How does the neural network translate down to the individual cars?

[IamGaryGnu]

Hi all,
Here's a pretty geeky question in case anyone knows the answer.
I used to program in ia former life - and i can envision subroutines for say, autopilot, and then FSD - "if you see an object, avoid it" - okay, kinda got that.
Then Tesla bought a multi billion dollar or whatever supercomputer with neural networking/AI - and my understanding is they throw a hundred thousand left turns at it, and the thing can then eventually recognize a good left turn - something like that.
I've seen google answers on neural learning - put in a bunch of inputs you want, and eventually you'll get output.
okay, fine.
the question i'm struggling with is, "okay, my car doesn't have direct access/link to the Tesla supercomputer" - going one step further, my belief is every time there's a firmware update or at least an FSD beta update, all of the FSD code is downloaded to the car, without need for an internet connection while driving - okay, great - but HOW is the neural computer's "output" translated into something the car can use - does it actually have as part of the output code to make subroutines, or what?
again, i can sort of see how for the massive supercomputer, it learned from the examples. but how does that get reflected into our "offline" cars?
thanks for anyone who can shed light on this - one of my curiosities, along with whether the light stays on in the fridge when i close the door

 

[GtiMart]

I'll give a general AI/ML answer and others can add details later. Typically you prepare a dataset, you label your dataset, you build your neural network and you do the "training" part using your prepared dataset. I believe it's that training that is done on the super computers. Once training is done, and its behavior verified with test datasets, you can "run" the model that has been generated. Treat that as an output from the supercomputer, that is transferred elsewhere to be used. I believe the model gets to run on the onboard AI computer, currently at Gen 3 if I am not mistaken.

EDIT. You later learn some more from the execution on the cars, and that is information that you send back to the mothership so that data scientists can analyze. Some of that information will be used to re-train the model again, to enhance it. That newer model output will eventually make its way into the car with a software update.

 

[IamGaryGnu]

I'll give a general AI/ML answer and others can add details later. Typically you prepare a dataset, you label your dataset, you build your neural network and you do the "training" part using your prepared dataset. I believe it's that training that is done on the super computers. Once training is done, and its behavior verified with test datasets, you can "run" the model that has been generated. Treat that as an output from the supercomputer, that is transferred elsewhere to be used. I believe the model gets to run on the onboard AI computer, currently at Gen 3 if I am not mistaken.

EDIT. You later learn some more from the execution on the cars, and that is information that you send back to the mothership so that data scientists can analyze. Some of that information will be used to re-train the model again, to enhance it. That newer model output will eventually make its way into the car with a software update.

thanks - this is a lot of what i was looking for - my real question is how does it go from the mothership saying, "okay, this is behavior we want" to code the car can understand - i.e. i think the Teslas run on Linux. Fine. that means we need some sort of "neural net to Linux translation" - at least in my understanding.
I haven't been able to find anything about this anywhere, though admittedly I haven't watched the full FSD videos that i think are 3 hours long. I'm not sure if it mentions that last NN to car programming aspect.
thanks again!

 

[GtiMart]

I learned those aspects by working with our AI/ML team but also through a few training videos. I did not go as far as the details of how to run the model. With that said, if you put it very simply: the training assigns weights to the different nodes in the neural net. You only need to transfer those weights (Values) to the place where you want to run the model. Obviously the model structure must be known, but it seemingly takes a lot less power to run such a model than it does to train. You can look at things like PyTorch, Tensorflow etc if you want to learn more. Here, I'm getting out of my knowledge zone so will let others bring additional details.