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

FSD Beta 10.69

This site may earn commission on affiliate links.
Yes, this is a flaw, noted earlier. It’s a “must fix.”

It must engage the creep network earlier, not wait so long to begin the creep, and creep in a more predictable (slower) fashion. This will result in faster, more predictable, more human creeps.
Another "must fix" is the abysmal behavior on unmarked residential roads, where the car stops for any approaching car. When I saw Chuck Cook's video showing this, I wasn't too concerned because my neighborhood streets are plenty wide enough for two cars to comfortably pass. But, 10.69.1.1 broke something here and it needs to be addressed.
 
Another "must fix" is the abysmal behavior on unmarked residential roads, where the car stops for any approaching car. When I saw Chuck Cook's video showing this, I wasn't too concerned because my neighborhood streets are plenty wide enough for two cars to comfortably pass. But, 10.69.1.1 broke something here and it needs to be addressed.
There must be a threshold on the width of the road as (although not useful) on unmarked 4-lane roads, it doesn't stop when a car comes towards me. In fact, it moves slightly towards right as I would have done. Previous version used to drive dead center, then violently move right when a car appears in front of me... 😅

Should be able to find the right balance easily...
 
I do have problems with FSD beta not changing lanes to a straight-through lane when the one it is in becomes a turn-only lane. Fix one problem and create another?
Did 10.12 switch out of that lane before? I also noticed some differences in lane selection with 10.69 where previously it would switch out from further away. It seems like there's at least 3 different types of turn lane behaviors:
  1. planning ahead previously heavily based on map data and now assisted by "deep lane guidance" in 10.69
  2. selecting a lane immediately at an intersection
  3. committing to the allowed movement of the current lane
Items 2 and 3 are very closely related and most likely had general network training with autolabeled fleet data for 10.69 to better predict painted arrows. There doesn't even seem to be a release note specific to this as Tesla is probably incrementally improving the accuracy of many predictions with "operation vacation" automatically determining that a misprediction of say a turn arrow at 100 feet away was obviously wrong when viewing from 50 feet away. I have a hunch that the neural network mispredicts left vs right arrows as they are more obviously "not straight" so some signal detecting the curviness of the painted road marking boosts both "left" and "right" predictions even though practically we know there shouldn't be a left turn lane on the right side (… most of the time…).

The new changes for item 1 seems to have come with giving the new guidance module more time/distance to make its decision to avoid the existing 10.12 problem of unnecessarily switching lanes. So now that this is handled more with neural networks, it should be part of the "standard" process of generally improving predictions with more training data from mispredictions, and as the confidence of predictions increases, Tesla can adjust the distance thresholds of when this module takes priority.
 
Any model s refresh owners noticing the audio system sounds different (worse) after the update? I can't put my finger on what sounds wrong but it's very obviously different, family noticed it as well.

Model 3 and occasional when a system sound comes on the stereo goes out for several seconds. Sometimes it comes back distorted and is then fine seconds later. Some times it just comes back fine but with the delay of no sound.
 
Running red lights is obviously unacceptable. Since it's been brought up a couple of times for .69, it's worth mentioning that my car doesn't have .69 yet, but tried to run a red light at a right turn last week.

The car ahead of me had stopped and then turned, and my car seemed to be playing follow the leader instead of observing the red light. This is an intersection where I've never had any prior problems with this.
 
  • Informative
Reactions: GranularHail
Did 10.12 switch out of that lane before? I also noticed some differences in lane selection with 10.69 where previously it would switch out from further away. It seems like there's at least 3 different types of turn lane behaviors:
  1. planning ahead previously heavily based on map data and now assisted by "deep lane guidance" in 10.69
  2. selecting a lane immediately at an intersection
  3. committing to the allowed movement of the current lane
Items 2 and 3 are very closely related and most likely had general network training with autolabeled fleet data for 10.69 to better predict painted arrows. There doesn't even seem to be a release note specific to this as Tesla is probably incrementally improving the accuracy of many predictions with "operation vacation" automatically determining that a misprediction of say a turn arrow at 100 feet away was obviously wrong when viewing from 50 feet away. I have a hunch that the neural network mispredicts left vs right arrows as they are more obviously "not straight" so some signal detecting the curviness of the painted road marking boosts both "left" and "right" predictions even though practically we know there shouldn't be a left turn lane on the right side (… most of the time…).

The new changes for item 1 seems to have come with giving the new guidance module more time/distance to make its decision to avoid the existing 10.12 problem of unnecessarily switching lanes. So now that this is handled more with neural networks, it should be part of the "standard" process of generally improving predictions with more training data from mispredictions, and as the confidence of predictions increases, Tesla can adjust the distance thresholds of when this module takes priority.
I believe that 10.12.2 did select the correct lane at these intersections. I should note that this particular issue does not occur often for me, as I need to change lanes to get into most turn-only lanes. It's really a crummy design when your lane suddenly becomes a turn lane. But, the car must deal with it.
 
First couple drives on 10.69.1.1:
* Seems to respect the correct side of the road better on unmarked streets. Much more natural positioning, though it can still swerve to the center in random places.
* One situation on a particularly narrow street with parked cars on both sides and an oncoming car. We were about to meet on a stretch of road where there were no parked cars on one side, allowing more room to pass, but the oncoming car would have to move over first. The beta abruptly stopped in an awkward place, somewhat blocking the other car. Slowing down was the correct behavior but it stopped probably 30 feet short of where I would have, where the road was widest.
* Turning right on a red light, sharp turn; it drew the creep line, but didn’t commit soon enough so I took over and completed the turn manually due to the approaching cars.
* Simple UPL with an oncoming car at a medium distance - hesitated too much and swung to the right due to the sharp angle of the turn. I disengaged and completed the turn because there were other cars in the right lane and I wasn’t sure if it was about to get stuck in an awkward spot due to the oncoming car.
* Way way too slow approaching stop signs, and then starting from a stop. 10.12.2 was a half step forward in assertiveness but this one is too timid again.
* Maybe placebo effect but it did feel smoother accelerating and slowing down when following other cars.
* Probably biggest “wtf” was the abrupt stop on a 2-lane road with a double-parked car in the right lane and me in the left lane. There were a couple people loading the trunk with no intention of walking into traffic and there were cars behind me, so I disengaged before it dropped too much speed.
 
I'm normally pretty bullish on FSD but this release has been really disappointing. On my incredibly short 1.5 mile commute home FSD:

  • Tried to run 2 clearly red lights (no greens in view of the cameras)
  • Took a full 10 seconds to reach 29mph after each of the 5 stop signs.
  • Veered into the bike line when the road slightly curved then jerked back into the lane
  • Crept completely into a very visible intersection (blocking both directions if cars were there) then floored it to complete the turn
  • Didn't slow down at any speed bumps, used to do this fine
  • Chickened out of a simple right on green causing cars behind to brake hard.
This drive had probably 8 disengagements, it's normally about 2. 10.12.2 was much better for me, I've already recalibrated the cameras with no improvement.
I find this fascinating. I had the exact opposite experience with 10.12.2. I used it for 3 days then shelved it. It was un drivable for me. Different locations I guess.
 
It’s good to see feedback that is consistent with my experiences. And it is good to see videos that show the same behavior.

I realize I do nothing but complain, which is incredibly tiresome to everyone including myself, but at least I can say that everything is going as I expected with 10.69.1.1. I feel very satisfied with what we have - that it meets my current expectations.

I do sincerely hope that the polish that Elon says is coming in 10.69.2 actually happens. That would be awesome. It would actually be pretty great if this were something you could use to perform basic driving tasks between traffic lights and around turns, while monitoring it carefully.

Unfortunately, that’s not the way it is right now. Basically it is something that unless you are specifically testing it (a perfectly legitimate and fulfilling activity in its own right - and even fun as you watch the vehicle dive for a curb as it goes for a turn lane 😂 ), it is not useful. You have to immediately turn it off except in the most trivial driving situations (where it does moderately well - marked roads, no turns, no traffic, no pedestrians, no traffic control). In these situations, my wife will not notice it is being used - most of the time (and if it screws up once it has to be turned off).

And unfortunately I think we have to be cognizant of Elon’s prior pronouncements, and realize that the probability of it being polished in 10.69.2 is vanishingly small.

We can hope of course that “this time will be different.” But definitely not counting on it. I think we have another year or so of polishing before it starts to be more usable. Maybe longer. It depends on what it is that limits their current performance, which I think is very poorly understood by everyone at the moment.

True wide release certainly looks like it will have to wait for a long time, unless they want to do another Smart Summon trick.

And I fervently hope I am wrong. Would love to see enormous capability improvements sometime very soon!

Let’s get over 90% on CULT, with traffic, in 10.69.2! Even that would be a significant achievement at this point.
 
It’s good to see feedback that is consistent with my experiences. And it is good to see videos that show the same behavior.

I realize I do nothing but complain, which is incredibly tiresome to everyone including myself, but at least I can say that everything is going as I expected with 10.69.1.1. I feel very satisfied with what we have - that it meets my current expectations.

I do sincerely hope that the polish that Elon says is coming in 10.69.2 actually happens. That would be awesome. It would actually be pretty great if this were something you could use to perform basic driving tasks between traffic lights and around turns, while monitoring it carefully.

Unfortunately, that’s not the way it is right now. Basically it is something that unless you are specifically testing it (a perfectly legitimate and fulfilling activity in its own right - and even fun as you watch the vehicle dive for a curb as it goes for a turn lane 😂 ), it is not useful. You have to immediately turn it off except in the most trivial driving situations (where it does moderately well - marked roads, no turns, no traffic, no pedestrians, no traffic control). In these situations, my wife will not notice it is being used - most of the time (and if it screws up once it has to be turned off).

And unfortunately I think we have to be cognizant of Elon’s prior pronouncements, and realize that the probability of it being polished in 10.69.2 is vanishingly small.

We can hope of course that “this time will be different.” But definitely not counting on it. I think we have another year or so of polishing before it starts to be more usable. Maybe longer. It depends on what it is that limits their current performance, which I think is very poorly understood by everyone at the moment.

True wide release certainly looks like it will have to wait for a long time, unless they want to do another Smart Summon trick.

And I fervently hope I am wrong. Would love to see enormous capability improvements sometime very soon!

Let’s get over 90% on CULT, with traffic, in 10.69.2! Even that would be a significant achievement at this point.

This always seems to be the case with each update. Elon hypes up next version, YT Celebrities get the build making the update look pretty promising, @powertoold shows up to say how mindblowing it looks from videos🤣, then it rolls out wide and reports are a mixed bag...some things it does better, some areas it now messes up that it didn't before....sigh...
 
Running red lights is obviously unacceptable. Since it's been brought up a couple of times for .69, it's worth mentioning that my car doesn't have .69 yet, but tried to run a red light at a right turn last week.

The car ahead of me had stopped and then turned, and my car seemed to be playing follow the leader instead of observing the red light. This is an intersection where I've never had any prior problems with this.

This has been an ongoing issue for me. Right turn on controlled 4-way. As soon as you make a right there is another light. Literally one car length long.

Beta ALWAYS tries to run it. Every time. It doesn't check, or expect so soon another light.

There is a tall building on the corner so you can't even see the next light until you've started to complete the turn.

If I can see it and react then the car should have no problem !

Report it each time.
 
YT Celebrities get the build making the update look pretty promising
This part is false. The YouTube videos always look exactly like what we get. This time was no exception with 10.69. (And by extension 10.69.1.1.)

Every. Single. Issue. Was readily apparent. You just had to really watch the videos and feel them. It’s definitely something that sneaks by people though. Watch that regen bar, watch the speed, watch the car positioning, and feel the behavior.

That’s the one lesson: watch the real-time videos and assess, to see what we’re getting. You’ll know exactly what you’re going to get. No box of chocolates.
 
  • Like
Reactions: nvx1977
This part is false. The YouTube videos always look exactly like what we get. This time was no exception with 10.69.

Every. Single. Issue. Was readily apparent. You just had to really watch the videos and feel them. It’s definitely something that sneaks by people though.

If it's such a subtle issue with speeds and being slightly slow, is it that difficult to fix?

Were you testing out 10.2 when it first came out? It felt like the Indiana Jones ride at Disneyland, on rails. It was wild.
 
If it's such a subtle issue with speeds and being slightly slow, is it that difficult to fix?
It’s not subtle. That’s why I said it was readily apparent! I have no idea if it is difficult to fix - it depends on the root cause.
Were you testing out 10.2 when it first came out? It felt like the Indiana Jones ride at Disneyland, on rails. It was wild.
Yes, whatever the first release to SS 100 was. I’ve said multiple times the current FSD version is way better.
 
It’s not subtle. That’s why I said it was readily apparent!

Well, the "feeling" of the effects is apparent in the car, but on a macro decision level, it seems "obvious" that Tesla will reduce these little false slowdowns over time, as they polish up the occupancy and lane guidance networks.

It was only about 4-5 months ago when we were getting those massive slowdowns anywhere near pedestrians.

Now we get stuff like this, where it seems a bit too assertive around pedestrians (IMO):

 
Well, the "feeling" of the effects is apparent in the car, but on a macro decision level, it seems "obvious" that Tesla will polish these little false slowdowns over time, as they polish up the occupancy and lane guidance networks.
I mean, that is easy to say, but if you look at the arc of history, I have “concerns.”

You’re chalking it up to occupancy and lane guidance, but I would put those low on the list of likely root causes here. I have no idea why these issues exist.

Also, you seem to be talking about the false slowdowns, which is another issue, which I was not actually discussing.