Starting to regret FSD pre-purchase in a major way

[Daniel in SD]

You are the one who said "I only engage autopilot on divided highways, the type of road with the lowest accident rates per mile."

So surely you have a source for your claim regarding the rate per mile of accidents on different types of roads, right? I asked to see it.




No, the reason is because EAP is not designed to handle all situations. And never is intended to be.

That's the major difference between Level 1/2 systems and level 3/4/5 systems.

Level 1/2 systems will always require driver attention to handle the situations the system will not ever be able to handle. If a situation it can't handle comes up, the driver is expected to handle it.

The "maximum" safely level of 1/2 systems is "expects the driver to be actively paying attention and ready to avoid things the car is not capable of avoiding at any moment"

If the driver doesn't pay attention, the car still may run into something.


Level 3 goes a step further- your immediate attention is no longer required...the vehicle will handle situations that call for an immediate response like emergency braking.... it will ask the driver to take over but with time for them to safely do so only when leaving the domain(s) in which it is able to operate (for example the Level 3 car Audi sells only offers level 3 on certain roads below certain speeds)... the car only becomes "unsafe" if the car is exiting its domain and the driver then STILL doesn't ever take back over. This is a "It's ok to be reading a book or watching a movie while driving" state

Level 4 goes beyond that, it will not only handle situations that call for emergency response on its own, if it finds something it can't handle and you don't take over it will safely stop the car someone safe for you. This is a "it's ok to actually be asleep" state. The car is always safe without the driver needing to do anything.

Level 5 is just level 4 that works "everywhere" instead of specific domains.


EAP is a level 2 system.



Also incorrect.

While they'll certainly leverage some EAP code, FSD will require significant additional code that EAP-only cars won't be running and additional hardware for current cars for FSD as well (at least the next gen processor so far, possibly more at some point).





It's not, because that's not how software works.

FSD isn't just "really good EAP"

EAP isn't ever meant to have the "level of safety" you seem to imagine for it. It's not a Level 4/5 system, it's not ever supposed to be, and was never sold as one.

EAP will continue to be developed to be better at the things it's actually meant to do. "avoid every possible accident in all situations" isn't one of them.

It's actually much harder to find stats on this than I thought it would be. Most of the focus is on fatal accidents. But I'll give it a shot.
The police reported accident rate in the US is about 2 per million vehicle miles traveled.
6.3 million police reported accidents in 2015 https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812318
3.1 trillion vehicle miles traveled in 2015 Press Release: U.S. Driving Tops 3.1 Trillion Miles In 2015, New Federal Data Show, 2/222016 | Federal Highway Administration
Now the problem is I can't find collision data by road type for the entire country I'm not about to add up all the state data.
In California (where most Teslas are. haha) it's:
0.82 per million vehicle miles for freeways
1.12 per million vehicle miles for undivided 2 lane highways. http://www.dot.ca.gov/drisi/library/tasas/2015CollisionDataonCaliforniaStateHighway.pdf
I can't find stats for cities but I bet it's at least as bad as the country average judging by CA insurance rates!

Well I think we've found the our fundamental disagreement. I think FSD on the highway is just really good EAP. Tesla advertises it as being able to take you from onramp to offramp without driver intervention. Clearly they are trying hard to make EAP as safe as possible (how else do you explain phantom braking?). They investigate autopilot crashes and try to figure out how to avoid them in the future. I just don't see them branching the code to make FSD safer on the highway unless there is a hardware update required.

 

[LM-5]

Tesla aims for new neural net computer in production in 6 months, results in 500-2000% increase in ops/sec, says Elon Musk

 

[daniel]

I think we'll probably see some level 3 features much earlier.

There's no such thing as "level 3 features." Level 3 (like any other level) is a description of what the car can do. Either it meets that description or it does not.

For level 3 it's enough if the car can recognize situations it can't handle and alert the driver. This is why the "traffic jam pilot" in some European Audi models is level 3: while the system can only be used in a limited set of circumstances, it can recognize if these circumstances are about to change and alert the driver. Recognizing critical situations should be easier than being able to handle them autonomously.

Okay. Fair point. There are cars that are Level 3 but only allow "eyes off the road" under limited circumstances. EAP only engages under certain conditions. Recently (still on v8 at the time) and still now (on v9) my car increased the range of conditions where autosteer could be engaged, making it better. Even Level 4 is specifically domain limited. So I guess I'd judge the usefulness of a Level 3 car on the range of conditions where "eyes off" was allowed.

The reason you have to pay attention is because the car might run in to things. Once the EAP code is mature enough to be used for FSD it will no longer crash in to things. I'm trying to understand how this is going to work from a software development point of view. Will they stop developing EAP at a certain level of safety and reserve all future safety improvements for FSD? It just doesn't seem realistic to me.

I don't think that's what will happen. I think the FSD package will get functions that EAP does not, such as making turns based on NAV and being able to handle situations that EAP cannot. Level 2 is "Eyes on the road at all times." Maybe EAP will get as good at accident avoidance as FSD, but without the ability to stop for stoplights so you still have to keep your eyes on the road. Or maybe EAP will become Level 3 on the highway only but FSD becomes Level 3 in city driving.

Tesla has stated that all cars will have all safety features, so I doubt that accident avoidance will be left out of some cars when it's included in others.

 

[Knightshade]

It's actually much harder to find stats on this than I thought it would be.

I know, that's why I asked for a source on the claim, since I've looked before and there's not a lot of great data on it nationally....I was hoping you knew some source I didn't

In California (where most Teslas are. haha) it's:
0.82 per million vehicle miles for freeways
1.12 per million vehicle miles for undivided 2 lane highways. http://www.dot.ca.gov/drisi/library/tasas/2015CollisionDataonCaliforniaStateHighway.pdf

ok... so based on that--- .82/1.12=0.73

And the Tesla numbers of AP vs non-AP driving is 1.92/3.34=0.57

In other words- the ratio of AP to non-AP accidents in Teslas is lower than the ratio of all statewide accidents freeways vs undivided 2 lane roads.

So based on the best data you could find on the difference between roads, Teslas numbers prove AP makes the car safer than manual driving.... (and that is assuming ALL AP use is freeways- which we know it's not... meaning the amount it makes driving safer is even higher)

Well I think we've found the our fundamental disagreement. I think FSD on the highway is just really good EAP.

That's odd, since Tesla themselves is pretty clear about that not being the case.

Tesla advertises it as being able to take you from onramp to offramp without driver intervention

No, they don't.

What they say EAP will do:

"Your Tesla will match speed to traffic conditions, keep within a lane, automatically change lanes without requiring driver input, transition from one freeway to another, exit the freeway when your destination is near, self-park when near a parking spot and be summoned to and from your garage."

Guess what it says immediately after that?

"That said, Enhanced Autopilot should still be considered a driver's assistance feature with the driver responsible for remaining in control of the car at all times"

So it does not tell you it does everything the whole way without ever needing driver intervention- on the contrary it tells you it will assist with many aspects of the trip, but you must be ready to intervene at all times.


Now, what does FSD say?

"The system is designed to be able to conduct short and long distance trips with no action required by the person in the driver's seat"

Those aren't the same feature. And Tesla is pretty clear about it.

. Clearly they are trying hard to make EAP as safe as possible (how else do you explain phantom braking?).

You explain phantom braking by the system not being good enough yet at recognizing shadows yet (see also how v9s display thinks cars are in your lane when they cast shadows into it)

I just don't see them branching the code to make FSD safer on the highway unless there is a hardware update required.

They don't need to "branch" the code.

The EAP items that FSD leverages will remain (and likely keep improving) but there will be additional code that AP only cars don't run to do all the additional things FSD does but EAP does not.

One really obvious hardware difference is the new computer. AP only existing cars won't get it.

But it'll be required to be able to process 8 cameras worth of visual data in real time.


EAP is not FSD. It's not meant to be, and it never will be.

 

[Eno Deb]

There's no such thing as "level 3 features." Level 3 (like any other level) is a description of what the car can do. Either it meets that description or it does not.

I would call Audi's "traffic jam pilot" a level 3 feature. It does not mean that all its driver assist features are level 3. Similarly, I expect that Tesla will eventually allow the driver to take the hands off the wheel under a limited set of circumstances while still requiring constant supervision in others, long before there is "full self driving".

 

[Knightshade]

I would call Audi's "traffic jam pilot" a level 3 feature. It does not mean that all its driver assist features are level 3. Similarly, I expect that Tesla will eventually allow the driver to take the hands off the wheel under a limited set of circumstances while still requiring constant supervision in others, long before there is "full self driving".

TJP from Audi makes the car level 3 under the circumstances under which TJP operates.

Level 3, by definition, only works under certain circumstances... ditto level 4....(the main difference between 3/4 is that 4 doesn't ever require the human to do anything for safety even in non-emergency situtaitons- a level 4 car if needed can safely pull itself over and park if it has to- while a 3 expects a human can eventually take over if it leaves its domain of operation)

The only unrestricted self driving level is 5.

 

[Daniel in SD]

ok... so based on that--- .82/1.12=0.73

And the Tesla numbers of AP vs non-AP driving is 1.92/3.34=0.57

In other words- the ratio of AP to non-AP accidents in Teslas is lower than the ratio of all statewide accidents freeways vs undivided 2 lane roads.

So based on the best data you could find on the difference between roads, Teslas numbers prove AP makes the car safer than manual driving.... (and that is assuming ALL AP use is freeways- which we know it's not... meaning the amount it makes driving safer is even higher)

Well, I was unable to find the state level data for overall accidents per million vehicle miles. For the US it's 2 per MVM which is way higher than 0.82! (and of course includes freeways, so the city number is higher still). I'm willing to bet that CA has a higher accident rate than the country as a whole.
Fatality rates in CA are 0.27 per 100 million vehicle miles on the freeway and 0.99 per 100 million miles overall. The freeway is the safest place to drive. Believe whatever you want but I think the onus is on Tesla to prove that autopilot is safer and if they're not controlling for the type of road their data is garbage.

"That said, Enhanced Autopilot should still be considered a driver's assistance feature with the driver responsible for remaining in control of the car at all times"

So it does not tell you it does everything the whole way without ever needing driver intervention- on the contrary it tells you it will assist with many aspects of the trip, but you must be ready to intervene at all times.

So you don't think FSD "features" will have a similar warning? The warning is there because it's still in beta. I don't believe that Tesla will allow autopilot to get in to crashes simply to maintain differentiation from FSD. I agree that if it turns out to be a hardware limitation then there will be a difference in safety.

 

[Knightshade]

Well, I was unable to find the state level data for overall accidents per million vehicle miles. For the US it's 2 per MVM which is way higher than 0.82! (and of course includes freeways, so the city number is higher still). I'm willing to bet that CA has a higher accident rate than the country as a whole.
Fatality rates in CA are 0.27 per 100 million vehicle miles on the freeway and 0.99 per 100 million miles overall. The freeway is the safest place to drive. Believe whatever you want but I think the onus is on Tesla to prove that autopilot is safer and if they're not controlling for the type of road their data is garbage.

I'm reluctant to call it "garbage" when it appears to be better than the data literally any other car maker provides about the safety of their vehicles.

And when nobody seems to be able to actually show the conclusions are incorrect, other than hand waving and insisting they are "sure" the Tesla data isn't good.

I also don't see how as a practical matter anybody can doubt AP reduces accident rates... thanks to how much press they get we always know about ones ON AP, and we can count on one hand the # of accidents it has "caused" (and I put that in quotes because even in those few it was generally the driver failing to pay attention or act that really "caused" it).

But Tesla supposed has over a billion miles drive on autopilot... that's a lot of miles of things like TACC keeping a car from hitting another one becuase the inattentive driver looked down at their phone and other such nonsense that happens daily around here.

So you don't think FSD "features" will have a similar warning? The warning is there because it's still in beta.

The warning I quoted from the EAP description is not there because it's in Beta. The warning is inherent to the feature set

EAP will always require driver control and attention. That's what the product is

And your not accepting that fact seems the core of your confusion between EAP and FSD.

I don't believe that Tesla will allow autopilot to get in to crashes simply to maintain differentiation from FSD.

This is a nonsensical statement nobody has made.

it isn't going to "allow" things to crash because you didn't pay for FSD.

The car might crash because when using EAP the driver is still required to intervene when EAP can't handle a situation- and if he doesn't, the car might crash. That's explicitly what the feature tells you before you even order it.

Tons of other car companies have their own version of TACC... they're generally much more limited... for example a recent one I drove turns itself off below about 25-30 mph... if you aren't paying attention and traffic slows down ahead of you you might crash.

Is that car maker "allowing" the car to crash by turning off that feature at 30?

 

[aaron0k]

I'm reluctant to call it "garbage" when it appears to be better than the data literally any other car maker provides about the safety of their vehicles.

Unless there is an independent study performed, I'd advise reading manufacturer reports with a tablespoon of salt instead of a gallon of kool-aid.

These 11 car models have had zero fatalities (when the study was published).

Audi A6 Quattro
Audi Q7 Quattro
BMW 535i
BMW 535xi xDrive
Jeep Cherokee 4x4
Lexus CT 200h
Lexus RX 350
Mazda CX-9
Mercedes-Benz M-Class 4Matic
Toyota Tacoma Double Cab Long Bed 4x4
Volkswagen Tiguan

If I understand your argument, considering Tesla has had several documented fatalities... and these other models do not... they must be safer than the Model S/X even with AP?

 

[Knightshade]

Unless there is an independent study performed, I'd advise reading manufacturer reports with a tablespoon of salt instead of a gallon of kool-aid.

These 11 car models have had zero fatalities (when the study was published).

Audi A6 Quattro
Audi Q7 Quattro
BMW 535i
BMW 535xi xDrive
Jeep Cherokee 4x4
Lexus CT 200h
Lexus RX 350
Mazda CX-9
Mercedes-Benz M-Class 4Matic
Toyota Tacoma Double Cab Long Bed 4x4
Volkswagen Tiguan

If I understand your argument, considering Tesla has had several documented fatalities... and these other models do not... they must be safer than the Model S/X even with AP?

...I'm not clear what study you're talking about?

The one I quoted the Tesla numbers from was simply tesla releasing their own data across their entire fleet- which AFAIK no other manufacturer does.

I find it hard to believe nobody has ever died in such common vehicles as an RX350 or a Cherokee though.

 

[Dana1]

Ships have had autonomous positioning for about 40 years. It’s called dynamic positioning. It’s allows many different types of ships to perform tasks better than human. A container ship can have a transatlantic route in and travel two weeks on it’s on. A pipe layer can follow an ROV across an oilfield reeling pipe out, an offshore vessel can remain 10 meters from a platform under the crane offloading supplies, or a drill ship can remain in a 6 diameter circle above a well drilling a mile+ below.

The systems can adjust for stronger winds/current, don’t care it’s the end of a 12 hour shift. But there is always a human there to take over for the weird, oddball judgment calls it’s hard to code. It’s okay to hit the plastic bag that flies in front of the windshield. It’s not okay to hit the baby carriage. The bird is undesirable but better than swerving in front of a semi.

Many accidents with dynamic positioning are due to inadequate intervention by the human, even after the system provides warnings. But the systems provide more benefit that downside.

EAP/ FSD strike me as going down the same road. It’ll probably advance quickly, like DP did, in some aspects. We already see the cars responding to posted speeds when on TACC/Autosteer. I think this is a precursor step to autonomy to see how well the car can choose a speed. Unfortunately the mapped dataset isn’t quite up to snuff yet. If it can compare speed limit signs and mapped data that might help. If it can compare those two data points to the speed of cars around it that might also help. As we use the system we find the flaws and the engineers can go back to the drawing board.

 

[Dana1]

Unless there is an independent study performed, I'd advise reading manufacturer reports with a tablespoon of salt instead of a gallon of kool-aid.

These 11 car models have had zero fatalities (when the study was published).

Audi A6 Quattro
Audi Q7 Quattro
BMW 535i
BMW 535xi xDrive
Jeep Cherokee 4x4
Lexus CT 200h
Lexus RX 350
Mazda CX-9
Mercedes-Benz M-Class 4Matic
Toyota Tacoma Double Cab Long Bed 4x4
Volkswagen Tiguan

If I understand your argument, considering Tesla has had several documented fatalities... and these other models do not... they must be safer than the Model S/X even with AP?

I’d question this, particularly for models like the Jeep that have been around longer than airbags have.

 

[Daniel in SD]

I have no idea whether EAP is safer than driving without it. I use it and like it. I'm just saying that you have to somehow account for which roads are being driven on with autopilot vs. without it. They also don't talk about the severity of accidents. The information Tesla has provided is worse than useless because it is misleading.

Sorry, I still believe that Tesla's ultimate goal with EAP software is to have zero preventable accidents. If they have the software to do that they will include it with EAP. This isn't a question of the definitions of the levels of autonomy to me.

 

[Daniel in SD]

I’d question this, particularly for models like the Jeep that have been around longer than airbags have.

Obviously they're talking about the current model.

 

[Knightshade]

They simply stated the accident rate per mile for AP and non-AP.

Those are facts they are reporting (and that no other manufacturer reports at all).


And you can of course believe whatever you wish- but the definitions, and limits, of EAP are pretty clearly spelled out, in writing, on Teslas own website, before you even purchase the option.

So expecting that they're lying about it for no apparent reason doesn't seem very reasonable of a belief.

 

[Daniel in SD]

They simply stated the accident rate per mile for AP and non-AP.

Yep. And that is misleading because it could cause people to conclude that autopilot is safer.

And you can of course believe whatever you wish- but the definitions, and limits, of EAP are pretty clearly spelled out, in writing, on Teslas own website, before you even purchase the option.

So expecting that they're lying about it for no apparent reason doesn't seem very reasonable of a belief.

They don’t make any claims to the safety of autopilot in their description. I admit that I am making an assumption that they will try to make it as safe as they can.

 

[aaron0k]

...I'm not clear what study you're talking about?

Stronger economy means higher death rate

 

[Knightshade]

Stronger economy means higher death rate

that is... an oddly specific study...

the table you cited there is labeled-

"Fewer than 8 driver deaths per million registered vehicle years, 2014 and equivalent earlier models, 2012-15"


First thing I'd note is you listed 11 models with no deaths-

Audi A6 Quattro
Audi Q7 Quattro
BMW 535i
BMW 535xi xDrive
Jeep Cherokee 4x4
Lexus CT 200h
Lexus RX 350
Mazda CX-9
Mercedes-Benz M-Class 4Matic
Toyota Tacoma Double Cab Long Bed 4x4
Volkswagen Tiguan

That's not QUITE what your source says.

For example that Lexus RX350? Looking closer you see it's 0 for the RWD one, and 2 deaths for the AWD one.

So I was correct, the RX350 doesn't have 0 deaths, even in that limited time period... it's just that all the deaths were in AWD models not RWD one (and AFAIK most RX350s sold are AWD)

That VW with 0? Also only the 2WD model. The 4WD one has -30- deaths. Guess which most of their sales is? 30 was the average for the entire industry- making the generic VW listing even more misleading.

The Cherokee is accurate, but that's because it was brand new that year- so unlike other cars where a 3 year period is reported they only got 1 year of data.



On top of that- since the data looks at only a tiny slice of time, all before EAP even existed, and mostly in a period where very few teslas were on the road at all.... so few in fact that Tesla does not even appear in the study AT ALL...it's not especially useful info for this discussion.

Sadly they haven't published any new study since that 2014 model year one. They're due to soon though, and it should include Teslas (though likely not the Model 3)

 

[aaron0k]

On top of that- since the data looks at only a tiny slice of time...

Sorry to have baited you with that one. Kind of my point. People can make claims about any data point they want, without nuance or context and by leaving out the stuff that doesn't support the mission. Is Tesla doing that? One has to decide for themselves.

At least we all now know that you're more likely to die in a Tesla than a Jeep Cherokee.

 

[aaron0k]

Was Tesla reading this thread?

Announcing they're no longer going to sell vapor-ware?

Tesla removes ‘Full Self-Driving Capability’ package from its options due to ‘confusion’