The author seems unaware of how well recent Apple laptops run LLMs. This is puzzling and puts into question the validity of anything in this article.
And as far as I understand it, an Air with an M3 is perfectly capable of running larger models (albeit slower) if it had the memory.