How large language models work, without the maths · Pulse Beat
Millions of people use chat assistants now, and many have a picture in their heads that is partly right and partly wrong. A clearer picture helps you use them better and trust them appropriately.
Millions of people use chat assistants now, and many have a picture in their heads that is partly right and partly wrong. A clearer picture helps you use them better and trust them appropriately. You do not need any maths for this one.
The core trick: predict the next piece
A large language model is trained to do one thing: given some text, predict what comes next. Not the next sentence in one leap, but the next small chunk, called a token, which is often a word or part of a word. Then it adds that token to the text and predicts the next, and the next, until it stops.
That sounds too simple to produce an essay, a poem or working code. The surprise of the last several years is how much capability emerges when this simple task is done at huge scale on huge amounts of text, with a model of enormous size.
What "training" means
Imagine showing a student billions of passages with a word hidden and asking them to guess it. Each time they guess wrong, they adjust a little. After enough examples, they pick up grammar, facts, writing styles, reasoning patterns and even some logic, because all of those help predict what comes next.
The model itself is a very large collection of numbers, called parameters or weights. Training adjusts these numbers. When you finish, the model is those numbers, a file of fixed values, not a database of stored sentences. It does not look things up; it computes a likely continuation based on patterns it absorbed.
From text predictor to assistant
A model trained only to continue text will happily continue a rude message with another rude message. To turn it into a helpful assistant, developers add further stages.
Instruction tuning. The model is trained on examples of questions and good answers, so it learns the format of being helpful.
Learning from feedback. People, or other models, compare different answers and indicate which is better. The model is nudged toward answers people prefer, which shapes tone, honesty and safety behaviour.
Guidelines and guardrails. Further rules and checks discourage harmful or misleading outputs.
This is why assistants feel polite and structured: that is trained behaviour on top of the raw predictor.
Why it sometimes makes things up
This is the most important thing to understand. The model produces what is plausible, not necessarily what is true. If the text pattern calls for a citation, it can generate something that looks exactly like a citation even if no such source exists. This is often called a hallucination, though the better way to see it is that fluent text and correct text are different things, and the model is trained for the first.
Models have improved, and connecting them to search or documents can ground answers in real sources. But the safe habit stays the same: treat output as a draft from a fast, well-read, occasionally overconfident assistant, and check anything that matters.
What it does well
- Rewriting, summarising and changing the tone of text.
- Explaining a concept at different levels and giving examples.
- Brainstorming options and structuring ideas.
- Drafting code, converting between formats, and explaining errors.
- Translating and language practice.
Where it is weak
- Precise facts and figures, especially obscure ones or recent ones past its training data.
- Arithmetic and exact logic without tools, though tools help.
- Knowing what it does not know. It may answer confidently when a person would say they are unsure.
- Consistency across a long task, where small errors can pile up.
- Anything needing real-world verification, such as checking whether a website exists today.
Context: the memory of a conversation
A model can only consider a limited amount of text at once, called its context window. Everything in the conversation so far, plus any documents you paste, must fit. Within that window it can use what you said; outside it, nothing is remembered unless the product builds extra memory around it. When a conversation seems to forget an earlier detail, that is often why.
Prompts: why wording matters
Because the model continues from what you give it, clear input produces better output. Helpful habits include saying who the answer is for, what format you want, giving an example of what good looks like, and supplying the relevant material instead of hoping it knows. Asking it to explain its reasoning step by step often improves results on problems with several steps, and also lets you spot where it went wrong.
Privacy and responsibility
Anything you type may be stored or used according to the service's policy. Do not paste passwords, private documents or other people's personal data into a tool without understanding where it goes. For school and work, check the rules about disclosure. You remain responsible for what you submit under your name.
A good mental model
Think of a model as a very fast colleague who has read an enormous library, writes beautifully, never gets tired and sometimes confidently misremembers. You would not submit their work without reading it, and you would not argue with a good idea just because it came from them. Use it to think faster, not to stop thinking.