Andrew discusses the limitations of language models in understanding human pragmatic inferences. He highlights the potential of RL models from human feedback in enhancing pragmatic understanding. Grounding problem in large language models is explored, emphasizing the need for better integration into real-world contexts.
