Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models