Preface — Teaching Machines to Reason

📍 Where we are: Right at the door of the first of five volumes. Before we prove a single theorem or train a single model, let us agree on what neuro-symbolic AI actually is — and why neither logic nor learning, on its own, is enough.

Welcome. This is the first of five books about one of the oldest ambitions in computer science: building a machine that can reason — that can take what it knows and work out what must follow. Two great traditions have chased that goal for seventy years, and each captured half of it. Symbolic logic reasons exactly: it proves conclusions step by step and shows its work — but it is brittle, and falls silent the instant a fact is missing or a name is misspelled. Neural learning reasons robustly: it generalizes from examples and always ventures an answer — but it is opaque, and can be fluently, confidently wrong. Neuro-symbolic AI is the project of joining the two halves so that one system reasons like the first and generalizes like the second, at the same time. That project is what these five books are about, and this one lays its foundation.

You need no background to start. If you have never written a proof, never trained a model, never heard the words Horn rule or gradient descent, you are exactly the reader we wrote for. We define every specialized term the first time it appears. We begin, in the very next chapter, with the plainest idea in mathematics — a collection of things — and build every later idea on top of it.

The simple version

Imagine you want a machine to answer questions about a small university — who advises whom, which paper cites which, who works where. You could hire a lawyer, who answers only when she can walk you through an airtight chain of evidence, and who says nothing at all when a document is missing. Or you could hire a detective, who has seen thousands of cases and always offers a hunch — even for a question he has never heard — but who can be confidently wrong. The lawyer is symbolic logic; the detective is neural learning. Each is brilliant exactly where the other is helpless. Neuro-symbolic AI is the attempt to put them in the same room and get the lawyer's certainty and the detective's reach at once. This book builds both experts from scratch, on one tiny world, so you can watch precisely what each one does — and see why joining them is so hard.

What this chapter covers

A quick map of this preface: who these books are for and the promise they make — a single path from freshman to the research frontier across five volumes, with this one as the foundation; the honest thesis, that neuro-symbolic AI joins two complementary halves and that neither half suffices alone; the one small "academic world" — five people, three papers, two institutions, a handful of topics, and the relations that link them — threaded through every volume and re-read five ways, with everything runnable in a pure-Python companion; the five volumes in a paragraph and the five parts of this volume; and the house conventions — the plain-English analogy that opens each chapter, a term defined on first use, a running honesty section, and code you can actually run — that carry through every page.

Who these books are for

These books are for a curious beginner who wants to understand neuro-symbolic AI properly, not just admire it from a distance. Volume 1 assumes only high-school mathematics and a willingness to think slowly; by the final volume, the same reader is standing at the research frontier, reading papers and weighing claims. Nothing is imported as a black box — when we need logic, we build logic; when we need a neural network, we build one small enough to trace by hand. This is the foundation volume, and its job is to make sure everything the later volumes stack on top of it rests on ground you laid yourself.

The thesis: two halves of one problem

Here is the conviction the whole series turns on, stated plainly. Reasoning has two halves, and we have known how to build each one separately for decades. The symbolic half is exact but brittle: a proof is certain and auditable, yet it collapses without warning the moment the world contains a gap, a typo, or a fact nobody wrote down. The neural half is robust but opaque: a learned model degrades gracefully on messy, incomplete data and always returns something, yet it cannot tell you why, and it will hand back a confident answer that is simply false. Crucially, these two failure modes are not merely different — they are mirror images: where logic goes silent, learning still guesses; where learning hallucinates, logic can refuse the impossible. That is both the reason to combine them and the difficulty, because the obvious ways of wiring a prover to a learner tend to destroy the very guarantee, or the very robustness, you were trying to keep. Naming that tension precisely — two cultures with complementary strengths — is what turns "here are two techniques" into "here is a reason to unite them."

The running example: one small academic world

Rather than scatter a new toy problem into every chapter, this series learns and reasons over one shared knowledge base, small enough to hold in your head and rich enough to carry every idea in the series. It is a tiny academic world: five people — alice, bob, carol, dave, and erin — three papers p1, p2, and p3, two institutions mit and cmu, and a few research topics, linked by five relations — advises, authored, cites, affiliated, and about. From just 23 stated facts, a handful of rules grind out a settled world of 47 facts — twenty-four of them derived by the machine, including that alice grand-advises carol along the advising chain alice → bob → carol → erin, and that p3 reaches p1 along the citation chain p3 → p2 → p1. Every idea in this book lands on this exact world, and none of it is hand-waving: the world lives in a pure-Python companion, examples/logic, where kb.py states the facts, a forward-chaining engine derives the rest, a backward-chaining prover answers questions with a proof you can read, and a from-scratch embedding places the same symbols on a map.

One knowledge base, five lenses

That single world is the spine of the whole series, because each of the five volumes re-reads it a different way while changing nothing about the world itself. Volume 1, Foundations (this book), reads it as logic and learning built from scratch — facts forward-chained into a model, goals proved by hand, symbols placed as points. Volume 2, Symbolic Reasoning, reads the same facts as a formal ontology and hands them to a reasoner that classifies the world. Volume 3, Neural Representation, learns the entities and relations as geometry at scale, so that a hierarchy becomes literal containment. Volume 4, Neuro-Symbolic Integration, answers multi-hop questions — like who does alice grand-advise — by geometry when a proof is out of reach. Volume 5, The Reasoning Frontier, adds time and uncertainty, asks whether stated confidences can be trusted, and measures the joint system the series builds toward. Because the world is pinned down to the last fact, any difference you see between the volumes has to be a difference in method, not in data — which is exactly what makes the series' claims comparable rather than anecdotal.

Two halves of one problem, joined over a single shared world: exact-but-brittle symbolic logic and robust-but-opaque neural learning meet over the academic-world knowledge base, while the five-volume ladder climbs from this foundation toward the reasoning frontier. Original diagram by the authors, created with AI assistance.

The five parts of this volume

Volume 1 is one continuous climb from the plainest mathematics to the neuro-symbolic idea, in five parts:

1. Part I — Logic from Scratch. Sets, relations, and functions; order and lattices; then propositional and first-order logic — the vocabulary for saying what is true, laid one stone at a time.
2. Part II — Reasoning as Computation. Turning truth into mechanism: forward chaining that grinds out every consequence, backward-chaining proof that chases a single goal down to the facts, and the fixpoint that ties both together.
3. Part III — Learning from Scratch. The other half: what learning even is, the neural network that bends to fit patterns no straight line can match, and the embedding that turns symbols like alice and advises into points a network can compute with.
4. Part IV — The Neuro-Symbolic Idea. The two halves face to face: two cultures with mirror-image strengths, the field's six named ways to combine them, and the running example on which every method is measured.
5. Part V — The Verdict. The honest reckoning: what logic and learning each genuinely deliver, what their fusion buys today, and where — on exactly this shared world — the real frontier still lies.

How to read this book

A few conventions repeat on every page, so it is worth meeting them once.

• The simple version. Every chapter opens with a :::tip[The simple version] box that states the whole idea in an everyday analogy — a lawyer and a detective, a contact list, a family tree — before any formalism arrives. Read only those boxes and you still get the story.
• Every term defined on first use. No word is assumed: the first time a term appears it is defined in plain language on the spot, and a Key terms box at the end of each chapter gathers them so you never have to scroll back.
• The unsolved part. Every chapter carries a short honesty section that names what is not solved — the gap, the open problem, the place the field is still stuck. This book is more useful for admitting those early than for discovering them late.
• Code you can run. Every code block is a true excerpt of a real file in the examples/logic companion, never pseudocode dressed up to look real; where a snippet is illustrative rather than runnable, it says so.

Where this leads

We have the whole shape now: two halves that need each other, one small world to test them on, five volumes and five parts to climb. The foundation begins where mathematics itself does — with the idea of a collection of things. The first chapter, Sets, Relations, and Functions, lays the ground floor: what a set is, how a relation is nothing more than a set of pairs, and how composing advises with itself gives grand-advising — the first time our academic world tells us something nobody wrote down. Everything else in the series is built on top of that, one honest stone at a time.