12 min · up to 60 XP

What is an LLM, really?

The one mental model that explains everything AI does well, and everything it gets wrong.

It's 4:55pm on a Tuesday. Priya, a marketing manager, has a board summary due at 5. She types "summarise this report in 5 bullet points" into an AI assistant, pastes the report, and eight seconds later has something better than she'd have written in an hour.

Her first thought: how does it do that? Her second, a week later, after it invented a statistic in front of her CFO: what IS this thing? Below is a working model of the machine. You're the engine now.

Priya turns a messy report into a clean summary with help from an AI assistant. — Priya gets the useful answer first. Now the lesson asks what kind of machine produced it.

A model chooses the next token from several weighted candidate tokens. — The machine builds the answer one step at a time, by choosing what is most likely to come next.

The token machine

Each round, the model weighs up candidate next words (the bars are its confidence). The temperature dial controls how it chooses: Precise always takes the favourite, Wild gambles on long shots. Generate the whole sentence, then run it again on a different temperature and watch the same machine change personality.

Temperature

The best time to send a marketing email is

What the model is weighing up next

“mid-morning”

50%

“Tuesday”

28%

“3am”

12%

“pancakes”

10%

So what did you just operate? LLM stands for Large Language Model: the machine behind ChatGPT, Claude, Gemini and Copilot. It has read a colossal amount of text (books, articles, websites, code) and learned one skill to superhuman level: predicting what comes next.

You already know a tiny version of this. Your phone keyboard suggests your next word. An LLM is that idea scaled up a billion-fold, so far that prediction starts to look like reasoning, writing, and conversation.

Training material becomes internal patterns, while a database sits disconnected from the reply path. — Training teaches patterns. At reply time, the model completes from those patterns; it is not checking a live fact store.

Key insight

An LLM doesn't know things the way you do. It has absorbed the patterns of everything it read. When the pattern matches reality, it looks like knowledge. When it doesn't, it looks like a very confident lie.

Four versions of a prompt add more context and produce increasingly focused outputs. — Same model, better starting material. The output improves because the pattern it is completing becomes more specific.

Play with it

Priya needs an email improved. Drag the slider to give the AI more and more context, and watch what the same model does with each version of her request.

Drag the slider through every setting to see the effect.

What Priya types“Improve this”

Usefulness

12%

Happy to help! Could you share the text you'd like me to improve, and let me know what kind of improvement you're after?

Nothing in, nothing out. The model isn't being difficult; there is literally no text for it to complete against.

Why is a "next-word predictor" so capable? Because to predict text well, the model had to internalise the structures inside text: grammar, logic, arguments, the shape of a good executive summary. And it completes what you started, which is why your input is the biggest lever you have.

One more thing to experience. A real working session isn't one perfect question. It's a conversation. It's the day before the board meeting. This time you're Priya: you choose every message she sends.

Human facts flow into the model and come out as a structured board summary. — You bring reality. The model brings shape. Without facts, it can only imitate the shape of a good answer.

You drive the conversation

Steer Priya through prepping her board summary. There's no single right path, but watch how differently the machine behaves depending on what you feed it.

Happy to help with the board summary. Do you want to paste in the Q3 marketing results, or shall I sketch a general structure first?

A good model asks for what it can't see. Remember: it has no access to Priya's files, inbox or dashboards.

Your move: what do you say?

Only cards inside the context window connect to the model; outside files are out of reach. — When the model replies, the context window is the room. Everything outside it is not available unless you put it there.

Tap the diagram

Last check before the machine is yours. When the model writes its next reply to Priya, tap everything it can actually see.

There's more than one. Tap all of them.

Found 0 of 3

Tip

Try this today: take one email or summary you'd normally write by hand and run the context slider experiment for real: ask lazily first, then again with full context and constraints. The gap between the two outputs will teach you more than any article.

Five connected stations summarise prediction, patterns, context, facts, and the context window. — An LLM predicts from patterns. Your context shapes the prediction. Your facts keep the answer grounded.

Key takeaways

An LLM is a prediction engine: you built sentences from weighted next-word guesses, and temperature is just how boldly it gambles.
There is no fact database inside: 'knowledge' is patterns absorbed from training text.
Output quality tracks input quality. You slid the same model from useless to send-ready.
In conversation, you supply facts, it supplies form, and “what did you assume?” is your power move.
At reply time the model sees only its context window. You tapped it: system prompt, earlier chat, your message. Training data isn't live input.