How is it useless? If it has an NPU it is literally built to run models. You're just extremely biased in what you consider to be a useful ML model. For example

How is it useless? If it has an NPU it is literally built to run models. You're just extremely biased in what you consider to be a useful ML model. For example, for some strange reason you think only LLMs exist. The model must be as big as possible or else it is pointless. Training custom non-LLM models for specific tasks so they run on a resource constrained device? You must be insane.