AI Glossary

Accuracy is the most basic measure of a classification model's performance: the share of all its predictions that were correct — though on imbalanced data, where one outcome is rare, a high accuracy can be deeply misleading.

An activation function is a small mathematical operation applied inside each unit of a neural network that decides how strongly the unit responds, and crucially introduces non-linearity so the network can learn complex, curved patterns rather than only straight-line relationships.

Active learning is a machine learning strategy in which the model itself chooses which unlabeled examples would be most useful to learn from, and requests human labels for just those — reducing how much data must be labeled to reach a given level of accuracy.

An adversarial attack is a deliberate attempt to fool an AI model by feeding it carefully crafted input — often altered in ways too subtle for a human to notice — that causes the model to make a confident but wrong decision, such as misreading an image or ignoring a safety rule.

Agentic AI refers to AI systems that can pursue goals with a degree of independence — planning steps, making decisions, and taking actions through tools over time — rather than simply responding to one request at a time.

An AI agent is a system that doesn't just answer a question but pursues a goal on your behalf — breaking it into steps, using tools like search, software, or other programs, and adjusting as it goes until the task is done.

AI alignment is the effort to ensure that an AI system's goals and behavior genuinely match human intentions and values — that it does what we actually mean and want, not just what we literally said or what it was narrowly trained to optimize.

An AI assistant is a software helper, usually powered by a large language model, that you interact with in everyday language to get things done — answering questions, drafting text, summarizing information, and helping with a wide range of tasks across work and daily life.

An AI coding assistant is a software tool, usually built into a code editor, that uses a large language model to help programmers write software — suggesting and completing code as they type, generating whole functions from a plain-language request, explaining unfamiliar code, and pointing out likely bugs.

An AI copilot is an AI assistant built directly into a piece of software — an email client, a code editor, a spreadsheet — that suggests, drafts, rewrites, or completes work alongside you as you use the tool, with you accepting or rejecting each suggestion.

AI detection is the use of software to estimate whether a given piece of content — most often text, but also images, audio, or video — was generated by AI rather than created by a human, typically by analyzing statistical patterns rather than reading any embedded marker.

AI ethics is the study of the moral questions raised by artificial intelligence — what is fair, harmful, or acceptable in how AI systems are built and used — and the principles meant to guide AI toward benefiting people rather than harming them.

AI governance is the set of policies, rules, standards, and oversight processes — within organizations and across society — that guide how AI is developed, deployed, and used responsibly, and that hold people accountable for its effects.

AI memory is the ability of an AI system to store and recall information beyond a single exchange — remembering facts, preferences, and past interactions so it can carry context forward across a long task or many separate conversations.

AI planning is the process by which an AI system works out a sequence of actions to get from where things are now to a desired goal — choosing the steps, their order, and how to adapt them when circumstances change.

AI safety is the field concerned with making AI systems behave reliably and avoid causing harm — designing, testing, and governing them so their behavior stays beneficial and under control, both today and as they grow more capable.

AI search is a way of searching where an AI system reads relevant sources and writes a direct, conversational answer to your question, instead of only returning a ranked list of links for you to open and read yourself.

An algorithm is a precise set of step-by-step instructions for solving a problem or completing a task — the fixed procedure a computer follows to turn an input into a result.

Anomaly detection is a machine learning task that learns what "normal" looks like in a set of data and then flags the rare points that depart sharply from it, surfacing unusual events without anyone having to define every unusual case in advance.

Area under the curve (AUC) is a single number that scores how well a classification model separates two categories across every possible decision threshold — most often the area under the ROC curve, where 1.0 is perfect and 0.5 is no better than guessing.

Artificial general intelligence (AGI) is a hypothetical kind of AI that could understand, learn, and handle virtually any intellectual task a human can, rather than being limited to the one narrow job it was built for.

Artificial Intelligence (AI) is the field of computer science focused on building systems that can perform tasks that would normally require human intelligence — things like understanding language, recognizing images, making decisions, and solving problems.

Artificial narrow intelligence (ANI), also called weak or narrow AI, is AI built to perform one specific task or a tightly limited set of tasks — describing essentially every AI system in real use today, as opposed to the hypothetical general intelligence that could handle any task.

Artificial superintelligence (ASI) is a hypothetical level of AI whose intellectual ability would far exceed the best human minds across essentially every domain — not merely matching human intelligence, as artificial general intelligence would, but decisively surpassing it.

An attention mechanism is a technique that lets a neural network weigh which parts of its input matter most for each piece of output, focusing on the relevant pieces rather than treating everything equally.

An autonomous agent is an AI system that pursues a goal over many steps with little or no human involvement at each step — planning its own actions, using tools, judging the results, and deciding what to do next on its own.

An autonomous vehicle is a car, truck, or other vehicle that can sense its surroundings and drive itself with little or no human input, using cameras, other sensors, and machine learning software to steer, accelerate, brake, and navigate.

Backpropagation is the core algorithm used to train neural networks, which works out how much each internal connection contributed to a wrong answer and adjusts them all slightly to make the network's next prediction a little more accurate.

Batch size is the number of training examples a machine learning model looks at together before it updates itself once during training — a setting chosen in advance that affects how fast, how smoothly, and how much memory training takes.

A benchmark is a standardized test — a fixed dataset and scoring method — used to measure how well an AI model performs at a task and to compare different models fairly on the same yardstick.

BERT (Bidirectional Encoder Representations from Transformers) is an influential language model that reads text in both directions at once to understand the meaning of words in context, designed for comprehension tasks like search and classification rather than generating text.

Bias in AI is when a system produces systematically unfair or skewed results — favoring or disadvantaging certain groups or outcomes — usually because it learned from data that reflected human prejudices or was unrepresentative of the people it affects.

Chain-of-thought prompting is a technique that asks a language model to work through its reasoning step by step before giving a final answer, which improves accuracy on problems that need several stages of thinking.

A chatbot is a software program that holds a conversation with people in everyday language, through text or voice, answering questions and helping with tasks as if you were messaging a person.

ChatGPT is an AI assistant developed by OpenAI that uses a large language model to hold conversations, answer questions, write content, analyze information, and help with a wide range of tasks through a simple chat interface.

Classification is a machine learning task in which a model sorts each input into one of a fixed set of predefined categories — deciding, for example, whether an email is spam or not, or which of several diseases an X-ray most likely shows.

Claude is a family of large language models and AI assistants built by the company Anthropic, used through a chat app and through developer tools to hold conversations, write and analyze text, work with code, and carry out multi-step tasks.

Clustering is an unsupervised machine learning task that automatically groups a set of data points into clusters, so that items in the same group are more similar to one another than to items in other groups — without being told in advance what the groups should be.

Code generation is the use of AI to produce working software code from a description of what the code should do, usually written in plain language, turning an instruction like "sort this list" into the actual programming statements that carry it out.

Computer vision is a field of artificial intelligence that enables computers to extract meaningful information from digital images and video, and to act on what they find.

A confusion matrix is a table that summarizes how a classification model performed by laying out, for each category, how many predictions were correct and exactly how the wrong ones were wrong — separating, for example, false alarms from missed cases.

Constitutional AI is a training approach, developed by the company Anthropic, that aligns an AI model's behavior to a written set of principles — a "constitution" — by having the model critique and revise its own responses against those rules, reducing the amount of human labeling of harmful outputs needed.

Context engineering is the practice of deciding what information an AI model is given before it responds — instructions, relevant documents, memory, and tool results — and how that information is selected, arranged, and fit within the model's limited context window.

The context window is the maximum amount of text an AI language model can take in and consider at one time — everything you've said, everything it has replied, and any documents you've added — measured in tokens, beyond which the oldest material drops out of view.

Conversational AI is technology that lets people interact with software through natural back-and-forth dialogue — by text or voice — understanding what's said, keeping track of the conversation, and responding in a human-like way.

A convolutional neural network (CNN) is a type of neural network designed to work with images, which learns to detect visual features like edges, textures, and shapes by scanning small patches across a picture and building them up into recognizable objects.

Cosine similarity is a measure of how alike two vectors are based on the angle between them rather than their length, widely used in AI to judge how similar in meaning two pieces of text or data are.

Cross-validation is a technique for testing how well a machine learning model will perform on new data by repeatedly splitting the available data into training and testing portions in different ways, then averaging the results for a more reliable estimate.

DALL-E is a text-to-image AI system developed by OpenAI that generates original pictures from written descriptions, and was one of the systems that first brought AI image generation to wide public attention.

Data augmentation is the practice of expanding a training set by creating modified copies of the data you already have — flipping, cropping, or rewording existing examples — so a model sees more variety and learns to handle the real world without anyone having to collect more data.

Data labeling is the process of attaching correct answers or tags to raw data — marking what each example is — so that the labeled examples can be used to train and evaluate supervised machine learning models.

A dataset is an organized collection of data — such as a table of records, a folder of images, or a body of text — gathered together so it can be analyzed or used to train and test an AI model.

A decision tree is a machine learning model that makes a prediction by asking a sequence of simple yes/no questions about the data, branching at each answer until it reaches a final decision — much like a flowchart learned automatically from examples.

A decoder is the part of a neural network that generates an output one piece at a time from an internal representation of meaning, producing each new element in light of what it has already generated — the component behind text-generating models.

Deep learning is a branch of machine learning that uses neural networks with many layers to learn complex patterns from large amounts of data, powering most of today's advanced AI systems.

A deepfake is a piece of synthetic media — typically a video, image, or audio clip — in which AI has been used to convincingly replace or fabricate a real person's face, voice, or actions, making them appear to say or do something they never actually said or did.

A diffusion model is a type of AI that creates images and other content by starting from random visual noise and gradually cleaning it up, step by step, into a finished result — having learned how to do that by studying how real images dissolve into noise.

A digital twin is a virtual model of a real-world object, system, or process that is kept in sync with its physical counterpart using live data, so it can be monitored, analyzed, and used to simulate and test changes without affecting the real thing.

Dimensionality reduction is a set of machine learning techniques that reduce the number of input variables (dimensions) describing each data point, compressing the data into fewer variables while preserving as much of its meaningful structure and variation as possible.

Direct preference optimization (DPO) is a training method that teaches a model from human preference comparisons directly, without building a separate reward model or running reinforcement learning — a simpler alternative to reinforcement learning from human feedback.

Dropout is a technique for training neural networks that randomly switches off a fraction of the network's units on each training step, which discourages the network from over-relying on any single part and helps it generalize better to new data.

Edge AI is the practice of running artificial intelligence directly on or near the device where data is produced — a phone, camera, car, or sensor — instead of sending that data to a distant cloud server for processing.

Embeddings are a way of representing words, images, or other data as lists of numbers, arranged so that items with similar meaning end up with similar numbers — letting a computer measure how related two things are even when they share no words in common.

An encoder is the part of a neural network that reads an input and compresses it into a rich internal representation of its meaning, which the rest of the system can then use for understanding tasks or pass to a decoder to produce an output.

Ensemble learning is a machine learning approach that combines the predictions of several models instead of relying on one, pooling their answers so the group is more accurate and more reliable than any individual model on its own.

An epoch is one complete pass of a machine learning model through the entire training dataset during training; models typically train for many epochs, seeing the same data repeatedly so they can keep refining what they've learned.

Explainable AI (XAI) is the field and set of techniques concerned with making an AI system's decisions understandable to people — producing human-readable reasons for why a model reached a particular output, rather than leaving it an opaque "black box."

The F1 score is a single number that combines a classification model's precision and recall into one balanced figure, designed so the model can only score well when both — its rate of false alarms and its rate of misses — are kept low together.

Fairness in AI is the principle and practice of building AI systems whose decisions do not unjustly disadvantage people based on characteristics such as race, gender, age, or disability, along with the techniques used to measure and reduce such unfair treatment.

Feature engineering is the process of selecting, transforming, and creating the input variables — the "features" — that a machine learning model learns from, so the data is presented in a form that makes the underlying patterns easier for the model to detect.

Feature extraction is the process of turning raw data into a smaller set of informative measurements — the "features" — that a machine learning model can actually learn from, capturing what matters in the data while discarding noise and redundancy.

Federated learning is a way of training a shared machine learning model across many devices or organizations without collecting their data in one place — each keeps its data locally and sends back only what it learned, so the raw data never leaves home.

Few-shot learning is when an AI is given just a handful of examples of a task and uses them to perform it, rather than needing thousands of training examples to learn what to do.

Fine-tuning is the process of taking an AI model that has already been trained on broad, general data and training it a bit further on a smaller, focused set of examples, so it specializes in a particular task, domain, or style.

A foundation model is a large, general-purpose AI model trained on a huge, broad sweep of data, built to serve as a reusable base that can be adapted to many different tasks rather than being made for just one.

A frontier model is one of the most advanced AI models available at a given time — a large, general-purpose system at the leading edge of capability, typically built by a major AI lab at great expense and pushing the boundary of what AI can do.

Function calling is the structured mechanism by which a language model requests a specific external tool or function — naming it and supplying the inputs in a machine-readable format — so that surrounding software can run it and return the result for the model to use.

Gemini is a family of large language models and AI assistants built by Google, used through a chat app and through developer tools to hold conversations, answer questions, write and analyze text and code, and work across text, images, audio, and other formats.

Generalization is a machine learning model's ability to perform well on new, unseen data rather than only on the examples it was trained on — the true measure of whether it has learned a useful pattern instead of memorizing.

A generative adversarial network (GAN) is a type of AI made of two neural networks trained in competition — one generating fake data and the other trying to spot the fakes — which together push the generator to produce increasingly realistic results.

Generative AI is a category of artificial intelligence that can create new content — including text, images, audio, video, and code — by learning patterns from existing data.

Generative engine optimization (GEO) is the practice of writing and structuring online content so that AI answer engines — chatbots and AI-powered search that generate a written answer — are more likely to surface, quote, and cite it when responding to a user's question.

Generative Pre-trained Transformer (GPT) is a family of large language models, originally developed by OpenAI, whose name describes how they work: they generate text, are pre-trained on huge amounts of data, and are built on the transformer architecture.

Gradient descent is the optimization method most AI models use to learn — repeatedly nudging their internal settings in the direction that reduces error, taking small steps downhill until the model's mistakes are as low as they can get.

A graphics processing unit (GPU) is a specialized computer chip designed to perform many calculations at once, originally built to render graphics but now the workhorse hardware for training and running AI models because of its ability to do massive numbers of operations in parallel.

Grounding is the practice of tying an AI's responses to verifiable, real-world information — such as trusted documents, data, or live sources — so its answers are based on actual facts it can point to rather than on plausible-sounding guesses from memory.

Guardrails are the rules, filters, and checks placed around an AI system to keep its behavior within safe, acceptable bounds — blocking harmful or off-limits outputs and keeping the system on task.

A hallucination in AI is when a system, especially a chatbot, produces information that sounds confident and plausible but is actually false, made up, or not supported by any real source.

Human in the loop is an approach in which a person stays actively involved in an AI system's operation — reviewing, approving, correcting, or overriding its decisions — rather than letting the AI act entirely on its own.

A hyperparameter is a setting chosen *before* a machine learning model starts training that controls how the training process runs — such as how fast the model adjusts or how long it trains — as opposed to the internal values the model learns by itself from the data.

Image segmentation is a computer vision task that labels every pixel in an image according to what it belongs to, producing a precise outline of each object or region rather than a rough bounding box.

Inference is the stage at which a trained machine learning model is actually put to use — taking in new input and producing an output, such as a prediction or a generated reply — as distinct from the earlier training stage where the model learned.

Instruction tuning is a training stage in which a model is fine-tuned on many examples of instructions paired with good responses, teaching it to follow what a user actually asks rather than merely continuing text.

Interpretability is the degree to which a human can understand the internal logic of an AI model — how it actually arrives at its outputs — together with the techniques used to make complex, opaque models more readable to people.

A jailbreak is a deliberately crafted input that tricks an AI model into ignoring its own safety rules and producing responses its makers designed it to refuse, by disguising or reframing the request so the model's guardrails don't recognize it as off-limits.

A knowledge base is an organized, machine-readable store of information — facts, documents, rules, or structured relationships — that a system can draw on to answer questions or make decisions, rather than relying only on what a model learned during training.

Knowledge distillation is a technique for training a small, fast AI model to copy the behavior of a large, capable one, so the smaller model captures much of the bigger model's skill while being cheaper and quicker to run.

A knowledge graph is a way of storing information as a network of entities — people, places, things, concepts — connected by labeled relationships, so that the facts and the links between them can be navigated and reasoned over by both software and people.

A large language model is a type of AI system trained on massive amounts of text that can understand, summarize, translate, and generate human language with remarkable fluency.

Latent space is the internal, multi-dimensional space in which an AI model represents data as points, arranged so that the meaningful features and relationships in the data correspond to positions and directions within that space.

The learning rate is a setting that controls how big an adjustment a machine learning model makes each time it corrects itself during training — a small but pivotal hyperparameter that decides whether the model learns smoothly, too slowly, or never settles at all.

Llama is a family of large language models built by Meta, distinguished from most rivals by being released as open-weight models that developers, researchers, and companies can download and run on their own systems, rather than only using through one company's app.

LLMOps (large language model operations) is the set of practices and tools for deploying, running, monitoring, and improving large language models in real-world applications — a specialized branch of MLOps focused on the particular challenges of these very large, generative AI models.

Long short-term memory (LSTM) is a type of recurrent neural network designed to remember information over long sequences, using built-in "gates" that decide what to keep, what to update, and what to discard — overcoming the tendency of plain recurrent networks to forget earlier inputs.

A loss function is a formula that measures how far an AI model's prediction is from the correct answer, producing a single number that training works to make as small as possible, and defining what counts as a good or bad result.

Low-rank adaptation (LoRA) is an efficient fine-tuning technique that adapts a large model by training a small set of added values rather than adjusting all of the model's original parameters, drastically cutting the cost of customizing a model.

Machine learning is a branch of artificial intelligence in which systems learn from data to make predictions and decisions, improving their performance over time without being explicitly programmed for every step.

Machine translation is the use of AI to automatically convert text or speech from one human language into another, aiming to preserve the original meaning rather than translating word by word.

Midjourney is a popular generative AI service that creates images from written descriptions, known for producing especially polished, stylized, and artistic results from a short text prompt.

Mixture of experts (MoE) is a neural network design that splits a model into many specialized sub-networks and, for each input, activates only the few most relevant ones — giving the model a huge total size while keeping the cost of each prediction low.

MLOps (machine learning operations) is the set of practices and tools for reliably deploying, running, monitoring, and maintaining machine learning models in real-world use, bringing engineering discipline to the whole lifecycle of a model after it has been built.

A model card is a short, standardized document that accompanies an AI model and describes it in plain terms — what it does, what data it was trained on, how well it performs, who it's intended for, and its known limitations and risks — so that people can use it responsibly.

The Model Context Protocol (MCP) is an open standard that defines a common way for AI assistants to connect to external tools, data sources, and services, so any compliant app can plug into any compliant tool without custom, one-off integrations.

Model deployment is the step of taking a trained machine learning model and putting it into live use, where it can take in real input from users or other systems and return outputs — moving the model out of the lab and into an actual product or service.

Model drift is the gradual decline in a machine learning model's accuracy over time after deployment, caused by the real world changing so that the patterns the model learned during training no longer match the data it now sees.

Model monitoring is the practice of continuously tracking a machine learning model after it has been deployed — watching its accuracy, its inputs, and its behavior in real use — to catch problems like declining performance early, before they cause harm.

A multi-agent system is a setup in which several AI agents, often with different roles or specialties, work alongside one another — coordinating, dividing tasks, or checking each other's work — to tackle a problem that's hard for a single agent to solve alone.

Multimodal AI is artificial intelligence that can work with more than one kind of information at once — such as text, images, audio, and video — understanding and connecting them within a single system rather than handling just one.

Named entity recognition is an AI technique that scans text and automatically identifies and labels the named things in it — such as people, organizations, places, dates, and amounts — pulling structured information out of unstructured writing.

Natural language generation is the part of AI that produces fluent, human-readable text — turning data, structured information, or an internal representation of meaning into sentences a person can read naturally.

Natural language processing (NLP) is a field of AI that enables computers to read, understand, and generate human language in all its complexity and variation.

Natural language understanding is the part of AI focused on working out the meaning and intent behind human language — extracting what a person actually wants and the key details they mention — rather than just processing the words on the surface.

A neural network is a computing system loosely modeled on the human brain, made up of interconnected nodes that process data and learn patterns from examples.

A neural processing unit (NPU) is a specialized chip, often built into phones, laptops, and other devices, designed to run artificial intelligence tasks efficiently and with low power use — a category of hardware also known more generally as an AI accelerator.

Neuromorphic computing is an approach to designing computer chips whose structure and operation are modeled on the brain — using interconnected artificial neurons that communicate in spikes and process information in a highly parallel, event-driven way — aiming for far greater energy efficiency on certain AI tasks than conventional chips.

Object detection is a computer vision task in which an AI finds the individual objects in an image or video, drawing a box around each one and labeling what it is — identifying both what objects are present and where they are located.

On-device AI is artificial intelligence that runs entirely on a single user device — typically a phone, laptop, or wearable — performing its processing locally on that device's own hardware rather than relying on a remote cloud server.

One-shot learning is a machine learning approach in which a model learns to recognize a new category from just a single example, by drawing on broader knowledge it already has rather than needing many labeled samples of that category.

An ontology, in computing and AI, is a formal, machine-readable description of the concepts in a subject area and the relationships between them, defining the categories, properties, and rules that let software interpret information consistently rather than treating it as plain text.

Open weights describes an AI model whose trained parameters — the learned numbers that make it work — are released publicly for anyone to download and run, while the training code, data, and full method behind it are typically not shared.

Open-source AI refers to artificial intelligence systems released under a license that lets anyone freely use, study, modify, and share them — ideally including not just the trained model but the code, and where possible the training details, behind it.

Optical character recognition is technology that converts images of text — such as scanned documents, photos, or printed pages — into machine-readable text that a computer can search, edit, and store.

Optimization, in machine learning, is the process of systematically adjusting a model's internal settings to make a chosen measure of error as small as possible (or a measure of performance as large as possible), and it is the mechanism by which a model actually learns from data.

In AI, orchestration is the coordination of the many moving parts of an AI system — models, tools, data sources, and individual steps — so they run in the right order and pass results between each other to complete a larger task.

Overfitting is when a machine learning model learns its training data too closely — memorizing the noise and quirks of those specific examples rather than the general pattern — so it performs well on data it has seen but poorly on new, unseen data.

Parameters are the internal numerical values a model adjusts during training — the learned settings that store what the model has learned and determine the output it produces from a given input.

Post-training is the collection of training stages applied to a model after its large initial pretraining phase, shaping a raw, broadly capable model into one that follows instructions, behaves helpfully, and avoids harm.

Precision and recall are two metrics for judging a classification model: precision asks how many of the items it flagged were genuinely correct, while recall asks how many of the genuinely correct items it managed to find — and improving one often costs the other.

Pretraining is the first and largest training phase of a model, in which it learns broad, general patterns from a huge amount of data, building the foundational competence that later, more focused training stages refine into a finished, usable system.

Principal component analysis (PCA) is a technique that reduces data with many variables down to a few new combined variables — the principal components — chosen to preserve as much of the data's variation as possible, making it simpler to analyze, visualize, and feed to a model.

A process reward model is a model trained to score the individual steps in another AI's reasoning — judging whether each step is sound — rather than only judging the final answer, so a system can be guided toward solutions that are right for the right reasons.

A prompt is the input you give an AI system — usually the question, instruction, or piece of text you type or speak — that tells it what you want and that it responds to.

Prompt engineering is the practice of crafting and refining the instructions you give to an AI system to get more accurate, useful, and consistent results.

Prompt injection is an attack in which hidden instructions are planted inside the outside content an AI reads — a web page, document, or email — so the AI treats them as commands and acts on the attacker's behalf instead of the user's.

Pruning is a technique for shrinking an AI model by removing the parts that contribute little to its results — such as weak connections or unused units — so the model becomes smaller and faster while keeping most of its accuracy.

Quantization is a technique for shrinking an AI model by storing its internal numbers at lower precision — for example, using whole numbers instead of long decimals — so the model takes up less memory and runs faster, with only a small loss of accuracy.

Quantum machine learning is a research field that combines quantum computing with machine learning, exploring whether quantum computers — which process information using the unusual rules of quantum physics — can run certain learning tasks faster or better than conventional computers.

A random forest is a machine learning model that builds many different decision trees on random slices of the data and combines their predictions by vote or average, producing results that are more accurate and stable than any single tree.

A reasoning model is a large language model trained to work through a problem step by step before answering, spending extra effort to think the question through rather than replying with its first instinct.

A recommendation system is software that predicts what a particular person is likely to want — products, content, connections — and surfaces those suggestions, by learning from that person's past behavior and from the behavior of others like them.

A recurrent neural network (RNN) is a type of neural network built to handle sequences — such as text, speech, or time-series data — by processing one element at a time while carrying forward a memory of what came before, so earlier inputs can influence later ones.

Red teaming is the practice of deliberately attacking or stress-testing an AI system to uncover its weaknesses, harmful behaviors, and ways it can be misused, so those flaws can be fixed before the system is released to the public.

Regression is a supervised machine learning task that predicts a continuous numerical value — a price, a temperature, an amount — by learning the relationship between input variables and that number from past examples where the answer is known.

Regularization is a set of techniques that discourage a machine learning model from becoming too complex or relying too heavily on any single detail, helping it generalize to new data and reducing overfitting.

Reinforcement learning (RL) is a type of machine learning in which a system learns by trial and error, taking actions and adjusting its behavior based on rewards or penalties it receives, rather than being shown the correct answers in advance.

Reinforcement learning from human feedback (RLHF) is a training technique that uses people's judgments about which AI responses are better to teach a model to produce more helpful, honest, and appropriate output, rather than relying on data labels alone.

Responsible AI is the practice of designing, building, and deploying AI systems so that they are fair, transparent, accountable, safe, and respectful of privacy, with clear human responsibility for the outcomes they produce.

Retrieval-augmented generation (RAG) is a technique that lets an AI look up relevant information from an outside source — like a company's documents or a database — and use what it finds to answer your question, instead of relying only on what it memorized during training.

A reward model is a separate AI model trained to predict how a human would rate any given response, so that its scores can stand in for slow, expensive human judgment and guide the training of another model at scale.

Robotics is the field concerned with designing, building, and programming robots — physical machines that can sense their surroundings and carry out actions in the real world — increasingly using artificial intelligence to perceive, decide, and adapt rather than only repeating fixed movements.

Self-attention is a mechanism inside transformer models that lets each element of a sequence — such as each word in a sentence — weigh its relationship to every other element in the same sequence, so the model can interpret each part in light of the whole.

Semantic search is a way of searching that matches on meaning rather than exact words, so a query returns relevant results even when they use different wording — because the system compares the underlying ideas, not just the letters.

Sentiment analysis is an AI technique that determines the emotional tone of a piece of text — typically whether it is positive, negative, or neutral — so opinions expressed in writing can be measured automatically and at scale.

Sequence-to-sequence is a family of machine learning approaches that takes one ordered sequence as input — such as a sentence — and produces another ordered sequence as output, even when the two differ in length, using one component to read the whole input and another to generate the result.

Similarity search is the task of finding the items in a collection that are most alike a given item — usually by comparing number-representations of their meaning and returning the closest matches, even across millions of entries, fast.

A small language model (SLM) is a language model with far fewer internal parameters than the largest models, trading some breadth of capability for the ability to run cheaply, quickly, and often directly on a phone or laptop.

Softmax is a mathematical function used at the output of many neural networks that converts a list of raw, unbounded scores into a set of probabilities that are all positive and add up to one, turning the model's internal numbers into a confidence spread across possible answers.

Speech recognition is the AI capability of identifying the words in spoken audio, converting the sounds of human speech into recognized words a computer can act on — whether to transcribe them, obey a command, or trigger a response.

Speech-to-text (STT) is technology that converts spoken language into written text, automatically transcribing what someone says into words on a screen.

Stable Diffusion is a popular open text-to-image AI model that generates pictures from written descriptions, notable for being openly available so anyone can download, run, and build on it themselves rather than only using it through one company's service.

Stochastic gradient descent (SGD) is a method for training machine learning models that updates the model's settings using the error from one small, randomly chosen batch of data at a time, rather than the whole dataset at once, making training far faster and more practical at large scale.

Supervised learning is a type of machine learning in which a model learns from examples that are already labeled with the correct answer, so it can work out the relationship between an input and its right output and then apply that to new, unlabeled cases.

A support vector machine (SVM) is a supervised machine learning algorithm that classifies data by finding the boundary that separates the categories with the widest possible gap between them, using only the trickiest borderline examples to define where that boundary sits.

Synthetic data is artificially generated information — produced by an algorithm or simulation rather than recorded from real-world events — that is used in place of, or alongside, real data to train, test, or protect the privacy of machine learning systems.

Synthetic media is any content — images, audio, video, or text — that has been generated or significantly modified by AI rather than captured or created entirely by a human, ranging from AI-generated artwork and voices to fully fabricated photorealistic scenes.

A system prompt is a set of background instructions given to an AI model before a conversation begins, setting its role, behavior, tone, and rules — usually written by whoever built the application rather than the end user, and working quietly behind the scenes.

A tensor processing unit (TPU) is a specialized chip designed by Google specifically to speed up the kind of mathematics that neural networks rely on, built to perform machine learning calculations more efficiently than a general-purpose processor or even a graphics processing unit.

Text-to-image is a type of generative AI that creates an original picture from a written description, turning a sentence describing what you want into a brand-new image that matches it.

Text-to-speech is technology that converts written text into spoken audio, generating a natural-sounding human voice that reads the words aloud.

Text-to-video is a type of generative AI that creates an original short video clip from a written description, turning a sentence describing a scene into moving footage that didn't exist before.

Time-series forecasting is the use of historical data points recorded in time order to predict future values, identifying patterns such as trends and repeating cycles in the past so a model can estimate what comes next.

A token is the small chunk of text — often a whole word, a piece of a word, or a punctuation mark — that an AI language model reads and generates, since these models work in tokens rather than in letters or whole sentences.

Tokenization is the process of breaking text into tokens — the small chunks, often whole words or pieces of words, that an AI language model actually reads and works with, since models operate on these units rather than on raw letters or whole sentences.

Tool use is an AI model's ability to reach beyond its own text output and operate external resources — running a search, calling software, querying a database, or executing code — to gather information or take action it couldn't manage on its own.

Training data is the collection of examples an AI model learns from during training — the photos, text, sounds, or other information it studies to discover the patterns it will later use to make predictions.

Transfer learning is a machine learning approach where a model developed for one task is reused as the starting point for a related task, so the knowledge it already gained is carried over instead of training a new model from scratch.

A Transformer is a type of neural network architecture that processes language by learning which parts of a text are most relevant to each other, and is the foundation on which most modern AI language systems are built.

Transparency in AI is the principle that the workings of an AI system — what data it uses, how it makes decisions, who built it, and what its limits are — should be open and visible to the people who use it, are affected by it, or are responsible for overseeing it.

Underfitting is when a machine learning model is too simple to capture the real patterns in its training data, so it performs poorly even on the examples it was trained on, let alone on new ones.

Unsupervised learning is a type of machine learning in which a system is given data with no labels or correct answers and left to find patterns, groupings, or structure in it on its own.

A variational autoencoder (VAE) is a type of generative AI that learns to compress data into a compact, organized summary and then rebuild it, in a way that also lets it produce brand-new examples by sampling from that learned summary space.

A vector database is a database built to store embeddings — lists of numbers representing the meaning of text, images, or other data — and to quickly find the entries whose meaning is most similar to a given query.

A vision-language model (VLM) is an AI model that can take in both images and text together and respond in language — describing a picture, answering questions about it, or reasoning about what it shows.

Voice AI is artificial intelligence that interacts through spoken language — understanding what people say, responding in a natural-sounding voice, and powering hands-free tools like voice assistants, phone systems, and dictation.

Watermarking, in the context of AI, is the practice of embedding a hidden, machine-detectable signal into AI-generated content — an image, audio clip, video, or block of text — so that it can later be identified as synthetic and, sometimes, traced back to the system that produced it.

Weights are the adjustable internal numbers of a neural network that determine how strongly each connection influences the next — the values a model tunes during training and the place where everything it has learned is ultimately stored.

A word embedding is a representation of a word as a list of numbers that captures its meaning, positioned so that words used in similar ways end up with similar numbers — letting software work with meaning rather than just spelling.

Workflow automation is the practice of stringing a series of tasks together so they run automatically in sequence, without someone doing each step by hand — and increasingly, with AI handling steps that once needed human judgment.

A world model is an AI system's learned internal representation of how an environment behaves, used to predict what will happen next in response to actions, so the system can plan and rehearse decisions internally rather than relying solely on real-world trial and error.

Zero-shot learning is when an AI performs a task it was given no examples for, relying entirely on its general knowledge and the instruction it's given rather than on demonstrations of that specific task.