Well, one glaring issue with the assumption of the quality of LLM output being mostly dependent on a large volume of examples online would be Sturgeon's law.