Every LLM Request Has Two Halves. Only One Uses Your GPU Cores

Paste a long document into ChatGPT and hit enter. Nothing happens for a second or two. Then the answer starts appearing, word by word, at a steady pace until it finishes. You have seen this hundreds of times. Most people never think about it. But those are two completely different things happening inside one request, running on two different parts...