Tried that for a bit before.

Utterly terrible on less than top-tier hardware, which I have neither the budget or desire to acquire.

I'm not going to be locked into a system of perpetual FOMO.

I'm not missing out. Besides, I'm dead weight when it comes to this sort of stuff. Useful for ballast and that's about it.

Things that might change my opinion: Open-source models that I can train on my own dataset. (Which would be mostly books I've read that aren't trash.) Effective enough models for simple, local useful things like voice recognition to be able to be lazy about home automation stuff.

Basically, no black box stuff. No crazy hardware needed to carry on a better than 1960s robot level tasks (actually difficult).