AI glossary

One-line definitions for the words that get tossed around in every AI article.

The basics

• AI (Artificial Intelligence) — software that does things that used to require a human, like writing or recognizing images.
• LLM (Large Language Model) — the kind of AI behind ChatGPT, Gemini, and Claude. It reads and writes language.
• Model — a specific trained AI brain. GPT-4, Gemini 2.0, and Claude 3.5 Sonnet are models.
• Prompt — whatever you type into the AI.
• Response — what the AI types back.
• Chat or Conversation — a back-and-forth session. The AI remembers the whole chat.

How it works

• Token — a chunk of text the AI reads, roughly 4 characters. "Cat" is one token, "antidisestablishmentarianism" is several.
• Context window — how much text the AI can remember at once, measured in tokens. Bigger = can read longer documents.
• Training — the months-long process of teaching a model by showing it tons of text.
• Parameters — the internal dials inside a model. Billions of them. More usually means smarter, but slower.
• Inference — when the AI is actually answering you (the opposite of training).

What can go wrong

• Hallucination — when the AI makes something up that sounds true but isn't.
• Bias — the AI inherits patterns from its training data, including unfair ones.
• Jailbreak — a clever prompt that tricks the AI into ignoring its rules.

The new vocabulary

• RAG (Retrieval-Augmented Generation) — the AI looks something up before answering. How tools "use your documents."
• Agent — an AI that takes actions on your behalf, not just answers questions. Books your flight, runs your code.
• Multimodal — handles more than text. Images, audio, video.
• Fine-tuning — extra training on specific data to specialize a model.
• Embedding — turning text into a list of numbers a computer can compare for similarity. Powers AI search.
• MCP (Model Context Protocol) — a standard way for AI tools to connect to your apps and data.

The labels on the box

• GPT — OpenAI's family of models (powers ChatGPT). Stands for Generative Pre-trained Transformer.
• Gemini — Google's family of models.
• Claude — Anthropic's family of models.
• Open source — model weights are public; anyone can run them. Llama (Meta) and Mistral are popular ones.
• Closed — only the company can run it; you use it through their app or API.

That's it. You now speak AI.