How Language Models Work

Take the lid off the black box.

Thirteen short interactive lessons, three to five minutes each. No math required, no hype allowed. You’ll click buttons, break sentences, and leave knowing exactly what the robot is actually doing. Works with zero internet.

Short on time? The five marked core lessons are the whole story in about fifteen minutes. The rest go deeper.

1. 01Next-token prediction
   It’s guessing the next word. That’s the whole trick.→
2. 02Tokens
   The model reads pieces, not letters.→
3. 03Embeddings
   Meaning becomes geometry.→
4. 04Attention
   Every word watches every other word.→
5. 05Training
   A trillion tiny nudges.→
6. 06Scale
   The same recipe, a much bigger pot.→
7. 07The assistant act
   From autocomplete to answering.→
8. 08The context window
   Its whole memory is the prompt.→
9. 09Hallucination
   Confidently finishing a sentence that was never true.→
10. 10Prompting
    Steering by choosing the words it continues.→
11. 11Limits
    The failures are the mechanism showing through.→
12. 12Tools
    When predicting isn’t enough, hand it better text.→
13. 13The whole picture
    One loop, end to end.→

Press M anywhere for the lesson map. Presenting? Toggle presenter mode from the map, or add ?mode=presenter to any lesson URL.