I’ve heard it said before on HN that this is not true in general because more tokens in familiar patterns helps the model understand what it’s doing (vs. very terse and novel syntax).

Otherwise LLMs would excel at writing APL and similar languages, but seems like that’s not the case.