Artificial Intelligence
Software that performs tasks normally associated with human cognition, such as recognising patterns, generating language, or making predictions.
Reference // Plain Language
AI Glossary
The useful meaning behind the vocabulary, without the ceremonial fog. Every term an AI system uses, groups by where it sits in the stack.
Section 01
Core Concepts
Machine Learning
An approach where a system learns patterns from data instead of following rules written by hand.
Deep Learning
Machine learning using neural networks with many layers, which learn useful features directly from raw data.
Neural Network
A layered arrangement of simple mathematical units whose connection strengths are adjusted during training.
Model
The trained artefact, meaning an architecture plus its learned parameters, that turns input into output.
Algorithm
A defined procedure. In machine learning it usually means the method used to train or run a model, not the model itself.
Training
Adjusting a model's parameters so its outputs better match the examples it is shown.
Inference
Using a trained model to generate a prediction or output from new input.
Parameters
The learned numerical values that shape how a model transforms input into output.
Weights
The individual parameter values on connections between units, often used interchangeably with parameters.
Supervised Learning
Training on examples that are labelled with the correct answer.
Unsupervised Learning
Finding structure in data that carries no labels.
Self-Supervised Learning
Training where the labels come from the data itself, such as predicting a hidden next word.
Reinforcement Learning
Training by trial and feedback, where actions that earn reward become more likely.
Transfer Learning
Reusing a model trained on one task as the starting point for another.
Section 02
Architecture and Training
Transformer
The architecture behind most modern language models, built around attention rather than strictly sequential processing.
Attention
A mechanism that lets a model weigh which parts of the input matter most for each part of the output.
Foundation Model
A large model trained broadly, then adapted to many downstream tasks.
Pre-Training
The initial large-scale training phase that builds general capability.
Post-Training
Everything done after pre-training to shape behaviour, including instruction tuning and preference optimisation.
Fine-Tuning
Additional training that adjusts a model toward a narrower task, style, or domain.
Instruction Tuning
Fine-tuning on instruction and response pairs so a model follows directions rather than merely continuing text.
RLHF
Reinforcement learning from human feedback, where human preference ratings train a reward signal used to shape the model.
DPO
Direct preference optimisation, which tunes a model on preferred versus rejected pairs without training a separate reward model.
LoRA
Low-rank adaptation, which fine-tunes a small set of added parameters instead of the whole model.
Gradient Descent
The optimisation method that nudges parameters in the direction that reduces error.
Loss Function
The measure of how wrong a prediction is, which training tries to minimise.
Epoch
One complete pass through the training data.
Batch Size
How many examples the model processes before each parameter update.
Learning Rate
How large a step the training process takes when updating parameters.
Overfitting
When a model memorises its training data and performs worse on anything new.
Underfitting
When a model is too simple or undertrained to capture the real pattern.
Checkpoint
A saved snapshot of a model's parameters partway through training.
Quantization
Reducing the numeric precision of parameters to shrink a model and speed up inference.
Distillation
Training a smaller model to reproduce the behaviour of a larger one.
Mixture of Experts
An architecture that routes each input to a subset of specialised sub-networks rather than the whole model.
Section 03
Language and Generation
Large Language Model
A model trained on large text collections to predict and generate token sequences.
Small Language Model
A compact model designed to run cheaply, often on local or edge hardware.
Token
A chunk of text processed by a language model, often part of a word rather than a whole word.
Tokenization
Splitting text into the tokens a model can process.
Context Window
The amount of input and generated material a model can consider at one time.
Autoregressive
Generating output one token at a time, each conditioned on what came before.
Temperature
A setting that controls randomness. Lower values make output more predictable.
Top-P Sampling
Choosing the next token from the smallest set whose probabilities sum to a threshold. Also called nucleus sampling.
Top-K Sampling
Choosing the next token from the K most likely candidates.
Logits
The raw scores a model assigns to possible next tokens before they become probabilities.
Perplexity
A measure of how surprised a model is by text. Lower generally means a better fit.
Embedding
A numerical representation that places related items near one another in a learned space.
Vector
An ordered list of numbers. Embeddings are vectors.
Stop Sequence
A string that tells the model to end its output.
Max Tokens
A cap on how much output a model may generate in one response.
Streaming
Returning output progressively as it is generated rather than all at once.
Section 04
Prompting and Interaction
Prompt
The instructions, context, examples, and material supplied to a model.
System Prompt
Instructions that set a model's role, rules, and boundaries for a whole session.
Prompt Engineering
Designing prompts deliberately, treating the request as a specification rather than a wish.
Zero-Shot
Asking a model to perform a task with no worked examples.
Few-Shot
Supplying a small number of worked examples to demonstrate the pattern you want.
Chain of Thought
Prompting a model to work through intermediate steps before committing to an answer.
Structured Output
Requiring output in a defined shape such as JSON so another system can consume it safely.
Grounding
Tying an answer to supplied source material rather than the model's own recall.
Context Engineering
Deciding what information goes into the context window, in what order, and what is left out.
Role Prompting
Telling the model what perspective or expertise to adopt.
Section 05
Retrieval, Memory, and Data
Retrieval
Fetching external material for a model to use while answering.
Retrieval-Augmented Generation
Combining retrieval with generation so answers rest on supplied documents. Usually shortened to RAG.
Vector Database
A store that indexes embeddings so similar items can be found quickly.
Semantic Search
Searching by meaning rather than exact keyword match.
Chunking
Splitting documents into passages small enough to retrieve and fit in context.
Reranking
Reordering retrieved candidates with a more precise model before they are used.
Knowledge Graph
A structured network of entities and the relationships between them that a system can query.
Memory
State carried between turns or sessions so an assistant can refer back to earlier context.
Training Data
The examples used to fit a model's parameters.
Synthetic Data
Machine-generated training data, used when real data is scarce, expensive, or sensitive.
Data Provenance
The documented origin, licensing, and handling history of a dataset.
Ground Truth
The verified correct answer used to judge a model's output.
Section 06
Agents, Tools, and Orchestration
Agent
A system that uses a model in a loop with tools, state, and a goal.
Tool Use
Letting a model call external functions, APIs, or services to act beyond generating text.
Function Calling
A structured mechanism by which a model requests a named tool with typed arguments.
Model Context Protocol
An open protocol for connecting models to tools and data sources in a consistent way. Usually shortened to MCP.
Orchestration
Coordinating models, tools, and steps into a reliable workflow.
Multi-Agent System
Several agents with distinct roles cooperating on a single task.
Control Loop
The cycle in which an agent plans, acts, observes the result, and decides what to do next.
Human in the Loop
Requiring a person to review or approve before a consequential action proceeds.
Guardrails
Enforced constraints on what a system may accept, generate, or do.
Sandbox
An isolated environment that limits what code or an agent can reach.
Idempotency
Designing an action so repeating it causes no extra effect, which matters because agents retry.
Section 07
Other AI Domains
Natural Language Processing
The field concerned with machines working with human language.
Computer Vision
Extracting meaning from images and video.
Speech Recognition
Converting spoken audio into text. Also called automatic speech recognition.
Text to Speech
Generating spoken audio from written text.
Optical Character Recognition
Extracting text from images or scanned documents.
Multimodal
A system that works across more than one data type, such as text, images, or audio.
Diffusion Model
A generative model that produces images or audio by repeatedly removing noise.
Generative Adversarial Network
A pair of models trained against each other, one generating and one judging.
Recommender System
A system that ranks items by predicted relevance to a particular person.
Anomaly Detection
Identifying data points that deviate from an expected pattern.
Classical Machine Learning
Non-neural methods such as decision trees and gradient boosting, still the better tool for much tabular data.
Section 08
Evaluation and Reliability
Hallucination
A plausible-looking output that is unsupported, incorrect, or invented.
Evaluation
Structured measurement of whether a system meets defined criteria. Often shortened to eval.
Benchmark
A standard dataset and scoring method used to compare models.
Accuracy
The share of predictions that are correct. Misleading when one class dominates.
Precision
Of the items flagged positive, how many truly were.
Recall
Of the items that truly were positive, how many were found.
F1 Score
The harmonic mean of precision and recall, used when both matter.
Confusion Matrix
A table of correct and incorrect predictions broken down by class.
Regression Test
A repeatable check that a behaviour fixed once has not broken again.
Drift
Gradual divergence between live data and what a model was trained on, degrading performance over time.
Observability
Being able to see what a system did and why, through logs, traces, and metrics.
A/B Test
Comparing two variants on live traffic to measure which performs better.
Section 09
Infrastructure and Operations
GPU
A processor with many parallel cores, the standard hardware for training and serving models.
TPU
Google's custom accelerator designed specifically for machine learning workloads.
Accelerator
Any specialised hardware that speeds up model computation.
FLOPs
Floating point operations, a measure of computational work.
VRAM
Memory on an accelerator. It usually determines the largest model you can run.
Latency
How long a single request takes to return.
Throughput
How much work a system completes per unit of time.
Inference Endpoint
The network address where a served model accepts requests.
Serving
Running a trained model in production so it can answer requests.
MLOps
The practices for deploying, monitoring, and maintaining models in production.
GenAIOps
The same discipline applied to generative systems, covering prompts, retrieval, and evaluation as well as models.
Batch Inference
Processing many inputs together offline rather than one at a time in real time.
Caching
Reusing prior computation or results to cut cost and latency.
Rate Limit
A cap on how many requests a client may make in a given period.
Section 10
Safety, Security, and Governance
Responsible AI
Practices that keep systems fair, transparent, accountable, and safe.
AI Governance
The policies, roles, and controls by which an organisation manages AI risk.
Alignment
Work to make a model's behaviour match human intent and stated values.
Bias
Systematic error that produces unfair or skewed outcomes for particular groups.
Fairness
Whether outcomes are equitable across groups, measured explicitly rather than assumed.
Explainability
The ability to give a human-understandable account of why a system produced an output.
Interpretability
The degree to which a model's internal workings can be understood directly.
Transparency
Disclosing how a system works, what data it used, and where its limits are.
Model Card
A published summary of a model's intended use, training data, performance, and limitations.
Prompt Injection
An attack where hostile instructions hidden in content cause a model to ignore its own rules.
Jailbreak
A prompt crafted to bypass a model's safety constraints.
Data Poisoning
Corrupting training data so a model learns attacker-chosen behaviour.
Model Inversion
Extracting information about training data by probing a model's outputs.
Membership Inference
Determining whether a specific record was part of a model's training data.
Differential Privacy
A mathematical guarantee limiting how much any single record can influence a result.
Red Teaming
Deliberately attacking a system to find failures before an adversary does.
Shadow AI
Unapproved AI tools used inside an organisation without oversight.
Least Privilege
Granting only the minimum access required, which bounds the damage when something is compromised.
EU AI Act
European legislation that regulates AI systems according to their risk category.
NIST AI Risk Management Framework
A voluntary United States framework for identifying, measuring, and managing AI risk.