Guides / Concepts explained

What Is an LLM? Large Language Models Explained Simply

An LLM is an AI model trained on huge amounts of text to predict and generate language. How it works, tokens and context windows, limits, and what to ask for.

Mythex Team · 2026-09-29 · 5 min read

An LLM, or large language model, is a type of AI trained on enormous amounts of text so that it can understand and generate language. Given some text, it predicts what should come next, one small piece at a time, and that simple ability turns out to be enough to answer questions, write, summarise, translate and write code. The models behind ChatGPT, Claude and Gemini are all LLMs.

Why it matters when you build with AI

If you build with an AI app builder, you're working with an LLM twice over. First, an LLM is doing the building: it reads your request, plans changes and writes the code. Second, you may want an LLM inside your app — a chatbot, a summary button, a writing assistant.

Understanding the basics helps in both cases. It explains why the AI sometimes misunderstands a vague request, why it can confidently produce something wrong, why long conversations can make it lose track of earlier details, and why adding AI features to your app costs money per use. You don't need the maths — just a sound mental model.

An everyday analogy

Imagine someone who has read an unimaginably large library — books, websites, manuals, code, forums — and has become extraordinarily good at continuing any piece of writing in a way that fits. Give them the start of an email and they'll finish it in your tone. Ask a question and they'll write the kind of answer that usually follows such a question.

But they have no notes to check, and the library closed on a certain date. They're recalling patterns, not looking things up. Most of the time the patterns are right. Sometimes they fill a gap with something that merely sounds right, with exactly the same confidence.

That's an LLM: a very well-read, very fluent writer with an imperfect memory and no built-in fact-checker.

How an LLM works, in plain terms

1. Training. The model is shown vast amounts of text and learns, by adjusting billions of internal numbers (its parameters), to predict the next piece of text. The design most LLMs use is called a transformer, introduced by Google researchers in 2017.
2. Tuning. The raw model is then refined to follow instructions, hold a conversation and avoid harmful output, often using human feedback.
3. Using it (inference). When you send a prompt, the model reads it and generates a reply one token at a time, each time choosing a likely next token given everything before it.

It doesn't have a database of facts inside it, and it doesn't keep learning from your chat. Everything it "knows" in a conversation is either from training or from what's in front of it right now — its context window.

A worked example: a support reply

Say you run an online plant shop and paste this into an LLM:

A customer writes: "My order #4412 arrived with a broken pot. What can you do?" Our policy: we replace damaged items free within 14 days if the customer sends a photo. Write a friendly reply.

Here's what happens:

• The text is split into tokens, maybe a hundred or so.
• The model reads the whole thing: the customer's message, your policy and your instruction.
• It generates a reply token by token: "Hi! I'm so sorry your pot arrived broken…" and asks for a photo, mentioning the free replacement.

Now try the same request without the policy. The model still writes a confident, friendly reply — but it might promise a refund, or a 30-day window, because those are common patterns in support emails it has seen. That's a hallucination, and it shows the most practical lesson about LLMs: they're only as accurate as the information you give them. Apps that need reliable answers feed the model the right source material each time, a technique called RAG.

Key terms explained

Common mistakes and misconceptions

• "It looks things up." A plain LLM doesn't. Some apps add web search or document retrieval, but the model itself is generating from patterns.
• "If it sounds sure, it's right." Confidence and accuracy are unrelated. Check names, numbers, dates, prices and anything legal or medical.
• "It remembers everything I said." Only what fits in the context window. In very long conversations, early details can drop out or get less attention. Restate what matters.
• "It's thinking like a person." It can reason through problems impressively, but it has no goals, beliefs or awareness of consequences. Treat its output as a draft to review.
• "Bigger or newer is always better." For a simple, high-volume task, a smaller, cheaper, faster model is often the right choice.
• Sending private data without checking. Before you paste customer data into any AI service or build it into an app, check the provider's data-use and retention terms.

What to ask your AI builder for

When the LLM is building your app:

• "Before you change anything, tell me what you understood and what you'll do."
• "Here's the exact policy text; use it word for word rather than inventing details."
• "Don't guess at facts like prices or dates. Leave a placeholder and list them for me."

When you want an LLM inside your app:

• "Add a [summarise] button that calls [provider]'s API from the server. Keep the API key in a secret, never in the browser."
• "Put our policies in the prompt so answers stick to them, and say 'I don't know' when unsure."
• "Limit how long answers can be, and cap how many requests one user can make per day."
• "Log what each request cost so I can see usage."

For a step-by-step walkthrough, see how to add AI features to your app.

LLMs and Mythex

Mythex's building agent runs on an LLM. Everyone gets the same Standard agent — there's no model picker — and building uses Mythex credits. AI features inside the app you build are separate: you bring your own API key from a provider such as OpenAI or Anthropic, store it as a project secret, and the provider bills you for that usage directly. The docs recipe add AI to your app shows the prompt to use.