Update 2024.44.25.3...could this be for CT?

[rtfitch]

“Something just popped up in the Update Environment—could it be for us?

We don’t want to get too ahead of ourselves, but it’s so much fun keeping our eyes peeled for the Holiday update.

Ho, Ho, Howdy!”

Sponsored

 

[BBradac9]

Let's hope so!

 

[dalton108]

No, next incoming is 2024.39.15 but it's not the Holliday update but 13.2 FSD

Source? When did the plan change from only 13.3 going wide?

 

[dalton108]

me

have no idea what plan you speak of, not privy to any planning

IRC Elon said that 13 would not be going out wide until 13.3. He’s not always right about these things so, I’ll take you at face value.

 

[CyberTahoe]

have no idea what plan you speak of, not privy to any planning

Tesla AI on X October 31: "We are targeting a wide release with v13.3 with most of the above improvements for AI4 vehicles around Thanksgiving!"

then again, we're long past Thanksgiving so decent chance plans have changed.

 

[carsly]

thinking the “holiday update” might be Memorial Day

 

[65SoYoLO]

Groundhog day.

 

[rtfitch]

No, next incoming is 2024.39.15 but it's not the Holliday update but 13.2 FSD

Sorry about starting this thread then, thought I saw a glimmer of hope!

 

[dalton108]

Sorry about starting this thread then, thought I saw a glimmer of hope!

NAN! Keep hope alive Red One!

 

[65SoYoLO]

Sorry about starting this thread then, thought I saw a glimmer of hope!

Hey I'm still hoping. I am guessing on the 22nd. Hey, only a nickel but it's still a bet!

 

[Obie_fl]

Pretty sure this is an update to the software that has been rolling out to the older AI3 cars.

 

[Zane Edwards]

I think we will get it before Christmas

 

[D101010]

I think we will get v13 with the holiday update. The best Tesla deserves the best holiday update ??

 

[LiveFreeOrDieBob]

Would be super disappointing if we didn't get the Cyber Sleigh and Sit Happens before Christmas, but I suspect they are struggling with FSD v13 for Cybertruck. I have a sinking feeling that the combination of dynamic steering ratio and rear steering are a challenge to train using data from any of the other S3XY cars, which would mean CT FSD updates are always going to be way behind since they need to be trained or maybe optimized for certain scenarios with only CT-sourced training corpus where all the other cars can share data between the different models. My theory comes from the observation that my Beast's FSD does amazing at everything EXCEPT low-speed tight turns where the Steer-by-Wire is most aggressive, where the CT cuts corners either hitting curbs or driving on the wrong side the road when turning onto perpendicular streets without center lines. Both of those symptoms feel to me like the model is requesting a certain steering input, but that steering angle input is way too much for the situation and for some reason it isn't able to employ a direct feedback loop based on video data to correct it mid-turn - at least most of the time. I also have seen the opposite, where FSD is approaching an offramp and dials in a certain steering angle for the offramp, but as the truck slows that steering angle input stays the same, causing the truck to track out to the outside of the ramp until it gets close to the curb and then applies more steering input to go to the inside of the ramp. If FSD isn't capable of using real-time camera input to determine how close it is to the edge of the road on a tight turn, and from that, dynamically adjust the steering input, then it'll have to have some sort of translation function between the FSD and the Steering that takes in speed, requested steering angle, and rear-steering on/off, and updates the steering angle received by the steering modules based on those inputs to get the correct steering angle.

 

[D101010]

Would be super disappointing if we didn't get the Cyber Sleigh and Sit Happens before Christmas, but I suspect they are struggling with FSD v13 for Cybertruck. I have a sinking feeling that the combination of dynamic steering ratio and rear steering are a challenge to train using data from any of the other S3XY cars, which would mean CT FSD updates are always going to be way behind since they need to be trained or maybe optimized for certain scenarios with only CT-sourced training corpus where all the other cars can share data between the different models. My theory comes from the observation that my Beast's FSD does amazing at everything EXCEPT low-speed tight turns where the Steer-by-Wire is most aggressive, where the CT cuts corners either hitting curbs or driving on the wrong side the road when turning onto perpendicular streets without center lines. Both of those symptoms feel to me like the model is requesting a certain steering input, but that steering angle input is way too much for the situation and for some reason it isn't able to employ a direct feedback loop based on video data to correct it mid-turn - at least most of the time. I also have seen the opposite, where FSD is approaching an offramp and dials in a certain steering angle for the offramp, but as the truck slows that steering angle input stays the same, causing the truck to track out to the outside of the ramp until it gets close to the curb and then applies more steering input to go to the inside of the ramp. If FSD isn't capable of using real-time camera input to determine how close it is to the edge of the road on a tight turn, and from that, dynamically adjust the steering input, then it'll have to have some sort of translation function between the FSD and the Steering that takes in

speed, requested steering angle, and rear-steering on/off, and updates the steering angle received by the steering modules based on those inputs to get the correct steering angle.

Would be super disappointing if we didn't get the Cyber Sleigh and Sit Happens before Christmas, but I suspect they are struggling with FSD v13 for Cybertruck. I have a sinking feeling that the combination of dynamic steering ratio and rear steering are a challenge to train using data from any of the other S3XY cars, which would mean CT FSD updates are always going to be way behind since they need to be trained or maybe optimized for certain scenarios with only CT-sourced training corpus where all the other cars can share data between the different models. My theory comes from the observation that my Beast's FSD does amazing at everything EXCEPT low-speed tight turns where the Steer-by-Wire is most aggressive, where the CT cuts corners either hitting curbs or driving on the wrong side the road when turning onto perpendicular streets without center lines. Both of those symptoms feel to me like the model is requesting a certain steering input, but that steering angle input is way too much for the situation and for some reason it isn't able to employ a direct feedback loop based on video data to correct it mid-turn - at least most of the time. I also have seen the opposite, where FSD is approaching an offramp and dials in a certain steering angle for the offramp, but as the truck slows that steering angle input stays the same, causing the truck to track out to the outside of the ramp until it gets close to the curb and then applies more steering input to go to the inside of the ramp. If FSD isn't capable of using real-time camera input to determine how close it is to the edge of the road on a tight turn, and from that, dynamically adjust the steering input, then it'll have to have some sort of translation function between the FSD and the Steering that takes in speed, requested steering angle, and rear-steering on/off, and updates the steering angle received by the steering modules based on those inputs to get the correct steering angle.

If they were directly translating video in to controls I would agree. But they have a patent on a virtual camera that they create. If they just take the desired path and use degrees of rotation. They can let a vehicle specific algorithm figure out how to translate that to wheel rotation. I think it is just a numbers game. More AI4 2 wheel steering vs 4 wheel.

Sponsored