Large Language Model

A Large Language Model — LLM for short — is an AI system that has learned how language works from a very large amount of text. It doesn’t know rules in the classical sense; instead, it has absorbed statistical patterns from billions of sentences. Anyone who asks it a question gets an answer based on these patterns. The result often sounds as fluent as if a human were responding. Well-known LLMs include GPT-4 by OpenAI, Gemini by Google, and Llama by Meta.

Why LLMs mark a technical turning point

Before LLMs, AI systems had to be built individually for each task. One system for translations, another for summaries, yet another for programming code. LLMs can do all of that — with the same model, without any retooling.

The reason is sheer size. A modern LLM has billions of parameters — these are numerical values that the model adjusts during training in order to better predict language. Beyond a certain size, abilities emerge that no one explicitly programmed in: logical reasoning, simple arithmetic, analogical thinking. Researchers call this “emergent abilities” because they seem to appear out of nowhere.

This makes LLMs a tool that can hardly be confined to a single domain — and that is both their strength and a source of risk.

From text fragment to answer

An LLM works with small text building blocks called tokens — these are usually word parts or short words. When you enter a question, the model breaks it down into tokens and then predicts, step by step, which token is most likely to come next. This is how the answer is formed — one token after another, recalculated each time.

Training works differently: the model is given a text in which the last word is hidden, and it has to guess it. If it gets it wrong, the internal parameters are corrected slightly. This happens billions of times, on texts from the internet, from books, from scientific papers. In the end, the model has a fine sense of which words and ideas belong together.

Important: the model does not “understand” in the human sense. It optimizes probabilities. This explains why LLMs sometimes make confidently false statements — a phenomenon experts call hallucination.

LLMs in products and in reporting

ChatGPT, probably the best-known AI chatbot, is based on an LLM. So is Microsoft's Copilot, which is integrated into Word and Excel. When search engines nowadays display entire answer paragraphs instead of lists of links, there is almost always an LLM behind it.

In the finance and technology industries, the term frequently comes up when companies talk about automation, customer service bots, or internal knowledge search. Many companies further train an existing LLM on their own documents — this is called fine-tuning — so that the model can better answer domain-specific questions.

A common misconception in reporting: LLM and AI are often equated. However, LLMs are only one subtype of AI — albeit currently the most visible one. There are many AI systems that have nothing to do with language, such as image recognition systems or fraud detection algorithms.