Atlas

Large Language Models

Positional Encoding

Attention has no sense of order. This gives it one.

You can jump straight in, but this star assumes The Transformer. Not recommended as a first stop.

Four layers of depth

Each layer ends with a quiz. Finish layer 4 and you own this concept.

  1. L1IntuitionAttention has no idea what order words came in unless you tell it. This is the story of how you tell it.6m
    40 XP
  2. L2MechanicsThe four schemes, mechanically: how each one actually computes a position signal and injects it into attention.10m
    70 XP
  3. L3CodeImplementing sinusoidal, learned, ALiBi, and full rotary position embeddings from scratch.14m
    110 XP
  4. L4FoundationsThe rotation matrix behind RoPE, why it produces a genuinely relative dot product, and the arithmetic of extending context.14m
    180 XP

33 stars in the atlas.