I do find that current LLMs are quite bad at design problems and answering very specific questions for which they may lack sufficient training data. I like them for general Q&A though.
A different architecture or an additional component might be needed for them to generalize better for out-of-training-distribution questions.
