An LLM (large language model) is an AI model trained on huge amounts of text to generate text. It is what sits behind ChatGPT, Claude or Gemini. It works by scoring the odds of each possible next token, picking one, adding it and repeating. That is how it writes answers, summaries or code.
The cat climbed onto the roof
6 tokensA question from the app. Pick an answer.
How does a model that predicts the next token generate text?
It predicts the next token, adds it and predicts again, until it is done.
In this model, each generated token helps predict the next. The app decides how to display the text.
It generates tokens; it does not just pick answers from a list. It may repeat text from its data.
LLM stands for large language model. "Language" because it works with text. "Model" because it is a trained computational system, not a database of answers. "Large" because it has a huge number of parameters and was trained on a huge amount of text. There is no official number that makes a model "large": it is a label, not a measurement.
If you give it "The cat climbed onto the", the LLM computes probabilities for what comes next. With example numbers: roof 45%, wall 30%, sofa 20%, theorem 1%. It picks one according to its settings, adds it to the text and computes again with the longer sentence. A three-paragraph answer is hundreds of those rounds.
To predict, it does not look only at the last word: it looks at everything in view, your message and what it already wrote, up to a limit called the context window. That is why the same question, with more context, usually gets a better answer.
First comes pretraining: the model reads a huge amount of text and practices guessing the next token. Each miss nudges its parameters a little, billions of times over. Then comes fine tuning: with example conversations and people rating answers, it learns to follow instructions and reply like an assistant. All of that happens before you use it: chatting with it does not change the model, and what it "remembers" from the chat is in its context window.
GPT, from OpenAI, is the one ChatGPT uses. Claude is from Anthropic, Gemini is from Google and Llama is from Meta, which publishes its weights so anyone can run it as a local model. In Brazil, the startup Maritaca AI makes Sabiá, trained with a focus on Portuguese. Note: ChatGPT is the app and GPT is the model; the same app can switch models.
Generative AI is the umbrella: any AI that creates content, whether text, images, audio or video. The LLM is the part that works with text (and code). An image generator is generative AI but not an LLM. Many models today mix both: they read photos and write text.
An LLM does not look the answer up online unless the app gives it that tool: it writes what usually comes next based on what it learned. A sentence can sound perfect and be false. Ask it for the last train time in your city and it may give you one with full confidence without having that data. That is called a hallucination.
Many tokenizers were built with more English text than text in other languages. That is why the same sentence in Spanish or Portuguese often splits into more tokens, like Spanish "subió" or Portuguese "telhado" (climbed, roof) cut in two. It matters in two places: APIs charge per token, almost always in US dollars, and the context window fills up sooner.
"LLM and AI are the same thing." An LLM is one kind of AI; spam filters and face recognition are AI too, and they are not LLMs. "It has the answers stored." No: it generates them token by token every time. "If it gets something wrong, it is broken." Making things up comes from how it works, not from a fault.