The ability to tell letter occurrence in a word is not a good measure for the quality of a large language model.