AI Glossary of Terms

A classic pathfinding recipe developed in 1968 that finds the shortest path between two points. It was famously used to help Shakey the Robot navigate rooms.

An AI program designed to act like a helper that can make decisions, plan steps, and take actions on its own to reach a specific goal.

The science of making computer systems smart so they can perform tasks that normally require human brainpower, like solving puzzles or understanding speech.

A step-by-step set of rules or instructions (like a kitchen recipe) that a computer follows to solve a problem or complete a task.

The process of making sure an AI's goals, decisions, and actions match what is helpful, safe, and honest for human beings.

A software bridge that allows two different computer apps to securely talk, share data, and work together.

The math procedure a neural network uses to learn. It looks at its errors, goes backward through its cells, and adjusts their settings to do better next time.

A setting inside an artificial neuron that determines how easy or hard it is to get that neuron to activate (or 'fire') its output.

Unfairness in an AI's results or decisions, usually caused by training it on data that is one-sided or leaves people out.

A computer program built to have text or voice conversations with humans, making it feel like you are chatting with a real person.

AI systems whose code, weights, and training recipe are kept secret by the company that built them, usually only usable through a paid subscription.

The field of AI that helps computers 'see' and recognize objects, faces, or actions in pictures and videos.

The maximum amount of text (measured in tokens) that an AI can look at, read, and remember at one time during a single conversation.

A classic science pioneered in 1948 by Norbert Wiener that studies feedback loops and communication in both animals and machines.

A trick to make training datasets bigger by slightly changing existing examples, like rotating, flipping, or zooming into pictures of cats.

A powerful type of Machine Learning that uses networks of artificial brain cells stacked in many layers to find patterns in complex data.

A type of AI that generates images by starting with random fuzzy static noise and gradually cleaning it up until a detailed drawing appears.

A standard set of metadata terms used worldwide to describe digital museum archives and resources, like creator, publisher, date, and type.

A way of converting words, pictures, or concepts into long lists of numbers (vectors) so that items with similar meanings are grouped close together.

A simple neural network layout where signals only travel forward in one direction, from the input cells directly to the output cells.

Showing an AI model a few completed examples of a task inside your prompt to teach it how to solve a new, similar problem.

Taking an AI that is already generally smart and giving it extra training on a specific subject (like medicine or law) to make it an expert in that area.

Two AI neural networks that compete: one (the Generator) tries to draw fake pictures, and the other (the Discriminator) tries to spot the fakes, making both smarter.

AI models that can create brand new contents, such as writing essays, drawing artwork, writing computer code, or composing music.

A computer chip designed to perform thousands of mathematical calculations at the same time, making it perfect for training complex AI neural networks.

An optimization method that helps an AI lower its mistakes. It takes small steps down a mathematical slope until it finds the point of fewest errors.

A modern, memory-efficient reinforcement learning algorithm that scores a group of answers against each other instead of using a separate critic model.

When an AI makes up a false fact and presents it confidently as if it were true, similar to a computer dream or mistake.

Rule-of-thumb shortcuts or educated guesses that help computers solve problems quickly, even if the answer isn't guaranteed to be 100% perfect.

A massive AI model trained on billions of pages of text so it can understand, write, summarize, and translate human languages.

A branch of AI where computers learn patterns from datasets and experiences on their own, instead of being programmed with rigid, manual rules.

The automatic translation of text or speech from one human language (like French) to another (like Japanese) using computer software.

An efficient AI design that acts like a network of specialists. Instead of using its entire network for every query, it only turns on the best 'expert' cells for that task.

AI systems that can understand and connect different types of data at once, like reading text, looking at images, and listening to audio.

A computer program structure made of interconnected cells (artificial neurons) that processes data like a simplified model of the biological brain.

The study of helping computers understand, read, and write human languages, bridging the gap between computer code and human talk.

An AI model whose inner settings and parameters (weights) are published openly on the internet, allowing anyone to download, inspect, and host it.

When an AI memorizes its training data so perfectly that it cannot handle new, different examples it hasn't seen before.

The settings or 'knobs' inside an AI model that are fine-adjusted during training. More parameters generally allow the model to learn more complex facts.

A simple mathematical model of a single biological neuron invented in 1958. It is the grandfather of modern deep learning nodes.

The initial step of training a model on massive datasets of text or images, giving it general knowledge and language skills before fine-tuning.

The written question, instruction, or text you type into an AI interface to guide it on what to write or create.

The skill of carefully choosing and organizing words in a prompt to get the most accurate, creative, or helpful response from an AI.

An early reinforcement learning method where an AI updates a grid of numbers (Q-values) to learn the best moves in a maze or simple game.

Giving an AI a search tool or a database of documents so it can look up real-time, accurate facts before formulating its final answer.

A neural network that has loops in its connections, letting information persist so it can process sequences of text or time-series data.

Training an AI by giving it positive points/rewards when it takes a good action and negative points when it fails, similar to game scoring.

Improving an AI model's safety and helpfulness by having human testers rate its responses and using those scores to guide its behavior.

A mechanism that lets a model look at all words in a sentence at once and calculate which words are most mathematically related to each other.

A training trick where an AI teaches itself by hiding parts of its data (like blanking out words in a sentence) and trying to guess the hidden parts.

A theoretical future point in time when AI becomes smarter than humans, resulting in rapid, self-improving changes we cannot predict.

AI infrastructure and models trained, hosted, and operated by a nation using its own hardware to protect its cultural values and data.

Training an AI by showing it labeled examples (e.g. showing 1,000 pictures labeled 'dog' and 'not dog') so it learns to classify them on its own.

Data generated by computer models rather than real-world events, used to train AIs when real data is scarce or private.

A small chunk of a word (like a syllable or character combination) that an AI model uses to break down and read human text.

A custom silicon microchip built by Google engineered specifically to run high-speed tensor math for AI training and inference.

A groundbreaking neural network design introduced in 2017 that uses self-attention to process sentences all at once, powering all modern LLMs.

Adjusting settings and parameters in a model to optimize its training, speed, or accuracy on a specific benchmark task.

A test proposed by Alan Turing in 1950 to see if an AI's text chat is so human-like that a judge cannot distinguish it from a real human.

When an AI model is too simple to capture the patterns in its training data, resulting in poor accuracy and bad predictions.

Training an AI on a pile of unlabeled data, letting it discover hidden structures, categories, and patterns entirely on its own.

A list of numbers that represents coordinates in a multi-dimensional space, used by computers to calculate similarities in meanings.

A number that acts like connection strength between virtual brain cells in a neural network, deciding how much signal passes through.

When a pre-trained AI solves a completely new task on the first try without having seen any specific examples of it beforehand.