Large language models power chatbots, coding assistants and search. Here's how they actually work — what they're good at, where they fail, and the vocabulary you need.
A large language model — an LLM — is a program trained to predict the next chunk of text. That sounds almost too simple to explain ChatGPT, Claude or Gemini, yet prediction at enormous scale turns out to produce something that reads, reasons and writes. This guide covers how that works, the words you’ll keep hearing, and where these systems quietly break.
At its core, an LLM reads everything before a point and estimates the most likely text to come next. Train that objective on a vast slice of the internet, books and code, and the model has to internalise grammar, facts, tone and reasoning patterns just to predict well. Nobody hand-codes those rules; they emerge from the data.
The text is broken into tokens — roughly word fragments. “Unbelievable” might be three tokens. The model works in tokens, prices are quoted in tokens, and its memory limit is measured in tokens too.
Modern LLMs are built on an architecture called the transformer. Its key trick is attention: for every token, the model weighs how relevant every other token is. When it processes “the bank of the river,” attention lets it connect “bank” to “river” and rule out the financial meaning.
You don’t need the maths. The practical takeaway is that transformers handle long-range context well and scale efficiently on modern hardware — which is why they won.
The model reads a massive corpus and learns to predict text. This is where raw capability comes from, and it costs the most — millions of dollars in compute for a frontier model.
A pre-trained model is knowledgeable but unruly. A second stage, often using reinforcement learning from human feedback (RLHF), teaches it to follow instructions, be helpful, and refuse harmful requests. This is the difference between a model that completes text and an assistant that answers you.
The context window is how much text the model can consider at once — the prompt plus its own reply. Early models held a few thousand tokens; current ones hold hundreds of thousands. A bigger window means you can paste a whole contract or codebase and ask questions about it. But everything outside the window is invisible: the model has no memory of past chats unless that text is fed back in.
This is the part marketing tends to skip.
If you can’t verify an LLM’s answer, treat it as a confident first draft, not a source.
An LLM is a powerful, fluent, unreliable assistant. The skill is using it where fluency helps and verification is cheap — and staying sceptical where it can confidently mislead you. Understand the prediction at the centre of it, and most of the behaviour stops being mysterious.