I’ve been trying to understand how LLMs like GPT-4 work on a deep level. Instead of a typical explainer, I wrote a short fantasy story where each part maps to a real component of the model — from tokenization to attention to training.

It’s called The Oracle of Lexiconia, and it hides the technical mapping until the end. An attempt to make the abstract intuitive.