lesson 1 of 9 · how a model works

One pass, one token.

A model writes one token at a time. Each token is one full pass through the model, and the next token starts the pass again.

new words

Pass: One run of the text through the whole model, from the first step to the last. It gives one new token.
Weights: The numbers inside the model, set by training. A pass reads them and never changes them.
LM head: The last step, LM for language model. It gives every token the model knows a score: how well it fits next.
Inference: Running a finished model to get an answer. Only training changes the weights.

the lesson

Your message, the pages your AI read, and the conversation so far go in as one text. The model cuts it into tokens. Each token becomes a list of numbers, a vector. The vectors go through a stack of layers. The last step scores every token the model knows, and one is picked.

Claude's shape is not published. Llama 3 8B, an open model, has a published shape, so this series counts on it:

Step	Goes in	Comes out	Lesson
Tokenizer	text	token numbers	2
Embedding matrix	one token number	a vector of 4,096 numbers	3
32 layers	one vector per token	one changed vector per token	4, 5
LM head	the last token's vector	128,256 scores	6
Pick	the scores	one token	6

That is one token. It joins the end of the text, and the whole pass runs again for the next one. An answer of 500 tokens is 500 passes. Claude writes its answers one token at a time too, which is why you see them appear piece by piece.

Two facts follow. The model keeps nothing between calls: each call sends the whole conversation again, and the model reads it from the first token. What a product calls memory is text that someone puts back into the window. And the weights never change while it runs: a conversation cannot teach the model anything; only a new training run can.

One pass of an example text. The cut into tokens is an example; the sizes are Llama 3 8B's.

try it

Open a new Claude window. Type this and press Enter:

Do you remember anything from a conversation we had yesterday? If something seems remembered, where did that text come from?

you should see

No memory of its own: a new session starts from an empty window.
Anything it seems to know about you came in as text at the start, such as an instruction file like CLAUDE.md, a page it read, or a summary.