How Large Language Models Work, in Plain English

People talk about large language models as if they were either magic or a trick. They are neither. The real explanation is simpler than the magic and more interesting than the trick, and you do not need any mathematics to follow it. This is the plain-English version.

The single mechanism that made modern models possible is attention, the technique that lets the network weigh how much each word in a passage should influence every other word.

It is also background for the bigger argument I make about how AI is changing human language.

One sentence: it predicts the next word

A large language model is a system trained to do one thing: given some text, predict what comes next. That is the entire core skill. Show it “the capital of France is” and it predicts “Paris,” not because it looked anything up, but because that continuation is overwhelmingly common in the text it learned from.

Everything else, answering questions, writing code, holding a conversation, is that same next-word prediction, repeated one chunk at a time, with each new word fed back in to predict the next.

How it learns

The model is trained by being shown enormous amounts of human writing with the next word hidden, guessing it, and being corrected. Repeat that billions of times and the model is forced to build an internal representation of how language works: grammar, facts, styles, the shape of an argument, the format of an answer.

Nobody programs in the rules of grammar or a database of facts. They fall out of the single goal of predicting text well. This is the surprising result that the 2017 transformer paper set in motion: a simple objective, at a large enough scale, produces general-seeming competence. The engine that makes it work is the transformer architecture, which lets the model weigh how every word in a passage relates to every other rather than reading strictly left to right.

Why it is useful

Because so much of human knowledge is written down, a system that has absorbed the structure of that writing can be genuinely helpful:

• It can rephrase, summarise, and translate, because those are transformations of text it has seen done.
• It can answer many questions, because the answers are statistically encoded in how the relevant words relate.
• It can follow instructions, because it has seen countless examples of instructions being followed.

Why it breaks

The same design explains the failures, which are not bugs so much as the flip side of the method:

1. It makes things up. When the most plausible next words are not the true ones, the model produces them anyway, with the same confidence. This is usually called hallucination.
2. It can be out of date. A base model knows only what was in its training data, frozen at a point in time, unless it is connected to live tools.
3. It has no built-in sense of certainty. Fluent and wrong looks exactly like fluent and right, because only fluency was trained directly.

Each strength and each failure is the same mechanism seen from a different side:

Behavior	Why it happens	What to do about it
Helpful answers	Knowledge is statistically encoded in how words relate	Lean on it for the shape of an answer
Hallucination	A plausible continuation beats a true one when they differ	Verify any fact that matters
Out of date	Training data is frozen at a point in time	Connect live tools for current information
No sense of certainty	Only fluency was trained, not truth	Remember fluent-and-wrong looks like fluent-and-right

The practical rule that falls out of this is the one I keep coming back to: trust a model on the shape of language, verify it on facts about the world. For the deeper version of why that split exists, see do large language models actually understand language.

What this means

Once you see that a language model is a next-word predictor that learned the structure of human writing, the mystery dissolves without the usefulness going away. It is not a mind and not a search engine. It is a model of language itself, which is exactly why it belongs in a conversation about where language is going.

That conversation is the subject of my book, Building Your First Brain, which is free for the first 1,000 readers and assumes no technical background.

Further reading

• “Attention Is All You Need”, the architecture behind today’s models.
• For where this fits in the long arc of communication, the evolution of language from speech to code.

Frequently asked questions

Does a language model search the internet for answers?

Not by default. A base model generates text from patterns learned during training; it has no live access to the web unless it is explicitly connected to a search tool. That is why it can answer instantly, and also why it can be confidently out of date.

Why do language models make things up?

Because their only job is to produce plausible next words, not to check facts. When the most plausible continuation is not the true one, the model still produces it, fluently. Fluency and accuracy are separate properties, and only fluency was trained directly.

Do language models think?

Not in the way people do. They have no goals, memory of you between sessions by default, or experience. They are very good at modelling the structure of language, which can look like thinking, and the question of whether that counts is genuinely debated.