You laugh, but the requirements are actually very different between computer vision and other pre-2023 "AI" tasks and LLMs.

LLMs are hopelessly memory bandwidth bound. Most other models are compute bound.

AI goes through cycles where whichever field is currently making the most advances gets all the hype. But just because LLMs are currently flashy doesn't mean that object detection, image segmentation, facial recognition, image generation, text-to-speech, speech-to-text or anomaly detection are suddenly less useful.

I think BlueCruise depends on tbe cloud. Not sure about other implementations like Tesla's