▲ The glossary
AI, in plain English.
Plain-English AI terms for working professionals — what each concept means, why it matters to your work, and the Nowrap tools and news where it shows up.
Concepts.
- AI agentAn AI that doesn't just answer questions but takes a goal, makes a plan, and uses tools to carry it out across multiple steps.
- Knowledge cutoffThe date an AI's training data ends — after which it knows nothing unless you tell it or it can search.
- MultimodalAn AI that works across more than just text — reading images, audio, and video, not only words.
- PromptThe instruction you give an AI — and the single biggest lever you have over the quality of what it gives back.
Models.
- Foundation modelA large, general-purpose AI model trained at huge scale that other tools and products are built on top of.
- Large language modelThe kind of AI — trained on vast amounts of text — that powers chatbots and writing assistants by predicting the next word.
- Open-weight modelAn AI model whose trained parameters are published, so anyone can download, run, and adapt it themselves.
- Reasoning modelA model that works through a problem step by step before answering, trading speed for accuracy on hard tasks.
Techniques.
- Few-shot promptingShowing an AI a couple of worked examples in your prompt so it copies the pattern instead of guessing.
- Fine-tuningFurther training a general AI model on your own examples so it adopts a specific style, format, or domain.
- GroundingTying an AI's answer to verifiable sources, so it reports what the evidence says instead of what it vaguely recalls.
- Retrieval-augmented generationA technique that feeds an AI your own documents at question time, so its answers are grounded in real sources instead of memory.
- System promptThe standing instructions that set an AI's role and rules before you ever type your first message.
- TemperatureA setting that controls how predictable or creative an AI's output is — low for consistent, high for varied.
- Tool callingThe ability for an AI to use external tools — run a search, send an email, query a system — instead of only producing text.
Infrastructure.
- APIA doorway that lets one piece of software talk to another — how apps and agents plug into an AI model.
- Context windowHow much an AI can "hold in mind" at once — the working memory that limits how much it can read or remember in one go.
- EmbeddingA way of turning text into numbers that capture meaning, so an AI can find things by what they mean rather than the words they use.
- InferenceThe act of running a trained AI model to get an answer — the step you actually pay for and wait on.
- Model Context ProtocolA common standard that lets AI assistants plug into your tools and data without a custom integration for each one.
- TokenThe chunk of text an AI reads and writes in — and the unit you're actually billed by.
- Vector databaseA database built to store embeddings and find items by meaning — the retrieval engine behind most RAG systems.
Risk & Evaluation.
- BenchmarkA standardised test used to measure and compare AI models — and a number to read with healthy skepticism.
- GuardrailsThe rules and checks bolted around an AI to keep its output inside safe, compliant, on-brand bounds.
- HallucinationWhen an AI states something false with complete confidence — the failure mode that matters most in professional work.
- Prompt injectionAn attack where hidden instructions buried in a document, email, or webpage hijack an AI into ignoring its real task.