Find your way to be trained or even get certified in TMAP.
Start typing keywords to search the site. Press enter to submit.
To effectively navigate the field of Artificial Intelligence (AI), it’s essential to grasp a set of foundational concepts and understand how they interconnect. AI is a broad domain that encompasses several subfields, each with its own techniques and objectives. This chapter outlines the core terms and their relationships, highlighting the key concepts while briefly clarifying the supporting terminology.
The following terms are commonly used in the context of intelligent machines. The below figure depicts the following conceptual relationships:
Together, these terms form a conceptual map of the AI landscape. Understanding them is crucial for engaging with AI technologies whether building, testing, or governing those technologies.
AI is the overarching discipline concerned with building systems that mimic human cognitive functions. This includes capabilities like understanding language, recognizing images, solving problems, and making decisions. AI systems are designed to perform tasks that typically require human intelligence and can be either rule-based or data-driven.
Machine learning is a branch of artificial intelligence (AI) that enables systems to learn from data and improve over time without being explicitly programmed. It uses algorithms to identify patterns, make predictions, or take actions based on input data. There are different types of machine learning, including supervised, unsupervised, and reinforcement learning. In supervised learning, models learn from labeled data; in unsupervised learning, they find hidden structures in unlabeled data. Reinforcement learning involves learning through trial and error. Machine learning powers many modern technologies, such as recommendation systems, image recognition, and fraud detection. It forms the backbone of many generative AI systems used today.
Deep Learning is a specialized area within ML that employs neural networks with many layers to capture and model complex data patterns. These architectures are particularly effective in handling high-dimensional data such as images, audio, and natural language, enabling state-of-the-art results in perception and generation tasks.
Generative AI refers to systems that can create new content text, images, audio, or even code often indistinguishable from human-generated artifacts. These systems learn from existing data and generate novel outputs based on learned patterns, rather than simply classifying or labeling inputs.
Agentic AI refers to the use of AI entities that work on behalf of a human and have autonomy and authority to take actions without detailed step-by-step instructions.
LLMs are a category of generative AI models designed specifically to understand and generate human language. Built using transformer architectures, they are trained on vast amounts of text and can perform a range of language tasks, from answering questions to summarizing documents or composing creative writing.
Diffusion models are a class of generative models used to create data, such as images or text, by simulating a reverse process of gradual noise removal. They start by adding random noise to data over many steps, then learn how to reverse this process to reconstruct the original data. During generation, the model starts from pure noise and iteratively denoises it to produce realistic outputs. Diffusion models have gained popularity due to their ability to generate high-quality and diverse content. They are widely used in generative AI systems, such as image synthesis and text-to-image tools like DALL·E and Stable Diffusion.
To contextualize the core concepts, several additional terms are important:
Algorithm: A sequence of computational steps used in all AI systems to process data and learn from it.
Neural Network: A neural network is a computational model inspired by the human brain, consisting of layers of interconnected nodes (neurons) that process data and learn patterns. It is the foundation of many AI systems, especially deep learning. Neural networks are implemented in software frameworks (e.g. TensorFlow, PyTorch), but they require powerful hardware, especially GPUs or TPUs, for efficient training and inference. The software defines the model and learning process, while the hardware accelerates computation. Together, they enable tasks like image recognition, language processing, and generative AI.
Transformer Model: The neural architecture behind most LLMs, using mechanisms like attention to process language more effectively. Transformers are central to the success of LLMs, made effective through the attention mechanism, enabling these models to dynamically focus on the most relevant parts of an input sequence when processing data.
Foundation Model: A large pre-trained model like Generative Pretrained Transformer (GPT) or Bidirectional Encoder Representations from Transformers (BERT), adaptable to various downstream tasks.
Multimodal Model: A model capable of processing multiple data types simultaneously (e.g., text + image).
Generative Adversarial Network (GAN): A generative model architecture involving two competing networks to produce realistic outputs.
Hyperparameter and Parameter: Settings that govern model training; the former are pre-set, while the latter are learned.
Transfer Learning: Reusing a model trained on one task for a related task, improving efficiency.
Vectorization: The transformation of data into numeric formats usable by machine learning models.
Guardrails: Guardrails in AI are mechanisms that ensure systems behave safely, ethically, and within intended boundaries. They can include rules, filters, or human oversight to prevent harmful or biased outputs. Guardrails are essential for the responsible deployment of generative AI.
Responsible AI: Responsible AI refers to designing, developing, and deploying AI systems in a way that is ethical, transparent, and aligned with human values. It emphasizes fairness, accountability, and minimizing harm. The goal is to ensure AI benefits individuals and society as a whole.
Hallucination: Hallucination in AI refers to when a model generates information that is false, misleading, or not grounded in real data. This is common in generative systems that prioritize fluent responses over factual accuracy. It poses risks in high-stakes domains like healthcare, law, or software engineering.
Supervised learning: A machine learning approach where models are trained on labeled data, learning to map inputs to known outputs. It’s commonly used for tasks like classification and regression. The model improves by minimizing the difference between predictions and actual labels.
Unsupervised learning: Involves training models on unlabeled data to discover hidden patterns or groupings. It’s used for clustering, dimensionality reduction, and anomaly detection. The system learns without explicit feedback.
Reinforcement learning: A trial-and-error approach where an agent learns by interacting with an environment, receiving rewards or penalties. The goal is to maximize cumulative reward over time. It’s often applied in robotics, gaming, and decision-making systems.
Prompt: The input given to a generative model to initiate a response or generation. This will be explained in more detail in the module, Prompt Engineering.
Bias in AI: Refers to systematic errors or prejudices in model behavior caused by imbalanced, incomplete, or skewed training data. These biases can lead to unfair, inaccurate, or discriminatory outcomes when the AI is deployed.
Explainable AI (XAI): Refers to methods and techniques that make the behavior and decisions of AI systems transparent, understandable, and traceable to humans. It helps build trust, supports accountability, and enables effective validation and debugging.
Confabulation: In the context of artificial intelligence refers to the phenomenon where an AI system lacks certain training data or user input and generates information that is plausible sounding but false, inaccurate, or entirely fabricated. These outputs are presented as if they are factual, even though they are not grounded in the AI’s training data or the real world (also see “Hallucination”). Confabulation typically arises because AI models, especially large language models, are designed to produce coherent and contextually appropriate responses based on statistical patterns in their training data, rather than verifying factual correctness or truly understanding of the content. More about confabulation can be found in Section “Prompt Crafting“.
The whole world seems to talk about nothing else than Generative AI and agentic AI. So, the superficial observer may think that all AI is GenAI and agents.This is not true. In our department with AI specialists, we see that about half of the work is indeed GenAI, but the other half of the work concerns applying traditional AI (such as training specific models with machine learning) or data engineering. A classic example (since 2018) is Sogeti’s Artificial Data Amplifier (ADA) solution to generate synthetic data for all kinds of purposes.
When GenAI first became available for the general public, many believed everything should be done with it from now on. By now we have already learned that there still is room for traditionalforms of AI. Object detection, as an example, can very well (or even better) be done with specifically trained AI models, rather than with generic GenAI models.Traditional AI models have qualities like built for a specific task, e.g. classification where you don’t need to generate, and only require a task specific classification, trained on non-public (company) data for your specific problem. This is faster, less compute power is needed, and results are more interpretable and explainable etc. Another advantage of applying traditional machine learning models is that there is the possibility of “freezing” the model after tests have shown that the machine learning process has resulted in an AI model with the right level of quality. After the model is frozen, it will keep performing in the same way, resulting in a much more reliable IT system largely because of the removal of the non-deterministic system characteristics.
Vibe Coding is an AI-based development approach where code is generated based on natural language prompts, allowing creators often with limited technical background to focus on the conceptual and creative aspects of software design rather than on implementation details.
It emphasizes the overall “vibe” or feel of the application over precise control of the underlying code, with the humans typically accepting AI-generated outputs without thorough review. While this accelerates prototyping and lowers the barrier to entry, it also introduces various risks such as shallow technical understanding, increased technical debt, and potential security vulnerabilities.
Vibe Testing is an AI-based software testing approach that validates the intended behavior and user experience of applications, without deep understanding of the underlying code and the principles of quality and testing.
It focuses on natural language prompts to assess whether applications function as expected from an end-user perspective. While this accelerates validation in rapid development contexts, it can lead to incomplete test coverage, overlooked security issues, accumulated technical debt, and many remaining quality risks.
Amplified Coding (Amp Coding) is a structured, expert-in-the-lead, GenAI-augmented coding practice where developers interact with AI tools to co-create high-quality, maintainable code.
It differs from vibe coding by emphasizing explainability, code correctness, and traceability throughout iterative development cycles. Amp coding combines natural language prompting and agentic support, with human quality gates, automated checks, and context-aware feedback to deliver robust and production-ready software.
Amplified Testing (Amp Testing) is a GenAI-enabled testing discipline that extends the principles of Amp Coding to the management, design, and execution of tests, with the expert remaining in control.
It leverages AI to generate purposeful, traceable testware that aligns with pursued business value and quality risks, enabling teams to verify, validate and explore both functional and nonfunctional requirements continuously.
Amplified Quality Engineering (Amp QE) is an approach to quality engineering empowered by modern practices and automation, especially but not exclusively, by GenAI, that amplifies the capabilities of team members involved in highperformance IT delivery.
Amp QE enhances all activities of the software lifecycle monitor, plan, design, build, test, integrate, deploy and operate by embedding various quality measures directly into the IT delivery process. Rooted in the TMAP definition of quality engineering, Amp QE integrates AI-driven tools and practices to improve quality at speed and scale across teams and systems. It enables continuous feedback, traceability, and risk-based decision-making, elevating quality from a responsibility to a shared capability. In the next Part we will dive deeper into how various roles can leverage GenAI to contribute to Quality Engineering.
AmpQE overview
