It adds weight to my suspicion that the real key to being "productive" with generative transformers is not caring that the output's wrong.