What is AI, actually?

Here is the sentence this whole primer unpacks: modern AI is a machine that has been shown a great many examples and adjusted, step by tiny step, until its predictions on those examples stopped being wrong. Everything else – the chatbots, the image tools, the systems that finish your sentences – is that one sentence wearing different costumes.

Three words in it are carrying all the weight. The model is the machine itself: no gears, just arithmetic with an enormous number of adjustable settings. Training is the adjusting: the process that sets those numbers by showing the model examples and correcting it. And prediction is the output – the only thing the machine ever produces, whether it arrives looking like a caption, a diagnosis, or the next word of a poem. Each of these words gets a chapter of its own, and by chapter 5 they will have assembled into an honest picture of what happens when you type a question and something types back.

Three promises about what this primer will not do. It will not tell you whether any of this amounts to a mind – nobody knows, and where that matters the primer will say so plainly. It will not predict where the field is going – nobody knows that either, least of all the people who sound most certain. And it will not discuss any particular product. Products change by the month; the machinery underneath all of them is what this primer teaches, and its essentials have held still for decades.

The habit to resist is built into the language. “The model knows my name.” “It understood the question.” “It thinks the answer is Paris.” Every one of those sentences borrows a word from minds and lends it to arithmetic, and every loan quietly misleads.

Here is the better picture. A model is not a thing that knows; it is a thing that has been shaped. A river bed does not know where the sea is, yet water finds the sea through it, because years of water have carved it into a shape that sends water that way. A key filed until it turns in a lock holds no opinion about the lock. The model is the river bed, not the river; the key, not the locksmith. Training carved it – millions of small adjustments, each nudging its behaviour a little closer to the examples – and what remains afterwards is shape: a very long list of numbers, each one a knob that training turned and then left where it lay. The field calls these numbers parameters; this primer will mostly say knobs, because that is what they are.

You may notice that this picture says nothing about the brain. That is deliberate. The most popular analogy in circulation – that these systems are artificial brains built from artificial neurons – misleads far more than it explains, and chapter 3 exists to take it apart carefully.

Every system this primer covers, whatever it appears to do, has the same skeleton: examples go in, something adjusts, predictions come out.

It is worth drawing once, in words, so the later chapters can point back at it. On the left, a pile of examples – photographs with labels attached, sentences with their next word covered up. In the middle, the model: plain arithmetic, studded with knobs. On the right, its predictions, compared against the answers the examples carry – and the size of each miss fed back to nudge the knobs before the next example arrives. Around and around, until the predictions stop improving.

The chapters walk this skeleton left to right. Chapter 1 takes the left box: why showing examples won where writing rules failed, and what the training loop actually does when it runs. Chapter 2 takes the feedback arrow: how being wrong becomes a number, and how a machine walks that number downhill. Chapters 3 to 5 open the middle box – the unit it is built from, how words are turned into the numbers it can chew, and the loop that turns next-word-prediction into conversation. Chapter 6 closes on what this skeleton cannot do: the failures that follow from its shape, which is all of the famous ones.

One register note before you start. Nothing ahead is magic, and nothing is trivial – and you should be suspicious of anyone whose explanation leans on either word. Where the field genuinely does not know something, the primer will not smooth it over. Being careful is the whole value on offer here.