AI Glossary (2026)

AI Agent

An LLM-powered system that takes actions in a loop - reading data, calling tools, executing tasks - to reach a goal.

AI Autocomplete

Inline LLM-generated completions in a code editor or text field, accepted with a single keystroke.

Chain of Thought

A prompting technique that asks the model to reason step-by-step before answering, materially improving accuracy on hard problems.

Context Window

The maximum number of tokens an LLM can read in a single request, including the prompt, files, and the response.

Diffusion Model

A generative model that creates images (or video, audio) by iteratively denoising random noise into a coherent output.

Embeddings

A numerical representation of text (or image, audio) as a high-dimensional vector that captures semantic meaning.

Evals

Test suites that score LLM output against benchmarks or human-rated criteria, the unit tests of LLM systems.

Few-shot Learning

Showing the model a small number of examples in the prompt to teach a desired pattern, without changing model weights.

Fine-tuning

Continued training of a pre-trained model on a smaller domain-specific dataset to specialise its behaviour.

Foundation Model

A large, general-purpose model trained on broad data that serves as the base for many downstream applications.

Guardrails

The combined safety, format, and policy filters that constrain what an LLM can output in production.

Hallucination

When an LLM produces output that is fluent and confident-sounding but factually wrong or invented.

Inference

The act of running a trained model on new input to produce output. Distinct from training. Inference cost dominates production budgets.

Large Language Model (LLM)

A neural network trained on huge text corpora that produces human-like text by predicting the next token, one token at a time.

LoRA (Low-Rank Adaptation)

A fine-tuning method that trains small "adapter" matrices instead of updating full model weights, dramatically cheaper than full fine-tuning.

Mixture of Experts (MoE)

A model architecture where only a subset of parameters activates per token, giving large total capacity at lower inference cost.

Multimodal

A model that accepts or produces multiple input/output types - text plus images, audio, or video.

Open-source LLM

A foundation model whose weights are publicly available, letting anyone run it locally or self-host without paying per-token.

Prompt Engineering

The practice of writing inputs to an LLM that reliably produce the desired output, often through structure, examples, and constraints.

Prompt Injection

A class of attacks where untrusted input embedded in an LLM prompt overrides the system instructions.

Quantization

Reducing the numerical precision of model weights (e.g., 16-bit to 4-bit) to shrink memory and speed up inference at minor quality cost.

RAG (Retrieval-Augmented Generation)

A pattern that retrieves relevant documents and feeds them to an LLM as context, instead of relying on the model's training data.

RAG vs Fine-tuning

The standard build-time decision: ground the LLM in retrieved documents (RAG) or specialise its weights with more training (fine-tuning).

RLHF (Reinforcement Learning from Human Feedback)

A training technique that uses human preference data to align an LLM's output with what people actually want.

System Prompt

A hidden instruction at the start of an LLM conversation that sets persona, style, and constraints for all subsequent turns.

Temperature

A model parameter from 0 to ~2 that controls how random the LLM's token sampling is. Lower temperature equals more deterministic output.

Tokens

The chunks of text an LLM reads and produces - usually 3-4 characters each. Pricing and limits are measured in tokens.

Transformer

The neural network architecture introduced in 2017 ("Attention Is All You Need") that powers every major LLM today.

Vector Database

A database optimised for storing and searching high-dimensional vectors by similarity rather than exact match.

Zero-shot Learning

Asking the model to perform a task without any examples - just instructions. Frontier models do this reasonably well in 2026.