FSD Beta 10.69

[AlanSubie4Life]

Thread quality and 10.69 quality appear to be converging. Law of the universe?

Not that I contributed to thread quality earlier with incessant whining about stopping and pedantic details of a beer bet. That being said, things have come to pass as I expected, and good to see most agree the stopping behavior is unacceptable (not a surprise of course). It’s important to set reasonable expectations! No disappointment.

I am not sure why Elon said 10.69 would solve CULT 100%. Should we give him a pass? Will they stay on 10.69.X until we get “perfect” performance on CULT? For what sample size?

(Of course, some will claim the last two videos were perfect. But two “perfect” (with failures) videos with sample size 15 does not imply perfect. It’ll be interesting to see how 10.69.3 does! Hoping it is even more bold!)

I am hoping .3 and subversions will bring me another delicious beer, just in time for Oktoberfest.

 

[JulienW]

Not important but Beta drove me from my gym through midtown and to the Atlanta Supercharger (where I live but through an extra couple of intersections and roads) with 0 disengagements (a couple of accelerator interventions) for the first time that I remember happening since Beta release. And traffic has heaver than average (Friday), though didn't have any "deal killer" edge cases to creep up.

 

[momo3605]

Thread quality and 10.69 quality appear to be converging. Law of the universe?

Not that I contributed to thread quality earlier with incessant whining about stopping and pedantic details of a beer bet. That being said, things have come to pass as I expected, and good to see most agree the stopping behavior is unacceptable (not a surprise of course). It’s important to set reasonable expectations! No disappointment.

I am not sure why Elon said 10.69 would solve CULT 100%. Should we give him a pass? Will they stay on 10.69.X until we get “perfect” performance on CULT? For what sample size?

(Of course, some will claim the last two videos were perfect. But two “perfect” (with failures) videos with sample size 15 does not imply perfect. It’ll be interesting to see how 10.69.3 does! Hoping it is even more bold!)

I am hoping .3 and subversions will bring me another delicious beer, just in time for Oktoberfest.

I finally figured out that the regen bar shows braking on it (i started paying attention after your posts), and man, slowing down for a light at any reasonable speed, it brakes so heavily, it uses the brakes EVERY time. Honestly, i'd probably use FSD Beta more if the slowing alone was smoother. Just get the basics right...I don't mind taking over at complex intersections.

 

[AlanSubie4Life]

(i started paying attention after your posts), and man, slowing down for a light at any reasonable speed, it brakes so heavily, it uses the brakes EVERY time.

Not EVERY time. Try an uphill stop. Also every now and again it’ll pull a miracle out of its a** and make a nice smooth stop even on the flat. Downhill is a lost cause usually though (not always, see video).

But yes, high speed stops from 45-50mph add even more challenge for it, and I think make a smooth stop less likely.

It’s absurd.

Very glad they added the braking display. It makes me even more upset at every stop. But to be clear, it’s not even the brake use that is the issue - it’s the smoothness. If they were more smooth they could avoid brakes most of the time (goes hand in hand).

I would definitely use it way more if it could drive smoothly. It’s just unpleasant and annoying to use on wide well-marked parkways with traffic lights, the way it is.

A big part of using no brakes is predicting lights. Hopefully Tesla will add this soon. Right after they smooth out the stopping.

 

[myshenaini]

I haven't even had a chance to try out .2 yet but judging by what I've read here, I don't think I have to.

 

[myshenaini]

Maybe they should start offering: How to park in a single spot subscription too.

View attachment 856041

This is called "parking like a boss". Tesla needs to add this option, opposite end of the spectrum from Chill mode.

 

[momo3605]

Not EVERY time. Try an uphill stop. Also every now and again it’ll pull a miracle out of its a** and make a nice smooth stop even on the flat. Downhill is a lost cause usually though (not always, see video).

But yes, high speed stops from 45-50mph add even more challenge for it, and I think make a smooth stop less likely.

It’s absurd.

Very glad they added the braking display. It makes me even more upset at every stop.

I would definitely use it way more if it could drive smoothly. It’s just unpleasant and annoying to use on wide well-marked parkways with traffic lights, the way it is.

A big part of using no brakes is predicting lights. Hopefully Tesla will add this soon. Right after they smooth out the stopping.

What confuses me is that Elon says 10.69 has MAJOR architecture changes. Yet, if you watch how it drives, it mostly makes the exact same dumb mistakes FSD has always made. It seems to me like it's mostly the same thing, just with a new median network added in(creep was always there but not visualized)....and maybe some minor tweaking that happened to make turns and creeping a bit smoother.

If it takes 3 months of dedicated effort on ULTs to get them to ~80%, how long before they even cover 80% of edge cases? Forget 90 or 99% at this point.

 

[AlanSubie4Life]

What confuses me is that Elon says 10.69 has MAJOR architecture changes. Yet, if you watch how it drives, it mostly makes the exact same dumb mistakes FSD has always made. It seems to me like it's mostly the same thing, just with a new median network added in(creep was always there but not visualized)....and maybe some minor tweaking that happened to make turns and creeping a bit smoother.

If it takes 3 months of dedicated effort on ULTs to get them to ~80%, how long before they even cover 80% of edge cases? Forget 90 or 99% at this point.

Adding the occupancy network (which was there in some form but not being used and now it appears to have better resolution) was a big step forward and fundamental.

I’m less concerned about the details and more concerned about fundamental capabilities, for example:
1) Detection range (and velocity estimation at range)
2) Knowing what it can’t see
3) Predicting other users’ trajectories
4) Ability to localize.
Etc.

… so when we see some of these capabilities improve that is a good step forward. It looks like the car can see 130m/450ft at least, which is good, not good enough though. Hopefully we’ll get indications soon it can see further.

Anyway big changes and huge improvements in capability in general I would expect to have very little impact on the quality of the product at this point. So nothing inconsistent there. It’s just got so far to go.

Having the capabilities means it is possible to improve some of these things even if they do not improve at all.

 

[Yelobird]

What confuses me is that Elon says 10.69 has MAJOR architecture changes. Yet, if you watch how it drives, it mostly makes the exact same dumb mistakes FSD has always made. It seems to me like it's mostly the same thing, just with a new median network added in(creep was always there but not visualized)....and maybe some minor tweaking that happened to make turns and creeping a bit smoother.

If it takes 3 months of dedicated effort on ULTs to get them to ~80%, how long before they even cover 80% of edge cases? Forget 90 or 99% at this point.

If all cars were stationary it might be simple but this is Not a typical video game solution. If it was easy Ever car on the road would be autonomous.

 

[momo3605]

Adding the occupancy network (which was there in some form but not being used and now it appears to have better resolution) was a big step forward and fundamental.

I'm pretty sure this was being used in some form before too. FSDb would jog around unvisualized street poles for instance, and also avoid those outdoor eating setups (also not visualized)

 

[sleepydoc]

What confuses me is that Elon says 10.69 has MAJOR architecture changes. Yet, if you watch how it drives, it mostly makes the exact same dumb mistakes FSD has always made. It seems to me like it's mostly the same thing, just with a new median network added in(creep was always there but not visualized)....and maybe some minor tweaking that happened to make turns and creeping a bit smoother.

If it takes 3 months of dedicated effort on ULTs to get them to ~80%, how long before they even cover 80% of edge cases? Forget 90 or 99% at this point.

You're confusing the system architecture with the actual code. You can put a ton of work into redesigning the system without it being evident to the end user.

 

[Daniel in SD]

You're confusing the system architecture with the actual code. You can put a ton of work into redesigning the system without it being evident to the end user.

It seems challenging to reproduce the exact same bugs in a new architecture though.
Unless most of the bugs are intentional...

 

[sleepydoc]

A big part of using no brakes is predicting lights. Hopefully Tesla will add this soon. Right after they smooth out the stopping.

That's actually pretty advanced. I'll typically gauge my braking so I hit the light as it turns green (or so I approach the cars in front of me right as they're speeding up.) Doing so requires looking at the light, looking at the number of cars waiting, estimating how long the light will take before it turns and also knowing the pattern of the light (does the turn lane go first or last? When does your lane turn green relative to the other lanes?)

I had a case yesterday where it kept driving up to a bunch of cars waiting at a light and waited far too long to brake, necessitating a pretty aggressive stop. I'd be happy with a more reasonable braking distance.

 

[momo3605]

Doing so requires looking at the light, looking at the number of cars waiting, estimating how long the light will take before it turns and also knowing the pattern of the light (does the turn lane go first or last? When does your lane turn green relative to the other lanes?)

Of course what you suggest is challenging, but i don't think they need to go down that route...Audi has had traffic light predicting for years now...and if that's some kind of patented technology, then slowing down a little earlier would also require less aggressive braking. Not sure if this is a camera distance limitation though

 

[sleepydoc]

Of course what you suggest is challenging, but i don't think they need to go down that route...Audi has had traffic light predicting for years now...and if that's some kind of patented technology, then slowing down a little earlier would also require less aggressive braking. Not sure if this is a camera distance limitation though

What does Audi’s system do and how does it work? I had a 2017 A4 and it wasn’t available on that model when I bought it. I suppose it’s possible to include information on traffic light Patterns in the map database but even that would be problematic since like timing frequently changes throughout the day.

You’re right though. Slowing down for a red light should not be that difficult. The cameras are supposed to be able to see a couple hundred meters and I routinely see the red light in the display long before the car starts braking so distance is not an issue.

 

[uscbucsfan]

What does Audi’s system do and how does it work? I had a 2017 A4 and it wasn’t available on that model when I bought it. I suppose it’s possible to include information on traffic light Patterns in the map database but even that would be problematic since like timing frequently changes throughout the day.

You’re right though. Slowing down for a red light should not be that difficult. The cameras are supposed to be able to see a couple hundred meters and I routinely see the red light in the display long before the car starts braking so distance is not an issue.

From my experience, it's a calculation on when to let off the accelerator, how much regen to use vs. braking. When the car sees a red light, it appears that the acceleration slows, not completely and tries to gradually slow down until it realizes it's too late and needs to apply brakes. Those tolerances from full acceleration to light acceleration (slowing) to full regen to braking need to adjust...seems easy, but sometimes it seems they focus on very specific tasks instead of working on everything.

 

[GregalseT]

The main area of improvement I would like to see progress towards is lane assignation. Missing turns b/c FSD chooses an incorrect lane is maddening, and silly given that the map data usually shows that a turn is coming up.

Curious as to why this is such a seemingly difficult problem.

 

[HookBill]

Has anyone gotten

"WELCOME TO 11"

My son got this message this morning when he got in his 3 this morning. Then his screen went black. now he doesn't even have FSD. Any resolution to get it back. I have not talked to him about resets yet but believe he at least did the two thumb wheel reset.

Redirect Notice

A hairball left over from past grooming.

 

[PACEMD]

It seems challenging to reproduce the exact same bugs in a new architecture though.
Unless most of the bugs are intentional...

Of course, it's intentional, and a lot of work. It's not easy to perfect phantom braking to continue to occur randomly and unpredictably........that's some work for your ass right there........

 

[PACEMD]

He was definitely up to no good - and Tesla's new motivation NN detected it nearly immediately. If you enable the new enhanced visualizations in the settings menu, it now includes a "mood" function which changes the color of people based on their intentions...

So you can legally shoot them if they have a busted tail light, right?