Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[Usain]

Guess we should see some FSD 12.4.2 videos in the next few hours:

Tesla Finally Launches FSD V12.4.2 to Customers with Vision-Monitoring Changes

FSD V12.4.2 goes to employees, with updates to normal driving, and updates to the language of the driver management system

www.notateslaapp.com

 

[Usain]

Guess we should see some FSD 12.4.2 videos in the next few hours:

Tesla Finally Launches FSD V12.4.2 to Customers with Vision-Monitoring Changes

FSD V12.4.2 goes to employees, with updates to normal driving, and updates to the language of the driver management system

www.notateslaapp.com

Sooner than I thought. Here's one:

 

[Waiting4M3]

Found it:

https://www.nasdaq.com/articles/tesla-tsla-q4-2022-earnings-call-transcript

so it's doable, just with some cost. I suspect if they were commenting as a free upgrade without involving robotaxi application, then the cost may be prohibitive, but if it can enable robotaxi, the calculus may be completely different

 

[EnzoXYZ]

Was posted on the Dutch TMC forum.
BYD has some domain name work to do.

View attachment 1061334

byd.de

Damn it. Thanks, now Google/FB and etc. thinks I'm in the market for a new 'product' yikes!

 

[Singuy]

Well I have worked for a long time on neural networks - and I also have common sense reading of the literature and understanding of model limitations.

Elon admitted that to get big changes in reliability they had to increase model parameter count by 5x.

If that model does not get them to at least 50,000 miles per critical disengagement (and say more like 10k - 20k), then they will need another 5x increase in model parameter count.

HW3 will not be able to handle that increase if it's already near capacity. No amount of additiona smart compression and quantization can handle that.

We got a Tesla employee in da house! Someone who actually is working on FSD using Hw3!

Oh wait, just someone pulling speculations out of his rear end.

 

[jerry33]

so it's doable, just with some cost. I suspect if they were commenting as a free upgrade without involving robotaxi application, then the cost may be prohibitive, but if it can enable robotaxi, the calculus may be completely different

It could be, but if it costs $20,000+ (lots of labour, I'd imagine) purchasing a new Tesla would make more sense even for Robotaxi.

 

[EQC_]

...

Elon admitted that to get big changes in reliability they had to increase model parameter count by 5x.

...

Whether intentional or unintentional, your use of the word "admitted" has a very negative connotation. People "admit" a mistake or a wrongdoing or a problem. It would be much more neutral to just say "Elon said...."

Beyond that, unless I missed some other tweet, Elon didn't even "admit" needing the 5x increase in parameters to get "big changes in reliability" as you said. He may have meant that, or he may not....but Elon didn't specify.

We know that future models will be "better" in a wide range of ways...and reliability will be part of that. But there are many other metrics that might benefit from the the increased number of parameters in the model. For example, future versions will likely auto-select parking spots and/or park in garages without additional driver input. Future version may be able to reverse out of parking spots and/or garages too. Future versions will likely add better interpretation and response to school zones, school buses, and emergency vehicles with lights/sirens. Presumably, this sort of thing, and other new behaviors, would require additional parameters in the model...and such things aren't really what many folks think of as "reliability" and/or "miles between intervention" metrics (although some probably do...).

Elon's full quote:

Sorry for the delay. This release had far fewer interventions, but suffered in driving smoothness.

Part of the issue was too much training on interventions and not enough on normal driving.

It’s like a doctor training too much on patients in the emergency room vs training on preventative care.

Our next-gen AI model after this has a lot of promise: ~5X increase in parameter count, which is very difficult to achieve without upgrading the vehicle inference computer.

That last/bolded part is the relevant one -- Elon said the next-gen version has "a lot of promise"...and that can relate to a wide range of things: maybe interventions, maybe new capabilities, maybe better smoothness and fewer pointless lane changes.

Sorry for the rant...but a lot of people (here, journalists, etc.) do their own interpretation and additions to Elon's statements, and then act like Elon said things he didn't. This goes into the same bin of pet peeves where I also include people claiming Elon made and broke "promises."

 

[insaneoctane]

Guess we should see some FSD 12.4.2 videos in the next few hours:

Tesla Finally Launches FSD V12.4.2 to Customers with Vision-Monitoring Changes

FSD V12.4.2 goes to employees, with updates to normal driving, and updates to the language of the driver management system

www.notateslaapp.com

But I like to wear sunglasses when I drive...

 

[jkirkwood001]

So much petty arguing on a day we're up 6%! It's so tiring when two people make a point on TMC, usually about something speculative which can't be proven anyway, and then hammer away again and again expecting the other person to say "you're right, I'm an idiot, you're so smart." Keep waiting. In silence, please.

 

[GhostSkater]

Elon agreeing with this, but not clear if that means they can cram a bigger NN into HW3 due to this or not

https://twitter.com/x/status/1807746333807124613

 

[SunCatcher]

But I like to wear sunglasses when I drive...

I'm pretty sure there's a way around that problem.

 

[Cult Member]

But I like to wear sunglasses when I drive...

Forget that you 'drive' anymore, you 'supervise'.

 

[cliff harris]

I'm not a computer vision expert but this is my speculation:

The chances that the arrangement of data in the in-car model is perfect for driving right now is minimal. There are definitely areas where its insufficiently trained (lane selection), but we cannot tell (by definition) those areas where its possibly over-trained. If the car is identifying a stop light 99.999999999% of the time, and with a simplified model of stoplight detection its still 99.9999999%*, then they can definitely tweak the model to bias more towards focusing on where its bad, and less on where its already extremely good.
In other words, you don't have to always make a model more data rich at the inference side to get improvements, unless the distribution of inference already perfectly matches the requirements.
In even different words: The inference model has scope to get better, without getting bigger, and thus could still improve on the same hardware, even if its using 100% of the hardware now.

*this would already be better than humans, who definitely run stop lights some times by mistake, or almost do so, but then spot other cars moving, or would have done so, but are behind another vehicle etc.

 

[Musskiah]

I have been using Tesla Vision on my MY (w/USS) since v12.3.6 in early June. This is far superior to a 360 degree camera view from a usefulness standpoint. Sure Tesla could create a bird-eye 360 view of the surroundings in full color. It's (1) not as usefull as a 3D rendering that can be tilted and rotated & (2) too much unecessary information in full color. The heat map overlayed on the 3D layout of the surroundings to indicate proximity is quick and easy to interpret at a glance and intuitive. It is also very convincing evidence that, at slow speeds, Tesla vehicles are already able to predictably navigate their surroundings to avoid obstacles of all kinds. Good luck to "the competition." I have yet to see any other competitor reproduce such a remarkable and impressive visualization. Let's see if Xpeng, NIO, BYD or other Chinese copycats can step to Tesla Vision - I submit not.

 

[Zaddy Daddy]

We got a Tesla employee in da house! Someone who actually is working on FSD using Hw3!

Oh wait, just someone pulling speculations out of his rear end.

Is that it? No substance here. Do you think the parameter count is going up 5x "for fun"?

When V12.5 or 12.6 comes out and does amazingly well - maybe 5,000 - 10,000 miles per critical disengagement (like 20x - 40x 12.3.6), I hope people still realize it will need another 10x-20x to get to robotaxi level.

I really hope that's not a contentious statement here. Robotaxis need to go at least 100,000 miles if not 2x or 3x to be better than humans.

If they need to get to that level - and they needed to 5x parameter count to get part of the way there - surely you would think parameter count would go up even further to get "all the way" there?

Whether intentional or unintentional, your use of the word "admitted" has a very negative connotation. People "admit" a mistake or a wrongdoing or a problem. It would be much more neutral to just say "Elon said...."

Beyond that, unless I missed some other tweet, Elon didn't even "admit" needing the 5x increase in parameters to get "big changes in reliability" as you said. He may have meant that, or he may not....but Elon didn't specify.

We know that future models will be "better" in a wide range of ways...and reliability will be part of that. But there are many other metrics that might benefit from the the increased number of parameters in the model. For example, future versions will likely auto-select parking spots and/or park in garages without additional driver input. Future version may be able to reverse out of parking spots and/or garages too. Future versions will likely add better interpretation and response to school zones, school buses, and emergency vehicles with lights/sirens. Presumably, this sort of thing, and other new behaviors, would require additional parameters in the model...and such things aren't really what many folks think of as "reliability" and/or "miles between intervention" metrics (although some probably do...).

Elon's full quote:



That last/bolded part is the relevant one -- Elon said the next-gen version has "a lot of promise"...and that can relate to a wide range of things: maybe interventions, maybe new capabilities, maybe better smoothness and fewer pointless lane changes.

Sorry for the rant...but a lot of people (here, journalists, etc.) do their own interpretation and additions to Elon's statements, and then act like Elon said things he didn't. This goes into the same bin of pet peeves where I also include people claiming Elon made and broke "promises."

Um, well you may not know but the CEO has stated for many years your HW3 vehicle will turn into a robotaxi someday. I didn't claim that.

Assuredly there will be other features and metrics that improve and do take up *some* of that added complexity, but the majority of the complexity is added to solve the sparse, almost stochastic edge cases in everyday driving. It's just like every other massive deep learning model developed by other tech companies. They keep getting bigger to get better.

Tesla is not some magic unicorn tech company that is going to toot out the same sized model with 20x the performance. That is what HW3 Truthers are now betting on.

Elon agreeing with this, but not clear if that means they can cram a bigger NN into HW3 due to this or not

https://twitter.com/x/status/1807746333807124613

Likely not much more.

Keep in mind, this really doesn't affect the company - all real potential robotaxi revenue is going to be gated by the robotaxi vehicle and the infrastructure that has to be built to support it. And FSD development is a positive toward this goal overall.

 

[unk45]

I have been using Tesla Vision on my MY (w/USS) since v12.3.6 in early June. This is far superior to a 360 degree camera view from a usefulness standpoint. Sure Tesla could create a bird-eye 360 view of the surroundings in full color...

Many of us remember back when many of us were pleading for 'bird's eye' view even if Tesla had to pay for using the patent. I was thinking of that today, I was one of those, and reflecting on how much more useful the v12.3.6 is than what I had wished for. Since my other non-Tesla does have a bird's eye view I also note that Tesla has now more precise location measurement than ever before. AutoPark still is exaggerating each correction but, in my use anyway, is compulsive about positioning very precisely in my garage space. Right now the only missing link is the ability to navigate within my garage, where I am on the fourth floor. At this rate even this may come soon.

 

[Opus BC fan]

It is tranquil after hours. I think everyone inclined to gamble on stocks has placed their bets...

 

[hobbes]

I'm pretty sure there's a way around that problem.

View attachment 1061361

Someone else had the same idea

 

[Musskiah]

Many of us remember back when many of us were pleading for 'bird's eye' view even if Tesla had to pay for using the patent. I was thinking of that today, I was one of those, and reflecting on how much more useful the v12.3.6 is than what I had wished for. Since my other non-Tesla does have a bird's eye view I also note that Tesla has now more precise location measurement than ever before. AutoPark still is exaggerating each correction but, in my use anyway, is compulsive about positioning very precisely in my garage space. Right now the only missing link is the ability to navigate within my garage, where I am on the fourth floor. At this rate even this may come soon.

I don't have Tesla Vision on my 2021 MX (vertical screen). Super bummer.

Do you mean you want to navigate your car in your garage while you are on the fourth floor? Or your garage is on the fourth floor?

 

[Chunky Jr.]

New head of production at Giga Berlin...

Tesla has taken an important step to optimize its production processes at Giga Berlin. The company has appointed an experienced manager, Christian Adrian, as its new head of production.
...
The new leader has already been working at Tesla since October 2020. He will succeed Heiko Steinmetz, who resigned in mid-May for personal reasons.
...
The reorganization follows recent personnel changes and is aimed at improving collaboration within teams. “We wanted to bring the teams closer together by bringing together the powertrain and vehicle organizations,” Thierig explained. These measures should help improve efficiency and achieve the company’s long-term goals.
...

Tesla appoints new head of production of Giga Berlin

Tesla has taken an important step to optimize its production processes in Giga Berlin. The company has appointed an experienced manager, Christian Adrian, as its new head of production.

driveteslacanada.ca