AI Product Managers Handbook, Second Edition (Early Access) (Irene Bratsis)（Z-Library）

【One-Line Pitch】 A practical roadmap for product managers who must build, ship, and evolve AI-driven products without becoming machine-learning engineers. Best for PMs moving from traditional software into AI/ML, and for technical leads who need a shared vocabulary with product. 【Book Arc】 - **Opening (~0%–10%)**: Sets the frame — what AI product management is, why it differs from traditional PM work, and the "optimal flow" of an AI process from data through training to deployment. - **Early (~10%–35%)**: Productizing AI outputs. Contrasts AI and traditional software on scalability, margins, uncertainty, B2B vs. B2C, domain knowledge, experimentation, and MLOps/AIOps reliability; introduces vertical customization (FinTech, healthcare, marketing, manufacturing, education, cybersecurity) and value metrics like OKRs and KPIs. - **Middle (~35%–65%)**: The technical foundation. Infrastructure and tooling, databases/warehouses/lakes/lakehouses, deployment strategies, definitions of AI vs. non-AI, the four learning types (supervised, unsupervised, semi-supervised, reinforcement), and concrete algorithms from Naive Bayes and SVM to linear/logistic regression and k-means. - **Late (~65%–80%)**: Deep learning, LLMs, GANs, and generative AI — including the black-box explainability problem, autonomous feature learning, and why frontier models are so compute- and data-expensive. - **Ending (~80%–100%)**: Product design for AI-native products, benchmarking performance/cost/growth, managing and evolving AI products, industry trends, and career development for the AI PM (from breaking in to maturing as an "evergreen learner"). 【Key Takeaways】 - **AI PM work is a translation role** (Early): You sit between what engineering builds and what marketing communicates, so you must know the difference between ML, DL, and the umbrella term "AI" at the model level — not just the buzzword level. - **AI products differ from traditional software in specific, manageable ways** (Early): Scalability, profit margins, uncertainty, domain knowledge, experimentation, MLOps reliance, performance evaluation, and feedback loops are the axes where planning must change. - **All current deployed AI is narrow (ANI), not AGI** (Middle): The book is blunt that sentient AI does not exist, but untested, biased, or poorly managed ANI applications already cause real harm — ethics and safety are PM responsibilities, not afterthoughts. - **The four learning types organize every ML model** (Middle): Supervised, unsupervised, semi-supervised, and reinforcement learning differ mainly in labeling and reward structure; knowing which camp a model belongs to tells you what data and evaluation it needs. - **Simple algorithms still dominate production** (Middle): Naive Bayes, SVM, linear and logistic regression, and k-means remain workhorses because they are versatile and dependable — deep learning is not automatically the right answer. - **Deep learning's opacity is a product risk** (Late): Neural networks do their own feature selection and resist explanation, so the importance of explainability varies by product context and must be weighed deliberately. - **Generative AI is advanced deep learning in two families** (Late): GANs for image generation and LLMs trained on terabytes of data; their prompt-based accessibility is what made them reach non-technical users, but their cost and complexity are formidable even for tech giants. - **Vertical context changes the product** (Early): FinTech, healthcare, marketing, manufacturing, education, and cybersecurity each impose different domain knowledge, value metrics, and customization needs. 【Reading Tips】 - Deep-read the early chapters on how AI products differ from traditional software — that contrast is the book's core decision framework and the part most PMs will use weekly. - Skim the algorithm catalog (Naive Bayes through k-means) for vocabulary and intuition rather than implementation detail; you need to converse with engineers, not replace them. - Treat the explainability and ethics discussion in the DL/generative AI section as a checklist for your own product's risk review. - Use the verticals chapter as a reference when your product targets a specific industry; read your own vertical closely and skim the rest. - The career chapters at the end are best read after the technical material, once you can articulate what kind of AI PM you want to be. 【Coverage Limits】 The excerpts are heavily front- and middle-weighted: they cover framing, productization, learning types, algorithms, and generative AI well, but the later chapters on design, benchmarking, managing evolving products, and career development are only visible through the table of contents and brief mentions. Specific case-study details and chapter-level arguments in the final third are not covered by these excerpts.