AI is full of jargon that makes simple ideas sound complicated. Here's the vocabulary you'll actually encounter, explained the way a friend would — bookmark it.
The basics
- AI (Artificial Intelligence) — software that performs tasks we'd normally consider "smart," like understanding language or recognizing images.
- Machine Learning (ML) — the main way modern AI is built: instead of being programmed with rules, the system learns patterns from lots of examples.
- Neural network — the brain-inspired structure most AI learning happens in; layers of simple units that adjust as they learn.
- Generative AI — AI that creates new content (text, images, audio, video) rather than just classifying or predicting.
- Model — the trained "brain" itself. ChatGPT, Claude and Gemini are products built on top of models.
Large language models
- LLM (Large Language Model) — an AI trained on huge amounts of text to understand and generate language. It powers chatbots like ChatGPT and Claude.
- Token — the small chunks (roughly ¾ of a word) that models read and write in. Pricing and limits are usually counted in tokens.
- Context window — how much text a model can "hold in mind" at once. Bigger windows let you feed in whole documents.
- Parameters — the internal dials a model tunes during training; more (usually) means more capable, but not always.
- Multimodal — a model that handles more than one type of input/output — e.g. text and images and audio.
Talking to AI
- Prompt — the instruction or question you give an AI.
- Prompt engineering — the craft of writing prompts that get better results.
- System prompt — hidden instructions that set an AI's role and rules before your conversation starts.
- Few-shot / zero-shot — giving the AI a few examples (few-shot) vs. none (zero-shot) to guide its answer.
- Temperature — a setting for randomness: low = focused and predictable, high = creative and varied.
Behaviors & limits
- Hallucination — when AI confidently states something false. Always verify important facts.
- Bias — skewed outputs that reflect imbalances in the training data.
- Grounding — tying answers to real sources so they're accurate and checkable (see Perplexity).
- Guardrails — safety limits that stop a model from producing harmful content.
- Cutoff / knowledge cutoff — the date a model's training data ends; it may not know newer events unless it can search.
How models get smarter
- Training — the process of learning patterns from data.
- Fine-tuning — extra training on a narrower dataset to specialize a model.
- RLHF (Reinforcement Learning from Human Feedback) — using human ratings to make models more helpful and safe.
- RAG (Retrieval-Augmented Generation) — giving a model access to a knowledge base so it answers from real documents, not just memory.
- Embeddings — turning text into numbers that capture meaning, so AI can find "similar" things.
- Vector database — where those embeddings are stored for fast semantic search (the backbone of RAG).
The new frontier
- AI agent — AI that doesn't just answer, but acts — using tools, browsing, running steps to complete a task.
- Agentic — describes AI systems built to take autonomous, multi-step action.
- Copilot — an AI assistant embedded in a tool to help you work (e.g. coding copilots).
- Inference — the act of a model actually running to produce an answer (as opposed to training).
- Open vs. closed model — open models can be downloaded and self-hosted; closed models are accessed via a company's API/app.
- API — the interface developers use to plug a model into their own apps.
That's the core vocabulary. Want to put it to use? Start with our beginner's guide to AI or browse the tool directory.