Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
It appears that the "tester" group are not being updated to 2024.14.x . You own update history would tell if you are in the tester group or not. The testers and the main branch have had different sequences of updates. The latest tester update appears to have been 2024.3.25 and they appear to being held there.

If the pattern continues, several releases of 12.4.x will go to testers before the main branch receives a version of 12.4.

Time will tell. But Tesla won't tell before time. ;-)
I noticed that black tesla got updated to 2024.14.x. Isn't he one of the original testers?

 
Had a new behavior for me this morning - was in the right lane for an upcoming freeway entrance. Traffic was backed up a little, and there was a strip mall with a few cars waiting to turn right. FSDS slowed down, and left a big gap for a waiting car, and once he went, FSDS pulled up and filled in the gap. Was a very polite thing to do.
I've seen (as others) this behavior. It was one of the nice improvements in V12, it isn't perfect but does pretty well.
 
Isn't there a rumor that Tesla already has a Robotaxi specific version of FSD?
If Robotaxi is a much different form factor (smaller) than S3XY, it would presumably require its own specific network, similar to how Cybertruck apparently requires it. Also, if the Robotaxi sensor suite is significantly different from S3XYC (e.g. +radar, +lidar, +ultrasonics, +more/different cameras), it would need to be specifically trained. My own feeling is that Tesla will have to significantly overbuild the Robotaxi sensor suite in order to have any hope of getting it ready on anything like Elon's timeline. Pure vision non-geofenced L4 (on a HW4-equivalent sensor suite) is still at least 6-8 years away I think, if it's even possible at all. (99.99% reliability is likely achievable, but getting from there to the required 99.99999% is not obvious at all.)
 
additional training can make these decisions more sharp / opinionated with a higher drop-off to the next best action, e.g., 70%/30% is a 40% difference instead of 60%/40%'s already 20% difference
Yeah just wonder what happens when they do all that and get a super sharp 50/50. I guess eventually it becomes a very rare event.

Who knows. I can’t imagine it working any time soon given the extremely primitive nature of AI right now. Got a decade or two probably to really be good (starting to realize my 5-10 year prediction which only has 5 years left was fantastically optimistic).

Humans are insanely good. It’s just so impressive.
 
Last edited:
  • Like
Reactions: OxBrew and Mardak
So it’s probably supported?
Seems to be yes but unclear what's the current success rate for correctly responding to painted turn arrows, lane use control signs, other signs like no-right-on-red or right-lane-must-turn or more broadly any visual information including hand signals or face/body orientation, etc.
Presumably initial 12.x releases focused on driving fundamentals like keeping distance from objects, stop-and-go, etc. and these examples happened to also include road markings, signs, etc. resulting in some support resulting in non-zero / low probability of success even without dedicated training. Potentially focused training on certain scenarios is what Elon Musk was referring to with "three significant improvements to FSD" -- perhaps the high level directive to the team even if the actual number and timing is off. My guess is that this focused training would be in addition to general training incrementally reducing interventions on a wide range of scenarios.
One hope is that even if the current probability of success for any particular scenario is low, it should mean that the current neural network architecture can support the correct behavior and "just" needs more training to be more consistent. Although it's a separate question of whether the neural network architecture has capacity to be consistent across all scenarios as well as if there are better architectures that allow for faster learning, better generalization, etc.
 
Any good theories why they removed TACC with from v12? Will it return in a later version? Seems like a horrible decision to remove it.
Several threads stated about it. Here is one you can rummage through.

 
Any good theories why they removed TACC with from v12? Will it return in a later version? Seems like a horrible decision to remove it.
It isn't removed, you just can't use TACC if you have FSD enabled. The reason being is that mode confusion has caused a lot of accidents where people thought they double-pulled to get into FSD, but they only got one, or TACC was enabled on the first, and FSD rejected the second leading to accidents. (Also, that taking steering control would only stop Autosteer, leaving TACC engaged.) Their solution was to force the single-pull activation mode for FSD to eliminate mode confusion. (One of the things that NHTSA has been concerned about.)

It is very possible that they will force single-pull for basic Autopilot as well, given that NHTSA is complaining about mode confusion again. If they do, I hope they make an alternate way to enable TACC without having to switch your profile configuration. (But with limited buttons/stalks available it could be challenging to come up with a good alternate activation method.)
 
Took my first drive on 12.3.6. Did a 40 some mile loop that has interesting problems. Also added the post office at the beginning. The only time that I had to take over. It tried to go in the out driveway. You know how post offices are!

On a highway on version 11 (I'm assuming it's that since the speed limit indication changed to absolute rather than auto) it did the same thing I've done a number of times. A right turn that's hard to see and have to slow down suddenly once you do see it.

Then I went in an area that has a number of hills with very sharp crests. Two of them have a stop sign shortly thereafter. Amazingly the thing slowed down for each of the crests and slowed down before it even got to the stop signs. In both cases almost exactly as I would have. Perhaps even more conservatively. (I have it set to average.) I'm assuming it must have the stop signs from the navigation, rather than seeing it.

It did a great job of managing the speed. I thought I might be adjusting it by pushing up and down on the stock, but I don't think it had anything to do with it.

Overall, pretty amazing! World's better than 11, as everyone has noted.
 
I noticed that black tesla got updated to 2024.14.x. Isn't he one of the original testers?
That is his wife's Model X. He has used it in the past to demo current software updates past the FSD Beta. So his S is still on the early path 24.3.25.

Screenshot 2024-05-29 at 5.23.41 PM.png
 
I am currently a subscriber. I did previously subscribe last summer, 12 months ago and let the sub lapse.
Recently my three Ts have been getting the latest code.
There is your answer. You can't let it laps AT ALL if you want to stay early access. Tesla probably uses this as a measuring stick as to how interested and commented to testing you are. Also maybe a little reward for paying the price too. So you will now heave to wait until 12.4.x is released as the standard FSDS.