I wonder how much utility of such special-purpose languages will keep getting diminished as AI coding becomes a staple in programming and software engineering.

With current LLM I am not sure. Using macro, you can often spare a lot of explicit code boilerplate. So token efficiency. On the other hand, will LLM shine in such a paradigme, where picking the right abstraction is more useful than generating large amount of lines?

Maybe it's a lake of training set, but just roaming the fine article, it already mention a LLM generated project which is more looking like Java idiomatic rather than something authors would consider elegant Rhombus code.