If choosing the "wrong" model, or not wording your prompt in just the right way, is sufficient to not just degrade your output but make it actively misleading and worse than useless, then what does that say about the narrative that all this sort of work is about to be replaced?