The LLM Field Manual

The book

Most explanations of large language models either skip the mechanism entirely or lose you in linear algebra before you have built any intuition worth keeping. This book takes a third path. Marcus Hale works through sixteen interconnected concepts — from how a token is defined to why the largest model is not always the right model — using plain language and mental models grounded in work like Vaswani et al.'s foundational attention paper and Anthropic and OpenAI's public research. The aim is not to make LLMs feel impressive; it's to make them feel legible.

What you'll learn

• How the transformer's attention mechanism routes information across a sequence, and what that means for how a model processes a prompt — not as text, but as vectors shaped by training
• What tokens are, why they are the true unit of computation, and how tokenization boundaries explain cost, arithmetic failures, and other behaviors that otherwise look arbitrary
• The generation mechanism behind hallucination — why confident wrong answers are a structural outcome of how models produce text, and what that means for working with them in practice
• What temperature, top-p, and other sampling parameters actually control in the output distribution, and when tuning them is and isn't the lever you want
• How tool use, function calling, and chain-of-thought reasoning extend what a model can do without modifying its weights
• How to evaluate whether a model is genuinely better for a specific task, beyond benchmark scores that may not reflect real-world performance

Who it's for

Developers who have shipped something with an LLM and found themselves reasoning about it by feel rather than by understanding. Also for technical readers who want a clear-eyed account of the current model landscape — where Claude, GPT, Gemini, and the leading open-weight models stand in 2026, and which claims about where the field is heading are grounded in mechanism versus marketing.