AI glossary

The following AI glossary explains the most important basic terms of artificial intelligence in simple language.

Agent

An agent is a system that can perceive surroundings, make decisions and act autonomously to reach certain goals.

An AI-Agent can be a software program or also a physical robot. Agents are distinguished by their ability to act flexible on new situations and to learn from new experiences.

An intelligent customer service agent can, for example, understand customer enquiries, call up information from a knowledge data base and solve the problem of the customer independently or to forward it to a human employee.

Data Bias

Data Bias is created when training data of a AI-models contain systematic distortions or prejudices. These inequalities are learnt and reproduced by the AI, which may lead to discriminatory results.

When, for example, an AI-system is trained for application with data of male employees, it could learn incorrectly to prefer male applicants.

Deep Learning

Deep Learning is a method of the machine learning that uses artificial neuronal networks with a lot of layers, similar to the structure of a human brain. These networks can notice quite complexe patterns in large data amounts like pictures, sounds or texts.

The face recognition is an application example.

Generative AI

Generative AI is a form of the artificial intelligence that is programmed on creating new contents, for example texts, pictures, music or videos. It is trained by huge amounts of data to recognize patterns and to create something completely new on this basis.

This technology is the engine behind applications like chat GPT that lead human conversations or create creative texts on command.

GPT

GPT stands for Generative Pre-trained transformer and describes a family of Large Language Models.

Generative means that the model can create new contents.

Pre-trained means that it has already been pre-trained with a huge amount of data.

Transformer is the underlying architecture.

These models are the driving force behind applications like chatGPT and can be used for a wide range of tasks like creating texts or pictures as well as programming.

Hallucination

A hallucination is an answer generated by generative AI that sounds convincing, but is objectively wrong or that is not based on your training data. The model invents information, as it is designed for that purpose to generate plausible word sequences. It is irrelevant for the algorithm, if it is right or wrong.

An example for hallucination is a chatbot that writes a detailed biography about a non-existent person or an AI-tool for creating videos that invents tourist attractions in foreign countries.

Inferencing

Inferencing (German: "Inference" or "Conclusion") describes the application of acquired knowledge through an AI model to react to new, unknown data. This is the active part of the AI, where it performs her actual job.

After the model has been trained, for example, on what a cat looks like, it can analyze an unknown picture and draw the conclusion (inference), that a cat can be seen.

AI (Artificial intelligence)

Artificial intelligence, often abbreviated as AI or KI (German: "Künstliche Intelligenz"), describes the ability of computer systems to perform tasks that normally require human intelligence. This includes learning from experience, understanding language, recognizing patterns, and making decisions.

AI has already been present in many areas of our everyday lives, from personalized recommendations for streaming-services to intelligent assistants on our smartphones. After publishing chat GPT in 2022, the generative AI with tools for text, pictures or sound generation has become more of a focus.

Context size

The context size, also described as "context window", indicates how much information can be considered by a language model, when it generates an answer. A larger context size means that the model can “remember” longer sections of text or conversation threads and thus provide more coherent and relevant responses.

The context size determines how long a chatbot can "remember" what was said or written at the beginning after multiple messages.

Large Language Model

A Large Language Model (LLM) is an AI model that has been trained on understanding and generating human language on a large scale. These models are trained with huge amounts of text data from the Internet and books to write human-like texts, answer questions and translate languages.

A popular example for an application, that is based on a LLM, is chat GPT.

Machine Learning

Machine Learning (German: "Maschinelles Lernen") is a subarea of AI, in which computers learn from data without being explicitly programmed for each task. Algorythms are trained with huge amounts of data to recognize patterns and, on this basis, make predictions for new, unknown data.

Imagine you show thousands of pictures of cats to your computer. Through machine learning, the system eventually learns on its own to recognize a cat in a new image.

Model

A model in the sense of artificial intelligence describes a software program that has been trained with huge amounts of data to recognize patterns, make decisions or to generate new contents.

The most popular form of a model is the Large Language Model that creates new, textual content based on text input.

Multimodality

Mulitmodality describes the ability of an AI system to process and understand information from different input data - such as texts, pictures, videos and sound at the same time.

A multimodal model can, for example, look at a picture and create a suitable textual description or understand a spoken question and answer with a relevant picture.

Neuronal network

An artificial neuronal net is a computer model that is inspired by the functionality of the human brain. It consists of a lot of processing units, which are called neurons, that are connected with each other and that are arranged in layers.

Through a training process, the network learns by adjusting the connection strengths (also known as weights) between the neurons so that it performs a specific task better and better.

Here is an example for better comprehension: Imagine you want to teach a child to differ apples and pears. You show him lots of pictures and tell him what each picture shows. A neuronal net learns it in a similar way: It is fed with thousands of pictures of apples and pears and it adjusts its internal "adjusting screws" until it can differ apples and pears on new, unknown pictures on its own.

Parameter

Parameters are the internal variables of an AI model that are adjusted during the training process. You can think of them as the model's adjustment screws, whose values determine how the model solves certain tasks.

The number of parameters is often a measure for the complexity and performance of a model. Modern and large language models have billions, partially trillions, of parameters.

Predicative AI

Compared to generative AI, the predicative AI creates no new contents, but classifies or differs between different kinds of data. It learns to categorize data and to recognize different categories. It can make a decision or predictions on this base.

This kind of AI is often used for exercises like spam detection in emails or the diagnosis of diseases on the basis of medical pictures.

Prompt

A prompt is the instruction or question given by a user of AI to trigger a certain answer or action. The quality of the prompt has a huge influence on the result that is delivered by the AI. The clearer and more detailed the instructions, the better the response will usually be.

Quantization

Quantization is a process, where the precision of numbers is reduced in an AI model to make it smaller and faster.

This reduces the storage requirements and the energy consumption of the model.

Reinforcement Learning

Reinforcement Learning is a learning process through trial and error. An AI system, a so-called agent, learns through reward and punishment, in an environment to make decisions. The agent gets a reward for the right decision and a punishment for wrong decisions, whereby he learns to develop the best strategy.

An AI, that plays a video game, learns, for example, through reinforcement learning. It tries out different moves and is rewarded for reaching a new level (reward), while learning to avoid this mistake when it loses a life (punishment).

Supervised Learning

Supervised Learning is a method that trains an AI model based on data that is already labeled with the right answers. The algorythm learns by comparing its predictions with the correct results and by correcting itself, similar to a pupil that studies with a solution book.

This method is often used for predictions of stock prices or for the recognition of credit card fraud.

If you want to teach to the AI to recognize cats, you have to show it, for example, thousands of cats that are marked with "cat" or "no cat". The AI is able to identify cats on the new pictures on its own after the training.

Token

A token is the smallest unit, in which a text is broken down by a language model, to process it. A token can be a complete word, a part of a word or an individual sign.

These tokens are like building blocks for the AI from which it composes and understands language.

Training

The training is the process where the AI model learns to fulfill a certain task by feeding it with a huge amount of data. During the training, the model adjusts its internal parameters to recognize patterns in data and to make its predictions more and more accurate.

You can imagine it like studying for an exam: The more data the model works through, the better its performance will be in the actual test.

Transformer

The transformer architecture is a special design for neuronal networks. This design is particularly good at understanding relations in sequential data, for example texts. Compared to older model, a transformer can consider all parts of a sentence at the same time, whereby it can capture the context better.

This technology is the basis for the most modern and large language models such as GPT.

Unsupervised Learning

The AI model receives data without any labels or given answers. The task of the AI is to find patterns, structures or anomalies in the data. This method is quite useful to analyze large amounts of data.

For example, an online shop could use unsupervised learning for analyzing the purchasing behaviour of its customers.

Author: Stefan Bohn Stefan Bohn has been employed at Thomas-Krenn.AG since 2020. Originally based in PreSales as a consultant for IT solutions, he moved to Product Management in 2022. There he dedicates himself to knowledge transfer and also drives the Thomas-Krenn Wiki.

Translator: Alina Ranzinger Alina has been working at Thomas-Krenn.AG since 2024. After her training as multilingual business assistant, she got her job as assistant of the Product Management and is responsible for the translation of texts and for the organisation of the department.