Autonomous Car Progress

[stopcrazypp]

Do people actually think there are ancient AP 2018 code being used by Xpeng's SOTA system right now? Seriously? The Tesla software in 2018 doesn't do 1% of what Xpeng software does today.

Its like accusing someone who comes out today with a legit IPhone 15 competitor for stealing the source code from the original IPhone.
Or me getting an F on a test and accusing someone who got a A for copying me.

The Tesla 2018 code base and Xpeng's current system codebase are literally worlds apart. I hate that this keeps getting brought up. 5 years from now people will still be bringing this up as cope. As though Xpeng is using code from 10 years ago.

I'm not making any claims about today's code, given a long time has passed since then (Tesla's code is also way different from 5 years ago), just pointing out that the point about lidar doesn't make logical sense as an explanation that the code was useless to XPeng back then given XPeng still used cameras.

I should note however, functionality wise, NOA was already released in late 2018.
Introducing Navigate on Autopilot | Tesla

XPeng did not release their equivalent NGP until early 2021.
XPeng To Release Navigation Guided Pilot (NGP) in Q1 2021

 

[hooty]

this keeps getting brought up. 5 years from now people will still be bringing this up as cope.

Nailed the psychology.
It’s borne from an uncritical acceptance of the words from a perceived hero

 

[diplomat33]

Mobileye delivered an informative keynote on their AV approach and progress.

 

[Bladerskb]

SMART licenses Mobileye Supervision. Special Edition Model Launching in Germany in Q4.
Still no licenses by Tesla.

Navigation smart Pilot (NSP) Set to Launch in Q4 smart to Release New Special Edition Model Equipped with Mobileye SuperVision™ System

On August 25th, smart officially unveiled its ADAS evolution strategy roadmap. The smart Pilot Assist 2.0 version, which includes the Navigation smart Pilot (NSP) feature, is planned to launch officially in the fourth quarter of 2023.The special edition model, equipped with 12 cameras (12V)...

media.smart.com

 

[flutas]

Do people actually think there are ancient AP 2018 code being used by Xpeng's SOTA system right now? Seriously? The Tesla software in 2018 doesn't do 1% of what Xpeng software does today.

Its like accusing someone who comes out today with a legit IPhone 15 competitor for stealing the source code from the original IPhone.
Or me getting an F on a test and accusing someone who got a A for copying me.

The Tesla 2018 code base and Xpeng's current system codebase are literally worlds apart. I hate that this keeps getting brought up. 5 years from now people will still be bringing this up as cope. As though Xpeng is using code from 10 years ago.

If you think 5 year old code is so irrelevant, then ask your director to open source your code from 5 years ago. I work in SW engineering as well (Sr, just like you). I know the reaction from my boss, even though we are literally a month away from pausing work to rewrite our project from the ground up. Just because code is "old" doesn't mean it's not relevant and can't be learned from or just straight copied.

 

[Bitdepth]

If you think 5 year old code is so irrelevant, then ask your director to open source your code from 5 years ago. I work in SW engineering as well (Sr, just like you). I know the reaction from my boss, even though we are literally a month away from pausing work to rewrite our project from the ground up. Just because code is "old" doesn't mean it's not relevant and can't be learned from or just straight copied.

What i find fascinating is that 3 years ago, Xpeng were doing things Tesla couldn't do. Tesla Auto Park couldn't park using just lines in a parking lot, it required cars beside it to park. Xpeng also has Valet parking where the car can remember where you parked in a parking facility and auto park whenever you drive into the parking lot. All they needed to do was look at what Tesla couldn't do and offer better. No source code needed. Source code from what Tesla was doing in 2018 would have no use to what people are doing today in terms of machine learning.

 

[stopcrazypp]

What i find fascinating is that 3 years ago, Xpeng were doing things Tesla couldn't do. Tesla Auto Park couldn't park using just lines in a parking lot, it required cars beside it to park. Xpeng also has Valet parking where the car can remember where you parked in a parking facility and auto park whenever you drive into the parking lot. All they needed to do was look at what Tesla couldn't do and offer better. No source code needed. Source code from what Tesla was doing in 2018 would have no use to what people are doing today in terms of machine learning.

Tesla's parking functions were never advanced, there were plenty of existing auto park implementations that were more advanced. I don't think the market really cared however.

The feature of interest would instead be AP and NOA, which the rest of the industry had not replicated yet at the time, but Tesla released in late 2018. As noted above, XPeng did not release an equivalent NGP feature until early 2021.

So it's fairly specious to suggest the AP code would have useless to XPeng at the time, simply based on them having some more advanced parking functions.

 

[flutas]

What i find fascinating is that 3 years ago, Xpeng were doing things Tesla couldn't do. Tesla Auto Park couldn't park using just lines in a parking lot, it required cars beside it to park. Xpeng also has Valet parking where the car can remember where you parked in a parking facility and auto park whenever you drive into the parking lot. All they needed to do was look at what Tesla couldn't do and offer better. No source code needed. Source code from what Tesla was doing in 2018 would have no use to what people are doing today in terms of machine learning.

You do know that most of the NoA stack isn't ML though right? You're trying to narrow down the claim there by moving from NoA like features, to "machine learning."

They use ML for perception of the world around them, but NoA is 300k+ lines of hand-crafted rules in C++. THAT is valuable, because it gives them a foundation to work from without having to invest the time to understand every situation, Tesla already did that for them.

Side note, 2018 < 2020, So I don't understand why you think "All they needed to do was look at what Tesla couldn't do and offer better." when most of the functionality you list wasn't available before the supposed source code theft. Take the stolen work, bolt on a few features as best you can and suddenly it's doing things Tesla "couldn't" (or in reality they probably realize don't have the mass market appeal in the US and aren't worth it).

 

[Bladerskb]

If you think 5 year old code is so irrelevant, then ask your director to open source your code from 5 years ago. I work in SW engineering as well (Sr, just like you). I know the reaction from my boss, even though we are literally a month away from pausing work to rewrite our project from the ground up. Just because code is "old" doesn't mean it's not relevant and can't be learned from or just straight copied.

It is irrelevant, this isn't a phone app or a website where old code can be reused or "straight copied".

You do know that most of the NoA stack isn't ML though right? You're trying to narrow down the claim there by moving from NoA like features, to "machine learning." They use ML for perception of the world around them, but NoA is 300k+ lines of hand-crafted rules in C++. THAT is valuable, because it gives them a foundation to work from without having to invest the time to understand every situation, Tesla already did that for them.

This is hugely incorrect. NOA in 2018 was the most horrific released software before FSD Beta. I find it funny how people always talked about how amazing NOA had gotten better. But never actually watch the old videos to see how bad it was at release. We even have a whole thread about how useless it was.

For example Tesla's NOA did not use a prediction network, they used a primitive cut in detector. NOA stacks today are night and day compared to what Tesla had in 2018. Complete different architecture and a full stack of prediction networks and partial ML in the planner.

Another example, NOA wasn't "300k+", that's city streets. NOA was orders of magnitude lower.

 

[flutas]

It is irrelevant, this isn't a phone app or a website where old code can be reused or "straight copied".

This is hugely incorrect. NOA in 2018 was the most horrific released software before FSD Beta. I find it funny how people always talked about how amazing NOA had gotten better. But never actually watch the old videos to see how bad it was at release. We even have a whole thread about how useless it was.

For example Tesla's NOA did not use a prediction network, they used a primitive cut in detector. NOA stacks today are night and day compared to what Tesla had in 2018. Complete different architecture and a full stack of prediction networks and partial ML in the planner.

Another example, NOA wasn't "300k+", that's city streets. NOA was orders of magnitude lower.

All this to say "I don't care about IP theft" while simply dismissing arguments as "it's irrelevant."

Again, the strategy here is

1: Minimize the claim ("It is irrelevant", "no use")
2: Deflect to some other slightly related topic, that you made up just so you can win (aka: strawman).

Sounds about right for someone with your level of hubris lmao.

Not even worth arguing when the argument presented is "so what."

 

[Bitdepth]

All this to say "I don't care about IP theft" while simply dismissing arguments as "it's irrelevant."

Again, the strategy here is

1: Minimize the claim ("It is irrelevant", "no use")
2: Deflect to some other slightly related topic, that you made up just so you can win (aka: strawman).

Sounds about right for someone with your level of hubris lmao.

Not even worth arguing when the argument presented is "so what."

You would make sense if there was IP theft. There wasn't.

Tesla sued a former employee for copying source code before they quit which the employee does not deny but denies taking the source code to Xpeng.

Elon Musk claimed Xpeng was using their old source code and a 3rd party audited both source code and found nothing was copied from Tesla. Tesla dropped suit after the former employee settled with them. You don't need to be making up stories. This is not the first time corporations have sued former employees for stealing IP after they quit. Waymo sued Levandowski and Uber for stealing Lidar IP and Uber settled because they in fact did use Waymo Lidar IP taken by Anthony Levandowski after he quit. That has not been proven the case with Tesla however.

 

[Bladerskb]

All this to say "I don't care about IP theft" while simply dismissing arguments as "it's irrelevant."

Again, the strategy here is

1: Minimize the claim ("It is irrelevant", "no use")
2: Deflect to some other slightly related topic, that you made up just so you can win (aka: strawman).

Sounds about right for someone with your level of hubris lmao.

Not even worth arguing when the argument presented is "so what."

As was detailed above. There was no IP theft.
And even then the code in question in today's standard would be rubbish and would not be used.
Its not hubris, its called keeping up with SOTA ML and ADAS and AV Tech stack from alot of companies.

 

[Doggydogworld]

Kyle Vogt did a conference session the other day. Some comments spawned stories in the press. Seeking Alpha has the entire transcript. I highlighted these point on Reddit:

• "largest in terms of fleet size, scaling most quickly"
• Stickiness / retention matches Uber in only 2nd year (vs. 14 years)
• Sensationalist media Human drivers are worse, wah, wah
• Their study showed half as many collisions, 75% fewer severe/injury collisions
• In 91% of their actual wrecks the other party was the "primary contributor"
• Service in 5 cities, testing/ramping ops in 10 more
• Origin "just days away from the last regulatory approval which would let us start production and almost immediately start putting these vehicles on the road."
• Origin designed to last a million miles
• Joystick system to drive Origin off the end of the assembly line (and for stall rescues?)
• "half of the car owners in the U.S., actually their cost is around $1 per mile or more"
• 2nd gen Origin in 2025 with custom silicon and $1/mile unit economics
• "multi trillion-dollar market"
• Working to make mapping easier and cheaper, but also trying to dump "complicated and somewhat expensive maps and instead consume readily available map data"
• "highly disruptive" announcement coming in November. Also more biz model detail.
• "starting in 2025, the hardware produced then is capable of those very attractive unit economics," (but isn't Origin this year/early 2024?)
• Three years to get permits in CA, three days in TX

 

[Bladerskb]

Kyle Vogt did a conference session the other day. Some comments spawned stories in the press. Seeking Alpha has the entire transcript. I highlighted these point on Reddit:

• "largest in terms of fleet size, scaling most quickly"
• Stickiness / retention matches Uber in only 2nd year (vs. 14 years)
• Sensationalist media Human drivers are worse, wah, wah
• Their study showed half as many collisions, 75% fewer severe/injury collisions
• In 91% of their actual wrecks the other party was the "primary contributor"
• Service in 5 cities, testing/ramping ops in 10 more
• Origin "just days away from the last regulatory approval which would let us start production and almost immediately start putting these vehicles on the road."
• Origin designed to last a million miles
• Joystick system to drive Origin off the end of the assembly line (and for stall rescues?)
• "half of the car owners in the U.S., actually their cost is around $1 per mile or more"
• 2nd gen Origin in 2025 with custom silicon and $1/mile unit economics
• "multi trillion-dollar market"
• Working to make mapping easier and cheaper, but also trying to dump "complicated and somewhat expensive maps and instead consume readily available map data"
• "highly disruptive" announcement coming in November. Also more biz model detail.
• "starting in 2025, the hardware produced then is capable of those very attractive unit economics," (but isn't Origin this year/early 2024?)
• Three years to get permits in CA, three days in TX

I believe he’s referring to their own compute chip rather than using Nvidia because they are overpaying them due to low volume

 

[Bladerskb]

Serious question, can 'FSD' even handle this?

 

[Dewg]

Serious question, can 'FSD' even handle this?

No, it can't, but US traffic laws are different.

 

[stopcrazypp]

Serious question, can 'FSD' even handle this?

The system didn't handle that fully either in this situation. They had to use accelerator intervention the whole way, the driver kept saying he had to have his foot on the accelerator the whole way though the pedestrian sections otherwise they would be there the whole day and never be able to cross. The warning message that kept popping up seems to be the accelerator intervention warning (I don't have FSD, but AP has a similar warning when you use accelerator intervention, saying that the car won't slow down automatically while pedal is depressed).

Only on the last right turn into an emptier road did the system find the gap by itself, although he was tempted to intervene.

 

[Bitdepth]

Im sure if FSD beta was trained on driving data from that region it can attempt to handle it. How well is up for debate.

 

[stopcrazypp]

Im sure if FSD beta was trained on driving data from that region it can attempt to handle it. How well is up for debate.

FSD Beta is rumored to start rolling out in China soon, so perhaps we will see people trying in those areas:
Tesla is taking steps to launch FSD Beta in China

Last I checked, Tesla was using Baidu maps as their maps provider in China, which seems to be a subpar provider (most seem to be using Amap instead). Not that FSD Beta here necessarily does a good job of routing in the US (it seems it now uses Mapbox, which is subpar against the leaders like Google).

There are also recent issues in China with Teslas being banned from certain parking lots and data limitations. Article mentions Tesla is unlikely able to gather data relevant to FSD Beta for use for training in US based supercomputers. It's something they would have to work out (perhaps with a cluster in China itself).
Tesla reassures Chinese users on data security amid spying concerns | TechCrunch

 

[Bitdepth]

The feature of interest would instead be AP and NOA, which the rest of the industry had not replicated yet at the time, but Tesla released in late 2018. As noted above, XPeng did not release an equivalent NGP feature until early 2021.

Almost all manufacturers had AP at least some combination of it in certain models and trims of their vehicles. Autopilot is LKA, LCA, TACC, Lane Change Assist etc.

LKA, TACC, Autosteer has been in Mercedes vehicles since 2014.

While I enjoy the way Tesla is very nimble and allows consumers to try out what they are working on like NOA, it does not mean other manufacturers are idle not doing anything. They are very risk averse in terms of introducing features to consumers.

2014 Mercedes working on ADS.

2013 BMW Connected Drive ADS

Just to blow your mind, Toyota in the 90s had voice activated FSD capability. Lol

There are also recent issues in China with Teslas being banned from certain parking lots and data limitations. Article mentions Tesla is unlikely able to gather data relevant to FSD Beta for use for training in US based supercomputers. It's something they would have to work out (perhaps with a cluster in China itself).
Tesla reassures Chinese users on data security amid spying concerns | TechCrunch

Not just Tesla but also iPhones were banned in certain government-controlled spaces as well. They'll have to work with local suppliers maybe or find other workarounds if special permission isn't given to them which who knows they might with Tesla having factories in China.