AI Glossary
Clear, precise definitions for essential concepts across foundation models, architecture, and agent systems.
Context Window
The total amount of information (in tokens) an AI model can process in a single request, including input and history.
Fine-Tuning
Further training of a pretrained model on a smaller, task-specific dataset so it adapts to a particular domain, format or tone.
Hallucination
A confident but factually wrong model output, caused by generating plausible text rather than retrieving verified facts.
Inference-Time Reasoning
Spending extra computation at answer time — an internal chain of thought — to improve accuracy on hard problems, as used by reasoning models such as OpenAI o1 and DeepSeek R1.
Mixture of Experts (MoE)
A machine learning technique where different specialized sub-networks (experts) are used for different parts of an input.
Multimodal
A model that accepts or produces more than one kind of data — for example text plus images, audio or video.
Open Weights
A model whose trained parameters are published for download, allowing self-hosting and fine-tuning, even when the licence is not strictly open source.
Parameters
The weights and biases that a model learns during training, roughly correlating to its complexity and capacity.
Quantisation
Storing model weights at lower numeric precision to cut memory use and speed up inference, usually with a small quality cost.
Retrieval-Augmented Generation (RAG)
A framework for retrieving data from external sources to ground LLM responses in factual information.
Tokens
The basic units of text processed by an LLM, representing sequences of characters, words, or sub-words.