What is an LLM? Explained simply, no maths

LLM stands for “Large Language Model.” It’s the engine behind ChatGPT, Claude and Gemini. The acronym is everywhere, yet rarely explained. Here’s the essential, without a single equation.

The idea in one sentence

An LLM is a system trained to predict the next word in a text, with such accuracy that what follows looks like a considered answer.

It is not a database that “looks up” an answer somewhere. There is no stored record saying “here is the capital of France.” During its training, the model saw billions of sentences where this kind of information appears, and it learned statistical regularities: after “the capital of France is,” the word “Paris” almost always follows. It generates its answer word after word (more precisely, “token” after token, a token being a fragment of a word) by relying on these regularities.

A simple analogy

Picture your phone’s autocomplete, but immeasurably more powerful and trained on a significant fraction of everything that has ever been written and digitised. Your keyboard suggests the next word based on the last three you typed. An LLM does the same, but taking a whole paragraph into account, sometimes a whole document, and with a far finer grasp of context, subject and expected style.

Where its abilities come from

Three ingredients explain why these models have become so good:

• The size of the data. Massive volumes of text — books, websites, code, conversations — expose the model to almost every subject and every writing style.
• The size of the model. The more parameters a model has (internal settings adjusted during training), the more nuance it captures. This is partly what distinguishes a “fast” model from an “advanced” one at the same provider.
• Post-training tuning. A raw model just predicts the continuation of a text. Additional work — refining it on examples of good answers, with human feedback — turns it into an assistant that actually answers a question rather than simply continuing a text at random.

What this explains, very concretely

Understanding this mechanism sheds light on behaviours that otherwise seem strange.

Why it sometimes “makes things up” with confidence. The model does not check a fact in a database: it produces the most plausible continuation. On a niche or poorly documented topic, plausible and true can diverge — this is what we call a hallucination. We break down this mechanism and how to guard against it in why AI “hallucinates”.

Why the prompt matters so much. Since everything hinges on the text provided as input, the quality and precision of your instruction directly change the quality of what follows. A vague prompt yields a vague continuation.

Why it “forgets” the start of a long conversation. The model only retains what fits within its context window, the span of text it can process at once. Beyond that, the earliest exchanges gradually drop out of its field of view.

Why its knowledge has an expiry date. By default, an LLM only knows what was in its training data, cut off at a given date. For recent news, it needs access to web search or a document you provide.

What an LLM is not

It is neither a conscious intelligence, nor a search engine, nor a reliable calculator by nature. It is an extraordinarily capable text-generation system, whose accuracy depends on what it has learned and how it is prompted. For a fuller tour of the related vocabulary, our 10-word glossary complements this one.

Key takeaway

An LLM predicts the next word from regularities learned across immense volumes of text — nothing more mystical, but nothing less powerful. This simple mechanism explains both what makes these tools so useful day to day and why they remain fallible on precise facts. Knowing this requires no technical skill; it is enough to use them with discernment.