Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

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

This site may earn commission on affiliate links.
The situation now feels like: Tesla did exaclty know what they wanted to do for a year, but couldn‘t, due to lacking training capacity. Now the training capacity is online and all the queued tasks get done.
Tesla often know what to do, and then do it, but it takes time, and investors get panicky during the process. Everyone panicked that the cybertruck was taking too long, everyone panicked that 4680 ramp was 'disappointing'. The same is true of FSD, only more so. I expect to see the same with the bot. In 2025-2026 this forum will be filled with people bemoaning that the bot is a failure, and there are still none of them making cars. Then maybe 2027 the stock will double again and the teslabot will 'from out of nowhere' be a huge hit.

But specifically regarding training capacity: Maybe, but also there is a lot of other stuff you need to line up. Tesla have messed around with various revisions of hardware, and placements of cameras, and different algorithms before they got everything just right, and THEN, they needed tons of compute. I've been a coder for 43 years, and the first half dozen attempts to design a program and build it are all inferior. Tearing it down and rebuilding is an inevitable part of the process. We have watched Tesla go through it very publicly. Now I think the current architecture is production-ready. Good times to hold the stock :D.
 
They should get a message on their phone via the app telling them to close it...

If they don't do that, then ban them from booking Robotaxis in future.....

Next step is the car needs to plead with nearby pedestrians to close the door :)

If none of that works, it is a difficult problem, but out of the hypothetical 1000, cases per day the solution above probably works for at least 900, for the other 100 a Robotaxi employee or Optimus needs to travel to the scene and fix the problem.

If the car can't leave a parking spot because a door is open, there is a good chance someone else wants that spot and will close the door. If no one wants the spot the the car waits for Optimus to jog over, preferably in 1980s jogging gear with a headband,. :)
This is one of the many reasons I contend that the early robotaxi trials will be geofenced. At the end of the day, someone has to physically attend to the robotaxi if there is a problem.

It's also one of the reasons why the Gen 3 vehicle will be better suited to the task. For current Teslas, they were designed with hardware that would one day give them the ability to drive autonomously. But there is more to robotaxi than autonomous driving. For Gen 3, Tesla has begun to design solutions for those issues arising from an unattended vehicle, such as automatically closing a door that was left open. Wireless charging will almost definitely be included in Gen 3 as well.
 
I think there's a possibility that Dojo can be imported into China for Tesla use only as it's the only US owned enterprise. High chance they can get away with it. Also technically they are not selling chips to China as Dojo is not being sold to anyone except it's hardware developed inhouse. I think this will give Tesla a huge advantage over peers in China when it comes to end to end training(after they steal Tesla's homework) since no one is allowed to buy Nvidia hardware.

I don't think that's correct?

the export restrictions said:
(a) General prohibition. In addition to the license requirements for items specified on the CCL,
you may not export, reexport, or transfer (in-country) without a license any item subject to the
EAR described in paragraphs (a)(1) through (4) of this section when you have “knowledge” at
the time of export, reexport, or transfer (in-country) that the item is destined for a destination, end use, or type of end user described in paragraphs (a)(1) through (4) of this section, unless excluded by paragraph (a)(5) of this section.

See the bold- it doesn't appear to care who owns the company in China, just that it's going TO China.

Further down it defines "destination" as "located in or destined to China or Macau" again without regard to who owns the HW or the company.

The one exclusion paragraph only excludes one specific aspect of the production definition not the use or destination parts.

Full rules here:




it's also not correct nobody in China can buy Nvidia HW-- they just can't buy the most advanced stuff--- Nvidia is already making toned-down versions that are exportable- this obviously has the negative impact of needing more of it to get the same amount of compute.
 
Last edited:
I don't think that's correct?



See the bold- it doesn't appear to care who owns the company in China, just that it's going TO China.

Further down it defines "destination" as "located in or destined to China or Macau" again without regard to who owns the HW or the company.

The one exclusion paragraph only excludes one specific aspect of the production definition not the use or destination parts.

Full rules here:




it's also not correct nobody in China can buy Nvidia HW-- they just can't buy the most advanced stuff--- Nvidia is already making toned-down versions that are exportable- this obviously has the negative impact of needing more of it to get the same amount of compute.
Didn't they ban the toned down version as well?
 
If Tesla has managed extreme improvements from V12.3.1 to V12.3.2 I wonder how they did it. We know that they are no longer compute constrained from Elons tweets. But how would this have worked? Suddenly 10k H100 came online at once? Doesn't seem plausible. And it's not only that the improvement is impressive, it also seems that the interval between releases has gone from 1 month to 2weeks. Maybe Dojo got a leap in performance or a large number of them were made and installed at the same time, that seems more likely to give a step response in performance than nVidia chips.

Maybe the new system just need failure cases from the real world to rapidly fix these from release to release.

Or maybe they found some nice tweak to the cost function for the end2end neural network.

Anyway this new rate of improvement is imo the interesting thing, not the performance of V12.3.2. Think of where this is going over a year with FSD, optimus and alien dreadnaught. That's the important question...

Well shoot. NOW what are we gonna tease Elon about??
 
Ladies and Gentlemen, I saw our first Cybertruck in the parent community in carpool this morning!

Everyone was in awe. :)

giphy.gif
 
If Tesla has managed extreme improvements from V12.3.1 to V12.3.2 I wonder how they did it. We know that they are no longer compute constrained from Elons tweets. But how would this have worked? Suddenly 10k H100 came online at once? Doesn't seem plausible. And it's not only that the improvement is impressive, it also seems that the interval between releases has gone from 1 month to 2weeks. Maybe Dojo got a leap in performance or a large number of them were made and installed at the same time, that seems more likely to give a step response in performance than nVidia chips.

Maybe the new system just need failure cases from the real world to rapidly fix these from release to release.

Or maybe they found some nice tweak to the cost function for the end2end neural network.

Anyway this new rate of improvement is imo the interesting thing, not the performance of V12.3.2. Think of where this is going over a year with FSD, optimus and alien dreadnaught. That's the important question...
You're seeing some significant improvement, but it's at a time where significant code changes are being merged in. Based on the historic pattern they will finish the major merges, and then run into another wall/white trailer.
 
  • Like
Reactions: saniflash
FSD deploys in another country and stock goes down. 🙃

I bet (think) it's folks exiting before the Q1 deliveries burn. They seem to exit earlier each time for a head start on the action.
I guess the market doesn´t immediately translate beta testers in one more country into profits. We are still very much in a Show Me The Money phase in general and in particular in relation to Tesla, which has failed to deliver on its communicated targets for quite some time now. Compare that with NVIDIA where surprises on the upside seem to be the recurring pattern (so far at least).
 
FSD deploys in another country and stock goes down. 🙃

I bet (think) it's folks exiting before the Q1 deliveries burn. They seem to exit earlier each time for a head start on the action.
Just why?

You know the SP moves based on the options market and current manufactured narrative. And the only reason they bother with the latter is as a veiled attempt to cover up their shenanigans.
 
Just why?

You know the SP moves based on the options market and current manufactured narrative. And the only reason they bother with the latter is as a veiled attempt to cover up their shenanigans.
Plus the people here who are "waiting for 140"... no?

Edit: Unless they are part of the cover up. :eek: Right on this thread, can you believe the nerve of these people!
 
Last edited: