What are large language models (LLMs)?

What are large language models (LLMs)?

A large language model is a machine-learning model trained to capture patterns in language at substantial scale. Many widely used LLMs use transformer architectures and learn from large collections of text and other training material.

"Large" is a descriptive term rather than a fixed threshold. Models differ in architecture, parameter count, training data, and capabilities.

Learning through prediction

A common training objective is to predict the next token from earlier tokens. A token may be a whole word, part of a word, punctuation, or another text unit.

Repeated prediction across many examples encourages the model to learn patterns involving grammar, facts, writing styles, and relationships between ideas. The resulting knowledge is represented across numerical parameters, rather than organized as a conventional table of verified facts.

One model, many language tasks

Because many tasks can be expressed through language, an LLM can support summarization, translation, classification, drafting, and code generation.

For instance, the same model might summarize a product review and then extract the reviewer's complaints into a structured list. Instructions and examples help specify which operation the application needs.

A base model may receive additional training to follow instructions and behave usefully as an assistant. The assistant product can also add search, files, memory, and other tools.

Capability does not guarantee correctness

An LLM can generate text that sounds convincing while containing an incorrect date, a nonexistent reference, or a faulty calculation. It may also lack relevant information outside its available context.

Use a model for the tasks it can perform reliably, supply appropriate context, and evaluate the outputs that matter. Fluency is valuable, but it is only one dimension of quality.