An LLM is a sentence-completion machine, but supercharged to a dizzying scale. You know the smartphone keyboard feature that suggests the next word as you type? A large language model is the same basic principle, trained on a gigantic fraction of everything humanity has written, until it becomes capable of writing a novel, fixing code, or explaining quantum physics. All the magic lies in that little gesture repeated billions of times: guessing what comes next.
It's the most ubiquitous acronym in tech news. LLM here, LLM there. You're told Claude is an LLM, GPT is an LLM, that such-and-such Chinese model is a bigger LLM than the others. But if you were asked, right now, to explain what that means, could you? Don't worry, by the end of this article, yes.
The simple definition
LLM stands for Large Language Model, in French grand modèle de langage. Let's break down the three words. Model: a computer system that has learned patterns from data. Language: it works on text, words, sentences. Large: it's trained on colossal amounts of text, and itself contains an astronomical number of internal settings.
Its core function is surprisingly simple to state: predict the most likely next word in a sequence. You give it "The sky is", it calculates that "blue" is very likely, "clear" plausible, and "tractor" very unlikely. It picks, adds the word, then starts again including that new word, over and over. That's it. Absolutely everything an LLM does stems from this repeated prediction mechanism.
How can something so simple do so much?
Here's the real mystery, and it's fascinating. How can predicting the next word produce reasoning, working code, translations, poems? The answer lies in scale and in a discovery no one had fully anticipated.
To become truly good at predicting the next word across billions of different texts, the model is, in a sense, forced to learn things about the world. To complete "The capital of France is" correctly, it has to have absorbed a form of geographical knowledge. To complete a mathematical proof correctly, it has to have grasped a logic. Next-word prediction, pushed to sufficient scale, forces the emergence of abilities that resemble understanding. Researchers call these emergent abilities: skills that weren't explicitly programmed, but arise when the model gets big enough.
An LLM doesn't "understand" in the human sense, and it doesn't consult a database of facts. It manipulates probabilities over words. That's why it can produce a false answer with total confidence: what it generates is the most plausible text, not necessarily the most true. This is also the root of the hallucination phenomenon, which we explored in our article dedicated to hallucinations. Keeping this nuance in mind completely changes how you trust these tools.
The ingredients of an LLM
Three elements give rise to an LLM. First, training data: immense amounts of text from books, websites, code. Then, the architecture, almost always what's called a Transformer, a 2017 invention that allowed models to account for context much more finely. Finally, parameters: the model's internal settings, adjusted during training, which today number in the hundreds or even thousands of billions.
It's this last figure that often serves as a size indicator. When you read that a model has "1.5 trillion parameters", that refers to the number of these internal settings. The more there are, the more nuances the model can in principle capture, but at the cost of higher computational expense. Note, however, that a larger number of parameters doesn't automatically guarantee a better model: data quality and training matter just as much.
What to remember
An LLM is thus a probabilistic machine that, by repeatedly predicting the next word over a vast share of written knowledge, has developed abilities that resemble understanding and reasoning. It's neither a brain, nor an encyclopedia, nor an oracle. It's a statistical tool of staggering power, with real strengths and equally real limitations.
Understanding this is already knowing how to use it better: leveraging its fluency to write, brainstorm, code, while keeping a critical mind about what it claims, because at its core, it's always looking for the most plausible word, not necessarily the truth. This dual awareness — marvelling at the capability while staying lucid about the mechanism — is probably the most useful skill to cultivate in the age of LLMs.