In this guide · 7 sections
Short answer
An LLM (large language model) is AI software trained on vast amounts of text to predict the next word. Repeating that prediction lets it write, summarise and answer questions. It has no built-in fact-checking, so it can sound sure while being wrong. ChatGPT is an app; GPT models are the LLMs inside it.
What an LLM is, in one paragraph
A large language model is a program built from billions of numbers (called parameters or weights) that has been trained on a vast amount of text. Given some text, it predicts what text is likely to come next. Ask it "Write a polite reminder for an unpaid invoice" and it produces, piece by piece, the kind of text that typically follows such a request.
OpenAI's explanation of how its models are developed describes exactly this: during training, the model learns how words typically appear together and uses that to "predict the next most likely word when generating a response, one word at a time". Text is handled in small units called tokens, which can be whole words, parts of words or punctuation.
| Term | Plain meaning |
|---|---|
| Model | The trained program itself, such as GPT-6 Astra or Claude Opus 5.5 |
| Parameters / weights | The billions of numbers adjusted during training |
| Token | A chunk of text, roughly a word or part of a word |
| Prompt | What you type in, including any instructions or documents |
| Context window | How much text the model can consider at once |
| Hallucination | A confident answer that is false or made up |
How an LLM learns: training vs using
There are two very different stages, and mixing them up causes a lot of confusion about privacy and accuracy.
Pre-training: reading a vast library
The model is shown huge amounts of text and repeatedly asked to fill in the next word. Early guesses are random; over time its parameters are adjusted so the guesses improve.
Post-training: learning to be an assistant
Human trainers and further training teach it to follow instructions, refuse harmful requests and answer helpfully, rather than just continue text.
Using it: no learning by default
When you chat with a model, it doesn't permanently learn from that conversation in real time. Whether your chats are later used to train future models depends on the vendor's settings and your plan.
OpenAI also notes that models "do not store or retain copies of the data they are trained on"; they adjust numbers to reflect patterns. That is why an LLM can't reliably quote a specific document from training, and why it needs your documents supplied at question time to answer about your business. Our guide on embeddings and vector databases explains how that is done.
Why it sounds confident when it's wrong
An LLM is built to produce plausible text, not to check facts. If the most plausible-sounding continuation is wrong, it will write it in the same calm, fluent style as a correct answer. OpenAI's help article Does ChatGPT tell the truth? lists typical errors: incorrect facts and dates, fabricated quotes and references, and overconfident answers to ambiguous questions.
There is also built-in randomness. Because several next words can be plausible, the same question can produce different answers on different tries.
Read more in AI hallucinations: the business risk.
Models vs apps: ChatGPT is not the same as GPT
People use product names and model names interchangeably, but they are different layers.
| Vendor | App you use | Models inside (October 2026) |
|---|---|---|
| OpenAI | ChatGPT | GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna |
| Anthropic | Claude | Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 |
| Microsoft | Copilot | Models from Microsoft's partners and its own |
| Gemini app | Gemini models |
The app adds features around the model: web search, file uploads, memory, connectors to your email or drive, and admin controls. The model decides how good the writing and reasoning are. When comparing tools, ask both questions: which model does it use, and which app features and data protections come with the plan? Our guide to LLM benchmarks explains how models are compared.
Every model also has a knowledge cutoff. Anthropic's models overview lists a reliable knowledge cutoff of June 2026 for its current models. Anything after that is unknown to the model unless the app looks it up.
What this means for using AI in your business
Good fits for an LLM
- First drafts of emails, posts and proposals
- Summarising long documents or meeting notes
- Answering customer FAQs from your own approved content
- Rewording, translating and tidying text
Poor fits without safeguards
- Quoting prices or policies from memory
- Legal, tax or medical advice sent unchecked
- Anything needing exact, up-to-date facts without a source
- Pasting customers' personal data into free consumer tools
The pattern that works: give the model your own approved information, keep a human reviewing anything external, and use business plans whose terms say your data isn't used for training. How much text a model can take in at once matters for long documents; see context windows explained.
Do it yourself or get help?
Anyone can start with a business plan of ChatGPT, Claude or Copilot and a short internal rule: no customer personal data in prompts, and a human checks everything sent out. That covers most small-business uses.
If you want an assistant on your website that answers customers from your own content rather than general knowledge, our AI Website Assistant (S$790) sets one up, limited to your approved information, with a handover to a person when it can't answer.
Frequently asked questions
LLM stands for large language model. Large refers to the size of the model, built from billions of adjustable numbers, and the huge amount of text it was trained on. Language model means it works by predicting text. ChatGPT, Claude, Gemini and Copilot are all powered by LLMs.
Strictly, ChatGPT is an app built on top of LLMs. The language models inside it are OpenAI's GPT models. The app adds features such as chat history, file uploads, web search and admin controls. The same distinction applies to Claude, where the app runs Anthropic's Claude models, and to Microsoft Copilot.
The model itself doesn't. It answers from patterns learned during training, which stop at a knowledge cutoff date. Some apps add a web search tool that fetches current pages and passes them to the model. If an answer depends on recent facts, check whether the tool actually searched and look at its sources.
Because they generate the most plausible-sounding text, not verified facts. When the model lacks the right information, a confident but false answer can still be the most likely-looking continuation. Vendors call this hallucination. Supplying your own documents, asking for sources and having a person check important outputs all reduce the risk.
Not in real time; the model doesn't permanently change while you chat. But vendors may use conversations to train future models depending on your plan and settings. Business and enterprise plans from major vendors generally exclude your data from training by default, while free consumer plans often allow it unless you opt out.