I believe GPT has been demonstrated to be usable for machine translation of long passages without "losing the thread." As long as the input has structure, and GPT just has to "follow along" the structure of the input when generating the output, then it the output is structured intelligibly.

Also, I've seen GPT used in recursive fill-in-the-blanks approaches, where it does equally well, since the "skeleton" of its answer is already there.

I imagine the task of "translating" between natural language, and a set of expert-system assertion predicates (or vice-versa) would work very well for GPT, and be "scalable."

I'd love someone to go ahead and try this experiment, actually. It could potentially be meta-learnable in just a few prompts.