AI Dictionary

A hypothetical AI that could reason and learn any task as well as a human. It does not exist yet.

Example: Today's models are great at specific tasks, but none of them is AGI yet.

An AI that does not just answer, but can take several steps on its own to finish a task.

Example: An agent that researches flights, compares prices and builds you an itinerary without you asking step by step.

The study of how to use AI fairly, safely and responsibly, without harming people.

Example: Deciding whether AI can be used to make hiring decisions is a question of AI ethics.

An ordered series of steps a computer follows to solve a problem or finish a task.

Example: The recipe that decides which videos an app shows you is an algorithm.

The 'door' that lets one program talk to another program or service automatically.

Example: A weather app uses an API to ask another service for the forecast data.

Computer programs that do tasks only a human used to be able to do, like understanding text, recognizing a photo or making decisions.

Example: When you ask Claude to write you an email, that is AI at work.

Making a repetitive task happen on its own, without a human redoing it by hand every time.

Example: A welcome email that sends itself when someone subscribes is automation.

It's a standard test run on different AI models to compare how well they perform, so everyone gets measured with the same yardstick.

Example: When a new model launches and they say it 'beat the others at math', they're almost always talking about a specific benchmark.

When the AI repeats unfairness or imbalance that already existed in the data it learned from.

Example: A model that only saw resumes from one type of candidate may favor them without meaning to.

Huge amounts of data, so large that special tools are needed to analyze them.

Example: Every click from every user of an app over a year is a Big Data case.

When the AI 'thinks out loud' step by step before giving the final answer, which usually improves accuracy.

Example: Asking it to 'explain your reasoning step by step' triggers a chain of thought.

A program you chat with by typing, that responds like it were a person.

Example: The support chat on a store that replies instantly is usually a chatbot.

An AI whose code and exact workings are not public; you only use it through the company that made it.

Example: Claude is a closed-source model: you use it through Anthropic, you cannot download it.

Someone else's computers, connected to the internet, that you use to store data or run programs without owning the hardware.

Example: When you use Claude, the AI runs in the cloud, not on your phone.

The branch of AI that lets a computer 'see' and interpret images or video.

Example: Your phone's face unlock uses computer vision.

How much text the AI can 'remember' at once within a conversation.

Example: If you paste a very long document, some models start forgetting the beginning: they ran out of context.

It's how much text the AI can 'remember' at once in a conversation: what you already wrote, what it answered, the documents you pasted in. Once it fills up, it starts forgetting what came earlier.

Example: If you paste a 50-page contract into Claude and keep asking it questions for hours, at some point it may start forgetting details from the beginning because it ran out of context window.

An AI that helps you do a task while you stay in charge, suggesting or completing things for you.

Example: An AI that suggests your next line while you code is acting as a copilot.

The right for your personal info not to be used or shared without your permission, even when training or using AI.

Example: Reading the privacy policy before uploading your documents to an AI protects your data privacy.

The collection of examples (text, photos, numbers) used to train an AI.

Example: Millions of web pages are part of the dataset an LLM was trained on.

A fake video, audio or photo made with AI that makes it look like someone said or did something that never happened.

Example: A video where a celebrity seems to announce something they never recorded is a deepfake.

The most common technique behind image generators: it starts with pure noise and gradually 'cleans' it into an image.

Example: Higgsfield, which we use on this site, runs on a diffusion model.

An invisible or visible signal added to AI-generated content to flag that it is not 100% human made.

Example: Some AI images carry a digital watermark that only other software can detect.

A way of turning words or text into numbers so the AI can compare how similar they are in meaning.

Example: Thanks to embeddings, a search engine understands that 'car' and 'automobile' mean nearly the same thing.

Giving the AI a few examples inside the same prompt so it understands better what you want.

Example: You show it 3 examples of how you want your post to sound before asking for the fourth.

Taking an already trained AI and giving it extra, more specific training for one task or style.

Example: Fine-tuning a general model so it only answers legal questions about your local area.

A large, general model trained first, on top of which more specific versions are later built.

Example: Claude starts as a foundation model, then gets tuned for tasks like writing code or chatting.

The general name for all AI that creates original content: text, images, music, video.

Example: Claude, image generators and voice clones are all examples of generative AI.

A type of AI that creates new content (text, image, audio) instead of just classifying or predicting.

Example: A model that writes you a brand new story is a generative model.

A computer chip very good at doing many calculations at once, key for training and running AI.

Example: Training a large model can need thousands of GPUs working together.

The limits put on an AI so it does not say or do harmful or inappropriate things.

Example: Claude refusing to give you dangerous instructions is a guardrail working.

When the AI gives you an answer that sounds confident but is false or made up.

Example: You ask for a historical fact and it gives you a date that does not exist. That is a hallucination.

A setting an engineer decides before training the AI, like how many times it will review the data.

Example: How many training 'rounds' happen is a hyperparameter set ahead of time.

AI that creates brand new images from a text description.

Example: You type 'an astronaut cat, watercolor style' and it gives you back the image.

The moment the AI uses what it already learned to give you an answer, without training anymore.

Example: Every time you ask Claude a question and it answers, that is inference.

An attempt to trick the AI with wording tricks so it skips its safety rules.

Example: Asking the AI to 'act with no rules' so it says something it normally would not is a jailbreak.

The time it takes the AI to start answering you after you ask.

Example: A low-latency app feels instant; a high-latency one feels slow.

The kind of AI behind Claude or ChatGPT: it learned from huge amounts of text, so it can chat, write and answer questions.

Example: Claude is an LLM. That is why it can hold a full conversation with you.

A way of building AI where the program learns from lots of examples, instead of a human writing rules one by one.

Example: A system that learned to spot spam by seeing millions of emails, not because someone told it 'if it says YOU WON it's spam'.

A standard that lets an AI connect directly to your tools and data (like your email, your calendar, or a database), instead of only working with what you type in the chat.

Example: When Claude Code reads files from your project or searches GitHub on its own, it's thanks to an MCP connected behind the scenes.

An AI that understands and works with more than one type of content: text, images, audio or video.

Example: You send a photo of your fridge and it suggests what to cook: that is multimodal.

The math structure, inspired by the brain, most modern AI is built on.

Example: Neural networks are what 'learn' to recognize a cat in a photo.

The branch of AI focused on getting computers to understand and generate human language.

Example: A program detecting whether a review is positive or negative is using NLP.

Tools that let you build something (an app, a site) without writing a single line of code.

Example: Building your online store by dragging blocks, without programming, is no-code.

Software whose code anyone can see, use and modify, for free.

Example: Some AI models are open source and anyone can download and tweak them.

When the AI 'memorized' its training examples instead of learning the pattern, so it fails on new cases.

Example: A model that recognizes the 10 dogs it trained on perfectly, but fails on a brand new dog.

An add-on piece that gives the AI a new ability, like searching the web or reading a PDF.

Example: A calendar plugin lets an AI assistant actually schedule you an appointment.

What you type to the AI to ask for something: a question, an instruction or a task.

Example: 'Write me an Instagram post about my coffee shop' is a prompt.

The craft of writing your instructions to the AI well, so it gives you the best possible result.

Example: Adding 'answer as if you were a teacher' to your prompt is a prompt engineering technique.

It's making an AI model lighter by reducing the precision of its internal numbers, so it takes up less space and runs faster, with barely any loss in quality.

Example: Thanks to quantization, a large model can run on your laptop or even your phone, instead of needing a supercomputer.

When the AI first looks up info in your documents or online, then uses that to answer better.

Example: An assistant that searches your company manual before answering your question is using RAG.

It's an AI that takes a moment to "think" step by step before answering you, instead of blurting out the first thing that comes to mind. That's why it tends to do better with math, logic, or code.

Example: You ask Claude to solve a tricky math problem, and in reasoning mode it first works through its steps internally before giving you the final answer.

The field that combines AI with physical machines so they can move and act in the real world.

Example: A robotic arm that learns to sort boxes uses robotics plus AI.

A safe, separate space where AI or code is tested without risk of affecting anything real.

Example: Before launching an AI agent with access to your email, it gets tested in a sandbox.

AI that turns what you say into written text.

Example: When you dictate a voice message on your phone and it comes out as text, that is speech recognition.

Same as speech recognition: turns spoken audio into written text.

Example: Automatically transcribing a work call is speech-to-text.

A hypothetical AI that would surpass human intelligence at nearly everything. It is a debated idea, not current reality.

Example: In sci-fi movies, the 'AI that becomes smarter than everyone' is a superintelligence.

Fake but realistic information that an AI creates to train another AI, instead of using real people's data. It's used when real data is scarce or when using it would raise privacy issues.

Example: A hospital that wants to train an AI to detect illnesses might use synthetic medical records (made up by another AI but statistically similar to real ones) so it doesn't expose actual patient data.

Invisible instructions given to the AI before your message, to set how it should behave.

Example: A company can set a system prompt telling its chatbot 'always answer in a formal tone'.

A setting that controls how creative or predictable the AI's answer is. Low = safer, high = riskier.

Example: High temperature gives you wilder business name ideas; low gives you the obvious one.

AI that turns written text into a spoken voice.

Example: The narrator that reads an article out loud to you uses text-to-speech.

A small chunk of a word the AI uses to read and write text. One word can be one or several tokens.

Example: 'Understanding' might split into tokens like 'Under' + 'standing'.

The process where the AI learns, by showing it tons of examples until it picks up the pattern.

Example: Training a model on photos of dogs and cats until it learns to tell them apart.

Using what an AI already learned on one task to help it learn a different but similar task faster.

Example: A model that already knows how to recognize animals learns specific dog breeds faster.

The architecture (the 'internal design') used by most modern LLMs, including Claude.

Example: The 'T' in GPT stands for Transformer, the tech behind these models.

It's coding by describing what you want to an AI in plain language, instead of writing the code yourself line by line.

Example: You tell Claude 'build me a page with a contact form' and in minutes you have a working site, without touching a line of code.

Asking the AI to do a task without giving it any prior example, just the instruction.

Example: 'Classify this email as spam or not' without showing examples first is zero-shot.