Almost ready with FSD Beta V9

[aronth5]

It sure will be more enjoyable to watch V9 beta videos (good or bad) then to follow this thread. Lighten up everyone.

 

[Bladerskb]

They're not though.

They stopped using it in mid-2017.

They NEVER used it on the 3 or Y which are the majority of cars they've ever built.


You've been shown links to news stories proving this.

You've been shown links to the Tesla parts catalog proving this.

Why keep lying about it anyway?

There are tens of thousands of Teslas with the Bosch radar which Elon/Tesla says is L5 and thousands who paid for FSD which Tesla has to support.
Till tesla comes out and cuts off support to order cars, then This is also the radar Tesla uses.

A completely worthless and useless radar. Just as much as the continental ARS 410 radar.

Both are 2D radar. This is why Waymo said they won't sell their system because they would have to support old hardware and make their new software support it. They would rather do subscriptions.

Tesla attempts to use a ~2010 radar that was meant for ACC, for self driving.
Brags about it constantly. Then years later throws its hands up because the radar is useless, then says Radar is ultimately a crutch.
They are stuck with old ancient hardware. Not just their radar but their 1.2 MP cameras of different sort.

Waymo on the other hand has a 2020 SOTA 4D imaging radar that they are deploying in 2021 and replacing their entire fleet with.

 

[Knightshade]

Tesla keeps telling us FSD is subject to regulatory approval.

Nope.

They keep telling us future features may be. And thus may impact rolling them to customers.

Not the stuff it already does.

You think they couldn't turn off FSD claiming some regulator told them they had to?

Not really.

All but one of its existing features have been in widespread use nation-wide since roughly 2017.

And they're all L2- something that nobody regulates in the US at all.

I mean sure, the license isn't revoked

Thanks for admitting your claim they could revoke it at any time was factually incorrect.

Now if we could get that wrong-about-the-radar guy on that train!

, but the features it buys you could dramatically change at any time

The existing features were part of the purchase agreement, so not really.

. Tesla has already removed features before with software updates.

Which features that were specifically called out as part of a paid package were removed for the existing owner that bought them with software updates?

I'm confused now- I have two pre 2019 cars. I keep getting told by FSD apologists that the delays are totally cool, and that Tesla never needs to deliver full hands off driving at a safety level 2X that of a human, despite that being the claim on the website at the time of purchase, and that they have sneaky language such that I should have read the fine print and know I'm not owed anything. Now I hear they are required to deliver all features. So *when* do I get full hands off driving on short and long trips like was promised?

Not sure why you're asking me, since I've never said any of the things you're claiming...someone... has said.

So I guess if anyone did say those specific things you'd need to ask the people who said it.

What I said in reply to you was Tesla is required to deliver all EXISTING features immediately to anyone who buys FSD, as part of that purchase (or as immediately as sending the signal to the car takes anyway).

For post ~3/19 buyers they're also obligated to deliver at a minimum L2 city-streets driving (what is currently making the rounds as FSDBeta, but is only actually city street driving- as of 8.2 at least the remaining code was unchanged).

Pre 3/19 buyers are owed significantly more (at minimum L4 driving with very little ODD limitations), but Tesla has some considerably leeway on the timelines for delivering it.


I've also said- (to disagrees for a certain few) that if they ever admit they're not ABLE to deliver L4, they'll owe the pre-3/19 buyers refunds (but not necessarily the post 3/19 ones)


And guess what their most recent 10Q does for the very first time?

Admit that's at least a possibility that would have material impact on the company (since it'd be 3-5k refunds for a fair # of folks- but the hit to future revenues obviously)

 

[Knightshade]

There are tens of thousands of Teslas with the Bosch radar

That's not what you, repeatedly, and factually incorrectly, originally claimed though.

You claimed that was the CURRENT radar.


Are you finally admitting your claim was wrong?

which Elon/Tesla says is L5 and thousands who paid for FSD which Tesla has to support.

None of that is correct either.

Nothing in Teslas official documentation around FSD says L5. Nothing during the purchase process promised those folks L5.

It did promise them L4, but not that they were obligated to use radar to get there.

In fact- those folks already got a free CPU upgrade because Tesla determined the car needed it. At the time the 2->2.5 sensor upgrade happened Tesla told 2.0 owners they'd ALSO get sensor upgrades if it was determined those were needed. For free. (So far that has not proven to be needed)

On top of that Tesla recently announced it does not need radar at all to reach their FSD goals.


So your weird obsession with ancient radar they haven't used in years, and now don't need AT ALL continues to be frankly bizarre.

 

[ZeApelido]

Luckily we won't have to argue about which radar the cars have, since they won't be using it.

 

[Bladerskb]

That's not what you, repeatedly, and factually incorrectly, originally claimed though.

You claimed that was the CURRENT radar.


Are you finally admitting your claim was wrong?

Yes they currently use Bosch MRR and Continental ARS 410. That's what their fleet comprises of.
Which part of that don't you understand?

None of that is correct either.

Nothing in Teslas official documentation around FSD says L5. Nothing during the purchase process promised those folks L5.

It did promise them L4, but not that they were obligated to use radar to get there.

In fact- those folks already got a free CPU upgrade because Tesla determined the car needed it. At the time the 2->2.5 sensor upgrade happened Tesla told 2.0 owners they'd ALSO get sensor upgrades if it was determined those were needed. For free. (So far that has not proven to be needed)

On top of that Tesla recently announced it does not need radar at all to reach their FSD goals.

Nothing in Teslas official documentation around FSD says L4. Nothing during the purchase process promised those folks L4.
On the other hand What Elon has repeated in press releases, including official publication is that the car is L5.

So your weird obsession with ancient radar they haven't used in years, and now don't need AT ALL continues to be frankly bizarre.

WTF they use the radar. The radar output in those teslas are not ignored.
Whats with Tesla fans and spreading misinformation?

 

[Cheburashka]

Luckily we won't have to argue about which radar the cars have, since they won't be using it.

Exactly. I have FSD and also the junk Bosch radar.

I had an old Infiniti that used a camera + laser for cruise control & lane keeping and it worked wonderfully well. Never had to keep my foot by the gas when in ACC because there was no phantom braking, unlike the Tesla. Oh yeah, this was in 2007.

So Tesla can't get cruise control to work reliably in all these years yet Level 5 is just around the corner.

 

[Bladerskb]

On top of that Tesla recently announced it does not need radar at all to reach their FSD goals.

Tesla's radar is trash and not meant for autonomous driving, that's like saying I don't need my sandals to run a 10 seconds 100 meter dash while others are running in custom PUMAs and then claiming you are better than them because they use shoes and you ditched your shoe (sandals) which you call a crutch.

 

[gearchruncher]

Which features that were specifically called out as part of a paid package were removed for the existing owner that bought them with software updates?

Launch mode (part of the Ludicrous paid upgrade) was limited in how many times it could be used via a SW update that was not present when the car was delivered, nor indicated in any documentation by Tesla:

Tesla reverses recent limit on lifetime full-speed launches (updated)

Tesla uses over-the-air software updates to quickly add new features to its electric cars, or to correct problems without the need for customers to bring their cars to service centers. But one apparent software update rankled more than a few Tesla owners: it limits the number of times Launch...

www.greencarreports.com

Tesla drops range on battery packs via software:

Tesla owners see battery range drop after software update to 'improve longevity'

Many owners of older Tesla Model S and Model X vehicles are reporting their max range dropping significantly after a...

electrek.co

All but one of its existing features have been in widespread use nation-wide since roughly 2017.

And they're all L2- something that nobody regulates in the US at all.

But this is the crux of the problem- in 2017 there was EAP and FSD. FSD was clearly sold as hands off, no nags, drive anywhere, 2X as safe as a driver. It appears for you FSD is what it does today,. For someone that bought in 2018, all the stuff it does today is part of EAP except stop lights/signs. The question is when are they ever going to get close to delivering what was advertised in 2016, 2017 and 2018?

 

[Soda Popinski]

Yes they currently use Bosch MRR and Continental ARS 410. That's what their fleet comprises of.
Which part of that don't you understand?

Knightshade is talking about what Tesla is currently producing, and you are talking about what is currently in the existing Tesla fleet, which must also be capable of FSD at some point according to Tesla.

 

[Knightshade]

Nothing in Teslas official documentation around FSD says L4. Nothing during the purchase process promised those folks L4.

Yes, it did.

Screen shots of what was promised pre 3/19 buyers appear earlier in the thread for your reference.

They describe a system that is, at a minimum, L4 per SAE standards.

On the other hand What Elon has repeated in press releases, including official publication is that the car is L5.

Can you cite these "official publication" press releases?

WTF they use the radar. The radar output in those teslas are not ignored.
Whats with Tesla fans and spreading misinformation?

Once again you accuse someone posting facts of spreading misinformation- while doing exactly that yourself.

Your misinformation is out of date. Tesla previously announced they would NOT be using radar in FSD.

FSD Beta V9.0. Step change improvement is massive, especially for weird corner cases & bad weather. Pure vision, no radar.

Guess this is yet ANOTHER bit of info you missed.

Did you just arrive from 2015 or something? That would explain a lot

 

[Knightshade]

Launch mode (part of the Ludicrous paid upgrade) was limited in how many times it could be used via a SW update that was not present when the car was delivered, nor indicated in any documentation by Tesla:

...you know they removed that limit a little while later, right?

But this is the crux of the problem- in 2017 there was EAP and FSD. FSD was clearly sold as hands off, no nags, drive anywhere, 2X as safe as a driver.

Not quite.

Drives ALMOST anywhere.

It literally included that wording.

That's why what they promised is an L4 system, not L5.

It appears for you FSD is what it does today

Again you appear to be attributing something someone else (I guess?) said to be something I said. I never said that.

,. For someone that bought in 2018, all the stuff it does today is part of EAP except stop lights/signs.

Well, and the free HW3 upgrade that EAP buyers in 2018 don't get so the visualizations and stuff that come with that.

I can't recall if the (relatively new) speed limit reading/reacting is HW3 exclusive or not too, though it's not an explicitly listed FSD or EAP feature either way.

The question is when are they ever going to get close to delivering what was advertised in 2016, 2017 and 2018?

I agree that's a question. (though it's only fully owed to the folks who bought before they changed what they were selling in early 2019)

"soon" is the perpetual answer it seems.

Hence why I pointed out the new risk language in the 10Q.

 

[gearchruncher]

Well, and the free HW3 upgrade that EAP buyers in 2018 don't get so the visualizations and stuff that come with that.

Some EAP cars are HW3 from the factory. They threw HW3 in my Model X when it was in for service once (didn't ask, no FSD). It has the new visualizations. It's not tied to you paying for FSD, it's tied to what computer they have. They never advertised visualizations as part of EAP so they don't "owe" it to everyone but they also don't block it.

...you know they removed that limit a little while later, right?

Yes, after a huge uproar from customers. But it's clear that they can do it, have done it, and customers being pissed may not be enough for them to turn something back on in the future, especially if it becomes clear it's a liability risk like FSD behaviors easily could be. They did not go back on the battery range change.

Agreed on FSD being advertised is an L4 system. But "soon" has been so long now, that the most likely outcome of this is that it's functionally never in the lifetime of these cars, and I think it's important that someone buying in 2021 knows that.

As for the 10Q- Damages could be more than the $0-$5000 that 2016-2019 buyers paid for FSD. The car was advertised as capable even if FSD wasn't purchased. People may be able to demonstrate damages greater than the amount Tesla charged for the SW feature. If I sell you a $5000 computer with a $1000 video card option that I promise can run Flight Simulator 2020, if it turns out not to run the software it doesn't mean I only have to take back the video card and claim that because the rest of the computer can run chrome it's all good. I have no idea how that will go, but I have to imagine some people/lawyers will argue that even for people that never paid for FSD. The car was advertised as capable in the most basic configuration.

 

[Bladerskb]

My apologies if whatever you have going on with the author of the page got in the way of that fact.

I have no problem or quarrel with anyone. Other than the fact that Tesla fans continously spread misinformation and balant lies.

"Custom Neural Network Chip
The Winner of Autonomous Driving
Currently $200,000 Waymo vehicles are giving autonomous rides to people in Arizona. The trunks of these cars are filled with computer servers that eat up massive amounts of energy. This means that the trips these cars provide are limited to 30 minutes, and are not able to be ‘go anywhere’ robocars."

Why Tesla Will Dominate Autonomous Driving

Autonomous driving is the next frontier in technology. We believe there's a clear winner in the race.

www.voltequity.com

I cited that link because it contained links to several ACTUAL sources-- including the specific, peer reviewed study from 2019 proving that vision, using years old mid-range cell phone processing, is good enough to do distance within a few percent of LIDAR.

And I showed you results of the latest's peer review paper and its not yet close.

So modern dedicated hardware running dedicated large NNs, for that purpose certainly can.

The 2019 study used none of those, it ran DSCnet, and then a "lite" version that was less resource intensive, as it was running on VERY weak HW.

With our
own CNN framework running on a garden-variety 4-core
ARM A72 processor (found in millions of smartphones today) and without using any type of GPU or accelerator
The study found it was within 4.6% absolute. Slightly over 6% when they dropped to the "lite" version of the code to reduce resource use- but again the compute they had available for this was massively less than Tesla has available for this.

So when in my original post I say the difference was a few percent even running on very weak HW- it factually was

The inference hardware doesn't matter because models don't scale with inference hardware You can increase parameters and use more complex architecture, diverse dataset, more data which require stronger hardware. However you still only get a percentage point give or take if you are extremely lucky. Which still fails to reach Lidar accuracy. You don't extrapolate and say that a scaled down model reach 80% so if you just scale up the model with 10000x more hardware, you get 99.99%. It doesn't work like that.

No, they're using vision point clouds.

You misunderstood my question. You train pseudo lidar models using photometric loss using self supervision methods but you can also increase its accuracy by using priors (lidar data) in a semi supervised way. If Tesla's Vidar is trained using only photometric loss, it will be inferior to Vidar models trained with camera and lidar data.

Watch this video, it will explain it.

But here, maybe this will set you off a bit less.... it's an excerpt from the Karpathy stuff really showing the point cloud stuff they're doing with the cameras for depth.

https://twitter.com/x/status/1120815737654767616

Everyone uses pseudo lidar (Vidar) for their camera system. For mobileye this is already in production on the EyeQ5 well before Tesla. Infact its the foundation of the door to door L2 system anywhere in china on the 2021 Zeekr. Its called FalconEye ViDAR (visual detection and ranging) system. The system uses 11 cameras, however the car does has one forward radar.

Toyota

Waymo

Mobileye (11m26s)

 

[Knightshade]

I have no problem or quarrel with anyone. Other than the fact that Tesla fans continously spread misinformation and balant lies.

Except everything you've claimed is a lie so far turned out to be factrually correct.

Only lies I've seen are:

YOU claiming they still intend to use radar for FSD in the future, then being directly proven wrong but not admitting you were wrong.

YOU claiming the radar on current Teslas is the Bosch unit from 2010, then being directly proven wrong but not admitting you were wrong.


YOU claiming Tesla in "official publications" says FSD is L5, then when asked to support your claim, ignoring the question because your claim is also not true.


And so on...


Again I've deleted your 73 page rant about that website I only provided because it had links to the peer reviewed papers being discussed (and I said that was why I linked to it) but you keep being fixated on the websites own conclusions instead of the thing I was actually talking about.... except for the couple of places you again get your basic facts wrong-

And I showed you results of the latest's peer review paper and its not yet close.

It was within 4.6% in 2019 using basically a cheap cell phone as a processor.

What is "not yet close" exactly?

The inference hardware doesn't matter because models don't scale with inference hardware You can increase parameters and use more complex architecture, diverse dataset, more data which require stronger hardware. However you still only get a percentage point give or take if you are extremely lucky.

Again this is factually untrue.

Simply using the lite version (for only marginally less capable HW) versus the regular (again cheap cell phone quality HW) on the non-lite one got them a difference of roughly 2.0%

Not "a percentage point if you are lucky"

And even then the differences between vision and lidar were already only single digits so it's not like you had a 20% gap to close anyway.


And Teslas HW is many many orders of magnitude more powerful than either thing

Which still fails to reach Lidar accuracy. You don't extrapolate and say that a scaled down model reach 80% so if you just scale up the model with 10000x more hardware, you get 99.99%. It doesn't work like that.

Nor did I use those nonsense numbers.

I pointed out a cheap 2019 cell phone CPU got to about 95% of LIDAR. In 2019.

With a fairly tiny set of training data and fewer cameras too.

So expecting HW that's vastly more powerful, with more self-supervised training from more inputs, and with vastly larger NNs running should get you significantly more accurate results.

And even posted a video, from Tesla, showing some of their results with exactly that.

Karpathy himself calls this out at an ML conference last year-

Here's Karpathy demoing this for you-including explaining the self supervised learning the multiple cameras enables

He also mentions they've found the remaining gap between lidar and this approach has kept shrinking smaller and smaller in Teslas own work (and reminder, a year before the video it was already only about a 5% gap with crap hardware, and THIS video where he mentions the continually shrinking gap is itself over a year old now)

 

[rxlawdude]

New Bosch Mid-range Radar Sensor

Bosch is currently readying a new mid-range radar sensor (MRR) for series production. It works in the same way as the current LRR3 long-range radar sensor in the 77 GHz frequency band, which means it can record the distance and position of obstacles with great precision, but is far less costly...

www.electronicspecifier.com

Now, document where that radar is what is installed in 2015-2021 Teslas.

 

[rxlawdude]

Anyone who bought a AP2.0 tesla before August 2017 has the bosch radar. A car Tesla/Elon claims will be L5!
Stop spreading misinformation!

So you compound the lie. Lovely. Tag team of frick and frack.

 

[rxlawdude]

They threw HW3 in my Model X when it was in for service once (didn't ask, no FSD). It has the new visualizations. It's not tied to you paying for FSD, it's tied to what computer they have.

I'd love to read the mind of service folks when you roll in.

 

[Bladerskb]

Except everything you've claimed is a lie so far turned out to be factrually correct.

Everything I have claimed to be a lie IS a lie

YOU claiming they still intend to use radar for FSD in the future, then being directly proven wrong but not admitting you were wrong.

No where did I make such statement. Infact i said in the post.

"The argument debates devoid into radar is awesome. Funny now that Tesla is not using radar, the new updated page will now say "radar is a stupid crutch". Proving there's no intelligent thought, its just people regurgitating whatever Elon tells them."

I examined the absolute nonesense article you asked me to read because you thought it would educate me. Yet every single sentence in that article was false.
Yet this is where you get your information. You pass off the article to others and spread misinformation.

Using Tesla fanatics logic: Who cares about the future? FSD Beta is fake and not real. A marketing gimmick.
Until real regular customers has it in their hands. Stop posting about some mystical software version 9.

To be frank, to date, the only one who has a vision only system is Mobileye.

YOU claiming the radar on current Teslas is the Bosch unit from 2010, then being directly proven wrong but not admitting you were wrong.

Teslas currently uses Bosch and Continental units and currently supports both. You can't face that fact.
Tesla uses radar and those radar includes Bosch and Continental as confirmed by @Cheburashka
Till there is a software release to all customers which @verygreen can get his hands on and analyze.
I don't believe any marketing nonsense elon pushes.

YOU claiming Tesla in "official publications" says FSD is L5, then when asked to support your claim, ignoring the question because your claim is also not true.

You claim that Tesla has claimed L4, where is it? This IS NOT L4. This doesn't say the functionality doesn't require the human driver in the driver seat paying attention. It just says no input.

It was within 4.6% in 2019 using basically a cheap cell phone as a processor.

What is "not yet close" exactly?





Again this is factually untrue.

Simply using the lite version (for only marginally less capable HW) versus the regular (again cheap cell phone quality HW) on the non-lite one got them a difference of roughly 2.0%

Not "a percentage point if you are lucky"

And even then the differences between vision and lidar were already only single digits so it's not like you had a 20% gap to close anyway.


And Teslas HW is many many orders of magnitude more powerful than either thing




Nor did I use those nonsense numbers.

I pointed out a cheap 2019 cell phone CPU got to about 95% of LIDAR. In 2019.

With a fairly tiny set of training data and fewer cameras too.

So expecting HW that's vastly more powerful, with more self-supervised training from more inputs, and with vastly larger NNs running should get you significantly more accurate results.

This isn't how any of this works.

And even posted a video, from Tesla, showing some of their results with exactly that.

And I posted videos showing results from Waymo, Mobileye and Toyota. What's your point?

Karpathy himself calls this out at an ML conference last year-

Here's Karpathy demoing this for you-including explaining the self supervised learning the multiple cameras enables

He also mentions they've found the remaining gap between lidar and this approach has kept shrinking smaller and smaller in Teslas own work (and reminder, a year before the video it was already only about a 5% gap with crap hardware, and THIS video where he mentions the continually shrinking gap is itself over a year old now)

If your logic was right then it would have been solved already instantaneously. cell phone CPU vs 144 TOPs NN processor is like *1000000000x increase. Why is AK saying the gap is closing rather than the gap has closed? The same is the case with everyone else. Why ain't they saying they have solved it. Why is EVERYONE saying Vidar still doesn't match Lidar?

Stop spreading misinformation.

 

[DanCar]

https://twitter.com/x/status/1387897099983990793

Instead of wavy lines, will we get a gray area to denote probablity area, and center is most probable?

https://twitter.com/x/status/1387893690354515970