That probably plays into it as well. I have yet to see any convincing evidence that contradicts LLMs being mere pattern parrots.
My personal benchmark for these models is writing a simple socket BPF in a Rust program. Even the latest and greatest hosted frontier models (with web search and reasoning enabled!) can only ape the structure. The substance is inevitably wanting, with invalid BPF instructions and hallucinated/missing imports.
imho these tools are great i fyou know what you're doing, becasue you know how to smell test the output, but a footgun otherwise.
It works great for me, but it is necessarily an aid learning tool more than a full on replacement, someone's still gotta do the thinking part, even if the llm's can cosplay -reasoning- now
Low level programming means actual -thinking- about the system, resources, and language decisions etc
Even humans struggle with it, Its much easier to build a website than say a compiler, for anyone, humans and llm's included