Just in
TechJD Earnings: Q3 Net Income Rises 23% on 3.5% Revenue GrowthTechBank Outage Today: Chase, Bank of America, Wells Fargo Users HitTechBlackstone Amazon Data Center Deal: BXDC Takes 75% of 13 Sites for $5BTechNationwide Emergency Alert Test Set for Nov. 17 at 2:20 p.m. ETTechXPeng Earnings: Q3 Revenue Rises 19% to $3.62 BillionCryptoBitmine Ethereum Holdings Hit 6.08 Million ETH, Total Falls to $11.8B
AI · Guide

What Is a Large Language Model? A Plain English Guide

A large language model is an AI program that writes by predicting the next piece of text. Here is how tokens, training and attention work, and where LLMs go wrong.

USA Publishing Group6 min read
Share
Purple chat speech bubbles on a translucent screen, illustrating how a large language model holds a conversation
Photo: kuu akura / Unsplash

Key takeaways

  • A large language model breaks text into tokens, roughly three quarters of an English word each, and writes by predicting the next token over and over.
  • The word large refers to the huge amount of training text and the billions or even trillions of parameters the model adjusts while it learns.
  • Because LLMs produce statistically likely text rather than looking up verified facts, they can hallucinate, so specific facts and figures should always be checked.
In this story 6 sections
  1. How a large language model works, in simple terms
  2. What makes a language model "large"
  3. LLM vs AI: how the terms fit together
  4. What large language models are used for
  5. Limits and common myths
  6. How to use this knowledge

A large language model (LLM) is a very large AI program, trained on huge amounts of text, that writes by predicting which piece of text should come next.

That one sentence answers the question most people type into Google when they ask what is a large language model, but it hides a lot. The same trick powers chatbots, coding assistants and the AI summaries at the top of search results, and knowing how it works makes you better at using these tools.

How a large language model works, in simple terms

Think about the autocomplete on your phone keyboard. You type "See you" and it offers "tomorrow" or "soon." An LLM does the same thing at an enormous scale. Google's machine learning course describes LLMs as essentially autocomplete mechanisms that can predict thousands of tokens in a row, which is how a short question turns into several paragraphs of answer.

Three ideas explain most of what happens under the hood.

Tokens: the pieces of text a model reads

LLMs do not read whole words the way you do. They break text into tokens, which can be a full word, part of a word or a single character. According to Google's LLM introduction, one token in English is roughly four characters, or about three quarters of a word, so 400 tokens is around 300 words. That is why AI services often price their plans and set their limits in tokens rather than words.

Prediction: one token at a time

When you send a prompt, the model calculates how likely each possible next token is, picks one, adds it to the text and repeats. IBM's explainer calls LLMs giant statistical prediction machines that repeatedly predict the next word in a sequence. The model has no plan for the whole paragraph in advance. It builds the answer piece by piece, and each new token is shaped by everything before it.

Attention: working out what matters in a sentence

Most modern LLMs are built on a design called the transformer, introduced in a 2017 research paper titled "Attention Is All You Need." Its key feature, self attention, lets the model weigh how much each word in your text affects the meaning of every other word. Google's course uses a good example: in "The animal didn't cross the street because it was too tired," attention helps the model link "it" to the animal rather than the street.

What makes a language model "large"

The "large" refers to two things: the amount of training text and the number of parameters. Parameters are the internal numbers the model adjusts while it learns, a bit like the millions of tiny dials on a mixing desk. IBM notes that LLMs can have billions or even trillions of these weights, and points to GPT-3, released in 2020, which had 175 billion parameters.

AWS describes the training process as the model iteratively adjusting those values until it correctly predicts the next token, using sources such as Common Crawl and Wikipedia. Google's course says training takes months and enormous amounts of computing power and electricity, so only a few companies build the biggest models.

Training usually happens in two stages:

  • Pretraining, where the model reads a vast pile of text and learns to fill in or continue it, with no human labeling each example.
  • Fine tuning, where the pretrained model is shaped for a job, for instance answering questions politely, following instructions or writing code. IBM lists methods such as supervised fine tuning and reinforcement learning from human feedback, where people rate answers so the model learns what a helpful one looks like.

LLM vs AI: how the terms fit together

People often ask whether an LLM is the same thing as AI. It is not. AI is the broad field of making computers do tasks that normally need human judgment. Machine learning is a branch of AI where systems learn patterns from data. Deep learning is a type of machine learning that uses large neural networks. An LLM is one kind of deep learning model, specialized in language.

So every LLM is AI, but plenty of AI is not an LLM. The system that recognizes faces in your photo library or recommends your next show may use completely different models. A product like ChatGPT, which we cover in our guide to what ChatGPT is, is an app built on top of LLMs, with extra features such as memory, file uploads and web search layered on.

What large language models are used for

You probably run into LLMs more often than you realize. The sources above list the same core uses again and again:

  • Writing and rewriting: emails, cover letters, product descriptions, social posts.
  • Summarizing long documents, meeting notes or articles.
  • Translating between languages.
  • Answering questions in chatbots and customer support systems.
  • Writing and explaining computer code.
  • Sorting and classifying text, such as tagging reviews as positive or negative.

Tools that let an LLM take actions, like booking a meeting, are usually called AI agents, covered in AI agents explained.

Limits and common myths

Myth: the model looks up facts

A plain LLM does not search a database of verified facts. It produces text that is statistically likely. That is why both Google and IBM warn that LLMs hallucinate, meaning they generate answers that are false or misleading while sounding perfectly confident. Some products connect a model to outside sources (IBM describes this as retrieval augmented generation), which helps, but does not remove the problem.

Myth: it knows what happened this morning

A model learns from data collected up to a certain date. Unless the app adds live web search, it may not know about recent news, price changes or new product versions.

Myth: it is neutral

LLMs learn from text written by people, so they can pick up and repeat the biases in that text. IBM notes that they can reflect and even amplify those biases. NIST, the federal standards agency, published a generative AI risk profile in July 2024 to help organizations manage risks like these.

Limit: memory per conversation

Each model has a context window, the maximum number of tokens it can consider at once. Paste in a very long document and the model may lose track of the early parts, or the app may refuse the upload.

How to use this knowledge

Once you know an LLM is a prediction engine, a few habits follow naturally. Give it clear context, because it can only work with what is in the conversation; our guide to writing better AI prompts shows how. Treat any specific fact, number, quote or citation as a lead to check, not a final answer, and use the methods in our AI research and fact checking guide. And keep private details, like account numbers or medical records, out of chats unless you know how the service stores them.

A quick test before you trust an AI answer: ask yourself whether you could verify it in two minutes from an official source. If you can, do it. If you cannot, treat the answer as a draft. For more plain English explainers on chatbots and the models behind them, browse our AI section.