A large language model (LLM) works by predicting what piece of text is most likely to come next, one small chunk at a time, using patterns it absorbed from an enormous amount of writing during training. By default it does not look facts up, it does not keep a list of true statements, and it does not understand the world the way a person does. That single mechanism explains most of what these tools are good at and most of the mistakes they make.
If you have ever watched your phone keyboard suggest the next word, you have seen a tiny cousin of the same idea. The difference is scale: a model that has read far more text than any human could, and that can keep a long conversation in view while it writes. Below we open the box in plain language, without maths, so you can decide when to trust the output and when to check it.
What is a large language model, really?
Strip away the marketing and an LLM is a very large mathematical function. You give it a sequence of text; it gives back a list of possible next pieces, each with a probability. The software picks one, adds it to the text, and asks the function again. Repeat a few hundred times and you have a paragraph. Every answer you have ever read from a chatbot was built this way, piece by piece, left to right.
The pieces are called tokens. A token can be a whole common English word, part of a longer word, a punctuation mark, or a fragment of a Vietnamese syllable. How text is cut into tokens depends on each model's design. In many models, Vietnamese text is split into more tokens than an English text of the same meaning, which can make Vietnamese conversations use up the model's working memory a little faster. It is a detail, but it explains why very long Vietnamese documents sometimes get cut off or summarised less carefully.
Where does the model’s knowledge come from?
Where do the probabilities come from? From training. During training, the model is shown huge amounts of text and repeatedly asked to guess the next token. Each time it guesses wrong, its internal numbers, called parameters, are nudged slightly so the right answer becomes a little more likely next time. After an enormous number of these nudges, the parameters hold a compressed picture of how language tends to flow: grammar, common facts, styles of writing, typical arguments, even patterns of code.
A model fresh out of that stage is good at continuing text but not at being helpful. So developers add a second stage, often called fine-tuning, where the model learns from examples of good answers and from people rating its replies. This is what turns a raw text predictor into an assistant that follows instructions, refuses some requests, and answers in a polite tone. It also shapes habits you notice, like the tendency to agree with you or to wrap up every answer with a neat summary.
Notice what is missing from this picture: there is no step where the model checks a fact against reality. Some products add that step from the outside, by letting the model search the web or read your files before answering. That helps, but it is a feature bolted onto the product, not something the model does by nature.
How does the model turn your question into an answer?
Your message is cut into tokens. The question, any earlier messages in the chat, and hidden instructions from the app are all converted into one long sequence of tokens before the model sees anything.
Everything must fit in the context window. The context window is the maximum amount of text the model can look at in one go. Paste in more than that and something gets dropped, usually the oldest part of the conversation.
The model looks at all tokens at once. A mechanism called attention lets each part of the text weigh which earlier parts matter most. That is how it links “it” in your last sentence to the contract you mentioned ten lines ago.
It scores every possible next token. The output is not a single answer but a ranking of thousands of candidates, from very likely to almost impossible.
The app samples from that ranking. A setting often called temperature controls how adventurous the choice is. Low temperature picks the safest token; higher temperature allows less obvious choices, which is why the same question can get different answers.
The chosen token is added and the loop repeats. The model never plans the whole paragraph in advance in a fixed way. Each new token is chosen with the text so far in view, so an early wrong turn tends to be carried forward.
Tools may step in between turns. In products with search or file reading, the app may fetch documents, paste them into the context, and then let the model write. The quality of the answer then depends heavily on what was fetched.
Nothing is remembered after the chat ends, unless the product saves it. The model itself does not learn from your conversation in real time. Any “memory” feature is the app storing notes and feeding them back in later.
Why does it sound so sure even when it is wrong?
Fluency and accuracy are two different skills, and the model was trained mostly on the first. A sentence that reads smoothly, uses the right jargon, and has a confident rhythm is exactly the kind of text it has seen millions of times. Whether the content is true is a separate question that the prediction process does not ask.
Take a simple office example. Ask for the article of a regulation that covers late payment, and the model knows what such an answer usually looks like: a document type, a number, a short quote. It can produce all three in the right shape even when it has no reliable memory of the real one. The result looks authoritative precisely because the shape is perfect.
Training for helpfulness can make this worse. People rating answers often preferred replies that were complete and decisive over replies that said “I am not sure”. Developers now work hard to reward honest uncertainty, and newer assistants hedge more often, but the pull toward a tidy, confident answer has not disappeared.
The practical lesson is to judge each claim by how checkable it is, not by how confident it sounds. Explanations of general ideas are usually fine. Specific names, numbers, dates, quotes and legal references deserve a second source every time.
One more trap: the model is shaped to be agreeable. If you push back on a correct answer, it may apologise and switch to a wrong one. Treat a quick change of mind as a warning sign, not as proof that you were right.
What LLMs are good at, and where they fall short
Good: rewriting and changing tone. Turning a blunt note into a polite email, or a long report into five bullet points, plays directly to its strength with language patterns.
Good: first drafts and outlines. It gets you past the blank page quickly. The draft is rarely final, but editing is faster than starting from nothing.
Good: explaining common concepts. Well-documented ideas such as compound interest, how a VPN works, or what a balance sheet shows are usually explained clearly, because the training data covers them many times.
Good: translation and bilingual work. It handles everyday English–Vietnamese translation well and can adjust register, though legal and medical wording still needs a qualified human.
Weak: precise facts and fresh news. Its knowledge stops at a training cut-off date and is blurry for rare details. Without a search tool, anything recent or very specific is risky.
Weak: exact counting and long calculations. Because it works in tokens rather than digits and letters, it can miscount characters or slip in arithmetic. Use a calculator or spreadsheet for numbers that matter.
Weak: local and niche knowledge. Information about a small Vietnamese company, a ward after the 2025 administrative reorganisation, or a niche regulation may be thin or outdated in its training data.
Risky: decisions with real consequences. Medical, legal, tax and financial choices need a professional. The model can help you prepare questions, but it should not be the one deciding.
Frequently asked questions
Does an LLM search the internet when it answers?
Not by itself. The model answers from patterns learned in training; only products that add a search tool fetch live pages, and they usually show sources.
Does a chatbot learn from my conversation?
The model does not update itself while you chat. However, some providers may use conversations to train future versions unless you turn that off, so check the data settings.
Why do I get different answers to the same question?
The app samples from a ranking of possible next tokens instead of always taking the top one, so wording and sometimes content change. If answers on a factual point keep changing, that is a sign the model is unsure.
Is a bigger model always more accurate?
Bigger models usually handle language and reasoning better, but they can still invent details confidently. Accuracy on a specific task depends on training data, tools and how you ask, so test with your own real examples.
Is it fine to write in Vietnamese with AI tools?
Yes, major assistants handle Vietnamese reasonably well for everyday tasks. Local facts, slang and very recent Vietnamese information are weaker spots, and long Vietnamese documents may hit length limits sooner.
Does an LLM actually understand what it says?
It captures a lot of structure about language and the world, enough to reason through many problems, but it has no experience and no built-in way to check truth. Treat it as a powerful text tool, not as a mind.
