Why?
What problem are we trying to solve, and why isn't the current approach enough?
A free learning resource from Othermind
Understand how LLMs work.
One idea at a time.
A practical, plain-language guide to large language models—from the first token to training, reasoning, and real-world systems. No research background required.
A model, piece by piece
01TextWhat goes in
02Tokens & embeddingsRepresent meaning
03Attention & transformersBuild context
04Training & inferenceLearn and generate
The learning method
Not a list of parts.
A chain of ideas.
Each concept arrives because the previous step leaves a problem unsolved. That way, the architecture becomes something you can reason about—not just a set of terms to memorize.
How?
See how the next building block addresses that problem, first intuitively and then technically.
What?
Follow the result with small worked examples, real numbers, diagrams, and the engineering view.
Inside the handbook
From first principles
to production AI.
Start at the beginning and build a working mental model, or jump to the chapter you need. The guide grows from fundamentals into the systems and decisions that shape modern AI.
Build a GPT
Tokenization, embeddings, attention, transformer blocks, and how they fit together.
Train it
Training data, loss, gradients, backpropagation, and the training loop.
Run, adapt, and shrink it
Inference, the KV cache, fine-tuning, LoRA, RLHF, DPO, and quantization.
Explore modern architectures
Transformer families, attention variants, Mixture of Experts, and multimodal models.
Reason, serve, and scale
Reasoning models, Qwen and DeepSeek, model serving, speed, hardware, and scale.
Put AI into practice
Production AI systems, agents, retrieval, decision models, calibration, and evaluation.
Made to learn from
Read it in order. Keep it as a reference. Or study a chapter with an AI assistant as your tutor.
Open by design
Learn in the format
that works for you.
The source is written in plain Markdown, so it’s easy to read, search, reuse, and share. A full PDF is also available for offline reading.
Read online ↓
Go chapter by chapter in the public repository, with references and worked examples alongside the text.
Download the PDF ↗
Get the complete 355-page handbook as a single document for reading offline.
Learn with an AI assistant ↓
Use the README and one Markdown chapter to turn an assistant into an interactive tutor.
Read on Othermind
The handbook,
right here.
Choose a chapter to load the latest Markdown directly from the public GitHub repository. The source stays there; this reader fetches the chapter when you open it.
The chapter reader
Choose where to begin.Select any chapter. Its latest version will load from GitHub.
A practical study guide
Learn with an AI assistant.
One chapter at a time.
The handbook is designed to work as a source for an interactive tutor. Give your assistant a small, focused set of files, then ask it to guide you through the ideas and check your understanding as you go.
Start with the question you want to answer.
Read in order the first time, beginning with The Big Picture. Later, use the topic index to find a subject.
Attach the README and one chapter.
Use the Markdown files: the handbook is about 140,000 tokens, so a focused chapter is easier for an assistant to work with than the entire source or PDF. Attach the README.md and optionally include AGENTS.md for guidance on using the material.
Ask, answer, and follow the thread.
Ask for an explanation, a worked example, or a comparison. Answer the tutor’s check questions, and ask it to slow down, use another analogy, or show the numbers when a step is unclear.
Separate lasting ideas from current details.
For recent model names, prices, or software versions, ask the assistant to search current sources and distinguish those findings from what the handbook says.
A PROMPT TO BEGIN
Use the attached README and chapter as your main sources. Teach me the chapter in its problem-first order: what problem we have, why the previous step is not enough, how this idea solves it, and what it produces. Start with intuition, then bring in the math and engineering details. Take one idea at a time, use the chapter’s examples and numbers, and ask me a short question after each section. Wait for my answer before continuing. If I’m confused, explain it another way. For newer facts or specific model names, prices, and versions, search current sources and clearly separate them from what the handbook says.Attach the handbook README and the chapter you want to study, then send this prompt.