AI Engineering Building Applications with Foundation Models (Chip Huyen)（Z-Library）

【One-Line Pitch】 A practical, end-to-end guide for software engineers and technical leaders who want to build production-grade AI applications using foundation models, covering everything from model selection and evaluation to prompt engineering, RAG, finetuning, and deployment. 【Book Arc】 - **Opening (~0%–5%)**: Introduces AI engineering as a discipline distinct from traditional ML engineering, framing the model-as-a-service shift that lowers barriers to entry. Sets up the "new AI stack" and the central role of evaluation in mitigating failure risks. - **Early (~5%–20%)**: Builds foundational knowledge of foundation models—how they work, their capabilities and limitations—and establishes a rigorous evaluation methodology, including metrics for open-ended tasks and the emerging AI-as-a-judge approach. - **Middle (~20%–50%)**: Moves into core application techniques: prompt engineering (zero-shot, few-shot, chain-of-thought), retrieval-augmented generation (RAG), and agent design. Emphasizes a progression from simple to sophisticated methods. - **Late (~50%–80%)**: Covers finetuning and dataset engineering, explaining when to finetune versus prompt, how to curate and clean data, and the trade-offs involved. Includes inference optimization for latency and cost. - **Ending (~80%–100%)**: Ties everything together with AI engineering architecture patterns and user feedback loops, discussing how to design systems that learn from real-world usage and iterate safely. 【Key Takeaways】 - **AI engineering is a new discipline** (Early): Unlike traditional ML engineering, it focuses on composing and orchestrating pre-trained models rather than training from scratch. This shifts the core skill from model building to system design and evaluation. - **Evaluation is the linchpin of AI applications** (Early): Open-ended models fail in unpredictable ways, so rigorous evaluation—using benchmarks, human feedback, and automated judges—is essential. The AI-as-a-judge approach is rapidly growing as a scalable alternative to human annotation. - **Prompt engineering is the first lever to pull** (Middle): Techniques like few-shot prompting, chain-of-thought, and role prompting can dramatically improve output quality without any model changes. Start here before considering more expensive interventions. - **RAG and agents extend model capabilities** (Middle): Retrieval-augmented generation grounds responses in external knowledge, while agents enable multi-step reasoning and tool use. These patterns are the building blocks for most practical AI applications. - **Finetuning is a targeted tool, not a default** (Late): Finetuning is most valuable when you need to change model behavior, style, or domain knowledge. It requires careful dataset engineering—curation, cleaning, and balancing—to avoid degrading general performance. - **Dataset engineering is as important as model choice** (Late): The quality and composition of your training and evaluation data often matter more than the model architecture. Garbage in, garbage out applies with full force to foundation models. - **Inference optimization is a cost and latency game** (Late): Techniques like quantization, distillation, and batching can reduce inference costs by orders of magnitude. These decisions directly impact whether your application is economically viable at scale. - **Architecture and feedback loops close the loop** (Ending): Production AI systems need thoughtful architecture for logging, monitoring, and user feedback. This data is the fuel for continuous improvement—both for prompts and for future finetuning rounds. 【Reading Tips】 - **Skim the opening chapters** (~0–10%) if you already have ML experience; the value is in the framing of AI engineering, not the basics. Deep-read the evaluation chapters (3–4) since they are the book's most distinctive contribution. - **Treat Chapters 5–7 (Prompt, RAG/Agents, Finetuning) as a decision tree**: Read them together to understand when to use each technique. The progression from simple to sophisticated is the book's core practical framework. - **Pay special attention to the AI-as-a-judge material** in the evaluation chapters—this is a fast-moving area and the book's treatment is both current and practical. - **Use the dataset engineering and inference optimization chapters as reference material** rather than reading them cover-to-cover. They are dense with specifics you'll want to consult when you hit those problems. - **The epilogue and architecture chapter are worth a careful read** before you deploy anything; they contain the hard-won lessons about user feedback and system design that most tutorials skip. 【Coverage Limits】 The excerpts focus on the book's structure and high-level themes; detailed technical content, code examples, and specific case studies from the chapters are not covered in this guide.