Explain it from memory

Can you explain how a model answers with nothing in front of you?

In the first lesson you typed “What is a cat?” and could not say what happened next. This is the same question again, with no diagram, no notes and no phone.

You have built a tokenizer, embeddings, attention, a Transformer, a training loop, a sampler, a KV cache, a retriever, a fine-tuning run and an agent. Each part made sense on the day you built it.

Explaining the whole pipeline is harder. Imagine sitting across from someone who has never studied any of this, with nothing on the table between you, and telling them how text goes in and an answer comes out.

Recognising is not the same as knowing

Pause and predict Which makes something stick better: reading it again, or trying to recall it and failing?

A skeleton to hang it on

Do not start with twelve stages. Twelve are hard to hold as a flat list. Start with four moves and fill in each one as you explain it. If you blank during the challenge, come back to these four.

Pause and predict Before you look: if you had to split the whole pipeline into four moves, what would they be?

And for everything built around the model, one question sorts it out: does it change the prompt, the weights, or the code around the model?

What you should remember

1. The pipeline: text → tokens → ids → embeddings + position → attention inside stacked Transformer blocks → logits → probabilities → sampling → next token → repeat.
2. The model is a function from the tokens so far to next-token probabilities. Its knowledge is implicit in parameters set by gradient descent on next-token prediction.
3. RAG changes the prompt. Fine-tuning changes the weights. Tools add a loop in which your code acts and the model only writes text.
4. It is fluent and it makes things up for the same reason: it generates what is likely, and nothing in the mechanism checks truth.
5. Understanding fades unless retrieved. Come back and do this challenge again in a week.