Comment by dpark
3 hours ago
They aren’t trained on compressed plaintext so I’m not sure of the relevance there. But regardless it’s my understanding that’s modern models are trained with orders of magnitude more storage than their parameters require. But it’s possible I’m incorrect. This is getting to the fringe of my knowledge of concrete LLM details.
The relevance is because the LLMs are storing information, not the explicit text, so we want to know how much actual information they need to store (this being an information theoretic argument). The representation in the parameters doesn't necessarily need 1 parameter per character, if the text is highly redundant.