Glossary
86 terms in plain language. Star the ones you keep having to explain, they are usually the ones that decide the room.
86
- Artificial intelligence (AI)
- Systems designed to perform tasks associated with human intelligence, such as prediction, classification, generation, or planning.
- Machine learning
- A method in which a system learns patterns from examples instead of receiving every rule explicitly.
- Model
- A learned mathematical system that maps inputs to outputs.
- Large language model (LLM)
- A model trained on large collections of text to predict and generate language.
- Generative AI
- AI that produces new text, images, audio, code, or other outputs.
- Agentic system / AI agent
- A model connected to tools, memory, instructions, and action loops so it can complete multiple steps.
- Automation
- Using a system to perform a task with limited direct human intervention.
- Augmentation
- Using AI to support rather than replace human work.
- Human-in-the-loop
- A person reviews, approves, or intervenes in an AI-assisted process.
- Human-on-the-loop
- A person supervises a system without reviewing every action.
- Human-out-of-the-loop
- A system acts without a person reviewing each decision or action.
- Training data
- Examples used to adjust a model's parameters during training.
- Dataset
- An organised collection of data used for training, testing, or analysis.
- Data provenance
- Information about where data came from and how it was collected or changed.
- Data lineage
- The history of how data moves through systems and transformations.
- Structured data
- Data organised into defined fields, rows, categories, or schemas.
- Unstructured data
- Material such as text, images, audio, or video without a fixed tabular structure.
- Synthetic data
- Artificially generated data designed to resemble real data.
- Ground truth
- A reference label or outcome treated as correct for evaluation.
- Knowledge base
- A curated collection of information that a system can retrieve.
- Parametric knowledge
- Information encoded in a model's learned parameters.
- Factual knowledge
- Information about facts or relationships that a model may have learned from its training data.
- Retrieval-augmented generation (RAG)
- A system retrieves documents or records before generating an answer.
- Local knowledge
- Knowledge grounded in the experience of a particular place or community.
- Indigenous knowledge
- Knowledge systems developed and maintained by Indigenous peoples and communities.
- Data sovereignty
- The principle that data is governed according to the laws, rights, and authority of the people or place it concerns.
- Data minimisation
- Collecting and retaining only the data necessary for a stated purpose.
- Sensitive data
- Data whose disclosure or misuse could cause harm.
- Input / prompt
- The instruction or material given to a model.
- Output
- The text, prediction, classification, image, or action produced by a system.
- Inference
- Running a trained model to produce an output.
- Context window
- The amount of text or other material a model can consider at one time.
- Token
- A unit of text or other data processed by a language model.
- Embedding
- A numerical representation of text, images, or other data used to compare relationships.
- Classification
- Assigning an item to one or more categories.
- Prediction
- Estimating a likely outcome from available data.
- Forecast
- A prediction about a future state or event.
- Recommendation
- An output that suggests a course of action.
- Optimisation
- Searching for the best result according to a defined objective and constraints.
- Objective function
- The formal quantity a system tries to maximise or minimise.
- Proxy
- An indirect measure used in place of a harder-to-measure goal.
- Benchmark
- A test used to compare model performance.
- Evaluation
- Measuring a model or system against defined criteria.
- Uncertainty
- The degree to which an output may be wrong or incomplete.
- Confidence score
- A numerical estimate associated with a prediction or classification.
- Calibration
- The relationship between predicted confidence and actual accuracy.
- Explainability
- The ability to provide an understandable account of how an output was produced.
- Interpretability
- The extent to which a system's internal behaviour can be understood.
- Black box
- A system whose internal operation is difficult to inspect.
- Hallucination
- A fluent but unsupported, false, or fabricated model output.
- Bias
- Systematic error or unequal performance associated with data, design, or use.
- Fairness
- Approaches for evaluating and addressing unequal treatment or outcomes.
- Robustness
- The ability to perform reliably under changed or imperfect conditions.
- Adversarial attack
- An input deliberately designed to cause a system to fail or behave unexpectedly.
- Neural network
- A computational system made of connected layers that transform inputs into outputs.
- Perceptron
- An early mathematical model of a single brain cell, and a foundation for later neural-network research.
- Transformer
- A neural-network architecture that uses attention to process relationships among tokens.
- Weights
- The numbers a model learned during training. In practice, the weights are the model.
- Open weights
- Weights published so that anyone can run, inspect, or host the model themselves.
- Closed weights
- Weights kept on the provider's servers and reachable only through their service.
- Pretraining
- Initial training on a large body of data, usually by predicting missing or subsequent content.
- Fine-tuning
- Further training on a narrower dataset or task.
- Post-training
- Processes applied after pretraining to improve usefulness, behaviour, or alignment.
- Reinforcement learning (RL)
- Training through feedback that rewards some actions and penalises others.
- Reinforcement learning from human feedback (RLHF)
- Reinforcement learning using human preferences or evaluations.
- Reasoning model
- A model or system optimised to spend additional computation on multi-step problems.
- Chain of thought
- Intermediate reasoning text generated during problem solving.
- Scaling laws
- Observed relationships between model performance, model size, training data, and computing power.
- Mixture of experts (MoE)
- An architecture containing multiple specialist subnetworks, only some of which are used for each input.
- Alignment
- Designing a system so its behaviour reflects specified goals or values.
- AI safety
- Research and practices aimed at reducing harmful or uncontrolled system behaviour.
- AI governance
- The rules, institutions, processes, and practices that shape AI development and use.
- Responsible AI
- Developing and using AI with attention to social, ethical, legal, and environmental impacts.
- Value-sensitive design
- Designing technology with explicit attention to human values and affected interests.
- Sovereignty-aligned AI
- AI designed and governed to preserve the agency of a defined community or place.
- Fine-grained permissions
- Restrictions on which tools, files, data, or actions a system can access.
- Sandbox
- An isolated environment in which a system can operate with limited access.
- Human approval gate
- A required human authorisation before an action proceeds.
- Audit log
- A record of system inputs, outputs, actions, and changes.
- Model card
- Documentation describing a model's intended use, limitations, and evaluation.
- System card
- Documentation covering a broader deployed system, including risks and mitigations.
- Model drift
- A change in performance as data, behaviour, or conditions change over time.
- Versioning
- Keeping track of changes to models, data, prompts, and software.
- Rollback
- Reverting a system or deployment to an earlier version.
- Self-hosting
- Running a model on hardware you or your institution control.
- Zero data retention
- A contractual promise that a provider stores nothing you send it.