The following is written to be as penetrable as possible for a semi-technical audience, and includes an app to explore the output of an LLM. As a result some technical explanations are necessarily imprecise though I've worked hard to not say anything outright wrong. It also comes in two halves, with the latter introducing more technical details...