FSD v12.x (end to end AI)

[AlanSubie4Life]

"The miles driven by Teslas on FSD Beta have resulted in fewer accidents per million than the miles driven by Teslas without active safety features." That's a fact, given the data.

Sure.

we can absolutely say that Tesla drivers with FSD Beta and Autopilot are both safer than Tesla drivers without Autopilot.

But this does not follow, in any sense of the word “safe.”

As has been pointed out, it is also safer to drive with your headlights off (when this was possible).

Normally when talking about safer, we want to know whether or not you are less likely to be in an accident or not, in normal everyday driving. You want to compare yourself with access to FSD, for all your driving, vs. yourself, without access to FSD.

We don’t know. It is probably possible for Tesla to make a decent comparison.

You can argue that you think the data is fabricated.

I don’t think it is fabricated.

You can argue that the additional safety is caused by other factors that are spuriously correlated with FSD Beta.

I think that definitely there are confounding factors. They can probably be teased out with access to the data, but we don’t have that.

 

[kpanda17]

Do we have drive results for 12.3.2.1?

 

[kpanda17]

guys, wait, Tesla drivers are a unique group in the larger population, we are not considered normal (think higher education, experience, skill, safer, etc)
now think about what we see driving on a daily basis, I call it the ME!!! selfish driver, out of traffic lines, slowing, speeding, no seeing, etc
that group of dangerous drivers are dangerous and def a good size of the population or the auto body shops would not be so busy
if FSD is applied to us, the unique group (us), it might not be safer than we are vs the current code version, but overtime FSD will be safer
if FSD is applied to these dangerous group of drivers from the total population, the result will be safer transport for them and all of those around them
we have to agree with this analysis, just makes sense
FSD is looking very promising

 

[lzolman]

Ok ok, but what about Rust? C++ is probably why the old code kept misbehaving…
In my own domain, I actually find C (or C++ with a C ABI) to be the most useful as you can drop it in almost anywhere and usually without having to drag in a build system for the ride.

I don't even know Rust, but it seems to be well-regarded. In fact, the leader of the team I've been doing C++ training for said their next project will be in Rust rather than half C++ for low-level stuff and half .NET stuff for the rest. I'm retiring, so my brain is going to remain Rust-free (if not rust-free).

 

[aronth5]

I submit that FSD needs to be better than the best human driver ever to even be a good ADAS system.

Thats a challenge since there are human drivers that drive over 50 years without an accident of any kind

 

[Twiglett]

My car is fine. It has modern extremely capable hardware, capable of full autonomy: HW3.

Drive on a road at 50-60mph with traffic lights every half mile or so and report back.

I am curious if HW4 does better, actually, since I believe part of this could potentially be usable perception range. But I do think it is more likely a software limit, not perceptoon.

I can answer that question. I have a HW3 Model 3 and a HW4 Model Y
Traffic light behaviour on HW4 is noticably better. It spots red lights and other traffic hazards much sooner than HW3 ever did. Same for highway, HW4 spots brake lights further away than HW3.

 

[willow_hiller]

But this does not follow, in any sense of the word “safe.”

As has been pointed out, it is also safer to drive with your headlights off (when this was possible).

You need to take the statistics in the context of the entirety of the situations.

Driving on FSD Beta, on the roads that FSD Beta is typically driven on, with the disengagements that drivers have been making, is safer than not.

Just like driving during the day (with the headlights off), is safer than driving at night.

 

[AlanSubie4Life]

Driving on FSD Beta, on the roads that FSD Beta is typically driven on, with the disengagements that drivers have been making, is safer than not.

Show me the data that demonstrates this.

 

[SidetrackedSue]

Thats a challenge since there are human drivers that drive over 50 years without an accident of any kind

I'm just 2 years away from that. I intend to make it to that accomplishment, which is one of the reasons I don't use FSDb on city streets.

 

[willow_hiller]

Show me the data that demonstrates this.

It's in the screenshot above. You even agreed with me that there are fewer accidents per million miles on FSD Beta than for Teslas without active safety features. Are you trying to claim that fewer accidents does not mean more safe?

Again, I'm not making a casual statement. I'm not saying FSD Beta is the cause of the increased safety (just like the headlights are not the cause of night).

I'm only saying that the situations in which FSD Beta is used, and the drivers that use it, and the way in which they use/disengage it, are objectively safe. They don't cause an undue number of accidents per million miles.

 

[SidetrackedSue]

As has been pointed out, it is also safer to drive with your headlights off (when this was possible).

Wasn't the fact that it was not safer to drive with one's headlines off the reason daytime running lights were made mandatory?

 

[SidetrackedSue]

About that FSD safety chart:

First, are the demographics of tesla owners now the same as the general population? With the drop in the price (and the rise in the price of comparable ICE vehicles) that might now be true. It wasn't when I bought my car but it might have been in 2022, the date of the chart.

Secondly, When I was a lot younger, there was an anecdote surrounding a stat about how less likely you were to die in Soviet hospitals than in American hospitals. It wasn't that they were better at treating patients, it was that when there was no hope of recovering the patient was sent home to die.

When we get into the looking at the safety record of FSDb, I have to wonder if disengagements are the equivalent of sending patients home to die. If FSDb gets into a situation that can lead to an accident and it or the driver disengages but the accident still happens, does that count against FSD or not?

We don't know, because the data hasn't been released.

 

[AlanSubie4Life]

It's in the screenshot above.

No, it’s not.

Wasn't the fact that it was not safer to drive with one's headlines off the reason daytime running lights were made mandatory?

I think accident rates were higher when headlights were on (when it was common to not have them on during the day). This is because there were more accidents at night.

 

[sleepydoc]

Thats a challenge since there are human drivers that drive over 50 years without an accident of any kind

And there are human drivers who can’t go more than 50 days without an accident

 

[Usher]

My car is fine. It has modern extremely capable hardware, capable of full autonomy: HW3.

Drive on a road at 50-60mph with traffic lights every half mile or so and report back.

I am curious if HW4 does better, actually, since I believe part of this could potentially be usable perception range. But I do think it is more likely a software limit, not perceptoon.

I drive on a lot of roads at 55-60mph with traffic lights every few miles. HW4 does not do better. It's bad enough that I'm working on some modifications to my own car to automatically override the uncommanded braking events. The really frustrating ones are braking hard for cars crossing my path, except the braking doesn't start until after the visualization shows the car has already cleared my lane completely.

 

[sleepydoc]

Show me the data that demonstrates this.

Why? You won’t believe it.

 

[PRLM3]

Completely agree. The 5 mph limit on TACC is ridiculous, but at least FSD would respect whatever speed I set previously. Now, it goes whatever speed it wants to go, and that's typically either the speed limit, or within 3 mph of the speed limit. Often times, it'll go several mph under the speed limit for no apparent reason. In my neighborhood, it likes to go around 15, even though the limit is 25. Paying $12k for the current functionality is really pissing me off. Simple things like this speed limit issue should not be happening.

Agree completely. How they could release this and even brag about it is inexplicable to me. If it's really fixed in the next .1 release, why not just wait?

Also, look at the specific defect: consistently going lower than the speed limit. This is *not* how humans drive. So if the whole premise is "based on millions of video clips of humans driving" ... how do you figure? Also, how is something like this "fixed" with the current video/AI model if you're not writing explicit code to correct behavior? What, they took video clips of people driving faster? Doesn't add up.

Same old same old: 1 step forward, 2 steps back.

 

[FSDtester#1]

This forum for discussion is awesome, and all the different opinions are what make Tmc great.

I love hearing why I am wrong, which is frequent. The best way to learn is to make mistakes, not to always be successful.

With that said, some of these back and forth, which I have included myself, against my better, or lack of judgment, are waaaaaaay too subjective.

Someone the other day brought up the old joke about what is, is.
Loved it!

What one person likes, another can't stand. Welcome to the real world !

I have a Plaid, and the faster my car takes off in a safe manner, on or off Fsd, the better. My opinion. Most sane people disagree. No problem.
It's my car, and I will drive any way I want within the limits of the law.
Which by the way, I had to find out if it is legal to floor it at a light.
Every officer I asked said go for it. Just don't do it on wet pavement or exceed the speed limit. I couldn't believe it. So I asked, "What if you were in your patrol car behind me at a light and I did that". They all said they would laugh and do nothing unless I was speeding.

@AlanSubie4Life has taken his time for years to give his opinions on how the car should respond. Look at his post count.Huge! Most of us OG testers have gone back and forth with him countless times for years now. He and I, with some of the other usual posters have run off a lot of people I imagine.

He is always a gentleman and courteous, something I could do a better job of.

With that said, his and my style of driving are probably way different, as well as what we want our cars to do in different situations.
Who is right? Me, Alan??

Probably somewhere in the middle but who knows.

My point is that lately, all the back and forth is getting into the weeds about the definition of the word "Is" or "drive"

A little more objectivity I think will make the discussions better.

Rant over, back to your normally scheduled programming.

 

[enemji]

About that FSD safety chart:

First, are the demographics of tesla owners now the same as the general population? With the drop in the price (and the rise in the price of comparable ICE vehicles) that might now be true. It wasn't when I bought my car but it might have been in 2022, the date of the chart.

Secondly, When I was a lot younger, there was an anecdote surrounding a stat about how less likely you were to die in Soviet hospitals than in American hospitals. It wasn't that they were better at treating patients, it was that when there was no hope of recovering the patient was sent home to die.

When we get into the looking at the safety record of FSDb, I have to wonder if disengagements are the equivalent of sending patients home to die. If FSDb gets into a situation that can lead to an accident and it or the driver disengages but the accident still happens, does that count against FSD or not?

We don't know, because the data hasn't been released.

Sue - This is not human vs FSDb, but rather human vs the joint duo of human+FSDb. FSDb enables the human to be safer.

 

[kpanda17]

back to business, who has 12.3.2.1 results....