I'm increasingly convinced that this is the #1 reason why frontier LLMs are still not capable of one-shotting high quality code They need to spell out their reasoning process in the completion in order to get the right result so when you style transfer that reasoning to code you get cluttered code
I've long suspected that when LLMs "code" (meaning: one shot a coding solution) what they're actually doing is a style transfer of their thinking process to code (no different than translating their inner abstract thoughts to English/French) In other words, "coding" for them is a low-rank adaption