So what is an LLM, really?
A large language model (LLM) is, at heart, a computer program that takes in some text and produces more text. You type a question, a request, or the start of a story, and it writes a continuation that reads as if a knowledgeable person wrote it. That is the whole job description. Everything impressive an LLM does — answering questions, drafting emails, explaining code, translating — is this one ability, write plausible text, pointed at different tasks.
More precisely, an LLM is a kind of generative text model: generative because it creates new text rather than just classifying or scoring it, and text because words in and words out are all it deals with. It does not have a database of answers it looks things up in. It generates each response fresh, one piece at a time, based on patterns it absorbed from an enormous amount of writing.
You've already met its tiny cousin
When your phone keyboard suggests the next word as you type — you write "I'll be home in five" and it offers "minutes" — you are watching a baby version of the same idea. That is the heart of the autocomplete-to-LLM story: an LLM is autocomplete that has been scaled up almost beyond recognition, fed vastly more text, and given a far more powerful brain.
The leap from phone autocomplete to a useful assistant is enormous, but the kind of thing happening is the same: look at the text so far, and predict what comes next. Hold on to that picture. It is the single most useful intuition in this entire domain, and the next guide unpacks it fully.
The whole idea in one line: given the words so far, the model assigns a probability to every possible next word.
Why is it called "large"?
The word large points to two kinds of bigness. First, the model has a huge number of internal dials, called parameters — often tens or hundreds of billions of them. These numbers are what get adjusted during learning, and they store everything the model "knows" about how language works. Second, it learned from an immense pile of training data: a large fraction of the books, articles, and web pages humanity has put online.
Put those together — many parameters, lots of data — and you get a system that has, in effect, distilled patterns from a sizeable chunk of human writing into a single program. That scale is exactly why a plain "predict the next word" machine ends up able to do so much. It is a flavor of language model taken to an extreme, which is why people also just call it a large language model and study it as its own thing.
A log-log chart showing loss falling steadily as model size and data scale up.
What it feels like to use one
In practice you open a chat box, type in plain language, and get an answer in plain language. There is no special syntax to learn. You can ask follow-up questions, paste in a paragraph and ask for a summary, or say "explain this like I'm twelve." The model treats your whole message as text to continue, sensibly.
- You type a message — a question, a task, or some text to work on.
- The model reads everything so far, including the conversation's earlier turns.
- It writes a reply by repeatedly predicting the most fitting next piece of text.
- You read it, refine your request, and the loop continues as a conversation.
Where this track is heading
This first track stays at the level of clear intuition — no math, no jargon you can't picture. Over the next four guides we'll see the one core idea (predicting the next word), how a raw predictor becomes a polite chatbot, what is actually "inside" a model, and an honest accounting of what LLMs do well and where they fail. By the end you'll have a sturdy mental model that the rest of the domain — tokens, attention, training, prompting — builds on.