Coming soon · Book Profile
The Book of Why - The New Science of Cause and Effect
A manifesto for the Causal Revolution showing how causal diagrams and the mathematics of counterfactuals let us answer 'why' questions that statistics alone never could.
A profile of this book is on the way.
What it’s about
The Book of Why argues that data, no matter how big, cannot by itself tell us about cause and effect; we need a model of reality. Judea Pearl traces the history of causal inference from Galton and Pearson's blind spot, through Sewall Wright's path diagrams, Bayesian networks, the smoking-cancer debate, and the development of do-calculus, to show that causal questions occupy three rungs of a 'Ladder of Causation': association (seeing), intervention (doing), and counterfactuals (imagining). Using intuitive examples—the Monty Hall problem, Simpson's paradox, confounding, colliders, mediation, and instrumental variables—the book equips readers with the conceptual tools (causal diagrams, the back-door and front-door criteria, the do-operator) to reason rigorously about causation. It is at once a popular science narrative, a defense of human causal intuition, and a roadmap for building machines that genuinely understand why.
The through-line
- Who it’s for
- A scientist, analyst, student, or curious thinker who wants to answer 'why' questions and make trustworthy causal claims from data.
- The problem
- Standard statistical training offers no rigorous language for cause and effect, leaving causal questions unanswerable or answered wrongly. They feel intellectually frustrated and uncertain, fearing they are being fooled by paradoxes, confounding, and spurious correlations.
- The plan
- Climb the Ladder of Causation: learn to distinguish seeing, doing, and imagining.
- Draw a causal diagram that encodes your assumptions about who listens to whom.
- Use the back-door and front-door criteria to identify which variables to adjust for.
- Apply the do-operator, instrumental variables, or mediation analysis to estimate causal effects.
- Reason counterfactually to answer 'what would have happened' questions.
- The payoff
- You confidently distinguish causation from correlation and avoid adjustment errors. · You can estimate causal effects even without a randomized experiment. · You resolve paradoxes by reasoning from the data-generating process.
See our guide
Related profiles we’ve built
- The AI Marketing Canvas: A Five-Step AI Plan for Marketers →
- Artificial Intelligence →
- Artificial Intelligence - A Very Short Introduction →
- Effective Data Science Infrastructure →
- Human + Machine →
- Time Series Forecasting Using Foundation Models →
- Understanding Deep Learning →
- Big Data_ A Very Short Introduction (Very Short Introductions) →
Additional reading
- Designing Machine Learning Systems (DMLS) · Chip Huyen
The author explicitly positions 'AI Engineering' as a companion book to DMLS. DMLS covers the principles of traditional ML engineering (feature engineering, model training, tabular data), which this book builds upon by focusing on the new paradigm of foundation models.
- Artificial Intelligence: A Modern Approach · Stuart Russell and Peter Norvig
The book references this classic AI textbook for its definition of an 'agent' as a perceiving and acting entity, providing the foundational academic context for the 'Agents' section in Chapter 6.
- OpenAI Prompt Engineering Guide · OpenAI
Official documentation from the model provider on how to craft effective prompts, recommended by the author.
- Anthropic's Prompt Engineering Guide · Anthropic
Official guide from another major LLM provider, offering best practices for prompt design.
- dair-ai/Prompt-Engineering-Guide · DAIR.AI Community
A comprehensive, community-maintained GitHub repository covering a wide range of prompt engineering techniques.
- GPT Researcher · assafelovic
An open-source project that inspired the book's research summarization engine, recommended for further exploration of building production-grade LLM applications.
- Building Effective AI Agents · Anthropic
An article recommended by the author for a deeper understanding of the distinction between agentic workflows and autonomous agents.
- Model Context Protocol (MCP) Official Website · Anthropic and Community
The primary source for the MCP specification, tutorials, and ecosystem resources, essential for building or consuming MCP servers.
- FastMCP 2 Documentation · jlowin and community
The official documentation for the recommended Python SDK for building MCP servers, used in the book's example.
- Knowledge Graph Enhanced RAG · Tomaž Bratanič and Oskar Hane
Recommended by the author for readers interested in diving deeper into combining knowledge graphs with RAG.