What is an LLM?

Before you write a single line of code, it helps to know what you are actually talking to. A large language model is not a search engine and not a database. It is a next-token predictor that has read a very large amount of text. Once that clicks, most of its strange behaviour stops being surprising.

Overview

A large language model (LLM) is a neural network trained to answer one question over and over: given the text so far, what comes next? You give it a prompt, it predicts the most likely continuation one chunk at a time, and that stream of chunks is the answer you read.

Everything else — writing code, translating Telugu to English, summarising a contract — is the same prediction trick applied to text that happens to look like those tasks. The model learned the patterns by reading, not by being programmed with rules for each task.

Key ideas

Prediction, not retrieval

The model does not store your documents and look them up. It stores statistical patterns in its weights. When it "remembers" a fact, it is reconstructing something it saw many times during training. This is why it can be fluent and wrong at the same time — a plausible continuation is not always a true one. That failure mode is called hallucination.

Training happened in the past

A model has a knowledge cutoff: the date after which it saw no new data. It does not know today's news unless you give that information in the prompt. This is the single most common beginner confusion. The fix is to feed fresh context, which is exactly what retrieval-augmented generation does (you will build that in Week 3).

Bigger context, same idea

Models differ in how much text they can consider at once — the context window. A larger window lets you paste more background, but the core behaviour does not change. It is still predicting the next token, just with more to look at.

When to reach for an LLM

LLMs are strong at tasks with fuzzy inputs and language-shaped outputs: drafting, summarising, classifying, rewriting, extracting fields from messy text, and answering questions over context you supply. They are weak at exact arithmetic, anything needing live data, and tasks where a wrong answer is expensive and hard to verify. For those, pair the model with a tool (a calculator, a database, an API) rather than trusting it alone.

Quick recap

• An LLM predicts the next chunk of text; every task is that prediction in disguise.
• It reconstructs patterns from training data, so it can be fluent and wrong (hallucination).
• It has a knowledge cutoff and no live data unless you supply it in the prompt.
• Use it for language-shaped, fuzzy tasks; pair it with tools for exact or live work.
• The context window sets how much text it can consider at once, not how it thinks.