Glossary
The AI glossary
Plain-English definitions of the AI terms you keep seeing — one clear sentence each, then a link to see it in a lesson.
Language modelThe kind of AI behind chat assistants. At heart it does one thing: predict the next word, over and over, to build text.TemperatureA dial for risk. Low keeps answers safe and predictable; high makes them surprising and creative.MCP (Model Context Protocol)An open standard (Anthropic, 2024) that lets any AI app connect to any tool through one shared "port," instead of a custom connector per pair.HostThe AI app you actually use — a chat app, an IDE, an agent. It runs the conversation and decides which tools to call.ClientThe connector inside the host that opens and manages one link to a single server.ServerA small program that exposes one tool or data source — your files, Slack, a database — to any MCP host.JSON-RPCThe simple text-based request-and-reply format MCP uses to pass messages between an AI app and a tool.TrainingShowing an AI an enormous amount of text and having it practise predicting the next word, over and over, until the patterns sink in. The squeezed-down result is the model.Knowledge cutoffThe point in time when an AI's training data ends. It knows little or nothing about events after that date.HallucinationWhen an AI states something false as if it were true — a plausible-sounding guess that fills a gap in what it actually knows. It isn't lying; it doesn't know it's wrong.PretrainingThe first training stage: reading a huge slice of the internet to absorb knowledge and language. It produces a base model.Base modelA model fresh from training that can only continue text. It has read a lot, but hasn't yet been taught to behave as a helpful assistant.FinetuningA second training stage that teaches a base model how to behave — answering helpfully, politely, safely — by showing it many example conversations.Context windowThe amount of text an AI can hold in view at once — your whole conversation so far. It re-reads this on every reply and has no memory beyond it; when it fills up, the oldest part drops off.Open weightsA model whose trained insides are published, so anyone can download and run it themselves — with full control and privacy. The opposite is a closed model, reached only through a company's service.Reasoning modelA model that works through a problem step by step before answering, instead of replying instantly. Slower, but stronger on hard, multi-step questions.AgentAn AI you give a goal rather than step-by-step instructions. It works in a think-act-observe loop — using tools and reacting to each result — until the goal is reached.Tool useAn AI taking action in the world through a connected tool — searching the web, sending an email, querying your files — rather than only producing text.Text-to-imageMaking a brand-new image from a written description, rather than photographing, drawing, or finding one. You type what you want; the tool generates it.DiffusionHow most image generators work: they start from a field of random noise and remove it step by step, guided by your words, until a clear picture is left.PromptThe words you give an AI to say what you want — a question, an instruction, or a description of an image to make.Image briefAn image prompt written like a brief you'd hand a photographer or designer: subject, style, composition, and lighting all specified up front, rather than discovered through back-and-forth chat.Aspect ratioThe shape of an image's frame — how wide it is versus how tall, like a square (1:1), a wide banner (16:9), or a tall phone screen (9:16). An image tool composes the whole picture for the shape you choose, so it's a creative decision, not just a crop.Style referenceHanding an image tool an example picture and asking for that look, instead of describing the style in words. The tool borrows the vibe — colours, texture, mood — and applies it to your subject.SeedThe single number that fixes an image tool's starting patch of random noise. The same prompt with the same seed always produces the exact same picture — so a seed is how you reproduce or make small controlled edits to a result you like.InpaintingRepainting just one masked area of an existing image while leaving the rest untouched — the way to fix or swap a single part without regenerating the whole picture. Its outward twin, extending a picture past its edges, is outpainting.Image-to-imageGiving an image tool a starting picture — a photo, a sketch, an earlier result — instead of starting from pure noise. Your picture becomes the base, and your words steer how far it changes.Control referenceA simple structure image — a pose skeleton, an edge outline, or a depth map — you hand an image tool so the result keeps that exact shape. The tool invents the surface (faces, colour, light) but follows your structure. The technique is often called ControlNet.UpscalingEnlarging an image and inventing the fine detail it never had — sharper edges, finer texture — so a small, soft result holds up at large sizes. Usually the final step, once the picture is right.Text-to-videoDescribing a scene in words and having an AI tool generate a short moving clip from scratch, frame by frame — rather than filming it or editing existing footage.Image-to-videoHanding the tool a single still picture — a photo you took or an image you generated — plus a short note on how it should move, so it keeps your picture as the first frame and only adds the motion.KeyframeA frame you pin at a chosen moment — often the first and last frame of a clip. You set those anchors and the tool generates the motion that flows between them.Temporal consistencyHow well every frame of a clip agrees with the ones around it — the same face, colours and shapes from start to finish. It's the hardest thing for video tools, and it slips the longer a clip runs, so the subject slowly 'drifts'.Lip-syncMatching a face's mouth shapes to the sounds in a piece of speech, frame by frame, so a still portrait or avatar looks like it is really saying the words.DubbingReplacing a video's spoken dialogue with a new recording — often in another language — and re-syncing the mouth to the new words, so the same footage 'speaks' the new track.TokenThe unit an AI actually reads and writes: a chunk of text — often a whole common word, sometimes just a piece of a longer or rarer one. A model sees your text as a sequence of these chunks, never as individual letters.TokenizationThe step that splits your text into tokens before the AI reads it. Common words become a single token; rarer or longer ones break into several pieces.TransformerThe design behind almost every modern AI model: a tall stack of near-identical layers that pass your text upward as numbers, refining them at each step until the top layer can predict the next token.ParametersThe billions of adjustable numbers inside a model — its "weights" — that were tuned during training. They are what the model learned, and running the model means passing your text through all of them.AttentionHow a model works out which earlier words each word should listen to. For every word it looks back over the text, weighs the others by how relevant they are, and pulls in meaning from the ones that matter — that's how it figures out what a word like "it" refers to. It's the core mechanism inside a transformer.Scaling lawsThe finding that a model's error drops in a smooth, predictable way as you scale it up — more parameters, more training data, more compute. The curve is regular enough that you can forecast how good a bigger model will be before you build it.Emergent abilitiesSkills a model suddenly gains once it crosses a certain size — reliable spelling, arithmetic, following multi-step instructions — that it simply didn't have when smaller. Nobody programs them in; they appear as a side effect of scale.In-context learningTeaching a model a task at runtime by putting a few worked examples straight in the prompt, instead of retraining it. The model spots the pattern in your examples and continues it — also called few-shot prompting. Nothing about the model changes; the examples only steer it for that one message.Post-trainingEverything done to a base model after pretraining to turn it into a helpful assistant — showing it example conversations, then letting it practise with reinforcement learning. It shapes how the model behaves, not what it knows.Reinforcement learningTraining by trial and reward instead of by copying. The model attempts a problem many times, and the attempts that reach a good outcome are reinforced — made more likely — so it discovers its own successful strategies rather than imitating ours.Chain of thoughtThe step-by-step reasoning a model writes out before its final answer. Thinking in the open spreads the work across many tokens and lets the model catch and fix its own mistakes on the way.RLHFReinforcement learning from human feedback: people rank answers, a reward model learns their taste, and the assistant is tuned to score well by it. It's how models pick up hard-to-specify qualities like being helpful, kind, and safe.Reward modelA learned judge that scores an answer the way people would, trained on human preference rankings. It stands in for a human rater so the AI can be improved automatically — but it's only an imperfect copy of real human taste.Reward hackingWhen a model games its scorer — finding answers the reward model rates highly without actually being better, because the reward is only a proxy for what people really want. Push the optimisation too hard and this creeps in.RefusalWhen an AI declines to answer instead of guessing — ideally because it recognises the question is outside what it reliably knows. Good models are trained to do this, so a well-placed "I'm not sure" is a feature, not a failure.GroundingAnswering from sources you put in front of the model — a document, a search result, your own notes — rather than from its fuzzy internal memory. Grounded answers are more accurate and can point to exactly where they came from.Jagged intelligenceThe uneven shape of AI ability: superhuman at some tasks and surprisingly poor at others, with no smooth line between them. The same model can draft a polished essay yet miscount the letters in a word.CalibrationHow well an AI's expressed confidence matches how often it's actually right. Most chatbots are poorly calibrated — they sound just as sure when they're wrong as when they're right.VerificationChecking an AI's answer against something independent — its cited source, a second search, a re-ask, or your own test — before you rely on it.MultimodalAn AI that can take in more than typed words — images, voice, a screenshot, a PDF — and read them all as part of the same conversation. It turns each one into tokens it can work with.MemoryA separate store, outside any single chat, where an assistant can save facts about you — your name, your preferences — and quietly slip them into future conversations so you don't repeat yourself.Few-shotGiving the AI a few worked examples inside your prompt so it copies the pattern, instead of describing the pattern in words. Zero-shot is no examples, one-shot is one, few-shot is several.DelimiterA visible boundary — quotes, dashes, or a labelled tag like <review>…</review> — that fences pasted text off from your instructions, so the AI can tell the material from the commands.Zero-shotPrompting with just a description and no examples — the plain, everyday way most people ask. It's fine for simple tasks; add examples (one-shot, few-shot) when the exact shape or style matters.Self-consistencyAsking the AI the same hard question several times, then taking the answer that comes up most often. Because each attempt reasons a little differently, the majority vote is more reliable than any single try — but only when the question has one right answer.Prompt chainingBreaking a big task into a short sequence of prompts, where each step's output feeds the next. Each prompt stays small and focused, so each result is sharper — and you can catch a mistake before it spreads down the chain.Prompt injectionAn attack where hidden or malicious text inside content an AI reads is treated as instructions, hijacking what it does.Custom GPTA version of ChatGPT you set up once for a specific job — with its own instructions, reference files, and skills — then save and reuse or share. No coding required.Code InterpreterA built-in tool that lets ChatGPT write and run real code (usually Python) to calculate, analyse data, or make charts — so it computes the answer instead of guessing it.Custom instructionsStanding notes you give ChatGPT once — who you are and how you'd like replies — that it applies to every new chat, so you don't repeat your preferences each time.Deep ResearchA mode where ChatGPT spends several minutes browsing many sources on its own, then returns a longer, organised report with references — research delegated, not typed.CanvasA side panel for writing or code that you and ChatGPT edit together — change one line in place and leave the rest, instead of regenerating the whole thing.JailbreakA prompt crafted to trick an AI past its safety rules — often by disguising a harmful request as roleplay or a hypothetical.DeepfakeSynthetic audio, image, or video generated by AI to convincingly impersonate a real person.Data leakageWhen sensitive information you share with an AI leaves your control — logged, reviewed, or used to train future models.Artificial general intelligenceA hypothetical AI that matches human ability across essentially all tasks — unlike today's narrow systems. It does not exist yet.Narrow AIAI that is good at one specific task (translation, image generation, chat) but has no general ability outside it — all AI in use today is narrow.Content CredentialsA signed record (the C2PA standard) attached to a file showing how it was made and whether AI was involved — a 'nutrition label' for media.APIThe metered, pay-per-use way to reach an AI model from your own code or product — billed per token, unlike a flat consumer subscription.Prompt cachingA discount where the model reuses an already-processed chunk of your prompt instead of re-reading it, cutting the cost of repeated input (often to about 10% of the normal rate).Reasoning tokensHidden 'thinking' tokens a reasoning model generates before its visible answer. You don't see them, but they're billed at the output rate — which is why reasoning costs more.Batch processingSending many non-urgent requests together for the model to handle when convenient, in exchange for a large discount (often about 50%) versus real-time calls.ArtifactsA side panel in Claude where its output — a document, app, chart, or diagram — appears as a living thing you edit and refine in place, instead of re-asking in the chat.Extended thinkingA mode where the model takes extra time to reason step by step before answering. It helps on hard problems and is wasted on easy ones, so you turn it up only when the question is tricky.ConnectorsSecure links that let an assistant reach your outside tools and services — mail, files, calendar — so it can use your real data in an answer instead of guessing.GemA custom, reusable version of Gemini with its own saved instructions — a role, tone, and rules — so a repeated task starts already set up.NotebookLMA Google tool that answers questions only from documents you upload, citing the exact source — so its answers are grounded in your material, not the open web.QuantizationShrinking a model by storing its numbers more coarsely (like rounding), so it takes less memory and runs faster — with only a small drop in quality. It's what lets big open models fit on a normal laptop.DistillationTraining a small model to copy a big one's behaviour, so you get much of the quality in a far smaller, faster model that's easier to run yourself.BiasSystematic unfairness in an AI's outputs, learned from patterns in its training data — it can favour or disadvantage particular groups, and is mitigated rather than simply switched off.EU AI ActThe European Union's law that regulates AI by risk tier — banning a few uses outright, imposing strict checks on high-risk ones, requiring disclosure for some, and leaving most everyday tools alone.LLM OSA vision (popularized by Andrej Karpathy) of the language model as the CPU of a new kind of computer — coordinating tools, memory, other models, and the web around it, rather than acting alone.SuperintelligenceA hypothetical AI that far surpasses the best humans at essentially everything — a level beyond AGI. It does not exist, and whether or when it could is contested.StemsThe separable layers a song is built from — vocals, drums, bass, melody — each on its own track, so you can edit one without touching the others.Royalty-freeMusic you license once (a fee or subscription) and can then use within the license's terms, with no further per-play royalties — licensed, not free of rules.Generative fillSelecting an area of an image and describing what should fill it, so the tool repaints only that region and leaves the rest untouched.OutpaintingExtending an image beyond its original edges, with the tool inventing plausible content for the newly added area.VectorAn image stored as shapes and instructions rather than a grid of dots, so it stays perfectly crisp at any size (e.g. an SVG logo).ProvenanceThe traceable origin and edit history of a piece of media — where it came from and how it was made. Checking provenance (via Content Credentials or a watermark) is more reliable than judging by how it looks.Text-to-speech (TTS)Turning written text into natural-sounding spoken audio; you supply the words and pick a voice, tone and pace, and the tool generates the speech — no recording needed.Voice cloningRecreating a specific person's voice from a short sample, so any new text can be spoken in that voice; it requires the person's clear consent — cloning without permission enables impersonation and fraud.Transcription (speech-to-text)Turning spoken audio into written text, usually with a timestamp on each line and often speaker labels; it powers subtitles, meeting notes and searchable recordings — the reverse of text-to-speech.De-identifyTo remove the details that could identify a person — name, date of birth, address, ID or record numbers — from information before sharing it, so the useful content stays but the identity doesn't.WorkflowA fixed, predefined sequence of steps the AI follows — more reliable and repeatable than an agent that decides its own path.Human in the loopA checkpoint where a person must approve an AI's action before it happens — used for risky or irreversible steps.No-codeBuilding software or automations by connecting visual blocks instead of writing code, so non-programmers can create working tools.WebhookA way for one app to instantly notify another when something happens, by sending a small message to a waiting URL — the 'reverse' of asking again and again.Self-hostingRunning software on your own server or computer instead of a company's cloud, so you control the data and the cost.NodeOne step or block in a visual workflow — it does a single job (a trigger, an action, or a transform) and passes its result to the next.RAG (retrieval-augmented generation)Giving the AI your own documents to pull answers from, so it responds from real sources instead of memory alone.VibecodingBuilding an app by describing what you want in plain language and letting an AI write and run the code — a term coined by Andrej Karpathy.BackendThe behind-the-scenes part of an app — the server, database and logic that store data and do the work, as opposed to the screen you see.DeploymentPublishing your app or workflow so it runs live on the internet and other people can actually use it.MoatA durable advantage that makes a product hard to copy — proprietary data, deep integration, distribution, or a niche you own — as opposed to the rented model everyone shares.Platform riskThe danger of building on one provider's model whose prices, features, or rules you don't control and that can change overnight.GEOGEO (generative engine optimization) — structuring your content so AI assistants quote it when answering; the AI-era twin of SEO.Scaled content abuseMass-producing many low-value, unoriginal pages or posts mainly to game search rankings or a feed — which platforms and search engines demote, no matter whether a person or an AI made it. The fix is originality and real value, not less volume.RepurposingReshaping one core piece of content into several formats for different channels — a long video into shorts, a thread, a newsletter — instead of creating each from scratch.AI FluencyThe skill of working well with AI, framed by Anthropic as four habits (the 4Ds): Delegation, Description, Discernment, and Diligence — choosing what to hand over, describing it clearly, judging the output, and using it responsibly.Innovation theaterBuilding flashy AI demos to look innovative rather than to ship real value — impressive on stage, but with no owner, success metric, or path to production.Data flywheelA loop where using a product generates data that makes it better, which draws more users and more data — a compounding advantage a rival who only rents the model can't copy.Foundation modelA large, general model pretrained on broad data that products build on — you call it via an API (buy), shape it with your data (adapt), or, rarely, train your own (build).EvalsA fixed set of test cases you run an AI product against to measure how well it does the job — and to catch when a change quietly makes it worse (a regression).North-star metricThe single business outcome a product is steered by — the number that actually matters (like retention or resolved tickets), as opposed to a technical score that may not move it.EmbeddingsA way of turning text into a list of numbers (a vector) that captures its meaning, so text with similar meaning ends up close together. Embeddings are what let retrieval search by meaning instead of exact words.Vector databaseA database built to store embeddings and quickly find the vectors nearest to a query — the retrieval half of a RAG system at scale. It uses an index to avoid comparing against every stored vector on each search.GPUA Graphics Processing Unit — a chip with thousands of simple cores that do easy maths all at once. AI models run on GPUs because their core work — multiplying long lists of numbers — is exactly the kind of massively parallel maths a GPU does far faster than a CPU.BenchmarkA standardized test used to compare AI models. Useful for rough ranking, but easy to game or cherry-pick — a high benchmark score is a claim to check, not proof of real-world skill.Automation biasThe human tendency to over-trust output from an automated system: accepting an AI’s answer because it came from a machine, and under-checking it. The reflex that turns a confident wrong answer into a costly mistake.