The AI glossary

An AI agent is a program that takes a goal and works through the steps on its own — making a plan, using tools, browsing the web, or calling other apps — without you micromanaging each move. Think of it as an AI intern you hand a task to, instead of typing every instruction yourself.

Chain-of-thought is a prompting technique that gets the model to spell out its reasoning step by step before giving the final answer. It noticeably improves accuracy on math, logic, and multi-step problems — simply asking for "step-by-step" reasoning is the simplest version.

The context window is how much text (measured in tokens) a model can consider at once — its working memory. A bigger window means it can handle longer conversations and documents; once you exceed it, the model starts forgetting the beginning.

Diffusion is the technique behind most AI image generators: the model starts from random noise and gradually removes it, guided by your prompt, over many steps. That's why tools like Midjourney reveal the picture in visible stages.

Embeddings turn words, sentences, or images into lists of numbers (vectors) that capture meaning, so "dog" and "puppy" land close together while unrelated words stay far apart. They're the backbone of semantic search, recommendations, and RAG.

Few-shot prompting means giving the model a handful of examples of the task inside your prompt, then asking it to do the same thing. It works like showing a template: the model copies the pattern without any extra training.

Fine-tuning trains an already-built model a bit more on your own examples — your support tickets, your brand voice, your data — to specialize it. It's far cheaper and faster than training from scratch, but the results depend heavily on the quality of your examples.

A foundation model is a large, general-purpose model trained on massive data that other products build on top of — GPT, Claude, and Gemini are examples. "Foundation" is literal: apps, features, and fine-tunes all sit on top of it.

Guardrails are the safety rules and filters wrapped around a model: blocking harmful outputs, keeping answers on-topic, and enforcing policies. They give developers a way to control what the AI can and can't do in a real product.

A hallucination is when an AI confidently makes something up — fake facts, invented links, or quotes that don't exist. It happens because models predict what sounds plausible, not what is verified true, so always double-check important claims.

Inference is simply running a trained model to get an answer — you type a prompt, the model "infers" a response. It happens every time you use ChatGPT, as opposed to training, which is the one-time process that built the model.

LLM stands for Large Language Model: an AI system trained on enormous amounts of text so it can understand and generate human language. ChatGPT, Claude, and Gemini are all LLMs.

A multimodal model handles more than one type of input or output — text plus images, audio, or video. Describing a photo you upload, or generating a video from a text prompt, are both multimodal tasks.

Parameters are the billions of adjustable numbers inside a model that hold everything it learned during training. More parameters usually means a bigger, more capable model — and a more expensive one to run.

Prompt engineering is the skill of writing clear, well-structured instructions to get better answers from AI. Examples, assigned roles, and explicit constraints are its main tools.

Quantization shrinks a model by storing its numbers with less precision (for example, 8-bit instead of 32-bit), so it runs faster and fits on smaller devices. The trade-off is a slight dip in quality for a big win in speed and size.

RAG (Retrieval-Augmented Generation) is a technique where the model first looks up relevant information in your documents or a database, then uses what it found to write its answer. It's how AI assistants answer questions about your own files without hallucinating as much.

Red-teaming means deliberately trying to break a model's safety with tricky prompts and edge cases, in order to find weaknesses before real users do. It's the AI equivalent of hiring hackers to test your security.

RLHF (Reinforcement Learning from Human Feedback) trains the model to prefer answers that humans rate highly. It's what turns a raw text-predicting model into a polite, helpful chatbot instead of an autocomplete that just blurts out likely text.

Temperature is a setting (usually 0–2) that controls how random a model's answers are: low temperature gives focused, predictable replies; high temperature gives more creative, varied ones. For factual work, keep it low.

Tokens are the chunks of text (roughly three-quarters of a word each) that AI models read and write. Everything is measured in tokens: pricing, speed, and the size of the context window all use them.

Training is the long, expensive process of teaching a model from data — adjusting its billions of parameters until it predicts well. Training happens once (costing millions for big models); inference happens every time you use it.

The transformer is the neural-network design behind nearly all modern AI, introduced by Google in 2017. Its "attention" mechanism lets models weigh which words matter most to each other — the breakthrough that made LLMs possible.

A vector database stores and searches embeddings by meaning rather than exact keywords. It's the storage layer behind RAG and semantic search — Pinecone and Weaviate are well-known examples.

Zero-shot means asking the model to do a task with no examples at all — just the instruction. LLMs can often handle it thanks to their broad training, though adding a few examples (few-shot) usually improves accuracy.