Lab 03

LLM Anatomy

A conceptual map of how a language model turns input into output: tokens, embeddings, attention, a probability distribution, one token at a time.

Conceptual model. This describes the published architecture of transformer language models in general. It is not a view inside any specific product, and none of it exposes a model's internal state or hidden reasoning — no such view is available from the outside.

The journey of one token

Stage 1 of 8 · Input

You give the model some text. That text is the entire input — nothing else.

Analogy

Handing someone a page and asking them to continue writing it.

Why it matters

Everything the model will use has to be in this text. There is no side channel.

1 / 8

Reading the result

What just happened: You are at Input. The row above shows where that sits in the loop that produces a single token.
Why it matters: Everything the model will use has to be in this text. There is no side channel.
Where you meet this: Every chat interface you have used runs this loop for each fragment of text it streams back to you — which is why answers appear a piece at a time.