I agree. Isn't this just utilizing the representation learning that's happened under the hood of the LLM?