Large language models went from a research topic to something people use daily in an extraordinarily short time. Millions now ask them questions, have them draft emails, use them to write code and rely on them to explain things.
What almost nobody has been told is what the thing on the other side actually is. The gap between what people assume is happening and what is happening explains most of the confusion — why the model invents a plausible-sounding fact, why it can be argued out of a correct answer, why it is superb at some tasks and unreliable at others, and why “it made that up” is a description of normal operation rather than a malfunction.
This explanation involves no mathematics. It is about the shape of the thing.
What a language model does
A large language model does one thing: given some text, it predicts what text should come next.
That is genuinely the whole mechanism. You provide a prompt; the model produces the most likely continuation, one piece at a time, each new piece informed by everything before it.
It is easy to hear that and conclude it is trivial. It is not, because doing that prediction well at enormous scale turns out to require something that behaves a great deal like understanding. To reliably predict the next word in “the capital of France is”, you need to have absorbed something about France. To continue a paragraph of legal argument convincingly, you need to have absorbed a great deal about how legal arguments are structured.
But the objective remains prediction. The model is not consulting a database of facts. It is producing text that fits.
How it learns
Training involves showing the model enormous quantities of text — a large portion of the public internet, books, code, articles — and repeatedly asking it to predict hidden portions. Each time it is wrong, its internal parameters adjust slightly.
Do this billions of times across billions of parameters and patterns emerge: grammar, facts, reasoning structures, writing styles, the shape of a recipe, the conventions of an apology.
Fine-tuning then adapts the raw model toward being helpful, following instructions and declining harmful requests. Part of this typically involves human feedback, where people rate responses and the model is adjusted toward the preferred ones.
This second stage is why a model is conversational rather than simply continuing your sentence, and it is also where much of its characteristic tone comes from.
Tokens, not words
Models process tokens rather than words — chunks that may be a whole word, part of one, or a punctuation mark. “Understanding” might be two tokens; a common word is usually one.
This matters practically because pricing and limits are expressed in tokens, and because it explains certain oddities. A model may struggle to count letters in a word or reverse a string, since it never sees individual letters in the way you do — it sees chunks.
The context window
The context window is how much text the model can consider at once, including your prompt, any documents you supply and its own response so far.
Everything outside that window does not exist for the model. In a long conversation, earlier parts may fall out of view, which is why a model can appear to forget something you established twenty messages ago.
Context windows have grown enormously, and larger ones allow whole documents to be considered at once. But there is always a limit, and performance can degrade with material buried in the middle of very long inputs.
Why they get things confidently wrong
This is the most important thing to understand, and it follows directly from the mechanism.
The model generates text that is plausible, not text that is verified. It has no separate store of facts to check against and no internal signal distinguishing “I know this” from “this is the sort of thing that would appear here”.
So when asked about something obscure, it does not fail with an error. It produces the most likely-looking continuation — which may be a citation that does not exist, a statistic that was never measured, or a confident account of an event that did not happen. This is usually called hallucination, and it is not a bug being fixed but a consequence of how the thing works.
The tone makes it worse. The model writes fluently and confidently regardless of whether it is right, because fluent confident text is what its training material looks like. Human writers signal uncertainty; a model’s fluency carries no such signal.
Practical implication: treat output as a well-informed draft from someone who never says “I am not sure”. Verify anything that matters — names, numbers, dates, citations, legal and medical claims, and anything you will act on.
What they are genuinely good at
Being fair about the strengths, because they are real.
Drafting and rewriting. Producing a first version, adjusting tone, shortening, expanding.
Summarising material you supply, where the source is in front of it.
Explaining concepts at a chosen level of difficulty.
Translating and adapting between registers.
Code, particularly boilerplate, and explaining unfamiliar code.
Structuring — turning notes into an outline, generating options.
Brainstorming, where being wrong occasionally costs nothing.
Notice the pattern: they are strongest where the material is provided, where fluency is the point, or where errors are cheap and easy to spot.
What they are weak at
Facts they were not clearly trained on, especially anything local, obscure or recent.
Precise arithmetic, unless using a calculator tool.
Current events, unless connected to search.
Counting and character-level operations.
Knowing what they do not know.
Anything requiring genuine accountability.
That first point deserves emphasis for Nigerian users. Models are trained overwhelmingly on English-language material about Europe and North America. Ask about Nigerian regulations, local institutions, market prices or place-specific detail and the risk of confident invention rises sharply, because the training material is thinner. Verify local specifics with particular care.
Practical use
Give context. More relevant detail produces better output. Say who it is for, what tone you want, how long it should be.
Supply the source material rather than relying on recall. A model summarising a document you paste is far more reliable than one recalling what it thinks the document says.
Ask for reasoning on anything complex, which makes errors easier to spot.
Iterate. Treat the first response as a draft and refine it.
Verify anything consequential. Especially names, figures, citations and anything local.
Do not paste confidential data into a service without knowing how it handles it.
The short version
A large language model predicts what text comes next, trained on enormous quantities of writing, then tuned to be helpful. That single mechanism explains both its fluency and its failures.
It produces plausible text, not verified text, and it cannot distinguish between the two. Confidence in its tone tells you nothing about accuracy.
Use it for drafting, summarising, explaining and structuring, where it genuinely excels. Verify anything factual, and be especially careful with Nigerian or otherwise local specifics, where the training material is thinnest and the invention most confident.
AI writing tools have multiplied, and most do the same few things. Here is how to tell the categories apart, what each is genuinely useful for, and where they let you down.
AI video generators produce impressive clips and disappoint in predictable ways. Here is what they can genuinely do, where they fail, and what they cost to use seriously.
ChatGPT and Gemini are close enough that the honest answer depends on what you do. Here is where each one leads, and what matters if you are paying from Nigeria.
The iPhone App Store is full of AI productivity apps, most of which wrap the same models. Here is what genuinely helps, what to skip, and how to avoid paying twice.
AI tools can save a small business real time, or become another subscription nobody uses. Here is what genuinely helps, what to skip, and how to start without spending much.
