What actually happens when ChatGPT responds? A look under the bonnet of modern AI

Improving formulations, translating texts, checking complicated documents – AI tools do all that nowadays in seconds. Fascinating, but how does it actually work? The good news: You don’t need a degree in computer science to grasp the basic idea.

Chatbot, LLM, Transformer: What’s what?

When people talk about ChatGPT, Claude or Gemini, many terms get mixed up. Which is the product, which is the technology? A chatbot is primarily the application, the interface through which you interact with the AI. Behind it works what’s known as a Large Language Model, or LLM for short: a language model that has been trained on vast amounts of text and has learned to understand and generate language.

And the Transformer? That’s the architecture on which these language models are built, a blueprint developed by Google in 2017 that virtually all major language models use today. If the chatbot is the car and the LLM is the engine, then the Transformer is the design drawing of this engine. And it’s precisely this blueprint that is the focus of the new course Understanding Transformer-based Large Language Models on the AI Campus.
 

It all started with a clever idea

The story of the Transformer begins with a very specific challenge: How do you teach a computer to turn one sequence into another? For example, an English sentence into a German one, or a question into an appropriate answer.

Until then, common systems worked with so-called Recurrent Neural Networks. These function a bit like the game of Chinese whispers: Each word is processed one after the other, and the accumulated information is passed on to the next station. For short sentences, this works well. But as with whispers, the longer the chain, the more gets lost along the way.

A research team at Google then had an elegant idea: What if the model no longer passes words along one by one, but can process all words simultaneously? This idea gave rise to the Transformer, and the associated paper was titled with what soon became a catchphrase: “Attention Is All You Need.”
 

Attention, or: How words give each other meaning

“Attention” is translated as “Aufmerksamkeit” in German. And that’s exactly how you should imagine the mechanism. When you read the sentence “Ich sitze auf einer alten Bank im sonnigen Garten” (“I am sitting on an old bench in the sunny garden”), you immediately know that “Bank” here means a bench to sit on. The meaning is revealed through the context: The words “alten”, “sonnigen” and “Garten” make it clear what is meant.

That’s exactly what the attention mechanism does: For each word in the sentence, it calculates how much attention it should pay to every other word. The result is a much richer picture of what a word means in its respective context. The crucial advantage: All this happens in parallel, not one after the other. The Transformer doesn’t read word by word, like a slow reader with their finger on the page, it grasps the whole sentence at once, like a quick glance at a page.

This attention, this trick of processing words simultaneously while weighing contextual relationships, is the core of the Transformer architecture. It is the one great idea on which most LLM products are based, from GPT-5 to DeepSeek and Claude.
 

From specialist trick to universal genius

What was developed for transforming sequences of words turned out to be astonishingly versatile. Researchers found that the Transformer approach could not just translate, but also summarise, answer questions, write texts and even programme – as long as it was supplied with enough data and built to a sufficient scale. That’s how the Large Language Models that are changing our everyday lives today came about. The core architecture remained essentially the same. What changed: the scaling, the training methods, and clever optimisations to the attention mechanism, like the ability to use multiple “attention heads” at once to focus on different aspects of a sentence.

It’s a bit like inventing a surprisingly powerful engine in 2017, and since then, all manufacturers have competed to build the best vehicles from it. Some optimise fuel consumption, some focus on top speed, others on everyday usability. But the engine is basically the same.
 

LLMs simply explained on the AI Campus

This is exactly where the new course Understanding Transformer-based Large Language Models comes in. In compact video units, we explain the central building blocks to you, from positional encoding through self-attention to the block structure – without maths or programming. The course is aimed at everyone who wants to not just use LLMs, but also understand them. Because those who know the engine drive more safely.