The AI Glossary from A to Z

Every term this blog uses, in one place, each explained in one or two sentences. Not a translation - an explanation.

We keep the English terms as they are. The industry uses them this way, translating them only makes readers less sure of themselves, and nobody types the translated version into a search box either.

So this page does not translate: it explains. Every term gets one or two sentences, in plain language, with the practical consequence rather than the textbook definition.

Every article on the blog links here. If you hit a word you do not know, this is where it lands you.

Jump to a section: Basics · Prompting · Models & market · Data & RAG · Agents · Images · Video & audio · Coding · Marketing · Law & security

Basics 34 terms

• AGI (artificial general intelligence) — The hypothetical point where one system matches human performance on any intellectual task. There is no agreed measure for it.
• AI (artificial intelligence) — The umbrella term for systems that perform tasks which used to require human thinking.
• alignment — The work of making a model behave in line with human intent and values.
• benchmark — A standardised test used to compare models. A high score does not mean it fits your task.
• closed model — A model reachable only through an API or app. Its weights are not downloadable.
• completion — The model's output - what it generates in response to your prompt.
• context window — The maximum amount of text a model can hold at once - your question, your attachments and its answer together.
• deep learning — The branch of machine learning built on many-layered neural networks. Almost every current AI model is one.
• foundation model — A large general-purpose model that specific applications are built on top of.
• frontier model — The most capable models available at a given moment. The field reshuffles every two to four months.
• generative AI — The branch of AI that produces new content - text, images, audio, video, code.
• guardrail — A built-in limit stopping a model from producing certain content or taking certain actions.
• hallucination — When a model states something false with full confidence - a consequence of probabilistic generation, not a bug.
• inference — Running a finished model - you ask, it answers. Every answer costs compute.
• knowledge cutoff — The date the model's training data ends. Anything later it only knows if it searches or you supply it.
• leaderboard — A ranking of benchmark results. The most-used models are often not at the top.
• LLM (large language model) — A large language model - trained on huge amounts of text, generating output by predicting the next token.
• machine learning — Systems that derive rules from examples instead of following rules a person wrote.
• model — The trained system itself - the complete set of weights.
• multimodal — A model that handles more than one kind of input - text, image, audio, video.
• neural network — A layered network of mathematical functions loosely inspired by the brain. Training means tuning the weights between layers.
• open-weight — A model whose weights can be downloaded and run on your own hardware. Not the same as open source - the training data is usually not public.
• parameter — An adjustable number stored in the model. The count is a rough size indicator, not a quality one.
• post-training — The refinement after pre-training that makes a model useful and safe, including RLHF.
• pre-training — The first and longest phase of training, where the model acquires general language and world knowledge.
• prompt — The instruction you give the model - not just a question, but the whole input.
• red teaming — Deliberately attacking your own system to find where it breaks before someone else does.
• SLM (small language model) — A small language model that runs on a laptop or phone - narrower knowledge, but cheap, fast and offline.
• temperature — A setting controlling how predictable the output is - low is consistent, high is more creative and erratic.
• token — The unit of text a model works in - roughly three quarters of a word. Pricing and length limits are both counted in tokens.
• tokenizer — The component that splits text into tokens. Hungarian text usually consumes more tokens than the same content in English.
• training — Building the model by adjusting weights over a huge dataset. A one-off, very expensive operation.
• transformer — The 2017 architecture behind today's language models. Its key idea is attention - every word can look at every other word.
• weights — The parameter values produced by training. Publishing them is what makes a model open-weight.

Prompting 20 terms

• chain-of-thought (CoT) — Asking the model to show its working, not just the answer - measurably better on multi-step tasks.
• context engineering — The successor to prompt engineering - deciding what belongs in the context window, not just how to phrase the request.
• delimiter — A marker separating instructions from data inside a prompt.
• few-shot — Two to five examples in the prompt - the fastest way to teach a format or style.
• grounding — Tying a model's answer to a supplied source instead of letting it work from memory.
• jailbreak — A prompt crafted to get around a model's built-in limits.
• JSON mode — A setting that guarantees the response is valid JSON.
• meta-prompting — Having the model write or improve the prompt instead of struggling with the wording yourself.
• negative prompt — What you do not want in the output - weak for text, effective for images.
• one-shot — Giving exactly one worked example alongside the request.
• persona prompting — Assigning the model a role - helps with vocabulary and tone, does not improve factual accuracy.
• prompt engineering — Writing instructions deliberately so the model returns something usable.
• prompt injection — An attack where a page, email or document hides an instruction that the model follows as if it were yours.
• prompt library — An organised collection of proven prompts for a team, with naming rules and an owner.
• prompt template — A prompt skeleton with fill-in slots, reused many times.
• structured output — Output in a fixed format - JSON, a table, a list - so it can be passed on to the next tool.
• system prompt — The background instruction that applies to the whole conversation - role, tone, rules.
• tree of thoughts — An extension of chain-of-thought where several lines of reasoning run in parallel and one is chosen.
• user prompt — The specific request you type, interpreted inside the frame the system prompt sets.
• zero-shot — Asking for something with no examples, just a description of the task.

Models & market 17 terms

• AI search — Search that returns a synthesised answer with citations instead of a list of links.
• answer engine — Another name for an AI search engine, common in marketing usage.
• API — Programmatic access to a model without a user interface - what any custom integration needs.
• context caching — Storing a repeated part of the prompt so you do not pay for it on every request.
• context rot — The way a model uses earlier information less reliably as a conversation grows very long.
• data residency — Where your data is physically stored and processed.
• deprecation — Retiring a model. Your prompts often break when it happens.
• distillation — Training a smaller model using a larger one - how the cheap fast variants are made.
• extended thinking — A mode where the model reasons at length before answering.
• inference provider — A company that runs other people's models and sells access, often cheaper or faster than the maker.
• latency — The delay before the first character of the answer arrives - critical for chat and voice.
• mixture of experts (MoE) — An architecture where only part of the model activates per request, making a very large model relatively cheap to run.
• model router — A service that puts many models behind one access point and picks between them per task.
• rate limit — A cap on how many requests or tokens you may send in a given period.
• reasoning effort — A dial telling the model how long to think before answering. In 2026 this is a setting, not a separate model.
• sovereign AI — AI running under a given country's or region's jurisdiction - not the same as a server located there.
• test-time compute — The idea that spending more compute at answer time improves the result.

Data & RAG 18 terms

• chunk overlap — Deliberate overlap between consecutive chunks so a split does not cut an idea in half.
• chunking — Splitting documents into retrievable pieces. Bad splits are the most common cause of bad answers.
• embedding — A list of numbers representing the meaning of a piece of text - similar meanings land close together.
• fine-tuning — Further training an existing model on your own examples. Good for style and format, poor for facts - use RAG for those.
• GGUF — The common file format for running models locally. The Q4 variant is usually the sensible trade-off.
• hybrid search — Keyword and meaning-based search combined - usually the single biggest RAG improvement.
• knowledge base — The organised set of documents a system answers from.
• local LLM — A language model running on your own machine - private and offline, but less capable.
• LoRA — A cheap fine-tuning method that trains only a small add-on layer - widely used for style and character in image generation.
• quantization — Compressing a model to lower-precision numbers so it fits smaller hardware. Too much degrades quality.
• RAG (retrieval-augmented generation) — Letting the model retrieve the relevant passages from your documents before answering - your knowledge without retraining.
• re-ranking — A second pass where a more precise model reorders the retrieved results.
• retrieval — The step where relevant passages are fetched, before any answer is generated.
• self-hosted — Running the software on your own servers rather than a vendor's.
• semantic search — Searching by meaning rather than exact words - it finds matches even when you used different wording.
• vector — The mathematical form of an embedding - a point in a high-dimensional space.
• vector database — A database that stores embeddings and quickly finds the closest ones by meaning.
• VRAM — The memory on your graphics card - the real limit on which models run fast for you.

Agents 23 terms

• A2A (agent-to-agent) — A standard letting agents from different vendors talk to each other.
• agent loop — The repeating cycle an agent runs - think, call a tool, read the result, decide the next step.
• agent memory — What an agent retains between runs - and the separate questions of how long and what to forget.
• agentic — Describes a system that acts on its own rather than only generating text.
• AI agent — An AI system that acts rather than just answers - calling tools, checking results and retrying until the goal is met.
• browser agent — An agent that drives a browser - opening pages, filling forms, collecting data.
• computer use — An agent operating software the way a person does - looking at the screen, clicking, typing. For when there is no API.
• durable execution — The property that a long run resumes from where it stopped instead of starting over.
• eval — Repeatable quality measurement of your system against a fixed set of real tasks.
• function calling — The technical mechanism behind tool use - the model emits a structured request your code executes.
• human in the loop (HITL) — A design where a person approves certain steps - place it before risky actions, not before every one.
• least privilege — The principle of granting an agent only the access its task needs - excessive permissions are now the top agent risk.
• MCP (Model Context Protocol) — The standard for connecting an AI system to external tools and data - now the common layer of the agent ecosystem.
• MCP server — The component exposing one specific tool or data source over MCP.
• multi-agent — Several agents working on one task. Coordination has a real cost, so more is not always better.
• no-code — Tools that let you build automation by clicking, without writing code.
• observability — Being able to see afterwards what the system did and why - traces, steps, cost, failures.
• orchestration — Coordinating agents and steps - who does what, in what order, and what happens on failure.
• sandbox — An isolated environment where an agent can run without reaching the rest of your systems.
• skill — A packaged procedure an agent loads when relevant.
• subagent — A helper agent given one subtask in its own clean context window - the gain is context hygiene, not speed.
• tool use — When a model operates an external tool - searching, calculating, writing files, calling an API.
• workflow — A fixed sequence of steps. Unlike an agent, a workflow decides nothing - it executes.

Images 19 terms

• aspect ratio — The image's proportions - set it for the final placement instead of cropping later.
• C2PA / Content Credentials — An industry standard for content provenance, embedding a signed history in the file.
• character consistency — Keeping the same person or product recognisable across many images or shots.
• ControlNet — An add-on model forcing a pose, outline or depth map onto generation - still standard in the open-weights world.
• diffusion model — The dominant image-generation method - starting from noise and refining it into a picture step by step.
• image-to-image — Turning an existing image into a new one while partly keeping its composition.
• in-context editing — Editing by conversation or by marking the image, without building a mask.
• inpainting — Regenerating a selected area inside an image - removing, replacing or retouching.
• layout control — Getting the model to put things where you want them on the canvas - the weakest capability of today's image models.
• outpainting — Extending an image beyond its original frame, continuing the style.
• photorealism — Making generated images indistinguishable from photographs.
• prompt weighting — A technique for making some parts of an image prompt count more than others.
• provenance — The origin history of a piece of content - who made it, with what, and when.
• seed — The starting number for randomness. The same seed and prompt reproduce the same image.
• style reference — A reference image the model takes style from, but not content.
• SynthID — Google's invisible marking on its generated content, designed to survive re-encoding and editing.
• text-to-image — Generating an image from a written description.
• upscaling — Increasing resolution by inventing detail. What the photo never captured is guessed, not recovered.
• watermark — A visible or hidden mark showing where content came from.

Video & audio 15 terms

• avatar — A generated speaking figure that delivers your script.
• clip length — How much video one generation produces - typically eight to twenty seconds today.
• dubbing — Replacing a video's audio track with another language.
• image-to-video — Starting a video from a finished image - almost always the better route when the look matters.
• keyframe — A specified first or last frame, with the model generating the motion between.
• lipsync — Matching mouth movement to speech - what makes convincing dubbed video possible.
• music generation — Generating music from a description. Commercial usability depends on the model's licensing, not its price.
• native audio — When the video model generates sound itself rather than having it added later - standard in 2026.
• storyboard — A visual plan of the shots before production - what replaces guesswork in AI video.
• STT / ASR — Turning speech into text. Worth testing separately for smaller languages.
• talking head — The format where one person speaks to camera - what AI avatars handle best today.
• text-to-video — Generating video from a written description.
• TTS (text-to-speech) — Turning written text into spoken audio.
• voice cloning — Copying a specific person's voice from a short sample - legally and ethically fraught without consent.
• world model — A model that generates an explorable, responsive environment rather than a fixed clip.

Coding 15 terms

• AGENTS.md — A project-level instruction file for AI tools - conventions, commands, prohibitions.
• autocomplete — Line-level code completion as you type - the lowest-risk mode of AI coding.
• CLI agent — A coding agent that runs in the terminal - better for automation and remote machines.
• code review bot — Automated review on every change - worth it only if it finds real defects instead of noise.
• coding agent — An agent that writes, runs and fixes code in your project rather than only suggesting lines.
• diff — The line-by-line list of changes - always review AI-written code through it.
• IDE — The integrated development environment where a programmer works - where AI features usually live.
• prompt-driven development — A way of working where the specification is the prompt and code follows from it.
• pull request — A bundle of changes submitted for review and merging - how agents hand in their work too.
• repository indexing — Processing a codebase so the AI tool can find the relevant files by itself.
• SKILL.md — A packaged procedure an agent loads when the matching task comes up.
• slopsquatting — An attack where someone registers the non-existent package name that AI tools keep inventing.
• SWE-bench — A benchmark of real bug fixes used to rate coding models - the score is not the daily experience.
• technical debt — The accumulated cost of quick solutions - AI produces it faster than ever.
• vibe coding — Building software by describing what you want without reading the code produced. Fine for prototypes, not for production.

Marketing 16 terms

• AEO (answer engine optimization) — Another name for GEO - the two abbreviations describe the same work.
• AI literacy — The level of understanding staff need to use AI responsibly - EU rules require supporting its development.
• AI Overviews — Google's generated summary at the top of results - clicks drop sharply where it appears.
• brand visibility — How often your brand shows up in AI answers.
• citation — When an AI answer references your page - the main GEO metric.
• conversion — When a visitor takes the action you wanted. Visitors from AI search tend to convert better.
• GEO (generative engine optimization) — Optimising to be cited by AI search engines - a different game from classic SEO.
• llms.txt — A proposed file summarising a site's content for AI tools. No evidence yet that major engines use it.
• machine readability — Whether a machine can read your page's content - the most concrete, measurable part of GEO.
• pillar page — The comprehensive entry article for a topic that the rest of the cluster links to.
• pilot — A trial deployment before real use - where most AI projects get stuck permanently.
• ROI — Return on investment. With AI the usual problem is that nobody measures it at all.
• SEO — Search engine optimisation - ranking well in a classic results list.
• structured data — Markup telling machines what each part of a page is - product, article, author, price.
• topic cluster — A group of articles organised around one pillar page - how a site earns authority on a topic.
• zero-click — A search where the user clicks nothing - now the majority of searches.

Law & security 15 terms

• AI inventory — A register of which AI systems the company uses, for what, and who owns them.
• data leakage — Confidential data leaving the organisation through an AI tool.
• deepfake — Generated audio, image or video convincingly imitating a real person - labelling is mandatory in the EU.
• DLP — Data loss prevention - controls that stop confidential data from being sent out.
• EU AI Act — The EU's AI regulation. Transparency duties apply since August 2026; high-risk rules moved to late 2027.
• GDPR — The EU's data protection regulation - it governs what personal data you may put into an AI tool at all.
• GPAI — General-purpose AI model under EU law, with its own set of provider obligations.
• high-risk AI — The EU's more strictly regulated category - hiring, credit, education, critical infrastructure.
• ISO 42001 — A certifiable standard for AI management systems, increasingly asked for by larger customers.
• NIST AI RMF — A voluntary US framework for managing AI risk - not certifiable, but a useful checklist.
• OWASP LLM Top 10 — The ten most important security risks for AI applications - prompt injection has topped it for years.
• shadow AI — AI tools staff use without approval, on personal accounts. Banning does not work - inventorying does.
• TDM (text and data mining) — The copyright exception model training is argued to rely on - European courts are starting to narrow it.
• training data — The material a model was trained on - the subject of most copyright litigation.
• transparency obligation — The duty to disclose AI involvement - in chatbots, generated content and deepfakes.