AI glossary

A large language model, or LLM, is a neural network trained on very large amounts of text to predict which piece of text comes next.

A token is the basic unit of text that a language model reads and writes: a whole word, a fragment of a word, a number or a punctuation mark.

The context window is the maximum amount of text, counted in tokens, that a language model can consider at one time.

Temperature is a setting that controls how random a language model is when it picks the next token.

A reasoning model is a language model trained to work through a problem in intermediate steps before it gives its final answer.

In AI, a hallucination is an answer that sounds fluent and confident but is false, unsupported or invented.

Retrieval-augmented generation, or RAG, is a technique in which the passages most relevant to a question are searched in a collection of documents and handed to a language model, which then writes its answer from them.

A system prompt is a set of standing instructions given to a language model before the conversation begins, defining its role, tone, format and rules.

Top-p, also called nucleus sampling, is a setting that limits a language model to the smallest group of likely tokens whose probabilities add up to a chosen share, then picks the next token from that group only.

A mixture-of-experts model is a neural network split into many specialized sub-networks called experts, with a router that sends each token to only a few of them.

Quantization is a technique that stores a model’s numbers with fewer bits, for example 8 or 4 instead of 16, so the model takes less memory and runs faster.

Embeddings are lists of numbers, called vectors, that represent the meaning of a piece of text, an image or a sound.

An AI agent is a system built around a language model that pursues a goal over several steps, deciding what to do next, using tools and checking the results.

Tool calling, also known as function calling, is a feature that lets a language model ask your program to run a function and use the result in its answer.

MCP, the Model Context Protocol, is an open standard that lets AI applications connect to tools and data sources through one common interface.

An uncensored model is a language model that has not been tuned to turn down requests arbitrarily or to bury its answers under warnings, so it replies frankly to what was asked.

A refusal is when a language model declines to carry out a request, usually with a short statement that it cannot or will not help.

Guardrails are the rules, filters and trained behaviors that an AI maker adds to a model to limit what it will say or do.

An open-weight model is an AI model whose trained parameters, called weights, are published so that anyone can download and run them.

Alignment is the work of making an AI model behave the way its makers and users intend: following instructions, being useful and avoiding what is judged harmful.

RLHF, short for reinforcement learning from human feedback, is a training method in which people rank a model’s answers and the model is then adjusted to produce the kind of answers they preferred.

Fine-tuning is the process of continuing to train an already pretrained model on a smaller, specific dataset so that it adapts to a task, a style or a domain.

A diffusion model is an image generator that starts from random noise and removes it step by step until a picture matching a text description appears.

A negative prompt is a second text description that tells an image model what to steer away from, such as blur, extra fingers or a washed-out look.

Image upscaling enlarges a picture to a higher resolution while trying to keep it sharp.

A seed is a whole number that fixes the random noise an image model starts from, so the same seed with the same prompt and settings can draw the same picture again.

The CFG scale, short for classifier-free guidance and also called guidance, is a setting that decides how strictly an image model follows your text prompt.

A LoRA, short for low-rank adaptation, is a small add-on file that adjusts a large AI model so that it learns a style, a subject or a skill without being retrained in full.

Inpainting is an image-editing technique in which an AI model redraws only a chosen region of a picture, guided by a text prompt, while the rest of the image stays as it is.

Background removal is the automatic separation of the main subject of a picture from its backdrop, so the subject can be placed on a transparent or different background.

Image-to-video is a way of making a short clip by giving an AI model a still picture to start from, plus a sentence that says how things should move.

Local-first software keeps your data on your own device as the main copy and treats servers as optional helpers.

An anonymizing proxy is an intermediary that sends a request on your behalf so that the service at the other end does not learn who you are.

Data retention is the length of time a service keeps the information it receives or creates before deleting it.

IndexedDB is a database built into web browsers that lets a site store large amounts of structured data on the visitor's own device.

Zero-log describes a service that does not keep logs of what you ask or say: no stored record of your prompts, replies or activity beyond what it strictly needs to run.

End-to-end encryption means data is locked on the sender's device and can only be unlocked on the intended recipient's device, so every server in between carries nothing but unreadable ciphertext.

An OpenAI-compatible API accepts the same requests and returns the same responses as OpenAI's API, so existing code and tools can use another provider by changing little more than the address and the key.

A rate limit is a cap on how many requests an account or key may send to a service within a period of time.

Credits are a prepaid unit of account that a service uses to price usage, so each request spends a small number of them according to what it consumed.