An LLM, short for large language model, is an AI system trained on an enormous amount of text so it can read and write language. Claude, GPT, Gemini and Llama are all LLMs. You give one a piece of text, called a prompt, and it writes a continuation: an answer, a summary, an email or a block of code.
How a large language model produces text
Text is first split into tokens, which are whole words or pieces of words. During training the model sees billions of examples and adjusts its internal numbers, called parameters, to get better at predicting the next token. Modern LLMs use the transformer design, described in the 2017 research paper "Attention Is All You Need", which lets the model weigh how every token relates to every other token in the input.
After this pretraining, companies tune the model on conversations and human feedback so it follows instructions and declines harmful requests. When you chat with it, it generates one token at a time, each chosen from the probabilities it predicts. The amount of text it can consider at once is its context window.
What that means in practice
Because an LLM predicts plausible text rather than looking facts up, it can state wrong things confidently. This is called hallucination. Its built-in knowledge also stops at a training cutoff date. Products fix both gaps by feeding the model fresh information at question time, a technique called RAG, or by giving it tools such as web search.
An example for a local business
A plumbing company pastes its price list and service area into an LLM and asks it to draft ten FAQ answers for its website. The draft takes seconds. The owner then checks each answer, because the model may have filled gaps with guesses, such as an emergency call-out fee the company never charges.
Where LLMs are useful for a small business
Drafting copy, replying to routine emails, summarising reviews, powering a website chatbot and generating code are all common uses. They work best on tasks where a human reviews the output or a mistake is cheap.
Common misunderstandings about LLMs
- Treating output as fact-checked. Verify names, prices, laws and figures before publishing.
- Assuming it knows your business. It only knows what is in the prompt or what it was trained on. Give it your details.
- Pasting sensitive data carelessly. Check the provider's data policy before sharing customer records.
- Publishing raw drafts at scale. Mass-produced, unedited AI pages add little value and can hurt search performance.
Related terms
- Prompt engineering: writing inputs that get better answers.
- RAG: grounding a model in your own documents.
- AI agent: an LLM that can take actions with tools.
- Chatbot: the most familiar product built on one.