Inside the Mind of the Machine

• Room one. The first shapes: edges, colours and curves, and the ordered vocabulary a network learns before it learns anything else.
• Room two. From building block to understanding: how simple detectors are assembled into something more complicated, and how researchers rebuilt that mechanism by hand.
• Room three. One idea, many forms: a single place that holds a concept across image, word and drawing, and how that same ability can be fooled by a handwritten note on an apple.
• Room four. The machine copies itself: induction heads, small circuits that recognise a pattern and continue it of their own accord.
• Room five. More thoughts than space: superposition, where a model holds more concepts than it has neurons, and how dictionary learning drew millions of readable features out of a real model.
• Room six. The dial of a thought: what happens when you turn up a single concept inside a model and watch the behaviour move with it.
• Room seven. The biology of thinking: planning ahead to a rhyming word, mental arithmetic along two parallel paths, a language-independent space for thought, refusal and invention.
• Room eight. The limits of the gallery: why only a fraction of the neurons is genuinely understood, why a feature can mislead, and what is still not understood.
• Stories and cases: the bridge that thought it was a bridge, the apple that said it was an iPod, and the rabbit that was already there before the line began.
• Looking under the hood yourself.
• Frequently asked questions.
• Pitfalls and myths.
• Glossary.
• Sources and accountability, with the full source list.

For years a model like this was a closed box. We could see what went in and what came out, but not what happened in between. Then a young field built something like a microscope, and for the first time a person could look at the inner workings of an artificial brain in a concrete, measurable way. This book is a guided tour of what they found there. Take that place that lights up at the spider, the word and the cartoon character. One idea, in three very different forms, strikingly close to the way a single cell in the human brain responds to everything connected with one particular person. And then the other side of it, in the same space. Stick a handwritten note with the word "iPod" onto an ordinary apple, and the model sees an iPod. That scrap of paper was, as the researchers soberly noted, no more technology than pen and paper, and yet it shifted the entire judgement. Further on, researchers turned a single dial inside a real model, the feature for the Golden Gate Bridge, and the model began to call itself the bridge. Not science fiction, but a measurable intervention with a visible consequence. This ebook takes you through eight rooms of findings like these, laid out as a museum tour. No mathematics on the walls, but real discoveries, from the first shapes a network learns to recognise to the biology of thinking itself. This is not a how-to manual and not a scare story. With every finding you see exactly what was measured, and where the measurement stops and the interpretation begins. Wonder, with both feet on the ground.

• The first shapes. Neurons that learn to see edges, colours and curves, and a dog-head detector that the network assembled for itself out of one left-facing and one right-facing part.
• One idea, many forms. How a single place in the network holds on to a concept across image, word and drawing, surprisingly close to the concept cells in the human brain. And how that same ability can be fooled by nothing more than a scrap of paper.
• The machine copies itself. Induction heads, small circuits that recognise a pattern and continue it of their own accord.
• More thoughts than space. Superposition, where a model holds more concepts than it has neurons, and how researchers used dictionary learning to draw millions of readable features out of a real model.
• The dial of a thought. What happens when you turn up a single concept inside a model and watch the behaviour move with it.
• The biology of thinking. A model that plans ahead towards the rhyming word at the end of a line of verse, that does mental arithmetic along two parallel paths at once, that uses a language-independent space for thought, and that sometimes refuses and sometimes invents.
• The limits of the map. Why only a fraction of the hundreds of thousands of neurons is genuinely understood, and why a feature can also be a misleading interpretation.

This is for the curious professional who uses AI every day and finally wants to see what is inside it. For the generalist who finds it fascinating how such a system actually works. And just as much for anyone who has to set up governance, policy or oversight and needs an honest picture instead of hype, along with a sense of where the certainty ends and the interpretation begins. No technical background is required. All you need is an interest in the question of what is really happening in there. Would you like to walk around a little before you read? Feel free to look around the first room before you decide. The tour begins with the simplest shape a machine can recognise, and builds up calmly from there.

This was written for the reader who distrusts AI hype, and built deliberately so that you come away with a picture that is accurate rather than a picture that sells. The book rests on twenty-nine verifiable sources, and they are not there as decoration but as foundation. Every finding can be traced back to the research it came from. Just as important is the separation we hold to throughout. What was measured, we call a measurement. What is an interpretation of that measurement, we call an interpretation. The two never blur into one another. That distinction is the heart of it. These are real, measurable findings, and at the same time they are interpretations, not a literal photograph of a thought. A feature can be a genuine insight and still give a misleading impression, and we do not conceal that; we make it part of the tour. You will not read a claim here that AI is conscious or has intentions, and you will not read sensationalism. Nor will you read that you will understand everything afterwards. What you do get is this: you see with your own eyes what science has uncovered so far.

The book rests on twenty-nine verifiable sources, and every finding can be traced back to the research it came from. Throughout, a measurement is called a measurement and an interpretation is called an interpretation, including where the evidence runs out. • Around 44 pages • Written in English, also available as a Dutch edition • 2026 edition, published by BNK KAM • Eight rooms, with five diagrams and imagery • Immediate download as PDF and Word