Improved automatic language detection would be even nicer. I.e. BlueSky _may_ recognize it's some language X I write in and consequently may propose to switch. But usually I just fail to see it. Besides, it's cumbersome when I think it could be automatic.
I have no doubt also you wouldn't have looked into it.
I've been also looking if some the newer models could be very little and purpose built for that. Maybe e.g. something done with github.com/microsoft/Bi.... Or have dictionaries of "most used words" in usual languages and "guess accurately".