AI moves fast. New terms surface weekly, and it's hard to know which ones matter. We built this glossary as a practical reference — the concepts our clients ask about most, organised into modules you can scan in minutes.
The core hierarchy
Before diving into specifics, understand how the foundational concepts nest together.
- Artificial Intelligence (AI): Systems that perform tasks typically requiring human cognition — reasoning, learning, perception, decision-making.
- Machine Learning (ML): An AI subset where systems learn patterns from data rather than following explicit programmed rules.
- Deep Learning (DL): Using multi-layered neural networks to learn hierarchical representations from large datasets.
- Neural Network: A computational model using interconnected nodes (neurons) organised in layers to process information.
1. AI Agents
- AI Agent: A system that perceives its environment, reasons, and takes autonomous actions toward goals.
- Agentic AI: AI capable of multi-step planning, tool use, and independent execution.
- Tool Use: AI's ability to invoke external functions, APIs, or services to complete tasks.
- Chain-of-Thought (CoT): A prompting technique that elicits step-by-step reasoning.
- RAG (Retrieval-Augmented Generation): Combining retrieval from external knowledge with generation for grounded outputs.
- MCP (Model Context Protocol): An open standard for connecting AI models to external tools and data sources.
- Agent Memory: Persistent context that lets agents remember past interactions across sessions.
- Agentic Workflow: A sequenced multi-step process where agents coordinate tasks autonomously.
- Computer Use: AI's ability to control a computer interface (mouse, keyboard, screen) directly.
- Spec-Driven Development: Writing detailed specifications that agents execute, instead of writing code.
- Human-in-the-Loop (HITL): A governance pattern where humans review and approve AI decisions at critical steps.
- LLM Orchestrator: A framework (e.g., LangChain, Semantic Kernel) that chains models, tools, and APIs into coordinated pipelines.
- Prompt Chaining: Taking the output of one prompt and feeding it as input to the next to solve multi-step problems.
2. Model Architecture
- Transformer: Architecture using self-attention mechanisms; the foundation of modern LLMs.
- Large Language Model (LLM): A neural network trained on massive text data to understand and generate language.
- Foundation Model: A large model trained on broad data, adaptable to many downstream tasks.
- Parameters: Learnable weights in a model that adjust during training.
- Hyperparameters: Configuration settings fixed before training (learning rate, batch size, etc.).
- Training Data: The dataset used to teach a model patterns and relationships.
- Epoch: One complete pass through the entire training dataset.
- Batch Size: The number of samples processed before updating model parameters.
- Loss Function: Measures the difference between predicted and actual outputs; guides optimisation.
- Gradient Descent: An optimisation algorithm that iteratively adjusts parameters to minimise loss.
- Backpropagation: An algorithm for computing gradients by propagating errors backward through the network.
- Reasoning Model: LLMs specifically trained to "think" step-by-step before answering.
- Small Language Model (SLM): Compact, specialised models optimised for specific tasks and edge deployment.
3. Generative AI
- Generative AI: AI that creates new content — text, images, audio, video, code.
- Prompt: Input text or instruction that directs a generative model's output.
- Token: The basic unit of text processing (word, subword, or character).
- Context Window: The maximum token length a model can process in one inference.
- Temperature: A parameter controlling output randomness; higher = more creative, lower = more deterministic.
- Inference: Using a trained model to generate predictions on new inputs.
- Diffusion Model: A generative model that learns to reverse a noise-adding process to create data.
- Autoregressive Model: Generates output sequentially, each token conditioned on previous ones.
- Prompt Engineering: The skill of crafting inputs to reliably elicit desired outputs from generative models — part art, part science.
4. Learning Paradigms
- Supervised Learning: Training on labelled data where inputs are paired with correct outputs.
- Unsupervised Learning: Training on unlabelled data to discover hidden patterns or structures.
- Reinforcement Learning (RL): Learning through trial and error, optimising actions based on reward signals.
- Self-Supervised Learning: The model generates its own labels from input data (e.g., predicting masked words).
- Transfer Learning: Applying knowledge from one task or domain to improve performance on another.
- Fine-Tuning: Adapting a pre-trained model to a specific task using targeted training.
- Zero-Shot Learning: Performing tasks without task-specific training examples.
- Few-Shot Learning: Learning from very limited examples (typically 1–10).
- LoRA (Low-Rank Adaptation): An efficient fine-tuning method that updates a small number of parameters, making adaptation dramatically cheaper and faster.
- RLAIF (Reinforcement Learning from AI Feedback): Using a "teacher" model to generate preference signals, reducing dependence on expensive human labellers.
5. NLP & Language
- Natural Language Processing (NLP): AI focused on understanding, interpreting, and generating human language.
- Embedding: A dense vector representation capturing semantic meaning of text, images, or other data.
- Semantic Search: Search based on meaning and intent rather than keyword matching.
- Tokenisation: Splitting text into tokens for model processing.
- Attention Mechanism: Allows models to weigh the importance of different input parts when producing output.
- RLHF (Reinforcement Learning from Human Feedback): A training method using human preferences to align model behaviour.
6. Vision & Multimodal
- Computer Vision: AI that interprets and analyses visual information from images and video.
- Multimodal AI: Models that process and relate multiple data types (text, image, audio).
- OCR (Optical Character Recognition): Converting images of text into machine-readable text.
- Image Segmentation: Partitioning an image into distinct regions or objects.
- Object Detection: Identifying and locating objects within images.
7. Model Behaviour & Problems
- Hallucination: A model generating plausible but factually incorrect or fabricated content.
- Overfitting: A model memorises training data, failing to generalise to new inputs.
- Underfitting: A model too simple to capture underlying patterns in data.
- Bias: Systematic errors or unfair outcomes due to flawed data or design.
- Alignment: Ensuring AI systems behave according to human values and intentions.
- Jailbreaking: Attempts to bypass a model's safety constraints through adversarial prompts.
- Prompt Injection: Malicious inputs designed to override a model's instructions.
- Drift: Gradual degradation of model performance as real-world data diverges from training data.
- Prompt Leaking: Unintentional exposure of system prompts or confidential instructions through user queries.
- Automation Bias: The human tendency to over-trust AI outputs, even when they are clearly wrong — a governance risk that goes beyond hallucination.
8. Evaluation & Metrics
- Benchmark: A standardised test for comparing model performance.
- Accuracy: The proportion of correct predictions out of total predictions.
- Precision: Of positive predictions, the proportion that are actually correct.
- Recall: Of actual positives, the proportion correctly identified.
- F1 Score: The harmonic mean of precision and recall.
- Perplexity: Measures how well a language model predicts text; lower = better.
- BLEU Score: A metric for evaluating generated text against reference translations.
- Ground Truth: The verified correct answer used for training or evaluation.
9. Safety & Ethics
- Explainable AI (XAI): Methods making AI decision-making interpretable to humans.
- Red Teaming: Adversarial testing to identify vulnerabilities and failure modes.
- Constitutional AI: A training approach using principles to guide model behaviour.
- Guardrails: Constraints preventing harmful or undesired model outputs.
- Data Privacy: Protecting personal information used in AI training and inference.
- Model Card: Documentation describing a model's capabilities, limitations, and intended use.
- Sandboxing: Isolating AI systems so their actions can't affect production environments without explicit permission.
- Copyright & IP Infringement: The ongoing legal debate over whether training on publicly available data constitutes "fair use."
- AI Sovereignty: The movement for nations or organisations to host their own models and data locally, reducing dependence on a handful of global providers.
10. Infrastructure & Deployment
- API (Application Programming Interface): An interface allowing software to interact with AI models programmatically.
- Latency: Time delay between input and model response.
- Throughput: The number of requests a system can handle per unit time.
- GPU (Graphics Processing Unit): Hardware accelerating parallel computations for AI training and inference.
- TPU (Tensor Processing Unit): Google's custom AI accelerator chips.
- Quantisation: Reducing model precision (e.g., 32-bit to 8-bit) to decrease size and speed inference.
- Distillation: Training smaller models to mimic larger ones, preserving performance.
- Edge AI: Running AI models locally on devices rather than cloud servers.
- Vector Database: A specialised database that stores and retrieves high-dimensional embeddings (powers RAG).
- Inference Cost: The compute cost of running a model per query — the real ongoing expense of AI in production.
- Vibe Coding: Building software by describing what you want in natural language, rather than writing code.
- NPU (Neural Processing Unit): Dedicated AI accelerator built into modern laptops and smartphones, enabling on-device Edge AI.
- KV Cache: A memory optimisation technique that stores previously computed key-value pairs, making LLM text generation significantly faster.
- Compute Governance: How organisations manage, allocate, and budget their limited GPU and cloud-compute resources.
11. Advanced Concepts
- Latent Space: A compressed representation space where similar concepts cluster together.
- Emergent Abilities: Capabilities appearing in large models not present in smaller versions.
- Scaling Laws: Predictable relationships between model size, data, compute, and performance.
- In-Context Learning: A model adapts to tasks from examples provided in the prompt, without weight updates.
- Synthetic Data: Artificially generated data used for training or augmentation.
- Mixture of Experts (MoE): An architecture where different subnetworks specialise in different inputs.
- AGI (Artificial General Intelligence): Hypothetical AI matching human-level reasoning across all cognitive domains.
- ASI (Artificial Superintelligence): Theoretical AI surpassing human intelligence in all areas.
- World Model: AI's internal representation of how the world works, enabling prediction and planning.
Keep learning
This glossary is a living document — we update it as the landscape shifts. If you're ready to move from terminology to implementation, start with Discovery and we'll diagnose where you sit on the AI maturity curve and what the right next move looks like for your firm.