The AI glossary you can actually understand

Tired of people throwing around terms like LLM, prompt injection and AGI as if it were everyday talk? Relax – here is the guide that explains the most important AI terms in English but with clear Swedish explanations, all categorized. A perfect mix of geekiness and clarity. M

🧠 General about AI

• AGI (Artificial General Intelligence) – An AI that can perform any intellectual task, like a human.
• AI (Artificial Intelligence) – Technology where machines simulate human intelligence.
• Artificial life – AI models that simulate biological life and evolution.
• Autonomy – The ability of AI to act without human intervention.
• Bias – Built-in prejudices in AI systems caused by skewed data.
• Black box – When it is impossible to understand how AI arrived at a decision.
• Closed-source – Software or AI model whose code is not available to the public.
• Cognitive computing – Systems that attempt to mimic human thinking.
• Ethics – Moral considerations regarding the use of AI.
• Explainability – How easy it is to understand how AI reasons.
• Hallucination – When AI fabricates false information.
• Intelligence explosion – A hypothesis that AI can improve itself exponentially.
• Knowledge representation – How AI organizes and stores information.
• Model interpretability – How easy a model is to understand and explain.
• Narrow AI – An AI with expertise within a limited domain.
• Open-source – Software that is freely available and modifiable.
• Singularity – A hypothetical point where AI develops faster than we can control.

📚 Träning och inlärning

• Backpropagation – AI’s way to correct itself during training.
• Batch size – How much data is used per training step.
• Bias-variance tradeoff – Balance between too much and too little learning.
• Cross-validation – Technique for testing AI models on multiple data splits.
• Data augmentation – Artificially increasing the amount of training data, for example by rotating images.
• Dataset – An organized set of data used for training.
• Epoch – One pass through the entire dataset during training.
• Few-shot learning – AI that learns from only a few examples.
• Fine-tuning – Training an existing model with new data.
• Gradient descent – A method to minimize errors during training.
• Hyperparameter – Predefined values that control the AI’s learning.
• Loss function – A measure of how wrong the AI model is during training.
• Mini-batch – A small subset of data used in training.
• Overfitting – When the AI memorizes the training but cannot generalize.
• Pretraining – Initial training before the model is used practically.
• Reinforcement learning – AI that learns through rewards and punishments.
• Supervised learning – Training with answers where the AI gets the correct results.
• Training data – Data used to teach AI to perform correctly.
• Underfitting – When the AI has not learned enough to perform well.
• Zero-shot learning – AI handles tasks it has never seen before.

💬 Language and comprehension (NLP)

• Attention mechanism – A technique where AI focuses on relevant parts of the text.
• Autoencoder – A model that learns to represent data efficiently.
• BERT – AI model that reads text in both directions for better understanding.
• BLEU score – A measure to assess the quality of translated text.
• Context window – The amount of text AI can hold in memory.
• Coreference resolution – When AI understands which words refer to the same thing.
• Inference – AI’s process of drawing conclusions based on its knowledge.
• LLM (Large Language Model) – A very large language model trained on massive amounts of text.
• Named Entity Recognition (NER) – When AI identifies names, places, dates etc. in text.
• Natural language processing (NLP) – Technology that enables AI to understand and generate language.
• POS tagging – Labeling parts of speech (noun, verb etc.).
• Prompt engineering – The art of formulating good AI questions.
• Prompt injection – Tricking AI by hiding instructions in text.
• Semantic analysis – Understanding the meaning of text, not just the words.
• Semantic search – Searching for meaning rather than exact words.
• Sentiment analysis – Analyzing whether the text is positive, negative, or neutral.
• Token – A small unit of text that AI processes.
• Transformer – AI architecture that makes language models efficient and context-aware.

🎨 AI-generated content

• 3D generation – AI that creates three-dimensional objects or environments.
• Audio synthesis – AI-generated music, voice, or sound.
• Diffusion model – Model that creates images by gradually removing noise.
• GAN – Two AI models competing to create realistic content.
• Generative AI – AI that creates new content (text, image, music).
• Image generation – AI that creates images from, for example, text commands.
• Multimodal AI – AI that handles multiple data types simultaneously.
• Neural style transfer – AI transfers the style of one image to another.
• StyleGAN – A specific GAN model used to create faces.
• Text-to-image – Technology where text descriptions are converted into images.

🤖 Interaction with users

• Chatbot – An AI system you can chat with.
• Conversational AI – AI that maintains coherent conversations.
• Dialogue management – Technology that controls how AI handles a conversation.
• Embodied AI – AI integrated into robots or physical devices.
• Multiturn conversation – AI that remembers and responds over multiple messages.
• Speech recognition – AI that converts speech to text.
• Turing test – A test where AI tries to appear human.

🚨 Safety and risks

• Adversarial attack – Tricking AI with manipulated input.
• Alignment – Ensuring AI’s goals and behavior align with human values.
• Anomaly detection – AI detects unusual or dangerous patterns.
• Deepfake – AI-generated content that makes it look like someone did or said something they did not.
• Explainable AI (XAI) – AI designed to be understandable.
• Kill switch – The ability to shut down AI in dangerous situations.
• Model collapse – When an AI loses capability due to poor training.
• Red teaming – Security testing where AI is exposed to attacks.
• Safety constraints – Restrictions that prevent harmful AI behavior.
• Synthetic data – Fake but realistic data, used for training without violating privacy.