What Are LLMs, Really? A Plain-English Explainer

Everyone talks about LLMs, but few explain what they actually are without drowning you in math. Here is the plain-English version.

A really, really good next-word predictor

At its core, a large language model does one thing: it predicts the next word (technically, the next "token") in a sequence. You give it some text, and it asks, "based on everything I have seen, what word most likely comes next?" Then it does that again, and again, building a response one piece at a time.

That sounds simple — almost too simple to explain ChatGPT. But when you scale that prediction up across trillions of words of training data and billions of internal parameters, something remarkable happens: the model learns grammar, facts, reasoning patterns, and even style, all as a side effect of getting good at prediction.

What "training" actually means

The model reads an enormous amount of text and adjusts billions of internal numbers (parameters) so its predictions get more accurate. It is not memorising the internet word for word — it is learning the patterns inside language. That is why it can write about things it never saw exactly, by combining patterns it did learn.

Why it "hallucinates"

Because the model predicts plausible text, not verified truth, it will sometimes produce something that sounds completely right but is wrong. It is not lying — it has no concept of lying. It is generating the most likely-sounding continuation, and sometimes the most likely-sounding answer is simply false. This is why you always verify anything important an LLM tells you.

Context windows: the model's short-term memory

The "context window" is how much text the model can consider at once. Early models held a few pages; modern ones like Gemini 3 Pro hold the equivalent of entire books. A bigger window means the model can keep track of more of your conversation or document before it starts forgetting the start.

Tokens, not words

Models read in tokens — chunks that are often part of a word. "Building" might be one token; an unusual word might be three. Pricing and limits are measured in tokens, which is why a long document costs more to process.

What LLMs are good and bad at

Why this matters to you

Understanding that an LLM is a prediction engine — not a search engine and not a truth machine — changes how you use it. You stop expecting perfect facts and start using it as a fast, tireless first draft that you then check and shape. That mental model is the difference between people who get great results from AI and people who get burned by it.