AI glossary

45 terms

AgentAgents

A system in which a model repeatedly chooses actions, observes the results and continues toward a goal.Go deeper: Agent

AlignmentSafety

Making an AI system pursue the goals and values its designers and users actually intend.Go deeper: Alignment

BenchmarkEvaluation

A fixed test set used to compare models. Useful for trends, weak evidence for your specific task.Go deeper: Benchmark

Chain-of-thoughtTraining

Having a model write intermediate reasoning steps before its answer, which often improves multi-step problems.Go deeper: Chain-of-thought

ChunkingBuilding

Splitting documents into passages before embedding them. Chunk size strongly affects retrieval quality.Go deeper: Chunking

Constitutional AISafety

Training a model to critique and revise its own outputs against a written set of principles.Go deeper: Constitutional AI

ContaminationEvaluation

When test questions leaked into training data, inflating scores without real capability.Go deeper: Contamination

Context windowFoundations

The maximum number of tokens a model can consider at once, including the prompt and its own output.Go deeper: Context window

DistillationTraining

Training a smaller model to imitate a larger one, keeping much of its quality at lower cost.Go deeper: Distillation

DPOTraining

Direct preference optimisation: learning from preference pairs directly, without a separate reward model.Go deeper: DPO

EmbeddingFoundations

A list of numbers that represents the meaning of text, an image or other data, so that similar items sit close together.Go deeper: Embedding

Eval setEvaluation

Your own collection of inputs and expected behaviours, used to measure an application and catch regressions.Go deeper: Eval set

Few-shot promptingBuilding

Including a handful of worked examples in the prompt to show the model the pattern you want.Go deeper: Few-shot prompting

Fine-tuningTraining

Further training of a pre-trained model on a narrower dataset to change its behaviour or specialise it.Go deeper: Fine-tuning

HallucinationFoundations

Fluent output that is false or unsupported. It is a natural consequence of predicting likely text, not a rare bug.

Human-in-the-loopAgents

Requiring a person to approve or correct certain actions before an AI system carries them out.Go deeper: Human-in-the-loop

Instruction tuningTraining

Fine-tuning on instruction–response pairs so a model follows requests rather than just continuing text.Go deeper: Instruction tuning

JailbreakSafety

A prompt crafted to make a model ignore its safety rules.Go deeper: Jailbreak

LLM-as-judgeEvaluation

Using a model to grade outputs. Scalable, but it has biases toward length, position and its own style.Go deeper: LLM-as-judge

LoRATraining

Low-rank adaptation: fine-tuning small added matrices instead of all weights, making customisation cheap.Go deeper: LoRA

MCPAgents

Model Context Protocol: an open standard for connecting AI applications to tools and data through servers.Go deeper: MCP

Mixture of expertsFoundations

An architecture that routes each token to a few specialised sub-networks, so only part of the model runs per token.

Model cardSafety

A document describing a model’s intended use, training data, evaluations and known limitations.

ParametersFoundations

The learned numbers inside a model. More parameters often, but not always, means more capability.Go deeper: Parameters

Pre-trainingTraining

The first, largest training stage, in which a model learns to predict text from a broad corpus.Go deeper: Pre-training

PromptBuilding

The input given to a model: instructions, context, examples and the question itself.Go deeper: Prompt

Prompt cachingBuilding

Reusing the processed form of a repeated prompt prefix to cut latency and cost.

Prompt injectionSafety

Malicious instructions hidden in content a model reads, such as a web page or document, that try to hijack its behaviour.Go deeper: Prompt injection

QuantisationTraining

Storing weights at lower numeric precision to cut memory and speed up inference, usually with a small quality loss.

RAGBuilding

Retrieval-augmented generation: fetching relevant documents and placing them in the prompt so answers are grounded in them.Go deeper: RAG

ReActAgents

A pattern that interleaves reasoning steps with actions and observations in one loop.Go deeper: ReAct

Red-teamingEvaluation

Deliberately probing a system for harmful, unsafe or broken behaviour before others find it.Go deeper: Red-teaming

RerankingBuilding

A second pass that re-scores retrieved passages with a more accurate model before they reach the prompt.Go deeper: Reranking

RLHFTraining

Reinforcement learning from human feedback: training a reward model on human preferences, then optimising the model against it.Go deeper: RLHF

Scaling lawsFoundations

Empirical relationships between model size, data, compute and performance, used to plan training runs.Go deeper: Scaling laws

Self-attentionFoundations

A mechanism that lets each token weigh every other token in the input when building its representation.Go deeper: Self-attention

Structured outputBuilding

Constraining a model to reply in a fixed format such as JSON that matches a schema.Go deeper: Structured output

System promptBuilding

Standing instructions that set a model’s role, rules and format for a whole conversation.Go deeper: System prompt

TemperatureFoundations

A sampling setting that controls randomness. Lower values give more predictable output, higher values more varied output.

Test-time computeTraining

Spending extra computation while answering, for example longer reasoning or multiple samples, to improve results.Go deeper: Test-time compute

TokenFoundations

The unit of text a model reads and writes, often a word piece. Pricing, speed and context limits are all counted in tokens.Go deeper: Token

Tool callingAgents

A model emitting a structured request to run a function, such as a search or a calculator, whose result is fed back to it.Go deeper: Tool calling

TransformerFoundations

The neural network architecture behind modern language models, built around self-attention.Go deeper: Transformer

Vector databaseBuilding

A store that indexes embeddings for fast similarity search, commonly used for retrieval.Go deeper: Vector database

Vision-language modelBuilding

A model that takes images and text together, for example to read a chart or describe a photo.Go deeper: Vision-language model