Applied AI

Course map · S1 How LLMs work · Episode 5 of 12

The transformer

S1 E5 Explainer · 16 slides · 2 October 2026

  1. Slide 1 of 16, Cover.
  2. Slide 2 of 16, Why it matters.
  3. Slide 3 of 16, Mental model.
  4. Slide 4 of 16, Concepts.
  5. Slide 5 of 16, Concepts.
  6. Slide 6 of 16, Concepts.
  7. Slide 7 of 16, Data.
  8. Slide 8 of 16, Concepts.
  9. Slide 9 of 16, Data.
  10. Slide 10 of 16, For engineers.
  11. Slide 11 of 16, Where it breaks.
  12. Slide 12 of 16, For engineers.
  13. Slide 13 of 16, In practice.
  14. Slide 14 of 16, For leaders.
  15. Slide 15 of 16, Decision rule.
  16. Slide 16 of 16, Series map.

Lesson notes

The transformer stacks attention and feed-forward layers over token embeddings, and its shape explains why writing took 22 times as long per token as reading.

Decision rule

Choose the smallest model that passes your own tests, and budget latency on output length.

Ideas in this lesson

  • Next-token prediction
  • Attention
  • Latency

Take it further