Reasoning

Generating the next token from enough context is one thing. Determining whether an answer actually follows—through evidence, constraints, search, model checking, and verification—is another.

Reasoning: How AI Solves Problems explores that decisive transition.

Through an unfolding story, the book follows an AI system confronting a seemingly simple routing problem. The obvious answer fails a hidden constraint. Another possibility depends on missing information. A third introduces uncertainty and competing objectives. Fluency alone cannot solve the problem. The system must represent it correctly, explore alternatives, test its conclusions, expose its assumptions, and know when the available evidence is not enough.

This journey becomes the book’s central framework:

Represent the problem. Search the possibilities. Verify the conclusion.

Each concept begins with story and intuition, then progresses through visual explanations, practical architectures, algorithms, and the mathematics beneath them. New readers can follow the narrative and mental models. Builders can study the mechanisms and system designs. Advanced readers can examine formal logic, probabilistic reasoning, search, causal inference, model checking, theorem proving, training, and evaluation.

The book also challenges comforting assumptions. A longer explanation is not necessarily a better one. Agreement among several model outputs is not independent evidence. A formal proof can validate an inference without validating its premises. A verifier can relocate an error instead of eliminating it. A confident answer may still be the most polished expression of an unresolved problem.

The recurring Reasoning Ledger makes evidence, assumptions, candidates, checks, and uncertainty visible. Decision Boundaries show where a system must stop, abstain, request information, or hand responsibility to a human owner.

This is Volume III of AI, Bottom Up. Token explains how language becomes computable. Context explains how AI selects what matters. Reasoning explains how that context becomes a defensible conclusion.

But a conclusion—even a well-verified one—does not change the world.

A decision is not an action.

That is where Reasoning ends and the next volume, Agent, begins: with authority, tools, consequences, observation, and recovery.