Post-training is where a next-token predictor becomes an assistant with a character, a set of values and the ability to reason through hard problems. This guide covers that process at research depth: the mechanisms, the derivations, the failure modes and the open questions as of late 2026.
Each idea arrives in the same order: intuition in one sentence → an analogy (and where it breaks) → an interactive you can poke → the math, with annotated equations you can hover → a short code sketch → pitfalls → what the frontier says → the questions worth asking about it. Every chapter ends with practice questions, flashcards and a verified reading list.
Full path (≈ 3–4 weeks at 1–2 h/day): chapters in order.
Short path: 00 → 04 → 05 → 06 → 03, then skim the "In one breath" box and the "Check yourself" questions of every other chapter.
Companion: The Alignment Field Guide covers why these techniques are not enough on their own: reward hacking, model organisms, oversight, interpretability, control.
Works offline: every library (D3, GSAP, KaTeX) ships with the repository. Your progress (chapters read, self-assessments) is stored only in this browser's localStorage.