The book · Plain English · No maths degree required
A language model never knows the answer. It only predicts the next token
- token41%
- word27%
- step14%
- move6%
That single idea explains almost everything about ChatGPT, including why it sometimes confidently makes things up. This book takes it from there.
How LLMs Actually Work, large language models explained, from hardware to hallucination.
The book for professionals who use ChatGPT, Claude or Gemini every day and want the real explanation, the genuine mechanism, from tokens to transformers, agents to alignment. No code. No condescension. No hype.
★★★★★ 5.0 on Amazon · Paperback & Kindle · By Alexander Lazutin
You use AI every day. Do you know what it's actually doing?
Millions of people now write, code, research and decide with ChatGPT, Claude, Gemini and Copilot, while treating the machine underneath as magic. That gap is where mistakes happen: misplaced trust, missed opportunities, and confidently wrong answers passed on as fact.
Spot the made-up answers
LLMs don't look facts up. They predict plausible text. Learn why fluent, confident answers can be false, and the tell-tale patterns that give hallucinations away.
See inside the black box
Tokens, training data, probabilities and attention, the whole pipeline from your prompt to the model's answer, told through analogies you'll actually remember.
Use AI with judgement
Once you understand what the model can and can't do, you know when to trust it, when to verify, and how to get dramatically better results, with any AI tool.
Inside the book: the full contents
This is not a book about how to use AI tools. It's a book about how they work, at the level of mechanism, written for smart professionals who want a genuine explanation rather than a simplified version of one. Fifteen chapters in three parts, plus a practical guide on where to go next.
Introduction · The moment we're in
How language models came to exist, what drove the field to this point, and what you will understand by the last page.
Part one · The raw materials
- The Hardware: What Runs an LLMData centres, GPUs, TPUs, energy, and the physical scale of building and running a frontier model.
- The Data: What LLMs Are Trained OnWhere training data comes from, how it is filtered, why its composition shapes the model, and the copyright questions it raises.
- Tokens: The Atom of LanguageWhat a token is, how tokenisation works, and why the quirks of this process explain some of the model's strangest failures.
Part two · The engine
- Embeddings: The Geometry of MeaningHow tokens become vectors, why meaning has geometry in high-dimensional space, and what the embedding space reveals about bias.
- The Transformer: The Architecture That Changed EverythingAttention, queries, keys, values, multi-head mechanisms, layer stacking, and the 2017 paper that made everything else obsolete.
- Generation: How a Model Actually Produces TextThe autoregressive loop, temperature, sampling, context windows, system prompts, and the mechanics of prompt engineering.
- How a Model Learns: Training From ScratchPre-training, loss functions, gradient descent, the engineering reality of a training run, the pioneer labs, and who funds them.
- From Predictor to Assistant: Fine-Tuning and AlignmentInstruction tuning, RLHF, Constitutional AI, sycophancy, and the gap between appearing helpful and being helpful.
- Scale and Emergence: What Happens When Models Get BiggerThe scaling laws, the Chinchilla correction, emergent capabilities, in-context learning, and the debate about what emergence means.
Part three · Capabilities, limits, and the landscape
- Beyond Text: Images, Audio, Video, and Multimodal ModelsDiffusion models, CLIP, speech recognition, voice synthesis, music generation, Sora, and what synthesis actually means.
- RAG, Memory, and Giving Models Access to the WorldThe knowledge cutoff problem, retrieval-augmented generation, vector databases, long-context trade-offs, and memory architectures.
- What LLMs Genuinely Can't Do, and WhyHallucination explained mechanistically, reasoning limits, the consistency problem, benchmarks, and a framework for calibrated trust.
- Agents and AI CoworkersWhat an agent is, tool use, the ReAct pattern, failure modes, and how ChatGPT, Claude, Gemini, Copilot, and GitHub Copilot differ.
- Open, Closed, Small, and Everywhere: The Model Landscape in 2026Open weights versus closed models, quantisation, distillation, Mistral, prompt injection, jailbreaking, and the geopolitics of AI.
- The Best Uses, the Real Picture, and What to Make of It AllWhere LLMs genuinely excel, where they predictably fail, the skills that matter now, and a clear-eyed view of the technology.
Ending · Where to go from here
Tools to experiment with, courses worth taking, papers worth reading, and people worth following.
Covers GPT, Claude, Gemini, Llama and Mistral, tokenisation, attention, RLHF, Constitutional AI, RAG, diffusion models, AI agents, scaling laws and more. See the full description on the Amazon listing.
Written for two kinds of reader
The curious professional
You work in technology, finance, law, consulting, healthcare, anywhere AI is changing how work gets done. You use these systems, make decisions about them and explain them to others. This book replaces vibes with mechanism.
Teams & L&D leaders
Rolling out AI across the business? Policies fail when people don't understand the tool. Give every employee the same clear mental model, order the book in bulk with volume pricing.
“The people who use these systems most effectively are not the ones who trust them most. They are the ones who understand them best.” , from How LLMs Actually Work
Think you already understand AI? Prove it.
Ten quick questions on tokens, training and hallucination. Most people score under 6/10. See where your mental model of ChatGPT breaks, then let the book fix it.
Start reading for free
Sample the book's approach with our plain-English explainers, the same ideas, in miniature.
How do LLMs work? A plain-English explanation
From your prompt to the answer in five steps: tokens, embeddings, attention and prediction, with zero jargon left undefined.
Why do LLMs hallucinate (and why ChatGPT makes things up)
The honest answer: hallucination isn't a bug being fixed next release. It's a consequence of how the technology works.
Frequently asked questions
Do I need a technical background to read this book?
No. It's written in plain English for curious professionals, managers and students. Every concept, from tokens and training to prediction and hallucination, is explained with everyday analogies, not equations.
Is this a book about how to use ChatGPT, or how it works?
How it works. Prompting tips date quickly; understanding the machine underneath doesn't. When you know why a model predicts, guesses and hallucinates, better use follows naturally, with any AI tool, not just today's.
Why do LLMs hallucinate, and does the book cover it?
It's a core theme. LLMs generate the most statistically likely next word rather than checking facts, so a fluent, confident answer can still be false. The book explains why this is built in, and how to spot and reduce it. There's a taster in our free explainer.
Can my company order the book in bulk for staff training?
Yes, bulk orders for L&D programmes, AI-literacy rollouts and onboarding are available directly, with volume pricing. Head to the For Business page.
Where can I buy How LLMs Actually Work?
On Amazon, in paperback and Kindle.