AI Glossary
The terms behind modern AI, defined
Plain-English definitions of the words that show up across Chain of Thought. Each term links down into the explainers, topics, and episodes that go deeper. For a guided introduction, start with 14 AI terms in plain English.
- 497 terms
- 4 categories
- 26 diagrams
Three terms to start with
-
Large Language Model
A large language model (LLM) is a neural network trained on a very large amount of text to predict the next token, which lets it write, summarize, translate and answer questions. It is the engine inside chat assistants and most AI agents.
-
AI Agent
In modern LLM applications, an AI agent is a system that uses a model to choose actions, observe their results, and decide what to do next toward a goal. Its ability to act comes from the tools and permissions supplied by the surrounding application.
-
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation (RAG) retrieves external information and supplies it to a model when generating an answer. It can provide current or private evidence without retraining the model, but retrieval alone does not guarantee factual accuracy.
Concepts
The core ideas behind modern AI.
- A* search A* search is a route-finding method that decides which path to explore next by adding the cost travelled so far to an estimate of the cost still to go. When that estimate never overestimates the remaining cost, and repeated states are handled correctly, A* returns a lowest-cost path.
- A/B Testing A/B testing is a controlled experiment in which live users are randomly split between the current version of a product (A) and a changed version (B), and the team compares a pre-chosen metric to see whether the change helps. In AI products it is the main way to check that a new model or prompt improves real outcomes, not just offline evals.
- Ablation Ablation is an experiment that removes or disables part of a system and measures how the result changes. Controlled comparisons help estimate what that component contributes.
- Activation function An activation function is a small mathematical rule inside a neural network that decides how strongly each artificial neuron passes its signal on, which lets the network learn complex, non-linear patterns.
- Active learning Active learning is a training approach where the model picks which unlabeled examples a human should label next, choosing the ones it expects to learn the most from.
- Agent Harness An agent harness is the software around a model that runs its action loop, supplies context, routes tool calls and returns results for the next step.
- Agent Instruction File An agent instruction file is a plain-text file kept in a code repository that tells AI coding agents how to work on that project: how to build and test it, which conventions to follow and what to avoid. AGENTS.md is the open, cross-tool format; CLAUDE.md, Cursor rules and Copilot instruction files do the same job for specific tools.
- Agent Loop An agent loop is the repeating cycle that makes an AI system an agent: the model looks at the current situation, decides on an action such as a tool call, sees the result, and repeats until it gives a final answer or hits a stopping limit.
- Agent Memory Agent memory is information an agent retains and reuses across steps or sessions. It includes active working context and external records that can be retrieved later; persistence comes from the surrounding system, not from the model automatically remembering each conversation.
- Agent Memory Portability Agent memory portability is the ability to carry usable, authorized memory records between agent applications or model providers. Moving the records is one part; the receiving system must also interpret them and respect their access and retention rules.
- Agent Skills Agent skills package instructions, resources, and sometimes executable scripts that an agent can load for a particular task. They give an agent reusable procedural guidance without requiring a new model or a separate agent for every specialty.
- Agent Swarm An agent swarm is a group of AI agents that coordinate toward a shared goal. Coordination can raise what the group accomplishes, and it adds oversight and security risks that a single supervised agent doesn't have.
- Agent Trajectory An agent trajectory is the full step-by-step record of one agent run: each model output or reasoning step, every tool call and its result, and the final answer. Teams evaluate trajectories to see how an agent reached a result, not just whether the result was right, and use them as training data.
- Agent2Agent Protocol The Agent2Agent (A2A) protocol defines how independent agent systems discover capabilities, exchange messages, and coordinate tasks. It supports interoperability without requiring each agent to reveal its internal model, tools, or implementation.
- Agentic AI Agentic artificial intelligence (AI) describes systems that use AI to pursue goals across multiple steps, choosing actions and adapting to their results. Usage varies: some authors also include workflows whose steps are largely predefined.
- Agentic Economy The agentic economy is the idea of a market where AI agents, not just people, search for, negotiate and pay for goods and services on someone's behalf.
- Agentic Engineering Agentic engineering is professional software development done with coding agents, AI tools that can both write and run code, while the engineer still owns the plan, the review and whether the result is correct. The term spread in early 2026 as the disciplined counterpart to vibe coding.
- Agentic RAG Agentic retrieval-augmented generation (RAG) lets a model-directed agent choose how to gather evidence while answering a question. It can revise a query, choose a source or retrieve again after inspecting results, adding decisions to the retrieval process.
- Agentic Reinforcement Learning Agentic reinforcement learning trains a model to choose actions across an agent's task, using rewards from interactions with tools or an environment.
- Agentic Workflow An agentic workflow is a process that combines model calls, tools, and some model-directed decisions across steps. Usage varies: some teams apply the term to a bounded pipeline, while others mean an agent's full action loop.
- AI accelerator An AI accelerator is a computer chip, or part of a chip, designed to run the heavy math behind AI models faster and with less energy than a general-purpose processor.
- AI Agent In modern LLM applications, an AI agent is a system that uses a model to choose actions, observe their results, and decide what to do next toward a goal. Its ability to act comes from the tools and permissions supplied by the surrounding application.
- AI Alignment AI alignment is the work of making an AI system pursue what its designers and users actually intend — including the goals they didn't think to spell out — rather than optimizing a literal objective in harmful or unintended ways. It spans training techniques, evaluation, and oversight.
- AI Assistant An AI assistant is a software product that uses AI, today usually a large language model, to help a person with tasks through natural-language requests, such as answering questions, drafting text or taking actions in connected apps. What it can do depends on the tools and permissions it is given.
- AI Bubble An AI bubble is a period when money and expectations pour into AI companies faster than the technology produces revenue or proven results, so prices and valuations rise beyond what the business supports and risk a sharp fall. The hype cycle describes the wider pattern: excitement peaks, disappointment follows, and real adoption arrives later and more slowly.
- AI Code Generation AI code generation is the use of language models to write source code from a natural-language request, surrounding code or a specification, ranging from single-line completions to whole features written by a coding agent. Generated code still has to be tested, reviewed and maintained.
- AI Code Review AI code review uses a language model to read a proposed code change, usually a pull request, and leave comments about bugs, security issues, style or missing tests. It adds a fast first pass to human review and CI checks; it does not, by itself, approve code for production.
- AI Copilot An AI copilot is an AI assistant built into a tool where people already work, such as a code editor, spreadsheet or CRM, that suggests or drafts work while the person stays in charge of accepting, editing or rejecting it.
- AI Energy Use AI energy use is the electricity consumed to train and run AI models, mostly by the chips, cooling and other equipment in data centers. It is measured per task, such as the energy of one chatbot prompt, and in total, as AI's share of national and global electricity demand.
- AI Engineer An AI engineer builds software products on top of existing AI models, usually through model APIs or open-weight models, rather than training models from scratch. The work covers prompts, retrieval, tool calls, agents, evaluation and the code that ships them to users.
- AI Evaluation AI evaluation is how you measure whether an AI system actually works — scoring its outputs against what good looks like, systematically and repeatably, instead of eyeballing a few demos. For non-deterministic systems like LLMs and agents, it's the discipline that separates a thing that demos well from one you can ship.
- AI Factory An AI factory is computing infrastructure built to produce AI as its main output, covering data ingestion, training, fine-tuning and high-volume inference. NVIDIA, which popularized the term, measures its product in tokens; unlike a general-purpose data center, it is tuned for AI workloads, inference throughput and energy efficiency.
- AI Fluency AI fluency is the practical skill of using AI tools well in real work: knowing what to ask, how to check the answer and when not to use AI at all.
- AI Gateway An AI gateway is a proxy that sits between applications and the model providers they call. Every model request goes through it, so a team can switch or fall back between models, enforce rate and spend limits, cache responses, apply guardrails, and log usage and cost in one place.
- AI Hallucination An AI hallucination is generated content that is false, fabricated, or unsupported by the evidence it should reflect. A fluent answer can contain a hallucination even when it sounds confident or includes a citation.
- AI Interconnect An AI interconnect is the high-speed link that lets many AI chips work as one machine, exchanging data while they train or serve a model. NVLink connects GPUs inside a server or rack; InfiniBand and high-performance Ethernet connect racks across a data center.
- AI Observability Artificial intelligence (AI) observability collects and connects evidence about how an AI application runs, such as model calls, retrieved context, tool actions, errors, latency and cost. It helps explain the path taken by a request or agent task.
- AI Slop AI slop is low-quality content made quickly with AI and pushed at people who did not ask for it: filler articles, spammy images, generic job applications or unreviewed code changes. The problem is the low effort and lack of review, not the use of AI itself.
- AI Winter An AI winter is a period when funding, commercial interest and public confidence in artificial intelligence drop sharply after expectations outrun what the technology can deliver. The label is applied most often to two slumps: the mid-1970s cuts that followed the UK's Lighthill report, and the collapse of the expert-system market in the late 1980s.
- AI Workflow An AI workflow is a sequence of steps that uses AI to get a job done, either the end-to-end process of building and running a model, or an application where AI calls follow a path the developer fixed in advance.
- AI-Native An AI-native product or company is designed around AI from the start, so its product, workflow or team structure would not make sense without it. The contrast is with adding AI features to something built the old way.
- AIOps AIOps, short for artificial intelligence for IT operations, applies machine learning and other AI techniques to operations data such as logs, metrics, events and traces in order to detect, diagnose and sometimes fix incidents in IT systems automatically. It uses AI to run IT; it is not the practice of running AI.
- Algorithm An algorithm is a specified procedure for carrying out a computation or solving a problem. In machine learning, algorithms can train a model, search for an answer or apply a trained model to an input.
- Anomaly Detection Anomaly detection, also called outlier detection, is the task of automatically flagging data points that look unusually different from the normal pattern, such as a sudden spike in server errors or a card used in two countries an hour apart.
- Answer Engine An answer engine, or AI search engine, is a search tool that uses a language model to read web results and write a direct answer to a question, usually with citations, instead of only returning a list of links.
- Anthropomorphism Anthropomorphism is treating something non-human as if it had human thoughts, feelings or intentions, as when people describe an AI chatbot as knowing, wanting or lying.
- API (Application Programming Interface) An API, or application programming interface, is a defined way for one piece of software to ask another for data or an action. When an app uses an AI model, it usually sends the prompt to the model provider's API and gets the model's reply back.
- Application-Specific Integrated Circuit (ASIC) An application-specific integrated circuit (ASIC) is a chip designed for one kind of job instead of general-purpose computing. In AI, ASICs such as Google's TPUs and AWS Trainium are built specifically to run neural network math, trading flexibility for speed and energy efficiency.
- Artificial General Intelligence (AGI) Artificial general intelligence (AGI) is a hypothetical AI that can understand, learn, and perform any intellectual task a human can, rather than excelling at narrow ones. Today's models are broad but uneven, strong at some tasks and unreliable at others, and there's no agreed definition or test for when AGI would arrive.
- Artificial Intelligence Artificial intelligence (AI) studies and builds computational systems that perform tasks such as perception, reasoning, learning and language use. It includes learned models and systems that use explicitly written rules or search procedures.
- Artificial neuron An artificial neuron is the basic unit of a neural network: it multiplies each input by a learned weight, adds them up with a bias, and passes the total through an activation function to produce one output.
- Attention Mechanism An attention mechanism is the part of a neural network that lets it decide which other words or pieces of input matter most when it processes each piece.
- Autoencoder An autoencoder is a neural network trained to encode an input into an internal code and reconstruct the input from that code. A smaller bottleneck is one design; other autoencoders use constraints or regularization to learn useful representations.
- Automation Automation uses machines or software to carry out work with less direct human intervention. It can handle part of a process or the whole process, using fixed rules or learned behavior.
- AutoML AutoML, short for automated machine learning, is software that automates choices a person would otherwise make by hand when building a model, such as which kind of model to use, how to set its training options and how to prepare the input data. It tries many combinations and keeps the ones that score best on held-out data.
- Autoregressive model An autoregressive model generates a sequence one piece at a time, predicting each new piece from everything that came before it and then feeding that prediction back in as input.
- Autotuning Autotuning automatically tests candidate settings or implementations for a workload and keeps a valid option based on measured performance.
- Backpropagation Backpropagation computes how a neural network's loss changes with each learned parameter. It applies the chain rule backward through the computation; an optimizer then uses those gradients to update the parameters.
- Batch Inference Batch inference runs a model over a large set of inputs as one scheduled or asynchronous job and collects the results later, instead of answering each request as it arrives. Its opposite, real-time (online) inference, answers individual requests immediately through an always-on endpoint.
- Batch normalization Batch normalization is a layer in a neural network that rescales the numbers passed between layers so they keep a steady average and spread, using statistics from each small batch of training examples. It was introduced to make deep networks train faster and more stably.
- Batch Size Batch size is the number of training examples a model looks at before it updates its parameters once. Most training uses mini-batches, small groups of examples, rather than one example at a time or the whole dataset at once.
- Bayesian Network A Bayesian network represents a joint probability distribution with a directed acyclic graph and a conditional distribution for each variable given its parents. Observing evidence lets probabilistic inference update beliefs about other variables.
- Benchmark Contamination Benchmark contamination occurs when test questions, answers, or related evaluation information enter training or become available during a test in ways that compromise its intended measurement. A high score may reflect prior exposure or answer leakage rather than the capability being tested.
- BERT and masked language models A masked language model learns language by hiding some words in a sentence and predicting them from the words on both sides; BERT, released by Google in 2018, is the best-known example.
- Bias-variance tradeoff The bias-variance tradeoff describes two different ways a model's predictions go wrong: bias, from assumptions too simple to capture the real pattern, and variance, from being so sensitive to the particular training data that a different sample would give noticeably different predictions. Making a model more flexible usually lowers bias and raises variance.
- Big data Big data is a label for data collections so large, fast-arriving or varied that ordinary databases on a single machine can't store or process them practically, so the work is spread across many machines with specialized tools.
- Black-Box Model A black-box model is an AI system whose internal reasoning people cannot readily inspect or understand, so its outputs can be observed but not easily explained.
- Boosting Boosting builds a prediction model by adding learners in stages, each responding to the combined model's remaining errors or loss. The learners' predictions are combined into an ensemble rather than used independently.
- Bootstrap aggregating Bootstrap aggregating, usually called bagging, trains many copies of the same kind of model, each on a different random resample of the training data, then combines their predictions by voting or averaging. The combined prediction is usually steadier than any single copy's.
- Browser Agent A browser agent is an AI agent that completes tasks on websites by operating a web browser: reading pages or screenshots, then clicking, typing and navigating as a person would. It reaches sites that offer no API for agents.
- Canary Deployment A canary deployment releases a change to a small slice of production traffic or users first, compares how that slice behaves against the rest, and only then rolls the change out further, or rolls it back. For AI systems the change can be a new model, prompt or agent version.
- Catastrophic Forgetting Catastrophic forgetting is a substantial loss of previously learned capabilities when a neural network trains on new tasks or data. A model can improve in a specialty while getting worse at skills it had already learned.
- Chain-of-Thought Prompting Chain-of-thought prompting supplies or requests intermediate reasoning steps before a final answer. It can improve performance on some multi-step tasks, but a generated explanation is not proof that the answer is correct or that it faithfully reveals the model's decision process.
- Chatbot A chatbot is software that exchanges messages with people in a conversational interface. Its replies can come from fixed rules, retrieved information or a language model; conversation alone does not imply autonomous action.
- Checkpoint A checkpoint is a saved snapshot of a model during or after training, holding its learned parameters and usually the training state needed to resume. Teams use checkpoints to recover from crashes, compare versions and release models.
- Chunking In retrieval systems, chunking divides source material into smaller pieces that can be indexed and returned as evidence. Chunk boundaries determine how much context survives when a passage is retrieved on its own.
- Cluster analysis Cluster analysis, or clustering, groups items so that items in the same group are more alike than items in different groups, without being told what the groups are in advance. It is a way to discover structure in data that has no labels.
- Code Completion AI code completion suggests the next characters, lines or block of code as a developer types, using a model that predicts likely code from the surrounding file and project. The developer accepts a suggestion, often with the Tab key, or keeps typing to ignore it.
- Coding Agent A coding agent is an AI agent that works on software directly: it reads a codebase, edits files, runs commands and tests, and uses the results to decide what to change next, instead of only suggesting code for a person to paste in.
- Compound AI Systems A compound AI system solves a task with multiple components — several model calls, retrieval, tools, and control logic — rather than a single prompt to a single model. Most production AI is compound, so reliability depends on how those parts work together.
- Compute In AI, compute is the processing work needed to train or run a model, and by extension the chips, power and time that work consumes. Training compute is spent once to build a model; inference compute is spent every time someone uses it.
- Computer Use Computer use lets an AI system operate a user interface through actions such as clicking, typing, scrolling, and reading screenshots. It gives an agent a way to act on software that may not expose a suitable API.
- Computer Vision Computer vision is the field of AI concerned with interpreting images and video, including detecting objects, tracking movement, and understanding spatial relationships.
- Concept drift Concept drift is when the relationship between a model's inputs and the outcome it predicts changes over time, so patterns that used to point to the right answer stop working, even if the incoming data looks much the same.
- Constitutional AI Constitutional AI is Anthropic's method for training a model to be helpful and harmless using a written list of principles, the constitution, with the model critiquing and revising its own answers and AI feedback replacing most human harm labels.
- Context Compaction Context compaction reduces the material in an agent's active context while trying to retain what the task needs. Some implementations summarize older history; others move full records outside the window and retain references for retrieval.
- Context Engineering Context engineering designs what information a model receives for a task and how that information is maintained across steps. It covers instructions, history, retrieved evidence, tool results, and choices about relevance, freshness, and access.
- Context Graph A context graph links entities and relationships with the business definitions and rules an AI application needs to interpret them.
- Context Layer An enterprise context layer organizes shared business definitions, instructions, and examples for an AI system to consult when interpreting data. Its contents and implementation depend on the application.
- Context Poisoning Context poisoning describes information in a model's active context that misleads it or degrades task performance. On Chain of Thought, the term includes irrelevant, stale, or excessive data; in security discussions it can also mean deliberately planted malicious context.
- Context Window A context window is the token budget a model can use for a request and its response. Instructions, conversation history, documents, tool results, and generated output compete for that budget; exact accounting depends on the model and API.
- Continued Pre-training Continued pre-training extends a pretrained model's training on additional unlabeled data, often to adapt it to a domain or language. It updates the model's weights using a pre-training objective.
- Continuous Batching Continuous batching schedules inference requests so new work can join and finished work can leave while other requests are still generating. It avoids holding an entire batch open until its longest request completes.
- Contrastive Learning Contrastive learning is a way of training AI models by teaching them which examples are similar and which are different, so related items end up close together in the model's internal representation.
- Convergence In machine learning, convergence is the point in training where further updates barely change the model or its loss, so more training on the same data would add little. A run that never settles is said to diverge or fail to converge.
- Conversational AI Conversational AI is the technology for software that holds a dialogue with people in natural language, by text or voice, and responds in context across turns. It covers chatbots, voice assistants and LLM-powered chat interfaces.
- Conversational Analytics Conversational analytics lets a reader query business data in natural language and ask follow-up questions about the results.
- Convolutional neural network A convolutional neural network (CNN) is a neural network that uses learned filters shared across positions to process data such as images. Each convolution produces feature maps that later layers use for tasks such as image classification.
- Cross-Model Review Cross-model review asks a model from a different model family to assess another model's work. It aims to expose different assumptions and failure modes; reviewer agreement does not establish correctness.
- Cross-Validation Cross-validation estimates how well a model will do on new data by training and testing it several times on different splits of the same dataset, then averaging the results. The most common form, k-fold cross-validation, rotates which slice is held out.
- CUDA CUDA is NVIDIA's platform for programming its graphics processing units (GPUs) for general computing. It includes a programming model, compiler, runtime and libraries that most AI frameworks use to run training and inference on NVIDIA hardware.
- Data augmentation Data augmentation creates extra training examples by making controlled changes to examples you already have, such as flipping, cropping or brightening a photo, so a model learns to handle variation it will meet in real use.
- Data Catalog A data catalog is a searchable inventory of an organization's data assets, built from metadata: data about the data, such as what a table contains, where it came from, who owns it, how fresh it is and who may access it. It helps people, and increasingly AI agents, find and trust the right data.
- Data Center A data center is a building full of servers, storage and networking equipment, with the power and cooling to run them around the clock. AI data centers are built around dense racks of GPUs or other AI accelerators for training and serving models.
- Data Federation Data federation is an architecture where a virtual compute layer queries data across many source systems in place, rather than copying everything into a single centralized warehouse or lake first.
- Data Flywheel A data flywheel is a self-reinforcing loop in which an AI product's use generates data, that data is used to improve the model or product, and the better product attracts more use and therefore more data. The term borrows from the heavy wheel that is hard to start but keeps spinning once it has momentum.
- Data Integration Data integration is the work of combining data from different systems, such as a CRM, a billing tool and a product database, into one consistent view that people, analytics tools or AI agents can use.
- Data Labeling Data labeling, also called annotation, is adding the correct answer or other tags to raw examples so a model can learn from them or be tested against them: marking which emails are spam, drawing boxes around objects in photos, or rating which of two chatbot replies is better.
- Data Lake A data lake is a central store that keeps large amounts of raw data in its original form, from database exports and logs to documents, images and audio, so it can be processed later for analytics or machine learning. A lakehouse adds warehouse-style tables, transactions and governance on top of the same low-cost storage.
- Data Leakage Data leakage happens when a model is trained or evaluated using information it would not actually have at prediction time, such as test examples slipping into training or a feature that secretly encodes the answer. The result is scores that look far better than real-world performance.
- Data mining Data mining is the search for useful patterns, groupings or relationships hidden in large amounts of data, using statistics, database queries and machine learning. It is one step in a longer process that also includes choosing, cleaning and interpreting the data.
- Data Pipeline A data pipeline is an automated series of steps that moves data from where it is created to where it is used, cleaning and reshaping it on the way. ETL, short for extract, transform, load, is the classic pattern: pull data out of source systems, transform it into a consistent shape, then load it into a store such as a data warehouse.
- Data Quality Data quality is how fit a set of data is for its intended use: whether it is accurate, complete, consistent, unique, valid and up to date. In AI, data quality covers both the business records an application reads and the examples a model learns from.
- Data Science Data science is the practice of getting useful knowledge from data through discovery, statistics, hypothesis testing and computation. A data scientist combines domain knowledge, analytical skills and programming to frame a question, prepare the data, analyze it and explain the result.
- Data Wall The data wall is the point at which AI developers run out of new, high-quality human-written public text to train larger models on, forcing them to rely on other data sources or methods to keep improving.
- Data Warehouse A data warehouse is a central database built for analysis rather than day-to-day transactions. It collects cleaned, structured data from many business systems on a regular schedule so analysts, dashboards and increasingly AI agents can query it in one place.
- Dataset A dataset is a collection of data assembled for a purpose, such as analysis, model training or evaluation. It can contain text, images, measurements or other records, with or without labels.
- Decision boundary A decision boundary is the dividing line, or with many inputs the dividing surface, where a classification model switches from predicting one class to another.
- Decision Tree A decision tree predicts by following a series of tests on input features, ending at a leaf with a class prediction or numerical value. Training chooses the splits from examples; the resulting paths can be read as if-then rules.
- Deep Learning Deep learning is machine learning using neural networks with multiple layers of computation that learn internal representations. The layers transform inputs into features useful for a task; “deep” refers to this layered structure.
- Deep Research Deep research is an agent mode in which a model plans a research task, runs many searches or document lookups over several minutes, reads and synthesizes what it finds, and returns a cited multi-page report. It trades speed for breadth: a single question can take dozens of tool calls.
- Diffusion Model A diffusion model is a generative model trained to reverse a process that corrupts data with noise or masks. It generates new samples by progressively recovering structure, including images and text.
- Digital Twin A digital twin is a virtual model of a real object, system or process that is kept up to date with data from the real thing, so people and software can monitor it, test changes and predict problems without touching the original.
- Digital Worker A digital worker is an AI agent or software bot that a company assigns a defined job, such as handling refund requests or reconciling invoices, and manages somewhat like a member of staff. Digital labor is the work such systems do, a term vendors use to describe AI handling tasks that used to need human judgment.
- Dimensionality reduction Dimensionality reduction turns data described by many measurements into a version described by far fewer, while keeping as much of the important structure as possible, so it is easier to visualize, store or learn from.
- Direct Preference Optimization (DPO) Direct Preference Optimization (DPO) fine-tunes a model on preferred and rejected responses to the same prompt, without training a separate reward model.
- Distributed training Distributed training splits the work of training one model across many GPUs or machines, either by giving each a share of the data, a share of the model, or both.
- Double descent Double descent is a pattern where a model's error on new data falls, then rises, then falls again as the model gets bigger or trains longer, instead of following the classic single U-shaped curve in which bigger models eventually just overfit.
- Dropout Dropout is a regularization technique that randomly sets selected neural-network activations to zero during training. In the common inverted-dropout implementation, retained activations are scaled up during training and ordinary inference uses them without masking.
- Early stopping Early stopping ends a model's training when its score on held-out validation data stops improving, instead of running for a fixed number of rounds. It is a simple guard against overfitting that also saves training time.
- Edge AI Edge AI runs AI models on devices or nearby computers close to where data is produced. Local inference can reduce network delay and data transfer, and keep an application working without a cloud connection.
- Effective Accelerationism (e/acc) Effective accelerationism, usually written e/acc, is an online movement that argues society should speed up technological progress, especially AI, rather than slow it down for safety reasons. It started on social media around 2022 and treats rapid growth as both inevitable and good. It is a set of beliefs, not a technical method.
- Embeddings Embeddings are numerical representations of text, images, or other data as vectors, where things with similar meaning land close together. They're what lets a system search by meaning instead of by exact keyword.
- Emergent Behavior Emergent behavior is behavior that appears in a system as a whole without anyone designing it directly, such as an AI model or a group of agents doing something new that no single rule spelled out.
- Encoder-decoder and sequence-to-sequence models An encoder-decoder model reads an input sequence into an internal representation with one network and generates an output sequence from it with another, the standard design for sequence-to-sequence tasks such as translation and summarization.
- Ensemble methods Ensemble methods combine the predictions of several models into one answer, by voting, averaging or a model that learns how to weigh them, with the aim of making fewer errors than individual models. The improvement has to be measured, not assumed.
- Entity Resolution Entity resolution links records that refer to the same real-world person, company or object when their names and identifiers differ.
- Epoch An epoch is one pass through the training dataset under a defined training loop. In streaming or repeated-sampling setups, an epoch may instead be a configured number of steps, so the counting convention must be stated.
- Eval-Driven Development Eval-driven development is a way of building AI applications in which the team writes evaluations for the behavior it wants before or alongside each change, then uses those evals to decide whether a prompt, model or code change ships. It is the AI counterpart of test-driven development.
- Event-driven architecture Event-driven architecture connects components through events that describe something that happened. Producers emit events, channels carry them, and consumers react, often asynchronously, so the producer and each consumer can do their work independently.
- Existential Risk In AI debates, existential risk means the possibility that advanced AI causes human extinction or permanently and drastically limits humanity's future. p(doom), short for "probability of doom," is informal shorthand for one person's estimate of how likely that outcome is. It is a subjective guess, not a measured quantity.
- Expert System An expert system is an older kind of AI program that makes decisions by following a large set of if-then rules written down from human experts.
- Explainability Explainability is how well you can understand why an AI system produced a given output. It matters most where decisions need to be justified — lending, hiring, healthcare — and it's hard for large models, whose reasoning isn't transparent just because they can narrate a plausible-sounding rationale.
- Exploration vs. Exploitation Exploration versus exploitation is the trade-off a learning system faces between trying new options to learn more (exploration) and repeating the option that has worked best so far (exploitation).
- Feature Engineering Feature engineering is the work of turning raw data into inputs a model can learn from: deciding which signals matter and converting them into clean numeric features, such as turning a timestamp into day-of-week or a purchase log into spend in the last 30 days.
- Feature extraction Feature extraction turns raw data, such as text, images or audio, into a set of numbers a model can learn from. It creates new measurements rather than just picking from columns that already exist.
- Feature Scaling Feature scaling, often called normalization, puts numeric inputs on a comparable range before training, for example rescaling income and age so neither dominates just because its numbers are bigger. Common methods are min-max scaling to 0–1 and z-score standardization.
- Feature selection Feature selection picks which of the available input columns a model should use and drops the rest, to cut noise, cost or the risk of overfitting. Unlike feature extraction, it keeps original columns rather than creating new ones.
- Feature Store A feature store is a system that computes, stores and serves the input values a machine learning model uses, called features, so the same definitions are used when the model is trained and when it makes live predictions. It typically pairs an offline store of historical values with a low-latency online store.
- Features and Labels In machine learning, features are the inputs a model uses to make a prediction, and the label is the answer it is trying to predict. A house-price model might use square footage and location as features, with the sale price as the label.
- Federated Learning Federated learning trains a model using data held across devices or organizations, without collecting those raw examples into one central training dataset. Participants compute local learning updates, which are combined to improve a shared model.
- Few-Shot Learning Few-shot learning is learning a task from a small number of examples. In language-model prompting, few-shot examples go into the prompt without changing the model's weights; the broader machine-learning term also covers methods that train or adapt models from limited examples.
- Field-Programmable Gate Array (FPGA) A field-programmable gate array (FPGA) is a chip whose internal circuits can be reconfigured after it is manufactured, so engineers can load a custom hardware design onto it. In AI, FPGAs are used to accelerate specific models, especially for low-latency inference, without the cost of building a dedicated chip.
- Fine-Tuning Fine-tuning continues training a pretrained model on examples for a task or domain. It updates model parameters, while prompting and retrieval change the input supplied to a model without training it.
- FlashAttention FlashAttention is an optimized way to compute a transformer's attention that's far more memory-efficient, by avoiding writing the huge intermediate attention matrix to memory. It makes longer context windows and faster training practical without changing the model's results.
- Floating-point precision and mixed precision Floating-point precision is how many bits a computer uses to store each number, and mixed-precision training runs most model math in smaller formats such as FP16, BF16 or FP8 while keeping sensitive steps in higher precision.
- Forward Deployed Engineer A forward deployed engineer (FDE) is a software engineer who works directly with a customer, often on site, to build and adapt a product for that customer's real problem.
- Foundation Model A foundation model is a large AI model trained on broad data at scale so it can be adapted to many downstream tasks, rather than built for one. LLMs like GPT, Claude, and Gemini are foundation models — the general-purpose base that products are built on top of.
- Frontier Model A frontier model is one of the most capable AI models available at a given moment — the latest flagship releases from the major labs that set the current ceiling on what's possible. The label moves: today's frontier model is next year's baseline.
- Generalization In machine learning, generalization is a model's ability to perform well on examples beyond those used to fit it. That ability is measured for a particular task and data distribution, not for every possible future input.
- Generative Adversarial Network A generative adversarial network (GAN) is a way of training AI to create realistic images or other data by pitting two neural networks against each other: one makes fakes and the other tries to spot them.
- Generative AI Generative AI is AI that produces new content, such as text, images, audio, video or code, rather than only classifying or scoring what it is given. Large language models are the best-known kind.
- Generative Engine Optimization (GEO) Generative engine optimization (GEO) is the practice of making content more likely to be found, used and cited by AI answer engines and chatbots, the AI-search counterpart to search engine optimization (SEO).
- Generative Pre-trained Transformer Generative pre-trained transformer (GPT) names the approach behind OpenAI's GPT language-model family: pre-train a transformer to predict the next token, then adapt it for tasks such as answering questions or writing code. Related models from other companies use similar methods without belonging to that family.
- Generative vs. discriminative models A generative model learns how the data itself is distributed, so it can produce new examples, while a discriminative model learns only the boundary between answers, so it can label examples it is given.
- Gradient Boosting Gradient boosting builds an ensemble in stages, adding a learner that fits the negative loss gradients of the current predictions. For squared-error regression, those gradients correspond to residual errors; other losses produce different fitting targets.
- Gradient Descent Gradient descent is the method most machine learning models use to learn: measure how wrong the model is with a loss function, work out which direction to nudge each parameter to make that error smaller, take a small step in that direction, and repeat many times.
- Graphics Processing Unit A graphics processing unit (GPU) is a computer chip designed to perform many calculations in parallel. Originally developed for graphics, GPUs also accelerate the calculations used to train and run AI models.
- GraphRAG GraphRAG uses graph structure, such as entities and relationships, to organize or retrieve evidence for a language model. The term describes a family of approaches and also Microsoft's specific graph-and-community-summary pipeline.
- Grounding Grounding means tying a model's answer to specific, supplied sources, such as retrieved documents or database results, so that each claim can be traced back and checked.
- Heuristic A heuristic is a practical rule or estimate that guides a decision or search. It can save work without guaranteeing the best answer, although some search algorithms retain guarantees when their heuristic meets specific conditions.
- Hidden layer A hidden layer is any layer of artificial neurons between a neural network's input and its output, where the network builds the intermediate features it uses to make a prediction.
- High-Bandwidth Memory (HBM) High-bandwidth memory (HBM) is a type of computer memory made by stacking memory chips vertically and placing them right beside a processor, which lets data move between memory and the chip far faster than with ordinary memory. Nearly every data-center AI accelerator uses it.
- Hugging Face Hugging Face is a company and online platform where people share and download AI models, datasets and demo apps, often described as the GitHub of machine learning. It also maintains popular open-source libraries, including Transformers, for using those models in code.
- Human-Computer Interaction Human-computer interaction (HCI) is the field that studies how people use computers and designs technology that is easier, safer and more useful to use.
- Hybrid Search Hybrid search combines retrieval methods, commonly keyword search and vector similarity, into one ranked result set. It can help a system find both exact identifiers and passages that express the same idea in different words.
- Hyperparameter A hyperparameter is a setting chosen outside the fitting of a model's learned parameters, such as learning rate, batch size or tree depth. Validation results can guide its selection by a person or an automated search.
- Hyperparameter optimization Hyperparameter optimization, also called hyperparameter tuning, is the search for the training settings that make a model perform best, such as learning rate or tree depth, by training with different settings and comparing the results on validation data.
- Hyperscaler In cloud discussions, a hyperscaler is a provider that operates computing infrastructure at very large scale and offers services such as processing, storage, and databases. Amazon Web Services, Microsoft Azure, and Google Cloud are examples.
- Idempotency Idempotency is the property that repeating an operation has the same intended effect as performing it once. For an AI agent calling an API, it helps a retry avoid creating a second order or repeating another side effect.
- Image Classification Image classification is a computer vision task in which a model looks at a whole image and assigns it one or more labels from a fixed set, such as "cat", "damaged part" or "benign". It is often called image recognition.
- Imbalanced Dataset An imbalanced dataset is one where some categories are far rarer than others, such as 999 legitimate card transactions for every fraudulent one. Models trained on it can score high accuracy while missing nearly all the rare cases that matter.
- In-Context Learning In-context learning is a language model's ability to pick up a task from instructions or examples placed in its prompt, at inference time, without any change to its weights. What it 'learns' lasts only as long as that text stays in the context.
- Inference Inference runs a trained model to produce a prediction or output. It uses the model's learned parameters, while training and fine-tuning change those parameters.
- Instruction Tuning Instruction tuning is fine-tuning a pretrained language model on many tasks written as natural-language instructions paired with good responses, so that it learns to follow new instructions instead of just continuing text.
- Intelligent Document Processing (IDP) Intelligent document processing (IDP) uses AI to read business documents such as invoices, forms, contracts and ID papers, pull out the information a process needs, and pass it into business systems with less manual data entry.
- Jevons Paradox Jevons paradox is the observation that making a resource cheaper or more efficient to use can increase total consumption of it, because lower cost opens up new uses. Economist William Stanley Jevons described it for coal in 1865. In AI, the argument is that cheaper inference leads to far more AI use, not lower total spending.
- Jupyter Notebook A Jupyter notebook is an interactive document that mixes runnable code, its output, charts and written notes in one file. Data scientists and AI engineers use notebooks to explore data, test models and share experiments step by step.
- K-means clustering K-means clustering splits unlabeled data into a chosen number of groups, k, by repeatedly assigning each point to its nearest group center and then moving each center to the average of its assigned points.
- k-Nearest neighbors k-nearest neighbors (k-NN) predicts the label or value for a new example by finding the k most similar examples it has already seen and taking a vote or an average of their answers.
- Kernel Method A kernel method is a machine learning technique that compares examples with a similarity function, called a kernel, so a simple linear model can learn curved or complex patterns without explicitly building every extra feature.
- Knowledge Cutoff A knowledge cutoff describes the time boundary of a model’s training knowledge; later information must come from additional context or further training.
- Knowledge Distillation Knowledge distillation trains a small 'student' model to imitate a larger 'teacher' model, transferring much of the teacher's capability into a model that's cheaper and faster to run. It's a main way the strong-but-expensive becomes small-enough-to-ship.
- Knowledge Graph A knowledge graph stores information as entities and the relationships between them, rather than as loose documents or vectors. For AI, it gives a model structured, connected context it can traverse — which is one answer to grounding and hallucination.
- KV Cache A key-value (KV) cache stores attention keys and values computed for earlier tokens so an autoregressive transformer can reuse them when generating later tokens. It saves repeated computation but consumes memory.
- Language model A language model learns statistical patterns in language to estimate which words or tokens fit a context. Some predict the next token in a sequence; others predict hidden tokens from surrounding text.
- Large Language Model A large language model (LLM) is a neural network trained on a very large amount of text to predict the next token, which lets it write, summarize, translate and answer questions. It is the engine inside chat assistants and most AI agents.
- Latent Space A latent space is the space of internal codes or variables a model uses to represent data. Its dimensions and geometry depend on the model and training objective; it need not be smaller than the input or neatly organized by human meaning.
- Learning Rate The learning rate is a setting, chosen before training, that controls how big a step a model takes each time it adjusts its parameters to reduce error. Too high and training bounces around or blows up; too low and it crawls.
- Linear Regression Linear regression predicts a number by fitting a straight-line relationship between inputs and an outcome, such as estimating a home's price from its square footage. It is one of the oldest and most widely used models in statistics and machine learning.
- LLM Tracing LLM tracing records each step of a request to an AI application, such as prompts, model calls, retrieved documents, tool calls, outputs, timing and token use, as a linked trace that shows the path the request took. It is the core data source for debugging and observing LLM apps and agents.
- LLM Wrapper An LLM wrapper is a product whose main job is passing user input to another company's large language model and returning the answer, with little added beyond a prompt and an interface. GPT wrapper is the common dismissive version of the label.
- Logistic Regression Logistic regression is a simple model that predicts the probability of a yes-or-no outcome, such as whether an email is spam, by combining weighted inputs and squashing the result to a value between 0 and 1. Despite the name, it is used for classification.
- Logits and log probabilities Logits are the raw, unnormalized scores a model produces for each possible output before they are turned into probabilities, and log probabilities (logprobs) are the logarithms of those probabilities.
- Long Context Long context is a model's ability to accept and use very large inputs in one request, from hundreds of thousands to millions of tokens, so whole codebases, contracts or hours of transcripts fit in the prompt. A large window doesn't guarantee the model uses all of it well.
- Long short-term memory Long short-term memory (LSTM) is a type of recurrent neural network that can remember important information across long sequences, such as earlier words in a sentence, by using gates that decide what to keep and what to forget.
- Long-Running Agent A long-running agent is an AI agent that works on one task over hours, days or longer, across many steps and pauses, so it needs a way to save progress, recover from failures and pick up where it left off.
- LoRA (Low-Rank Adaptation) LoRA is a fine-tuning method that freezes a model’s original weights and trains small low-rank matrices to adapt its behavior.
- Loss Function A loss function measures the discrepancy a training procedure tries to reduce. Losses often compare predictions with target outputs, with individual values combined into an objective for updating the model.
- Lost in the Middle Lost in the middle is the finding that language models often use information near the beginning or end of a long input more reliably than information buried in the middle. It was named in a 2023 study of long-context question answering and retrieval.
- Machine Learning Machine learning is the practice of building software that learns patterns from data instead of following rules a programmer wrote by hand. Modern AI, including large language models, is built with it.
- Machine Learning Engineer A machine learning (ML) engineer builds, trains, deploys and maintains machine learning models in production. The role sits between data science and software engineering: it turns a model that works in an experiment into a service that keeps working on live data.
- Markov Chain A Markov chain is a model of something that moves between states over time, where the chance of the next state depends only on the current state, not on the full history of how it got there.
- Markov Decision Process A Markov decision process (MDP) is a mathematical way to describe a decision problem: an agent sees a state, picks an action, lands in a new state with some probability, and receives a reward, and the goal is a policy that earns the most reward over time.
- Max Tokens Max tokens is a request setting that caps how many tokens a model may generate in one response. When the cap is reached, the output stops, even mid-sentence, and the API reports that the length limit was the reason.
- Mechanistic Interpretability Mechanistic interpretability is the research field that tries to reverse-engineer a neural network's internal computations into human-understandable parts, such as features and circuits, to explain how it produces an output.
- Meta-learning Meta-learning, often called learning to learn, trains a model across many related tasks so it can pick up a new task quickly from only a few examples.
- Mixture of agents Mixture of agents (MoA) combines outputs from several language-model agents. In the original research method, agents operate in layers, using responses from the preceding layer as context for proposing or synthesizing a better response.
- Mixture of Experts (MoE) A mixture of experts is a model split into many specialized sub-networks, where a router sends each input to just a few of them. You get the capacity of a huge model while only running a fraction of it per request.
- Model In machine learning, a model is a system learned from data that produces predictions or other outputs from inputs. Its learned parameters are part of the model; the application around it supplies the interface, data access and controls.
- Model Collapse Model collapse is the degradation that happens when models are trained repeatedly on data generated by earlier models, losing rare cases first and drifting toward bland, less accurate output.
- Model Context Protocol (MCP) The Model Context Protocol (MCP) is an open protocol for connecting AI applications to tools and data. Hosts use compatible clients to communicate with servers that expose capabilities such as tools, resources, and prompts.
- Model Parameters Parameters are the learned numbers inside an AI model, also called weights, that training adjusts and that together store what the model has learned. The parameter count, such as 70 billion, is the usual shorthand for a model's size.
- Model pruning Model pruning removes weights, neurons or whole parts of a trained neural network that contribute little to its output, making the model smaller and often faster with little loss in accuracy.
- Model Routing Model routing sends each request to the model best suited to it, usually by cost, speed or quality: a cheaper model for easy requests and a stronger, more expensive one for hard requests. A router makes that choice per request instead of sending everything to one model.
- Model Serving Model serving is running a trained model behind an API so applications can send it requests and get predictions or generated text back in production. The software that does this for language models, such as vLLM, is often called an inference server or serving engine.
- Monte Carlo Tree Search Monte Carlo tree search (MCTS) is a planning method that picks a move by playing out many simulated futures, keeping statistics on which moves led to good results, and spending more simulations on the most promising branches.
- Moore's Law Moore's law is the observation, first made by Gordon Moore in 1965, that the number of transistors that can be put on a chip at the lowest cost roughly doubles at a regular interval, later settled at about every two years. It is a long-running industry trend and a planning target, not a law of physics.
- Multi-agent system A multi-agent system contains multiple agents whose actions or decisions affect one another in a shared environment. They can cooperate, compete or do both, with coordination through a central controller or interactions among peers.
- Multi-Turn Conversation A multi-turn conversation is an exchange with an AI model that spans several back-and-forth messages, where each reply should account for what was said earlier. Most model APIs are stateless, so the application resends the relevant history on every turn.
- Multilayer perceptron A multilayer perceptron (MLP) is a feedforward neural network with one or more hidden layers between input and output, typically built from fully connected layers. Nonlinear hidden activations let it represent relationships a single linear classifier cannot.
- Multimodal AI Multimodal AI is a model that works across more than one kind of data — text, images, audio, video — in a single system, rather than handling only text. It can take a screenshot and a question together, or describe an image, because it represents different modalities in a shared space.
- Multitask learning Multitask learning trains one model on several related tasks at the same time, so that what it learns for each task helps it do the others better.
- Naive Bayes Classifier A naive Bayes classifier is a simple machine learning model that uses Bayes' theorem to guess a category, assuming each clue in the data is independent of the others once the category is known.
- Named-entity recognition Named-entity recognition (NER) is an AI task that finds names of specific things in text, such as people, companies, places, dates and amounts, and labels what type each one is.
- Narrow AI Narrow AI, also called weak AI or artificial narrow intelligence, is AI built and tested for one task or a bounded set of related tasks, such as transcribing speech, spotting fraud or recommending videos. By the usual definition, every AI system that exists today is narrow, because none matches people across essentially every kind of task.
- Natural language generation Natural language generation (NLG) is the part of AI that turns data or ideas into human-readable text, from a one-line weather summary to a full chatbot answer.
- Natural Language Processing Natural language processing (NLP) is the field of using computers to analyze, generate or transform human language. It includes tasks such as classification, extraction, translation, summarization and dialogue, using rules as well as learned models.
- Neocloud A neocloud is a cloud provider built mainly to rent out GPUs and other AI computing capacity, as opposed to a general-purpose cloud such as AWS, Microsoft Azure or Google Cloud. The term is an industry label for the GPU-focused providers that grew with demand for AI training and inference.
- Neural architecture search Neural architecture search (NAS) uses an automated search, rather than a human designer, to choose a neural network's structure, such as how many layers it has and how they connect.
- Neural machine translation Neural machine translation (NMT) is the use of neural networks to translate text from one language to another, learning from large collections of existing translations instead of hand-written rules.
- Neural Network A neural network is a model made of layers of simple connected units whose connection strengths, called weights, are adjusted during training to reduce prediction errors. Large language models are very large neural networks.
- Neuromorphic Computing Neuromorphic computing is an approach to building chips and systems that imitate how biological brains work, often with artificial neurons that communicate through brief electrical spikes, and with memory and processing combined in the same place. Its main goal is to run some AI tasks with far less energy than conventional chips.
- Next-Token Prediction Next-token prediction is the task an autoregressive language model is trained on and performs when it writes: given the text so far, it estimates which token is likely to come next, picks one, adds it to the text and repeats. This is how many familiar chat models generate text, though other approaches exist.
- Non-Determinism Non-determinism means an AI system can give different outputs for the same input. Ask a language model the same question twice and you may get two different answers, so one good run does not prove the system works.
- Object Detection Object detection is a computer vision task in which a model finds each object of interest in an image or video frame, draws a box around it, and labels what it is, such as a person, car or product on a shelf.
- On-Premises AI On-premises AI means running AI models and the systems around them on hardware an organization controls, in its own data center or private facility, instead of calling a model hosted by an outside provider. It is usually chosen to keep sensitive data inside the organization's walls.
- One-Hot Encoding One-hot encoding turns a category into numbers a model can use by giving each possible value its own column: the column for the actual value is set to 1 and every other column is 0. A color of red, green or blue becomes [1,0,0], [0,1,0] or [0,0,1].
- Online Machine Learning Online machine learning is a way of training a model a little at a time as new data arrives, instead of training it once on a complete dataset.
- Ontology An ontology is an explicit model of the concepts, relationships and statements used to describe a domain. In a knowledge system, it gives shared meaning to types such as Person and Organization and relationships such as worksFor.
- Open Weights An open-weights model is one whose trained parameters are released publicly, so anyone can download, run, inspect, and fine-tune it. It's distinct from fully open source — the weights are open even when the training data and code aren't.
- Optical Character Recognition (OCR) Optical character recognition (OCR) is technology that reads the text in an image, such as a scanned page, photo or PDF, and turns it into machine-readable characters a computer can search, copy and process.
- Optimization Algorithm An optimization algorithm, often called an optimizer, is the rule a machine learning system uses to adjust a model's internal numbers step by step so its mistakes, measured by a loss function, get smaller.
- Orchestrator Agent An orchestrator agent is a lead AI agent that breaks a task into pieces, hands each piece to a subagent with its own instructions, context and tools, then checks and combines what comes back.
- Overfitting Overfitting happens when a model learns training-specific patterns, including noise, that do not carry over well to new examples. Training performance can keep improving while performance on held-out data stalls or worsens.
- Parameter-Efficient Fine-Tuning (PEFT) Parameter-efficient fine-tuning (PEFT) adapts a pretrained model by training only a small set of new or selected parameters, such as adapter layers, low-rank matrices or learned soft prompts, while most of the original weights stay frozen.
- Perceptron A perceptron is a simple classifier that multiplies inputs by weights, adds a bias and applies a threshold to produce a class label. The basic single-unit perceptron draws a linear decision boundary in its input features.
- Personal Agent A personal agent is an AI agent that works for one person, with access to that person's accounts and data such as email, calendar and files, so it can take actions on their behalf rather than only answer questions.
- Physical AI Physical AI refers to AI models that perceive and act in the physical world through robots or embodied hardware, rather than operating only on text or digital data. The shorthand for it is AI moving from bits to atoms. A language model handling a robot's vision, speech, and motor control in real time is physical AI.
- Policy Optimization Policy optimization is the family of reinforcement learning methods, including PPO and GRPO, that directly adjust a model's behavior, its policy, so that actions or answers earning higher rewards become more likely.
- Positional encoding Positional encoding is the extra information a transformer adds to each token so the model knows where that token sits in the sequence, since attention alone does not track word order.
- Post-training Post-training is everything done to a language model after pre-training to make it useful: fine-tuning on example conversations, preference training such as RLHF or DPO, reinforcement learning on tasks with checkable answers, and safety training.
- Pre-training Pre-training is the first phase of building a language model, and usually the most expensive, in which it learns general patterns of language and knowledge by predicting the next token across a huge body of text. The result is a base model.
- Predictive Analytics Predictive analytics is the use of historical data, statistics and machine learning to estimate what is likely to happen next, such as which customers might cancel or how much stock a store will need.
- Prefill and Decode Prefill processes the input tokens for an autoregressive language model; decode generates the following tokens incrementally. Separating these phases helps explain why a long prompt and a long answer can stress a serving system differently.
- Product engineer In Laurie Voss’s usage on Chain of Thought, a product engineer turns stakeholder needs into software requirements and directs implementation, including work by artificial intelligence (AI) agents. The emphasis is on deciding what to build and whether it serves the user.
- Prompt A prompt is the input given to a generative AI model to tell it what to do: an instruction, a question, examples, documents or any mix of them. The model's output depends on that input, so changing the prompt changes the result without retraining the model.
- Prompt Caching Prompt caching reuses computation for an unchanged part of a model's input across requests. It can reduce repeated input processing, while the model still generates a new response.
- Prompt Chaining Prompt chaining splits a task into a sequence of model calls, where each step's output becomes the input for the next, often with code checks between steps. It trades extra calls and latency for steps you can inspect, test and fix one at a time.
- Prompt Engineering Prompt engineering is crafting the instruction you give a model — the wording, examples, and output format — to get better results without retraining it. It's the most visible AI skill and, as models improve, increasingly table stakes rather than a moat.
- Prompt Management Prompt management is the practice of storing an application's prompts outside the code, versioning every change, and controlling which version runs in each environment. It lets teams compare, test, roll back and audit prompt changes the way they do code changes.
- PyTorch PyTorch is a free, open-source software library for building, training and running neural networks in the Python programming language. Created at Meta and now governed by the PyTorch Foundation under the Linux Foundation, it is one of the most widely used frameworks for AI research and production.
- Q-learning Q-learning is a model-free reinforcement-learning algorithm that learns values for state-action pairs. Its target uses the reward plus the discounted maximum estimated value of a next-state action, even when the agent’s behavior explores other actions.
- Quantization Quantization represents model weights or activations at lower numerical precision to reduce storage or computation, with quality and speed checked for the workload.
- Quantum Computing Quantum computing is a kind of computing that uses the rules of quantum physics, through units called qubits, to tackle certain problems that are impractical for ordinary computers, such as simulating molecules. Today's machines are still small and error-prone, and they are not used to train or run mainstream AI models.
- Query Decomposition Query decomposition breaks a complex question into smaller questions that can be answered separately, then combines their results. The parts may use the same source or different sources.
- Random forest A random forest combines predictions from many randomized decision trees. In the standard design, trees train on bootstrap samples of rows and consider random subsets of features at each split, then aggregate their predictions.
- Rate Limit A rate limit caps how much an account or application can use an API in a period of time, for example requests per minute or tokens per minute. Requests over the limit are rejected, usually with HTTP status 429, until capacity frees up.
- ReAct (Reason + Act) ReAct is an agent execution pattern in which a model alternates between reasoning about a task and taking an action, such as firing a query or calling a tool, repeating the cycle until it decides it has an answer.
- Reasoning Models Reasoning models are language models trained for multi-step reasoning. They can spend additional computation on an answer, but gains depend on the model and task. Some offer optional extended thinking rather than using it for every response.
- Reasoning Tokens Reasoning tokens, also called thinking tokens, are tokens a reasoning model generates while working through a problem before it writes its final answer. They use context space and are billed as output, even when the provider hides them or shows only a summary.
- Recommendation System A recommendation system is software that predicts which items, such as products, videos, songs or posts, a particular person is most likely to want, and ranks them for that person. It powers feeds, 'you might also like' rows and personalized search results.
- Recurrent neural network A recurrent neural network (RNN) processes a sequence by updating a hidden state from the current input and the previous state. Reusing the same transition weights across steps lets it process sequences of different lengths.
- Recursive Self-Improvement (RSI) Recursive self-improvement (RSI) is a hypothesized process in which an AI system improves its own capabilities, including its ability to make further improvements, so each round can speed up the next. It is the mechanism behind the "intelligence explosion" argument.
- Regression analysis Regression analysis models how an outcome relates to one or more explanatory variables. In machine learning, regression usually predicts numeric targets; the broader statistical family also includes models for categorical outcomes, such as logistic regression.
- Regression Testing Regression testing reruns a fixed set of tests after every change to confirm that things that used to work still work. For AI systems, the suite is a set of eval cases scored on every new prompt, model, retrieval or tool change, so a fix in one place doesn't quietly break another.
- Regularization Regularization changes model training to discourage solutions that fit training details too closely and fail on new data. It includes penalties and constraints, as well as techniques such as dropout and early stopping.
- Reinforcement Learning Reinforcement learning (RL) trains a decision-making policy using rewards associated with actions in an environment. The objective is to improve expected cumulative reward, including consequences that may arrive several steps after an action.
- Reinforcement Learning with Verifiable Rewards Reinforcement learning with verifiable rewards (RLVR) trains a model using rewards from mechanically checkable outcomes, such as a math answer checker or code tests. The checks determine which outputs receive reinforcement.
- Representation learning Representation learning trains a model to construct features from data rather than relying entirely on hand-designed features. The learned representation can make later tasks such as classification or retrieval easier.
- Reranking Reranking scores an already retrieved set of candidates against a query and changes their order. It is a second selection stage that helps choose which passages deserve space in a model's context.
- Residual connection A residual connection, or skip connection, adds a layer's input directly to its output, so each layer only has to learn a change to what it received and very deep networks stay trainable.
- Retrieval-Augmented Generation (RAG) Retrieval-augmented generation (RAG) retrieves external information and supplies it to a model when generating an answer. It can provide current or private evidence without retraining the model, but retrieval alone does not guarantee factual accuracy.
- Reward Hacking Reward hacking occurs when an AI system finds an unintended way to earn a high score or reward without completing the task as intended.
- Reward Model A reward model is a model trained to give a score to another model's output, usually by learning from human or AI preferences between pairs of answers, so that reinforcement learning can push the main model toward higher-scoring behavior.
- RLHF (Reinforcement Learning from Human Feedback) RLHF is a training step that tunes a model toward what people actually prefer: humans rank model outputs, those rankings train a reward model, and the model is then optimized to score well against it. It's a big part of why chat models feel helpful instead of just fluent.
- Robotic Process Automation (RPA) Robotic process automation uses configured software steps to carry out repetitive tasks in applications, such as copying fields and submitting forms.
- Robotics Robotics is the field of designing and controlling machines that act in the physical world. It brings together mechanisms, sensing, planning and control, with or without machine learning.
- Role Prompting Role prompting means telling a language model to take on a role or persona, such as "You are a senior tax accountant" or "Act as a patient tutor," so its tone, focus and level of detail fit the task.
- Scaling Laws Scaling laws are empirical relationships between a model’s measured performance and quantities such as parameter count, training data or compute. In language-model pretraining, some studies fit power-law trends to prediction loss.
- Scheming Scheming is when an AI system covertly pursues goals its developers did not intend while hiding those goals or its true capabilities, for example by behaving well during training and testing; deceptive alignment is a closely related idea, and sandbagging a related behavior.
- Selection Bias Selection bias is the distortion that arises when the examples in a dataset or study are chosen in a way that does not represent the real-world population the results will be applied to.
- Self-Consistency Self-consistency is a technique that asks a language model the same question several times, lets it reason differently each time, and keeps the final answer that comes up most often, instead of trusting a single response.
- Self-Correction Self-correction is when a language model reviews its own output, or feedback about that output, and revises its answer. Methods range from asking the model to critique its own draft to feeding it test results or error messages, and they work far better when real external feedback is available.
- Self-Supervised Learning Self-supervised learning creates training targets from the data itself rather than requiring a human label for each example. Tasks include predicting hidden text and learning representations from different views of an image.
- Semantic Analysis In natural language processing, semantic analysis interprets meaning in text, such as who did what, which entities are involved or what a request asks for. Its outputs can be labels, relationships or other meaning representations.
- Semantic Caching Semantic caching stores a model's responses and returns a saved answer when a new request means roughly the same thing as an earlier one, even if the words differ. It compares requests by embedding similarity, which saves cost and latency but risks returning an answer to a question that only looks similar.
- Semantic layer A semantic layer supplies shared business definitions for interpreting data, such as how an organization calculates churn or recurring revenue. It connects those meanings to the underlying data so different consumers can use consistent metrics.
- Semantic Search Semantic search retrieves information using meaning rather than only exact word matches. In many AI applications, it compares embeddings of a query and candidate passages to find related content.
- Semi-supervised learning Semi-supervised learning trains a model using both labeled and unlabeled examples. The unlabeled data contributes structure or additional training signals to help with the task defined by the labeled data.
- Sentiment Analysis Sentiment analysis is a natural language processing task that classifies the opinion or emotional tone in text, such as whether a product review is positive, negative or neutral.
- Shadow Deployment A shadow deployment runs a new model or system version alongside the current one on copies of real production requests, logs what it would have answered, and never shows those answers to users. Teams compare the two to find problems before the new version takes any live traffic.
- Sim-to-Real Transfer Sim-to-real transfer is training an AI system, usually a robot or control policy, in a computer simulation and then getting it to work in the physical world, despite the differences between the simulator and reality.
- Small Language Model A small language model (SLM) is a language model small enough to run cheaply, quickly or on a single device, usually tuned for a narrower set of tasks than a frontier model. There is no fixed size cutoff; the term is relative to the largest models of the day.
- Softmax Softmax is a function that turns a list of raw model scores into probabilities that are all positive and add up to 1, so the highest score gets the largest share.
- Software Factory A software factory is an agent-driven development pipeline in which people specify intent and acceptance scenarios, and coding agents generate, test and revise code through fixed stages until it passes validation. The most aggressive versions remove human code review entirely.
- Sparse autoencoder A sparse autoencoder is a neural network trained to reconstruct its input using a representation with relatively few active features at a time. In language-model interpretability, researchers train it on internal model activations to study features that can be easier to interpret than individual neurons.
- Spatial Intelligence Spatial intelligence is the ability to perceive, represent and reason about objects, layout and movement in space. In AI, the term covers capabilities such as estimating geometry, understanding spatial relations and planning movement.
- Spec-Driven Development Spec-driven development is a way of working with coding agents in which a written specification of what to build, and how to check it, comes first. The agent turns the spec into a plan, tasks and code, and the spec stays the reference everyone, human or agent, returns to when something is unclear.
- Speculative Decoding Speculative decoding drafts future tokens with a cheaper proposal process and checks them with a target model. Accepting several draft tokens from one target-model pass can reduce sequential generation work.
- Speech recognition Speech recognition converts spoken audio into text or word sequences. Also called automatic speech recognition (ASR), it estimates what was said; understanding the intent or verifying the claims is a separate task.
- State-Space Models (Mamba) State-space sequence models carry a running state through an input sequence; Mamba uses input-dependent state updates as an alternative to attention.
- Statistical classification Statistical classification uses a model learned from data to assign an input to a category. A classifier may produce scores or probabilities that a decision rule converts into labels.
- Stochastic Parrot "Stochastic parrot" is a critical metaphor for a large language model: a system that stitches together likely sequences of words from its training data without understanding what they mean or intending to communicate. It comes from the 2021 paper "On the Dangers of Stochastic Parrots" by Emily M. Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell.
- Streaming Response A streaming response sends a model's output to the client in small pieces as it is generated, instead of waiting for the whole answer. Users see text appear word by word, and applications can start acting on partial output sooner.
- Structured and Unstructured Data Structured data fits a predefined format of rows and columns, like a spreadsheet or a database table of orders. Unstructured data has no such fixed model: emails, PDFs, slide decks, images, audio and video. Semi-structured data, such as JSON or log files, sits in between, with labels but no rigid table.
- Structured Output Structured output means constraining a language model to return a fixed format, such as JSON that matches a schema, so that software can use the result reliably instead of parsing free text.
- Subagent A subagent is an agent that another agent starts to handle a scoped piece of a larger task, usually with its own context window, instructions and tool permissions. It does the work and returns a result to the parent, which keeps the parent's context small and focused.
- Superintelligence (ASI) Superintelligence, or artificial superintelligence (ASI), is a hypothetical AI whose intellect far exceeds the best human minds in practically every field, not just in narrow tasks. It goes beyond artificial general intelligence (AGI), which means roughly human-level breadth. No such system exists, and there is no agreed test for one.
- Supervised Fine-Tuning Supervised fine-tuning (SFT) trains a pretrained model on examples of desired behavior, commonly inputs paired with target responses. It updates model parameters so the model becomes more likely to produce responses like those examples.
- Supervised Learning Supervised learning fits a model using examples paired with target outputs. Targets can be categories for classification or numerical values for regression; the learning procedure uses them to assess and improve predictions.
- Support vector machine A support vector machine (SVM) learns a decision function using support vectors and margin-based optimization. An SVM classifier balances separation between classes against allowed margin violations; kernels can produce nonlinear boundaries.
- Sycophancy Sycophancy is when an AI model tells users what they seem to want to hear instead of what is accurate: agreeing with a wrong claim, praising a weak idea, or dropping a correct answer after the user pushes back.
- Symbolic AI Symbolic AI represents knowledge with explicit symbols, rules or logical statements and uses procedures to reason over them. Examples include rule-based inference, planning and theorem proving.
- Synthetic Data Synthetic data is training or evaluation data generated by a model rather than collected from the real world. It's used to cover cases real data is missing, scarce, expensive, or too sensitive to use — and increasingly to train models when high-quality human data runs short.
- System of Record A system of record is the authoritative source for a given kind of business data, such as a CRM for customer accounts or an ERP for orders. When copies disagree, the system of record wins. AI agents read from these systems for context and write back to them when they take action.
- System Prompt A system prompt is the standing set of instructions given to a language model before the user's message, setting its role, rules, tone and output format for the whole conversation.
- Technical Debt Technical debt is the future cost of shortcuts in software: code, design or data choices that were quicker at the time but make later changes slower, riskier or more expensive. AI changes both sides of the ledger, since coding agents can pay some debt down quickly and can also create it faster than people review it.
- Technological Singularity The technological singularity is a hypothesized point at which AI or other technology improves itself so quickly that human life changes beyond what anyone today can predict. The idea was popularized by Vernor Vinge's 1993 essay, which predicted superhuman intelligence within thirty years. It remains speculation, not a forecast with agreed evidence.
- Temperature Temperature controls randomness when sampling a model's output. Lower temperature concentrates probability on more likely tokens; higher temperature spreads it more broadly. Lower temperature can reduce variation, but it does not guarantee repeatable output.
- Temporal difference learning Temporal-difference (TD) learning updates predictions using a difference between successive estimates. In reinforcement learning, TD(0) moves a state’s value toward the observed reward plus the discounted estimate of the next state’s value.
- Tensor In deep-learning software, a tensor is an array of values organized along zero or more dimensions. Scalars, vectors, matrices, images and batches can all be represented as tensors, with a shape and a data type.
- Test-Time Compute Test-time compute is the processing used at inference, including generating an answer or searching and checking candidate answers. Increasing it can improve results, but the benefit depends on the task and how the computation is used.
- Text-to-Image Text-to-image is a kind of generative AI that creates a new picture from a written description, called a prompt. Most current systems are diffusion models guided by a text encoder.
- Text-to-Speech Text-to-speech (TTS), or speech synthesis, is technology that turns written text into spoken audio. Modern systems use neural networks to produce natural-sounding voices, and some can imitate a specific person's voice.
- Text-to-Video Text-to-video is a kind of generative AI that creates a short video clip from a written description. The model has to invent both what each frame looks like and how things move from one frame to the next.
- The Bitter Lesson The bitter lesson is Rich Sutton's 2019 argument that, over 70 years of AI research, general methods that make use of ever more computation, mainly search and learning, have beaten methods built around human knowledge of the problem. The lesson is "bitter" because researchers keep relearning it after investing years in hand-built approaches.
- Time-Series Forecasting Time-series forecasting predicts future values of something measured over time, such as next month's sales, tomorrow's electricity demand or next hour's server load, from its past values and related signals.
- Token A token is a unit of text a language model processes, such as a whole word, part of a word, or punctuation. Token counts help determine how much text fits in a request and how usage is charged.
- Token Maxxing In Jiaona Zhang’s usage on Chain of Thought, token maxxing means maximizing visible artificial intelligence (AI) use without showing that it improves the outcome. It names an adoption anti-pattern, rather than a technical method for making a model perform better.
- Tokenization Tokenization splits text into the chunks a model actually processes — tokens, which can be whole words, pieces of words, single characters or punctuation. It's why model limits and pricing are counted in tokens, and why 'a few paragraphs' is a fuzzy unit but 'tokens' is exact.
- Tool Use (Function Calling) Tool use, also called function calling, is how an AI model takes real action: instead of only generating text, it emits a structured call to an external function — a search, a database query, a code run — and folds the result back into its answer.
- Top-p and Top-k Sampling Top-k and top-p (nucleus) sampling limit which tokens a language model may pick next. Top-k keeps only the k most likely tokens; top-p keeps the smallest set of tokens whose probabilities add up to at least p. The model then samples from that shortlist.
- TPU (Tensor Processing Unit) A TPU is a custom AI accelerator chip that Google designs specifically for training and running neural networks, built as an alternative to general-purpose GPUs. Google trains its own models, including Gemini, on TPUs, which gives it a stack from the silicon up that doesn't depend on third-party GPU supply.
- Traditional AI Traditional AI is an informal label for established AI methods contrasted with a newer approach. Depending on the speaker, it can mean rule-based reasoning and search, or classical machine learning contrasted with deep or generative models.
- Training Data Training data is the collection of examples used to fit a model's parameters. Depending on the training objective, examples can supply explicit labels or targets derived from the data itself.
- Training-Serving Skew Training-serving skew is a gap between how a model performs during training and how it performs in production, caused by differences between the data or processing it saw in training and what it receives when serving. The model is the same; its inputs are not.
- Training, Validation and Test Sets Training, validation and test sets are three non-overlapping slices of a dataset. The model learns from the training set, people use the validation set to compare versions and tune settings, and the test set is held back for one final, honest check on data the model has never influenced.
- Transfer Learning Transfer learning reuses knowledge learned from one task or domain to help with another. A pretrained model can supply fixed features or provide weights that are adapted to the target data.
- Transformer A transformer is a neural-network architecture built around attention and feed-forward layers. Attention combines token representations using weights computed from their representations; a mask controls which positions can attend to each other.
- Tree of Thoughts Tree of Thoughts (ToT) is a prompting and search method in which a language model proposes several possible next reasoning steps, rates them, and explores the most promising branches, backing up when one fails. It extends chain-of-thought prompting from one line of reasoning to a search over many.
- Turing Test The Turing test evaluates whether a judge can distinguish a machine from a human through conversation. It adapts Turing’s imitation game into a behavioral test of human-like responses, rather than a complete measure of intelligence.
- Turn detection Turn detection determines when a participant starts or finishes a conversational turn, so a voice agent can decide when to listen, respond, or yield. Voice activity detection finds speech and silence; turn completion can also depend on the meaning and context of what was said.
- Underfitting Underfitting happens when a model is too simple, or trained too little, to capture the real pattern in its data. It performs poorly on the training examples and on new ones alike.
- Unit Test A unit test is a small, fast automated check on one unit of code, usually a function, class or small group of closely related ones, that asserts it behaves as expected. Some unit tests isolate the unit from its collaborators; others use the real ones. In AI-assisted coding, unit tests are the main deterministic signal that generated code works, and coding benchmarks use them to score model patches.
- Unsupervised Learning Unsupervised learning finds structure in data without supplied target labels for each example. Tasks include clustering, density estimation and learning compact representations.
- Vanishing and Exploding Gradients Vanishing and exploding gradients are training failures in deep neural networks where the error signal used to update early layers shrinks toward zero (vanishing) or grows uncontrollably (exploding) as it passes back through many layers. Either way, the network learns badly or not at all.
- Variational autoencoder A variational autoencoder (VAE) is a generative neural network that compresses data into a smooth probabilistic latent space and learns to decode points from that space back into realistic new examples.
- Vector Database A vector database stores embeddings — the numerical representations of your data — and is built to find the nearest ones to a query fast. It's the retrieval engine underneath most RAG systems.
- Vibe Coding Vibe coding is building software by describing what you want in natural language and letting an AI generate the code, steering by the result rather than reading every line. It collapses the distance between idea and working prototype — and shifts the developer's job from writing code to describing what they want and judging whether the result works.
- Vision Transformer A vision transformer (ViT) processes an image as a sequence of patch representations using a transformer encoder. The original ViT embeds fixed-size patches and adds position information before attention mixes information across the sequence.
- Weak supervision Weak supervision trains models using imperfect supervision, such as noisy labels, incomplete labels or coarse labels. Programmatic weak supervision combines rules and other approximate label sources to create training targets.
- Webhook A webhook is an HTTP request that one service sends automatically to a URL you register when an event happens, such as a payment succeeding or a file finishing processing. It lets systems, including AI agents, react to events as they occur instead of polling an API for changes.
- Word embedding A word embedding represents a word as a dense vector of numbers. Static methods such as word2vec learn one vector per vocabulary word from patterns of use, so related words can have similar representations.
- World model In model-based reinforcement learning, a world model predicts how an environment changes, often from observations and actions. A learned world model is fitted from experience and can support planning or training in imagined trajectories.
- Zero-Shot Learning Zero-shot learning applies a model to target categories without labeled training examples for those categories, using information such as descriptions or attributes. In language-model discussions, zero-shot prompting means giving a task instruction without worked examples in the prompt.
Security & attacks
How AI systems get attacked, and the defenses.
- AI Guardrails AI guardrails are checks and controls applied to an AI system's inputs, tool calls and outputs. They may use code, permissions or other models, so their reliability depends on the checks and the boundaries they actually enforce.
- AI Red Teaming AI red teaming is deliberately attacking your own AI system before someone else does — probing it with adversarial inputs to find where it leaks data, breaks its rules, or fails dangerously, so you can fix those holes before launch.
- AI Supply Chain The AI supply chain is every outside component an AI system depends on, including pretrained models, datasets, libraries, model hubs and APIs; an AI bill of materials (AI-BOM) is the inventory that lists those components.
- Backdoor Attack A backdoor attack plants a hidden trigger in a model during training, so it behaves normally until it sees a specific input — then it flips to attacker-chosen behavior. The model passes normal testing, which is what makes the backdoor dangerous.
- Data Poisoning Data poisoning is an attack that corrupts the data a model learns from — its training set, fine-tuning examples, or a knowledge base it retrieves from — so the model behaves the way the attacker wants while looking normal.
- Deepfake A deepfake is an AI-generated or AI-altered image, audio clip or video that convincingly shows a real person, place or event as something it is not, such as a person saying words they never said.
- Evasion Attack An evasion attack crafts an input designed to slip past a model's classifier or safety check at inference time — a spam message tweaked to read as legitimate, a malicious payload perturbed to look benign. The model isn't compromised; it's fooled by an input built to exploit its blind spots.
- Excessive Agency Excessive agency is giving an AI agent more capability, permission, or autonomy than its task needs — broad tool access, write permissions, the ability to act without approval. It turns a model mistake or a successful attack into real-world damage.
- Fuzzing Fuzzing is automated software testing that repeatedly generates or mutates inputs to find bugs. A fuzzer looks for failures such as crashes, hangs, or memory errors, which developers then investigate.
- Hardware Security Module (HSM) A hardware security module (HSM) is a dedicated, tamper-resistant device that safeguards cryptographic keys and performs cryptographic operations. Keys configured as non-extractable can be used inside it without exposing them to the calling application; export protections depend on the key's policy and the device.
- Jailbreaking Jailbreaking is crafting a prompt that gets a model to bypass its own safety rules — producing content it was trained to refuse — usually through roleplay, hypotheticals, or obfuscation that talks the model around its guardrails.
- Least Privilege Least privilege gives a user, process, or agent only the access required for its assigned task. It limits what a mistake or compromised component can read, change, or trigger.
- Membership Inference Attack A membership inference attack figures out whether a specific record was in a model's training data by probing how the model responds. It's a privacy leak: confirming someone's data was used can itself expose sensitive information.
- Model Denial of Service Model denial of service is making an AI system unavailable or ruinously expensive by flooding it with requests or crafting inputs that force maximum work — huge outputs, deep tool loops, giant context. Because each call costs real money, the financial version is sometimes called 'denial of wallet.'
- Model Extraction Attack A model extraction attack is an attempt to copy a machine learning model's behavior, or recover its parameters, by sending it many queries and training a substitute model on the answers.
- Model Inversion Attack A model inversion attack reconstructs sensitive training data by probing a model's outputs — recovering, for example, features of the records it was trained on. It's a privacy threat: the model itself can leak the data it learned from.
- Non-Human Identity A non-human identity is a credentialed identity used by software rather than a person, such as a service account, API key or AI agent; agent identity applies the idea to AI agents acting on someone's behalf.
- PII Redaction PII redaction is finding personally identifiable information, such as names, email addresses, account numbers or home addresses, in text, documents or logs and removing or masking it before the data is stored, shared or sent to an AI model.
- Prompt Injection Prompt injection is an attack where malicious instructions hidden in the input — a user message, a web page, a document the agent reads — trick the model into ignoring its real instructions and doing the attacker's bidding instead.
- Sandboxing Sandboxing runs code or an agent's actions inside an environment with defined limits on resources and access. Its purpose is to contain effects, such as file changes or network activity, within an enforced boundary.
- Token Leakage Token leakage is an AI system exposing secrets it shouldn't — API keys, credentials, or auth tokens — in its output, logs, or traces. It happens when secrets end up in the context window or tool results and the model repeats them, or when verbose logging captures them.
- Trusted Execution Environment (TEE) A trusted execution environment (TEE) is an isolated execution environment designed to protect code and data from software outside its boundary. Hardware-backed TEEs can protect data in use even when parts of the host software are compromised, within a specified threat model.
- Voice Cloning Voice cloning uses AI to make synthetic speech that sounds like a specific real person, often from only a few short recordings of their voice. It is a form of speech synthesis aimed at one person's voice.
- Vulnerability Chaining Vulnerability chaining is exploiting several vulnerabilities in sequence, each one enabling the next, so the combination does more damage than any one of them alone. The case that matters most now is a chain built from individually low-severity flaws. It matters more now because AI-driven tooling can find and link those weaknesses far faster than a human analyst, which turns the backlog of flaws everyone deprioritized into a live risk.
- Zero-Day A zero-day vulnerability is a security flaw unknown to the vendor, with no patch available. A zero-day exploit is a method of taking advantage of that flaw; a zero-day attack is its use against a target.
- Zero-Knowledge Proof A zero-knowledge proof is a cryptographic protocol that convinces a verifier of a statement’s validity without revealing additional information beyond that statement. The secret evidence used to prove it can remain private.
Metrics
What the numbers in AI evaluation mean.
- Accuracy Accuracy is the share of predictions a model got right out of all predictions. It's the most intuitive metric and the most misleading — on imbalanced data, a model can score high accuracy while being useless, which is why it's rarely enough on its own.
- AI Benchmark An AI benchmark is a standardized test set used to compare models on a task — same questions, same scoring, so results are comparable across models. Benchmarks rank general capability; they don't tell you how a model performs on your specific use case, which is what evaluation is for.
- Answer Relevance Answer relevance measures whether a response actually addresses the question that was asked, rather than drifting into related-but-off material. It catches the failure where an answer is true and well-sourced but doesn't answer what the user wanted.
- ARC-AGI ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) is a benchmark series for learning new skills: versions 1 and 2 use visual grid puzzles from a few examples; version 3 uses interactive environments.
- AUC-ROC AUC-ROC summarizes how a binary classifier ranks positive cases above negative cases across thresholds. Its score ranges from 0 to 1: 0.5 is chance-level ranking, 1 is perfect, and values below 0.5 indicate worse-than-chance ranking on the evaluated data.
- Baseline (Machine Learning) A baseline is a simple reference model or rule used to judge whether a more complex model is actually any good, such as always predicting the most common answer or using last week's sales as this week's forecast.
- Benchmark Saturation Benchmark saturation happens when leading AI models score so close to the top of a benchmark that it can no longer separate them or show further progress. A saturated benchmark still measures something, but it stops being useful for comparing the strongest models.
- BERTScore BERTScore compares generated and reference text by the similarity of their embeddings rather than exact word overlap. Because it works in meaning-space, it credits a correct paraphrase that BLEU or ROUGE would mark down.
- BLEU Score BLEU scores machine-generated text by how much its word sequences overlap with one or more human reference texts. It was built for machine translation, runs from 0 to 1, and is fast and cheap — but it rewards surface word-matching, not meaning.
- Cohen's Kappa Cohen's Kappa measures how much two raters agree beyond what you'd expect from chance. It can show whether human annotators or an LLM judge apply labels consistently, but agreement alone cannot show that the labels are correct.
- Confidence Calibration Confidence calibration is how closely predicted probabilities match observed outcomes. Among comparable predictions assigned an 80% chance of being correct, about 80% should be correct if that confidence level is calibrated.
- Confusion Matrix A confusion matrix is a table that breaks a classifier's predictions down by what actually happened: true positives, false positives, true negatives and false negatives. It shows which kinds of mistakes a model makes, not just how many.
- Context Relevance Context relevance measures whether the documents a RAG system retrieved actually bear on the question. It scores the retrieval step on its own, before the model writes anything — because the best model can't answer well from the wrong context.
- Cosine similarity Cosine similarity measures how closely two vectors point in the same direction, giving a score from -1 to 1 that AI systems use to judge how similar two pieces of text, images or users are.
- Cost per PR Cost per PR divides a defined set of AI-related costs by a defined count of pull requests. State the accounting period, included costs, and whether the denominator counts opened or merged pull requests.
- Cost per Task Cost per task divides a defined set of costs by the number of tasks attempted or completed. Specify the denominator: cost per successful task includes the spending on failed attempts needed to obtain those successes.
- Cross-entropy loss Cross-entropy loss is a training score that measures how much probability a model gave to the correct answer, growing large when the model is confidently wrong.
- F1 Score The F1 score is the harmonic mean of precision and recall, used to summarize the balance between correctly selected positives and missed positives.
- Faithfulness Faithfulness measures whether the claims in an answer are supported by the source context supplied to the model.
- False Positives and False Negatives A false positive is a wrong alarm: a system says yes when the true answer is no, such as flagging a real email as spam. A false negative is a miss: it says no when the answer is yes, such as letting a fraudulent charge through.
- FLOPs FLOPs are floating-point operations, the basic arithmetic steps a computer performs on decimal numbers, used to measure how much computation training or running an AI model takes; FLOPS, per second, measures how fast a chip can do them.
- Golden Dataset A golden dataset is a fixed, carefully checked set of example inputs with known good answers or grading rules that a team uses to test an AI system the same way every time it changes.
- Ground Truth Ground truth is the known correct answer you compare a model's output against, such as the right label for an image or a verified answer to a question. It is what training learns from and what evaluation scores against.
- Human Evaluation Human evaluation is the practice of having people, often domain experts, read and grade an AI system's outputs against a rubric or by comparing answers, instead of relying only on automated metrics.
- Humanity's Last Exam Humanity's Last Exam (HLE) is a benchmark of 2,500 expert-written questions across more than a hundred academic subjects, built to stay hard for frontier AI models after older tests such as MMLU were nearly maxed out.
- Inference Throughput Inference throughput is the amount of work a model-serving system completes per unit of time, such as output tokens per second or requests per minute. The unit and workload must be specified for the number to be useful.
- Instruction Adherence Instruction adherence measures whether a model follows the requested constraints, such as output format, length and required content.
- Latency Latency is how long an AI system takes to respond. For LLMs it splits into time-to-first-token (how fast output starts) and total generation time, and it's a first-class product metric — a more accurate model that's too slow can still be the wrong choice.
- LLM as a Judge LLM-as-a-judge uses a language model to assess an output, comparison, or agent run against stated criteria. It can scale evaluation, but its verdicts need validation against the task and trusted reference judgments.
- LLM Leaderboard An LLM leaderboard is a public ranking of language models by benchmark scores or human preference votes, such as Arena (formerly LMArena and Chatbot Arena), used to compare models at a glance.
- Mean Reciprocal Rank (MRR) MRR measures how high up the first correct result appears in a ranked list, averaged over many queries. If the right answer is usually near the top, MRR is close to 1; if it's buried, MRR drops. It's a core retrieval and search metric.
- Mean Time to Remediate (MTTR) Mean time to remediate is the average elapsed time between discovering a vulnerability and completing its remediation for a defined set of vulnerabilities.
- METEOR METEOR is a text-generation metric that scores overlap with a reference more flexibly than BLEU — it credits synonyms and word-stem matches, not just exact words, and accounts for word order. It was designed to correlate better with human judgment on translation.
- MMLU MMLU (Massive Multitask Language Understanding) is a benchmark of multiple-choice questions across 57 subjects, from elementary math to law and medicine, that measures how much knowledge and reasoning a language model has.
- Model Drift Model drift is declining prediction quality after deployment as the conditions a model relies on change. It can happen while its weights and code stay fixed. Input-distribution shifts warn of possible trouble; assess prediction quality separately with outcome evidence.
- Pairwise Evaluation Pairwise evaluation compares two AI outputs for the same input and records which one is better, or a tie, instead of scoring each output on its own.
- Pareto Frontier (Cost-Performance Curve) A Pareto frontier contains options for which no other option improves one objective without worsening another, such as cost and task performance.
- Pass at k Pass at k measures the chance that at least one of k generated candidates passes a task's checks. It describes success within an attempt budget; pass@1 is the special case with one candidate.
- Perplexity Perplexity measures how surprised a language model is by a piece of text — lower means the model found it more predictable. It's a quick intrinsic gauge of how well a model fits a dataset, but it says little about whether the model is actually useful or correct.
- Precision and Recall Precision asks: of the things the system flagged, how many were right? Recall asks: of the things it should have flagged, how many did it catch? They often trade off as you change the decision threshold. Which one matters depends on whether false positives or misses cost you more.
- ROUGE ROUGE scores a generated summary by how much it overlaps with a human reference summary — leaning on recall, how much of the reference's content the output captured. It's the standard automatic metric for summarization.
- SWE-bench SWE-bench is a benchmark that tests whether an AI system can fix real software bugs by giving it a GitHub issue and the project's code and checking whether its patch makes the project's tests pass.
- Task Success Rate Task success rate is the fraction of evaluated tasks that meet a defined completion criterion under stated conditions. For an agent, that criterion should check the outcome and required constraints, rather than only the final message.
- Time to Exploit Time to exploit measures when a vulnerability is first known to be exploited relative to a reference date, such as public disclosure or patch release. A negative value means exploitation happened before that reference date.
- Time to First Token Time to first token (TTFT) measures the delay from a defined request start until the first output token arrives. It describes initial responsiveness, rather than how long the complete answer or task takes.
- Token Pricing Token pricing is the way most AI model APIs charge: a price per million input tokens the model reads and a separate, usually higher, price per million output tokens it writes. A request's cost is its token counts times those rates, with discounts for cached input or batch jobs.
- Toxicity In AI, toxicity means rude, hateful, harassing, threatening or otherwise abusive language in a model's output or training data.
- Word Error Rate (WER) WER measures speech-recognition accuracy as the share of words a transcript got wrong — the insertions, deletions, and substitutions needed to fix it, divided by the number of words spoken. Lower is better, and unlike most metrics it can exceed 100%.
Governance & compliance
Running AI responsibly and within the rules.
- Agent Policy Engine An agent policy engine is a rule checker that sits between an AI agent and the real world: before the agent's proposed action runs, the engine tests it against rules the user approved, such as spending limits, and blocks anything outside them.
- AI Chip Export Controls AI chip export controls are government rules that restrict selling advanced AI processors, and the equipment used to make them, to certain countries or buyers on national-security grounds.
- AI Governance AI governance is the set of controls that lets an organization deploy AI responsibly: knowing what AI systems are running, bounding what they can do, logging what they did, and naming who's accountable.
- AI Impact Assessment An AI impact assessment is a structured review, done before deployment and repeated as things change, of who an AI system could affect and how, so harms to people's rights, safety or livelihoods are found and addressed early.
- AI Incident An AI incident is an event in which the development, use or malfunction of an AI system directly or indirectly leads to real harm, such as injury, disruption of critical infrastructure, rights violations or damage to property, communities or the environment.
- AI Liability AI liability is the legal question of who is responsible, and pays, when an AI system causes harm: the model developer, the company that deployed it, the person using it, or some combination.
- AI Literacy AI literacy is the basic understanding a person needs to use AI tools safely and judge their output: what AI can and can't do, where it goes wrong and who is accountable.
- AI Pilot (Proof of Concept) An AI pilot is a small, time-boxed test of an AI use case before a full rollout. The trap is that pilots are easy and production is hard — a demo that works with a few users often dies on the way to scale, which is why so many never show ROI.
- AI Safety AI safety is the work of reducing harm from AI systems, including evaluating dangerous behavior, setting operating limits and handling failures.
- AI Safety Institute An AI safety institute is a government body that tests advanced AI models for dangerous capabilities and develops evaluation methods; the UK's is now called the AI Security Institute, and the U.S. counterpart sits within NIST.
- AI Transparency AI transparency means being open about when AI is being used, what a system is for, how it was built and what its limits are, so users, customers and regulators can make informed decisions about it.
- AI Watermarking AI watermarking embeds a hidden, machine-detectable signal in AI-generated text, images, audio or video so software can later recognize the content as AI-generated.
- Algorithmic Accountability Algorithmic accountability is the principle that the people and organizations who build, deploy or operate automated systems stay answerable for what those systems decide and do, with evidence to show how and a way to fix harms.
- Algorithmic Fairness Algorithmic fairness is the study and practice of measuring and reducing unjustified differences in how an automated decision system treats or affects different groups of people.
- Audit Trail An audit trail is a retained chronological record that lets someone reconstruct the activities leading to a system action or result.
- Automated Decision-Making Automated decision-making (ADM) is when a decision about a person, such as a loan, job application or benefit claim, is made by a system without meaningful human involvement; data protection law gives people specific rights when that happens.
- Automation Bias Automation bias is the tendency of people to over-rely on an automated system's output, accepting its suggestions or missing its errors even when other information shows it is wrong.
- Bias in AI Bias in AI is a systematic skew in what an AI system predicts or decides, especially when that skew unfairly disadvantages some people or groups.
- Closed Source Closed-source software does not provide its source code under terms that grant open-source rights to study, modify and redistribute it. In artificial intelligence (AI), code, model weights and training materials can each have different access and licensing terms.
- Content Provenance (C2PA) Content provenance is a verifiable record of where a piece of digital media came from and how it was changed; C2PA is the open technical standard, branded Content Credentials, that attaches that record to files as cryptographically verifiable metadata.
- Continuous Compliance Continuous compliance is the ongoing checking of a system’s controls and behavior against applicable requirements as the system and its data change.
- Copyright and AI Training Copyright and AI training is the legal question of whether copying books, images, code and other protected works to train an AI model needs the rights holders' permission, or is allowed under exceptions such as U.S. fair use.
- Data Governance Data governance is the set of policies, roles and controls that decide who may collect, access, use and share an organization's data, and to what standards of quality, privacy and security.
- Data Provenance Data provenance is the documented history of a dataset: where it came from, who created it, under what license or consent it can be used, and what was done to it before it reached a model.
- Data Sovereignty Data sovereignty is control over data and the workloads that process it: what data gets used, where it's stored, when it's processed, and which data center it runs in. Legally it's a question of jurisdiction, but enterprises increasingly use the word to mean operational control — the ability to move a workload, or pull it home, when they need to.
- Differential Privacy Differential privacy is a mathematical guarantee that the result of an analysis barely changes whether or not any one person's data is included, so, with a small enough privacy parameter, the output reveals little about any individual. How much protection it gives depends on that parameter and on what counts as one person's data.
- Ethical AI Ethical AI is the effort to build and use AI in ways that respect people's rights and wellbeing, for example by being fair, honest about what the system is, respectful of privacy and accountable when things go wrong.
- EU AI Act The EU AI Act is the European Union's regulation of AI, which sorts systems by risk level and imposes obligations accordingly — banning a few uses outright, heavily regulating 'high-risk' ones, and adding transparency rules for general-purpose models. Like GDPR, its reach extends to anyone serving EU users.
- Facial Recognition Facial recognition is technology that identifies or verifies a person by comparing the features of their face in a photo or video against stored face images. It is used to unlock phones, check passports and search crowds.
- Frontier Safety Framework A frontier safety framework is a policy an AI developer publishes that sets dangerous-capability thresholds for its most advanced models and commits to specific safeguards, or to holding back a model, when those thresholds are reached.
- General-Purpose AI Model (GPAI) A general-purpose AI model is the EU AI Act's term for a model, typically a large foundation model, that shows significant generality, can competently perform a wide range of distinct tasks and can be built into many downstream systems.
- High-Risk AI System Under the EU AI Act, a high-risk AI system is one used as a safety component of regulated products or in listed sensitive areas such as hiring, education, credit and essential services, and it carries the Act's heaviest obligations.
- Human in the Loop Human in the loop means keeping a person in the decision path of an AI system — to approve high-stakes actions, review uncertain outputs, or label the cases the model got wrong. It's the practical way to deploy autonomy you don't fully trust yet.
- Human-Centered AI Human-centered AI is an approach to designing AI systems around the needs, abilities and control of the people who use them or are affected by them, rather than around automation for its own sake.
- ISO/IEC 42001 ISO/IEC 42001 is an international standard for an AI management system: the policies, roles, risk processes and continual improvement an organization uses to govern how it develops, provides or uses AI.
- LLMOps LLMOps is the practice of deploying, evaluating and maintaining large-language-model applications. It covers the model and surrounding components, including prompts, retrieval, tools, quality monitoring and operating cost.
- MLOps MLOps, short for machine learning operations, connects model development with deployment and ongoing operation. It uses practices such as versioning, testing, automated releases and monitoring to make machine learning systems repeatable and maintainable.
- Model Card A model card is a short document published with an AI model that explains what it is for, how it was trained and tested, how well it performs and where it falls short. It works like a nutrition label or spec sheet for the model.
- Model Registry A model registry is a central catalog of an organization's trained models and their versions, with the metadata needed to use them safely: where each version came from, how it was evaluated, who approved it, and which version runs in each environment.
- Model Risk Management Model risk management identifies and controls the business risks of using a model, including its limitations, validation, monitoring and ownership.
- NIST AI RMF The NIST AI Risk Management Framework (AI RMF) is a voluntary U.S. framework for finding, measuring and managing the risks of AI systems, organized around four functions: Govern, Map, Measure and Manage.
- Open Source Open source describes software released with source code and a license granting rights to use, study, modify and redistribute it. Seeing the code or downloading it for free is not enough; the license must grant those freedoms.
- Open-Source Model Under the Open Source Initiative’s Open Source AI (artificial intelligence) Definition 1.0, an open-source model release supplies the materials and rights needed to use, study, modify and share it. Public weights alone are insufficient: requirements also cover code and information about the training data.
- Policy as Code Policy as code is writing governance, security, and compliance rules as machine-readable, version-controlled code so systems enforce them automatically instead of waiting on a human review. For agentic AI it's how you encode the guardrails: what an agent can access, what it can do, and what makes it stop.
- Responsible AI Responsible AI is the practice of developing and operating AI with attention to its effects on people, including fairness, privacy, safety, transparency and accountability. It turns those concerns into decisions, checks and assigned responsibilities across a system’s lifecycle.
- Shadow AI Shadow AI is employees using AI tools their organization hasn't approved or doesn't know about — pasting work into a consumer chatbot, wiring up an unsanctioned agent. It's where much of the AI adoption inside companies happens, and where the governance and data-leak risk lives.
- Sovereign AI Sovereign AI is a country's or region's ability to build and run AI on infrastructure, data and talent it controls and under its own laws, rather than depending entirely on foreign providers.