These have historically often been thought of separately: knowledge representation was the data, and language was the interface.
To at least some degree, maybe language is enough for both. Language doesn't just express concepts, but it contains them.
To predict the next word very accurately for a huge variety of texts, a model must have some amount of real understanding. Sure, it may look different than human understanding.
But there’s no way around this. GPT could not predict so accurately if it didn’t have any understanding the world.