AI glossary

Short definitions of words you’ll run into in AI news and on our benchmark pages, listed A to Z. Each term has its own link (for example, /learn/glossary/#context-window), so other pages can point straight to it.

Benchmark

A standard test used to compare AI models on the same questions or tasks, scored the same way. See What is an AI benchmark?, How to read AI benchmarks, and how we explain benchmark results.

Chatbot

An app you talk to in plain language, usually by typing. Most AI chatbots are powered by a large language model.

Contamination

When a benchmark’s test questions (or their answers) end up in a model’s training data. The model may then score well by remembering rather than reasoning, so the score overstates its real skill.

Context window

How much text a model can consider at once, including your question, any files you add, the conversation so far, and its own reply. It’s measured in tokens. Anything beyond the limit is left out or forgotten.

Elo rating (arena rating)

A score borrowed from chess. On “arena” style leaderboards, people compare two anonymous answers and vote for the better one; each vote nudges the two models’ ratings up or down. A higher rating means people preferred that model’s answers more often. Ratings are relative: they only mean something compared with other models on the same leaderboard.

Fine-tuning

Extra training that adapts an existing model to a particular job or style.

Free tier

A no-cost version of a paid service, usually with limits such as fewer messages per day, slower responses, or access only to smaller models. Limits can change at any time, so check the provider’s own pricing page.

Hallucination

When an AI states something false as if it were true, such as a made-up fact, quote, or source. Always check important facts.

Inference

Using a trained model to produce an answer. Every time you send a prompt, the model runs inference. Inference speed and cost are what you notice as an everyday user.

Large language model (LLM)

An AI model trained on large amounts of text to predict and generate language. It’s the technology behind most AI chat apps.

MMLU

Short for “Measuring Massive Multitask Language Understanding.” A widely used benchmark of multiple-choice questions across dozens of school and professional subjects. Many top models now score near the top of it, so newer, harder tests are often used alongside it.

Model

The trained AI system itself. One app may offer several models, and one model may be used by many apps.

Multimodal

A model that can work with more than one kind of input or output, such as text, images, or audio.

Open weights

A model whose trained parameters (its “weights”) are published so anyone can download and run it on their own hardware. “Open weights” doesn’t always mean “open source”: licenses vary, and the training data is often not shared. Check each license.

Parameters

The internal numbers a model learns during training. Model sizes are often described by parameter count (for example, “7B” means about 7 billion). More parameters usually need more computing power to run, but size alone doesn’t tell you how good a model is.

pass@1

A way of scoring coding benchmarks: the share of problems a model solves correctly on its first try. “pass@k” means at least one of k tries was correct, which gives a higher number, so always compare pass@1 with pass@1.

Prompt

The instructions or question you give an AI model.

Rate limit

A cap on how many requests or messages you can send in a period of time, such as per minute or per day. Free tiers usually have lower rate limits than paid plans.

Self-reported score

A benchmark result published by the company that made the model, rather than measured by an independent group. It may be accurate, but the test settings are chosen by the maker, so treat it as a claim until others confirm it.

Token

A small chunk of text, often part of a word, that models read and write. Prices, context windows, and rate limits are often counted in tokens. As a rough guide, a token is a few characters of English text.

Training data

The examples a model learned from before it was released.