Knowledge

AI Glossary

Over 120 key terms in artificial intelligence, explained in plain English. From the fundamentals through language models and AI agents to the EU AI Act and AI in marketing.

This glossary explains the most important concepts in artificial intelligence in clear language, for anyone getting into the field or looking up individual terms. Jump straight to a topic:

Foundations & Key Concepts

Artificial Intelligence (AI): The broad field of building systems that perform tasks normally requiring human intelligence, such as understanding language, recognising patterns, or making decisions.
Machine Learning (ML): A branch of AI in which systems learn patterns from data instead of being explicitly programmed for every rule.
Deep Learning: A machine-learning approach based on many-layered neural networks that excels at large, unstructured data such as images, text, and audio.
Algorithm: A precise, step-by-step procedure for solving a problem; the foundation of every AI method.
Model: The knowledge a system learns from training data and uses to process new inputs.
Training: The process in which a model learns from data by adjusting its internal parameters.
Inference: Running a trained model on new data, the actual day-to-day use as opposed to training.
Parameters: The internal values (weights) a model learns; large language models have billions of them.
Narrow AI vs. General AI: Narrow (weak) AI handles specific, well-defined tasks; general AI would match human flexibility across any task and remains hypothetical.
AGI (Artificial General Intelligence): A hypothetical AI with human-level, general problem-solving ability across many domains.
ASI (Artificial Superintelligence): A hypothetical AI that would surpass human intelligence across virtually all fields.
Turing Test: A test proposed by Alan Turing: if a person cannot tell machine from human in conversation, the machine is deemed intelligent.
Symbolic AI: The classic rule- and logic-based approach to AI, in contrast to data-driven machine learning.
Data Science: The discipline that combines statistics, programming, and domain expertise to draw insights from data.

Machine Learning

Supervised Learning: Learning from labelled examples, inputs paired with their desired outputs.
Unsupervised Learning: Learning from unlabelled data to discover structure or groupings on its own.
Reinforcement Learning: Learning through reward and penalty while an agent interacts with an environment.
Self-Supervised Learning: The model creates its own training signal from the data, for example by predicting the next word, the basis of modern language models.
Classification: Predicting a category, such as spam or not spam.
Regression: Predicting a continuous value, such as a price.
Clustering: Automatically grouping similar data points without predefined categories.
Feature: A single measurable property of the input data.
Label: The correct answer a supervised model learns from.
Training, Validation & Test Data: Splitting data into sets to learn from, tune on, and independently evaluate a model.
Overfitting: When a model memorises the training data and fails on new, unseen data.
Underfitting: When a model is too simple to capture the patterns in the data.
Bias-Variance Tradeoff: The balance between models that are too simple (bias) and too complex (variance).
Cross-Validation: A technique that splits the data multiple ways to estimate model quality more reliably.
Feature Engineering: Preparing and constructing meaningful features from raw data.
Decision Tree & Random Forest: Rule-based models, and their combination into a more robust forest of many trees.

Neural Networks & Deep Learning

Neural Network: A model inspired by biological neurons, built from interconnected nodes arranged in layers.
Neuron (Node): A compute unit that takes weighted inputs, sums them, and passes the result through an activation function.
Layer: A level of neurons; deep networks have many of them.
Weight: The learned strength of a connection between two neurons.
Activation Function: A non-linear function such as ReLU that lets a network learn complex relationships.
Backpropagation: The method that sends the error backward through the network to update the weights.
Gradient Descent: An optimisation method that nudges the weights step by step toward a lower error.
Loss Function: A measure of how far a prediction is from the target value.
Epoch: One complete pass through all the training data.
CNN (Convolutional Neural Network): An architecture built for images and spatial patterns.
RNN & LSTM: Networks for sequences such as text or time series; LSTMs retain longer-range context.
Transformer: The architecture behind modern language models; it processes sequences in parallel via attention.
Attention: A mechanism by which a model weighs which parts of the input matter for the current output.
GAN (Generative Adversarial Network): Two competing networks, a generator and a discriminator, that together produce realistic data.
Autoencoder: A network that compresses and then reconstructs data to learn compact representations.
Diffusion Model: A generative method that turns random noise into an image or audio step by step, the basis of many image generators.

Generative AI & Language Models

Generative AI: AI that creates new content: text, images, audio, video, or code.
Large Language Model (LLM): A model trained on vast amounts of text that generates human-like language.
Foundation Model: A large, broadly pre-trained model that serves as the base for many downstream applications.
GPT: Short for Generative Pretrained Transformer, the model family behind many well-known chatbots.
Token: The smallest unit a language model processes, usually a piece of a word that text is assembled from.
Tokenisation: Breaking text into tokens.
Embedding: Representing words, sentences, or images as a numeric vector in which similar things sit close together.
Context Window: The maximum amount of text, measured in tokens, that a model can consider at once.
Temperature: A setting that controls how creative or random a model’s outputs are.
Multimodal Model: A model that understands or generates several data types at once, such as text, image, audio, and video.
Hallucination: A confident-sounding but factually wrong or fabricated output from a language model.
Reasoning Model: A language model that explicitly thinks before answering, producing intermediate steps to solve harder problems more reliably.
Mixture of Experts (MoE): An architecture where only relevant sub-networks (experts) activate per request, making large models more efficient.
Text-to-Image / Text-to-Video: Generative systems that create images or video from a text description.
Text-to-Speech / Speech-to-Text: Converting text into spoken audio and vice versa.
Open-Weights Model: A model whose weights are freely available and can be run locally, distinct from fully open-source models that also release training data and code.

Prompting & Interaction

Prompt: The input or instruction you give a model to steer it.
Prompt Engineering: The craft of phrasing prompts so a model produces useful, reliable results.
System Prompt: A higher-level instruction that sets a model’s role and behaviour for the whole conversation.
Zero-Shot / Few-Shot: Solving a task with no examples, or with a few examples included in the prompt.
In-Context Learning: A model’s ability to pick up a task purely from examples in the prompt, without being retrained.
Chain-of-Thought: Prompting a model to reason step by step to tackle complex problems.
RAG (Retrieval-Augmented Generation): A method where a model retrieves relevant documents at query time and grounds its answer in them, reducing hallucinations.
Grounding: Anchoring answers in reliable, verifiable sources.
Prompt Injection: An attack in which hidden instructions trick a model into unwanted behaviour.
Guardrails: Rules and filters that keep a model from producing unsafe or unwanted output.

AI Agents & Automation

Agentic AI: AI systems that pursue goals on their own, planning, using tools, and carrying out multi-step tasks.
AI Agent: A program with a language model at its core that makes decisions and takes actions autonomously.
Function Calling (Tool Use): A model’s ability to invoke external tools or APIs in a targeted way.
MCP (Model Context Protocol): An open standard that lets AI applications connect to external tools and data sources in a uniform way.
Multi-Agent System: Several cooperating AI agents that split a task among themselves.
Orchestration: Coordinating multiple AI steps, tools, or agents into one end-to-end workflow.
Workflow Automation: Wiring AI into other systems, for example via n8n or Zapier, to run processes automatically.
Test-Time Compute: Extra compute a model spends thinking at run time to produce better answers.
Human-in-the-Loop: A workflow where a person reviews, approves, or corrects AI decisions.
Copilot (Assistant): AI that assists people with tasks rather than replacing them entirely.

Training & Optimization

Pre-Training: A model’s first, broad training run on large, general datasets.
Fine-Tuning: Further training a base model on specific data or tasks.
Transfer Learning: Reusing knowledge from a pre-trained model for a new task.
RLHF (Reinforcement Learning from Human Feedback): Aligning models with human preferences by having people rate outputs.
DPO (Direct Preference Optimization): A newer method that aligns models directly on preferred answers.
LoRA / QLoRA: Resource-efficient fine-tuning methods that train only a small number of extra parameters.
Distillation: A large teacher model transfers its knowledge to a smaller, faster student model.
Quantisation: Lowering a model’s numeric precision to save memory and run it faster, for example on local hardware.
Synthetic Data: Artificially generated training data that mimics real data without using actual personal information.
Data Annotation (Labelling): Tagging raw data with correct answers for supervised learning.
Alignment: The effort to make AI behaviour consistent with human values and intent.

Data & Infrastructure

Training Data: The data a model learns from; its quality strongly shapes the outcome.
Vector Database: A store for embeddings that enables fast similarity search, central to RAG.
GPU / TPU: Specialised processors that accelerate the compute-heavy training and running of AI.
Edge AI / On-Device AI: AI that runs locally on the device instead of in the cloud, faster and more privacy-friendly.
Local Model: A model run on your own hardware, for example via Ollama or LM Studio, without sending data to third parties.
API: An interface through which applications call AI services.
Inference Cost: The compute and monetary cost per model request in production.
Compute: The pooled processing power on which large AI models are trained and served.
AI Sovereignty: Building your own AI infrastructure and models to stay independent of large providers and keep control of your data.

Evaluation & Quality

Benchmark: A standardised test used to compare models on equal footing.
Accuracy: The share of predictions that are correct overall.
Precision & Recall: Precision is how many of the hits are correct; recall is how many of the relevant cases are found.
F1 Score: A combined measure of precision and recall.
Perplexity: A measure of how surprised a language model is by a text; lower is better.
Evaluation: Systematically checking model quality, often with test data or human judgement.
Explainable AI (XAI): Methods that make an AI’s decisions understandable to people.
Red Teaming: Deliberately probing a model for weaknesses and failure modes.
Confidence: How sure a model is that an output is correct, not to be confused with actual correctness.

Ethics, Law & Governance

EU AI Act: The European Union’s AI regulation, which classifies AI systems by risk and sets obligations for providers and deployers.
AI Literacy: Under the EU AI Act, the duty to ensure staff can use AI competently and safely.
GDPR & AI: The data-protection requirements for handling personal data in AI systems.
AI Governance: The organisational framework of rules, roles, and processes for responsible AI use.
Bias: Systematic skew in data or models that leads to unfair outcomes.
Fairness: The goal that AI systems do not unjustly disadvantage particular groups.
Transparency: Disclosing when and how AI is used and how it reaches its results.
Deepfake: AI-generated, deceptively real fakes of image, video, or voice.
Copyright & AI: Legal questions around training data and the use of AI-generated content.
Chief AI Officer (CAIO): A leadership role that owns an organisation’s AI strategy, governance, and compliance.
Watermarking: Usually-invisible marking of AI-generated content to make it traceable.

AI in Search & Marketing

Generative Engine Optimization (GEO): Optimising content for AI answer systems such as ChatGPT, Perplexity, or Google AI Overviews.
AI Overviews: AI-generated answer summaries shown directly in Google Search.
Answer Engine: A system that returns direct answers instead of just listing links, such as Perplexity.
AI Crawlers: Bots that AI providers use to collect web content for training and answers, such as GPTBot, ClaudeBot, or PerplexityBot.
llms.txt: A proposed standard by which websites tell AI systems, in a structured way, which content they may use.
AI Visibility: How prominently and how accurately a brand appears in AI-generated answers.
Structured Data (Schema): Machine-readable markup of content that helps AI and search engines understand it.

Putting AI to work in your organisation?

From EU AI Act classification and AI training to automation with AI, minoka supports you hands-on.

AI consulting