What is a large language model and how does it work?
The plain-English guide to the technology behind ChatGPT, Claude, and every AI that seems to understand you.
A large language model (LLM) is a type of AI trained on vast amounts of text so it can read, write, summarise, and answer questions in natural language. It works by learning patterns across billions of words, then predicting the most useful next word, sentence, or idea. Think of it as autocomplete that went to university.
What does "large" actually mean?
Large refers to two things at once: the size of the training data and the number of "parameters" inside the model. Parameters are the adjustable internal settings, a bit like tiny dials, that the model tunes during training to get better at predicting text. GPT-4 is estimated to have hundreds of billions of parameters, which is why running these models needs serious computing power.
More parameters do not always mean a smarter model. Researchers have found that carefully curated, smaller datasets can produce surprisingly capable models at a fraction of the size. The "large" label is becoming a convention more than a strict rule.
How does training actually work?
Training is the process of showing the model an enormous amount of text and having it repeatedly guess the next word, then adjusting those internal dials whenever it guesses wrong. Do that billions of times and the model starts to capture grammar, facts, reasoning styles, and even tone.
The dominant method behind modern LLMs is called the transformer architecture, introduced in a landmark 2017 paper. Transformers use a mechanism called "attention" that lets the model weigh how relevant each word in a sentence is to every other word, so it grasps context rather than just word order.
After that initial training, most commercial models go through a second step called RLHF (reinforcement learning from human feedback). Human raters score the model's answers, and those scores teach the model to be more helpful and less harmful.
What can an LLM actually do?
In practice, a well-trained LLM can draft emails, explain medical jargon, write code, translate languages, and summarise long documents. A real-life example: a GP practice in the UK used an LLM to draft patient follow-up letters, cutting admin time by roughly half, because the model could turn a doctor's bullet points into full, readable prose in seconds.
What it cannot do reliably is reason from first principles, access live information (unless given a tool to search the web), or guarantee factual accuracy. LLMs sometimes "hallucinate", meaning they confidently state things that are simply wrong. Anthropic's research on model behaviour covers why this happens and how builders try to reduce it.
How is an LLM different from a search engine?
A search engine finds and ranks pages that already exist. An LLM generates new text from scratch, drawing on patterns it absorbed during training. Search gives you a list of sources; an LLM gives you a synthesised answer with no guarantee it points to the original evidence.
The two are increasingly combined. Tools like retrieval-augmented generation (RAG) let an LLM pull in live documents before answering, mixing the best of both approaches.
What does it cost to use one?
| Model / service | Free tier | Paid tier starts at |
|---|---|---|
| ChatGPT (OpenAI) | Yes, GPT-4o mini | ~$20/month for Plus |
| Claude (Anthropic) | Yes, Claude 3 Haiku | ~$20/month for Pro |
| Gemini (Google) | Yes, Gemini 1.5 Flash | ~$20/month for Advanced |
| API access (build your own) | Usage credits on sign-up | Pay per token (fraction of a cent) |
For most casual users, free tiers cover everyday tasks. Heavy users or developers paying per token (a token is roughly three to four characters of text) can rack up costs fast on long documents.
Are there privacy catches I should know about?
Yes, and they matter. When you type a prompt into a commercial LLM, that text is sent to the provider's servers. Many providers state in their terms that they may use your conversations to improve future models unless you actively opt out. OpenAI's privacy controls let you turn off training data use, but the setting is not on by default for all users.
Do not paste personal data, patient records, financial details, or confidential business information into a public LLM unless you have checked the provider's data-processing terms carefully.
Common questions
Is a large language model the same as artificial general intelligence?
No. An LLM is a specialised system trained on text prediction. Artificial general intelligence (AGI) would mean a system that can learn and reason across any domain the way a human can, something researchers consider an open and unsolved problem.
Can an LLM learn new facts after it is released?
Not on its own. A standard LLM has a "knowledge cut-off", the date its training data ended, and it does not update itself when new events happen. Some systems add live search tools on top to get around this limit.
Why does an LLM sometimes make things up?
Because it is fundamentally a pattern-completion engine, not a fact-checker. If the pattern of a sentence points strongly toward a plausible-sounding answer, the model produces it, even when no reliable evidence exists. Research into this hallucination problem is active across major labs.
