What Is a Large Language Model (LLM)?

How a Large Language Model Works

At its core, an LLM does one thing: given some text, it predicts what text is most likely to come next, one small piece at a time. Those pieces are called tokens, which are whole words or fragments of words. Google's Gemini documentation, for example, notes that a token is roughly four characters of English text. When you ask a question, the model generates a response token by token, each one chosen based on your question and everything it has written so far.

That sounds too simple to produce useful answers, but predicting text well requires learning a great deal: grammar, facts mentioned frequently in the training data, how arguments are structured, how code works, and how different writing styles sound. "Large" refers to the scale of the model (billions of internal parameters) and of the training data.

Most current LLMs are built on the transformer, an architecture described in the 2017 paper "Attention Is All You Need." Its key idea, attention, lets the model weigh how relevant every earlier token is to the one it is predicting, which is what lets it keep track of meaning across long passages.

How LLMs Are Trained

Training happens in two broad stages:

1. Pretraining. The model reads a very large collection of text and code and learns to predict the next token. This is where its general knowledge and language ability come from, and it is the expensive part.
2. Post-training. The pretrained model is tuned with examples of good responses and with human and automated feedback so that it follows instructions, answers helpfully, and declines harmful requests. This is what turns a text predictor into an assistant.

Training a frontier model from scratch costs far more than any small business would spend, which is why almost every business uses a model someone else trained and customizes the inputs instead. Customization usually means better instructions, supplying your documents at question time, or, less often, a light fine-tuning step on your own examples.

Training ends at a cutoff date, so a model's built-in knowledge is frozen at that point. Products work around this by letting the model search the web or read your documents at the time of the question.

Key Terms Business Owners Will Hear

• Context window: the maximum amount of text, in tokens, the model can consider at once, including your instructions, any documents and the conversation so far. Current models handle long documents, but bigger inputs cost more and can dilute focus.
• Prompt: the instructions and input you give the model. Clear prompts with examples get much better results.
• Tool use or function calling: the model's ability to request an action, such as searching a database or booking an appointment, which your software then performs.
• Multimodal: models that also accept images, audio or video, not just text.
• Open-weight vs proprietary: some models can be downloaded and run on your own servers; others are used only through the vendor's service. See open-source vs proprietary LLMs.

What LLMs Do Well, and Where They Fail

Strengths: drafting and editing text, summarizing long documents, extracting structured data from messy input, classifying and routing requests, translating, answering questions from supplied material, and writing or explaining code.

Limits: they can state false things confidently (hallucination), they do not know your business unless you supply that information, their knowledge has a cutoff date, they can be steered by instructions hidden in the content they read, and the same question can produce slightly different answers each time. Exact arithmetic and counting are better handed to ordinary code, which models can call as a tool.

Researchers are still working to understand exactly how models arrive at their outputs. Anthropic's interpretability research, for instance, traces internal steps inside its models and has found evidence of planning ahead when writing, but this is an active field, not a settled one.

How Businesses Use LLMs in Practice

Most businesses meet LLMs in three ways: through chat assistants for staff (ChatGPT, Claude, Gemini, Copilot), through AI features inside software they already use, and through custom systems that connect a model to their own data and tools. The third is where automation happens, using techniques like retrieval-augmented generation and agents. Choosing a model matters less than people expect; the data, integrations and testing around it usually decide whether a project works. DR3AM AI automation focuses on that surrounding work.