AI Compendium
Practical AI use: create images, enhance photos, write texts, simplify everyday life – plus the question of how AI can run locally and data-privacy-friendly on your PC.
What exactly is AI?
Artificial Intelligence – or AI for short – is a collective term for software that solves tasks which previously required human intelligence: understanding and writing texts, recognizing and generating images, translating languages, finding patterns in data.
What everyone means by "AI" in 2026 are large language models (LLMs) like ChatGPT or Claude, and image models like Midjourney or Stable Diffusion. They are the result of years of training with gigantic amounts of texts, images, and data – and have evolved from toys into real tools over the past two years.
This compendium is practice-focused: what you can actually do with it – instead of research background.
What is it about?
30+ topics, categorized into three clusters based on application.
AI for Creative Tasks
Creating images, rescuing photos, generating videos, writing texts – AI as a creative toolkit.
- AI images with Midjourney & Co.
- Enhance old photos with AI
- AI videos: Sora, Runway, Avatars
- AI texts: ChatGPT, Claude, Gemini
AI in Everyday Life
Practical helpers for every day – from email filters to grandma on a tablet.
- AI tools for everyday use
- AI for seniors (accessibility-friendly)
- Remove background noise
- The best AI tools 2026
AI on your PC
How to use AI locally on Windows – without your data going to the cloud.
- AI locally on Windows
- Local models (Ollama, LM Studio)
- AI on older PCs
- Data privacy in AI
The most important terms
From Prompt to Hallucination – AI technical language explained simply.
The instruction given to the AI – a question, a command, a description. "Write me a poem about mountains" is a prompt. The art of formulating good prompts is called Prompt Engineering.
The smallest unit of text processed by an AI. Approximately 4 characters = 1 token. A German word is often 1–3 tokens. Models have token limits that restrict their processing capacity.
Large language model. The family that includes ChatGPT, Claude, Gemini, and Llama. Trained with gigantic amounts of text and can understand and generate language.
When AI invents things that sound plausible but are false. Sources that don't exist. Facts that should be true but aren't. The biggest weakness of current LLMs.
How much text the AI can simultaneously "keep in mind." Models in 2026 can handle up to 1 million tokens – that's entire books. Larger window = better processing of long documents.
AI that understands multiple types of input: text, image, audio, video. Most modern models (GPT, Claude, Gemini) are multimodal – you can show them a photo and ask questions about it.
A mathematical representation of text as a numerical vector. AI uses embeddings to capture the meaning of words and sentences – the basis for search functions and recommendations.
Retraining an existing model with custom data – e.g., company-specific knowledge. This adapts AI for specific applications without training a new model from scratch.
AI that loads and incorporates external documents in real-time. The trick behind "AI with access to your company documents" – without fine-tuning, always with current data.
AI that doesn't just answer, but acts itself: opens programs, clicks buttons, solves tasks in steps. A historic moment: In early March 2026, an AI model for the first time surpassed humans on the OSWorld benchmark (75% vs. 72.4% human baseline) – in autonomous computer operation.
General artificial intelligence that achieves or surpasses human capabilities in all areas. Not yet reached in 2026 – but hotly debated. Some believe we are close, others far away.
Proprietary: ChatGPT, Claude, Gemini – only usable through the providers. Open Source: Llama (Meta), Mistral, DeepSeek – freely available, runnable locally. Open source models are rapidly catching up technically.
The actual application of a fully trained AI model. When you ask ChatGPT a question, that's inference. Inference requires less power than training, but for large models, still a GPU.
What a model has learned from. For modern LLMs: millions of websites, books, scientific papers, code. The quality and diversity of the training data determine the model's quality.
The neural network architecture that has revolutionized the AI world since 2017. The "T" in GPT stands for it. Transformers made today's language models possible – through efficient sequence processing.
📜 AI & Engelmann: We have been using AI components in our products for years – most visibly in Photomizer (AI-based image enhancement) and Photo BlowUp (AI upscaling). However, the rapid development of the last two years opens up entirely new possibilities for private users – and these are exactly what we show here.
More from the Engelmann Blog
Tips, tutorials, and background knowledge about software, AI, and digital life – from over 30 years of software development.