In one sentence

A large language model is an AI system trained on huge amounts of text that can read and produce language, and is the engine behind most current AI tools.

The simple model

You have probably used one already, through an assistant like ChatGPT or a similar tool. When you type a question and get a fluent answer, a language model produced it. It works by predicting likely, sensible text in response to what it is given.

That last point matters more than it sounds. The model is producing a likely answer, not looking up a fact. Usually the likely answer is also correct. Sometimes it is not, and the model cannot tell the difference. This is why its output is probabilistic and why it can hallucinate.

Why you care

Language models are what make the interpretive parts of a workflow possible: reading an email, classifying a request, extracting fields, drafting a reply, answering from documents. Anywhere your workflow needs to understand or produce language, a model is likely involved.

Knowing what one is tells you what to expect: it is strong at language and interpretation, and it is not a reliable database, calculator or rule engine. Those jobs are better handled by deterministic software.

What it is good at and not good at

  • Good at: understanding and producing language, summarizing, classifying, extracting, drafting, answering from provided information.
  • Not reliable for: exact facts from memory, precise calculation, guaranteed consistency, or anything that must be exactly the same every time.

The art of designing an AI workflow is largely about using the model for what it is good at and using rules for the rest. See automation vs AI vs agent.

A small example

A team used a model to draft answers to customer questions, which it did well, and separately used a fixed rule to look up the exact order total, which the model should not be trusted to recall. Each tool did the part it was suited to.

What this does not solve

Understanding what a model is does not tell you which specific model to use or how it will perform on your task. Capability varies and changes over time. You still measure fit and performance through evaluation on real cases.