Section checkpoint

How LLMs Work (Visual First)

10 questions drawn from the 5 chapters in this section. If most land on the first try, you are ready for the next section. If not, the explanation after each answer says which idea to revisit.

Checkpoint

10 questions · first-try answers earn the most XP

1. Why do API prices and context limits get quoted in tokens rather than words?

2. What are the three broad training stages most instruction-tuned models go through?

3. Your RAG system returns a chunk that is semantically perfect but far too long to paste into the prompt. What is the cleanest fix?

4. Two different tokenizers split the same sentence into 6 and 9 pieces. Which statement is true?

5. What does the KV-cache store and reuse during text generation?

6. Why does the chapter say you do not need to derive softmax to work with these models?

7. What is the risk of over-tuning with preference optimisation?

8. Why does the chapter say "behaviour is baked, knowledge is rented"?

9. Why do you need a held-out eval set for a fine-tune?

10. Your fine-tuned model scores no better than your best prompt on the eval set. What is the honest conclusion?

What this checkpoint covers

Next section: Your First AI App