Books

Reliable Generative AI bridges business use and technical architecture. It teaches the foundations of prompt design, RAG, agentic workflows, tool use, structured outputs, safety patterns, and evaluation without assuming the reader is a software engineer. The focus is practical: understanding how AI workflows fail, how to design around those failures, and how to build systems that professionals can trust.

Una guida pratica a Claude Code, scritta con Claude Code.Oltre 300 pagine costruite sulla documentazione ufficiale Anthropic e verificate su Claude Code v2.1.311: dall'installazione a Plan Mode, CLAUDE.md, gestione del contesto, permessi, Skill, subagent, hook, MCP e plugin, con esempi reali su WordPress e Node.js invece di demo da slide. In apertura c'è una prefazione scritta da Claude Opus, che ha letto il manoscritto e lo ha commentato con franchezza.Il libro è gratis: pay-what-you-want con minimo zero, decidi tu se e quanto contribuire. La licenza Creative Commons BY-SA 4.0 ti permette di redistribuirlo e modificarlo, citando la fonte e mantenendo la stessa licenza.

Large language models can write code, solve problems and reason about the world, but how do they actually work? The Ghost in the Weights takes you inside the black box, exploring the tools and techniques researchers use to uncover the algorithms hidden in neural networks. Learn to trace circuits, decode features and turn opaque computations into something you can understand, test and investigate yourself.

Discover how artificial intelligence is transforming scientific research, from hypothesis generation to autonomous experimentation. This book explores the models, algorithms and computational tools driving a new era of discovery, offering researchers and engineers the knowledge to understand, build and apply AI-powered scientific systems.

What if you could make massive Mixture-of-Experts models smaller without retraining them from scratch? This book takes you from the fundamentals of expert pruning to practical implementation, covering proven methods, working code and real-world deployment challenges. Learn how to identify redundant experts, preserve model performance and build pruning solutions that work in production.

In Place of a PrefaceSometimes a thought arrives before the language to express it. It is already there: you can sense it, circle around it a few times, argue about it, put it to use - and still not quite understand what is happening. Until a thought has a name, it remains almost invisible. Not because it does not exist, but because we keep looking at it through old words.This work began at precisely that point. Not where a complete theory, a rigorous methodology and a settled system of concepts are already in place. Earlier: at the moment when familiar language began to fail, and it became clear that something important was happening, but the old words were no longer enough to name it.So what you have before you is not a conclusion. Not a treatise. Not even a theory in the strict sense of the word.It is a prelude to a hypothesis.The first recorded trace of a question still searching for its own form.That qualification matters. The first account of a new idea cannot be expected to contain everything that must come later: a fully operational definition, an empirical foundation, a classification of cases, a legal framework, an educational protocol and a definitive methodology. All of that will be needed. Perhaps that is exactly what the next step should be. But the first step works differently: it notices rather than proves; it opens rather than closes; it offers no finished map, but draws a first line where there may once have seemed to be nothing.Of course, we are not starting from scratch, nor do we claim that the question itself is wholly new. Long before today, others...

Вместо предисловияИногда мысль появляется раньше своего языка. Она уже есть: ее можно почувствовать, несколько раз обойти вокруг, спорить о ней, пользоваться ею — и все же не до конца понимать, что именно происходит. Пока у мысли нет имени, она остается почти невидимой не потому, что ее нет, а потому, что мы продолжаем смотреть на нее через старые слова.Эта работа возникла именно в такой точке. Не там, где уже есть законченная теория, строгая методология и окончательная система понятий. А раньше — в моменте, когда привычный язык начал давать сбой и стало ясно: происходит нечто важное, но назвать это прежними словами уже недостаточно.Поэтому перед вами не итог. Не трактат. И даже не теория в строгом смысле слова.Перед вами — пролог к гипотезе.Первый зафиксированный след вопроса, который еще только ищет собственную форму.Эта оговорка важна. От первого текста о новой мысли не- льзя требовать того, что должно появиться на следующих этапах: полной операционализации, эмпирической базы, классификации кейсов, юридической модели, образовательного протокола и окончательной методологии. Все это понадобится. Возможно, именно это и должно стать следующим шагом. Но первый шаг устроен иначе: он не доказывает — он замечает; не закрывает — открывает; не дает окончательную карту — проводит первую линию там, где раньше, возможно, казалось пустое место.

This is a hands-on, code-first book. Every example is real, tested, and runs against a local model with no cloud key required, except where a capability genuinely needs one. You get working Maven projects, not snippets, and a clear-eyed account of each stack's strengths, gaps, and surprises. The full source code is on GitHub under an open-source license What you'll learn:Build a working AI service in three stacks and grade them with one shared testGround answers in your own documents with RAG, and tune what retrieval returnsGive a model memory and tools, and reach external systems over the Model Context ProtocolAdd vision and audio, structured output, production observability, and securityGo from assistants to agents, and see where the JVM's agent story is headingWho it's for: Java developers and architects choosing an AI stack, or already committed to one and curious what the alternatives do differently. Comfort with modern Java and either Spring Boot or Quarkus is all you need. No machine-learning background required.

A generation loop fits in a few lines of Python, yet serving it to many users at once raises hard questions: who owns the cached state, when memory can be reused, and when a guessed token is worth checking. This book answers them through small experiments you can run on your laptop and the source code of DeepInfer, the open-source engine behind DeepSeek's production service.

vLLM and SGLang, side by side. Learn how LLM serving works, measure both engines fairly, tune them with evidence, and take a model from an empty GPU to a deployment with a cost per million tokens you can defend.

Your clients pay you for careful admin work. An AI agent can do the first draft of most of it overnight — if it knows exactly what "done" looks like and exactly what it must never touch.

You tried AI tools. They're fast. And you're still doing all the work — just in a chat window.

From Raw Model Weights to High-Performance C++20 and CUDA ExecutionModern large language models are only as efficient as the systems that execute them. Behind every generated token lies a complex interaction between model architecture, memory bandwidth, numerical precision, parallel computation, kernel execution, and runtime orchestration. Understanding these interactions is essential for engineers who want to move beyond high-level inference frameworks and build systems whose performance, architecture, and execution behavior they can control directly.

For PMs, presales, and managers who work with AI but don’t write code: how RAG, agents, fine-tuning, compliance, and security actually fit together.

The model is one box. This book is about every other box: retries, contracts, evals, budgets, guardrails and agents for backend engineers shipping LLM features to production.