The Visual Encyclopaedia · October 2026

The Visual Encyclopaedia
of AI

100 chapters. 10 topics. Everything that matters.
by Mat Siems
Part I

Foundations

What machine intelligence actually means.

Chapter 1 · Part I

What Is Intelligence?

Ask ten experts to define intelligence and you will receive eleven answers. There is still no single agreed definition. Some stress the ability to learn, others the capacity to reason, plan, remember or perceive. A crow bending a wire into a hook, a child learning a second language and a chess grandmaster calculating twelve moves ahead all look intelligent, yet they share remarkably little machinery on the surface.

The working definition that most AI laboratories reach for is pragmatic: intelligence is the ability to achieve goals across a wide range of environments. The phrase owes much to the researchers Shane Legg and Marcus Hutter, who in 2007 surveyed dozens of definitions and distilled their common core. It has a great virtue for engineers. It describes what a system can do, not what it is made of, and so it can be measured, compared and, in principle, built.

That shift is the founding move of artificial intelligence. For two thousand years the nature of mind was a puzzle for philosophers. AI recast it as an engineering programme: if intelligence is a set of capabilities, then each capability can be specified, tested and improved, whether the substrate is neurons or silicon. The question changed from "what is a mind?" to "what would a machine need in order to behave like one?"

The answer, broadly, is a bundle of abilities working together. Learning lets a system improve from experience. Reasoning lets it draw conclusions from what it already knows. Memory stores facts and skills, perception turns raw signals such as light and sound into meaningful objects, and planning strings actions together towards a goal. Human intelligence blends all five so seamlessly that it is easy to miss how distinct they are. Much of AI's history is the story of building them one at a time, and then discovering how hard it is to join them up.

Did you know? Psychologists and researchers have catalogued more than 70 distinct definitions of intelligence, and Legg and Hutter's 2007 survey collected around that many before attempting to unify them.

Crucially, AI targets capability, not consciousness. Whether a machine that answers questions brilliantly has any inner experience is a deep and open question, but it is a separate one. A thermostat pursues a goal without feeling warm; a navigation app finds the fastest route without caring whether it arrives. The engineer's question is narrower and more tractable: does the system reliably get the job done, across situations it has never seen before? That last clause, generality, turns out to be the hardest part of all.

Intelligence, in the machine age, is judged by what it can achieve.

The bundle of abilities we call intelligence Intelligence achieving goals in many environments Learning improves from experience Reasoning draws conclusions Memory stores facts and skills Perception light and sound to objects Planning chains actions to a goal Generality works in unseen situations Judged by what it achieves, not by what it is made of.
Fig 1 · What Is Intelligence?. The question that started everything.
Chapter 2 · Part I

The Turing Test

In 1950 the British mathematician Alan Turing published a paper in the philosophy journal Mind titled "Computing Machinery and Intelligence". It opened with a question, "Can machines think?", and then promptly set it aside as too vague to answer. In its place he proposed a game he called the imitation game, known ever since as the Turing Test.

The set-up is simple. A human judge holds conversations with two hidden partners, one a person and one a machine. Judge, human and machine communicate by text only, so voice, face and handwriting give nothing away. The judge may ask anything: arithmetic, poetry, gossip, the taste of strawberries. If, after a fair stretch of questioning, the judge cannot reliably tell which partner is the machine, the machine is said to pass.

The genius of the idea lay in what it left out. Turing did not ask what thinking feels like, or whether a machine has a soul, or whether its insides resemble a brain. He framed intelligence as behaviour, not biology: if the conversation is indistinguishable, on what grounds would anyone deny the machine the label they grant to other people? He anticipated the objections, from the theological to the mathematical, and answered them one by one in the same paper.

The test has attracted criticism ever since. In 1980 the philosopher John Searle offered his "Chinese Room" thought experiment, arguing that shuffling symbols according to rules could produce fluent answers without any understanding at all. Others pointed out that the test rewards deception, and that a machine might fail simply by being too quick at sums. Early programs exploited human generosity: ELIZA, written by Joseph Weizenbaum at MIT in the mid-1960s, mimicked a psychotherapist by reflecting users' words back at them, and some users confided in it anyway.

Did you know? Turing predicted that by around the year 2000 machines would play the imitation game so well that an average judge would have no better than a 70 per cent chance of spotting them after five minutes. He was roughly two decades out.

That vindication arrived with large language models. Modern chatbots routinely fool judges in short sessions, and in 2025 researchers at the University of California, San Diego reported that GPT-4.5, prompted to adopt a human persona, was picked as the human more often than the real people it was competing against. Passing, it turned out, said as much about human judges as about machines, and few researchers now treat the test as a finish line. Its lasting gift is the stance it introduced: judge a machine by what it does.

Turing turned a metaphysical riddle into an experiment anyone could run.

How the imitation game is played Hidden human Judge Machine Any question, by text arithmetic, poetry, gossip Typed reply no voice, face or handwriting Same kind of question the taste of strawberries Typed reply a real person, hidden Can't tell them apart: pass 2025: GPT-4.5 judged human Text only: the judge sees behaviour, never biology.
Fig 2 · The Turing Test. Can a machine fool a human?.
Chapter 3 · Part I

Narrow vs General AI

Every AI system deployed in the world today is, in an important sense, a specialist. Engineers call this narrow AI: software that can be superhuman at one task and useless outside it. A chess engine cannot drive a car. A system that spots tumours in mammograms cannot translate a menu, and a translation system cannot fold a towel. Each is a brilliant savant with a single trick.

Within their lanes these specialists are formidable. Narrow AI already exceeds human performance in dozens of domains: board games such as chess and Go, many video games, protein-structure prediction, speech transcription in good conditions and certain image-classification benchmarks. Often the margin is enormous. No human has beaten a top chess engine in a serious match for many years, and the best players now study engine analysis to improve their own play.

The dream that has animated the field since its founding is something else: artificial general intelligence, or AGI, a single system matching humans across the board. Such a system could learn a new job from a manual, switch from tax law to plumbing to poetry, and transfer what it learned in one area to another, the way a person who has learned to drive a car can quickly pick up a van. Generality, rather than raw skill, is the defining feature.

The line between narrow and general has blurred in recent years. Large language models write code, summarise contracts, pass professional exams and hold conversations on almost any subject, all from one set of trained weights. Some researchers argue they are early, uneven steps towards generality; others point out that they still stumble on tasks a child finds easy, lack reliable memory across sessions and cannot act competently in the physical world. Both camps agree that AGI is a goal, not a shipped product, and definitions of when it would count as achieved remain disputed.

Did you know? IBM's Deep Blue beat the world chess champion Garry Kasparov in 1997, yet it could not have played a game of noughts and crosses: everything it knew was chess.

The distinction matters beyond the laboratory. Narrow systems fail in narrow, predictable ways and can be tested against their task. A general system would be far harder to evaluate, because the space of things it might do, well or badly, has no obvious edge. That is why discussions of AI safety and governance so often hinge on how general a system really is, rather than on how impressive any single demonstration looks.

The specialists are already here; the generalist remains a horizon.

Specialists, today's models and the AGI dream Narrow AI One task only Superhuman at it Chess, Go, proteins Useless outside it Easy to test LLMs today One set of weights Code, exams, chat Trip on easy tasks Weak memory No physical skill AGI Any task Human breadth Transfers skills Hard to evaluate Still a goal Large language models blur a line that once looked sharp.
Fig 3 · Narrow vs General AI. Specialists and the dream of a generalist.
Chapter 4 · Part I

Symbolic AI

The first great paradigm of artificial intelligence treated thinking as the manipulation of symbols. Its founders, among them Allen Newell, Herbert Simon and John McCarthy, believed that a mind was essentially a machine for applying logical rules to statements about the world. Capture enough knowledge as symbols, add an engine for drawing conclusions, and intelligence should follow. The approach is known as symbolic AI, or, affectionately, "good old-fashioned AI".

A symbolic system has two main parts. A knowledge base holds facts and rules written by human experts, in a form such as IF fever AND rash THEN suspect measles. An inference engine chains these rules together, matching facts against conditions and adding new conclusions until it reaches an answer. Because every step is explicit, such a system can explain its reasoning, which remains one of the paradigm's lasting attractions.

Symbolic AI was the dominant paradigm from the 1950s to the 1980s. Early triumphs included the Logic Theorist of 1956, which proved theorems from Principia Mathematica, and programs that solved algebra problems and stacked virtual blocks. Its commercial peak came with expert systems in the late 1970s and 1980s. MYCIN, developed at Stanford, used several hundred rules to recommend antibiotics for blood infections and performed respectably against specialists in trials. Corporations rushed to bottle the know-how of their best employees in software.

The weakness was brittleness. Rules fail outside anticipated cases, and the real world is mostly unanticipated cases. A medical system that knew nothing about a patient's pregnancy, or a typing error in a test result, would cheerfully reach absurd conclusions. Writing the rules was slow and expensive, a problem dubbed the knowledge acquisition bottleneck, and the rule bases became tangled as they grew. Common sense, the vast stock of unspoken facts that people take for granted, proved almost impossible to write down; the Cyc project, begun by Douglas Lenat in 1984, spent decades trying.

Did you know? XCON, an expert system built in the early 1980s to configure orders for Digital Equipment Corporation's VAX computers, was estimated to save DEC around $40 million a year by the mid-1980s.

When the expert-systems market collapsed in the late 1980s, symbolic AI lost its crown, but not its relevance. Its ideas live on in search algorithms, planning software, database query languages and the formal verification of chips and code. Researchers are now actively exploring "neurosymbolic" hybrids that pair learned systems with explicit logic, hoping to combine the flexibility of one with the reliability of the other.

Symbolic AI showed that rules can reason, and that the world refuses to fit them.

How an expert system reaches a verdict 01 Human expert writes the rules by hand 02 Knowledge base IF fever AND rash THEN measles 03 Inference engine matches facts to conditions 04 Conclusion with a full chain of reasons new conclusions become facts Every step is explicit, so the system can explain itself.
Fig 4 · Symbolic AI. Intelligence as logic and rules.
Chapter 5 · Part I

The Learning Turn

Traditional programming is a recipe: a human writes the rules, the computer applies them to data, and answers come out. Machine learning inverts the recipe. Instead of writing the rules, the programmer supplies examples, data paired with the right answers, and lets the machine find the rules itself. Rules are discovered, not written.

Take the problem of recognising a handwritten digit. Describing in code what makes a "7" a seven, across every slant, scrawl and serif, is maddeningly hard. Showing a learning system tens of thousands of labelled sevens, and letting it adjust its internal settings until it sorts them correctly, turns out to be far easier. The famous MNIST collection of 70,000 handwritten digits, assembled in the 1990s, became a standard training ground for exactly this approach.

The idea is old. In 1959 the IBM engineer Arthur Samuel published a study of a checkers-playing program that improved by playing thousands of games against itself, and in doing so he popularised the very phrase "machine learning". The central claim was that learning systems improve with experience: the more games, examples or feedback they receive, the better they become, without anyone rewriting their code.

For decades learning methods coexisted with symbolic AI, often as a junior partner. The balance shifted in the 1990s and 2000s as computers grew faster and digital data piled up. Statistical methods began beating hand-built rules at speech recognition, spam filtering, machine translation and web search. Then, from 2012, deep learning with large neural networks swept through vision and language, and the learning approach became the default for almost every hard problem in AI.

Did you know? Arthur Samuel was not a strong checkers player himself, and his self-improving program learned to beat its own creator; in 1962 it won a widely publicised game against a capable human player.

The learning turn changed what counts as the valuable ingredient. In the symbolic era, the precious resource was expert knowledge, painstakingly typed in. In the learning era, data replaced hand-crafted knowledge as fuel, and data quality now matters more than clever code. A model trained on mislabelled, biased or unrepresentative examples will faithfully learn the mistakes. It also changed what engineers do all day: much of modern AI work consists of gathering, cleaning and labelling data, designing objectives, and measuring whether the learned behaviour really generalises beyond the examples.

Machines stopped being told the answers and started working them out.

From self-taught checkers to deep learning 1959 Arthur Samuel names machine learning 1962 Checkers win beats a capable player 1990s MNIST 70,000 digits 2000s Statistics wins speech, spam, search 2012 Deep learning the new default Rules are discovered from data, not written by hand.
Fig 5 · The Learning Turn. From programming answers to learning them.
Chapter 6 · Part I

Search & Optimisation

Beneath a surprising amount of artificial intelligence sits a single idea: search. A problem is framed as a space of possible answers, and the system explores that space looking for the best one. Chess moves, delivery routes, timetables, protein shapes and the weights inside a neural network are all found in essentially the same way, by trying candidates and keeping the promising ones.

The difficulty is size. The mathematician Claude Shannon estimated in 1950 that chess has roughly 10^120 possible games, a figure now called the Shannon number. No computer could ever examine them all. Even a modest problem, such as visiting 30 cities in the best order, has more possible routes than could be checked in the lifetime of the universe. Brute force alone is hopeless; intelligence lies in deciding where not to look.

That is the job of heuristics: rules of thumb that prune impossible search spaces down to manageable ones. A chess program does not consider sacrificing its queen on move one in every variation; it estimates which positions look good and spends its effort there. The A* algorithm, published in 1968 by Peter Hart, Nils Nilsson and Bertram Raphael, finds shortest paths by combining the distance travelled so far with an estimate of the distance remaining, and its descendants still power route-planning software and video-game characters. Game-playing programs use minimax search, assuming the opponent will reply with their best move, and alpha-beta pruning to skip branches that cannot change the result.

Optimisation is search with a score. Instead of looking for a goal state, the system seeks the option that maximises or minimises some quantity: profit, distance, error. The workhorse of modern machine learning, gradient descent, is search over weights. Picture a walker in fog on a hilly landscape whose height represents the model's error; at each step the walker feels the slope underfoot and moves downhill. Repeat billions of times across billions of dimensions and a neural network gradually settles into a configuration that makes few mistakes.

Did you know? There are more possible chess games than atoms in the observable universe: Shannon's 10^120 dwarfs the roughly 10^80 atoms that physicists estimate the cosmos contains.

The two great methods often work together. DeepMind's AlphaGo, which beat Lee Sedol at Go in 2016, combined Monte Carlo tree search with neural networks trained by gradient descent to judge which moves and positions were promising. Search provided the foresight; learning provided the intuition about where to look.

Much of what looks like cleverness is a well-pruned search.

The family of search methods in AI Search explore a space of answers Heuristic search rules of thumb prune A* (1968) Route planners Game characters Game-tree search chess has ~10^120 games Minimax Alpha-beta pruning Monte Carlo (AlphaGo) Optimisation search with a score Gradient descent Downhill in fog Network weights Intelligence lies in deciding where not to look.
Fig 6 · Search & Optimisation. Intelligence as finding the best option.
Chapter 7 · Part I

Knowledge Representation

Knowing something is not enough; a machine must store it in a form it can use. Knowledge representation is the study of how. The choice is far from neutral: the representation determines what reasoning is possible, much as a map's projection determines which distances are easy to measure. AI has tried many formats over seventy years, and each one makes some questions easy and others nearly impossible.

The oldest approach is logic: facts and rules written as precise statements that an inference engine can combine. "All birds have wings; Tweety is a bird; therefore Tweety has wings." Logic is exact and explainable, but awkward with exceptions (penguins, ostriches, Tweety's unfortunate accident) and with degrees of belief.

Graphs came next as a dominant format for linked facts. A knowledge graph links entities by relations: "Marie Curie" connects to "physicist" through occupation, to "Warsaw" through born in, and to "Nobel Prize" through award received. Google launched its Knowledge Graph in 2012 under the slogan "things, not strings", and its summary panels beside search results draw on it. Wikidata, the open database behind Wikipedia, holds over a hundred million items arranged the same way. Probabilistic representations, meanwhile, attach likelihoods to facts and links, so a system can say a diagnosis is 80 per cent likely rather than simply true or false.

The modern revolution replaced explicit symbols with numbers. Embeddings encode meaning as vectors: each word, image or document becomes a list of hundreds or thousands of numbers, positioned so that similar things sit close together. In 2013 the word2vec system showed that arithmetic on such vectors could capture relationships, so that "king" minus "man" plus "woman" landed near "queen". Large language models go further still: they store knowledge implicitly in their weights, billions of numerical parameters adjusted during training. Nowhere inside a model is there a line reading "Paris is the capital of France", yet ask it and the answer emerges.

Did you know? By 2020 Google reported that its Knowledge Graph held over 500 billion facts about 5 billion entities, built up in just eight years after launch.

Each format trades something away. Logic and graphs are transparent and editable but rigid; weights are flexible and astonishingly broad but opaque, which is why a model can state a falsehood with confidence and why researchers struggle to locate or correct a specific fact inside one. Systems that retrieve facts from a database or graph and hand them to a language model try to combine the strengths of both.

How a machine stores knowledge quietly decides how it can think.

What each way of storing knowledge trades Logic and graphs exact, explainable, brittle with exceptions Retrieval + LLM facts from a graph, fluency from a model Worst of both rigid and unreadable Embeddings, weights king − man + woman ≈ queen; facts hidden rigid Breadth and flexibility flexible ↑ Transparency opaque → editable How a machine stores knowledge decides how it can think.
Fig 7 · Knowledge Representation. How machines store what they know.
Chapter 8 · Part I

Probability & Uncertainty

The real world is noisy. Sensors misread, witnesses misremember, symptoms overlap and words carry several meanings. A system that demands certainty before acting will never act at all. Modern AI is therefore probabilistic: it weighs evidence and outputs likelihoods, not certainties, and treats belief as something that comes in degrees.

The grammar of that reasoning is Bayes' theorem, named after the English minister and mathematician Thomas Bayes. It describes how to update a belief in the light of new evidence. Start with a prior, how likely something seemed before; observe evidence; ask how probable that evidence would be if the belief were true or false; and combine the two into a posterior, the revised belief. The posterior then becomes the prior for the next piece of evidence, and the cycle continues.

A medical test shows why this matters. Suppose a disease affects one person in a thousand, and a test catches 99 per cent of cases but also flags 1 per cent of healthy people. A positive result feels damning, yet Bayes' rule shows that most positives come from the far larger healthy population, so the chance of actually having the disease is only around 9 per cent. Intuition, which tends to ignore the prior, gets this wrong; the arithmetic does not.

Spam filters were early Bayesian successes. In 2002 the programmer Paul Graham described a filter that learned, from a user's own mail, how often words such as "viagra" or "unsubscribe" appeared in junk versus genuine messages, then combined those clues into an overall probability that a new email was spam. In the 1980s Judea Pearl developed Bayesian networks, diagrams of linked causes and effects that let machines reason about uncertainty at scale; the work earned him the Turing Award in 2011. Self-driving cars use related methods to fuse imperfect readings from cameras, radar and lidar into one best estimate of where everything is.

Did you know? Bayes' rule was published in 1763, two years after Bayes's death, when his friend Richard Price presented it to the Royal Society: some 180 years before the first electronic computers.

Probability runs right through today's language models. Every LLM output is a probability distribution: at each step the model assigns a likelihood to every possible next word fragment, then picks one. A setting called temperature controls how adventurous that choice is. The confident tone of a chatbot can mask the fact that, underneath, it is always placing bets.

Machine belief is never a yes or no; it is a number that keeps moving.

Bayes' rule as a loop of belief Bayes' theorem Prior belief before evidence 1 in 1,000 has the disease Evidence a new observation test comes back positive Likelihood how probable if true? 99% caught · 1% false alarm Posterior revised belief only about 9% are ill 1 in 1,000 ill, test 99% accurate: a positive means only about 9%.
Fig 8 · Probability & Uncertainty. Reasoning when nothing is certain.
Chapter 9 · Part I

Compute, Data, Algorithms

Every AI capability is the product of three inputs: compute, the raw processing power used for training and running a model; data, the examples it learns from; and algorithms, the ideas that turn the first two into skill. Progress in any one multiplies the value of the others. A clever algorithm starved of data learns little; mountains of data are useless without the compute to digest them.

Of the three, compute has grown fastest. Analysts at the research group Epoch AI estimate that the computation used to train the largest models grew roughly four- to five-fold per year through the late 2010s and 2020s, equivalent to doubling about every six months. Much of that came from graphics processing units, chips originally designed for video games that happen to excel at the vast matrix multiplications at the heart of neural networks. The 2012 network AlexNet, which shocked the field by winning the ImageNet image-recognition contest, was trained on two consumer graphics cards; frontier models are now trained on tens of thousands of specialised chips in purpose-built data centres.

Data supplied the second surge. ImageNet, assembled by Fei-Fei Li's team and released in 2009, offered over 14 million labelled photographs. Internet-scale text made large language models possible: books, websites, code and encyclopaedias, filtered and deduplicated, amounting to trillions of tokens, the word fragments models read. Meta reported training its Llama 3 models in 2024 on about 15 trillion tokens. Labs increasingly worry that high-quality human-written text is finite, and are turning to synthetic data, licensed archives and new modalities such as video.

The third fuel arrives in bursts. Backpropagation, popularised in 1986, made it practical to train multi-layer networks. The transformer, introduced by Google researchers in the 2017 paper "Attention Is All You Need", was the key algorithmic unlock for language: it processes whole sequences in parallel, which suits modern chips, and scales gracefully as models grow. Studies suggest algorithmic improvements have also steadily cut the compute needed to reach a given level of performance.

Did you know? Training compute for the largest frontier models grew by a factor of over 100 million between 2012 and 2024, according to estimates compiled by Epoch AI.

These three fuels explain both the speed of recent progress and the shape of the industry. Compute is expensive, so it concentrates power among organisations that can afford vast data centres and the electricity to run them. Data raises questions of copyright and consent. Algorithms, at least, still travel freely through published papers.

Capability is compute times data times ideas, and all three keep compounding.

Three fuels and the milestones they enabled Capability compute × data × ideas Compute 4–5× more each year Data trillions of tokens Algorithms ideas that turn both to skill AlexNet, 2012 two gaming GPUs Llama 3, 2024 about 15T tokens Transformer, 2017 parallel, scales well Progress in one fuel multiplies the value of the others.
Fig 9 · Compute, Data, Algorithms. The three fuels of AI.
Chapter 10 · Part I

The Bitter Lesson

In March 2019 the computer scientist Rich Sutton, a pioneer of reinforcement learning, posted a short essay on his personal website called "The Bitter Lesson". It distilled seventy years of AI research into one uncomfortable claim: general methods that leverage massive computation always win, in the long run, over methods built on human-crafted knowledge.

Sutton's evidence came from the field's own history. In chess, researchers spent decades encoding grandmaster wisdom about pawn structures and king safety, only to be overtaken by Deep Blue's massive brute-force search in 1997. In Go, programs built on human insight stalled at amateur level until AlphaGo combined search with learning from self-play. In speech recognition, statistical methods based on hidden Markov models beat systems built on linguistic rules from the 1970s onwards, and were later swept aside by deep learning. Computer vision followed the same arc, as learned networks replaced hand-designed edge detectors and feature extractors.

The lesson is bitter because it stings. Hand-built knowledge repeatedly plateaus. It delivers satisfying early gains, flatters its designers' understanding of the problem, and then hits a ceiling, while the cruder, more general approach keeps improving as computers get faster. The driving force, Sutton noted, is the relentless fall in the cost of computation, sometimes summarised as Moore's law. A method that cannot absorb more compute is betting against the most reliable trend in technology.

Sutton named two methods that scale indefinitely: search and learning. Both turn extra computation directly into better performance, and neither requires the researcher to understand the problem as well as the solution eventually does. His prescription was to stop building in what we think we know about how minds work, and instead build systems that can discover such structure for themselves.

Did you know? "The Bitter Lesson" runs to under 1,200 words, yet it reshaped research agendas worth billions; Sutton later shared the 2024 Turing Award with Andrew Barto for founding reinforcement learning.

The essay became the scaling era's founding text, quoted in boardrooms as often as in laboratories. The rise of large language models, trained with comparatively simple objectives on ever more data and compute, reads like a vindication. Critics caution that the lesson describes a historical pattern, not a law of nature: scaling may meet limits of data, energy or cost, and clever ideas, such as the transformer itself, still matter. Sutton's point, though, was never that ideas are worthless, only that the best ideas are the ones that let computation do the work.

Bet on what scales, and expect it to win.

Seventy years of the bitter lesson 1970s Speech HMMs beat rules 1997 Deep Blue brute-force search 2012 Vision learned features 2016 AlphaGo search + self-play 2019 Bitter Lesson Sutton's essay 2024 Turing Award Sutton and Barto Methods that absorb more compute beat hand-built knowledge.
Fig 10 · The Bitter Lesson. Scale beats cleverness.
Free preview · Part I

Keep reading

You have read the first 10 of 100 chapters of The Visual Encyclopaedia of AI. The full book has every chapter and plate, as PDF and EPUB in English, French, Polish and Spanish.

The full book is coming soon to the shop.

Already bought it? Read the full edition online

The Visual Encyclopaedia of AI · First Edition, October 2026
100 chapters · 10 parts · one hundred diagrams
by Mat Siems · The Visual Encyclopaedia, No. 1 · 2026