Last updated: 2026-09-07

How LLMs work

Hitting Send does not fetch an answer from a shelf. The model writes chances for the next piece, a picker chooses, your loop glues it on, and the whole thing runs again.

Read in this order. Each page ends on a question the next page opens by answering. The math pages are not optional asides — they are the next page when the story needs a formula.

Reading order

  1. After Send, it is building a continuation
  2. The whole job is the next piece
  3. Tokens are not words
  4. Meaning as a list of numbers
  5. Math: the dot product
  6. Mix, clip, mix
  7. Math: mixing with a table of knobs
  8. Math: softmax
  9. Training is downhill on a measure of wrongness
  10. Math: loss
  11. Why looking-at-everyone beat reading-one-at-a-time
  12. Attention, by hand
  13. Math: why divide by √dₖ
  14. One transformer block
  15. Stack the blocks, pick a piece
  16. Document simulator, then assistant, then practice
  17. What this means when you build

Sources you do not need to finish: references.

Send →