Is it the redundancy? Or is it because markup is a more natural way to annotate language, which obviously is what LLMs are all about?
Genuinely curious, I don’t know the answer. But intuitively JSON is nice for easy to read payloads for transport but to be able to provide rich context around specific parts of text seems right up XML’s alley?