What Is a Large Language Model (LLM)? Why It Can Be Confidently Wrong

Confident and wrong: the morning after the final#

It's the morning after a cricket final, and Arjun's flatmates are arguing about the result. Arjun opens a chat assistant and asks two things. First: "Write a funny message inviting my friends to watch the replay tonight." It writes a brilliant one in seconds.

Then he asks: "Who won last night's final, and what did the captain score?" It answers at once, sure of itself: "India won by 6 wickets; the captain made 82 off 64 balls." Meera opens the real scorecard. Wrong winner. And the captain made 31, not 82.

Same assistant, same confident tone. One answer brilliant, one made up. So here's the puzzle this lesson solves: if it got the score wrong, how did it write such a good invite? What actually is a large language model? By the end you'll be able to say what an LLM is in one sentence, and you'll know which questions to trust it with and which to check.

Your phone keyboard, grown huge#

You already use a tiny version of this idea every day. Type "Happy" on your phone and the keyboard suggests "birthday". It didn't look anything up. It has simply seen "birthday" follow "Happy" thousands of times, so it offers the word that usually comes next.

The chat assistant Arjun used works on the same idea, grown huge. Instead of suggesting one word, it writes a whole reply, one likely piece after another. Underneath, it's a neural network: a model made of a very large number of adjustable numbers, the kind the previous lesson introduced.

How an LLM works, in three steps#

1. It reads. During training, the model is shown a huge amount of text: books, websites, code. For a sense of scale, Meta (the company behind Facebook) reports that its Llama 3 models were trained on over 15 trillion tokens (a token is a small piece of a word; the next lesson covers them). Assume each token is about three-quarters of a word, as in English text: then fifteen trillion tokens is about eleven trillion words, and at about a hundred thousand words a book, that's about a hundred million books. And the reading stops on a set date, its cutoff, so the model never sees anything newer.
2. It learns patterns, not pages. It doesn't keep pages it can look up later. Instead it adjusts billions of internal numbers, like tiny dials that decide which word it guesses next: each time it guesses wrong, training nudges the dials a little. (It can still memorise some text it saw very often, but there's no store of pages to search.) That's your phone keyboard again, at enormous scale.
3. It writes by continuing. Given your message, it picks a likely next piece, adds it to the reply, and repeats until the reply is done.

Notice what's missing from that loop. Nothing in it looks up a fact or checks one. To be fair, this is a simplification: some chat apps add a web search on top of the model and paste the results into its input. The model itself doesn't search.

Put the three steps together: a model trained on a huge amount of text that writes by predicting what comes next. That's a large language model, or LLM. "Large" because it learns from a huge amount of text and has billions of internal numbers. "Language model" because what it models is which words tend to go together.

Arjun's four questions, scored#

Let's score the assistant on Arjun's whole morning. He asked it four things. In this example, here's what it wrote and how each answer held up.

One cricket word: LBW, leg before wicket, is one way a batter gets out (roughly, their leg stops a ball that would have hit the stumps). And "won by 6 wickets" is simply a winning margin.

Walk through the third question the way the model does. After "Who won last night's final?", what usually comes next in the text it learned from? A team name, then "won by", then a number, then "wickets". So it writes "India won by 6 wickets". It sounds exactly like a real result because it's built from the shape of thousands of real results. It just isn't this result.

Now group the four. The two right answers are writing and explaining: tasks full of patterns the model saw again and again. The two wrong answers are specific facts it never read: training stops at a cutoff date, so last night's match came too late, its winner and its captain's score alike. That's the whole lesson in one table.

The catch: it's not a search engine#

Most people think an LLM is a smart search engine, or a giant database of facts: you ask, it finds the answer. But it stores no list of facts and runs no lookup. It writes the most likely-sounding continuation, so a fluent, confident answer can still be made up.

Pause and guess. Arjun asks it "Write a two-line birthday wish for my nani" and "What is my nani's date of birth?" Which answer should he trust, and why?

The wish. Birthday wishes are a pattern it has seen countless times. Nani's date of birth is a specific fact it never read, so any date it gives is a guess, however confident it sounds.

The interview answer, and what to remember#

If an interviewer asks "What is a large language model, and why can it be confidently wrong?", a strong answer is short: an LLM is a neural network trained on a huge amount of text to predict what comes next, and it writes a reply by repeating that prediction piece by piece. It learns patterns rather than storing a database of facts, and it doesn't look anything up, so for rare or recent facts it produces a plausible guess with the same confidence as a correct answer.

• Trust it with: drafting, rewording, summarising text you give it, explaining well-known ideas.
• Check it on: recent events, exact numbers, names, dates and anything about one specific person or place.

The sentence to remember: An LLM doesn't look things up; it writes what's most likely to come next, so trust its writing and check its facts. Next, we zoom in on the small pieces it reads and writes, called tokens, which explain a lot of an LLM's strange behaviour.

Practice

Check your understanding

FAQ

Frequently asked questions

Sources

Where the facts come from

• Google for Developers, Machine Learning Crash Course: Introduction to Large Language Models (what a language model is)
• OpenAI, Why language models hallucinate (September 2025)
• Meta AI, Introducing Meta Llama 3