September 27, 2026•2 min

Episode 0: Why Code LLMs From Scratch

m
mayo

What this series promises

Understand how LLMs (ChatGPT, Claude, etc.) work without needing a PhD in mathematics. Each episode starts from a simple question, explains the "why" before the "how", and uses Rust examples where they help.

Why "from scratch"?

Because reading articles is not enough. It's like cooking: watching a recipe doesn't teach you how to cook. Coding the 4 operations of attention yourself will teach you more than hours of reading.

What you will learn

  • How a word "looks at" the other words (attention)
  • Why Q, K and V are kept separate (specialization)
  • How meaning is turned into geometry (embeddings)
  • How to reuse an existing model (transfer learning)
  • How it all comes together to build a RAG

Who is this series for?

Anyone curious who can read simple code and wants to understand AI without getting lost in jargon. You don't need to be a mathematician, just patient.

The common thread

A single idea runs through the whole series: meaning can be represented by vectors, and those vectors can be compared by how close they are. Everything else follows from that.

  1. Step 2 of 7•5 min
    Episode 1: Attention Is Just 4 Mathematical Operations

    Behind Transformers and the attention mechanism there are only four simple operations: matrix multiplication, transposition, scaling and softmax. Here they are, in Rust.

  2. Step 3 of 7•2 min
    Episode 2: Q, K, V — Why 3 Matrices Instead of One?
  3. Step 4 of 7•3 min
    Episode 3: Embeddings — Absolute at the Start, Contextual at the End