They store the data on each car and query it from the mothership, then send only the most relevant samples back.

There is indeed way more data captured than can be sent back, or even stored long term on the car.

It's not live streaming all the data. They have an independently running AI that looks for events/problems and streams problematic areas. The cars do this by default.

In theory they could train the model in a distributed way on the cars themselves.

No this isn't really possible. The computation required for training is extensive compared to the computation required for simply running the system.

I disagree, federated learning is a very valid approach. The cars have a ton of computational power especially with HW3. (Although I think they don't have high precision floating point - the chips are for inference)

Federated learning is a lot like distributed training of a neural network.

The trouble with distributed training is that it's not as fast to converge as simply running more training steps. It is basically like increasing the batch size.

Also, it sounds like a general hassle for Tesla to use federated learning. I believe they need to be carefully auditing and labelling their training data for almost all their tasks. Perhaps some like depth-estimation don't require labelled data.

You could train and exchange data while charging.

"You are the product."