The AI Pocketbook (Emmanuel Maggiori)（Z-Library）

【One-Line Pitch】 A short, opinionated field guide that helps working engineers understand what today's AI actually does, where it helps, where it fails, and how to keep their careers on solid ground. Best for software, security, and infrastructure engineers who want practical judgment rather than hype. 【Book Arc】 - **Opening (~0%–10%)**: Sets the premise that AI is reshaping engineering work and previews the book's concerns—jargon, hallucinations, tool selection, job impact, and the fine print behind AI claims. - **Early (~10%–35%)**: Opens the black box: how LLMs generate text token by token, why tokens and vocabularies matter, how embeddings and the transformer architecture disambiguate meaning, and how machine learning fills in the model's parameters. - **Middle (~35%–55%)**: Covers ML fundamentals—supervised, self-supervised, unsupervised, and other paradigms—plus loss functions, overfitting, and stochastic gradient descent, clarifying what "generative AI" technically means. - **Late (~55%–85%)**: Moves to practice: why hallucinations happen and whether they can be fixed, how to select and evaluate AI tools without bias, and a checklist for deciding when to use or avoid AI in a task or product. - **Ending (~85%–100%)**: Addresses careers and context—which job characteristics resist AI, how engineers can stay relevant, and the less flattering side of AI such as exaggeration, copyright disputes, and dubious brain comparisons. 【Key Takeaways】 - **LLMs are next-token predictors, not reasoning engines** (Early): They generate text by repeatedly guessing the most likely continuation, which explains both their fluency and their failure modes. - **Tokens shape cost, fairness, and capability** (Early): Non-English text often consumes more tokens, making it more expensive and harder for the model to interpret—a bias that is not easily removed. - **Transformers solve context disambiguation** (Early): Attention mechanisms let a word's meaning shift based on surrounding words, which is why "bark" can mean animal or tree depending on context. - **Machine learning is a different programming paradigm** (Middle): Instead of writing explicit rules, engineers define architecture and let training data and loss minimization fill in the parameters. - **Overfitting is a core risk** (Middle): A model can score well on training data yet fail on new data, so low loss alone does not guarantee real-world performance. - **Hallucinations are inherent, not just bugs** (Late): The book treats them as a structural consequence of how LLMs work, and offers practical responses rather than promises of a quick fix. - **Tool selection needs disciplined evaluation** (Late): A structured method helps compare AI tools and avoid common biases when judging them. - **Not every task should use AI** (Late): A checklist helps decide whether AI belongs in a given task or customer-facing product—and when it will fail customers. - **Job resilience has identifiable traits** (Ending): Certain characteristics make roles harder to automate, and engineers can deliberately build relevance in the AI era. 【Reading Tips】 - **Deep-read the early technical chapters** if you want to understand why LLMs behave as they do; the token, embedding, and transformer explanations are the book's foundation. - **Skim the ML fundamentals** if you already know supervised learning and gradient descent—focus instead on how the book connects them to LLM training. - **Treat the late chapters as checklists**: the tool-evaluation method and the "when to use AI" checklist are the most directly actionable parts for daily work. - **Read the final chapter for context, not comfort**: it covers exaggeration, copyright, and brain comparisons that help you calibrate claims you'll hear at work. - **Keep the pocketbook as a reference**: its short format suits revisiting specific sections before adopting a tool or defending a technical decision. 【Coverage Limits】 This guide is based on stratified excerpts covering the book's structure, early technical foundations, and chapter summaries; specific examples, figures, and detailed arguments from later chapters are only partially represented. The excerpts do not cover the full text of the hallucination, tool-selection, and career chapters in depth.