Adventures through a GPT-4 conlang hallucination Large language models are surprisingly bad at learning models of languages