That is not correct. ONLY FSD Beta uses the new stack. TACC and NoA use the same legacy stack. When on FSD Beta you can even switch it to the old legacy NoA stack.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
New Tesla/EV owner and first-time FSD beta user - is this a joke?
- Thread starter zroger73
- Start date
That is not correct. ONLY FSD Beta uses the new stack. TACC and NoA use the same legacy stack. When on FSD Beta you can even switch it to the old legacy NoA stack.
EVNow
Well-Known Member
FSD can't solve mapping issues - Tesla needs to start a crowd-sourced way to update maps. But they won't do anything they think can't be easily scaled through automation - so they won't do it.
That's kind of what they are doing right now, no?
The fleet is reporting signs, poor road conditions, new roads, obstacles, wrecks, traffic, etc.
clydeiii
Member
Would you rather the car be too cautious or too aggressive?
Anyway, here's my observations, having been a near-constant FSDb user since late 2021. You install a new version of FSDb and it's super cautious at first. Slowly over the course of a few weeks, it gets less cautious. I don't know why or how this happens, but it definitely does happen, others have mentioned it as well here for years.
Another observation that gives me a lot of hope: I have been on 11.3.6 since it released. For most of those weeks, it was incorrectly stopping at three different intersections in my neighborhood in Baltimore. Like, it would just stop in the middle of the road, in three different spots, very consistently, as if there was a phantom stop sign that only the car could see (very likely bad map data). Despite not getting new map data and despite not getting new FSDb software, the car magically stopped stopping in these three intersections last week and now drives on those three roads perfectly. Presumably some sort of mapping data on the server side has changed here, but can't 100% speak to what changed. All I know is, every single time it stopped for no reason, I reported this as a bug.
The point is: the more you use FSDb, and the more you intervene during your specific commute, the better feedback Tesla gets about your specific route and, over the course of years, your commute will get better and better, but ONLY if you keep using it and reporting the bugs. If you give up, call it junk, say it isn't raodworthy, well, you have no one to blame other than yourself when your commute's route is still buggy in a year.
The folks who post to Youtube, you better believe they report bugs constantly. which explains why their drives look so good.
Would it be great if everything worked perfect out of the box? Of course! But we're on the bleeding edge of a technology no company has been able to truly solve yet, so yeah, you're a beta tester and the more you test, the better it will get. To me, the opportunit to help train the system is easily worth the $200/mo cost.
Anyway, here's my observations, having been a near-constant FSDb user since late 2021. You install a new version of FSDb and it's super cautious at first. Slowly over the course of a few weeks, it gets less cautious. I don't know why or how this happens, but it definitely does happen, others have mentioned it as well here for years.
Another observation that gives me a lot of hope: I have been on 11.3.6 since it released. For most of those weeks, it was incorrectly stopping at three different intersections in my neighborhood in Baltimore. Like, it would just stop in the middle of the road, in three different spots, very consistently, as if there was a phantom stop sign that only the car could see (very likely bad map data). Despite not getting new map data and despite not getting new FSDb software, the car magically stopped stopping in these three intersections last week and now drives on those three roads perfectly. Presumably some sort of mapping data on the server side has changed here, but can't 100% speak to what changed. All I know is, every single time it stopped for no reason, I reported this as a bug.
The point is: the more you use FSDb, and the more you intervene during your specific commute, the better feedback Tesla gets about your specific route and, over the course of years, your commute will get better and better, but ONLY if you keep using it and reporting the bugs. If you give up, call it junk, say it isn't raodworthy, well, you have no one to blame other than yourself when your commute's route is still buggy in a year.
The folks who post to Youtube, you better believe they report bugs constantly. which explains why their drives look so good.
Would it be great if everything worked perfect out of the box? Of course! But we're on the bleeding edge of a technology no company has been able to truly solve yet, so yeah, you're a beta tester and the more you test, the better it will get. To me, the opportunit to help train the system is easily worth the $200/mo cost.
Last edited:
clydeiii
Member
TACC and AP for single lane highway use or very infrequent lane changes worked much better for me before FSDb. The distance pacing and lane change frequency (even with 'minimal' selected) make it borderline unusable IMO. Not the mention the phantom braking frequency uptick. I get the impression I'm an outlier but I get at least one phantom break every 5 interstate miles on avg. Granted this is I-4 and the conditions are probably the among the most challenging in the U.S.
On the other hand, navigation choices and reaction speed on tough intersections have improved remarkably just over the past 6 months. I still don't see a path for it getting to a level where I will "trust" it 100% in the lifetime of my vehicle but we'll see.
On the other hand, navigation choices and reaction speed on tough intersections have improved remarkably just over the past 6 months. I still don't see a path for it getting to a level where I will "trust" it 100% in the lifetime of my vehicle but we'll see.
You use the term "production FSD" when there is no such thing.
So called production FSD is a combination of Autopilot, most accurately, TACC (beta), Autosteer (beta) and Navigate on Autopilot (beta). FSDb is an affection used by nobody but the folks on this forum as shorthand.
The issue is when someone notices how both Autopilot and FSD both say "beta" and both have warnings prior to use, but one is supposed to be "production beta" instead of "beta beta".
Where in the Tesla sales does it make that distinction? It certainly doesn't make that distinction in the manual or on the latest software releases.
When there isn't a difference made plain, then new drivers expect a great deal more polish than is delivered.
New owners also take their cues from company statements and those of the attention junkie CEO who are constantly crowing about just how amazing it is.
To take that from the current user manual
Autopilot convenience features are designed to reduce driver workload: (shortened list here) TACC, Autosteer, Navigate on Autopilot etc etc
Note that the word beta doesn't show up until you dig deeper or look at the screen. But every feature is documented as BETA (in those caps) including TACC.
So after all that - why would a new owner expect such difference when they have been having so much luck with TACC, Autosteer and NoA which they can already see is a beta product.
navguy12
Active Member
I dunno, I thought with my last update that that was the case. Obviously I’m wrong.
But why wouldn’t Tesla have TACC/NOA use the single stack at some near term point?
If people used TACC (with single stack) followed by NOA (with single stack) before going all in on FSDB, perhaps more users would take the FSDB plunge after being comfortable with the first two ADAS iterations.
At one time, that technique was called the building block approach.
And for the record, I tried FSDB on a stretch of HWY 7 westbound from Ottawa a few weeks ago and it was scared of all the pre-leafed out tree shadows making shadows perpendicular to the road direction.
So, I’m not surprised by the OP’s response to his exposures to FSDB.
Last edited:
Because the legacy stack is years old, WELL tested and WELL known (ain't broke so don't fix it with something that may actually be broke, until proven otherwise). The FSD Beta highway stack is VERY untested and novel. Tesla needs a lot of miles and a lot of real world edge cases tested on it and a lot of version improvements. Likely it will replace the legacy NoA stack in a year or two.
clydeiii
Member
Worse, some older Teslas cannot run single stack, while legacy TACC/AP NNs run just fine on HW2.5
navguy12
Active Member
I tried two more trips after work. The first trip involved a traffic circle, which was an epic failure. The car sat there while it was clear then when another car was coming around toward me, mine started creeping forward slowly while rapidly sawing the wheel back and forth grinding my tires into the pavement then when the oncoming car got closer, mine darted toward the center of the circle - not around it. Haha!
On the second drive, FSDb couldn't decide whether it wanted to turn right or go straight, so the projected path kept rapidly flickering back and forth and the turn signal kept turning on and off. Once the light turned green, the car suddenly turned right even though that was not the correct route. And, if it needed to turn right, It had plenty of opportunity to do so since right turns on red were not prohibited.
FSDb works okay in some situations, but those are the same situations where the standard Autopilot works. City driving - at least for me - is so bad that that FSDb has negative value. There's something wrong with almost every maneuver.
Surely, my experience isn't typical? We'll see what 11.4.x brings. So far, all I see is "disengaged - what happened?" at the slight majority of maneuvers and it seems like I spend more time recording what just happened than witnesses the car do something correctly. I do kind of suspect it doesn't know how to handle the non-perpendicular intersections around me in these small towns. Maybe. I dunno. I'll keep playing with it and hope it doesn't kill me.
On the second drive, FSDb couldn't decide whether it wanted to turn right or go straight, so the projected path kept rapidly flickering back and forth and the turn signal kept turning on and off. Once the light turned green, the car suddenly turned right even though that was not the correct route. And, if it needed to turn right, It had plenty of opportunity to do so since right turns on red were not prohibited.
FSDb works okay in some situations, but those are the same situations where the standard Autopilot works. City driving - at least for me - is so bad that that FSDb has negative value. There's something wrong with almost every maneuver.
Surely, my experience isn't typical? We'll see what 11.4.x brings. So far, all I see is "disengaged - what happened?" at the slight majority of maneuvers and it seems like I spend more time recording what just happened than witnesses the car do something correctly. I do kind of suspect it doesn't know how to handle the non-perpendicular intersections around me in these small towns. Maybe. I dunno. I'll keep playing with it and hope it doesn't kill me.
clydeiii
Member
Recalibrate cameras? I take traffic circles on my commute and it’s fine.
gx9901
Active Member
I just subscribed last night and got the 2023.12.10 download today. I immediately used FSD to take my daughter to an art class about a 30 minute drive away. It was heavy rush hour and the trip took almost 50 minutes, but I thought it did pretty well. There was one instance where it was signaling to turn left onto an on-ramp, but started going toward the right lane so I had to yank the steering wheel from it. Otherwise there weren't any glaring mistakes. It was mostly highway amid heavy traffic though, so maybe there weren't too many chances to screw up. But I was impressed with the lane changing maneuvers it did.
Anyway, I don't have high expectations for FSD and pretty much got it hoping I might have it for a road trip this weekend and then to play with it for a while, so I am planning to unsubscribe at some point. But I do like the better performance on highway compared to Autopilot, so that might keep me subscribed until the winter if I end up liking it enough.
Anyway, I don't have high expectations for FSD and pretty much got it hoping I might have it for a road trip this weekend and then to play with it for a while, so I am planning to unsubscribe at some point. But I do like the better performance on highway compared to Autopilot, so that might keep me subscribed until the winter if I end up liking it enough.
I’ve noticed that the last decent day of FSDb drives was this past Sunday. My weekly work routes (that I literally travel to the same places Monday-Wednesday (Thursday and Friday’s are always different)) have been absolutely horrible. I’ve had more disengagements since Monday this week than I’ve had in the last 3 weeks on 11.3.6. Explain that.
I sub'd yesterday and got 23.12.10 at around 5pm, so I took it out for a spin around my lake, as driving curving undulating backcountry roads is the norm around here. Also, I live on a dirt road, so it was interesting to see how well it handled that, and it was pouring rain!
I was pleasantly surprised at how well it worked. On the dirt road, it was going too fast, as I only drive about 15mph, and it wanted to drive 25mph, so I scroll wheeled the speed down. I set it for "chill", as I get used to it. I didn't have any other reasons to intervene, so I was impressed!
I was pleasantly surprised at how well it worked. On the dirt road, it was going too fast, as I only drive about 15mph, and it wanted to drive 25mph, so I scroll wheeled the speed down. I set it for "chill", as I get used to it. I didn't have any other reasons to intervene, so I was impressed!
clydeiii
Member
Whenever there's a change of behavior but no change of software, I look to changes of data. Tesla may have had a map update or a change of map service or some such thing, and the result is that the data for your route is worse. Failing that explanation, well... aliens.
Many were surprised to see an older version of FSDb (11.3.6) included with the 2023.12.10 update this week that made FSDb access available to the masses instead of 11.4.x. While we're all guessing, I have a conspiracy theory: Tesla may have intentionally pushed out a bad version so that those of us who are new to FSDb will be disappointed, frustrated, and frightened and have our expecations of the feature lowered. Then, 11.4.x will be pushed out in a matter of weeks and, even though far from perfect, will seem like "magic" compared to what we started with and we'll all be praising it far more than we would had we started with that version.
DISCLAIMER: This post is just for fun - I don't actually believe that Tesla would do something like that...or would they?
DISCLAIMER: This post is just for fun - I don't actually believe that Tesla would do something like that...or would they?
Similar threads
