The Glossary

An algorithm is like a recipe: a precise step-by-step guide that tells a computer how to solve a problem. In AI, algorithms are especially complex and can learn from data instead of just following rigid commands.

API (Application Programming Interface) is like a waiter in a restaurant: you order something, the waiter takes it to the kitchen and delivers the result back. APIs enable different programs to communicate with each other and exchange data – like when your weather app accesses data from a weather service.

AI is an umbrella term for computer programs that can perform tasks that normally require human intelligence – such as language understanding, pattern recognition, or decision-making. Modern AI systems "learn" from data instead of just following pre-programmed rules.

Bias (Voreingenommenheit) occurs when an AI reproduces certain prejudices due to its training data – for example, stereotypes regarding gender, origin, or professions. Developers are working with various methods to minimize this and create fair AI systems.

Chain of Thought (CoT) is a technique where you ask the AI to speak its thought process out loud. This often leads to better results on complex tasks because the AI works through the problem-solving process "aloud" instead of guessing the answer directly.

A chatbot is a computer program that can communicate with humans in natural language – written or spoken. Modern AI chatbots understand context, can answer complex questions, and even compose creative texts.

Deep Learning is an advanced form of machine learning that uses artificial neural networks with many layers. These "deep" networks can recognize extremely complex patterns – in images, speech, or text. Deep Learning is the basis for today's breakthroughs like self-driving cars or voice assistants.

Data processing describes all steps through which raw data is transformed into useful information – from collection and cleaning through analysis to visualization. High-quality data processing is crucial for AI systems, because "garbage in, garbage out": bad data leads to bad results.

Embeddings are a method of representing words, sentences, or entire documents as numerical vectors. Similar terms end up close together in the "vector space." This technique enables AI systems to understand semantic relationships – such as recognizing that "king" relates to "man" as "queen" relates to "woman."

Explainable AI deals with making AI decisions comprehensible to humans. Instead of a "black box" that delivers results without explanation, XAI shows which factors led to a decision. This is especially important in sensitive areas like medicine or finance.

Fine-tuning is the process of adapting an already pre-trained AI model (like GPT) to a specific task or domain. Instead of training a model from scratch, you take an "all-round genius" and teach it special skills – such as medical expertise or a specific writing style.

Foundation Models are large AI models that have been pre-trained on massive amounts of data and serve as the basis for many different applications. GPT-4, Claude, or Llama are Foundation Models – they can write texts, answer questions, translate, and much more, without being retrained for each task.

Generative AI can create new, original content – texts, images, music, code, or videos. Unlike analytical AI, which only evaluates data, generative AI is creative. Well-known examples are ChatGPT for texts, DALL-E for images, and GitHub Copilot for programming code.

GPT stands for "Generative Pre-trained Transformer" – an AI architecture specialized in understanding and generating human-like text. GPT models are first pre-trained on massive amounts of text and can then be fine-tuned for various tasks. ChatGPT is based on this technology.

Hallucination in AI means that the system outputs convincing-sounding but false or fabricated information. An AI might, for example, invent sources, make up facts, or imagine things that aren't true. That's why human verification is important.

Human-in-the-Loop (HITL) describes an approach where human experts are involved in the AI workflow – such as to label data, validate decisions, or intervene when uncertain. This combination of AI efficiency and human judgment often leads to the best results.

In-Context Learning is the ability of large language models to learn from examples in the prompt without the model needing to be retrained. Show the AI a few examples of your desired style or format, and it will imitate them – a powerful technique for precise results.

Iteration means improving a result step by step, rather than insisting on perfect results on the first try. When using AI, this is crucial: start with a first draft, give feedback, have the AI improve – repeat this until the result fits.

The Context Window determines how much text an AI can process at once – measured in tokens. A larger window means the AI can analyze longer documents or keep longer conversations in memory. Modern models have windows from 4,000 to over 100,000 tokens.

A Large Language Model is an AI system that has been trained on massive amounts of text to understand and generate natural language. GPT-4, Claude, and Llama are LLMs. They can write texts, translate, summarize, answer questions, and generate code.

Machine Learning is a subfield of AI where computers learn from data, recognize patterns, and make decisions – without being explicitly programmed for every situation. The more data, the better the models become.

An AI model is the result of the training process: a file (or several) that contains the learned patterns and parameters. You can think of a model as the "brain" of the AI that can be applied to new data after training.

Multimodal AI systems can process and understand different types of data simultaneously – text, images, audio, and video. A multimodal model could, for example, analyze a photo and write text about it, or describe a video.

NLP is a subfield of AI that deals with making human language understandable for computers. NLP enables machines to read, understand, interpret, and generate texts – the basis for chatbots, translation services, and voice assistants.

A neural network is a computer model loosely inspired by the human brain. It consists of interconnected "neurons" (nodes) organized in layers. Through training, the connection strengths adapt so that the network can recognize complex patterns.

The prompt is the input you give to an AI – a question, a command, or a task. The quality of the prompt has enormous influence on the result. Good prompts are clear, specific, and contain context.

Prompt Engineering is the ability to formulate precise and effective instructions to AI systems. Through skillful prompts – with context, examples, and clear instructions – you can dramatically improve the quality of AI outputs.

Reasoning refers to an AI's ability to think logically, analyze problems, and draw conclusions. Advanced models can perform complex thought processes, solve mathematics, or understand causal relationships – not just repeat patterns.

Retrieval means the targeted retrieval of relevant information from a knowledge database. In RAG (Retrieval-Augmented Generation), the AI first searches documents for relevant context and then generates its response – thus remaining factually correct and up-to-date.

System instructions are hidden instructions that tell the AI how it should behave – for example: "Be polite," "Answer briefly," or "Think step by step." They are transmitted to the AI before your actual question and shape the model's behavior.

Synthetic Data is artificially generated data created by AI systems to supplement or replace real data. This is especially useful when real data is scarce, expensive to obtain, or privacy-sensitive.

AI models don't understand words, but small text units called tokens. "Dog" might be one token, but "supermarket" might be two: "super" + "market." The number of tokens determines how much the AI can process at once and affects costs.

The Transformer is an AI architecture introduced in 2017 that revolutionized modern language models. It uses an "attention mechanism" to capture relationships between all words in a sentence simultaneously – the basis for GPT, BERT, and others.

Zero-Shot means that an AI model can solve a task for which it was not explicitly trained. A language model could, for example, translate texts into a new language or understand a new task based solely on its general training – without having seen specific examples for it.