Interesting. I wonder if it's just a matter of prompting properly or if GPT/GLM just doesn't have that training.