still not convinced that large language models can actually "understand" what's being said to them