One of the core ideas behind LLMs is that language is not a discrete space, but instead a multidimensional vector field where you can easily interpolate as needed. It's one of the reasons LLMs readily make up words that don't exist when translating text for example.