What Is a Large Language Model? Plain English Explained (2026) | LLM Basics Day 1

Welcome to LLM Basics — the five-day course that explains large language models the way I wish someone had explained them to me when I first started working in AI security. This zooms in on LLMs specifically. The LLM Hacking hub is where this course leads when you’re done. And the CEH practice exam has AI security questions you’ll be able to answer properly after just five days here.

What “LLM” Actually Stands For

LLM stands for Large Language Model. Three words. Let me explain each one.

“Large” — This means enormous. Like, unimaginably enormous. We’re talking about systems trained on more text than any human could read in thousands of lifetimes. The training data includes large chunks of the internet, millions of books, articles, code, conversations, and more. And the model itself has billions of numbers inside it storing everything it learned. “Large” is doing a lot of work in that name.

“Language” — This means it works with text. Words, sentences, paragraphs. It was trained on human language, and it produces human language. Not images. Not music. Not video (some systems do those things, but they’re different models). An LLM is specifically about text.

“Model” — This means it’s a mathematical representation. A model is a system that takes input and produces output based on patterns it learned during training. It’s software — a very specific kind of software that learned to work with language by studying examples.

Put it all together: a large language model is an enormous software system that learned how language works by reading billions of examples, and now it can produce language itself. That’s it. That’s the whole definition.

The One Idea That Explains Everything

Here’s the single most important thing I can tell you about LLMs. Everything else about them — why they’re useful, why they make mistakes, how people trick them — flows from this one idea:

That’s the whole mechanism. Predict the next word. Pick it. Repeat.

Let me make it super concrete. Say you ask an LLM: “What is the capital of France?” Here’s literally what happens:

It has your question. It starts generating a response. First word? Probably “The” — because most factual answers start with “The.” Second word? “capital” — because that’s what the question was about. Third word? “of” — because “capital of” is a very common phrase pattern. Fourth word? “France” — obviously. Fifth word? “is” — the sentence needs a verb. Sixth word? “Paris” — because billions of examples of this type of sentence in training data end with “Paris.” Done.

The result: “The capital of France is Paris.” It sounds like knowledge. It sounds like the AI knows the answer. But mechanically, it was just predicting the most probable next word at every step — and those predictions happened to be correct.

Here’s why this is the most important thing to understand: the LLM is always predicting. Even when there’s no right answer. Even when it doesn’t know. Even when the answer doesn’t exist. It still predicts the next most-probable word. And sometimes that word is wrong. Confidently, fluently wrong. This is why LLMs can make things up — a problem called “hallucination” that we cover in depth on Day 3.

Why LLMs Sound So Human

The most common question I get when explaining LLMs to beginners: “But it sounds so real! How can it just be predicting words?”

The answer is that human writing is actually very patterned. We don’t realise it because we’re inside the pattern, but from the outside, the way humans string words together follows very consistent rules. Grammar patterns. Topic patterns. Tone patterns. Conversation patterns. Response patterns. We learned them all as children by listening and reading.

An LLM learned those same patterns — but from billions of examples instead of a few thousand. It’s seen so many conversations, so many articles, so many explanations, that it learned every pattern humans use when they write. When it predicts the next word, it’s drawing on all those patterns simultaneously. The result is text that doesn’t just follow grammar correctly — it follows the subtle patterns of how a knowledgeable person would write about a specific topic in a specific style.

Think of it this way. You can tell a good chef’s cooking from a bad one, even if they used the same recipe. The good chef has cooked thousands of times and developed intuition about exactly how things should taste and feel. An LLM is like a chef who has “tasted” billions of pieces of writing and developed deep intuition about exactly how language should flow. The intuition is real. The billions of examples of human writing really are in there, compressed into billions of numbers.

But — and this is critical — the chef analogy has a limit. A good chef understands food. An LLM doesn’t understand language. It predicts language. There’s no understanding happening. There’s pattern matching at an enormous scale. When the patterns run out or get confused, the output stops sounding human. That’s why very specific questions about recent events or very unusual topics sometimes produce weird, wrong, or made-up answers.

ChatGPT vs Claude vs Gemini — What’s the Difference?

People use “ChatGPT” and “LLM” interchangeably, and that’s a bit like using “Google” and “the internet” interchangeably. One is a product built on top of the other.

Here’s the structure:

An LLM is the raw engine. It’s the trained model — the billions of numbers that represent everything it learned. You can’t really talk to a raw LLM directly. It’s like a car engine sitting on a bench. Powerful, but you need the rest of the car to do anything useful with it.

ChatGPT is a product built on an LLM. OpenAI trained an LLM called GPT-4 (and its successors). Then they built ChatGPT — the website you chat with — on top of that LLM. ChatGPT adds things like: the chat interface, memory of your conversation, safety filters, the system prompt that tells it to be helpful, and the infrastructure to serve millions of users. ChatGPT is the whole car. GPT-4 is the engine inside it.

Claude, Gemini, and Copilot work the same way. Claude is built on Anthropic’s Claude LLM. Gemini is built on Google’s Gemini LLM. Microsoft Copilot is built on GPT-4 (licensed from OpenAI). Each product wraps a different LLM engine with different instructions, interfaces, and safety systems.

Why does this matter? Because the same underlying LLM can power very different products depending on how it’s configured. A “helpful customer service chatbot” and a “creative writing assistant” might both use the same raw LLM — just with different instructions wrapped around it. Understanding this separation is key to understanding how LLMs can be manipulated through those instructions — something we explore properly in our Prompt Engineering course.

The Most Important Thing: It’s NOT Thinking

I want to be really clear about something before Day 1 ends. This is the thing most people get wrong, and it matters for everything else in this course.

An LLM is not thinking. It is not understanding. It has no idea what any of it means.

When ChatGPT writes you a paragraph about the French Revolution, it has not retrieved a memory of the French Revolution, thought about what’s important to say, and decided what to write. It has predicted the most probable sequence of words that follows your question, based on patterns in billions of training examples. The output looks like understanding. It is not.

Here’s my favourite analogy to make this real. Imagine a huge library with every book ever written on history. Now imagine someone who has read every single one of those books and can recite them perfectly — but who has never lived a single day in history themselves and has absolutely no experience of the world. Ask them a question about history and they can produce a perfect-sounding answer from memory. But they don’t actually know what any of it felt like. They can’t judge what’s important and what isn’t. They’re pattern-matching their memory, not reasoning.

That’s an LLM. The pattern matching is phenomenal. The depth of knowledge is astounding. But the understanding? Zero. None. Not happening.

Why does this matter? Because once you know this, the weirdness of LLMs starts to make sense. Why does it sometimes confidently make things up? Because it’s predicting probable words, not recalling true facts. Why can it be tricked with strange questions? Because tricky questions break the pattern-matching in ways that reasoning would handle. Why does it sometimes answer differently if you rephrase the same question? Because different words activate different patterns.

Why Understanding This Makes You Safer

You might be wondering why a cybersecurity website is teaching you LLM basics. Here’s my answer.

LLMs are everywhere now — and most people using them don’t understand them. That creates risks. Risks to you personally, and risks in the systems being built with them.

People who understand LLMs know: don’t trust it without checking. People who don’t understand LLMs often just believe whatever the confident-sounding text says. Hallucinations get shared as facts. Medical advice from a predicting machine gets followed without a doctor’s input. Legal advice gets acted on. Financial decisions get made.

From a security angle: I’ve seen phishing emails written by LLMs that were polished, personalised, and convincing — because LLMs are excellent at writing text that sounds legitimate. I’ve seen customer service chatbots manipulated into revealing information they weren’t supposed to share — because attackers understood the “predicting words” mechanism better than the chatbot’s designers did.

Understanding what an LLM is — really understanding it, not just knowing the buzzword — gives you a filter. When you read an LLM’s output, you know to ask: is this a reliable prediction or a made-up prediction? When you interact with a chatbot, you know what it can and can’t do. When someone tries to trick you using AI-generated content, you’re less likely to be fooled.

That’s why we’re here. Five days from now you’ll understand LLMs well enough to use them wisely and protect yourself from the ways they’re misused. Today was the foundation. Let’s keep building.