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

Really incoming FSD 12.1 ?

This site may earn commission on affiliate links.

Twiglett

Single pedal driver
Oct 3, 2014
4,459
5,556
Austin
Seems that Notateslapp has some inside news. . .


And from TeslaScope
Screenshot 2023-12-23 at 21.45.57.png
 
  • Like
Reactions: GSP and Todd Burch
Exciting!

What was the timeline of initial V11 (that brought the FSD stack to highway driving) release for employees until public release? 4 months, if I recall correctly?

I'm going to be pleasantly surprised if any Youtubers get V12 by April.
 
I think the newest event is it’s been moved into 44.30
Just running ahead of us on the outside of Tesla
I was at 38.9 and v12 was in 38.10
Tesla pushed me to 44.x
Now we’re together longer
There must be a plan for v12 and all of us in 44.30
Hopefully soon
 
So employees are testing V12? Will isn't that a case of what we already knew. Elon has already confirmed last month the obvious that V12 was being tested by employees.

View attachment 1002688
According to teslascope, V12 was pushed out to 30x more employees than the previous 100 (~3000 total now). So, the news is that they have greatly expanded the rollout beyond managers and execs. This should put in a lot of cars driven by 'ordinary' drivers.

I wonder if its the same version that went to the first 100 employees. That would be a good sign, though I'm sure there are still a few potholes to deal with.
 
  • Informative
Reactions: APotatoGod
According to teslascope, V12 was pushed out to 30x more employees than the previous 100 (~3000 total now). So, the news is that they have greatly expanded the rollout beyond managers and execs. This should put in a lot of cars driven by 'ordinary' drivers.

I wonder if its the same version that went to the first 100 employees. That would be a good sign, though I'm sure there are still a few potholes to deal with.
Don’t think so
We’re hearing v12.1
 
Initial rollout is not to "managers and execs" .. it's to people connected with AP team and may be they have an early access internal volunteer group. At least that's how we used to do in Microsoft.
Yeah it always seemed odd that Tesla would risk killing or injuring their execs before they risk the peons. I do think you may be more correct.
 
Yeah it always seemed odd that Tesla would risk killing or injuring their execs before they risk the peons. I do think you may be more correct.
It is about dogfooding. You want the actual engineers testing and evaluating because when they see problems they can better isolate what needs to be modified in the code. But again this is just a (hyperbole thread title) pure speculation thread and has 0 relevance on when we get V12. Everyday we are 1 day closer (but still MANY to go) and conjecture or rumors about what is happening internally is just that. So V12.1 has no relation to us and likely won't for a while.

2-3 weeks next thread "12.1.1 or 12.2 is coming" when employees get another updated build to test.
 
Initial rollout is not to "managers and execs" .. it's to people connected with AP team and may be they have an early access internal volunteer group. At least that's how we used to do in Microsoft.
Well, that's how teslascope reported it, as I recall. Nonetheless, it was a very small initial rollout that has expanded. At least, if teslascope has it accurately, and represents some progress.

The concern I still have is if this is an updated version, then it took about five weeks to turn it around. Assuming that this is typical, then predictions of an end of Jan or early Feb rollout to customers is almost certainly overly optimistic, because they would need to release the next update. That doesn't seem very likely.

But, we'll see how long it takes to expand to more employees and whether that will be a new version.
 
I think we should change our expectations for what "testing" means now.

Before, there was a lot of C++ code to tweak. Code written by humans will naturally have bugs, so testing meant exposing those bugs and fixing them. And also uncovering scenarios that required new code to be written.

But with v12, I suspect that "finding an issue" most likely doesn't mean that there's a problem in code anymore. It probably means they found a fairly common scenario that the neural network fails to handle. "Fixing a bug" now means sourcing more data into the training system.

So I suspect the process is now: Field Testing ==> Find Problems ==> Source Data from the fleet ==> Label that data ==> Feed the labeled data back into the next training run==> Field Testing.

I don't know if this means the process will be faster or slower. Certainly as more compute comes online training the network will be faster...but they will also be gathering more training data.

Most importantly, now that we are not writing code and instead just feeding data to neural networks, I think we'll only hit limitations based on neural network size. Something not working well? Give it more data. Still not working well? Give it more nodes/neurons and throw more data at it.
 
Something not working well? Give it more data. Still not working well? Give it more nodes/neurons and throw more data at it.
My gut tells me the edge cases will take longer to resolve - i.e. non-square UPL for example. Maybe not. Surely there will be edge cases.

I'm not sure E2E addresses lane choice today either which appears to be map-based. Lots of issues there which will look like FSD bugs when in fact it's a mapping bug.

I would not expect it to work better than what we have today for maybe 6 months post wide deployment. Maybe longer.

Just my $.02. Caveat emptor.
 
  • Like
Reactions: GSP
My gut tells me the edge cases will take longer to resolve - i.e. non-square UPL for example. Maybe not. Surely there will be edge cases.

I'm not sure E2E addresses lane choice today either which appears to be map-based. Lots of issues there which will look like FSD bugs when in fact it's a mapping bug.

I would not expect it to work better than what we have today for maybe 6 months post wide deployment. Maybe longer.

Just my $.02. Caveat emptor.
Um. Not going to disagree with you that much; this is more of a minor debate. Having taken what you said, however:

This is the point of the Dojo: Run a zillion different highly realistic scenarios, each slightly different, with outcomes that one wants. Yes, one can't run a hundred million scenarios, one runs out of time, and there's probably "sweet spot" testing that clears up a lot of potential bugs.

Still.. this is the glass half empty/glass half full forum.. this may be the breakthrough on training performance that Tesla has been looking for. We can hope, anyway.
 
My gut tells me the edge cases will take longer to resolve - i.e. non-square UPL for example. Maybe not. Surely there will be edge cases.

I'm not sure E2E addresses lane choice today either which appears to be map-based. Lots of issues there which will look like FSD bugs when in fact it's a mapping bug.

I would not expect it to work better than what we have today for maybe 6 months post wide deployment. Maybe longer.

Just my $.02. Caveat emptor.
Yep. There will always be edge cases throwing a ripple in the mix. They are extreme cases and tend to not be weighted as heavily out of regression fears. Time will tell if dojo can do any better but unfortunately it sounds like next gen dojo 2 chip isn't meeting expectations. Gotta wonder if v12 will be safe without a fully functioning gen 2 dojo chip.