I am working on a bit of writing about the challenge of learning about LLMs by using them, and here is a prime example. The image is a screencap of Google’s flan-t5-base model on HuggingFace. The diagram gives an answer to a question that when demoed live does not produce the expected answer.