AI Wiki
The language of AI, clearly explained.
147 terms from “agent” to “zero-shot” – briefly defined, clearly explained and linked together.
147 terms
- Agent-to-agent (A2A)Agents & automationAn open protocol through which AI agents from different vendors communicate and hand over tasks.
- Agentic AIAgents & automationUmbrella term for AI systems that plan, act and complete tasks independently over longer workflows.
- AI agentAgents & automationAn AI system that pursues a goal independently over several steps and uses tools to do so.
- AI avatarImage, video & audioAn artificially generated, speaking person in a video who presents a given text.
- AI chips (TPU, NPU)Data & infrastructureSpecialised processors built specifically for the calculations of neural networks.
- AI governanceAI in businessThe rules, roles and processes a company uses to steer and control its use of AI.
- AI in customer serviceAI in businessUsing chat and voice assistants that answer customer enquiries or support staff in doing so.
- AI literacyAI in businessThe knowledge and skills to use AI tools sensibly, critically and safely.
- AI modelFundamentalsThe trained “brain” of an AI system: a mathematical function that computes outputs from inputs.
- AI music generationImage, video & audioCreating music, vocals or sound effects from text descriptions.
- AI policyAI in businessAn internal document that sets out how employees may use AI tools.
- AI regulation in SwitzerlandLaw, ethics & safetySwitzerland relies on existing laws and targeted amendments rather than a comprehensive AI act of its own.
- AI safety and securityLaw, ethics & safetyMeasures that prevent AI systems from causing harm – through errors, misuse or attacks.
- AI strategyAI in businessA company’s plan for where and how it uses AI to achieve concrete business goals.
- AI upscalingImage, video & audioEnlarging images or videos with AI that plausibly fills in missing detail.
- AI use caseAI in businessA concrete application in which AI performs a clearly defined task better, faster or more cheaply.
- AlgorithmFundamentalsAn unambiguous sequence of steps used to solve a problem or complete a task.
- AlignmentLaw, ethics & safetyThe research and practice of aligning AI systems so that they act in line with human intentions and values.
- APIData & infrastructureAn application programming interface through which software can use an AI model directly – without a chat interface.
- Artificial general intelligence (AGI)FundamentalsHypothetical AI that can learn and perform practically any intellectual task as well as a human.
- Artificial intelligence (AI)FundamentalsUmbrella term for computer systems that perform tasks that previously required human intelligence.
- AttentionLanguage modelsA mechanism that lets a model determine, for each word, which other words in the context are relevant to it.
- BackpropagationMachine learningThe method a neural network uses to calculate how much each weight contributed to the error.
- BenchmarkFundamentalsA standardised test used to compare the performance of different AI models.
- BiasLaw, ethics & safetySystematic distortions in AI results that disadvantage certain groups or create a skewed picture.
- Build vs. buyAI in businessThe decision whether to develop an AI solution in-house or purchase a ready-made product.
- Chain of thoughtPrompting & usageThe model is guided to solve a problem in traceable intermediate steps.
- Change management for AIAI in businessDeliberately guiding people and processes when AI changes the way work is done.
- Chatbot and AI assistantFundamentalsA program that converses with people in natural language.
- ChunkingData & infrastructureSplitting documents into smaller sections so that they can be searched for RAG.
- ClassificationMachine learningA learning task in which inputs are assigned to one of several categories.
- ClusteringMachine learningAutomatically grouping similar data points without knowing the groups in advance.
- Coding agentAgents & automationAn AI agent that independently writes, changes, tests and fixes program code.
- Computer useAgents & automationAI agents operate computers or web browsers like a human – via the screen, mouse and keyboard.
- Computer visionFundamentalsField of AI that enables computers to analyse and understand images and video.
- Content credentials and watermarksLaw, ethics & safetyTechnical methods that make the origin and editing of digital content verifiable.
- Context engineeringPrompting & usageDeliberately assembling all the information a model needs for a task – not just the prompt.
- Context windowLanguage modelsThe maximum amount of text (in tokens) a language model can take into account in a single request.
- ControlNetImage, video & audioA technique for precisely steering image generators with templates such as poses, edges or depth maps.
- Convolutional neural network (CNN)Machine learningA neural network that uses small filters to detect local patterns in images.
- CopilotAgents & automationAn AI assistant embedded in an application that supports your work while you stay in control.
- Copyright and AILaw, ethics & safetyQuestions about whether AI may be trained on protected works and who owns AI-generated content.
- Custom assistantsPrompting & usageSelf-configured variants of a chat assistant with fixed instructions, your own files and tools.
- Data annotationData & infrastructureLabelling data with the correct answer so that a model can learn from it.
- Data protection and AI (FADP, GDPR)Law, ethics & safetyAnyone processing personal data with AI must comply with data protection law – the FADP in Switzerland, the GDPR in the EU.
- Deep learningMachine learningMachine learning with deep neural networks made up of many stacked layers.
- Deep researchAgents & automationA research agent that independently searches and evaluates many sources and produces a report with citations.
- DeepfakeLaw, ethics & safetyAn AI-manipulated or AI-generated image, video or audio that convincingly shows real people or makes them speak.
- Diffusion modelImage, video & audioA generative model that creates images (or video, audio) by gradually removing noise.
- DistillationMachine learningA small “student” model learns to imitate the answers of a large “teacher” model.
- Edge AI (local AI)Data & infrastructureAI models run directly on the device or in your own network instead of in a provider’s cloud.
- EmbeddingData & infrastructureA representation of text, images or other data as a list of numbers (vector) that captures its meaning.
- Energy consumption of AIData & infrastructureTraining and running AI models require a lot of electricity and cooling in data centres.
- EU AI ActLaw, ethics & safetyThe European Union’s AI regulation – the first comprehensive law regulating AI.
- Explainability (explainable AI)Law, ethics & safetyThe goal of making it understandable why an AI system arrives at a particular result.
- Few-shotPrompting & usageGiving the model a few examples in the prompt so that it adopts the pattern and format.
- Fine-tuningMachine learningFurther training a pre-trained model on your own data to specialise it for a task or style.
- Foundation modelLanguage modelsA large, broadly pre-trained model that serves as the basis for many different applications.
- Frontier modelLanguage modelsTerm for the most capable AI models at the leading edge of development.
- Function calling (tool use)Agents & automationA language model deliberately calls external functions – such as a search, a calendar or a database.
- GAN (generative adversarial network)Image, video & audioTwo neural networks compete: one creates fakes, the other tries to expose them.
- Generative AIFundamentalsAI that creates new content – for example text, images, video, music or program code.
- GPTLanguage modelsShort for “generative pre-trained transformer” – a family of language models from OpenAI.
- GPUData & infrastructureA graphics processor that performs many calculations in parallel and is therefore the workhorse of AI.
- Gradient descentMachine learningAn optimisation method that gradually changes parameters in the direction in which the error decreases most.
- GuardrailsLaw, ethics & safetySafeguards that check an AI system’s inputs and outputs and prevent unwanted content or actions.
- HallucinationLanguage modelsA convincingly worded but false or fabricated statement by an AI model.
- High-risk AILaw, ethics & safetyAI systems that, under the EU AI Act, may pose significant risks to health, safety or fundamental rights.
- Human in the loopAgents & automationA human reviews, corrects or approves an AI system’s results at defined points.
- Image-to-imageImage, video & audioAn existing image is used as a starting point and changed or reinterpreted by AI.
- InferenceFundamentalsUsing a fully trained model to compute an output for a new input.
- Inference costsData & infrastructureThe running costs of using an AI model – usually per token or per image, minute or request.
- Inpainting and outpaintingImage, video & audioInpainting fills or replaces marked image areas; outpainting extends an image beyond its borders.
- Instruction tuningLanguage modelsPost-training with examples of instructions and good answers so that a model follows requests reliably.
- Intelligent document processingAI in businessAI reads invoices, contracts, forms or delivery notes and turns their content into structured data.
- Labelling AI contentLaw, ethics & safetyThe obligation or good practice of disclosing that content or a conversation partner is AI-generated.
- Large language model (LLM)Language modelsA large model trained on vast amounts of text that can understand and generate language.
- LatencyData & infrastructureThe waiting time between a request to an AI model and the start or end of its answer.
- Latent spaceImage, video & audioA compressed, abstract space in which a model represents the essential features of data.
- LoRA (low-rank adaptation)Machine learningAn efficient fine-tuning method that trains only small additional matrices instead of the whole model.
- Loss functionMachine learningA formula that measures how far a model’s prediction is from the desired result.
- Machine learning (ML)Machine learningMethods in which computers learn from example data instead of following hard-coded rules.
- Measuring AI value (ROI)AI in businessAssessing whether an AI deployment pays off – effort and cost compared with the benefit achieved.
- Memory in AI assistantsAgents & automationFeatures that let an AI assistant remember information beyond a single conversation.
- Mixture of experts (MoE)Language modelsA model architecture in which only some specialised sub-networks (“experts”) are active for each request.
- MLOps and LLMOpsData & infrastructureMethods and tools for reliably developing, deploying, monitoring and improving AI models.
- Model Context Protocol (MCP)Agents & automationAn open standard through which AI applications access external data sources and tools in a uniform way.
- Multi-agent systemAgents & automationSeveral specialised AI agents work together on a task.
- MultimodalityFundamentalsThe ability of an AI model to process several types of data such as text, images, audio or video.
- Narrow AIFundamentalsAI designed for a specific task or a clearly defined domain.
- Natural language processing (NLP)FundamentalsField of AI that enables computers to process and generate human language.
- Negative promptImage, video & audioA specification of what should explicitly not appear in a generated image.
- Neural networkMachine learningA brain-inspired computing model made of many connected units (“neurons”) that pass on weighted signals.
- Next-token predictionLanguage modelsThe basic principle of language models: they compute the most likely next token step by step.
- Open-weight and open-source modelsFundamentalsModels whose trained weights can be freely downloaded and run on your own infrastructure.
- OrchestrationAgents & automationControlling how models, tools, data and agents work together in an AI system.
- OverfittingMachine learningA model has memorised its training data and performs poorly on new data.
- ParametersFundamentalsThe learned numerical values in an AI model that determine its behaviour.
- Pre-trainingLanguage modelsThe first and most expensive training phase, in which a model learns from huge amounts of general data.
- PromptPrompting & usageThe input – question, instruction or material – that you give an AI model.
- Prompt engineeringPrompting & usageDeliberately writing and refining prompts to obtain better and more reliable results.
- Prompt injectionLaw, ethics & safetyAn attack in which hidden instructions in inputs cause an AI system to do something unwanted.
- Proof of concept (PoC)AI in businessA small, time-limited trial that shows whether an AI idea really works in your own environment.
- Reasoning modelLanguage modelsA language model that works through intermediate steps internally before answering, to solve complex tasks better.
- Red teamingLaw, ethics & safetyTargeted attack and misuse tests that find weaknesses in an AI system before deployment.
- RegressionMachine learningA learning task in which a model predicts a continuous numerical value.
- Reinforcement learningMachine learningLearning by trial and error: an agent tries actions and is rewarded for good outcomes.
- Responsible AILaw, ethics & safetyPrinciples for developing and using AI that is fair, transparent, safe and human-centred.
- Retrieval-augmented generation (RAG)Data & infrastructureA language model first looks up relevant information in your own sources and bases its answer on it.
- RLHFLanguage modelsReinforcement learning from human feedback: a model learns from human ratings which answers are better.
- Robotic process automation (RPA)Agents & automationSoftware “robots” that automate rule-based clicks and inputs in existing applications.
- Role promptPrompting & usageAssigning the model a role or perspective, such as “You are an experienced tax adviser”.
- SeedImage, video & audioA starting value for the random number generator that allows a generated result to be reproduced.
- Semantic searchData & infrastructureSearching by meaning instead of exactly matching words.
- Shadow AIAI in businessEmployees’ use of AI tools without the company’s knowledge or approval.
- Small language model (SLM)Language modelsA compact language model that can run on your own hardware, laptops or smartphones.
- Speech-to-text (speech recognition)Image, video & audioAutomatically converting spoken language into written text.
- Structured outputPrompting & usageThe model answers in a predefined format such as JSON so that software can process the answer directly.
- SuperintelligenceFundamentalsA thought model of an AI that clearly surpasses humans in almost every domain.
- Supervised learningMachine learningLearning from examples where the correct answer (the “label”) is already known.
- Symbolic AIFundamentalsAn AI approach that represents knowledge in explicit rules and symbols instead of learning it from data.
- Synthetic dataData & infrastructureArtificially generated data that reproduces the properties of real data.
- System promptPrompting & usageAn overarching instruction that defines an AI assistant’s behaviour for the entire conversation.
- TemperaturePrompting & usageA setting that determines how random or predictable a language model’s answers are.
- Text-to-imageImage, video & audioGenerating images from a text description.
- Text-to-speech (TTS)Image, video & audioConverting written text into natural-sounding spoken language.
- Text-to-videoImage, video & audioGenerating short video clips from a text description or a starting image.
- TokenLanguage modelsThe smallest unit of text a language model works with – often a word or part of a word.
- Top-p (nucleus sampling)Prompting & usageA method that limits the choice of the next token to the most likely candidates.
- TrainingMachine learningThe process in which a model learns from data by gradually adjusting its parameters.
- Training dataData & infrastructureThe examples an AI model learns from – text, images, measurements or other data.
- Transfer learningMachine learningAn already trained model is used as the starting point for a new, related task.
- TransformerLanguage modelsThe network architecture introduced in 2017 on which almost all of today’s language models are built.
- Turing testFundamentalsA test proposed by Alan Turing in 1950 to check whether a machine can be told apart from a human in conversation.
- Vector databaseData & infrastructureA database that stores embeddings and very quickly finds the entries closest in meaning.
- Vendor lock-inAI in businessStrong dependence on a single AI provider that makes switching later expensive or difficult.
- Vision-language model (VLM)Language modelsA model that understands images and text together and can, for example, answer questions about a photo.
- Voice cloningImage, video & audioUsing AI to reproduce a specific human voice from short voice recordings.