AI terms and
platform language.

Application

Adoption

Adoption describes how quickly and strongly a new technology is actually used. An AI tool can be technically perfect, but if nobody uses it, it has no value. Adoption depends heavily on training and trust.

Research

AGI

Artificial General Intelligence

AGI refers to a hypothetical AI that could perform all mental tasks as well as a human: learning, thinking, planning, adapting. AGI does not exist today. When or whether it will arrive is disputed among researchers.

Research

AI Alignment

AI Alignment is the research into building AI systems so that they do what humans really want, not just what is written in the code. As AI grows more powerful, alignment becomes one of the most important safety questions.

Application

AI Assistant

An AI assistant is a program that helps you with tasks: writing texts, scheduling appointments, answering questions, or checking code. Voice assistants like Siri or Alexa, but also ChatGPT or Claude, are part of this.

Ethics

AI Governance

AI governance is the set of rules with which organizations use AI responsibly: policies, responsibilities, review processes, risk assessment. It ensures that AI is not used randomly or unsafely.

Tools

AI Platform

An AI platform is an online service that gives companies or developers access to AI models, tools and infrastructure. Instead of training models yourself, you use ready-made building blocks. Examples are OpenAI, Anthropic, Google AI or Hugging Face.

Ethics

AI Safety

AI safety covers all measures that make AI systems reliable and harmless, from protection against misuse to avoiding hallucinations. As the power of AI grows, AI safety becomes ever more important.

Application

AI Strategy

An AI strategy defines where a company wants to use AI and why. It sets goals, budget, responsibilities and order. Without a strategy, many isolated projects often emerge without any impact.

Fundamentals

Algorithm

An algorithm is a step-by-step set of instructions for the computer on how to solve a problem. Just as a cooking recipe has steps, an algorithm has precisely defined instructions. AI uses many such sets of instructions to carry out tasks.

Application

API

Application Programming Interface (Schnittstelle)

An API is a technical interface through which two programs talk to each other. Using the OpenAI API, for example, a company can integrate ChatGPT into its own tools without programming from scratch.

Fundamentals

Artificial Intelligence

Artificial Intelligence is the ability of computers to solve tasks that normally require human thinking, such as recognizing images, understanding language, or finding patterns. It learns from examples instead of following fixed instructions.

Fundamentals

Artificial Neural Network

A neural network is a computer program inspired by the human brain. It has many small computing units that pass signals along, similar to nerve cells. Through training it learns to recognize patterns and make decisions.

Language models

Attention Mechanism

The attention mechanism lets a model pay special attention to certain words while reading, for example the subject of a sentence. This way it understands relationships even over long distances. This invention was the key to modern language models.

Application

Automation

Automation means having computers handle repetitive work. With AI, automation becomes more intelligent: it can also take on tasks where people used to have to decide, such as pre-sorting customer inquiries.

Machine learning

Backpropagation

Backpropagation is the process by which a neural network learns: after each error, the error is sent back through the network, and the internal values are adjusted. This way the network gets better step by step.

Ethics

Bias

Bias (Verzerrung)

Bias means the AI systematically makes unfair decisions, for example because its training data was one-sided. The result: certain groups are disadvantaged. Recognizing and correcting bias is a core task of responsible AI development.

Application

Change Management

Change management is the art of leading change in organizations so that people come along. Particularly important with AI, because it changes the way people work. It is about communication, training and taking concerns seriously.

Language models

Chatbot

A chatbot is a program that communicates with people via text message or speech. Modern chatbots use large language models and often feel like a real conversation partner. There are simple bots for standard questions and smart ones for complex topics.

Tools

ChatGPT

ChatGPT is an AI chatbot from the company OpenAI, launched at the end of 2022. It was the first AI assistant to become widely known. ChatGPT can write texts, answer questions, program and much more, in many languages.

Machine learning

Classification

Classification means: an AI sorts inputs into categories. For example, it decides between 'spam' and 'not spam', or between 'apple', 'pear' and 'banana'. This is one of the most common tasks for AI.

Tools

Claude

Claude is an AI chatbot from the company Anthropic. It was developed with a special focus on safety and well-considered answers. Claude can process long texts, argue carefully and can be used in many languages.

Tools

Closed Source AI

Closed Source AI / Proprietary AI

Closed Source AI is AI whose code and model belong to the company that develops it. Outsiders can use it, but cannot look into the code or run the model themselves. This protects the business model, but limits transparency.

Machine learning

Clustering

In clustering, the AI groups similar things together without anyone saying in advance which groups exist. This automatically creates clusters, that is, accumulations of similar data points.

Computer Vision

Computer Vision

Computer Vision is the entire research field that teaches computers to understand images and videos. It includes image recognition, object recognition, face recognition and image generation. A core field of AI for decades.

Language models

Context Window

The context window is the amount of text a language model can read at once. It is measured in tokens. If the window is exceeded, the model forgets the beginning. Modern models have very large windows.

Computer Vision

Convolutional Neural Network

CNN, Faltungs-Netz

A CNN is a special kind of neural network for images. It looks at images piece by piece and gradually recognizes first lines, then shapes, then whole objects. For a long time it was the most important model for Computer Vision.

Tools

Copilot (Microsoft)

Microsoft Copilot

Copilot is Microsoft's AI assistant that is built into Word, Excel, Outlook, Teams and Windows. It helps with writing, analyzing data or preparing meetings. In the background, GPT models from OpenAI often run.

Ethics

Copyright and AI

Copyright and AI is about two questions: May AI be trained on copyrighted material? And who owns what an AI produces? Neither question is yet clearly regulated by law everywhere.

Tools

DALL-E

DALL-E is OpenAI's AI tool for image generation, built into ChatGPT. You enter a text description and DALL-E creates the matching image. The name is a blend of the painter Dalí and the Pixar robot WALL-E.

Fundamentals

Data

Data is all the information a computer can store and process: text, numbers, images, sounds, videos. AI needs data in order to learn. The more and the better the data, the better it can solve tasks.

Application

Data Pipeline

A data pipeline is the journey of data from the source to the analysis: collecting, cleaning, storing, processing. Without clean pipelines, no AI works reliably. It is the infrastructure behind every AI solution.

Ethics

Data Protection

Data Protection / Privacy

Data protection means safeguarding people's personal data against unwanted storage, sharing or analysis. AI systems often process a lot of data and must strictly follow data protection rules. In Europe this is governed by the DSGVO (GDPR).

Fundamentals

Dataset

A dataset is an organized collection of data used to train or test an AI. It can consist of images, texts, tables or sounds. The quality of the dataset strongly determines how well the AI works later on.

Fundamentals

Deep Learning

Deep Learning is a form of machine learning that uses especially deep, multi-layered neural networks. This allows an AI to recognize complex patterns, such as speech or images. Almost all well-known AI systems today use deep learning.

Ethics

Deepfake

Deepfakes are AI-faked videos, images, or audio that look deceptively real, for example a person saying something they never said. They are a growing risk for politics, media, and personal security.

Computer Vision

Diffusion Model

A diffusion model is an AI that develops an image step by step out of pure noise. During training it learned to remove noise, and can thereby generate new images. Stable Diffusion works this way.

Application

Digital Transformation

Digital Transformation (Digitale Transformation)

Digital transformation is the fundamental overhaul of a company using digital tools, not just swapping out technology, but reshaping processes, culture, and business models. AI is an important part of it.

Fundamentals

Discriminative AI

Discriminative AI does not create anything new, but distinguishes between categories. It sorts or rates inputs, that is, decides what something belongs to. This is the older form of AI and remains important in many areas.

Research

Edge AI

Edge AI means the AI runs directly on the device, smartphone, camera, sensor, instead of in a data center. The advantage: it works without the internet, is faster and more privacy friendly, because the data never leaves the device.

Language models

Embedding

An embedding is the translation of a word or text into a long list of numbers. These numbers capture the meaning. Similar words have similar embeddings. This lets the computer compute with meaning, for example to compare or search texts.

Ethics

Explainability

Explainability / XAI

Explainability means an AI can say why it made a particular decision. With simple models this is easy, with large neural networks it is hard. Researchers are working on methods to make AI decisions understandable.

Computer Vision

Facial Recognition

Facial recognition is a special form of image recognition. It detects individual faces and can even match them to a person. This has many applications, but it also raises strong data protection questions.

Ethics

Fairness

Fairness in the AI context means the model treats different groups equally fairly. What counts as fair depends on the use case and is often contested. There are several mathematical definitions that cannot all be satisfied at the same time.

Fundamentals

Feature

Feature (Merkmal)

A feature is a single characteristic the AI pays attention to, for example color, size, word length or pitch. The AI uses many features together to make a decision. Some are chosen by people, some the AI finds on its own.

Research

Federated Learning

Federated Learning is a method in which many devices train a model together without sharing their data. Each device learns locally; only the learning progress is merged centrally. This protects privacy.

Research

Few-Shot Learning

Few-Shot Learning means the AI learns a new task from just a few examples, often only three to five. These are provided in the prompt. Modern language models can do this surprisingly well without being retrained.

Research

Fine-Tuning

Fine-Tuning is the additional training of a finished model for a specific task or style. It needs less data than the basic training. This is how large general models are adapted to special applications.

Research

Foundation Model

A Foundation Model is a large, generally trained AI model that serves as a basis for many different applications. GPT, Claude and Llama are Foundation Models. They are often adapted for special purposes via Fine-Tuning.

Ethics

GDPR and AI

The GDPR is the European data protection law from 2018. Among other things, it requires AI systems to provide clear information, consent, purpose limitation and the right to erasure. Anyone using AI in Europe must follow these rules.

Tools

Gemini

Gemini (vormals Bard)

Gemini is Google's family of AI models, integrated into Google products like Search, Docs, or Mail. Gemini can understand text, images, video, and audio, so it is called a multimodal AI.

Fundamentals

Generative AI

Generative AI creates new content on its own: text, images, music, videos, or code. It has learned from many examples and combines what it learned into something new. ChatGPT and Midjourney are part of this.

Machine learning

Gradient Descent

Gradient Descent (Gradientenverfahren)

Gradient descent is a trick with which the AI takes small steps in the right direction during training. It looks for the point where the errors are smallest. This happens millions of times, until the model is good.

Language models

Hallucination

A hallucination is a made-up or false answer from an AI that sounds plausible. The model has no source for it, but rather invented the information. For important topics you should therefore always check the answers.

Machine learning

Hyperparameter

Hyperparameters are settings that a person defines before training, for example how fast the model should learn or how big it is. They change how the training runs. Finding the right values is a task of its own.

Computer Vision

Image Generation

Image generation is the ability of an AI to create new images, based on a text description or another image. Well-known tools for this are Midjourney, DALL-E or Stable Diffusion.

Computer Vision

Image Recognition

Image recognition is the ability of an AI to recognize what can be seen in an image, such as people, objects, animals, or scenes. It is used, for example, in photo apps, security cameras, or with medical images.

Application

Implementation

Implementation is the actual rollout of a solution in everyday work. The AI is connected, tested, taught, and built into workflows. It is the step where many projects fail when it is underestimated.

Application

Integration

Integration means building an AI into existing programs or processes. Only through integration does AI become useful in everyday life; otherwise it stays an isolated tool. This often happens via APIs.

Application

Interface

An interface is the connection between two systems through which data or commands flow. APIs are one form of interface. Screens or input forms are interfaces too, between humans and computers.

Fundamentals

Label

A label is a marking on a data point that reveals its correct answer. When you show the AI an image with the label 'dog', it knows what a dog looks like. Labels are the teachers of the AI during training.

Language models

Large Language Model

LLM, transformerbasiertes Sprachmodell

A large language model is a computer program that was trained on an incredible amount of text, books, websites, conversations. As a result it has learned how language works and can answer questions, write texts, or translate. ChatGPT is one such model.

Machine learning

Loss Function

Loss Function (Fehlerfunktion)

A loss function measures how far an AI's prediction is from the correct answer. The smaller the loss, the better. During training the AI tries to keep lowering this value.

Fundamentals

Machine Learning

Machine Learning (Maschinelles Lernen)

Machine Learning means a computer is not programmed directly; instead it receives many examples and learns the rules by itself. The more and the more varied the examples, the better the predictions. This is the core of modern AI.

Tools

Midjourney

Midjourney is an AI tool for image generation, known for its especially aesthetic, often artistic style. It runs through Discord or its own website. You enter an image description and get images within seconds.

Fundamentals

Model

An AI model is the result of the learning process: a program that has learned rules from many examples and can now make predictions on its own. You can picture it like a trained expert, except that it is software.

Research

Multimodal AI

Multimodal AI understands and creates several kinds of content, text, image, audio, video, within a single model. Instead of processing only text, it can, for example, look at a photo and talk about it.

Fundamentals

Narrow AI

Narrow AI is specialized in a single task: playing chess, translating texts, recognizing images. It is good at that one thing but cannot do anything else. All AI systems we use today are narrow AI.

Language models

Natural Language Generation

Natural language generation is an AI's ability to write texts on its own: answers, articles, stories, explanations. The model chooses, word by word, what fits the task. Today the texts are often hard to tell apart from human-written ones.

Language models

Natural Language Understanding

Natural language understanding means that an AI does not just recognize words but also grasps what is meant. This includes dealing with different phrasings, spellings or unclear sentences. Without language understanding, chatbots would be useless.

Computer Vision

Object Detection

Object detection goes one step further than image recognition: the AI finds not only what is in an image, but also where. It draws rectangles around the objects it finds. Important for self-driving cars and robotics.

Computer Vision

OCR

Optical Character Recognition (Texterkennung)

OCR is the recognition of text from images or scanned documents. A photo becomes searchable text. This way, even old or printed documents can be used digitally.

Tools

Open Source AI

Open Source AI means: the code and often the models too are freely available. Anyone can view, change and use them, even without licensing costs. This enables research, adaptation and trust, because outsiders can check what happens.

Machine learning

Overfitting

Overfitting means that an AI has memorized the training data instead of finding general rules. With new data it then performs poorly. It is a main problem during training and is detected through validation.

Tools

Perplexity

Perplexity AI

Perplexity is an AI search engine that gives answers in full sentences and cites sources. It combines a language model with a live search on the web. This makes its answers often more up to date than those of pure chatbots.

Application

Pilot Project

A pilot project is a small test before an AI is rolled out at scale. You try it out in one area, learn from the result and then decide about scaling. This lowers risk and costs.

Computer Vision

Pixel

A pixel is the smallest image point of a digital image. Thousands or millions of pixels together make up a photo. An AI sees an image as a table of pixel color values and looks for patterns in it.

Language models

Prompt

A prompt is the instruction or question you give to a language model. How good the answer turns out depends greatly on how clear and precise your prompt is. Writing good prompts is a skill in its own right.

Language models

Prompt Engineering

Prompt engineering is the art of writing the best possible instructions for AI. Professionals test many variants and find out which words and structures produce the best results. It is a new activity that emerged with language models.

Language models

Prompting

Prompting is the process of giving an AI a task. It is about choosing the right words so the AI understands what is wanted. Good prompting makes the difference between mediocre and good answers.

Research

Quantum Computing

Quantum computing uses the rules of quantum physics to solve certain tasks extremely fast. It is not yet ready for everyday use, but researchers are investigating how quantum computers could speed up AI tasks in the future.

Language models

RAG

Retrieval Augmented Generation

RAG means that an AI first looks something up in a database before answering, and then responds using what it found. This way it uses current or company-specific knowledge that it did not have during training. Hallucinations become less frequent.

Machine learning

Regression

Regression means: the AI predicts a number instead of choosing a category. For example, the price of an apartment, the temperature on the next day or the expected revenue. To do this it needs suitable data from the past.

Machine learning

Reinforcement Learning

Reinforcement Learning (Verstärkendes Lernen)

Reinforcement learning is learning through trial and reward. The AI tries things out, gets points for good actions and learns over time what works. This is how computers play chess or control robots.

Application

Scaling

Scaling means extending a tested solution to larger volumes or more users. What works with ten people may not work with ten thousand; scaling brings its own challenges.

Computer Vision

Stable Diffusion

Stable Diffusion is a well-known open-source diffusion model for image generation. Because the code is freely available, there are many variants and tools built on top of it. It was an important turning point for open AI image generation.

Fundamentals

Strong AI

Strong AI / AGI

Strong AI would be an AI that could do everything a human can: think, learn, adapt, solve problems of all kinds. No such AI exists today. Researchers debate whether and when it will ever emerge.

Machine learning

Supervised Learning

Supervised Learning (Überwachtes Lernen)

In supervised learning, the AI is given many examples that already include the correct answer. It learns to match each input to the right answer. This is the most common form of AI training and needs a lot of labeled data.

Machine learning

Test Data

Test data is a separate set of data used only for the final check of the AI. It shows how well the model really works on inputs it has not seen before. Training, validation, and test data are kept strictly separate.

Language models

Token

A token is the smallest unit into which a language model breaks down text. Usually that is three to four letters or a short word. The more tokens a model can process at once, the longer the texts it can understand.

Language models

Tokenization

Tokenization is the process of breaking text down into tokens, that is, into small building blocks the model can work with. The model learns in advance which token splits fit best, based on many texts.

Fundamentals

Training

Training is the name for an AI's learning process: you show it lots of examples, and step by step it adjusts until it masters the task well. It is similar to practicing for an exam, just with millions of examples.

Machine learning

Training Data

Training data is the material an AI learns from. Its quality and variety determine how good the model turns out to be. Poor training data leads to poor or biased AI results.

Language models

Transformer

Transformer-Architektur

Transformers are a design for AI models that was invented in 2017. They are especially good at understanding how words relate to each other, even across long texts. Almost all modern language models, from ChatGPT to Gemini, are based on transformers.

Ethics

Transparency

Transparency means it is clear what an AI does, what data it was trained on, and who operates it. Without transparency, no one can check whether the AI is fair, correct, and safe. It is the foundation of trust.

Machine learning

Underfitting

Underfitting means the AI has not learned enough from the data. It is too simple and delivers poor results on both training data and new data. The fix: more data, longer training, or a larger model.

Machine learning

Unsupervised Learning

Unsupervised Learning (Unüberwachtes Lernen)

Unsupervised learning works without answers. The AI is given only data and searches for patterns or groups on its own. Useful when you do not know what structures are hidden in the data.

Application

Use Case

Use Case (Anwendungsfall)

A use case is a concrete application: what is the AI actually used for? Instead of saying 'AI in marketing' in general, a use case describes, for example: 'AI writes social media posts from product data'.

Machine learning

Validation

Validation is testing an AI with data it did not see during training. This reveals whether it has really learned or only memorized. Without validation, you do not know whether a model is any good.

Application

Workflow

A workflow is a fixed sequence of work steps. With AI, workflows can be partly or fully automated. This saves companies time and lets people focus on the tricky parts.

Research

Zero-Shot Learning

Zero-shot learning means: the AI solves a task without having seen a single example, just from the description. This works because it already knows a lot from pretraining. For simple tasks it often works, for tricky ones less so.