Picture a reader who has walked past almost every library in the world. It did not memorize the books; that would be absurd. What it did was something subtler and more unsettling: it internalized the way words chase one another. It learned that "once upon a time" is rarely followed by "thermostat", that questions tend to end in a question mark, that a paragraph about grief has a different cadence from one about taxes. It did not read to hoard facts. It read to absorb patterns. That reader, imaginary, statistical, made of numbers, is the closest thing we have to a large language model.
01 · the mechanismGuessing the next word, a million times over
Beneath all the apparent sophistication, an LLM does one thing, obsessively: predict the next token. A token is a unit of text: sometimes a word, sometimes a fragment like "predic-" or a punctuation mark. Given a stretch of text, the model computes, across its entire vocabulary, which token is most likely to come next, and then the next, and the next, weaving a response as it goes.
It sounds modest, almost trivial. But a large idea is hiding in there. To reliably predict the word that follows "The Pythagorean theorem relates the legs of a triangle to the…", the model must have captured, somehow, something akin to geometry. Predicting well demands compressing the world. The task is simple; what it takes to solve it is not.
The task is humble (which word comes next?); what it takes to pull it off ends up looking suspiciously like knowledge.
02 · the learningNobody told it the answers
What is remarkable is how it learns. There is no army of humans labeling correct examples. The text labels itself: you take a sentence, hide the last word and ask the model to guess it. If it fails, its internal numbers are nudged slightly so that next time it fails a little less. This is self-supervision: the correct answer is already in the text, you just have to cover it up. Repeat this exercise over an enormous slice of everything humanity has written (books, articles, code, conversations) trillions of times.
Those "internal numbers" that get adjusted are the parameters, also called weights. They are, literally, the dials of the system: values that determine how much each signal influences the next. Learning is nothing more than finding, among billions of those dials, a configuration that predicts well. Today's models range, very roughly, from a few billion to hundreds of billions of parameters. They are not stored facts; they are the distilled memory of training, dissolved into arithmetic.
03 · the architecturePaying attention, and almost nothing else
How does the model decide which part of the text matters for predicting what comes next? This is where the piece that changed everything comes in. In 2017, a group of researchers published a paper with an almost arrogant title: Attention Is All You Need. They proposed an architecture, the transformer, whose heart is a mechanism called attention.1
Attention lets the model, as it processes each word, weigh all the others and decide which are relevant. In "the cat that was chasing the mouse was tired", to know who was tired, the model "attends" to cat, not mouse. It does this for every word, in parallel, across many stacked layers. Each layer refines the meaning. That ability to look at the entire context at once, instead of reading in single file, word by word, was what made it possible to train enormous models efficiently.
Attention is not a poetic metaphor: it is a mathematical operation that decides, for each word, how much every other word in the context weighs.
When we scale these models up (more data, more parameters, more compute) abilities appear that nobody programmed explicitly: translating, summarizing, reasoning in steps, writing code. They were not taught one by one; they emerged from training at scale. It is the most disconcerting observation in the field: quantity, past a certain threshold, turns into something that looks like quality.
04 · the limitWhat it is not
It is worth dismantling two misunderstandings, because false expectations are born from them. The first: an LLM does not understand the way you do. It has no beliefs, no intentions, no lived experience of what it says. It manipulates statistical patterns with astonishing skill, and that skill is sometimes indistinguishable from understanding, but the mechanism is different. Mistaking fluency for consciousness is the central mirage of this technology.
The second: when it answers, the model does not consult a database or search the internet. Everything it "knows" is encoded in its weights, compressed during training. That is why it can be wrong with poise: it is not retrieving a filed fact, it is generating the most plausible continuation. When the plausible coincides with the true, it gets it right; when it does not, it hallucinates with the same confidence. The model does not tell the two cases apart, because to it they are the same operation.
An LLM does not seek the truth. It seeks the most probable continuation. That the two so often coincide is our good fortune, not its purpose.
And yet, and here I return to the reader from the beginning, there is something genuinely new in all of this. A machine that, without understanding, has distilled so much of our language that it can keep us company in thought. It is not a mind. It is a statistical mirror of everything we have written, polished until it begins to speak. Treating it as an oracle is naive; treating it as a sharp instrument, with its edges and its limits, is the start of using it well.
Sources
- Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems (NeurIPS) 30. The publication that introduced the transformer architecture and the attention mechanism on which today's LLMs are based.
- On the emergence of capabilities with scale, see the research literature on scaling and emergent behaviors in language models (2020-2022), which documents how certain abilities appear only once particular thresholds of size and compute are crossed.