solve a lot of these problems as well.

> It gets safer overtime.

There is no evidence of this happening yet. It's still trying to drive people into concrete pillars. Remember that you only need to drive into a pillar once to die. Passing it 10,000 times before doesn't help when an update suddenly decides that the pillar is now a lane.

I'd be totally fine with Tesla testing this with trained drivers in suitable areas but giving this into the hands of amateurs is pure "move fast and break things", with "break" meaning ending lives.

The worst thing is: all of these videos feature sunny weather. How bad would it be in the rain? Ever tried to take picture in the dark with even light rain? Cameras are very bad at this.

Local maximums exist. If ML models were guaranteed to improve over time we would have no more problems to solve.

What are you proposing? If someone is overweight with high blood pressure and a higher risk for a heart attack we should take away their license?

These drivers are vetted and their usage is monitored, many are employees. In the long run automated driving will save many more lives than it takes. Just like automated elevators.

https://www.npr.org/2015/07/31/427990392/remembering-when-dr....

> What are you proposing? If someone is overweight with high blood pressure and a higher risk for a heart attack we should take away their license?

This does not seem like an important nit to pick, so I'm wondering why you're picking it. The number of people killed or injured because a driver had a heart attack at the wheel is vanishingly small compared to drunk or distracted driving.

So let's reframe your question as you should have asked it:

"If someone is driving drunk or distracted we should take away their license?"

Yes. Absolutely.

Clearly it's not that simple, otherwise we'd have done it already.

We… already have? Drunk driving is grounds for license suspension, heavy fines and jail time in (I’m pretty sure) every US state. We may still have drunk drivers, but it seems odd to suggest that we aren’t doing anything about it. Cops, judges and juries take it super seriously. If you get caught drunk driving in America: Do Not Pass Go, Do Not Collect $200.

>Mandatory breathalyzer tests before every engine start

We don't do that today and we are still allowing people to drive on the road today.

Exactly. That's why Tesla's approach is bad. There is no rush to let this technology out on the streets before testing it rigorously for flaws. But Tesla promised FSD for all current models with FSD hardware, so they have to move or they will potentially be on the hook for fraud. I call this irresponsible and dangerous.

Disagree, Tesla approach is the right approach. You want to test the technology as early as possible in real situation as much as possible.

> It's still trying to drive people into concrete pillars.

That in itself doesn't constitute evidence that it isn't getting safer. After all humans occasionally drive into concrete pillars too.

What matters is that the distance traveled without incident goes up in a super-linear fashion over time. It would be nice if someone had hard numbers on that.

I’d say if we compare Tesla self driving numbers with human drivers driving similar cars, and count every disengagement as an accident, it would be a fair comparison. Tesla wouldn’t come out looking very good with those numbers though.

These rules seem quite stacked, we could also count every human crossing a red light or texting while driving as an accident.

But that's besides the point. I was arguing about the trend line. It may seem silly now, but if the error rate is roughly a straight line on a logarithmic graph then we'll get there soon enough. If it isn't then tesla at very least would have to admit they have a problem and explain how they'll get out of it.

>> It gets safer overtime.

> There is no evidence of this happening yet.

With aircraft any crash or near-crash is investigated and then the software and procedures for all similar aircraft are updated. Over the decades this has resulted in a huge reduction in the rate of accidents. Isn't it likely that the same thing will happen with self-driving cars?

Eliminating drinking would get rid of a huge number of automobile accidents. Sober humans are incredible drivers and we don’t have self driving cars that are better than even poor humans, as long as they aren’t impaired.

> ...doesn't matter how well it does in 90% of situations...

Based on the full-length videos, I'd say it's more like 99.9% or even 99.99% for FSD 9.

Not sure if 99.99% refers to distance or time, but either way I'm not sure it's a good metric since it only takes a second or two to kill someone.

The same applies to easily distracted human drivers. 1.3 million people die in car accidents every year.

Apples to apples would be comparing miles driven on FSD autopilot to miles driven by human drivers in similar makes of vehicle on the same types of roads in the same weather conditions.

Human drivers don't steer towards concrete pillars every time they drive down the street.

Only when they’re drunk, or looking at a mobile phone, or having a heated argument!

The new software seen here hasn’t been released so there’s no data yet. We cannot possibly make an assessment based on a few hand-picked videos - a quick YouTube search will give you thousands of human driver examples that are even more dumb.

That might be acceptable for minor failures but not for catastrophic failure.

The rate of driving fatalities is ~1 per 60 million miles driven. At an average speed of 60 mph that would be ~1 per million hours. Even if we were to assume there would be only one fatality-inducing error per hour, 99.99% success rate would make FSD 9 100x more dangerous than the average driver. I do not remember the exact numbers, but that would make FSD something like 1000x more dangerous than the classic example of a deathtrap, the Ford Pinto.

That's your estimation but we'd need to wait with actual statistics

Why is Tesla going live with this (sure, call it private beta, but on public roads that makes no sense) without already disclosing all of those numbers to the public? If they're being this careless, bad estimates are all we have.

We have actual statistics on how well humans do with Tesla ADAS and it’s remarkably well. The accident rate numbers are hard to compare apples-to-apples since miles driven with all ADAS features active are different from miles driven without the full ADAS active, but the numbers show that the software is absolutely not “1000x more dangerous”.

> In the 1st quarter [2021], we registered one accident for every 4.19 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.05 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 978 thousand 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

> one accident for every 4.19 million miles driven

Not really a great metric. We should be looking at disengagements per MM.

If you're evaluating the system as a L5 autonomous driving system. Not if you're evaluating it as a safety assistance feature.

If it’s a safety assistance feature, it’s not full self driving is it? My crosstrek comes with plenty of safety assistance features too.

That’s true enough, but I’m not going to argue semantics. Tesla is explicitly clear in their product communication that what they’re selling as «full self driving» comes with the massive asterix of «someday» and is not currently an L5 autonomous driving system.

The part of the discussion that concerns their marketing is uninteresting and overdone; someone else can discuss that part.

The rub in Tesla's safety report is it compares Teslas to all makes of vehicles, including motorcycles and trucks, and all costs. A fair comparison would be to newer vehicles in a similar price range.

Their autopilot stat is also misleading because AP will most likely be engaged only in ideal weather conditions on divided roads.

Yes, this is the retort every time someone posts these numbers, which is why I said they are not apples to apples comparisons. But in the context of the OP comment that Tesla ADAS software makes driving 1000x more dangerous it is clear evidence to the contrary.

Mostly the Tesla numbers are comparing accident rates of just Teslas. You can ignore average accident rate across all vehicles if you like.

If would be truly great if the NHTSA could publish accident rates broken down is various ways like by road type and weather conditions. I’ve never been able to find this data.

I made no such claim that the actual success rate is 99.99%. I was replying to a person who supposed a 99.99% success rate based on their observation and I was pointing out that even an absurdly generous interpretation of a 99.99% success rate would constitute the most dangerously defective product in the car industry in half a century by multiple orders of magnitude.

The point being that the failure rates society accepts for cars are actually pretty amazingly low and human drivers are surprisingly good. These things are measured with 6-7 9s as a non-negotiable minimum, so positing hypotheticals at the 99-99.99% is highly non-representative of what is actually acceptable.

Anybody discussing things at that level is making a pretty fundamental category error and any conclusion that anyone might draw from those thought experiments are largely invalid due to the multiple order of magnitude difference between such a thought experiment and reality.

The Tesla safety numbers are pure marketing. They make no attempt to control for obvious problems like the kind of miles. Notice how they say autopilot engaged not autopilot equipped when talking about autopilot miles. I bet gm has great numbers for cruise control engaged accident rate too but Id hardly call it a safety device, more like a convenience.

Given the telemetry the car is equipped with they absolutely have the data to control for some of these factors yet they make no attempt to do so. Why is that? My take is that after doing so autopilot doesn't look so good.

No, it's not. In the full length video, there were 3 manual interventions in less than 30 minutes of drive time. If you're counting interventions per second, then sure that's 99.9%, but if you're counting safety critical interventions that would have led to an accident per mile then that rate is abysmal.

> Based on the full-length videos, I'd say it's more like 99.9%

You assumed numbers are time. Now assume numbers are manoeuvres.

In my opinion, a self-driving car which drives me smoothly for 6 hours then drives straight into a concrete pillar without warning isn't doing "remarkably well". That should be enough to get it pulled.

FSD occasionally requesting driver to take over in genuinely difficult situations would be completely fine.

The videos in the Twitter feed are nothing like that. The car makes potentially catastrophic blunders, like driving straight into a concrete pylon, with 100% confidence.

You know these are extract from multiple hours of video and it's a closed beta?

"The airplane rarely explodes! We're down to one explosion every 500 hours!"

a closed beta of a system that is "orders of magnitude better, completely reimagined"... videos like this certainly don't seem to show revolutionary growth in the usability of FSD, indeed "more of the same".

It doesn't matter how many hours of video there is, all it takes is hitting one pole to dramatically impact your life or the lives of others.

As a pedestrian and someone who shares the road with drivers who use FSD, I don't get to opt out of this "closed beta", and I certainly wouldn't care about how many hours of quality content it has on YouTube if it caused a car to hit me or drove me off the road.

I'm with you on the second point strongly.

But IMHO it's not full self driving if it requests the driver to take over even once.

If there's an insane storm or something then it's ok for FSD to know it should disable and then you have to drive 100% control. The middle ground is more like assisted driving which doesn't seem safe according to most HN comments.

Supposedly there are a "few thousand" FSD beta testers, and only a small fraction of them are videoing their drives and uploading them to YouTube. Beta 9 has existed for 2 days. This puts a pretty high lower bound on the serious error rate.

The consensus is that there far fewer than a few thousand

More than 2 thousands Tesla employees and at least dozens of non employees (source: teslamotorsclub.com)

For level 5 (which Elon promised), FSD has to go tens of thousands of miles without a collision-avoiding intervention. Right now it can't even do 10 miles.

Cherry picking is completely fine because with the limited number of beta users, even a few incidents is enough to show that FSD isn't nearly ready for level 5.

And it's pretty easy to make a car that can self drive on simple roads without a tonne of traffic, so being able to do that without intervention isn't much of an accomplishment.

People who drink and drive may very well be perfect sober drivers 99.9% of the time but that doesn't excuse the .1% of the time that they're running into things.

Also, this beta isn't "closed-to-the-public". The "public" is an active and unwilling participant in it.

If this happened all in one month of constant driving, I'd say it isn't fit even for limited closed testing in public traffic. It should be back at the closed circuit with inflatable cars. If it was cut down from just one or a few days of driving that's horrifying.

At least for me it's because these highlighted errors are so egregious and so obvious for humans. Don't swerve into the giant concrete polls.

The 99% of 'good' doesn't matter if you keep driving into giant barriers.

Reply on news.ycombinator.com