> In a few years todays high end AI models will run on your watch

Although possible with cpu power, I dont think you will ever get enough ram in a watch to run a decent local LLM.

I also dont think the high ram requirements for running them will come down at all.