We argue that claiming some system S does/doesn't understand some target T requires a clear specification of both S and T. If someone says "LLMs don't understand language," does that refer to a class of systems, a particular trained model, GPT3.5? And what's meant by language?
