Understand LLMs from scratch.

No deep-learning background needed. Read these in order for the whole story, or jump to whatever you're trying to understand today. Each one ends with a door into the deeper, math-heavy lessons — open it when you're ready.

1. 1What an LLM actually is→The one idea behind every language model: it repeatedly picks the next token that best fits, using weights shaped by its training data.
2. 2The life of a model→A blank network becomes an assistant in four stages — pretraining, instruction tuning, and two kinds of reinforcement learning.
   the four stages, in depth

   • Pretraining: learning the worldread →
   • Instruction tuning: learning to answerread →
   • RL for alignment: learning to behaveread →
   • RL for reasoning: learning to thinkread →
3. 3How a prompt becomes an answer→The two phases of inference — reading your prompt (prefill) and writing the reply one token at a time (decode) — and why the second is the slow one.
4. 42017 → 2026: how attention evolved→A guided tour of every major change to the attention block since 2017 — tagged by whether it bought quality, speed, throughput, or memory.