AI terms, explained.

A computer chip built to run AI math faster than a standard processor.

The idea of an AI that could do any thinking task a person can. Nothing on the market today works this way.

An AI that does multi-step tasks on its own instead of only answering, looping you in along the way.

Example: An agent could draft a reply to a client and send it once you give the okay.

Software that handles everyday requests, like email or appointments. People often use it to mean "AI agent," though an agent does more. We say "agent" for the thing you set up, and "assistant" only for the job it does.

Harmful or offensive language an AI might produce, usually picked up from its training data.

Research to make AI systems pursue goals that match what people want and value.

Treating AI as if it has human feelings or awareness because it talks like a person.

Example: A short AI reply can feel "cold," though the AI has no mood.

A company that develops Claude, an AI model built with a strong focus on safety.

A defined way for two programs to share information, like an AI assistant reading your calendar with permission.

A secret code that lets one program connect to another, such as an AI assistant to your inbox.

Example: Anyone holding the key could access the connected account, so guard it like a password.

A theoretical future AI that would outdo people in every field. No current technology works this way.

A method that helps an AI model focus on the parts of its input that matter most.

Example: In a long email chain, attention helps the AI find the request among the greetings.

A way of working where AI sorts through large amounts of information and people make the final calls.

A task that runs on its own instead of being done by hand each time.

Example: A rule sorts email by sender; an AI assistant could read each one and decide which need your attention.

A training step where a model traces an error back through its internal settings and adjusts them.

Unfair or inaccurate AI output caused by skewed or unrepresentative training data.

When an AI works through its reasoning step by step before giving a final answer.

A widely used AI chatbot from OpenAI, built mainly for conversation and answering questions.

Software built mainly to answer questions or chat, usually inside one app or website.

Example: A website chatbot might tell you store hours but can't book you an appointment the way an AI assistant could.

An AI model that links images with text, so it can match pictures to the words that describe them.

Using computing power and storage over the internet instead of a computer you own.

Example: An AI assistant "in the cloud" runs on remote servers you can reach from anywhere, with no equipment to keep up.

The processing power AI tasks use.

How much an AI can hold in mind at once, including earlier messages and attached files.

Example: In a very long chat, an AI may lose track of early details, like forgetting how a long meeting began.

AI you talk to in everyday language instead of commands or search terms.

Example: Instead of typing "weather NYC tomorrow," you ask "is it going to rain tomorrow?"

A type of AI model that is good at spotting patterns in images, like faces or road signs.

A system that keeps customer details, conversations, and history in one place.

A screen that sums up activity or status at a glance, often with charts or lists.

Making more training examples by slightly changing existing ones, such as flipping or rotating images.

A branch of machine learning that uses many-layered neural networks to find complex patterns in data.

When software or an AI system goes live for real use, after building and testing.

A tool that packs software with everything it needs into a "container" that runs the same on any computer.

A method AI image generators use to refine random visual "noise," step by step, into a finished image.

In a GAN (below), the part that tries to tell real data from AI-made fakes.

A research finding: as a model grows more complex, its accuracy can dip and then improve again.

A list of numbers that stands for a word, image, or other data, so similar meanings get similar numbers.

Example: A search for "checking in with customers" can find a document about "client follow-up."

Abilities that appear in an AI model as it grows, without anyone building them in.

Example: Large language models surprised researchers by writing working code without being trained specifically for it.

Training a model on raw inputs and wanted outputs, so it works out the steps in between itself.

A setting, such as an API key, kept outside a program's code so it can be changed without editing the code.

A framework for judging how responsibly a company uses AI, covering fairness, privacy, and accountability.

An older kind of AI run by hand-written "if-then" rules. Most modern AI learns patterns from data instead.

AI built to show its reasoning, so people can check why it reached a conclusion.

Extra, focused training that makes an already-trained model better at one task.

A model passing input through its layers to produce a result. It happens with every reply you get.

A large, broadly trained AI model that more specialized tools are built on.

What lets an AI use outside tools and data, like your calendar, beyond what it learned in training.

Two AI models that compete: one makes content, the other tries to spot fakes, and both improve.

AI that makes new text, images, audio, or video instead of only analyzing what exists.

Example: When an AI assistant drafts an email or summarizes a document, that is generative AI.

A tool developers use to track changes to code, undo mistakes, and work together.

A website where developers store and share code managed with Git, including many open-source AI projects.

A line of OpenAI language models that powers ChatGPT. People also say "GPT" loosely for any AI chatbot.

A chip first made for video game graphics that also excels at the math AI runs on.

How a model improves in training: many small corrections, each reducing its errors a little.

Example: Like walking downhill in thick fog, feeling for the slope and taking small steps until you reach the bottom.

Tying an AI's answers to real, checkable information, like your own files.

Example: A grounded AI assistant might check your actual calendar before suggesting meeting times.

When an AI confidently states something false or made up, so it sounds right but is wrong.

Example: An AI might invent a real-sounding phone number, so a person should check anything important.

The internal processing steps of an AI model, between input and output, that you never see.

A setup where a person checks and approves the AI's work before it is final.

Example: An AI assistant might draft a reply to a customer and hold it until you give the okay.

Choosing the settings that control how a model is trained, before training starts.

Example: Like setting the oven's temperature and timer before you bake.

An app developers use to write, test, and fix code, with their tools in one place.

Example: VS Code (below) is a popular IDE.

The stage where an already-trained AI model is used to produce a response.

Example: When you ask an AI assistant something and it answers, that is inference.

Extra training that teaches a model to follow directions like "summarize this" or "reply politely."

A link that lets an AI assistant, or other software, work directly with another account, app, or service.

The kind of AI behind most assistants and chatbots, trained on huge amounts of text to understand and write language.

How long an AI takes to respond after you ask.

An internal map a model builds, where ideas with similar meanings sit close together.

Example: This is why an AI can treat "car" and "automobile" as related without being told.

A score a model uses during training to measure how far its output is from the right answer.

The branch of AI where systems learn from data instead of being coded for every case.

Behind-the-scenes instructions given to an AI before a conversation starts, setting its role, tone, and limits.

Example: A system prompt might tell an AI assistant to always confirm before sending an email.

A model made of smaller specialized parts, with a router that picks which ones handle each input.

A program trained to find patterns in data, a mathematical stand-in for some real-world process.

Example: A weather forecaster, a translation tool, and the engine behind an AI assistant are all models.

Several AI agents working on different parts of one larger task, coordinated as a team.

An AI that can take in more than one kind of input, like text and images.

The branch of AI that understands and produces human language, typed or spoken.

A method for building a 3D scene from a set of ordinary 2D photos.

The basic structure behind most modern AI: layers of connected "nodes" loosely modeled on brain cells.

A tool developers use to download and manage ready-made code "packages" that others wrote.

The goal a model works to maximize or minimize during training. It is closely related to a loss function (above).

An open-source framework for building AI agents that take real action on a computer, like handling files or sending messages.

Software whose code anyone can view, use, or change. Proprietary software keeps its code private.

When a model learns its training data too exactly, so it does well on familiar examples but poorly on new ones.

Example: Like a student who memorizes practice answers, then struggles when the exam rewords them.

The internal values a model adjusts as it learns, like countless small dials.

Example: When a model is described as "70 billion parameters," these values are what is being counted.

A tool that keeps track of your contacts, past conversations, and pending follow-ups with each person.

Example: A personal CRM might remind you to follow up with a client next week without you setting a reminder.

An AI assistant tailored to your inbox, schedule, and contacts instead of a one-size-fits-all tool.

The first, broad stage of training a model on lots of general data, before any fine-tuning.

Example: Like general schooling before training for a career.

The request or instruction you give an AI, like "summarize this document" or "find three flights to Chicago."

Built-in protections that stop an AI from being tricked into producing harmful, false, or off-brand content.

Example: A rule might tell the AI to say "I'm not sure" instead of guessing.

Wording requests to an AI so they get more accurate or useful results.

A method where an AI checks your own documents before answering, instead of relying only on its training.

Example: Rather than guess at a return policy, an AI using RAG could pull up the policy itself and answer from it.

AI models built to work through hard problems in several steps before answering.

Deliberately trying to make an AI fail or misbehave, to find weak spots early.

Example: Like hiring a security firm to try a break-in before a real burglar does.

A training technique that keeps a model from overfitting, steering it toward broader patterns.

Training where an AI learns by trial and error, rewarded for good results and corrected for poor ones.

Example: Much like training a pet with treats.

A storage space for a project's code and files, often managed with Git and shared on GitHub.

Training where people rate an AI's answers and the AI adjusts toward the ones they prefer.

The field that works to prevent harm, misuse, or accidents from AI and keep it in line with human values.

AI that reads text and picks up its emotional tone, such as positive, negative, or frustrated.

A computer that stays on and connected, running software so it is ready whenever needed.

Example: An AI assistant that works around the clock typically runs on a server instead of a personal laptop.

A hypothetical point where AI progress speeds up beyond prediction and reshapes society. It remains speculative.

An add-on that gives an AI assistant a new ability, like checking the weather or reading a file type.

Training a model on examples that come with the right answers, like a textbook with an answer key.

An older approach to AI built on explicit symbols and logic rules instead of learning from data.

A messaging app, like WhatsApp or texting, that some AI assistants use to talk with you.

Example: If your assistant works this way, Telegram is just the "front door"; the AI isn't part of Telegram itself.

A setting that controls how predictable or creative an AI's output is.

Example: Lower gives steadier answers for facts; higher gives more varied ideas for brainstorming.

A free, open-source toolkit from Google for building and training AI models.

A text window for controlling a computer by typing commands instead of clicking.

The basic unit of text an AI processes, roughly three-quarters of a word on average.

Example: Tokens matter mainly for context window limits and AI service pricing.

A chip Google designed for AI work, similar in role to a GPU.

The examples an AI model learns from.

When what a model learned on one task gives it a head start on a related one.

Example: Like how learning one instrument makes a second one easier.

A model design from 2017 that weighs how the words in a sentence relate to each other. GPT and most current AI language tools are built on it.

Being open about how an AI works, what data it uses, and when you are talking to AI.

When a model is too simple to pick up real patterns, so it does poorly even on its training data.

Training a model on data with no answers attached, so it finds patterns and groups on its own.

A separate set of examples used to check how well a model is learning, before final testing.

Example: Like a practice exam before the real test.

A free, popular code editor from Microsoft.

An automatic message one program sends another the moment something happens, so nothing has to keep checking.

Example: A webhook could tell your AI assistant the instant a new form comes in.

A series of steps that together make up a larger task, often across several tools.

Example: An AI assistant's workflow for a new customer question might be: read it, check past cases, draft a reply, flag it for your approval.

A policy where what you share with an AI is used to produce a response and then discarded, with nothing stored afterward.

Example: It matters most when an AI handles sensitive details like your email or finances.

When a model does a task it was never trained on or shown examples of, using its general knowledge instead.

Example: Like playing a new board game well on the first try because you know games in general.