0 Comments // Reading Time: 8 min.
Artificial intelligence (AI) is no longer a topic for the future – it is already present in numerous applications in our everyday lives. But what is behind terms such as ‘AI model’, ‘deep learning’ or ‘strong AI’? And how do tools such as ChatGPT or Midjourney differ in the way they work?
In the coming weeks and months, we will be publishing a series of articles in which we compare different AI tools for specific areas of application and highlight their strengths and weaknesses. We will also take a look behind the scenes: How do AI models actually work, what types of artificial intelligence are there, and how do different learning methods differ?
Before we delve deeper into these topics, we will clarify the most important terms relating to AI in this first article. This will lay the foundations for a better understanding of the following articles and the many facets of artificial intelligence.
Artificial intelligence (AI) refers to the ability of machines to solve tasks that normally require human intelligence – such as learning, decision-making or problem-solving. This is based on algorithms, i.e. clearly defined instructions that specify exactly how a particular problem should be solved. Algorithms usually consist of a series of logical steps – similar to a cooking recipe – that a machine strictly follows.
Modern AI often uses machine learning, where algorithms are designed to independently recognize patterns from sample data and learn from them. Instead of being programmed for every decision, these systems improve through experience. A simple example: an AI that sees many photos of dogs and cats gradually learns to distinguish between the two animal species independently based on characteristics such as ear shape or coat structure.
The following table provides a compact overview of key concepts - from the basics to the types of AI and their practical application.
AI model
An AI model is the basis for both systems and tools and the foundation of all AI applications. It is a trained neural network that solves tasks such as text generation, image analysis or speech comprehension. Examples are: GPT-4 (text), DALL-E 3 (image generation), Whisper (speech recognition).
AI system
An AI system is a way of using AI and an overall technical solution that uses one or more models to perform tasks independently in the background, e.g. autonomous vehicle, diagnostic AI in medicine.
AI tool
An AI tool is also a form of AI usage. It is a user-friendly application that uses one or often several AI models and enables direct interaction with humans, e.g. ChatGPT, DeepL Translator, Midjourney.
Types of AI (by capabilities)
Klassifikation nach der Leistungsfähigkeit der KI, dazu gehören spezialisierte schwache KI, theoretische starke KI sowie hypothetische Superintelligenz. Die Einteilung nach Fähigkeiten unterscheidet KIs danach, wie flexibel und leistungsfähig sie sind. Hier zeigt sich der Unterschied zwischen spezialisierten, theoretisch generalisierenden und übermenschlichen Systemen.
Classification according to the capabilities of the AI, including specialized weak AI, theoretical strong AI and hypothetical superintelligence. The classification according to capabilities differentiates between AIs according to how flexible and powerful they are. This shows the difference between specialized, theoretically generalizing and superhuman systems.
| AI type | Description | Status & examples |
| Weak AI | Performs a specialized task without real understanding. | Image recognition, voice assistants such as Alexa, Siri |
| Strong AI | Can think flexibly and act independently in different contexts. | Hypothetical; so far only theoretical concepts |
| Superintelligenz | AI that surpasses human intelligence in all areas. | Pure future scenario |
Types of AI (by degree of development)
Classification according to the degree of cognitive development or functionality of an AI. These include Reactive AI (AI that reacts), Limited Memory AI (AI with limited memory capacity), Theory of Mind AI (AI that can fully understand emotions and human behavior), Self-Aware AI (AI with an awareness of itself). The following table presents the four stages of this development – from simple reactions to hypothetical self-awareness.
| AI type | Description | Status & examples |
| Reactive Machines | AI that only reacts to current stimuli: no data storage or learning possible. | Reality, e.g. Deep Blue (chess computer from IBM) |
| Limited Memory | AI that learns quickly from past experience and takes past data into account. | Reality, e.g. self-driving cars |
| Theory of Mind AI | AI that can understand emotions, intentions and beliefs. | Vision of the future |
| Self-Aware AI | AI with its own consciousness and self-awareness. | Theoretical concept |
Artificial intelligence comprises various sub-areas, which we briefly present in the following table.
Machine Learning (ML)
The ability of systems to learn from data and recognize patterns based on algorithms – without explicit programming. The more data is available and the more an application is used with machine learning, the better it becomes. ML requires structured data, such as tables with the various data and is used, for example, for optimizing processes, predicting probable values or behaviour (e.g. with customers), recognizing correlations or groups.
Deep Learning
Sub-area or type of ML with particularly deep, multi-layered neural networks. In contrast to traditional ML, deep learning can also process unstructured data such as images, music, videos or speech without having to define features for structuring beforehand. Deep learning therefore requires particularly powerful computers and a lot of time (up to several months) to analyze the data. It is particularly suitable in areas such as image processing, speech recognition and natural sciences.
Neural networks
Neural networks are part of deep learning and a general AI technology inspired by the human brain. They consist of artificial neurons/nodes and are the basis for deep learning, machine learning and many modern AI models. Each node is connected to another and there are always multiple layers: Input layer, hidden layer and output layer. The hidden layer can consist of many layers. Particularly deep neural networks are assigned to deep learning. Neural networks are used for complex problems such as language translation, weather forecasts, economic issues and many other areas (see Deep Learning).
Natural Language Processing (NLP), Natural Language Understanding (NLU), Natural Language Generation (NLG)
Language processing enables machines to understand and interpret human language. ML, deep learning and statistical methods are used for this purpose. NLU and NLG are sub-areas of natural language processing, which also uses many other methods to enable AI to understand, process, produce and translate human language. Language processing or NLP can be found in many AI applications such as virtual assistants (Alexa, SIRI), chatbots, processing customer opinions and in customer service, as well as spam filters and speech-to-text conversion, e.g. for mailbox messages that arrive in your inbox in text form.
Within machine learning, there are different learning methods and different types of algorithms, depending on whether labeled data, independent learning or feedback is used. These methods form the basis of many AI models.
Supervised Learning
Training method with labeled data: The training data sets consist of sample data containing both input values (input) and associated target values (output). The model learns to recognize correlations between input and output. Supervised learning is typically used for tasks such as classification (e.g. categorizing customer data) or forecasts (e.g. sales forecasts).
Unsupervised Learning
Pattern recognition without predefined output values: The algorithm receives unstructured/incomplete data and should independently recognize patterns or correlations – without a defined goal. Unsupervised learning is used to form clusters (groups of similar data points), uncover correlations or reduce variables, for e.g. in the personalization of offers or in shopping basket analysis.
Semi-Supervised Learning
Mixture of supervised and unsupervised learning: This method uses both a small set of labeled sample data with defined target values and a larger amount of unlabeled data. The model is first trained with a small amount of known data and then uses this basis to efficiently classify additional, unknown data. The advantage: significantly less manually created training data is required, which saves time and money. Semi-supervised learning is often used in similar application areas to supervised learning, such as the classification of texts or images.
Reinforcement Learning
Learning through reward and punishment in interactive environments: Initially, the algorithm is not given any instructions as to which actions are right or wrong. Through positive or negative feedback, it learns step by step which actions lead to the desired results. Over time, the model independently develops strategies to maximize rewards. The feedback arises automatically from a programmed reward function within the environment. Reinforcement learning is a central learning method for AI development and is used, for example, in autonomous driving and autonomous robotics.
Transfer Learning
Transferring previously learned patterns to new tasks: In transfer learning, an existing model that has been trained for a specific task is used to solve a related but new problem more quickly. Instead of starting from scratch, the new model builds on existing knowledge. This saves considerable amounts of training data, computing power and development time. Transfer learning is often used when only a small amount of specialized data is available, for example in medical image analysis or in the processing of small language corpora. Language corpora are large collections of texts or voice recordings that are used to train AI systems.
AI is a complex but fascinating field that ranges from technical basics to philosophical future scenarios. We come across terms such as neural networks, machine learning and superintelligence more and more frequently – and it is worth understanding what they mean. Because only those who are familiar with the central concepts can meaningfully classify and evaluate AI technologies.
In the upcoming articles in our AI series, we will go one step further: we will show how different tools are used in practice, compare their areas of application and explain which AI models are behind them. So if you want to better understand what AI can do today – and what it can't (yet) do – stay tuned.
Comments