>This could be right for the current architecture of LLMs, but you can come up with specialized large language models that can more efficiently use tokens for a specific subset of problems by encoding the information differently.

That's precisely what happens on the bad side of a S curve.