In the spirit of Why Can’t ChatGPT Draw a Full Glass of Wine?
How about this:
Credit for idea: Denis O. on LI (I shan’t link so as to encourage discussion here)
We all know that LLMs struggle with negative prompts which may be a fundamental flaw with the “attention” mechanism which is responsible for so much magic otherwise. Can anything be done about it?
And yes, sometimes it works and you can modify the prompt to make it more likely that there is no elephant, but the instruction is crystal clear as it is so why the lack of reliability?