FSD on Cybertruck Performs Significantly Worse than on Model 3, S, X (and is missing features)

[65SoYoLO]

Burn.

Sponsored

 

[igs]

@igs

Here ya go my man!

https://www.qwak.com/post/pytorch-vs-tensorflow



https://codersbarn.com/does-tesla-use-pytorch-or-tensorflow/



OMG...



Nooooo.... they don't... ?

your second paragraph just contradicted your first paragraph

 

[M0unt41nm4n]

your second paragraph just contradicted your first paragraph

How so?

TensorFlow is a powerful tool for building and training neural networks, but it’s also more complex than many other deep learning frameworks.It is used for major projects across the world and is used by companies such as Airbnb, Google, Uber, Tesla, and more.

When deciding between PyTorch and TensorFlow, Tesla opted for PyTorch as its primary deep learning framework.

They use both but their primary is PyTorch. Did I miss something?

I'm sorry @igs ... this is kind of fun. You aren't the sharpest tool in the box.

 

[igs]

How so?





They use both but their primary is PyTorch. Did I miss something?

I'm sorry @igs ... this is kind of fun. You aren't the sharpest tool in the box.

Like I said in my first post Tesla isn't using any of that anymore Dojo was dropped a looooooooooong time ago. Get with the times bro.

 

[jahansolu]

My Y and S that are hardware 3 are worse for me. I prefer it on the cybertruck.

How so?





They use both but their primary is PyTorch. Did I miss something?

I'm sorry @igs ... this is kind of fun. You aren't the sharpest tool in the box.

Holy... ok let me put this to rest once and for all.

It's evident that @M0unt41nm4n works within the AI industry. @igs , you have yet to prove any credibility within the AI domain to be able to criticize @M0unt41nm4n .

It doesn't really matter what language model the vehicle is using; as @M0unt41nm4n stated: It's in the release notes of FSD that it is indeed using end-to-end neural network.

Now to get back to the topic on hand: Yes, CT is performing less reliably than the other models, but in due time this will get better. Correct me if I'm wrong @M0unt41nm4n , one thing I have noticed with AI is this: It requires much less input to learn. What I mean by this is old AI required (let's say) 500,000 min of video footage, but now it can scale up to the same level with only 10,000. So, in due time, take the fact that CT was announced 1 year ago and it's already getting pretty close to what the other models had for years.

 

[M0unt41nm4n]

Holy... ok let me put this to rest once and for all.

It's evident that @M0unt41nm4n works within the AI industry. @igs , you have yet to prove any credibility within the AI domain to be able to criticize @M0unt41nm4n .

It doesn't really matter what language model the vehicle is using; as @M0unt41nm4n stated: It's in the release notes of FSD that it is indeed using end-to-end neural network.

Now to get back to the topic on hand: Yes, CT is performing less reliably than the other models, but in due time this will get better. Correct me if I'm wrong @M0unt41nm4n , one thing I have noticed with AI is this: It requires much less input to learn. What I mean by this is old AI required (let's say) 500,000 min of video footage, but now it can scale up to the same level with only 10,000. So, in due time, take the fact that CT was announced 1 year ago and it's already getting pretty close to what the other models had for years.

Yep you are 100% correct. How this works is they leverage the old models, which is common (i.e. for average AI folks, you download a model already trained on huggingface.com and use those to train your new info to tweak to your new info to make a more focused model).

So they use the models say, from the other Tesla models and they end up getting 95% of it right. But it just needs tweaking for the bigger/heaver/wider truck. So they can get the CT to do most of the others as others have stated in this thread. Hence my comment on clipping folks when you make the left hand turn. It gets it mostly right, but the model is generally based on a smaller vehicle. So its learning... thus why it will upload all that data, especially when you grab the wheel. They take that data and run it through their models to tweak it for the CT. Hence is a LOT faster because its based on all the other Tesla data. But its learning now. The good news is... it will be just as good in the not too distant future.

 

[Deleted member 27593]

I hope this isn’t too off topic. But if FSD needs time to adjust to a different sized truck, what does that mean for the robocab? Will the model also have to relearn to drive for the size / dimensions of that vehicle? And how long could it take for the model to be safe for autonomous robotaxi driving? Months?

Sorry for the derailment. I sniffed glue in kindergarten.

 

[M0unt41nm4n]

I hope this isn’t too off topic. But if FSD needs time to adjust to a different sized truck, what does that mean for the robocab? Will the model also have to relearn to drive for the size / dimensions of that vehicle? And how long could it take for the model to be safe for autonomous robotaxi driving? Months?

Sorry for the derailment. I sniffed glue in kindergarten.

Yes. But likely less time if its specs/size are similar to a model 3/Y/X.

 

[M0unt41nm4n]

Like I said in my first post Tesla isn't using any of that anymore Dojo was dropped a looooooooooong time ago. Get with the times bro.

Dojo was dropped?

https://techcrunch.com/2024/08/10/teslas-dojo-a-timeline/

https://techcrunch.com/2024/08/03/t...-plan-to-build-an-ai-supercomputer-explained/

Hrm... doesn't look like it to me. Looks like its running. But wait... there is more!

https://en.wikipedia.org/wiki/Tesla_Dojo

Dojo supports the framework PyTorch, "Nothing as low level as C or C++, nothing remotely like CUDA". The SRAM presents as a single address space.

Imagine that!

 

[igs]

Dojo was dropped?

https://techcrunch.com/2024/08/10/teslas-dojo-a-timeline/

https://techcrunch.com/2024/08/03/t...-plan-to-build-an-ai-supercomputer-explained/

Hrm... doesn't look like it to me. Looks like its running. But wait... there is more!

https://en.wikipedia.org/wiki/Tesla_Dojo



Imagine that!

Lol how long did it take you to Google that? Don't get your panties in a bunch bro.

 

[M0unt41nm4n]

?

 

[Protondecay123]

Since the FSD update, I have been 'testing' FSD out nearly every day. However, I have to frequently disengage due to harsh/ late stopping, unpredictable behavior in intersections, and riding the left side of the lane to the point of hitting a plastic barrier. This has been so disappointing that I decided to no longer use FSD on the truck.

Yet, since the update I have also had to chance to use FSD on the 3, X, and S. The experience on each has been significantly better. Driving is much smoother and actions are more natural. Furthermore, you can customize the driving style on each where you cannot on the cybertruck. Visualizations are also better and more detailed.

What is going on here?? Does the cybertruck run on an entirely different dataset or program? No excuse for FSD being this subpar only on the cybertruck

Since the FSD update, I have been 'testing' FSD out nearly every day. However, I have to frequently disengage due to harsh/ late stopping, unpredictable behavior in intersections, and riding the left side of the lane to the point of hitting a plastic barrier. This has been so disappointing that I decided to no longer use FSD on the truck.

Yet, since the update I have also had to chance to use FSD on the 3, X, and S. The experience on each has been significantly better. Driving is much smoother and actions are more natural. Furthermore, you can customize the driving style on each where you cannot on the cybertruck. Visualizations are also better and more detailed.

What is going on here?? Does the cybertruck run on an entirely different dataset or program? No excuse for FSD being this subpar only on the cybertruck

???
The FSD is significantly better on mine than my 2024 Model Y. Are you sure?
???

 

[SeattleTom]

Totally agree with OP.

CT FSD doesn’t stop smoothly (slamming on the brakes when approaching cars, stop signs, etc.). It cuts corners when turning, un safely entering the opposing traffic lane unnecessarily. It hugs the left side of the lane (sometimes even on straight roads) moving the truck needlessly close to opposing traffic.

It’s nuts they can’t figure this out.

 

[3cyberbeast]

Since the FSD update, I have been 'testing' FSD out nearly every day. However, I have to frequently disengage due to harsh/ late stopping, unpredictable behavior in intersections, and riding the left side of the lane to the point of hitting a plastic barrier. This has been so disappointing that I decided to no longer use FSD on the truck.

Yet, since the update I have also had to chance to use FSD on the 3, X, and S. The experience on each has been significantly better. Driving is much smoother and actions are more natural. Furthermore, you can customize the driving style on each where you cannot on the cybertruck. Visualizations are also better and more detailed.

What is going on here?? Does the cybertruck run on an entirely different dataset or program? No excuse for FSD being this subpar only on the cybertruck

Yup it’s shit!

 

[REM]

lol FSD is vision based it has no idea what inertia is nor does it care.

This is definitely not correct. FSD uses every applicable physics telemetry it can get it's hands on.

On another note, I believe the Cybertruck is modeling another FSD suite (maybe the Model X?), albeit with obviously parameter adjustments for weight, size, turn radius etc.

I use it far less than my M3 since there have been quite a few scenarios that made me look like a complete jerk.

Sponsored