What is a large language model (LLM)?

A large language model (LLM) is an AI model trained on vast amounts of text to predict the next token, which lets it write, summarize, translate and answer.

How it works

Google's Machine Learning Crash Course gives the core idea: "A language model estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens," where a token can be a word, part of a word or a single character (Google). Generate one likely token, append it, repeat, and you get sentences, code or a reply in a chat.

Modern LLMs are built on the Transformer architecture, introduced in the 2017 paper "Attention Is All You Need" (Vaswani et al.). Its attention mechanism lets the model weigh every part of the input when producing each token, instead of reading strictly word by word. "Large" refers to the scale of training data and model parameters.

Training usually happens in stages: pretraining on broad text to learn language and general knowledge, then tuning on examples and feedback so the model follows instructions and declines harmful requests.

What LLMs are good and bad at

They are good at language: drafting, rephrasing, summarizing, classifying, extracting fields, translating, and holding a natural conversation.

They have predictable limits:

• Knowledge cutoff. A model knows nothing after its training data ends, and nothing about your company that wasn't public and widely repeated.
• Hallucination. Without a source, a model may produce fluent, wrong answers; see hallucination.
• Context limits. It can only consider so much text at once, measured in tokens within its context window.
• Cost and speed grow with the amount of text processed.

Why it matters for your website

An LLM is the engine inside most website chat assistants, AI search boxes and AI agents. On its own it doesn't know your prices, policies or product details, which is why serious website assistants pair it with retrieval-augmented generation: the model supplies the language, your content supplies the facts.

When comparing tools, the model name matters less than what surrounds it: which content it answers from, whether it cites sources, what it does when it doesn't know, and when it hands off to a person.

Example

Ask a general-purpose LLM, "What does Acme's Team plan cost?" and it will either admit it doesn't know or guess, because that fact is specific and changes over time. Give the same model the text of Acme's current pricing page along with the question, and it can answer correctly and point to the page.

Related terms

• Tokens: the units LLMs read and write.
• Context window: how much an LLM can consider at once.
• Hallucination: confident answers without support.
• AI agent: an LLM that plans and uses tools.
• Retrieval-augmented generation (RAG): giving an LLM your facts at answer time.