What Is an LLM?

A large language model (LLM) is a neural network trained on massive amounts of text to predict and generate language. It learns statistical patterns — which words tend to follow other words, which phrases signal a question versus an answer, how explanations differ from arguments. It does not store facts like a database. It learns how text is written, then generates likely continuations based on what it has seen.

A fable — The Library of All Words: A scribe once tried to copy every book in the kingdom. She read histories, poems, legal codes, almanacs, and children's rhymes. She could not hold it all in memory, so she learned instead which words usually appeared near other words. She became fast at finishing sentences, answering questions, and writing in the style of any volume she had studied. When asked about the law, she answered in the tone of a lawyer. When asked about medicine, she drew on the language of healers. She was not the books — she was a model of how the books talked.

The training is not supervised by a teacher pointing to the right answer. The model learns by predicting the next word in billions of sentences, then adjusting its internal connections to get better at prediction. This is called self-supervised learning — it lets models train on raw text without humans labeling every example.

Scale changes what the model can do. A model trained on millions of sentences learns basic patterns. A model trained on billions learns nuances, reasoning steps, and how to follow instructions. A model trained on even more can reason across domains it has never seen explicitly labeled. This emergence of new capability with scale is why AI progress has been closely tied to compute.

LLMs are not databases. They do not retrieve facts on command. They generate likely language conditioned on everything they have seen. When the training data contains errors or biases, the model reflects them. When a question is ambiguous, the model may resolve it in unexpected ways.

The practical lesson is to match the tool to the task. LLMs excel at drafting, explaining, translating, summarizing, and reasoning through open-ended problems. They are weaker at providing verified facts without retrieval, handling recent events outside training data, and performing precise calculations. Strong AI products combine a capable model with retrieval, evaluation, and guardrails for high-stakes use cases.