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

[igs]

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.

That's the old version. The new one has no concept of what a car is, what a lane is, what a stop sign is, etc, let alone physics. It's just pixels in.

Sponsored

 

[M0unt41nm4n]

It's just pixels in.

Do you even know what that means? If you get pixels in... what then happens?

Here @igs ... pixels come in. It's run through a model(or models), made up of convolution neural networks (CNN) as well as pixel pooling. Those work together along with sliding windows of the entire image. This identifies, yes, stop signs, people, road, cars, other signs, etc.



There is a nice technical explanation of how it works. I will add that it also takes data from brakes, steering wheel, etc and feeds that to the model.

You are very wrong on all your claims.

 

[igs]

Do you even know what that means? If you get pixels in... what then happens?

Bro, stop googling random links that have nothing to do with anything just to pretend you know what you're talking about.

Remember when Tesla had teams of people tagging objects for FSD to recognize? That's all gone.

 

[M0unt41nm4n]

Bro, stop googling random links that have nothing to do with anything just to pretend you know what you're talking about.

Remember when Tesla had teams of people tagging objects for FSD to recognize? That's all gone.

Hahha... They don't need to tag objects to recognize. Thats all done! Their model knows what a stop sign looks like now.... OMG. Do you know why people had to do that? And I beg to differ that they still don't have to tag items now and then... they still need to do that when items get missed.

I give up... dude... you are totally clueless on how all of this works

 

[igs]

Hahha... They don't need to tag objects to recognize. Thats all done! Their model knows what a stop sign looks like now.... OMG. Do you know why people had to do that?

Hold up. You're telling me the "model" (as opposed to "NN" (hint) lol) now recognizes every object it will ever need to in every possible circumstance? That's quite an amazing feat they've accomplished then.

 

[M0unt41nm4n]

Hold up. You're telling me the "model" (as opposed to "NN" (hint) lol) now recognizes every object it will ever need to in every possible circumstance? That's quite an amazing feat they've accomplished then.

Did you read what I posted? Paste in the rest of the sentence, bud.

 

[igs]

Yo Elon, @M0unt41nm4n says your work is done. You can cancel the supercomputer cluster at Giga Texas. ?

 

[M0unt41nm4n]

?

 

[Crissa]

It's evident that @M0unt41nm4n works within the AI industry.

No, it isn't. They drop a few keywords here and there but...

...Their understanding or at least what they've shared here is shallow at best.

Learning for video games, for instance, depends highly upon predictable environments and repeatable actions. It's not the same as all as the AI that mastered Go.

And to be honest, I don't know what they're trying to argue here. The graph that something like tensorflow builds has predictive value, but you can't just tell it to adapt to a different set of inputs, it would be waay off.

Then again, the fact that the Cybertruck's FSD was mapped off to one side kinda implies they did something to the input data that it was trained upon that's off slightly to the side.

-Crissa

 

[Garee]

What do you mean when you say the cyber truck cannot be customized driving mode? I do it under dynamics setting.



Thank you
[/QUOTE]

 

[REM]

That's the old version. The new one has no concept of what a car is, what a lane is, what a stop sign is, etc, let alone physics. It's just pixels in.

Utter nonsense. At this point you are just trolling people.

 

[M0unt41nm4n]

No, it isn't. They drop a few keywords here and there but...

...Their understanding or at least what they've shared here is shallow at best.

Learning for video games, for instance, depends highly upon predictable environments and repeatable actions. It's not the same as all as the AI that mastered Go.

And to be honest, I don't know what they're trying to argue here. The graph that something like tensorflow builds has predictive value, but you can't just tell it to adapt to a different set of inputs, it would be waay off.

Then again, the fact that the Cybertruck's FSD was mapped off to one side kinda implies they did something to the input data that it was trained upon that's off slightly to the side.

-Crissa

I explained why its off slightly to the side. They use most of the data from the smaller vehicles. They need to take that as a base and train the bigger CT... which is exactly what they are doing. I was very clear on how this works.

What is being argued, is I explained how the AI works. The other dude is making sideways comments that have nothing to do with AI. Yes, @Crissa ... this is what I do for living

 

[M0unt41nm4n]

So I will do one post that explains this in detail in layman's terms. Its up to those who are interested (@Crissa I'm looking at you) to do their research for further details.

In AI, you build a model based on a number of methodologies. For images/video there are complex models based on what is called a CNN (Convolutional Neural Network <- Google that) and it is usually stacked with image/pixel pooling and it uses the concept of "sliding windows" which takes a single video image and uses the above to use a window of pixels, test it for recognition, and slides that window to the right and down. As it moves, each time it says "Do I recognize what is here and does is conform to something that interests me"? If it does, it flags it, and continues through the image. At the end it recognizes it and is able to make sense of it of the objects in the image. For those interested you can read up on it here:

https://machinelearningmastery.com/object-recognition-with-deep-learning/

What it ends up with is a model that identifies objects in the image. Not only does Tesla take a similar model to recognize objects, it also adds other features. Features are what is known as "input" information to yet another Model. This likely includes brakes, speed, steering, and many other data points in the vehicle. The output from the CNN above along with what I just explained goes into yet another model that makes sense of the controls. So it's able to learn based on the image and controls "what to do". i.e. if it sees you hit the brakes due to a vehicle being in front of you, it makes note of that and learns when its appropriate to hit the brakes. That ends up the output and a very very high level for how the AI works. Digging in, it's extremely complex. i.e. ImageNet, Google Inception, Resnet are made of very very complex models that stack the CNNs and the pooling. Why is this important? Because as an example, those models have been trained on a crap load of data, using years and years of computing resources (years meaning horizontal years), to build a model that is good at identifying objects. This is what creates a base model that can be used by individuals to produce image recoginition. An explanation can be shown here:

https://towardsdatascience.com/classify-any-object-using-pre-trained-cnn-model-77437d61e05f

Why is this important? Because training a complex model takes an insane amount of computing resources which is costly. An individual person with their laptop could not train a sophisticated model without an extreme amount of time. It literally takes computing years in the cloud to produce such a model. What's cool is once this model has been created, an individual can actually use that model and train it specifically because 98% just works...i.e. the complex part of object identification has been done.

Another example is LLMs. So many LLMs allow you to use a massively trained model, and then train it specifically on some documents that you have to be just an expert on those documents. It already understands language, so now you are teaching it the knowledge of the contents of your specific docs. You can also teach these LLMs to talk like another person, say "Homer Simpson".... the explanation is here:

https://replicate.com/blog/fine-tune-llama-to-speak-like-homer-simpson

It takes a base LLM, and tunes/trains it to speak like Homer Simpson. They didn't need to train it for millions of hours to relearn how to be a LLM, it reused a base model and tweaked it.

So how this works with Tesla is the following...

Ultimately, Tesla got where they are today by running their models on the Tesla S, 3, X, and Y vehicles over a very long time. They have trillions of records in data that trained for what we know today as "FSD". The key here is most of that data was trained on a smaller vehicle (the S, 3, X, and Y are all significantly smaller than the CT). The CT was able to get FSD quickly because it used that model that was trained by all the other vehicles as a base. This is also probably why Tesla waited a short time to release FSD for the CT as it had to train it at minimum to get it to mostly drive semi-decent. Imagine if it has to train it from scratch. Rest assured, if that was the case, you would be seeing FSD for probably 3-5 years just to gather the same amount of data it already had.

When you see some of the issues that the CT, has, you will probably note it makes some strange decisions. But when you look at it, it's actually making decisions that were probably more appropriate for a smaller vehicle. An example is making the left hand turn with a car stopped perpendicular in their left lane as you make the turn. Notice, many times you may actually clip the car and you may have to grab the controls. The FSD model on that thinks your CT is not as wide, so it cuts that turn a lot closer. This also may be the reason that it hugs the center line a bit more. The FSD model thinks its centered but due to the wide of the truck, it thinks it has more room when in reality it does not. Some folks say, it slams on the brakes when slowing behind another vehicle. Think about that... it is using a model for a vehicle that is 2000lbs lighter than the CT which doesn't need as much space to slow down.

The FSD on the CT right now is "good enough" to make 98% of the decisions correct. But it's learning. Thats why you get the upload of data to Tesla, where it takes video and other feature data and sends it to the mother ship to train on the current model. When it asks you to make a voice memo for them, it's likely that it may go to humans for tagging (supervised training). i.e. did it almost hit someone? Did it miss a speed sign? Did it miss its right turn? This may likely go to a human, and they will draw around a speed sign, or a stopped car, or a pedestrian you almost hit (hence @igs is wrong that nobody does it.... I can promise you there are humans tagging). Sometimes, they may run your video through another model to figure out what may have happened without human intervention in the essence of streamlining the issue, where that model can make heads or tails of what occurred.

What does this mean? It means your CT is using a base model trained on other vehicles and your driving actually is helping it get trained for its size/weight/etc on their new Dojo super computer. Links can be found earlier in this thread on how that works. What this also means is as they train this, you will get the more refined model, and your CT will make better decisions for its size/wight/mass, etc specific to that vehicle.

I hope this helps folks understand how this works and can see it's not "shallow". Its extremely complex and you all can go down the rabbit hole of starting with those links go really deep into his this stuff works. Its actually incredibly fascinating.

 

[igs]

Utter nonsense. At this point you are just trolling people.

You mean Elon is. ?‍

 

[igs]

Elon: "We never taught it what a car was, or what a person was, or a bicyclist"

Sponsored