Ultimately there's a relationship between the preciseness in which you want to control something and the underlying information, as conveyed in language, to describe such precision.
Whether you use plain english, or code - ultimately to do things of sufficient precision you will have to be equally precise in your description. I'm sure someone with more time and more knowledge on such things have already formalized this in some information theory paper, but...
The point I'm making here is that this is great because a lot of people are doing "simple" things, and now they will be able to do those things without understanding the idiosyncrasies of Blender APIs, but I'm convinced that this will ultimately turn into something equally difficult as blender APIs to do novel things. and WHEN (not if) that happens, I hope users are prepared to learn the Blender APIs, because it will be inevitable.
edit:
one other thought. I think "language models" are not the right solution ultimately. I think kind of like AI didn't boom until the proper compute was available even though theoretical models and algorithms existed, language models are the crud solution.
once we have a loseless way to simply "think" what we want, then a "large thought model" will have less trouble, as there will be less ambiguity in what you want to what is said.
right now it's thought -> language -> model.
later it will be thought -> model.