LLM

LLM stands for “Large Language Model.” This refers to a computer program that has read an enormous amount of text and derived patterns of language from it. It can therefore write sentences itself, answer questions, or summarize texts. The core of the matter is simpler than many suspect: the program only ever predicts which word fits best next. It repeats this prediction word by word until an entire answer has been formed. “Large” refers to two things here: the amount of text read and the number of internal adjustable settings the program tunes while learning.

Why language models suddenly changed everything

In the past, a computer needed a separate program written for every task. One for translations, another for spell-checking. An LLM covers many of these tasks with a single system. You simply describe the task in ordinary language, and the model tries to solve it.

This has economic consequences that are reported on daily in the news. Companies like OpenAI, Google, Anthropic, and Meta are investing billions in ever-larger models. Manufacturers of specialized chips also benefit from this, because training these models requires huge data centers. Entire stock prices now hinge on the question of how quickly this technology continues to develop.

At the same time, there is an important caveat. An LLM predicts probable words; it does not check facts. That’s why it sometimes invents information that sounds convincing but is false. Experts call this hallucination. Anyone using an LLM for homework or research should always double-check important information.

From a mountain of text to a finished answer

It all starts with training. The model is presented with vast amounts of text from books, websites, and forums. In each piece of text, a word is hidden, and the model has to guess which one is missing. At first it guesses randomly, but after every mistake its internal numerical values are slightly adjusted. These values are called parameters; large models have hundreds of billions of them.

It’s important to understand how the model actually processes text. It breaks sentences down into small building blocks called tokens. A token is often a word, sometimes just a syllable. Each building block is translated into a long series of numbers, because a computer can only calculate with numbers. The underlying architecture is called a transformer. Its trick is that every word can simultaneously look at every other word in the text.

After the basic training comes fine-tuning. Humans evaluate the model’s answers and show which ones are helpful and which are inappropriate. Only through this does a pure word machine become a useful assistant. A comparison helps: basic training is like reading an entire library, fine-tuning is like an etiquette course afterward.

LLMs in everyday school life, on the phone, and in business news

The best known are chatbots like ChatGPT, Gemini, or Claude. Behind each of these are LLMs from different providers. But you also encounter the technology without noticing it. Search engines summarize results, email programs suggest phrasing, translation services deliver smoother texts than before.

In software development, LLMs are especially widespread. Coding assistants write code suggestions as someone types. In companies, they answer customer inquiries or search internal documents. Smaller models now even run directly on smartphones, without a connection to a data center.

In financial news, the term usually appears in connection with costs and competition. There, the questions are things like: How expensive is it to operate a model per request? Who owns the necessary chips? And what rules does the legislature set, such as the EU’s AI Act? Anyone who understands what an LLM is can make much better sense of such reports.