How does a machine finish your sentence?

A hands-on tour of what’s inside a large language model. No equations, no code: every idea is a toy you can poke.

live · a model’s entire job, on loop

That’s the whole trick: guess the next word, append it, repeat. Everything below explains how the guessing gets so good.

The journey

  1. TokensChopping language into pieces a machine can count.→
  2. Meaning as a mapEvery word gets coordinates. Nearby means similar.→
  3. AttentionEvery word looks back and decides what matters.→
  4. Neurons & layersTiny pattern detectors, stacked until they understand.→
  5. Prediction & samplingTurning hunches into words: temperature, top-p, and dice.→
  6. A sense of orderHow a model knows “dog bites man” isn’t “man bites dog”.→
  7. Mixture of ExpertsA committee of specialists, and only two get to speak.→
  8. Speed tricksKV caches, shrunken numbers, and guessing ahead.→
  9. TrainingWhere the knowledge comes from, and why models make things up.→