In particular, most of these require always on audio and in some cases video scanning and recognition. My distant memory is that you can record audio on a fairly small processor and power budget, but video is expensive no matter what you do, and running the sort of pattern recognition or worse yet machine learning needed to actually do anything useful with it would be murder on your battery if you were unplugged and on performance even if you weren't. (This might change with the wide availability of hardware specifically designed for these tasks (TPUs), but we're not quite there yet and I don't know how general that will ever be)

> very annoying

They should use machine learning to learn how to not be annoying then. This is possible, it's just big companies have too much momentum to make useful products nowadays

> highly battery-draining

Having demand will will be motivation for better battery research.