Future of FSD after "traffic light & stop sign response" release

[Knightshade]

I would agree with this. I've just seen a few messages in recent days claiming that Tesla is no longer aiming for L3 or higher while selling FSD based solely on the changed language on the order page. I think they are still aiming for L3 at least, but are just tempering expectations on the order page for what the car can do now.

I think it'd be more accurate to say they're no longer selling (or perhaps promising?) higher than L2- to avoid lawsuits and refunds if that's all they're ever able to deliver... as evidenced by the Feb/March '19 change to the FSD sale page, and again the one from the last week or two.

But they would love to deliver more if they can figure out how.

 

[willow_hiller]

But they would love to deliver more if they can figure out how.

Absolutely. I think that's well put.

And I will say, in some regards, Tesla does bend over backwards for their customers. We laugh about FSD claims made in October 2016, but I'm pretty sure that once the HW2.0 to HW3.0 retrofits start going out en-masse, all of the cars made since that promise will be able to do the same things as cars made this year.

Can you imagine any other car manufacturer going to such lengths for what is comparatively so little money? Before I bought my Model 3 I test drove a Hyundai Kona EV (which the dealer was not interested in selling me, but that's a different story). If you want Traffic Aware Cruise Control on the Kona EV, you need to shell out an additional $8,200 plus tax since it's only available on the "Ultimate" trim...

 

[mikes_fsd]

But they would love to deliver more if they can figure out how.

Like was mentioned in the other thread.
They have stated a unified approach, from Elon to Karpathy, they know how!

But armchair analysts think that they know best. Nothing about their approach had fundamentally changed.
Their biggest change was realizing how much processing power they needed and designing the chip in house.

But since those computers are in the process of being upgraded I don't see massive roadblocks in their strategy.

Having sightly different cameras from 2.0 to 2.5 is not going to make much of a difference.
I fully expect 1080p upgrades to the cameras on future builds of all Tesla models.

Let them do their job.

You can blame Elon and Tesla for being late though, that's completely fair given they're published timelines.

 

[rxgigli]

Question - at this point, is it correct to say that Enhanced Auto Pilot has reached its maximum ability* and all future enhancements will only apply to FSD owners?
*current abilities will improve via updates but Enhanced Auto Pilot will not add any new features such as city driving, stop lights / signs, etc

 

[Knightshade]

Like was mentioned in the other thread.
They have stated a unified approach, from Elon to Karpathy, they know how!

Clearly, they do not- since they promised a coast to coast FSD drive in 2016, for 2017...and it's 2020 and NO feature is above level 2, let alone the whole thing being level 5.

In fact they have repeatedly said it's really hard and they are still figuring things out

So your blind faith is not supported by literally anything they've said in recent years.

But armchair analysts think that they know best. Nothing about their approach had fundamentally changed.

Except, it absolutely, publicaly, has.

They're doing a fundamental rewrite of the code right now.

The original approach did not work

Musk has mentioned this many times in recent months.


Not sure how you're still getting such basic facts wrong.

Their biggest change was realizing how much processing power they needed and designing the chip in house.

No, that was just ONE of the changes.

They changed the sensor hardware twice now (AP2->2.5, then adding heat to the front radar with the Y)

They're changing the fundamental AP code right now (still not done)

HW3 is yet another change.

So obviously whatever vision they had in 2016- when the promised everyone "every car made has ALL the HW it needs for FSD" has been wrong. Repeatedly.

There's no reason to think they somehow were wrong that many times but NOW they're magically right.


Other than, as I say, utterly blind faith.

Having sightly different cameras from 2.0 to 2.5 is not going to make much of a difference.

Tell that to the folks who couldn't get HW3 upgrades for months because they had AP2 and they hadn't figured out how to make it work yet.

And it's still unclear if there'll be any other reduced funcationality in the future. We KNOW there's OTHER differences not just in the radar but also in the amount of redundancy, so possibly those cars will NEVER be L5 even if the newer ones are due to these things.

I fully expect 1080p upgrades to the cameras on future builds of all Tesla models.

Why?

I thought you said they didn't NEED any more HW upgrades besides HW3?

If they find they NEED 1080p then how do you ALSO expect the hundreds of thousands of 720p cars to work properly?


Pick a lane man

 

[diplomat33]

Question - at this point, is it correct to say that Enhanced Auto Pilot has reached its maximum ability* and all future enhancements will only apply to FSD owners?
*current abilities will improve via updates but Enhanced Auto Pilot will not add any new features such as city driving, stop lights / signs, etc

Yes (IMO).

 

[mikes_fsd]

Clearly you do not understand what the definition of approach, vision or fundamental means.

Clearly, they do not- since they promised a coast to coast FSD drive in 2016, for 2017...and it's 2020

You can blame them/flog them/bitch on a forum about them being late (obviously you're doing that and then some).
Being late is a valid point to criticize them for.
Just because they are late delivering their initial vision, does not mean that vision has changed.
Just because you chose a different framework for your project (software) does not mean your goal of creating a widget has changed.
After all, they are solving a new problem that has not been solved before, and they are approaching it differently then most.

Nothing about their approach had fundamentally changed.

Except, it absolutely, publicaly, has.

They're doing a fundamental rewrite of the code right now.

The original approach did not work

On any given software stack, you eventually end up being forced to rewrite large portions (sometimes fully) of your application for many different reasons, a simple example Python <=2.6.x reaching end of life and migration to Python 3.x is not hot-swappable in most cases.

No, that was just ONE of the changes.

They changed the sensor hardware twice now (AP2->2.5, then adding heat to the front radar with the Y)

My "lane" has always been that incremental changes to the car will always happen, just like they happen in the non Autopilot suite for Tesla cars those changes will happen on Autopilot suite of hardware as well.
A radar is still a radar when it has a heating element, it just takes care of extreme cold climates. Will that make Autopilot on Model 3 not be capable of performing if Model Y is able to do FSD, hardly... but you might get a message when you start your Model 3 in the Arctic Circle that your radar needs cleaning or to warm up.
Does 720p to 1080p make a huge difference for Autopilot, hell no. A few more pixels is not going to screw up a NN that was trained on 720p images, that's not how NNs work. Now if they were trying to cheap out and degrade the sensor to 480p then you would be losing information not gaining.

If a 1080p sensor comes out that draws less power then the current 720p chips and at the same (or lower) price, you would be retarded not to upgrade to the new sensor.
Same with a radar that for the same price + power consumption can be as/more accurate further out, you should take the upgraded radar it will not screw up your algorithms at all.

So, all your arguments about AP2 -> 2.5 sensor suite changes are just silly.
Now, the computer upgrade HW2 to HW3 is not silly, that was Tesla learning their lesson on power requirements to process everything they would need to handle in the real world.

So, how about you stay in your lane, because I've been in my lane and my lane is pretty well defined. I don't need a "flavor of the day" release from a LiDAR manufacturer or a paper published by an academic that has never built $#!t in their life to tell me what is needed to "solve" FSD. There are people with actual real world data that are applying their knowledge and are showing progress by releasing their approach to the future of FSD (not a demo video) to the public of almost 1 million capable vehicles.

PS - when you're forced to rewrite a portion/all of your software for a particular project, you have an opportunity to take all lessons learned from prior years and make the re-written software 10x 50x or 100x better and more future proof.
But opportunities are just that, and they come at a cost... time, effort.

 

[Knightshade]

Clearly you do not understand what the definition of approach, vision or fundamental means.

I do. I think you've just made up rose colored definitions of them- and will now use goalpost moving and the No True Scotsman fallacy to dismiss every fact-based example of you being wrong about this.

You can blame them/flog them/bitch on a forum about them being late (obviously you're doing that and then some).
Being late is a valid point to criticize them for.
Just because they are late delivering their initial vision, does not mean that vision has changed.

No, the fact they've made multiple hardware changes, after saying they already have all the HW they need on the car each time, tells you their initial vision didn't actually work out.

Their vision to do all this with HW2.0 failed

Their vision to do all this with HW2.5 failed

Apparently now their vision to do all with without heated from radar is also a failure (though they've only fixed it on the Y so far).


The fact they're now fundamentally rewriting the code also tells you that.

Their ORIGINAL vision of how to solve the problem code-wise failed


Again there's no denying these facts other than by blind faith and ignoring their own words.

Just because you chose a different framework for your project (software) does not mean your goal of creating a widget has changed.

Yup- there we go moving goalposts.

See, what you said originally was "they know how"

Which has repeatedly been proven factually wrong.

NOW you're changing your story to "They know what they want their goal to be and just keep changing their ideas on HOW to get there"


So I guess they did not "know how" huh?



In fact there's still no evidence they "know how" now

Might turn out they need yet another hardware change.

Might turn out they need yet another software redesign.


Until they actually produce a working L4/L5 system you're running on literally nothing but blind faith and ignoring their repeated and admitted failures

After all, they are solving a new problem that has not been solved before, and they are approaching it differently then most.

Yup.

You seem unwilling however to admit they also have not yet solved it.

And have failed to do so several specific times, having to make HW and SW changes each time.

My "lane" has always been that incremental changes to the car will always happen, just like they happen in the non Autopilot suite for Tesla cars those changes will happen on Autopilot suite of hardware as well.
A radar is still a radar when it has a heating element, it just takes care of extreme cold climates. Will that make Autopilot on Model 3 not be capable of performing if Model Y is able to do FSD, hardly...

Unless it's cold/snowy, right?

but you might get a message when you start your Model 3 in the Arctic Circle that your radar needs cleaning or to warm up.

...which would mean it's NOT CAPABLE OF PERFORMING like the Y is.

Does 720p to 1080p make a huge difference for Autopilot, hell no

How do you know that?

. A few more pixels is not going to screw up a NN that was trained on 720p images, that's not how NNs work.

It's absolutely going to impact the distance at which it can recognize things, because higher resolution with the same FOV produces more detail, meaning better recognition sooner.

(it'll have knock on benefits like dashcam being better at capturing license plate data, but that's a minor benefit)

Because Now if they were trying to cheap out and degrade the sensor to 480p then you would be losing information not gaining.

So... wait... what?


Somehow going from 720 to 1080 does not gain info... but going from 720 to 480 loses it?


Again- pick a lane dude.

It seems even you don't agree with you (and rightly so!)

If a 1080p sensor comes out that draws less power then the current 720p chips and at the same (or lower) price, you would be retarded not to upgrade to the new sensor.

Also if you figure out you actually need 1080p sensors to hit your performance targets- regardless of if it draws less power or not.

Kinda like they figured out they were wrong about HW2.0 cameras being good enough for their goals and had to change them for HW2.5

Same with a radar that for the same price + power consumption can be as/more accurate further out, you should take the upgraded radar it will not screw up your algorithms at all.

I mean- except of course they did.

It's one of the reasons they had to delay HW2.0 peoples HW3 upgrades- the SAME CODE didn't "just work" for them like it did for HW2.5 radar/cameras.


So once again actual facts directly prove your imagination wrong.

So, all your arguments about AP2 -> 2.5 sensor suite changes are just silly.

If you find facts silly, that's really on you

Now, the computer upgrade HW2 to HW3 is not silly, that was Tesla learning their lesson on power requirements to process everything they would need to handle in the real world.

I mean- they actually already knew that up front.

They just lied to everyone and told them it was good enough for FSD at the time.

They knew the computer was incapable of processing all 8 cameras at full frame rate- that was basic math.

7x36 fps, plus 1x30 fps= 282 fps.

With a computer that maxed out at 200 fps processing.


They were desperate though since they'd lost Mobileye, so they took what they could get, lied about it being "all you need for FSD" and quickly began development on something they thought WAS actually good enough.


But in your imagination they "always knew how" despite all the facts showing how often they didn't.

So, how about you stay in your lane

I am- it's the lane of facts.

You should try it sometime!

, because I've been in my lane and my lane is pretty well defined

It sure is!

You probably wouldn't be a big fan of its label though

. I don't need a "flavor of the day" release from a LiDAR manufacturer or a paper published by an academic that has never built $#!t in their life to tell me what is needed to "solve" FSD. There are people with actual real world data that are applying their knowledge and are showing progress by releasing their approach to the future of FSD (not a demo video) to the public of almost 1 million capable vehicles.

Uh... you realize the only evidence Elon himself has ever presented of actual FSD is a demo video, right?

(two actually- though we know now the first one was hyper-gamed by doing dozens of runs to get ONE decent looking video)


You're a font of hilarity even though you seem unable to realize it

PS - when you're forced to rewrite a portion/all of your software for a particular project, you have an opportunity to take all lessons learned from prior years and make the re-written software 10x 50x or 100x better and more future proof.
.

Absolutely.

But they're an utter admission you did not know how to do it the first time.... (which also carries an implict possibility you still don't and might get it wrong yet again.

Despite your insistence they did know AND still do.

 

[mikes_fsd]

gotta love your "fact based" life when you get to make up $#!t on the fly.
What I said:

Does 720p to 1080p make a huge difference for Autopilot, hell no. A few more pixels is not going to screw up a NN that was trained on 720p images

vs. what you claim I said.

Somehow going from 720 to 1080 does not gain info... but going from 720 to 480 loses it?

FYI, gaining higher accuracy via improved sensors will not degrade performance of a NN. That is the whole argument of incremental sensor suite upgrades.

They just lied to everyone and told them it was good enough for FSD at the time.

First, the free upgrade to HW3 is the "lie"?? They knew they needed more power but the industry could not provide it, they got the best NVidia could put together and went with it. Taking the free upgrade expense on themselves. But you feel maligned.

It seems that you are a proponent of them just sitting on their ass and pumping out ~1 million cars until HW3 and until they had "figured it all out".
Elon builds rockets iteratively ... their cars components go through iterations that are not tied to model year. Autopilot is going to be done in similar fashion and that does not sit well with you.

Uh... you realize the only evidence

This is where your "facts" have led you, you do not see ~1 million cars getting regular updates and evolving/improving and providing feedback, all you see is that "it's not the promised FSD" and "I don't see how it could ever get there".

If you are so limited by our own ability to see the progress, step out of the way and let the ones that do not have such limits do their job.

/fin

 

[Knightshade]

gotta love your "fact based" life when you get to make up $#!t on the fly.
What I said:

vs. what you claim I said.

So then can you explain if you agree changing resolution changes the info provided- why you claimed that switching to higher res cameras wouldn't change how (or how well) the system works?

Because if they decide they NEED 1080p cameras- obviously it means all the cars with 720p cameras can't get the job done, right?

FYI, gaining higher accuracy via improved sensors will not degrade performance of a NN. That is the whole argument of incremental sensor suite upgrades.

...what?

Each time they've upgraded the sensors so far it hasn't been "cost" savings, it's been they realized they needed better sensors to actually accomplish their goal

First, the free upgrade to HW3 is the "lie"??

No- the fact they told people buying HW2 cars it has "all the hardware needed" for FSD was the lie.

They knew that wasn't true at launch- already showed you the math.

Then they lied again to the folks with HW2.5 cars saying THOSE cars CAME FROM THE FACTORY with all the HW needed for FSD.

Also a known lie (and even larger one since HW3 was well under development by then)

They knew they needed more power but the industry could not provide it, they got the best NVidia could put together and went with it.

Yes- and then lied by telling everyone it was already enough, when they knew for a fact it wasn't.


Not sure why you keep being an apologist for outright lies from the company?

Taking the free upgrade expense on themselves. But you feel maligned.

Hey, we lied to you repeatedly about the HW in your car being sufficient... but since we will eventually give you upgrades you don't get to point out we keep lying about things...


Nope, not how that works.

It seems that you are a proponent of them just sitting on their ass and pumping out ~1 million cars until HW3 and until they had "figured it all out".

Never said anything remotely like that.

But if they KNEW the HW was not good enough (and they did) then maybe, I dunno, actually be honest and don't LIE about it being good enough?

Just say "These cars have the best hardware available at the time- we expect to upgrade it further, for free, as need arises if you bought FSD"


Instead it was over 3 years after they released FSD before they first put in writing that anybody would get a free upgrade during FSD purchase- and the whole time they kept telling people buying HW2 and HW2.5 cars they ALREADY HAD all the HW they needed for FSD.



Why keep defending an objectively proven lie dude?

Elon builds rockets iteratively ... their cars components go through iterations that are not tied to model year. Autopilot is going to be done in similar fashion and that does not sit well with you.

Again making up stuff I never said.


I'm just pointing out the fact he has to keep changing HW proves he has not already solved this

Because if he had he'd start with the right hardware.


So again you original claim is directly disproven by his own words and actions.

He does not "know how to do this"

He "knows his goal" and keeps CHANGING how he gets there as each thing he tries doesn't work out- so he tries something ELSE.


Maybe the next time he'll get it right.

Maybe 10 years from now he STILL won't have it right.

We DO know he's been wrong a number of times, and his claimed target dates have been drastically wrong over and over.

you don't know either way despite your utter blind faith it's all handled.

This is where your "facts" have led you, you do not see ~1 million cars getting regular updates and evolving/improving and providing feedback

Sure I do.

I own one.

I also know that 3 years after level 5 FSD was supposed to be done, they're still not even managing level 3.

, all you see is that "it's not the promised FSD" and "I don't see how it could ever get there".

To L5? Correct, I don't see how.

Neither do you if you look at all known facts- you have to basically ignore the entire history of development to believe it at his point.

I do think they can PROBABLY reach level 3 on the highway... not sure they can in the city.

I don't see them handling L4 (let alone 5) without fundamental upgrades to the sensor suite though at a minimum... not "slightly higher resolution" to existing cameras- but more of them (and probably another radar or two).

There's still spots around the car the cameras can not see due to number and placement.

Those close-to-car blind spots are probably not super important for L3 highway driving- but they'd be pretty damn important for L5 city driving.

If you are so limited by our own ability to see the progress

No, you however seem very limited in your ability to see, admit, or acknowledge overpromising and underdelviering.

Which is what FSD has been since launch in 2016.

It's 2020 now, and only in the last couple of weeks has the first actual FSD feature launched.

(for us older EAP owners, every single feature has been available to us so far WITHOUT buying FSD- until the stoplight/stopsign thing)



If you wish to continue maintaining utter and absolute blind faith in Elon and FSD despite years of evidence that's not a sensible perspective, that's up to you.

None of that means there won't be progress and cool features. But it does mean your "they know how to do it" claim remains not just unproven, but actively disproven over and over again for the last several years.

They're still trying to figure that out

They explicitly have said they are still trying to figure it out

You have so much faith in them you won't even believe them telling you you're wrong!

 

[mikes_fsd]

Because if they decide they NEED 1080p cameras- obviously it means all the cars with 720p cameras can't get the job done, right?

This is how I know you're just trying to have argument for argument sake.
No one said they NEED 1080p cameras.
I said, if the 1080p cameras are as cheap or cheaper and use same power or less then the current one, you would be stupid not to upgrade.
The NN does not care that it is higher resolution, it's just 1s and 0s. Because it was trained to handle the information with less information, having more information is not going to hurt it.
Losing information is bad, getting more information is not.

You have fun with your "facts" I will just keep enjoying as the updates roll in.

As a side note, I started out with a Tesla on AP1, so I have lived through the great "parity with AP1" drama that was whipped up on here.
Of all ppl that I've talked to in real life, no one wants to go back to AP1, but on this forum you can still find a few sad souls that swear that AP1 is "da best".

But then again on this forum you still have ppl pining for Tesla OS 8.x, this truly is a wonderful place.

 

[diplomat33]

You have fun with your "facts" I will just keep enjoying as the updates roll in.

 

[NHK X]

Yes (IMO).

Well I would say yes we’ve reached max for feature set for EAP, but we haven’t reached max reliability yet and there is room for improvement in driving policy (ie lane merging nuance). I think it can and will continue to get much better even if the feature set doesn’t expand.

 

[MP3Mike]

Well I would say yes we’ve reached max for feature set for EAP, but we haven’t reached max reliability yet and there is room for improvement in driving policy (ie lane merging nuance). I think it can and will continue to get much better even if the feature set doesn’t expand.

Won't EAP get Reverse Summon?

 

[diplomat33]

Well I would say yes we’ve reached max for feature set for EAP, but we haven’t reached max reliability yet and there is room for improvement in driving policy (ie lane merging nuance). I think it can and will continue to get much better even if the feature set doesn’t expand.

Sure, there might be improvements in reliability but don't expect too much. Tesla is clearly focused on HW3 and FSD now.

 

[Knightshade]

Won't EAP get Reverse Summon?

I don't think it will.... "seek" of parking spots was only ever listed for FSD.

EAP included autopark but that was just parking itself when already near a spot- which is has done for years now.

 

[NHK X]

Sure, there might be improvements in reliability but don't expect too much. Tesla is clearly focused on HW3 and FSD now.

Don’t think it’s an all or nothing. Sure hardware will matter with the rewrite, which supposedly is optimized for the new FSD computer. But I haven’t seen anything saying it won’t run or improve experience on HW2.5. And even if that isn’t compatible I don’t see why better driving policy, and improvements to existing NOA can’t get better.

 

[Knightshade]

Don’t think it’s an all or nothing. Sure hardware will matter with the rewrite, which supposedly is optimized for the new FSD computer. But I haven’t seen anything saying it won’t run or improve experience on HW2.5.

The re-write requires full frame-rate from all cameras at all times to create a true 360-degree picture of surroundings.

HW2.x is incapable of doing that.

I would expect the re-write to be where 2.x and 3 officially fork ways as far as code base. (remember, the fundamental processors are different- right now HW3 is running 2.x code in emulation because it's so much more powerful it can afford to do this.... but the re-write is almost certainly HW3 native code that WILL NOT RUN on 2.x at all)

Do
And even if that isn’t compatible I don’t see why better driving policy, and improvements to existing NOA can’t get better.

Ask AP1 guys how much "better" their stuff got once development moved on to AP2

Granted there's some differences there- and I expect any code that isn't NN dependent might indeed get at least bits of it ported over to 2.x code for a little while-

But I absolutely would not expect any new features at all, just maybe occasional incremental improvements in existing behavior of existing features for a little while.

But ultimately as the % of the fleet that is HW2.x continues to shrink to a smaller and smaller minority- it won't make sense to keep putting any effort into development on it.

 

[Bladerskb]

A radar is still a radar when it has a heating element, it just takes care of extreme cold climates. Will that make Autopilot on Model 3 not be capable of performing if Model Y is able to do FSD, hardly... but you might get a message when you start your Model 3 in the Arctic Circle that your radar needs cleaning or to warm up.

Arctic circles lmao. The radar fails in the lightest snow and moderate rain and disables AP. You don't need extreme cold climates. This alone shows how misinformed you are.

Does 720p to 1080p make a huge difference for Autopilot, hell no. A few more pixels is not going to screw up a NN that was trained on 720p images, that's not how NNs work. Now if they were trying to cheap out and degrade the sensor to 480p then you would be losing information not gaining.

Another misinformed statement. Look below if you are interested in comparison of current gen cameras being introduced into some new cars today versus the last gen cameras that Tesla still puts in their cars today. And I haven't even gotten into comparison of low light, direct light and transition conditions.

 

[bedoig]

According to my calculations, even the old "1.27 million megapixel" version ought to be good enough to travel safely at somewhere between the speed of sound and the speed of light