Definition

A large language model (LLM) is a neural network with a very large number of parameters, trained on massive amounts of text. Its core skill is simple: given some text, predict what most likely comes next. On top of that objective, the model picks up grammar, factual knowledge, reasoning patterns and correspondences between languages, which lets it handle question answering, writing, translation, coding and much more.

“Large” refers to two things: parameter count (typically in the billions) and training data volume. Nearly all mainstream LLMs today are built on the Transformer architecture.

How It Works

An LLM usually goes through several stages before it’s ready to use:

Pretraining. The model learns from large-scale text (web pages, books, code) through self-supervised learning, predicting the next token from what came before. No human labels are needed, since the text itself supplies the answers. This stage uses most of the compute and gives the model broad language ability and world knowledge.

Instruction tuning. A pretrained model is good at continuing text but not necessarily at following instructions. Further training on many instruction/response examples teaches it to answer the way users ask.

Preference alignment. Methods such as reinforcement learning from human feedback (RLHF) push responses toward what people prefer: more helpful, more honest, less harmful.

Inference. At use time, the model splits input into tokens, predicts the next token, appends it, and predicts again, looping until generation ends. Sampling parameters like temperature control how random the output is.

Typical Applications

  • Conversational assistants: answering questions, offering suggestions, helping people think things through.
  • Content creation: drafting articles and copy, rewriting, summarizing.
  • Coding help: generating and explaining code, finding bugs, writing tests.
  • Translation and cross-lingual work: translating between languages, cross-language search.
  • Extraction and classification: pulling fields from contracts, résumés or tickets; tagging sentiment or topic.
  • Agents: acting as the “brain” that plans steps and calls external tools for complex tasks.

LLMs can hallucinate, producing plausible but false content. Where factual accuracy matters, they’re usually paired with retrieval-augmented generation (RAG), source citations or human review.

  • Transformer: the dominant architecture behind today’s LLMs.
  • Token: the basic unit models read and write, and the unit for billing and context length.
  • Context window: the maximum number of tokens a model can process at once.
  • Fine-tuning: further training a pretrained model on specific data.
  • RLHF: a training method that aligns models using human feedback.
  • Hallucination: when a model generates content that isn’t true.
  • Multimodal models: models that handle images, audio and other inputs besides text.