The Machine No One Understands

• The black box, and why it follows from the way these systems learn rather than from trade secrets
• The Golden Gate experiment, a single concept located inside a model and turned up until the behaviour moved with it
• The apple and the note, what an advanced model is really doing when it "sees"
• The Spider-Man neuron, one concept across photograph, word and drawing, and its parallel with concept cells in the human brain
• Superposition, and why a single neuron never neatly means one thing
• Explanations after the fact, when a tidy step-by-step account is not the real reason behind an answer
• The hopeful turn, millions of readable features drawn out of a real model
• Sources and accountability, 32 verifiable references with fact and interpretation kept apart

You probably use this machine every day. You ask a question, you receive a fluent and convincing answer, and you carry on with your work. But how did that answer come about? That is the uncomfortable secret behind the convenience. Modern AI is a black box, not because a company is hiding something, but because of how it is made. These systems are not programmed line by line. They are trained, and they learn their own strategies, hidden in billions of numbers that no one has written out by hand. Just how tangible this is becomes clear with an apple. Researchers stuck a handwritten note with the word iPod onto an ordinary apple, and an advanced model then identified that apple as an iPod with near-complete confidence. No hacking, no sophisticated equipment. In the words of the researchers themselves, it requires no more technology than pen and paper. What makes this remarkable is that the makers admit it themselves, in public. This book walks you through that terrain calmly, without panic and without conspiracy. It is wonder, backed by evidence.

• You will be able to explain why AI is a black box because of the way it learns, and not because of trade secrets.
• You will be able to say what the Golden Gate experiment actually shows, namely that a concept inside an AI has a findable and steerable location.
• You will understand why an apple with a note stuck to it was suddenly read as an iPod, and what that tells you about machine "seeing".
• You will know what a concept neuron is, and why the Spider-Man neuron looks so much like the concept cells found in the human brain.
• You will be able to explain superposition, and why simply looking inside a network is not enough when a single neuron never neatly means one thing.
• You will recognise when a tidy, step-by-step explanation from a model is not the real reason behind its answer.
• You will know why the story ends hopefully, now that researchers are able to draw millions of readable features out of a real model.
• You will be able to say, in your own words, exactly where today's certainty ends and interpretation begins.

This book is for the curious professional who uses AI every day and quietly wonders what is happening under the hood. It is for the generalist who does not want mathematics but does want the real story. And it is for the decision-maker or board member who is responsible for governance, risk or policy and is looking for an honest, well-evidenced basis rather than sensation. You do not need a technical background, only the wish to genuinely understand.

A great deal is said about AI, and most of it is either panic or sales talk. This book is deliberately the opposite. It is built on 32 real, verifiable sources, all of them listed in the book itself. Nothing is invented. Every quotation is a real quotation, every scene is a real experiment, every claim can be traced back. Fact and interpretation are kept neatly apart, so that you always know what is established and what is a reading of the evidence. The most surprising claim in this book, moreover, does not come from us but from the makers themselves. Google's chief executive describes it, in so many words, as a black box he does not fully understand. The chief executive of a leading AI lab writes that generative AI is grown rather than built. The chief executive of OpenAI acknowledges that interpretability has not yet been solved, and one of the founders of the field says that we did design the learning algorithm, but that we do not really understand how the models do what they do. The attitude of the book is wonder, neither fear nor dismissal. There is no promise that you will understand everything afterwards, but there is a clear and honest picture of what we do and do not know today. That is the tone of BNK KAM, calm, precise and free of noise.

The strongest evidence in this book is public and comes from the makers themselves: the head of Google on a black box he does not fully understand, the chief executive of a leading AI lab on generative AI being grown rather than built, and the chief executive of OpenAI on interpretability that has not yet been solved. Alongside those statements stand the experiments as they were actually carried out. The Golden Gate Bridge feature, located inside a real model and turned up until the model spoke of itself as the bridge. The Granny Smith apple with a handwritten iPod note on it, classified as an iPod with even higher confidence than before. The 32 references at the back let you look up each of these scenes at the original research. • Around 44 pages, with 5 diagrams and imagery • Written in English, also available in Dutch as De Machine Die Niemand Begrijpt • 2026 edition, published by BNK KAM • Immediate download as PDF and Word