Isn't an LLM essentially a stochastic ontology? Maybe that's why LLM's generalize so well to problems you wouldn't think would be amenable to next word prediction based on text analysis.