What Is an LLM? Large Language Models Explained in Plain English

AI is showing up in everything, from your web browser to software that previously seemed perfectly happy without it. Everything wants an AI summary now; every company wants AI in the name. Heck, I'm semi-surprised PEPS hasn't come out with a new AI flavor just hop on the bandwagon.

This is the first article in our series on understanding AI and finding practical uses for it. Before we start downloading software and running models on our own computers, let’s get comfortable with what is doing the work so we can make better decisions for how much of this new tech we want in our lives.

Like it or hate it, AI is here to stay, and with it comes a pile of abbreviations, product names, and marketing claims that can make learning the basics more confusing than it needs to be.

Let’s start with one you will hear constantly: LLM, short for large language model.

An LLM is a type of artificial intelligence trained on large amounts of text to learn patterns in language. Those patterns allow it to generate responses, explain ideas, summarize documents, translate languages, and write computer code.

That sounds useful. It is. However, understanding how it produces those answers also helps explain why it occasionally gives you instructions for a Windows setting that doesn't exist or advice you can't possibly follow.

What Does LLM Stand For?

LLM stands for large language model. Each word tells us something about what we are dealing with.

“Large” refers to the scale of the model, particularly its many learned numerical settings, called parameters. Training also involves substantial amounts of data and computing power. There is no magic number of parameters that officially makes a model large, and having more does not automatically make one better for every job.

“Language” includes ordinary writing, programming code, and other text. Some modern systems also handle images, audio, or video. You will see these described as multimodal, which means they work with more than one type of input or output.

“Model” is the trained mathematical system that processes your input and produces a response.

A point of note here is that the AI model you are using and the program you open to use that model are not necessarily the same thing.

Think of the model as the engine. The application provides the controls and the features that let you use that engine.

Where Do LLMs Get Their Training Data?

LLMs learn from large collections of material. Depending on the developer, that can include websites, books, articles, research papers, public discussions, and computer code. Some material is publicly accessible, some is licensed, and some is created specifically for training.

There are three broad sources:

• Publicly available material: Web pages, public code repositories, government documents, and other accessible content. Publicly accessible does not automatically mean public domain or free of copyright.
• Licensed and human-created material: Content obtained through agreements, along with examples, demonstrations, and feedback supplied by people working on the model.
• Synthetic data: Material generated by other AI models, such as sample questions, answers, or coding problems. Its usefulness depends on its quality and how it is checked.

Developers process these collections to remove duplicates, filter unwanted material, and prepare it for training. Those filters are imperfect, and developers do not all disclose their complete datasets. That makes it difficult to know exactly what an individual model encountered.

Your conversations may also be used to improve future models, depending on the service, account type, settings, and policies. That is separate from the model using what you just told it to answer your next question.

Also, uploading a document for an AI to summarize does not automatically mean that document becomes part of its permanent training. The application can supply it as temporary reference material. Whether the provider retains or later uses it is a separate policy question.

Is ChatGPT an LLM?

Services like ChatGPT are applications powered by AI models. The chat window, file uploads, browsing features, and other tools are parts of the application built around those models.

OpenAI develops the GPT (Generative Pre-trained Transformer) model family. ChatGPT is one way people use its models. Anthropic develops Claude models and offers the Claude assistant. Google uses Gemini as a name for both its model family and its assistant, which does not exactly help clear up the terminology… Frankly, I wish they had stuck with Bard.

You may also encounter model families such as Meta’s Llama, Alibaba’s Qwen, and models from Mistral AI and DeepSeek.

The names can wait. The useful distinction is that a developer builds a model, and software gives you a way to interact with it. Some applications let you choose between several models.

Changing the model can change the quality, speed, hardware requirements, and cost of the results, even when the chat window looks exactly the same.

What Can You Actually Use an LLM For?

For most people, the useful starting point is a task they already do.

You might ask an AI application to explain an error message, summarize a lengthy document, or turn that pile of rough notes into something another human can follow.

For example:

• Computer help: “Explain this Windows error in plain English. Start with steps that do not change anything, and tell me what information you need before suggesting a fix.”
• Understanding information: “Summarize this document, explain the unfamiliar terms, and point out any deadlines.”
• Writing and organization: “Clean up these notes, fix the spelling, and organize the ideas without changing my meaning or making me sound like a sales brochure.”

That last one is particularly useful to me. I generally know what I want to say. Getting it out of my head, through my fingers, and onto the page in a readable order can be another matter, espically being dyslexic. Spell check can only get you so far some days.

Having software help organize that mess is extremely useful.

How Are Large Language Models Made?

Developers do not manually program an answer to every possible question. Instead, they train a system to learn patterns from examples.

There are three broad stages.

First, developers prepare the training material. Depending on the model, this may include publicly available text, licensed material, computer code, human-written examples, and generated examples. Sources and methods vary between developers. The material needs processing to address issues such as duplication, formatting, and quality.

Next comes training. Text is broken into pieces called tokens. A token can be a word, part of a word, punctuation, or another small piece of text. The model processes these tokens and adjusts its parameters as it learns.

For generative models such as GPT, a central training task is predicting the next token.

The system makes a prediction, measures how far off it was, and adjusts its numerical settings. Repeat that across enormous amounts of text, and the model learns increasingly useful patterns involving grammar, facts, writing styles, code, and relationships between ideas.
So, for example, if I were to write this sentence

“She poured the coffee into a ___.”

An LLM assigns probabilities to possible next tokens. “Cup” or “mug” would likely fit better than “bicycle.”

Other language models use different training tasks, such as predicting missing text.

Finally, developers refine and evaluate the model. Additional examples, feedback, and training help it follow instructions and produce more useful responses. Testing examines its abilities, limitations, and behavior.

The result is a trained mathematical model. It does not need to keep an ordinary searchable copy of every page it encountered to generate an answer. However, models can memorize and reproduce some training material, so saying they never remember or repeat anything would also be wrong.

How Does an LLM Answer Your Question?

Your question or instruction is called a prompt.

Suppose you paste a Windows error message into an AI application and ask what it means. The application supplies the model with your prompt, relevant conversation context, and any other information its features provide.

The model processes that input and generates its answer one token at a time. Each new piece becomes part of the context used to produce what follows.

You will sometimes hear this described as “fancy autocomplete.” That is a useful introduction to next-token prediction, but it does not explain the full range of abilities that training can produce. The same underlying process can support a short email, a translation, computer code, or an explanation involving several steps.

There is an important catch: producing a likely answer does not establish that the answer is true.

Your Windows error might have several possible causes. The model may suggest a common explanation without enough evidence to know whether it applies to your computer. It can also combine details incorrectly and produce a troubleshooting step that sounds reasonable but is completely wrong.

The confident writing style is not a warranty.

Does an LLM Search the Internet?

Not automatically.

A model can generate answers using patterns learned during training and information supplied in its current context. That does not mean it just searched the web or checked the latest software documentation.

An application can connect the model to web search, uploaded documents, a calculator, or other tools. Those tools supply additional information or let the application perform actions.

This distinction explains why two applications using the same model can behave differently. One may have access to current documentation. Another may be working entirely from its training and your question.

When asking about something that changes frequently, such as software features or release information, check whether the application actually consulted a current source.
Then open the source and make sure it supports the answer.

What Was the First LLM?

Researching all this, I was surprised to find s no single universally accepted “first” LLM. Language models developed over decades, and what counted as “large” changed along with the available hardware.

A foundational milestone came in 2003, when Yoshua Bengio and colleagues published a neural language model that learned word representations while predicting the next word. It helped establish ideas used in later systems.

The phrase “large language models” also predates today’s chatbots by quite a margin. In 2007, Google researchers published Large Language Models in Machine Translation, describing statistical models trained on up to two trillion tokens. These used a different approach from the neural models powering modern AI assistants.

A major turning point arrived in 2017 with Attention Is All You Need, the research paper that introduced the transformer architecture. Attention helps a model account for relationships between different parts of its input. Transformers became the foundation for many modern LLMs.

OpenAI introduced its original GPT in 2018, now commonly called GPT-1. It was an important early example of the approach behind the GPT family, but calling it the first language model would skip a lot of work that came before it. But, for sure, it was the first with an actual chatbot associated with it, so I feel it's fair to call ChatGPT the first publicly useful LLM.

Why Do LLMs Make Things Up?

Think about a weather forecast. Meteorologists use observations, historical records, and computer models with hundreds of years of data to predict what the weather will do. Even with all that information, sometimes you end up mowing the lawn in the rain. In other words, prediction is not a guarantee. Nor is a prediction a fact.

There is a similar lesson with LLMs, although they are predicting something different. A weather model predicts atmospheric conditions. A language model predicts pieces of a response based on patterns learned during training and the information available in the conversation.

The problem is that a likely sequence of words can describe something that is completely wrong. A Windows troubleshooting step might sound right because it resembles thousands of other instructions, even though the menu option it tells you to click does not exist.

Invented facts, references, and other details are commonly called hallucinations. But they are really just bad guesses. They can happen when the model lacks information, relies on outdated material, misunderstands your question, or produces a plausible answer instead of recognizing that it does not know.

The problem with AI is that, unlike a forecast that gives you a 40 percent chance of rain, an AI answer often arrives as a confident statement from what people perceive as a trusted source. That confidence in its wording should not be confused with certainty about the facts.

The point being. Use it to help understand a problem and work through possible solutions. Check important details, especially before running commands, deleting files, or changing system settings or sending that letter to the boss about your quarterly numbers. The confident writing style is not a warranty.

Can You Run an LLM on Your Own Computer?

Yes, compatible models can run on your own hardware.

Cloud-based AI runs the model on a provider’s servers. Local AI runs it on your computer. A desktop application can use either approach, so installing a program does not automatically mean its AI processing stays on your PC.

Software such as Ollama and LM Studio can run compatible local models. LM Studio can also work offline once the necessary model files are downloaded. Both products offer additional connection options, so check which model and processing location you have selected.

Hardware requirements depend on the model and how it is stored and run. Smaller models can work on more modest systems. Larger ones may require considerably more memory and benefit from a capable graphics card. We will cover what those requirements mean before asking you to download anything enormous.

Local AI gives us a particularly interesting area to explore at MajorGeeks: software you can install, experiment with, and put to work on your own equipment. Pretty darn exciting. Also, running local models is a big step up in prvacy, but we’ll get to that later.

The Geek Verdict

LLM’s are going anywhere and understanding a few basics will give you the knowlegde to judge its usefulness in your life.

LLM’s learns patterns during training, uses those patterns and the context you provide to generate responses, and gains additional abilities through the applications and tools around it. It can save time, explain unfamiliar material, and help organize your work. It can also get things wrong while sounding remarkably pleased with its answer.

You do not need to become an AI researcher to use one. Start with a task you understand well enough to judge the result. That is a much better introduction than handing it something you cannot verify and hoping for the best.

Next, we will look at cloud AI versus local AI, why you might choose one over the other, and what you need to run a language model on your own PC.

Jim McMahon Jim McMahon, aka Corporal Punishment, is the founder of MajorGeeks.com. He has spent decades testing software, troubleshooting Windows, and helping users cut through the nonsense. He loves real freeware, hates bloatware, and runs on caffeine, sarcasm, and questionable choices.

comments powered by Disqus