What Is an LLM? A Plain Explanation With a Real Test

A large language model, or LLM, is a computer program that writes by predicting the next word, again and again, using probabilities it learned from enormous amounts of text. ChatGPT, Claude and Gemini are all built on one. The model works out which word is most likely to come next, adds it and repeats, without looking anything up, which is why it can sound fluent and still be wrong. To show what that looks like with real numbers, this site asked a small open model for its next-word probabilities on 18 October 2026, and the results are below.

What Is an LLM?#

An LLM is a very large statistical model of text, trained to predict the next piece of a sentence from the pieces before it. According to IBM (2026), the text is first split into tokens, which can be whole words, parts of words or single characters, and the model then calculates a probability for every possible next token. Language models are one branch of the wider field covered in what AI, machine learning and deep learning mean, and the AI glossary defines tokens, parameters and the other terms used here.

The "large" refers to size. IBM (2026) notes that these models can hold billions or even trillions of weights, the adjustable numbers that training sets, and that most are built on the transformer, a design introduced in 2017 whose self-attention lets the model weigh different earlier tokens when it chooses the next one. The idea of guessing the next word is older than any of this. Microsoft's Azure guide to LLMs points out that earlier statistical language models, such as n-grams, already predicted which words were likely to come next from the words before them. Modern models do the same job with far more text and far more of the sentence in view.

What Does Predicting the Next Word Actually Look Like?#

Predicting the next word looks like a ranked list of candidates, each with a probability, and the top candidate is often far from certain. This site ran the prompt "The capital of France is" through a fine-tuned copy of Hermes 3, an 8-billion-parameter model built on Meta's Llama 3.1, using Ollama on the author's own computer on 18 October 2026, and asked for the ten most likely next tokens. The chart below shows the top six.

Paris came first, at 23.7 per cent. The next candidates were "a" at 16.6 per cent and "the" at 8.3 per cent, and the list carried on through "known", "home" and "one". That spread says less about geography than about grammar, since plenty of true sentences begin "The capital of France is a..." or "The capital of France is home to...". This is the part most explanations skip. Every answer a chatbot gives is the end of a long chain of choices like this one, and each choice is made from a spread of options, not from a single stored fact.

According to IBM (2026), the model generates text one token at a time, calculating the probabilities for all the possible next tokens at each step, and it does not know the final answer in advance. The France result makes that concrete. When the model wrote "Paris", it had not yet decided what the rest of the sentence would say.

Why Does the Answer Change With the Words Before It?#

The answer changes because every token the model has already seen reshapes the probabilities for the next one, which is what people mean when they say an LLM uses context. In a second prompt, this site gave the same model one example sentence first, "The capital of France is Paris. The capital of Italy is", and the probability of "Rome" rose to 95.8 per cent.

The example sentence told the model what kind of text this was, a list of capitals, and the vague options almost vanished. "The" fell to 0.2 per cent, against 8.3 per cent in the first test. According to IMD (2026), a model processes the surrounding context before it calculates the most likely next token, and this pair of results is that process made visible. It is also the practical reason a more specific prompt usually gets a more specific answer. A prompt that shows an example of the wanted format narrows the spread before the model writes a word.

Why Can the Same Question Get a Different Answer?#

The same question can get a different answer when randomness is switched on, because the model then draws its next word from the likely candidates instead of always taking the first. IBM (2026) lists temperature as one of the settings that control this, describing it as the control for how random the generated text is. The first two tests used temperature 0, which takes the top token every time. For the third, this site asked the same model to finish "I take my coffee with milk and" twenty times at temperature 1.

"Sugar" won 9 of the 20 runs. The other 11 produced eight different words, from "cream" to "love" to "no". At temperature 0 the same prompt gives "sugar" every time, because sugar had the highest probability, at 25.9 per cent. Twenty runs is a small sample, so the counts should not be read as exact rates. The pattern is the point. A long reply is hundreds of picks like this one, so two answers to the same question can drift apart within a sentence or two.

Why Does an LLM Sound Sure When It Is Wrong?#

An LLM sounds sure because a confident, well-formed sentence is very likely text, whether or not it is true. Nothing in the numbers above marks "Paris" as checked and "home" as guesswork. Both are candidates with probabilities, and the model's job is to keep the text likely. According to IMD (2026), these errors are grammatically coherent but factually incorrect, and they are most dangerous in fields such as healthcare, finance and law. IBM (2026) describes the same failure as information that is false or misleading while sounding plausible, and calls it a hallucination.

The site's own test of AI hallucination shows what this looks like in a real chatbot. Asked 26 questions about things that do not exist, Gemini mostly refused to invent anything, but on one question it described a full court ruling three times out of ten. A reader cannot tell those answers apart by their tone, which is why names, figures and citations from any chatbot need checking against a source.

What Can an LLM Do, and What Can It Not Do?#

An LLM can do almost any job that can be framed as "continue this text", and it cannot check facts or understand them the way a person does. Elastic's guide to LLMs notes that sentiment analysis, question answering, translation, summarisation and text generation can all be handled by one model, because all of them reduce to predicting text. That breadth is why one chatbot can draft an email and then fix a spreadsheet formula in the same session.

The limits come from the same design. According to Yanjun Gao, quoted by the University of Colorado Anschutz (2026), "LLMs can imitate aspects of human thought, but they lack the emotional and experiential depth that defines human intelligence." A sensible working rule follows from the tests above. Treat an LLM as a strong first draft for language tasks and a weak source for facts. The probabilities themselves come from training, which how AI learns explains step by step, and a model the size of the one tested here can run on an ordinary home computer, a choice weighed up in local AI versus cloud AI.

What Does "Most Statistically Probable" Mean, in Plain Words?#

It means the model has learned, from huge amounts of writing, how likely each word is to follow the words before it, and when it writes it keeps choosing likely next words. Put in plain words, an LLM produces the most typical way of saying something given everything it has read. That makes its writing fluent and often useful. It also means the output leans towards the common and the average rather than the original or personal, and that it can be confidently wrong, because a likely sentence is not always a true one.

Common questions#

The questions below are the ones beginners ask most once they know an LLM predicts words.

Sources#

• What are large language models (LLMs)? (IBM Think, 2026)
• What are LLMs? A beginner's guide to large language models (IMD, February 2026)
• What is a large language model? (University of Colorado Anschutz, February 2026)
• What is a large language model (LLM)? (Elastic)
• What are large language models (LLMs)? (Microsoft Azure)
• Hermes 3, Llama 3.1 8B model card (Nous Research, Hugging Face)
• Ollama, the tool used to run the test

Last reviewed: 18 October 2026. The test figures are fixed to the model and settings named above and will not change. The explainer sources are re-checked quarterly.