AI comes with a lot of jargon. This glossary explains 523 terms in everyday words, with no maths needed. Each entry points to tools in our directory where you can see the idea at work. Looking for the tools themselves? Browse the full AI tools directory, the MCP servers list or the AI agents directory.
Showing two versions of something to different people to see which performs better. Teams use it to compare prompts, models and page designs.
See it in: ChatGPT
An experiment that removes one part of a system to see how much that part mattered.
See it in: Hugging Face
The share of answers a model gets right. It can mislead when one outcome is far more common than the other.
See it in: Hugging Face
A small formula inside a neural network that decides how strongly each unit responds. It lets the network learn curved, complex patterns.
See it in: Hugging Face
A way of training where the model asks people to label only the examples it finds most confusing.
See it in: Hugging Face
A small add-on layer trained on top of a frozen model so it learns a new job cheaply.
See it in: Hugging Face, ComfyUI
An input changed in a tiny way, often invisible to people, that makes a model give a wrong answer.
See it in: Hugging Face
An AI program that plans a task, uses tools and keeps working until it is done, instead of giving one answer and stopping.
A chatbot answers. An agent acts across several steps.
See it in: Genspark, Devin Desktop, Cline
A code library that gives developers building blocks for agents: memory, tools, planning and handoffs.
See our AI agents directory.
A process where a model makes decisions between steps, such as which tool to call next, rather than following a fixed script.
Notes an agent keeps between steps or sessions so it does not start from zero each time.
A tool that creates a working app from a text description.
A chat-style tool that helps with writing, questions and tasks on request.
A digital presenter that speaks your script, made from a photo or a recorded actor.
An agent that plans and edits code across a project, runs tests and fixes what breaks.
See it in: Cursor, Cline, Devin Desktop
A chat app built for friendly, ongoing conversation with a character.
See it in: Replika, Character.AI, Nomi AI
A tool that guesses whether text was written by a model. Results are unreliable, so treat them as a hint, not proof.
The study of what AI should and should not do, covering fairness, privacy, honesty and who is accountable.
A companion app marketed for romantic chat. Check the age rules, privacy policy and what data it keeps before using one.
The rules, laws and company policies that guide how AI is built and used.
Knowing enough about how AI works to use it well and to spot its mistakes.
A tool that joins meetings, records and writes up notes and action items.
See it in: Fathom, Otter.ai, Granola AI, Fireflies.ai
An assistant that works with you as you code, suggesting and explaining.
See it in: GitHub Copilot, Cursor
A search tool that reads results and writes one answer with sources.
See it in: Perplexity, Genspark
A tool that turns a topic or document into a slide deck.
An editor that cuts, captions or reformats video with help from AI.
A hidden or visible mark placed on AI-made content to show where it came from. Marks can often be removed.
See it in: Adobe Firefly Ai
A tool that drafts, rewrites and corrects text.
The work of making a model behave the way its builders and users intend, including refusing harmful requests.
The part of an image that stores transparency, so a subject can sit on any background.
See it in: Photoroom
AI that works quietly in the background, such as drafting notes from a conversation without being asked.
See it in: Granola AI
Adding labels to raw data, such as tagging objects in photos, so a model can learn from it.
See it in: Hugging Face
Software for adding labels to images, text or audio for training data.
See it in: Hugging Face
Spotting data points that look unlike the rest, such as a strange card payment.
See it in: Julius AI
A way for one program to talk to another. Most AI models can be used by apps through an API instead of a chat window.
See it in: Google AI Studio, Hugging Face
A front door that manages access, limits and billing for calls to several models.
See it in: Poe Ai
A secret code that proves your app is allowed to call an API. Keep it out of public code.
See it in: Google AI Studio
A style of evaluation where people vote between two anonymous answers.
See it in: Poe Ai
A separate panel where an assistant shows a document, page or app it built, so you can edit it beside the chat.
See it in: Claude
A hypothetical AI that can learn and do any intellectual task a person can. No current product is this.
The full name for a neural network: layers of simple units whose connections are tuned during training.
See it in: Hugging Face
The shape of a picture or video, such as 16:9 or 9:16. Image generators let you pick one.
See it in: Midjourney, CapCut
A reply written by the model in a chat history.
See it in: ChatGPT
The mechanism that lets a model weigh which words in the input matter most when producing each next word.
Creating sound, speech or music from a text description or a short sample.
See it in: Suno, Udio, ElevenLabs
Software that turns the speech in a video into subtitles on its own.
Producing text one token at a time, each based on all the ones before.
See it in: ChatGPT
A network that squeezes data into a short code and then rebuilds it. It is used for compression and cleaning noise.
See it in: Hugging Face
The human habit of trusting a machine's suggestion even when it is wrong.
See it in: ChatGPT
Tools that choose and tune a machine learning model for you, so you do not need to design it by hand.
See it in: Hugging Face
A model that produces output one piece at a time, using everything it has produced so far to pick the next piece.
Extra footage that covers the main video, such as cars driving while someone talks. AI tools can now create or suggest it.
See it in: CapCut, Runway Gen-3 Alpha
Cutting the subject out of a picture and deleting everything behind it.
See it in: Photoroom, Canva, Adobe Express
The method used in training that works out how much each internal setting contributed to an error, so it can be adjusted.
See it in: Hugging Face
An older way of reading text that counts words and ignores their order.
See it in: Hugging Face
A model after pre-training and before it is tuned to follow instructions or chat.
See it in: Hugging Face
An API option for sending many requests at once, usually cheaper but slower to return.
See it in: Google AI Studio
Running many inputs through a model together to save time and cost.
See it in: Ollama
How many training examples a model looks at before it updates itself.
See it in: Hugging Face
A way of generating text that keeps several candidate sentences alive and picks the best overall one.
See it in: Hugging Face
A fixed set of tests used to compare models. Scores are useful but can be gamed, so check how a model works on your own tasks.
When test questions leak into training data, making a model's score look better than its real skill.
See it in: ChatGPT
A language model from Google that reads text in both directions. It shaped search and many classifiers.
See it in: Hugging Face
A pattern in a model's output that favours some groups or views, usually inherited from its training data.
Steps taken to reduce unfair patterns in data or output.
See it in: Claude
A rectangle drawn around an object in an image to mark where it is.
See it in: Hugging Face
A setup where you paste your own API key into an app so usage is billed to your account.
See it in: Cursor, Open WebUI
An agent that opens pages, clicks and types like a person to finish a web task.
See it in: Genspark
How well a model's stated confidence matches how often it is actually right.
See it in: Claude
Something a model can do, such as writing code, reading images or following long instructions.
A tool that writes the text that goes under a photo or video.
When a model loses old skills after being trained on something new.
See it in: Hugging Face
Asking a model to write out its reasoning steps before the answer. It often improves results on maths and logic.
Keeping the same face and outfit across many generated images or video shots.
See it in: Midjourney, Kling AI, Leonardo.Ai
The API call that sends a conversation to a model and returns its next message.
See it in: Google AI Studio
A program you talk to in a conversation. Modern ones are powered by language models.
See it in: ChatGPT, Character.AI, Poe Ai
A tool for making a customer-facing chat assistant without much code.
See it in: Intercom Fin, Softr AI
A saved copy of a model at one point in training. People share checkpoints so others can reuse or fine-tune them.
See it in: Hugging Face, ComfyUI
Splitting a long document into smaller pieces so a retrieval system can find the right one.
See it in: Gemini Notebook AI, Open WebUI
Sorting inputs into categories, such as spam or not spam.
See it in: Hugging Face
A model that links images and text, so it can tell whether a caption fits a picture. Many image tools rely on it.
See it in: Midjourney, ComfyUI
A tool that cuts long videos into short, shareable clips.
A model whose weights are not shared. You use it only through the maker's app or API.
Running models on a company's servers and using them over the internet.
Grouping items that are alike without being told the groups in advance.
See it in: Julius AI
Suggestions that finish the line or block of code you are typing.
See it in: GitHub Copilot, Cursor, Zed
A feature that lets a model write and run code to calculate, chart or transform your files.
The locked-down space where a chat assistant runs code safely.
See it in: ChatGPT
A tool that suggests, explains or writes code inside your editor.
See it in: GitHub Copilot, Cursor, Zed
A plan for how an AI system combines memory, reasoning and action.
See it in: Genspark
The delay when a model or service has to load before it can answer.
See it in: Ollama
Adjusting colour and contrast to give a video a look. AI tools can match a reference.
See it in: CapCut, Runway Gen-3 Alpha
Assistants you use from a terminal window instead of a chat app.
See it in: Cline
Permission to use something, such as an AI image, in paid work. Terms differ by tool and plan.
See it in: Adobe Firefly Ai, Midjourney
The processing power used to train or run a model.
See it in: Ollama
A model's ability to look at a screen, move the mouse and type to operate software.
See it in: Claude
The branch of AI that works with images and video, such as recognising objects or reading text in photos.
See it in: Photoroom, Hugging Face
Creating output guided by an input, such as an image from a caption.
See it in: Midjourney
A number showing how sure a model is about an answer. Treat it with care, since models can be sure and wrong.
See it in: Claude
A table that shows which kinds of mistakes a classifier makes.
See it in: Hugging Face
A training method where a model critiques and improves its own answers against a written set of principles.
See it in: Claude
Metadata attached to a file that records how it was made, including whether AI was used.
See it in: Adobe Firefly Ai
A rule or model that blocks certain kinds of output.
See it in: ChatGPT
Using people or models to detect and remove harmful posts, images or comments.
See it in: Intercom Fin
Storing a long document in a model's memory for reuse across questions.
See it in: Gemini
Choosing what information goes into a model's context so it has what it needs and nothing distracting.
Another name for the context window: the amount of text a model can read in one go.
How much text a model can keep in mind at once, counted in tokens. Anything beyond it is forgotten.
Longer windows help with big documents but can cost more.
A serving method that adds new requests to a running batch, so GPUs stay busy and replies are faster.
See it in: Ollama
Steering what a model produces with extra inputs such as a pose, a sketch or a style reference.
An add-on for image models that makes them follow an outline, pose or depth map.
See it in: ComfyUI, Leonardo.Ai
What a chatbot remembers from earlier in the same chat.
Technology that lets software hold natural back-and-forth dialogue by text or voice.
See it in: ChatGPT, Intercom Fin
A network built to read images by scanning small patches. It powered the first big image-recognition gains.
See it in: Hugging Face
A name for an AI helper that works alongside you in an app instead of replacing you.
See it in: GitHub Copilot, Microsoft Copilot
Microsoft's tool for building your own copilots and agents.
See it in: Microsoft Copilot
The open question of who owns AI output and whether training on public work is allowed. Rules differ by country and are changing.
See it in: Adobe Firefly Ai
A large collection of text used for training or study.
See it in: Hugging Face
A way to measure how close two embeddings are by the angle between them.
See it in: Open WebUI
What a provider charges for each token, with separate rates for input and output.
An upscaler that invents new detail while enlarging a picture, rather than only sharpening.
See it in: Krea AI
Points you spend on generations in image and video tools instead of paying per use.
See it in: Midjourney, Kling AI
A group of agents working on one job together, a term used by some agent frameworks.
See it in: n8n
Testing a model several times on different slices of the data to check it holds up.
See it in: Julius AI
A version of ChatGPT set up with its own instructions, files and tools for one job.
See it in: ChatGPT
Creating extra training examples by changing the ones you have, for example flipping or cropping images.
See it in: Hugging Face
The work of tagging examples with the right answers so supervised learning is possible.
See it in: Hugging Face
When information from the test set slips into training, which makes results look better than they are.
See it in: Hugging Face
The steps that collect, clean and deliver data to a model or report.
Slipping bad examples into training data to damage or steer a model.
See it in: Hugging Face
Keeping personal information safe. Check what an AI tool stores, and whether it trains on your inputs.
How long a company keeps what you send it.
A collection of examples used to train or test a model.
See it in: Hugging Face
A model that makes a choice by answering a chain of yes-or-no questions.
See it in: Julius AI
The part of a model that produces output. Most chat models are decoder-only.
Tools and methods that try to spot AI-made faces or voices.
See it in: Deepfake Maker
Machine learning that uses neural networks with many layers. It powers most modern AI.
See it in: Hugging Face
A mode where an assistant searches many sources over several minutes and writes a long report with citations.
See it in: Perplexity, ChatGPT, Gemini
Fake video, audio or images of a real person made with AI. Making one without consent can be illegal.
Always label synthetic media.
See it in: Deepfake Maker, HeyGen
Removing noise, whether grain in a photo, hiss in audio or the noise a diffusion model starts from.
Putting a finished model or app where people can use it.
An image where each pixel shows how far away that point is. Tools use it to add depth or control a scene.
See it in: ComfyUI
A setting that opens up extra tools or access in an AI app for builders.
See it in: Cursor
A way to share results about a group while keeping any one person's data hidden.
See it in: Claude
An image or video model that starts from noise and removes it step by step until a picture appears.
See it in: Midjourney, ComfyUI, Runway Gen-3 Alpha
Squashing data with many features into fewer ones while keeping its structure, often for charts.
See it in: Julius AI
A simpler alternative to RLHF that trains a model straight from pairs of better and worse answers.
See it in: Hugging Face
Training a small model to copy the behaviour of a larger one, so it runs faster and cheaper.
See it in: DeepSeek, Hugging Face
Diffusion transformer: a design that pairs diffusion with transformer layers, used in many new video models.
Software that reads forms, invoices and contracts and pulls out the key fields.
See it in: AssemblyAI
Asking questions of a PDF or file and getting answers from its content.
See it in: Gemini Notebook AI, ChatGPT
Adjusting a model trained in one setting so it works in another, such as legal text instead of news.
See it in: Hugging Face
A faster, lower-quality setting for trying ideas before a full render.
See it in: Midjourney
A training trick that randomly switches off units so the network does not memorise.
See it in: Hugging Face
Replacing a video's speech with another language. AI dubbing can also match the original voice and lip movement.
See it in: HeyGen, ElevenLabs
Building a prompt in code from changing data, such as a customer's name.
Ending training when results stop improving to avoid overfitting.
See it in: Hugging Face
Running models on a phone, camera or device instead of in the cloud.
A rare or odd situation that a model or program may handle badly.
See it in: Claude
A list of numbers that captures the meaning of a piece of text or an image, so similar things sit close together.
See it in: Hugging Face
The length of the number list that stands for one item in an embedding.
See it in: Hugging Face
A model made to turn text or images into embeddings, used for search and recommendations.
See it in: Hugging Face
A skill that shows up only once a model gets big enough, such as following multi-step instructions.
Software that tries to read feelings from faces, voice or text. Accuracy claims should be treated with care.
See it in: Fathom
The part of a model that turns input into a rich internal representation.
See it in: Hugging Face
A model that handles a whole task directly, with no hand-built steps in between.
See it in: Gemini
Combining several models and averaging their answers for steadier results.
See it in: Hugging Face
AI products with admin controls, security and support for companies.
See it in: Microsoft Copilot, WRITER AI
A business tier that usually adds security controls and no training on your data.
Finding names, places, dates and companies in text.
See it in: Hugging Face
One full pass through the training data.
See it in: Hugging Face
The European law that sets rules for AI systems based on how risky they are.
See it in: ChatGPT
Short for evaluations: repeatable tests that check how a model or prompt performs.
See it in: Hugging Face
Testing how well a model does on tasks you care about.
See it in: Hugging Face
An older type of AI that follows hand-written rules from human specialists.
See it in: Harvey AI
How well a person can understand why a model gave a certain answer.
See it in: Claude
Replacing one person's face with another in a photo or video. Get consent first.
See it in: Deepfake Maker
A legal idea that allows some copying without permission. Whether it covers AI training is being decided in courts.
See it in: Adobe Firefly Ai
A measurable property of the data, such as the size of a house, that a model uses to decide.
See it in: Julius AI
Choosing or building the inputs a model learns from.
See it in: Julius AI
Training a shared model across many devices without collecting their raw data in one place.
See it in: Hugging Face
When a model's output affects the data it later learns from, which can reinforce its own mistakes.
See it in: ChatGPT
Learning a task from a handful of examples.
Giving a model a few examples in the prompt so it copies the pattern.
A feature that lets an assistant look through documents you upload.
See it in: ChatGPT, Gemini Notebook AI
Having precise settings over what a model makes, rather than one prompt box.
See it in: ComfyUI
Training an existing model a bit more on your own examples so it fits a job or style better.
See it in: Hugging Face, Google AI Studio
A video method where you give the start and end pictures and the model fills in the motion.
A faster, lighter way to compute attention that lets models handle longer inputs.
See it in: Ollama
A family of modern open image models known for sharp detail and good text in pictures.
A large model trained on broad data that can be adapted to many tasks.
Adding in-between frames to make video smoother or create slow motion.
See it in: CapCut, Runway Gen-3 Alpha
The no-cost level of a tool, usually with lower limits.
A pricing model with a free version and paid upgrades.
A setting that discourages a model from repeating the same words.
See it in: Google AI Studio
One of the most capable models available at a given time.
A model whose weights are locked during training, so only the added parts change.
See it in: Hugging Face
A feature that lets a model ask your app to run a specific function, such as checking the weather, and use the result.
See it in: ChatGPT, Claude, Google AI Studio
A pair of networks that compete: one makes fakes, the other tries to spot them. Older image tech, now mostly replaced by diffusion.
See it in: Hugging Face
A way of capturing a real scene from photos or video as millions of tiny soft blobs you can view from any angle.
See it in: Luma AI
The European privacy law that governs how personal data is collected and used.
See it in: Claude
Google's family of smaller open-weights models that can run on a laptop.
How well a model does on new data it has never seen.
See it in: Hugging Face
The full name for a GAN: two networks that train each other, a maker and a judge.
See it in: Hugging Face
AI that creates new content such as text, images, music or video.
See it in: ChatGPT, Midjourney, Suno
A feature that fills or changes part of an image from a text instruction.
See it in: Adobe Firefly Ai, Photoroom
Interfaces that an AI creates on the fly for the task you are doing.
A file format for running quantized models on ordinary computers.
A small, carefully checked set of examples used to test changes.
See it in: Hugging Face
Generative pre-trained transformer: the type of model behind ChatGPT.
See it in: ChatGPT
A directory of custom GPTs made by other people.
See it in: ChatGPT
The kind of chip used to train and run most AI models because it does many calculations at once.
The basic method for training: nudge each setting a little in the direction that lowers the error.
See it in: Hugging Face
Always picking the most likely next word. It is steady but can sound flat.
See it in: ChatGPT
The correct answer used to check or train a model.
See it in: Hugging Face
Writing answers that stay tied to supplied sources.
See it in: Perplexity, Gemini Notebook AI
Tying a model's answer to real sources so it makes fewer things up.
See it in: Perplexity, Consensus
A separate model that checks inputs and outputs for unsafe content.
See it in: Claude
Rules and checks around a model to stop unsafe or off-topic output.
See it in: Intercom Fin, Claude
A setting in image generators that controls how strictly the picture follows your prompt.
See it in: ComfyUI, Leonardo.Ai
When a model states something false with confidence. Always check names, numbers and quotes.
See it in: Perplexity, Consensus, Elicit
When one agent passes a task to another agent or to a person.
See it in: Intercom Fin
A web browser that runs with no window, used by programs and agents to visit pages.
See it in: Genspark
A large hub where people share models, datasets and demos.
See it in: Hugging Face
People rating a model's output, which is slow but catches what scores miss.
See it in: Claude
A setup where a person reviews or approves the AI's work at key moments.
Search that mixes keyword matching with meaning-based matching to give better results.
See it in: Perplexity
A setting chosen before training, such as learning rate, that changes how training behaves.
See it in: Hugging Face
Trying different settings before training to find the combination that works best.
See it in: Hugging Face
Software that describes a picture in words.
Deciding what category a picture belongs to, such as cat or dog.
See it in: Hugging Face
Creating pictures from a written description.
See it in: Midjourney, Ideogram, Leonardo.Ai
A reference picture you give a generator along with, or instead of, text.
See it in: Midjourney, Leonardo.Ai
Marking exactly which pixels belong to which object in a picture.
See it in: Photoroom
A tool that makes a small picture larger and sharper.
Turning a still picture into a short moving clip.
See it in: Kling AI, Runway Gen-3 Alpha, Pika AI
Training a system by showing it examples of an expert doing the task.
See it in: Hugging Face
A model picking up a task from examples in the prompt, with no retraining.
Running a trained model to get an answer, as opposed to training it.
What it costs to run a model for each answer, usually counted per token.
The field of finding the right documents for a question. Search engines and RAG both rely on it.
See it in: Perplexity
Editing part of an image by painting over an area and describing what should appear there.
See it in: Adobe Firefly Ai, Leonardo.Ai, Photoroom
A token you send to a model. Providers usually charge less for input than for output.
How well a model does exactly what you asked, including format and length limits.
Training a model on examples of instructions and good replies so it follows requests better.
Software that reads business documents and sorts or extracts what matters.
See it in: AssemblyAI
Working out what a user wants from what they typed, such as cancel an order or ask a price.
See it in: Intercom Fin
Research into what is happening inside a model and why it answers as it does.
See it in: Claude
A prompt designed to trick a model into ignoring its safety rules.
A setting that makes a model reply in valid JSON.
See it in: Google AI Studio
Finding the words people search for. AI tools can group and suggest them.
A collection of articles or documents a chatbot can answer from.
See it in: Intercom Fin
The date after which a model has no training data, so it does not know later events unless it can search.
See it in: Perplexity, ChatGPT
Another name for distillation: a small model learns from a big one.
See it in: DeepSeek
A map of facts and the links between them, used to answer questions precisely.
See it in: Connected Papers
Saved intermediate results that let a model avoid rereading the whole conversation for each new word.
See it in: Ollama
The correct answer attached to a training example.
See it in: Hugging Face
A popular toolkit for connecting models to tools, data and agents.
See it in: n8n
Detecting which language a text or recording is in.
See it in: DeepL AI, AssemblyAI
A model that predicts what text comes next. Bigger ones are called large language models.
A model trained on huge amounts of text to predict and generate language. ChatGPT and Claude are built on them.
The wait between sending a request and getting the first part of the answer.
The hidden internal space where a model represents concepts as numbers.
See it in: ComfyUI
One level of a neural network. Deep networks stack many layers.
See it in: Hugging Face
A ranked list of models by test score or by votes from users.
See it in: Hugging Face
How big a step a model takes when it updates during training.
See it in: Hugging Face
Matching a mouth's movement to spoken audio in a video.
A video where the mouth matches new audio, often used to dub or to animate a portrait.
See it in: HeyGen
Turning speech into text as it is spoken.
See it in: Otter.ai, Fireflies.ai
Meta's family of open-weights language models that people run and adapt themselves.
Short for large language model.
Using one model to grade another model's answers.
See it in: Claude
A model that runs on your own computer, so your data stays with you.
Tools designed to run on your own machine so your files stay private.
The raw scores a model gives to each possible next token before they turn into probabilities.
See it in: Hugging Face
The probability a model gave to each token it chose, sometimes shown for debugging.
See it in: Google AI Studio
A model's ability to read very long inputs, such as whole books or codebases.
Articles and videos that run long. AI can help draft, but check facts and keep the voice yours.
See it in: Jasper
A clip that repeats smoothly with no visible jump.
A light way to fine-tune a model by training a small add-on instead of the whole thing. Popular for image styles.
See it in: ComfyUI, Hugging Face
A score of how wrong a model is during training. Training tries to push it down.
See it in: Hugging Face
Building apps with visual blocks and a little code instead of writing everything by hand.
Software that learns patterns from data instead of following hand-written rules.
See it in: Hugging Face
Automatic translation between languages.
Software that sends emails, posts and follow-ups on a schedule.
A model trained by hiding words in a sentence and guessing them, as BERT was.
See it in: Hugging Face
A setting that limits how long a reply can be.
See it in: Google AI Studio
Model Context Protocol, an open standard that lets AI assistants connect to tools and data in a common way.
See our MCP servers directory.
A common score of prediction error that punishes big misses more than small ones.
See it in: Julius AI
A bot that records and summarizes your calls.
See it in: Fathom, tl;dv, Fireflies.ai, Granola AI
A feature that lets an assistant remember facts about you between chats.
A separate service that stores what an agent learns so it can use it later.
See it in: ChatGPT
Teaching a model how to learn, so it picks up new tasks from few examples.
See it in: Hugging Face
Blending digital objects into the real world through a headset or phone.
See it in: Luma AI
A design where only some parts of a large model run for each input, which saves cost.
See it in: DeepSeek, Mistral Vibe AI
The trained program that turns inputs into outputs. Different models suit different jobs.
See it in: Hugging Face
A short document describing what a model does, how it was trained and its limits.
See it in: Hugging Face
Quality loss that happens when models are trained again and again on AI-made content.
See it in: ChatGPT
Running the same prompt through several models to see which does better.
See it in: Poe Ai
Everything a model can see for one answer: instructions, chat, files and tool results.
See it in: Claude
When a provider retires an old model version. Apps that depend on it have to move to a newer one.
See it in: ChatGPT
When a model gets worse over time because the real world changes from its training data.
See it in: Julius AI
A site where people publish and download models.
See it in: Hugging Face
The record of where a model and its data came from.
See it in: Hugging Face
Sending each request to the cheapest model that can handle it.
The setup that keeps a trained model running so apps can send it requests.
See it in: Ollama, Hugging Face
How many parameters a model has. Bigger often means smarter but slower and costlier.
See it in: Ollama
A document describing how a model is meant to behave.
A young research area asking whether and how AI systems' interests, if any, should be considered.
See it in: Claude
A tool that lets you paint over part of a picture to say what should move.
See it in: Runway Gen-3 Alpha, Kling AI
Recording a person's movement so it can drive a digital character. AI can now do it from plain video.
See it in: Runway Gen-3 Alpha, Kling AI
Animated text, shapes and icons used in videos.
Several agents working together, each with a role such as researcher, writer or checker.
See it in: Genspark
A chat where earlier messages shape later answers.
A model that works in many languages.
A model that works with more than one kind of input, such as text, images and audio.
Creating songs or instrumentals from a text description.
A specific person, place, company or date mentioned in text.
See it in: Hugging Face
Controlling software by typing or saying what you want in plain words.
See it in: ChatGPT
Teaching computers to work with human language.
Words describing what an image generator should avoid.
See it in: ComfyUI, Leonardo.Ai
A system of connected layers loosely inspired by the brain, which learns by adjusting its connections.
See it in: Hugging Face
Using neural networks to draw or relight images and 3D scenes.
See it in: Luma AI
A synthetic voice made by a neural network that sounds natural.
See it in: ElevenLabs, Murf AI
An agent you set up with forms and drag-and-drop instead of code.
Tools that let you build AI apps by clicking and typing, with no programming.
A way of building by linking boxes on a canvas. ComfyUI uses it for image pipelines.
Random variation. In image models it is the static a picture starts from before it is cleaned up.
See it in: ComfyUI
Removing background noise from audio or grain from video.
Scaling data to a standard range so training is steadier.
See it in: Hugging Face
A document that mixes text, code and results, popular for data work. Some AI tools use the word for source-based chat.
See it in: Gemini Notebook AI, Julius AI
Finding and labelling the things in a picture and where they are.
See it in: Hugging Face
Tools that let you see what an AI app is doing, including prompts, costs and errors.
See it in: n8n
Optical character recognition: reading text from images and scans.
Models that run on a phone or laptop with no internet needed.
Giving a model a single example in the prompt before asking it to do the task.
A model with public weights that you can download and run.
Sharing methods and results publicly so others can check and build on them.
See it in: Hugging Face
A model whose trained numbers are published, even if the training data and code are not.
The terms that say how code or a model may be used, changed and shared. Read it before using one commercially.
See it in: Ollama, Hugging Face
A model whose weights are shared so anyone can run or adapt it. Check the licence for limits.
See it in: Ollama, Open WebUI, DeepSeek
An API that copies OpenAI's format, so apps built for it can switch to other models easily.
A setting that tells a provider not to use your chats to train its models.
A setting that stays off until you choose to turn it on.
See it in: ChatGPT
Adjusting a model's settings to push its error as low as possible.
See it in: Hugging Face
The routine that updates a model's settings during training.
See it in: Hugging Face
Coordinating several models, tools and steps so they work as one system.
Extending an image beyond its original edges.
See it in: Adobe Firefly Ai, Leonardo.Ai
The shape you ask a model's answer to take, such as a table, list or JSON.
See it in: ChatGPT
A token a model writes back. These usually cost more than input tokens.
When a model memorises its training data and does badly on new data.
See it in: Hugging Face
Trusting AI output without checking it.
See it in: ChatGPT
The subscription level with higher limits or better models.
One of the numbers inside a model that is learned in training. More parameters usually means a bigger, costlier model.
See it in: Hugging Face
Saying the same thing in different words.
Parameter-efficient fine-tuning: methods like LoRA that train only a tiny part of a model.
See it in: Hugging Face
A measure of how surprised a language model is by text. Lower means it predicts better.
See it in: Hugging Face
A character or role you give an assistant, such as a patient tutor.
See it in: Character.AI, Replika
Adapting answers or recommendations to one person's habits and preferences.
See it in: ChatGPT
A basic unit of sound in speech. Voice tools use them to pronounce words correctly.
See it in: ElevenLabs, Murf AI
Repairing old or damaged pictures with AI.
A web page for trying a model with different settings before you build anything.
See it in: Google AI Studio, Hugging Face
An add-on that lets an assistant use an outside service.
See it in: ChatGPT
A step that shrinks a picture or sequence into a smaller summary of its main features.
See it in: Hugging Face
Finding where a person's joints are in an image or video.
See it in: ComfyUI
The first, very large training stage where a model learns from broad data.
Two scores: precision asks how many flagged items were right, recall asks how many real ones were found.
See it in: Hugging Face
A model that estimates a future or unknown value from known data.
See it in: Julius AI
The text or instruction you give a model.
Reusing the already-processed start of a prompt so repeated requests are quicker and cheaper.
See it in: Claude
Breaking a job into steps where the output of one prompt feeds the next.
Writing prompts carefully to get better output: clear task, context, format and examples.
See it in: ChatGPT, Claude, Midjourney
An attack where text hidden in a page or file tells a model to ignore its real instructions.
A saved set of prompts you reuse.
See it in: ChatGPT
A site where people buy and sell prompts.
See it in: Midjourney
A reusable prompt with blanks you fill in each time.
Saving each version of a prompt so you can compare and roll back.
See it in: ChatGPT
Tools that catch spelling, grammar and tone problems.
Cutting parts of a model that matter little, so it gets smaller and faster.
See it in: Hugging Face
Shrinking a model by storing its numbers with less precision so it runs on smaller hardware.
The question or search you send to a system.
See it in: Perplexity
Systems that answer a question from a text or from stored knowledge.
See it in: Perplexity, ChatGPT
Retrieval-augmented generation: the model first looks up relevant documents, then answers using them.
See it in: Perplexity, Gemini Notebook AI, Open WebUI
A model that combines many decision trees and averages their votes.
See it in: Julius AI
A cap on how many requests you can send to an API in a set time.
See it in: Google AI Studio
An API for live back-and-forth voice or video conversations with a model.
See it in: ChatGPT
A model that spends extra time thinking through a problem before it replies.
Software that suggests items you may like based on what you and others did before.
See it in: Julius AI
An older kind of network that reads sequences one step at a time, now mostly replaced by transformers.
See it in: Hugging Face
Deliberately trying to make a model misbehave so its weaknesses can be fixed.
A picture you provide so a model copies its style, subject or layout.
See it in: Midjourney, Leonardo.Ai
When a model declines a request. Over-refusal is when it declines something that is actually fine.
Predicting a number, such as a price, rather than a category.
See it in: Julius AI
A check that new changes did not break something that used to work.
See it in: Cline
Methods that keep a model simple so it does not overfit.
See it in: Hugging Face
Training by trial and reward: actions that lead to good results are reinforced.
See it in: DeepSeek
Producing the final picture or video from a project.
See it in: CapCut, Runway Gen-3 Alpha
Being able to get the same result by following the same steps.
See it in: Hugging Face
Turning one piece, such as a long video, into many shorter formats.
A second pass that reorders search results so the most relevant come first.
See it in: Perplexity
A tool that finds, reads and summarizes papers.
See it in: Elicit, Consensus, Scite.ai, ResearchRabbit
Building and using AI with attention to fairness, safety, privacy and accountability.
See it in: Claude
Finding the most relevant pieces of information from a large collection to feed a model.
See it in: Perplexity, Gemini Notebook AI
The long form of RAG: look up documents first, then write an answer from them.
See it in: Perplexity, Gemini Notebook AI
Finding where a picture came from by searching with the image.
See it in: Perplexity
A model trained to score answers the way people would, used to guide RLHF.
Reinforcement learning from human feedback: people rank answers and the model learns to prefer the better ones.
Machines that sense and act in the physical world. AI helps them see, plan and move.
See it in: Hugging Face
Telling a model to act as a certain kind of person, such as an editor, to shape its answers.
The practice of reducing harm from AI, from bad advice to misuse.
How a model picks the next word from its list of likely options.
See it in: ChatGPT
A safe, walled-off space where code or an agent can run without touching your real files.
The finding that models tend to get better in predictable ways as data, size and compute grow.
Software that finds where shots change in a video.
Collecting data from web pages automatically. Check each site's terms first.
See it in: Genspark
A video made from a written script with an avatar or voice-over.
A number that fixes the starting randomness, so you can recreate the same image or result.
See it in: Midjourney, ComfyUI
A fixed number that makes a random process repeat the same way.
See it in: ComfyUI
Attention where every word in a sentence looks at every other word to build meaning.
Running a model or AI app on your own computer or server instead of someone else's cloud.
See it in: Ollama, Open WebUI, Jan
Learning from data by creating the answers from the data itself, such as hiding a word and guessing it.
See it in: Hugging Face
Microsoft's open toolkit for adding AI to apps.
See it in: Microsoft Copilot
Search that matches meaning instead of exact words.
See it in: Perplexity, Elicit
Working out whether text expresses a positive, negative or neutral feeling.
See it in: Fathom
A plan of the shots a video needs. AI can draft one from a script.
See it in: Runway Gen-3 Alpha
A compact model that is cheaper and faster, and can run on a phone or laptop.
Short suggested answers to a message.
See it in: Grammarly
A security audit report that companies show to prove they handle data carefully.
See it in: Notion
A formula that turns raw scores into probabilities that add up to one.
See it in: Hugging Face
A link or note showing where an answer's facts came from.
See it in: Perplexity, Consensus
AI that understands 3D space, such as where objects are in a room.
See it in: Luma AI
Working out who spoke when in a recording.
See it in: Otter.ai, Fireflies.ai, AssemblyAI
A speed trick where a small model drafts text and a large one checks it.
See it in: Ollama
A model that works with audio, either to listen or to speak.
See it in: ElevenLabs, AssemblyAI
Turning spoken audio into text.
See it in: Otter.ai, AssemblyAI, Fireflies.ai
Turning text into spoken audio. Also called text-to-speech.
See it in: ElevenLabs, Murf AI
Turning spoken words into written text.
See it in: Otter.ai, AssemblyAI, Descript
Single sign-on: logging in to many work tools with one company account.
See it in: Notion
A widely used open image model that can run on your own computer.
See it in: ComfyUI, Leonardo.Ai
A string that tells a model to stop writing when it appears.
See it in: Google AI Studio
A set of pictures that plans out a video, scene by scene.
See it in: Midjourney, Runway Gen-3 Alpha
Showing an answer word by word as it is produced instead of waiting for the whole thing.
Making a model reply in a fixed format such as JSON, so other software can read it.
See it in: Claude, Google AI Studio
An image that tells a generator which look to copy.
See it in: Midjourney, Krea AI
Applying the look of one image, such as a painting style, to another.
See it in: Adobe Firefly Ai, Krea AI
A tool that makes captions for a video.
Shortening a long text while keeping the main points.
See it in: ChatGPT, Gemini Notebook AI, Fathom
Training on examples that come with the right answers.
See it in: Hugging Face
A classic method that finds the best dividing line between two groups.
See it in: Julius AI
A model agreeing with you too much instead of giving an honest answer.
Data made by a model or program instead of collected from the real world, used to train or test other models.
See it in: Hugging Face
Images, audio or video made or changed by AI.
See it in: HeyGen, Synthesia, Deepfake Maker
Another name for the system prompt: standing instructions that frame every reply.
Hidden instructions that set a model's role and rules before the user says anything.
A setting that controls randomness. Low gives steady answers, high gives more varied ones.
See it in: Google AI Studio, ChatGPT
The rules you accept when you use a tool, including what happens to your data.
See it in: ChatGPT
Data kept aside and used only at the end to check how well a model works.
See it in: Hugging Face
Extra computing used while answering, so the model can think longer on hard questions.
Sorting text into categories, such as spam, topic or mood.
See it in: Hugging Face
A list of numbers that stands for the meaning of a passage.
See it in: Hugging Face
Changing wording while keeping the meaning.
A tool that condenses long text into a short version.
See it in: QuillBot, Gemini Notebook AI
Creating a picture from a text description.
See it in: Midjourney, Ideogram, Recraft
Software that reads text aloud in a synthetic voice.
See it in: ElevenLabs, Murf AI
Creating a video clip from a written description.
See it in: Kling AI, Luma AI, Hailuo AI, Runway Gen-3 Alpha
A cap on how much a reasoning model may spend thinking before it answers.
A tool that makes the cover image for a video.
Data recorded over time, such as daily sales or temperature.
See it in: Julius AI
Cutting video by moving clips along a timeline. Some tools now let you edit video by editing its transcript.
A small chunk of text, often part of a word, that models read and write. Pricing and limits are counted in tokens.
The most tokens a model can take in or give back in one request.
The tool that splits text into tokens before a model reads it.
See it in: Hugging Face
The personality of writing, such as friendly or formal. Many writing tools let you set it.
Another name for function calling: the model asks your app to run a tool and uses the result.
A model calling outside tools such as search, a calculator or your calendar to finish a job.
A way of limiting the model's next-word choices to the most likely group that adds up to a set chance.
See it in: Google AI Studio
Recording every step an agent takes so you can debug it later.
See it in: n8n
The examples a model learns from. Its quality shapes the model's strengths and flaws.
See it in: Hugging Face
One full attempt at training a model, from start to finish.
See it in: Hugging Face
Editing audio or video by changing the written text of what was said.
See it in: Descript
Writing out what was said in audio or video.
Reusing what a model learned on one task as the starting point for another.
See it in: Hugging Face
The model design behind most modern language and image AI, built around attention.
A model that converts text from one language to another.
A prompting method where a model explores several lines of reasoning and keeps the best.
See it in: ChatGPT
A training signal that pulls similar examples together and pushes different ones apart.
See it in: Hugging Face
A thought experiment: can a person tell a machine from a human in conversation?
See it in: ChatGPT
How unsure a model is. Good systems show it instead of hiding it.
See it in: Claude
When a model is too simple to learn the pattern in the data.
See it in: Hugging Face
Information with no fixed format, such as emails, photos and recordings.
See it in: AssemblyAI
Training on data without labelled answers, so the model finds patterns itself.
See it in: Hugging Face
Increasing an image or video's resolution and adding believable detail.
The cap on how much you can use a tool in a day or month.
The message a person sends, as opposed to the hidden system prompt.
See it in: ChatGPT
A slice of data used while building a model to tune it, separate from the final test set.
See it in: Hugging Face
A list of numbers. In AI it often stands for the meaning or features of something.
See it in: Hugging Face
A database built to store embeddings and find the most similar items quickly.
See it in: Open WebUI
Building software by describing what you want to an AI and steering the result, with little hand-written code.
Still read and test what it makes.
Creating moving clips from text, images or other video.
See it in: Kling AI, Luma AI, Pika AI, Runway Gen-3 Alpha
Models that watch a video and answer questions about what happens in it.
Making a video sharper or larger in resolution.
Showing how clothes would look on a person or a model using AI images.
See it in: Photoroom
A model that can understand images, such as reading a chart or describing a photo.
Searching with a picture instead of words.
See it in: Gemini
An assistant that talks and listens in real time, for example to answer phone calls.
See it in: ElevenLabs, Intercom Fin
Software you control by speaking.
Software that alters the sound of a speaking voice, in real time or after recording.
See it in: ElevenLabs
Making a synthetic voice that sounds like a real person. Only do this with the speaker's permission.
See it in: ElevenLabs, Murf AI
Spoken narration added to a video. AI voices can read a script in many languages.
See it in: ElevenLabs, Murf AI
A tool that removes a mark from an image. Check you have the right to edit the picture first.
See it in: Photoroom
An agent that can browse, click and fill in web pages for you.
See it in: Genspark, Perplexity
A feature that lets a model look things up online before answering.
See it in: Perplexity, ChatGPT
The learned numbers inside a model. Sharing the weights means others can run the model.
See it in: Ollama, Hugging Face
An open speech-to-text model that is good at transcribing many languages.
See it in: Otter.ai, AssemblyAI
A number list for a word, placed so words with similar meaning sit near each other.
See it in: Hugging Face
Linking apps and steps so repetitive jobs run on their own.
A canvas for linking apps, models and steps into an automation.
Assistants built into everyday work apps such as docs, mail and meetings.
See it in: Notion, Microsoft Copilot, Gemini
A model that learns how the world behaves so it can predict what happens next, used in video and robotics.
See it in: Luma AI, Runway Gen-3 Alpha
Making an image of something the model was never specifically trained on, from the prompt alone.
See it in: Midjourney
When a model handles something it was never explicitly trained on, using what it already knows.
Asking a model to do a task with no examples, just the instruction.
Missing a term? Tell us and we will add it in the next batch.