Why Large Language Models Won't Replace Human Coders - The New Stack