What Is an LLM? A Plain-English Explanation With Real Examples

What is an LLM? A large language model – is a program that predicts the next chunk of text, one chunk at a time, based on statistical patterns it absorbed from enormous amounts of writing. It does not look anything up and it does not check whether it is right. Understanding those two facts explains almost every strange thing you have seen an AI assistant do, including the confident wrong answers.

Most beginner explanations of large language models tell you they are like a friend who has read the entire internet. That analogy is comforting and slightly misleading, because it implies recall. What is an LLM actually doing is closer to improvisation than to memory – and the difference is the reason it can produce a flawless paragraph and an invented statistic in the same breath.

So this explanation of what is an LLM stays in plain English, but it does not simplify to the point of being wrong. Along the way it corrects two things that beginner pages get wrong often enough to be worth naming directly: what parameter counts really tell you, and where the boundary sits between a large language model and the chatbot you actually type into.

What is an LLM, in one sentence?

An LLM is a system trained to predict what text comes next, at a scale large enough that the prediction becomes useful. That is the whole large language model meaning, stripped of marketing. It reads your input, converts it into numerical chunks, runs those chunks through an enormous set of learned weightings, and produces the chunk it calculates as most likely to follow. Then it does it again, using its own output as part of the new input, until it reaches a stopping point.

Here is an analogy that fits the mechanism better than the well-read-friend version. Think of a jazz musician who has spent decades absorbing recordings. Play them four bars and they will continue convincingly, in the right key, in a style that matches. They are not recalling a specific record – they are producing what fits. Ask them which album taught them that phrase and they cannot tell you, and if you press for an answer they may give you a plausible one that happens to be wrong. That is a large language model. Fluent continuation, no lookup, no source, no internal check.

The “large” in the name is doing real work. Small versions of this idea have existed for decades and produced nonsense. What changed was scale – more text, more computing power, and a design called the transformer, introduced in a 2017 Google research paper, that let models weigh which earlier words matter most when predicting the next one. That weighting mechanism, called attention, is why modern systems hold a thread across paragraphs instead of losing it after a sentence.

How do LLMs work? The prediction loop, step by step

If you want to know how LLMs work without the mathematics, it comes down to four stages. This is the part of what is an LLM that most explanations rush. Two happen once, before you ever touch the thing. Two happen every time you press enter.

• Training. The model processes enormous quantities of text and adjusts billions of internal values until its predictions stop improving. This takes months of computation and happens once per model version. Nothing you type changes it afterwards.
• Tokenizing. Your input is chopped into tokens – chunks that are usually shorter than words. This is the stage most people never hear about, and it explains more odd behaviour than anything else on this list.
• Predicting. The tokens run through the trained weights, and the model produces a probability distribution across possible next tokens. It picks one, appends it, and repeats.
• Stopping. It continues until it hits a stop signal or a length limit, which is why an answer can end mid-thought when a conversation gets long.

Notice what is absent. There is no database query. There is no fact-checking pass. There is no moment where the system compares its answer against a source of truth, because there is no source of truth inside it – only the statistical residue of what it read during training.

Why tokens explain the weird failures

Tokens are the single most useful concept for predicting when large language models will struggle, and almost no beginner page connects them to anything practical. A token is a fragment of text, not a letter and not reliably a word. The model never sees your sentence as a string of characters – it sees a sequence of these chunks.

That has three consequences you have probably run into:

• Letter-level tasks go wrong. Counting the letters in a word, reversing spelling, or working with individual characters is genuinely hard for a system that never sees characters. This is not stupidity; it is the input format.
• Some languages cost more. Text in languages less represented in the tokenizer breaks into more tokens for the same meaning, which means higher API costs and faster consumption of usage limits for identical work.
• Long documents get truncated. The context window – the amount of text a model can hold at once – is measured in tokens, not pages. Attach enough material and the earliest part silently falls out of view.

Does size matter? The parameter myth

Almost every answer to what is an LLM reaches for parameter counts early, and this is where beginner explanations most commonly mislead, so it is worth being direct. Parameters are the adjustable values inside the model, set during training. GPT-3 had 175 billion of them, a figure OpenAI published in its 2020 research paper on the model. That number became famous, and a rule of thumb grew around it: more parameters, better model.

That rule of thumb no longer holds, and pages still teaching it are teaching a 2020 understanding of a 2026 field. Training data quality, training method, and post-training refinement now matter as much as raw size. Smaller modern models routinely outperform much larger older ones on tasks a normal person cares about. A parameter count tells you roughly how expensive a model was to build. It does not tell you whether it will answer your question well.

One related caution, because you will meet the number: the figure of roughly 1.7 trillion parameters for GPT-4 circulates widely across beginner articles and is not something OpenAI has published. Treat it as unverified. A page that repeats it as fact is a page that did not check, which tells you something about the rest of its numbers.

LLM vs chatbot: what is actually different

The llm vs chatbot distinction confuses people because the words get used interchangeably in marketing, and they are not the same layer of the stack. The model is the engine. The chatbot is the car built around it.

The practical implication matters more than the definition. When an assistant cites a live web page or reads your uploaded document, that is not the large language model remembering things – it is the product around it fetching material and pasting it into the input before the model sees it. This is why the same underlying model can seem knowledgeable in one app and clueless in another. The wrapper is doing the work.

It also explains the memory question. A model has no recollection of your previous conversation. The chatbot re-sends the earlier messages every time, which is why long conversations get slower, cost more, and eventually start losing their earliest details.

What is an LLM in practice? Real examples

You are almost certainly already using large language models, often without a chat window involved:

• Assistants you type into. ChatGPT, Claude, Gemini and Copilot are the recognisable front doors. Google’s assistant was launched as Bard and renamed Gemini, which is why older articles referencing Bard are a quick signal that a page has not been updated.
• Features embedded in tools you already had. Email reply suggestions, document summarisation in office software, smart search in note apps, and the code completion in a developer’s editor are all the same underlying technology without the branding.
• Things built on top by other companies. Most AI writing tools, support chatbots and summarisers are products layered over a model from one of a handful of providers. Two competing tools frequently run the same engine underneath.

That last point is worth holding onto as a consumer. When you compare AI products, you are often comparing wrappers, not intelligence – and the wrapper is where the price difference usually lives.

What is an LLM bad at?

This is the section most explainers skip. Every page will tell you to verify important information. Almost none gives you a rule for when to stop trusting it in the first place. Here is the boundary that matters: an LLM is unreliable exactly where a plausible answer and a correct answer look identical to it.

• Anything requiring a source of truth. Legal citations, medical dosages, financial figures, historical dates. The model will produce something correctly shaped whether or not it is right, and it has no mechanism for telling the difference.
• Arithmetic you cannot check. Prediction is not calculation. Some products route maths to a real calculator; if yours doesn’t, treat the answer as a draft.
• Current facts, without retrieval. A model’s training has a cutoff. If the product isn’t actively fetching live information, recent events are outside what it can know, and it may answer anyway.
• Anything where being confidently wrong is expensive. This is the general form of the rule. Where the cost of a fluent error is high, use the model to draft and a human or a real source to verify.

None of that makes the technology unreliable overall. It makes it a specific kind of tool – excellent at producing and transforming language, structurally incapable of knowing whether what it produced is true.

What to take away

If you remember three things from this explanation of what is an LLM, make them these. It predicts rather than retrieves, which is why it invents sources so convincingly. It sees tokens rather than letters, which is why it miscounts characters and why some languages cost more to process. And the model is not the product, which is why two assistants running the same engine can behave completely differently. Everything else about large language models follows from those three facts.