Glossary

The AI Glossary

The AI vocabulary, explained in plain English — a direct answer first, then just enough depth to actually use the term.

AGI (artificial general intelligence)

A hypothetical AI that matches or exceeds human ability across most cognitive tasks — not just one domain. No agreed definition or test exists, which is precisely why every 'AGI achieved' claim is contested.

Read definition →

AI agent

An LLM running in a loop with access to tools — it takes a goal, plans steps, executes them (run code, browse, edit files), checks results and keeps going until the task is done or it gets stuck.

Read definition →

Benchmark

A fixed, repeatable set of test tasks used to score AI models and compare them — like SWE-bench for coding or Arena Elo for human preference. Useful, gameable, and never the whole story.

Read definition →

Context window

The maximum amount of text — measured in tokens — that an LLM can consider at once: your prompt, the conversation so far, any documents, and its own reply all share this budget.

Read definition →

Diffusion model

A generative model that creates images (and video or audio) by learning to reverse a noising process: it starts from random static and removes noise step by step until a coherent output matching the prompt emerges.

Read definition →

Embedding

A vector — a long list of numbers — that represents the meaning of text, images or audio, so that similar things end up numerically close together. The backbone of semantic search and RAG.

Read definition →

Fine-tuning

Further training a pretrained model on your own examples so its weights change — teaching behavior, style or domain skills that prompting alone can't reliably produce.

Read definition →

Hallucination

When an AI model states false or invented information fluently and confidently — fake citations, nonexistent APIs, wrong facts — because it predicts plausible text rather than checking truth.

Read definition →

Inference

Running a trained model to produce outputs — as opposed to training, which creates the model. Every chat reply and API call is inference, and it's where the ongoing cost of AI lives.

Read definition →

Large language model (LLM)

A neural network trained on vast amounts of text to predict the next token in a sequence — the technology behind chatbots and coding assistants like ChatGPT, Claude and Gemini.

Read definition →

Mixture of experts (MoE)

A model architecture that splits capacity into many 'expert' sub-networks and routes each token through only a few — giving huge total size at a fraction of the compute per token.

Read definition →

Multimodal

Describes an AI model that works across more than one data type — text, images, audio, video — either understanding them as input, generating them as output, or both.

Read definition →

Open weights

A model whose trained parameters are published for anyone to download, run and usually fine-tune — as opposed to closed models reachable only through a vendor's API. Not necessarily open source.

Read definition →

Prompt

The input you give an AI model — instructions, questions, context and examples. The model's entire behavior in a session is steered by it, which is why writing good prompts is a real skill.

Read definition →

Quantization

Compressing a model by storing its weights in fewer bits (e.g. 4-bit instead of 16-bit) — cutting memory and speeding inference several-fold, at a usually-small accuracy cost.

Read definition →

RAG (retrieval-augmented generation)

An architecture that searches your documents for relevant passages and inserts them into the model's context at query time — so answers are grounded in retrieved facts rather than the model's training memory.

Read definition →

RLHF (reinforcement learning from human feedback)

A training stage where humans rank model outputs and the model is optimized toward the preferred ones — the technique that turned raw next-token predictors into helpful, instruction-following assistants.

Read definition →

System prompt

The standing instructions given to a model before any user message — defining its persona, rules, tone and tools. Users don't see it, but every reply is shaped by it.

Read definition →

Temperature

A sampling setting that controls how random a model's output is. Low temperature (~0) makes it deterministic and safe; higher values (~1+) make it more varied and creative — and more error-prone.

Read definition →

Token

The unit of text an LLM actually processes — a word, part of a word, or punctuation, averaging about ¾ of an English word. API pricing and context limits are both measured in tokens.

Read definition →