I have already seen the LLM hallucinate prompts from me - in this case, hallucinating being asked to switch to a different programming language - because it was

I have already seen the LLM hallucinate prompts from me - in this case, hallucinating being asked to switch to a different programming language - because it wasn't able to complete the task asked for in a satisfactory way instead of giving up and telling me it's not able to do it. If it doesn't already, I suspect training needs to include those no-solution scenarios and reward not overstepping bounds, or else we're going to see a lot more harmful side effects.