Suppose this was true, an extremely small language model would outperform small language models. Sounds like inventing a compression which compresses any input

Suppose this was true, an extremely small language model would outperform small language models. Sounds like inventing a compression which compresses any input by at least one byte. Sorry for the snark, but the linked article doesn't explain why this obvious inference is false in this case.