How LLMs are actually built — and why most "how it works" threads stop too early