I was pondering the other day if someone could convince one of the proprietary LLMs to spit out something like json-ld of its internal knowledge base to train other AIs.