I actually would love to see the output of an LLM that used letters or syllables or phonemes. The output when it makes mistakes would be absolutely wild.