720 terms · 311 full concepts · Updated Nov 4, 2026

AI Glossary, A to Z

Every term you are likely to meet in AI, in one plain sentence. Follow a link to read the full concept and its best free sources.

2

2:4 sparsity
A pattern where at most two of every four consecutive weights are nonzero, which NVIDIA tensor cores accelerate. Pruning and Sparsity

A

Acceptable use policy
A document that states what employees may and may not do with a technology, and the consequences of breaking the rules. AI Policy and Acceptable Use
Accuracy
The share of all predictions that were correct; weak when classes are imbalanced. Evaluation Metrics & Cross-Validation
Activation
The output of a layer that the backward pass needs in order to compute that layer's gradients. Gradient Checkpointing and Memory-Efficient Training
AdaGrad
An optimizer that gives each weight its own learning rate, shrinking it for weights that have already changed a lot. Optimizers: SGD, Momentum & Adam
Adoption rate
The share of eligible people who actually use a tool regularly, usually the first number a rollout is judged on. AI Adoption and Change Management
Age assurance
Methods a service uses to estimate or verify a user's age, from self-declaration to ID checks. Children, Teens and AI
Agent assist
AI that works beside a human agent, suggesting replies and pulling up answers, rather than talking to the customer directly. AI in Sales and Customer Support
Agent loop
The cycle of think, act, observe and repeat that drives an agent. Agentic Workflows & Patterns
Agent verification
A check that the agent placing an order is registered and acting for a real, consenting buyer. Agentic Commerce and Payments
Agent-ready checkout
A merchant endpoint that accepts a cart and payment from an agent over a protocol instead of a web form. Agentic Commerce and Payments
Agentic token
A payment credential issued to a registered agent in place of the real card number, with limits attached. Agentic Commerce and Payments
AI Action Plan
The July 2025 White House plan that sets federal priorities for AI, written after EO 14179 ordered it. US AI Policy and Executive Orders
AI companion
A chatbot designed for ongoing personal conversation, friendship or emotional support. Children, Teens and AI
AI copilot
An assistant built into a tool that suggests work while a person stays in control. Enterprise AI
AI Future Ready
MIT CISR's fourth stage, where AI is embedded in how the enterprise runs and creates value at scale. AI Maturity Models
AI hazard
An event or condition that could plausibly lead to an AI incident but has not yet caused harm. AI Incidents and Incident Databases
AI incident
An event in which the development or use of an AI system led to harm to people, property or society. AI Incidents and Incident Databases
AI literacy
Knowing enough about AI to use it well and judge its output. AI Ethics
AI Overviews
Google's AI-written summaries shown at the top of search results. AI in Marketing & Search
AI race
Competitive pressure that pushes companies or countries to deploy systems before they are shown to be safe. Existential Risk from AI
AI slop
Low-effort, mass-produced AI content that adds little value. Generative AI
AI winter
A period when AI funding and interest collapsed after hype outran results. History of AI
Airflow
A widely used tool for scheduling and monitoring data pipelines. Data Pipelines & ETL
Algorithmic impact assessment
A review, done before deployment, of who an AI system affects and how its harms will be found and reduced. Evaluating AI Vendors and Procurement
Alpha-beta pruning
Skipping branches of a game tree that cannot change the decision. Search Algorithms (A*, Minimax)
AlphaFold
DeepMind's system that predicts protein structures, recognized by the 2024 Nobel Prize in Chemistry. AI-Driven Drug Discovery
AlphaGo
DeepMind's system that beat a world Go champion in 2016 using deep learning and search. Reinforcement Learning (RL)
Ambiguous question
A request that maps to more than one valid query, such as which date column "last quarter" refers to. Text-to-SQL and Natural Language Querying
Answer engine optimization
Making content easy for AI answers to find and cite. AI in Marketing & Search
Answer relevance
A RAG metric asking whether an answer actually addresses the question. Retrieval-Augmented Generation (RAG)
Ant colony optimization
A method inspired by ants laying trails, used for routing problems. Evolutionary & Swarm Optimization
Anti-curriculum
Hard-to-easy ordering, sometimes tested as a control and occasionally found to help. Curriculum Learning
Any-to-any
A model whose inputs and outputs can each be any mix of text, image, audio or video. Any-to-Any (Omni) Models
Apache Kafka
A platform for moving streams of events between systems in real time. Data Pipelines & ETL
Apache Spark
A framework for processing very large datasets across many machines. Data Pipelines & ETL
Approval gate
A specific action, such as sending money or deleting data, that always pauses for a person's yes. Levels of Agent Autonomy
Approximate nearest neighbor
Finding vectors that are probably closest, much faster than checking them all. Vector Databases
Attention mask
A marker that tells a model which tokens to ignore, such as padding. Self-Attention
Attention sink
The first few tokens that a model learns to lean on heavily; keeping them makes windowed attention stable over long streams. GQA, MQA, Multi-head Latent and Sliding Window Attention
AUC
The area under the ROC curve: how well a model ranks positives above negatives. Evaluation Metrics & Cross-Validation
Augmentation
AI that helps a person do a task better or faster, as opposed to automation, which does the task instead of the person. AI and the Future of Work
AutoGen
A Microsoft framework for building conversations among multiple agents. Multi-Agent Systems & Orchestration
Automated decision-making
A decision made about a person by a system with no meaningful human involvement. Data Protection Law and AI (GDPR and beyond)
Automatic curriculum learning
Methods that pick the next task or environment for an RL agent based on its learning progress. Curriculum Learning
Automatic prompt optimization
Using a program or another model to test and improve prompts. Prompt Engineering
Automation bias
The tendency to over-trust a system's output and stop checking it. Human in the Loop
Autonomy level
A rung on a scale that says how many decisions the system makes without a person confirming them. Levels of Agent Autonomy
Autoregressive
Generating one token at a time, each conditioned on everything before it, the way GPT-style models work. Diffusion Language Models

B

Backdoor
A hidden trigger placed in a model that causes bad behavior on command. Data Poisoning
Backward chaining
Reasoning from a goal back to the facts that would prove it. Knowledge Representation & Reasoning
Bagging
Training models on random samples of the data and averaging their results. Ensemble Learning
Barge-in
When the user starts talking over the assistant and it has to stop and listen. Speech-to-Speech and Realtime Voice Models
Baseline
The measurement taken before a tool is introduced, without which no change can be attributed to it. Measuring AI Value and ROI
Batch API
A way to submit many requests to be processed later, often at a lower price. Batch vs Online Inference
Batch normalization
Rescaling a layer's activations within each batch to make training steadier. Normalization Layers (BatchNorm to RMSNorm)
Batch size
How many examples a model sees before it updates its weights once. Optimizers: SGD, Momentum & Adam
Behavior cloning
Supervised imitation of the dataset's actions; the simplest offline baseline. Offline Reinforcement Learning
Behavior policy
The policy, often a human or an older system, that produced the logged data. Offline Reinforcement Learning
Bellman equation
The recursive rule that links a state's value to the values of the states after it. Markov Decision Process (MDP)
Benchmark saturation
The point where top models score near the ceiling and the test stops telling them apart. ARC-AGI and Reasoning Benchmarks
Benign overfitting
Fitting training data perfectly, noise included, while still generalizing well. Double Descent
Best-of-N
Sampling N answers and keeping the one the reward model scores highest. Reward Models and Process Reward Models
Bi-encoder
A retriever that compresses each text into one vector so documents can be scored by a single dot product. ColBERT and Late Interaction Retrieval
Bias term
A learned constant added inside a neuron, so its output need not pass through zero. Neural Network
Binning
Grouping a continuous number into ranges, such as age bands. Feature Engineering
Black-box testing
Testing a model only through its inputs and outputs. Red Teaming AI Systems
Blue-green deployment
Keeping two live environments so that traffic can switch instantly and roll back. Model Deployment & Serving
BM25
A classic keyword ranking formula that is still strong in search. Hybrid Search & Reranking
Boost
Improve a vendor's model with your own data, for example through retrieval-augmented generation or fine-tuning. Build, Buy or Partner
Bounding box
A rectangle that marks where an object is in an image. Object Detection
Bradley-Terry model
The statistical model that turns pairwise preference judgments into a scalar score per response. Reward Models and Process Reward Models
Build
Create, run and maintain a custom AI system in-house for maximum control at the highest cost and effort. Build, Buy or Partner
Business case
The written argument for an AI investment that compares expected value with cost, risk and the alternatives. AI Strategy
Buy
Adopt a ready-made AI product that the vendor runs and maintains, trading control for speed. Build, Buy or Partner
Byte-pair encoding
A common tokenizer method that repeatedly merges frequent character pairs. Tokenization
Bytes per token
A compression ratio for a tokenizer: how many bytes of text each token represents on average. Token Density and Information Density

C

Cache hit
A request answered from the cache, so no model call is made. Semantic Caching
Calibrated trust
A state where users rely on the AI about as much as its real accuracy deserves, neither too much nor too little. Designing AI Products (Human-AI Interaction)
Canary release
Sending a new version to a small share of traffic first. Model Deployment & Serving
Candidate generation
The first stage of a recommender, which narrows millions of items to a few hundred. Recommendation Systems
Capability assessment
A structured review of people, data, technology and process strengths that places a firm on the model. AI Maturity Models
Capability Maturity Model
The 1980s software-process model whose five-level structure most later maturity models copy. AI Maturity Models
Cascaded pipeline
The older design of speech recognition, then a text model, then text-to-speech, each stage adding delay and losing tone. Speech-to-Speech and Realtime Voice Models
Case-based reasoning
Solving a new problem by adapting the solution to a similar past one. Symbolic AI & Expert Systems
CASP
The blind contest held every two years where AlphaFold 2 showed near-experimental accuracy in 2020. AI for Science
Catastrophic risk
A risk of harm on a societal scale, which may or may not be existential. Existential Risk from AI
Causal mask
A mask that stops a language model from looking at future tokens during training. GPT (Generative Pre-trained Transformer)
Censoring
A record where the event had not occurred by the last observation, so only a lower bound on its time is known. Survival Analysis
Center of excellence
A dedicated team that concentrates expertise, sets standards and helps the rest of the organization adopt a capability. AI Center of Excellence and Operating Model
Central limit theorem
The result that averages of many samples tend toward a bell curve. Statistics for ML
Chain of command
The ordered list of who a model obeys when instructions conflict, from the lab's own rules down to user requests. Model Specs and Behavior Policies
Chain rule
The calculus rule for differentiating nested functions, which is what backpropagation applies. Calculus for ML
Change data capture
Tracking changes in a database so that they can flow to other systems. Data Pipelines & ETL
Change management
The planned way an organization moves people from the old way of working to a new one, including communication, training, and support. AI Adoption and Change Management
ChatGPT
OpenAI's chat assistant, whose 2022 launch took generative AI mainstream. Conversational AI & Chatbots
Checkpoint
A saved copy of a model's weights at a point in training. Experiment Tracking
Checkpoint segment
A block of layers whose input is stored while everything inside the block is recomputed on demand. Gradient Checkpointing and Memory-Efficient Training
Children's code
The UK Age Appropriate Design Code, a set of standards for online services likely to be used by children. Children, Teens and AI
Chinese room
John Searle's argument that manipulating symbols is not the same as understanding. Narrow AI vs Strong AI
Chroma
A lightweight open source vector database popular for prototypes. Vector Databases
CI/CD
Automated pipelines that test and release code, extended in MLOps to models and data. MLOps (ML Operations)
Circuit formation
The phase where a general algorithm gradually forms inside the weights while the memorized solution is still in use. Grokking
Class imbalance
When one label is far rarer than others, so accuracy alone becomes misleading. Evaluation Metrics & Cross-Validation
Classifier-free guidance
A setting that controls how strongly an image generator follows the prompt. Diffusion Models
Claude
Anthropic's family of language models and assistant. Large Language Models (LLMs)
Claude Code
Anthropic's coding agent that works in a terminal or editor. Agentic Coding
Cleanup
The phase where weight decay removes the memorized solution, which is when test accuracy jumps. Grokking
Coherence
A score of whether a topic's top words tend to appear together in real text, used to judge topic quality. Topic Density and Topical Coverage
Collaborative filtering
Recommending items based on what similar users liked. Recommendation Systems
Common Crawl
A nonprofit's free archive of web crawl snapshots that most open pretraining corpora start from. Pretraining Data and Data Mixtures
Compaction
Summarizing an agent's conversation so far into a shorter record and continuing from that. Prompt and Context Compression
Compound AI system
A product built from several models, tools and retrieval steps rather than one model. The AI Stack
Compute governance
Regulating AI by tracking and controlling access to large amounts of computing power. AI Regulation & Governance
Confidence display
Showing how sure the system is, through scores, wording or ranking, so users weigh the output properly. Designing AI Products (Human-AI Interaction)
Confidence interval
A range that reflects the uncertainty around an estimate from a sample. Statistics for ML
Confident learning
A method that uses a model's predicted probabilities to estimate which labels in a dataset are wrong. Data-Centric AI
Confusion matrix
A table that counts correct and incorrect predictions for each class. Evaluation Metrics & Cross-Validation
Connectionism
The approach of modeling thinking through networks of simple units. History of AI
Conservative Q-Learning
A method that trains the Q-function to give low values to actions outside the dataset so the policy stays close to the data. Offline Reinforcement Learning
Constitution
Anthropic's natural-language statement of the values and priorities Claude is trained to hold. Model Specs and Behavior Policies
Constraint satisfaction
Finding values that meet all of a problem's rules, as in scheduling. Search Algorithms (A*, Minimax)
Container
A package holding software and its dependencies so that it runs the same anywhere. Model Deployment & Serving
Containment rate
The share of conversations a bot finishes without handing off to a person. AI in Sales and Customer Support
Contamination
Test or benchmark text that leaked into the training set and inflates scores. Data Deduplication and Pruning
Content moderation
Screening text or images for harmful material, by people or by models. AI Guardrails
Content provenance
Metadata or signatures that record where a piece of media came from and how it was edited. AI, Elections and Information Integrity
Context compaction
Summarizing earlier conversation so that an agent can keep working within its context window. Context Engineering
Context precision
A RAG metric asking how much of what was retrieved was truly useful. Retrieval-Augmented Generation (RAG)
Context rot
The drop in answer quality as a context window fills with stale or irrelevant tokens. Prompt and Context Compression
Context window
The maximum number of tokens a model can take in at once, including the prompt and its own output. Token Density and Information Density
Continuous batching
Adding and removing requests from a running batch to use the GPU fully. Inference Optimization
Controller
The organization that decides why and how personal data is processed, and carries the legal duties. Data Protection Law and AI (GDPR and beyond)
ControlNet
A way to guide image generation using a sketch, pose or depth map. AI Image Editing & Inpainting
Convex function
A bowl-shaped function with one lowest point, which makes optimization much easier. Gradient Descent
Convex hull of stability
The test for whether a predicted crystal is energetically stable relative to known compounds. AI for Climate, Weather and Materials
Convolutional layer
A layer that slides small filters across an image to detect local patterns. Convolutional Neural Network (CNN)
COPPA
The US Children's Online Privacy Protection Act, which governs collecting personal data from children under 13. Children, Teens and AI
Core knowledge priors
The basic concepts the puzzles assume every person has, such as objects, counting and symmetry. ARC-AGI and Reasoning Benchmarks
Core vs context
A test that asks whether a capability differentiates the business (build it) or merely supports it (buy it). Build, Buy or Partner
Coreset
A small subset of a dataset chosen so that training on it gives nearly the same model as training on everything. Data Hardness and Example Difficulty
Corrective RAG
A method that falls back to a web search or a new query when retrieved passages are judged weak. Agentic RAG
Corrigibility
A system's willingness to accept correction and being shut down. AI Alignment
Cosine similarity
A score of how alike two vectors are by the angle between them; common for comparing embeddings. Embeddings
Counterfactual explanation
Showing the smallest change to an input that would flip a decision. Explainable AI (XAI)
Cox proportional hazards model
A regression that multiplies a baseline hazard by a factor depending on covariates, without assuming a shape for the baseline. Survival Analysis
CrewAI
A framework for organizing agents into role-based teams. Multi-Agent Systems & Orchestration
Cross-attention
Attention in which one sequence reads from another, such as a decoder reading the encoder. Encoder-Decoder Models
Cross-encoder
A model that scores a query and a document together, used for reranking. Hybrid Search & Reranking
Cross-entropy
A measure of how far predicted probabilities are from the true labels; the standard classification loss. Loss Function
CTC
A training method that aligns audio with text without needing exact timing. Speech Recognition (ASR) & Whisper
CUDA
NVIDIA's platform for programming GPUs, a major reason for its lead in AI. GPU & TPU Computing
Curriculum
A schedule that controls which training examples or tasks the model sees at each stage. Curriculum Learning
Curse of dimensionality
As features multiply, data becomes sparse and distance loses its meaning. Dimensionality Reduction
Cursor
A code editor built around AI assistance and agents. Agentic Coding

D

DALL-E
OpenAI's text-to-image model family. Image Generation
Data card
Google's structured summary of a dataset's facts and decisions, built with the Data Cards Playbook. Datasheets, Data Cards and Model Cards
Data cascade
A compounding downstream failure caused by a data problem that was skipped early in a project. Data-Centric AI
Data catalog
An inventory that helps people find and understand an organization's data. AI Governance
Data classification
Labeling data by sensitivity, such as public, internal, or confidential, so rules can say what may go into which tool. AI Policy and Acceptable Use
Data governance
The rules, roles and controls that decide who may use which data, for what, and how its quality is kept. Data Strategy for AI
Data iteration
Repeating error analysis, data fixes and retraining as the main development loop instead of tuning the model. Data-Centric AI
Data lake
Cheap storage that holds raw data of any type. Data Pipelines & ETL
Data lineage
A record of where data came from and how it changed along the way. Data Pipelines & ETL
Data map
A plot of each training example by the model's mean confidence and variability across epochs, separating easy, ambiguous and hard regions. Data Hardness and Example Difficulty
Data mixture
The proportion of each source, such as web, code or books, in a pretraining corpus. Pretraining Data and Data Mixtures
Data monetization
Turning data into measurable financial value, whether by improving operations, enriching products or selling access. Data Strategy for AI
Data parallelism
Training with the same model on many machines, each handling different data. Distributed Training
Data processing agreement
The contract terms that say what the vendor may do with your data, including whether it can train on it. Evaluating AI Vendors and Procurement
Data product
A dataset or feed treated like a product, with an owner, documentation, quality targets and consumers. Data Strategy for AI
Data quality
How accurate, complete, timely and consistent a dataset is relative to what the model or decision needs. Data Strategy for AI
Data validation
Automated checks that a dataset matches expected types, ranges, counts and distributions before it is used. Data Quality for AI
Data wall
The point at which usable human-written text runs out for larger training runs. Pretraining Data and Data Mixtures
Data warehouse
A database organized for analytics on clean, structured data. Data Pipelines & ETL
Datasheet
A dataset document answering fixed questions about motivation, composition, collection, cleaning, uses and maintenance. Datasheets, Data Cards and Model Cards
DBSCAN
A clustering method that finds dense regions of any shape and marks the rest as noise. Clustering
Debate
A protocol in which two AI systems argue opposite sides of a question and a human judges the transcript. Scalable Oversight
Deceptive AI election content
AI-generated material that fakes a candidate or official, or misleads voters about when, where or how to vote. AI, Elections and Information Integrity
Decision boundary
The line or surface where a classifier switches from one label to another. Classification vs Regression
Deep Blue
IBM's chess computer, which beat world champion Garry Kasparov in 1997. History of AI
Deep research
An agent mode that searches and reads many sources, then writes a cited report. Agentic AI
DeepEval
An open source testing framework for LLM applications. LLM Evaluation & Benchmarks
Deepfake
Synthetic audio, image or video that shows a real person saying or doing something they did not. AI, Elections and Information Integrity
DeepSeek
A Chinese AI lab known for efficient open-weight reasoning models. Open-Weights & Small Language Models
Default behavior
What the model does when nobody has asked for something different, such as its tone or how it handles ambiguity. Model Specs and Behavior Policies
Delegated payment
The buyer's payment method is handed to the agent in a scoped, revocable form rather than as raw card details. Agentic Commerce and Payments
Demand forecasting
Predicting how much of each product customers will buy, by location and time, to decide what to stock. AI in Retail, Supply Chain and Manufacturing
Demographic parity
A fairness test asking whether groups receive positive outcomes at equal rates. AI Bias & Fairness
Denoising step
One pass in which the model predicts cleaner tokens for the whole sequence; sampling runs several of these. Diffusion Language Models
Dense vector
An embedding in which most values are non-zero and capture meaning. Semantic Search
Depth estimation
Working out how far away things are from images. Computer Vision
Dirichlet prior
The assumption in LDA about how spread out topics are across documents and words. Topic Density and Topical Coverage
Disaggregated evaluation
Reporting results separately for demographic or domain groups instead of one average. Datasheets, Data Cards and Model Cards
Discount factor
How much an agent values future reward compared with reward now. Markov Decision Process (MDP)
Displacement
Jobs lost because a technology does the work; usually measured against jobs created elsewhere. AI and the Future of Work
Distributional shift
The gap between the states and actions in the dataset and the ones a new policy would visit. Offline Reinforcement Learning
Docker
The most common tool for building and running containers. Model Deployment & Serving
Domain randomization
Varying simulator parameters during training so a policy transfers to the real world; widening the ranges gradually is a curriculum. Curriculum Learning
Domain reweighting
Adjusting mixture proportions, often with a small proxy model, to improve a larger model's results. Pretraining Data and Data Mixtures
Dot product
A single number that measures how much two vectors point the same way. Linear Algebra for ML
DPIA
A data protection impact assessment, a written review of risks required before high-risk processing. Data Protection Law and AI (GDPR and beyond)
Due diligence
The structured check of a vendor's claims, finances, security and model evidence before signing a contract. Evaluating AI Vendors and Procurement
Duty of competence
The professional rule that lawyers must understand the tools they use well enough to use them properly. AI in Law and Legal Work
DVC
Data Version Control: version control for datasets and models alongside code. Model Versioning & Registry

E

E-discovery
Finding and reviewing electronic documents for litigation, a task AI has helped with for over a decade. AI in Law and Legal Work
Early stopping
Ending training when results on validation data stop improving, to limit overfitting. Regularization
Effective model complexity
Nakkiran et al.'s measure: the largest number of samples a training procedure can fit to near-zero error. Double Descent
Eigenvalue
A number that says how much a matrix stretches a particular direction; central to PCA. Linear Algebra for ML
EL2N
The norm of the error between a model's early predicted probabilities and the true label, used to rank example importance. Data Hardness and Example Difficulty
Elastic net
A regularization method that combines the L1 and L2 penalties. Regularization
ELIZA
A 1960s program that imitated a therapist using simple pattern matching. History of AI
Embedding layer
The first layer of a language model, which maps each token to a learned vector. Embeddings
Embedding model
A model that turns text or images into vectors for search and comparison. Embeddings
Ensemble forecast
Many forecasts run from slightly different starting states to show the range of possible outcomes. AI for Climate, Weather and Materials
Episodic memory
An agent's record of specific past events and interactions. Agent Memory, Planning & Reflection
Epoch
One full pass of the training data through a model. Gradient Descent
Epoch-wise double descent
The same rise-then-fall pattern seen over training time for a fixed, large model. Double Descent
Epsilon-greedy
Taking the best-known action most of the time and a random one occasionally. Q-Learning & Deep Q-Networks (DQN)
Equalized odds
A fairness test asking whether error rates match across groups. AI Bias & Fairness
Error analysis
Sorting a model's mistakes into groups to find which slice of data to fix or collect next. Data-Centric AI
Escalation
Passing a conversation from the AI to a human when the issue is complex, sensitive, or outside policy. AI in Sales and Customer Support
ETL versus ELT
Whether data is transformed before loading into storage, or after. Data Pipelines & ETL
EU AI Act
The European Union's law that regulates AI systems by level of risk. AI Regulation & Governance
Euclidean distance
The straight-line distance between two points. k-Nearest Neighbors (kNN)
Evaluation set
A fixed collection of test inputs with expected outputs used to measure whether an AI feature is good enough to ship. AI Product Management
Evasion attack
Changing an input slightly at prediction time so that a model gets it wrong. Adversarial Machine Learning
Exact-match cache
A cache that only returns an answer when the new request is byte for byte identical to a stored one. Semantic Caching
Execution accuracy
The share of generated queries whose result set matches the result of the reference query. Text-to-SQL and Natural Language Querying
Executive order
A directive from the President to federal agencies; it is not a law passed by Congress and a later President can revoke it. US AI Policy and Executive Orders
Executive sponsor
A senior leader who owns an AI initiative, secures its budget and clears organizational obstacles. AI Strategy
Existential risk
A risk whose bad outcome would wipe out Earth-originating intelligent life or permanently and drastically curtail its potential. Existential Risk from AI
Exit clause
Contract terms that let the buyer leave, take its data and models, and switch suppliers without penalty. Evaluating AI Vendors and Procurement
Expectation
A declared, testable statement about data, such as a column having no nulls, used by tools like Great Expectations. Data Quality for AI
Experiment and Prepare
MIT CISR's first stage, where a firm explores AI with isolated experiments and no structured program. AI Maturity Models
Explainability
Showing a user enough about why the system produced an output for them to judge whether to act on it. Designing AI Products (Human-AI Interaction)
Exploding gradient
When gradients grow through many layers until training becomes unstable. Recurrent Neural Network (RNN)
Exploration versus exploitation
The tradeoff between trying new actions and repeating what already works. Reinforcement Learning (RL)

F

F1 score
A single number that balances precision and recall. Evaluation Metrics & Cross-Validation
Fair use
A US doctrine that allows some unlicensed uses of copyrighted work, judged on purpose, nature, amount, and market effect. AI, Copyright and Training Data
FAISS
Meta's open source library for fast similarity search. Vector Databases
Faithfulness
A RAG metric asking whether an answer is supported by the retrieved context. Retrieval-Augmented Generation (RAG)
False hit
A cached answer served for a question that looked similar but needed a different reply. Semantic Caching
False negative
A real positive case that the model missed. Evaluation Metrics & Cross-Validation
False positive
A case the model flagged that was actually negative: a false alarm. Evaluation Metrics & Cross-Validation
FastAPI
A Python framework commonly used to serve models as web APIs. Model Deployment & Serving
Feature scaling
Putting numeric features on similar ranges so that no one dominates by size alone. Feature Engineering
Federated model
An arrangement where business units own their own AI teams and budgets under shared organization-wide rules. AI Center of Excellence and Operating Model
Feedback loop
Ways for users to correct or rate outputs so the product and its model improve over time. Designing AI Products (Human-AI Interaction)
Feedforward network
A network in which information flows one way, with no loops. Multilayer Perceptron (MLP)
Few-shot examples
Question and query pairs placed in the prompt to show the model the house style of the database. Text-to-SQL and Natural Language Querying
Few-shot prompting
Including a few worked examples in the prompt to show the desired pattern. Prompt Engineering
First-order logic
A formal language for statements about objects and their relationships. Knowledge Representation & Reasoning
Fluid intelligence
The ability to solve problems you have never seen before, as opposed to recalling practiced skills. ARC-AGI and Reasoning Benchmarks
Forgetting event
A moment during training when an example that was classified correctly becomes misclassified again. Data Hardness and Example Difficulty
Forward chaining
Reasoning from known facts toward conclusions. Knowledge Representation & Reasoning
Frankenmerge
Stacking layers from different models into a deeper network, a merge that changes the architecture rather than averaging weights. Model Merging
Frequency penalty
A setting that discourages tokens in proportion to how often they have appeared. Text Generation & Sampling
Frontier developer
A company that trains the largest, most capable models, the group California's SB 53 targets. US AI Policy and Executive Orders
Frontier firm
Microsoft's term for a company that runs on human and AI agent teams rather than adding AI to existing roles. AI Adoption and Change Management
Full autonomy
An agent that writes and runs its own plans with no person able to stop a given step. Levels of Agent Autonomy
Full duplex
Both sides can speak and listen at the same time, as people do, rather than strict turns. Voice Agents
Full-duplex
Both sides can talk at once, so the model can be interrupted and can respond with small cues while it listens. Speech-to-Speech and Realtime Voice Models

G

Gemini
Google DeepMind's family of multimodal language models. Large Language Models (LLMs)
Gemma
Google's family of open-weight models. Open-Weights & Small Language Models
Generative engine optimization
Improving how a brand appears in AI-written answers. AI in Marketing & Search
Genetic algorithm
Evolving solutions through selection, crossover and mutation. Evolutionary & Swarm Optimization
GGUF
A file format for running quantized models on ordinary computers. Model Quantization & Compression
Gist tokens
A handful of learned tokens a model is trained to attend to in place of a full instruction. Prompt and Context Compression
GitHub Copilot
GitHub's AI pair programmer, built into common editors. AI Code Generation
Golden dataset
A hand-checked set of inputs and ideal answers used to test a system. LLM Evaluation & Benchmarks
Goodhart's law
When a measure becomes a target, it stops being a good measure. Reward Design & Reward Hacking
Govern, Map, Measure, Manage
The four functions of the NIST AI Risk Management Framework that organize how an organization handles AI risk. AI Policy and Acceptable Use
Graceful failure
Designing so that when the AI is wrong or unsure, the user can see it, recover quickly and keep working. Designing AI Products (Human-AI Interaction)
Gradient boosting
Building trees in sequence, each one correcting the errors of those before it. Boosting (XGBoost, LightGBM)
Gradient clipping
Capping the size of gradients so that one bad batch cannot blow up training. Backpropagation
Gradio
A Python library for building quick web demos of models. Model Deployment & Serving
Grafana
A popular tool for building monitoring dashboards. ML Monitoring & Observability
Greedy decoding
Always picking the single most likely next token. Text Generation & Sampling
Grounding
Connecting a model's answer to verifiable sources or data. Retrieval-Augmented Generation (RAG)
Grouped-query attention
A middle ground where each group of query heads shares one key-value head. GQA, MQA, Multi-head Latent and Sliding Window Attention
Guardrail metric
A number such as latency, error rate or revenue that a new model is not allowed to make worse. Offline and Online Evaluation
Guardrails
Rules and filters around a model that block unsafe, off-topic or policy-violating outputs. AI Product Management

H

Hard negative
A wrong candidate that the model currently scores almost as high as the right one. Hard Negative Mining
Hazard function
The instantaneous event rate at time t among those still at risk. Survival Analysis
HBM
High-bandwidth memory stacked next to AI chips, and often a supply bottleneck. GPU & TPU Computing
Helicone
An open source observability tool that sits between an app and its model provider. LLMOps
Hessian
The matrix of second derivatives, which describes the curvature of a loss surface. Calculus for ML
Heuristic
A practical rule of thumb that guides search toward good answers without guarantees. Search Algorithms (A*, Minimax)
HHH
Helpful, honest and harmless: a common shorthand for the goals of assistant training. Constitutional AI & RLAIF
Hidden layer
Any layer between a network's input and output. Multilayer Perceptron (MLP)
Hierarchical clustering
Building a tree of clusters by repeatedly merging or splitting groups. Clustering
Hill climbing
Repeatedly moving to a better neighboring solution until none is better. Search Algorithms (A*, Minimax)
Hinge loss
The loss behind support vector machines, which rewards a clear margin between classes. Support Vector Machines (SVM)
Historical bias
Unfairness already present in the world that is carried into training data. AI Bias & Fairness
HNSW
A graph-based index that makes vector search fast at large scale. Vector Databases
Holdout data
Any data deliberately kept out of training so it can be used to check the model. Evaluation Metrics & Cross-Validation
Holdout set
Data kept out of training so a model can be scored on examples it has not seen. Offline and Online Evaluation
Hub-and-spoke
An operating model where a central hub owns platforms, standards and governance while teams in business units deliver the work. AI Center of Excellence and Operating Model
Hugging Face Hub
A public site for sharing models, datasets and demos. Open-Weights & Small Language Models
Human authorship
The requirement that a person made the creative choices before a work can be copyrighted. AI, Copyright and Training Data
Human evaluation
People scoring outputs directly: slow and costly, but often the most trustworthy check. LLM Evaluation & Benchmarks
Human-in-the-loop
A design where a person reviews, approves or corrects AI output before it takes effect. AI Product Management
HyDE
Searching with a hypothetical answer generated by a model rather than the raw question. Retrieval-Augmented Generation (RAG)
Hyperparameter
A setting chosen before training, such as the learning rate. Hyperparameter Tuning

I

IID data
Examples that are independent and drawn from the same distribution, an assumption behind most ML. Statistics for ML
ImageNet
The large labeled image dataset whose 2012 contest result set off the deep learning boom. History of AI
Imputation
Filling in missing values with an estimate such as the median. Exploratory Data Analysis (EDA)
In-batch negative
Using the other examples' positives in the same batch as negatives, which is cheap but mostly easy. Hard Negative Mining
Incident database
A public, searchable collection of incident reports, usually built from news and official records. AI Incidents and Incident Databases
Inference engine
The part of an expert system that applies rules to facts to reach conclusions. Symbolic AI & Expert Systems
Influence operation
A coordinated campaign, often covert, that uses fake accounts and content to shift opinion. AI, Elections and Information Integrity
Information density
How much new, useful content a passage carries per token or per word. Token Density and Information Density
Infrastructure as code
Defining servers and services in version-controlled files, using tools like Terraform. MLOps (ML Operations)
Inner monologue
A technique where the model predicts text tokens alongside its audio tokens to keep what it says coherent. Speech-to-Speech and Realtime Voice Models
Input IDs
The numbers that stand for tokens when text is given to a model. Tokenization
Instance segmentation
Outlining each separate object, not only each class of pixel. Semantic Segmentation
Instrumental convergence
The idea that many goals lead an agent toward the same subgoals, such as acquiring resources. AI Alignment
Intended use
The tasks and settings a model or dataset was built and tested for, and by exclusion those it was not. Datasheets, Data Cards and Model Cards
Interconnect
The fast links, such as NVLink or InfiniBand, that join GPUs into one system. AI Data Centers & the AI Factory
Interference
The accuracy loss that happens when deltas from different fine-tunes cancel or distort each other. Model Merging
Interleaved data
Training examples where text, images and audio appear in sequence, as they do in a real conversation. Any-to-Any (Omni) Models
Interleaving
Mixing results from two rankers in one list and counting which one users pick; it needs far less traffic than a full A/B split. Offline and Online Evaluation
Interpolation threshold
The model size or training point at which the network can fit every training example exactly; test error peaks near it. Double Descent
Intersection over union
How much a predicted box overlaps the true box. Object Detection
Inverse design
Starting from the properties you want and searching for a molecule or material that has them. AI for Science
IO-aware
An algorithm designed around how much data moves between levels of GPU memory, not just how many arithmetic operations it does. FlashAttention and IO-Aware Attention
ISO/IEC 42001
An international standard for AI management systems. AI Governance
Iteration
One update of a model's weights, usually after one mini-batch. Gradient Descent
Iterative magnitude pruning
Train, remove the smallest-magnitude weights, reset the rest to their initial values, and repeat. The Lottery Ticket Hypothesis

J

Jacobian
The matrix of first derivatives of a function with several outputs. Calculus for ML
Jagged frontier
The uneven boundary of what AI does well; tasks that look similar can fall on opposite sides of it. AI Adoption and Change Management
Jailbreak
A prompt designed to make a model ignore its safety rules. Prompt Injection & Jailbreaking
JAX
Google's library for high-performance numerical computing and ML research. ML Frameworks (PyTorch, TensorFlow, JAX)
Jupyter notebook
A document that mixes live code, results and notes. Exploratory Data Analysis (EDA)

K

K-fold cross-validation
Splitting data into k parts and training k times, each time testing on a different part. Evaluation Metrics & Cross-Validation
Kaplan-Meier estimator
A step curve of survival over time built from observed event times while keeping censored records. Survival Analysis
Keras
A high-level interface for building neural networks. ML Frameworks (PyTorch, TensorFlow, JAX)
Kernel
A function k(x, x') giving the covariance between function values at two inputs; it encodes smoothness and length scale. Gaussian Processes
Kernel fusion
Combining several GPU operations into one program so intermediate results never round-trip through slow memory. FlashAttention and IO-Aware Attention
Kernel trick
A way to find non-linear boundaries by computing similarity in a higher-dimensional space. Support Vector Machines (SVM)
Keypoints
Specific landmarks in an image, such as joints on a body. Computer Vision
Knowledge cutoff
The date after which a model has no training data. Large Language Models (LLMs)
Kriging
The geostatistics name for GP regression, used for decades to interpolate mineral and soil measurements. Gaussian Processes
Kubeflow
An open source platform for running ML workflows on Kubernetes. MLOps (ML Operations)
Kubernetes
A system that schedules and scales containers across many machines. Model Deployment & Serving
KV cache
Stored attention values that spare a language model from recomputing earlier tokens. Inference Optimization

L

L1 regularization (lasso)
A penalty on the absolute size of weights that pushes many of them to exactly zero. Regularization
L2 regularization (ridge)
A penalty on squared weights that keeps them small without forcing them to zero. Regularization
Label consistency
Whether different labelers, or the same labeler on different days, give the same answer for the same kind of example. Data-Centric AI
Label error
A training or test example whose recorded answer is wrong. Data Quality for AI
Label noise
Wrong labels that make an example look hard to learn even though the input is ordinary. Data Hardness and Example Difficulty
Lakehouse
A design that adds warehouse features to a data lake. Data Pipelines & ETL
LangChain
A popular framework for composing LLM calls, tools and retrieval. Agentic Workflows & Patterns
Langfuse
An open source platform for tracing and evaluating LLM applications. LLMOps
LangGraph
A framework for building agents as graphs of steps with saved state. Multi-Agent Systems & Orchestration
LangSmith
A platform for tracing and evaluating LLM applications. LLMOps
Late interaction
Encoding query and document separately, then letting their token vectors interact only at scoring time. ColBERT and Late Interaction Retrieval
Latency
The delay between sending a request and getting a response. Inference Optimization
Latent diffusion
Running diffusion in a compressed space to make image generation much cheaper. Stable Diffusion & Text-to-Image
Latent state
A compact internal representation of a position that the model plans over, not a reconstruction of the board or screen. AlphaZero and MuZero
Layer normalization
Rescaling activations across features instead of across the batch; standard in transformers. Normalization Layers (BatchNorm to RMSNorm)
Lead scoring
Ranking prospects by how likely they are to buy, using past data. AI in Sales and Customer Support
Lead time
How far ahead a forecast looks; skill drops as lead time grows. AI for Climate, Weather and Materials
Learned dynamics model
MuZero's network that predicts the next latent state and reward from a state and an action, standing in for the game rules. AlphaZero and MuZero
Learning rate
How large a step the optimizer takes on each update; too high diverges, too low crawls. Gradient Descent
Learning rate schedule
A plan that changes the learning rate during training, usually lowering it over time. Optimizers: SGD, Momentum & Adam
Leave-one-out cross-validation
Cross-validation in which each single example takes a turn as the test set. Evaluation Metrics & Cross-Validation
Legitimate interest
A legal basis that allows processing when the controller's interest is not overridden by the person's rights. Data Protection Law and AI (GDPR and beyond)
Liar's dividend
The advantage a person gains by dismissing a real recording as fake because fakes are known to exist. AI, Elections and Information Integrity
LIME
A method that explains one prediction by fitting a simple model around it. Explainable AI (XAI)
Linear mode connectivity
Two trained copies of a network can be averaged without a loss increase, a property that predicts when tickets are stable. The Lottery Ticket Hypothesis
LISP
A programming language from 1958 that dominated early AI research. History of AI
Llama
Meta's family of open-weight language models. Open-Weights & Small Language Models
llama.cpp
A lightweight engine that runs language models on ordinary CPUs and laptops. Open-Weights & Small Language Models
LlamaIndex
A framework focused on connecting LLMs to your own data. Retrieval-Augmented Generation (RAG)
llms.txt
A proposed file that gives AI systems a guide to a website's content. AI in Marketing & Search
LM Studio
A desktop app for trying open models locally. Open-Weights & Small Language Models
Log loss
A classification loss that heavily penalizes confident wrong answers. Loss Function
Log-rank test
A hypothesis test for whether two survival curves differ. Survival Analysis
Logits
A model's raw output scores before they are turned into probabilities. Text Generation & Sampling
Long-term memory
Information an agent stores outside its context and retrieves across sessions. Agent Memory, Planning & Reflection
Loss curve
A plot of the loss during training, used to spot overfitting or a bad learning rate. Loss Function
Loss of control
A scenario in which humans can no longer correct, shut down or redirect an AI system. Existential Risk from AI
Lottery ticket hypothesis
The claim that a dense network contains a sparse subnetwork that, trained alone from its original initialization, matches the full network. Pruning and Sparsity

M

Mandate
In AP2, a signed record of what the buyer allowed, such as an intent mandate for the goal and a cart mandate for the final items and price. Agentic Commerce and Payments
Margin
The minimum gap a loss demands between the positive's score and the negative's score. Hard Negative Mining
Marginal likelihood
The probability of the data under the GP, maximized to choose kernel hyperparameters. Gaussian Processes
Markov chain
A random process in which the next state depends only on the current one. Probabilistic Graphical Models & HMMs
Masked diffusion
A discrete diffusion process where tokens are replaced by a mask token and the model learns to fill them back in. Diffusion Language Models
Masked language modeling
Training a model to fill in hidden words, as BERT does. BERT (Bidirectional Encoder)
Matplotlib
A widely used Python plotting library. Exploratory Data Analysis (EDA)
Matrix
A rectangular grid of numbers. Linear Algebra for ML
Matrix factorization
Breaking a user-item table into smaller factors that predict missing ratings. Recommendation Systems
Maturity stage
A named level on a scale, with the capabilities and practices an organization must show to be placed there. AI Maturity Models
Max tokens
A limit on how many tokens a model may generate. Text Generation & Sampling
Maximum likelihood
Choosing the parameters that make the observed data most probable. Statistics for ML
MaxSim
For each query token, the highest similarity to any document token; the document score is the sum of these maxima. ColBERT and Late Interaction Retrieval
MCP client
The component inside an AI app that connects to an MCP server. Model Context Protocol (MCP)
MCP server
A program that exposes tools and data to AI applications through the Model Context Protocol. Model Context Protocol (MCP)
Mean absolute error (MAE)
The average absolute gap between predictions and true values, less sensitive to outliers. Loss Function
Mean average precision
The standard accuracy score for object detection. Object Detection
Mean reciprocal rank
A search metric based on how high the first correct result appears. AI-Powered Search & Ranking
Mean squared error (MSE)
The average squared gap between predictions and true values. Loss Function
Medallion architecture
Organizing data in stages from raw, to cleaned, to ready for business use. Data Pipelines & ETL
Mel spectrogram
A picture of sound over time, scaled to match human hearing. Speech Recognition (ASR) & Whisper
Membership inference
An attack that works out whether a specific record was in the training data. Differential Privacy & ML Privacy
Memorization
A model reproducing training text verbatim, which rises when the text was seen many times. Data Deduplication and Pruning
Mental model
What a user believes the AI can do and how it works, which sets their expectations and their reaction to errors. Designing AI Products (Human-AI Interaction)
Metadata filtering
Limiting retrieval by fields such as date, source or permission. Retrieval-Augmented Generation (RAG)
Milvus
An open source vector database built for very large scale. Vector Databases
MinHash
A hashing trick that estimates how much two documents overlap without comparing them word by word. Data Deduplication and Pruning
Mini-batch
A small random group of training examples used for one update step. Gradient Descent
Mistral
A French AI company known for open-weight models. Open-Weights & Small Language Models
MLflow
An open source tool for tracking experiments and managing models. Experiment Tracking
Modality encoder
The part that converts an image, clip or waveform into tokens the shared transformer can read. Any-to-Any (Omni) Models
Model card
A short document that describes what a model is for, how it was tested and its limits. Responsible AI
Model inversion
An attack that reconstructs training data from a model's outputs. Adversarial Machine Learning
Model parallelism
Splitting one model across machines when it will not fit on one. Distributed Training
Model soup
A plain or weighted average of several fine-tuned checkpoints of the same base model. Model Merging
Model spec
A published document stating the intended behavior of a model family, used both to train it and to judge it. Model Specs and Behavior Policies
Model stealing
Copying a model's behavior by querying it many times, or taking its weights. Adversarial Machine Learning
Model-centric AI
Improving results by changing the model or its training recipe while the dataset stays fixed. Data-Centric AI
Model-wise double descent
The test-error curve seen as the number of parameters grows with everything else fixed. Double Descent
Modular addition
The toy task (a + b) mod p used in most grokking studies because its correct algorithm can be read off the weights. Grokking
Multi-armed bandit
A simple decision problem about which of several options to try next, with no states. Reinforcement Learning (RL)
Multi-head attention
Running several attention operations in parallel so that a model can track different relationships. Self-Attention
Multi-head latent attention
A design that stores a compressed latent vector per token and reconstructs keys and values from it. GQA, MQA, Multi-head Latent and Sliding Window Attention
Multi-hop question
A question whose answer needs facts from two or more documents combined. Agentic RAG
Multi-query attention
A design with one shared key head and one shared value head for all query heads, giving the smallest cache. GQA, MQA, Multi-head Latent and Sliding Window Attention
Multi-vector index
An index that stores many vectors per document, one per token, instead of a single pooled vector. ColBERT and Late Interaction Retrieval
MYCIN
A 1970s expert system that recommended antibiotics using rules. Symbolic AI & Expert Systems

N

Native multimodality
Handling every modality in one network rather than connecting separate models with text in between. Any-to-Any (Omni) Models
NDCG
A ranking metric that rewards placing the most relevant results at the top. AI-Powered Search & Ranking
Near miss
A failure that was caught or ran out of luck before anyone was harmed. AI Incidents and Incident Databases
Near-duplicate
A document that differs from another only by small edits, boilerplate or formatting. Data Deduplication and Pruning
Negative prompt
A description of what an image model should leave out. Image Generation
Neural audio codec
A network that compresses sound into a short stream of discrete tokens a language model can predict, such as Mimi or EnCodec. Speech-to-Speech and Realtime Voice Models
NIST AI RMF
A voluntary US framework for managing AI risk. AI Risk Management
No free lunch theorem
The result that no single algorithm is best across every possible problem. Machine Learning (ML)
Non-max suppression
Removing overlapping boxes so that each object is detected once. Object Detection
Normal distribution
The bell curve, in which most values sit near the mean. Statistics for ML
NP-hard
A class of problems for which no fast general solution is known. Search Algorithms (A*, Minimax)
NPU
A chip built to run neural networks efficiently, often found in phones and laptops. Edge AI & On-Device ML
Numerical weather prediction (NWP)
Forecasting by integrating physics equations on a supercomputer, the method ML models are benchmarked against. AI for Climate, Weather and Materials
NumPy
The core Python library for fast array math. ML Frameworks (PyTorch, TensorFlow, JAX)

O

Observer
In the five-level framework, the agent acts on its own and the person only watches the outcome. Levels of Agent Autonomy
Occam's razor
The preference for the simplest explanation that fits the evidence. Bias-Variance Tradeoff
Occupational exposure
A score for how much of a job's tasks a language model could do or speed up, used to compare jobs. AI and the Future of Work
Off-policy evaluation
Estimating how well a new policy would do using only logged data, needed because the policy cannot be tested live. Offline Reinforcement Learning
Offline-online gap
The difference between a model's score on a frozen dataset and its measured effect on real users. Offline and Online Evaluation
Ollama
A tool for downloading and running open models on your own computer. Open-Weights & Small Language Models
One-hot encoding
Turning a category into a list of zeros with a single one marking the category. Feature Engineering
Online softmax
A way to compute softmax over blocks one at a time while tracking running maximums and sums, so the full score matrix is never stored. FlashAttention and IO-Aware Attention
ONNX
An open format for moving trained models between frameworks and hardware. ML Frameworks (PyTorch, TensorFlow, JAX)
Ontology
A formal definition of the kinds of things in a domain and how they relate. Knowledge Graphs
Open source versus open weights
Open weights share the trained model; true open source also shares the data and training code. Open-Weights & Small Language Models
OpenAI-compatible API
An interface that copies OpenAI's request format so that models can be swapped with little code. Model Routing & AI Gateways
OpenTelemetry
An open standard for collecting traces, metrics and logs from software. ML Monitoring & Observability
Operating model
The agreed split of decision rights, funding, roles and processes between central and local teams. AI Center of Excellence and Operating Model
Operator
In the five-level framework, the person drives and the agent only responds to direct commands. Levels of Agent Autonomy
Opt-out
A rights holder's signal, such as a robots.txt rule, that its works may not be used for training. AI, Copyright and Training Data
Optical flow
Estimating how pixels move between video frames. Computer Vision
Orthogonality thesis
The idea that intelligence and goals are independent, so a capable system can hold any goal. AI Alignment
Out-of-distribution action
An action the dataset never recorded in a given state, whose value a Q-function can only guess. Offline Reinforcement Learning
Outcome supervision
Training the reward model only on whether the final answer was right. Reward Models and Process Reward Models
Outlier
A value far from the rest, which may be an error or the most interesting case. Anomaly Detection
Oversampling
Repeating or synthesizing examples of a rare class to balance training data. Data Augmentation

P

P-value
How likely a result at least this extreme would be if there were really no effect. Statistics for ML
Padding
Adding values around an image or sequence so that its size lines up. Convolutional Neural Network (CNN)
pandas
The standard Python library for working with tables of data. Exploratory Data Analysis (EDA)
Parallel decoding
Updating many positions in a single forward pass instead of one token per pass. Diffusion Language Models
Parameter
A value the model learns during training, such as a weight. Hyperparameter Tuning
Parental controls
Settings that let a parent limit or monitor a child's use of a product. Children, Teens and AI
Parquet
A column-based file format that is efficient for analytics and ML. Data Pipelines & ETL
Particle swarm optimization
A method in which candidate solutions move together, guided by the best found so far. Evolutionary & Swarm Optimization
Peak memory
The largest amount of GPU memory in use at any moment of a training step, the number checkpointing lowers. Gradient Checkpointing and Memory-Efficient Training
Peeking
Checking an experiment early and stopping when it looks good, which inflates the false positive rate. Offline and Online Evaluation
Perceptron
The earliest artificial neuron: a weighted sum followed by a threshold. Multilayer Perceptron (MLP)
Perplexity-based pruning
Dropping tokens a small model finds predictable, on the theory that they carry little new information. Prompt and Context Compression
Personal data
Any information relating to an identified or identifiable person. Data Protection Law and AI (GDPR and beyond)
pgvector
An extension that adds vector search to Postgres. Vector Databases
PII redaction
Removing personal details such as names and account numbers before processing or storing text. AI Guardrails
Pillar page
The broad page at the center of a topic cluster that links out to the detailed pages. Topic Density and Topical Coverage
Pilot
A small trial of an AI tool with a limited group, used to learn before a wider rollout. AI Adoption and Change Management
Pilot project
A small, time-boxed AI project run to test feasibility and value before committing to a full build. AI Strategy
Pilot-to-production gap
The common failure where a working prototype never becomes a system people rely on day to day. AI Strategy
Pinecone
A managed vector database. Vector Databases
Planner-executor
An agent design in which one step plans the work and others carry it out. Agentic Workflows & Patterns
Platform team
A team that runs shared infrastructure, tooling and guardrails so product teams can build AI features safely. AI Center of Excellence and Operating Model
Policy
The rule an agent follows for choosing an action in each state. Reinforcement Learning (RL)
Policy head
The network output giving a probability for each legal move, used to focus the search. AlphaZero and MuZero
Polynomial regression
Regression that fits a curve by adding powers of the input features. Linear Regression
POMDP
A decision process in which the agent cannot see the full state of the world. Markov Decision Process (MDP)
Pooling
Shrinking a feature map by keeping the strongest or average value in each area. Convolutional Neural Network (CNN)
Pose estimation
Locating a person's joints and limbs in an image or video. Computer Vision
Post-training
Later stages such as fine-tuning and RLHF that shape a pretrained model's behavior. RLHF (Reinforcement Learning from Human Feedback)
Posterior
What you believe about a value after updating on the data. Bayesian ML & Uncertainty
Posterior predictive
The Gaussian distribution of the function value at a new input after conditioning on the training data. Gaussian Processes
Power-seeking
The tendency of a capable goal-directed system to acquire resources and influence because they help with almost any goal. Existential Risk from AI
Precision
Of everything flagged as positive, the share that truly was. Evaluation Metrics & Cross-Validation
Precision at k
The share of the top k results that are relevant. AI-Powered Search & Ranking
Precision-recall curve
A plot of precision against recall across thresholds, useful for rare classes. Evaluation Metrics & Cross-Validation
Predictive maintenance
Using sensor data to predict when a machine will fail so it can be fixed before it breaks. AI in Retail, Supply Chain and Manufacturing
Preemption
When federal law overrides state law on the same subject. US AI Policy and Executive Orders
Presence penalty
A setting that discourages any token that has already appeared, nudging new topics. Text Generation & Sampling
Pretrained model
A model already trained on a large dataset that you can reuse or adapt. Transfer Learning
Pretraining
The first, largest training stage, in which a model learns language from huge amounts of text. Large Language Models (LLMs)
Principal component analysis (PCA)
A way to compress data onto the directions that hold the most variation. Dimensionality Reduction
Prior
What you believe about a value before seeing the data. Bayesian ML & Uncertainty
Prior over functions
Before seeing data, the GP says which functions are plausible, set entirely by its mean and kernel. Gaussian Processes
Probabilistic forecast
A forecast that gives a range of likely outcomes rather than one number, so planners can set safety stock. AI in Retail, Supply Chain and Manufacturing
Probabilistic output
A result that can vary between runs of the same input, which makes acceptance criteria harder to write. AI Product Management
Probability distribution
A description of how likely each possible value is. Probability & Bayes' Theorem
Procedural memory
Knowledge of how to do things, such as saved instructions and skills. Agent Memory, Planning & Reflection
Process supervision
Training the reward model on a label for each intermediate step of a solution. Reward Models and Process Reward Models
Product feed
A structured catalog the merchant publishes so agents can find and compare items. Agentic Commerce and Payments
Product risk
SVPG's four questions for any feature: is it valuable, usable, feasible and viable for the business. AI Product Management
Productivity J-curve
The pattern where a new general-purpose technology first lowers measured productivity, because of hidden investment, before raising it. Measuring AI Value and ROI
Program synthesis
Searching for a short program that turns each input grid into its output, a common ARC approach. ARC-AGI and Reasoning Benchmarks
Progress measure
A quantity computed from the weights that rises smoothly before the sudden jump, showing grokking is not a discontinuity inside the model. Grokking
Prolog
A logic programming language long used in symbolic AI. Symbolic AI & Expert Systems
Prometheus
An open source system for collecting and alerting on metrics. ML Monitoring & Observability
Prompt chaining
Passing one prompt's output into the next to break a task into steps. Agentic Workflows & Patterns
Prompt compression
Shortening a prompt by removing or rewriting low-information tokens before sending it to a model. Token Density and Information Density
Prompt template
A reusable prompt with slots filled in by software at run time. Prompt Engineering
Promptfoo
An open source tool for testing prompts and red teaming LLM apps. LLM Evaluation & Benchmarks
Propositional logic
Logic over whole true-or-false statements joined with and, or and not. Knowledge Representation & Reasoning
Protein structure prediction
Inferring a protein's 3D shape from its amino acid sequence. AI for Science
Proxy attribute
A feature, such as a zip code, that stands in for a sensitive trait. AI Bias & Fairness
Pruning mask
A binary pattern marking which weights are kept; the mask and the initialization together define the ticket. The Lottery Ticket Hypothesis
PyTorch
The most widely used deep learning framework in research. ML Frameworks (PyTorch, TensorFlow, JAX)

Q

Qdrant
An open source vector database written in Rust. Vector Databases
Quality classifier
A small model that scores documents so low-quality text can be dropped from a corpus. Data Deduplication and Pruning
Quality filter
A rule or classifier that keeps or drops documents based on language, length, boilerplate or educational value. Pretraining Data and Data Mixtures
Query rewriting
Rephrasing or expanding a question so that retrieval finds better results. Retrieval-Augmented Generation (RAG)
Query, key, value
The three vectors each token forms to compute attention: what it seeks, offers and passes on. Self-Attention
Qwen
Alibaba's family of open-weight language models. Open-Weights & Small Language Models

R

R-squared
The share of variation in the target that a regression model explains. Linear Regression
RACI
A chart naming who is responsible, accountable, consulted and informed for each AI decision or task. AI Center of Excellence and Operating Model
RAGAS
An open source framework for scoring RAG pipelines. LLM Evaluation & Benchmarks
Random forest
Many decision trees trained on random samples and features, then averaged. Decision Trees & Random Forests
Random variable
A quantity whose value depends on chance. Probability & Bayes' Theorem
Randomized controlled trial
An experiment that gives the tool to a random half of workers so the difference in output can be credited to the tool. Measuring AI Value and ROI
Rate limit
A cap on how many requests or tokens a user may send in a given time. Model Routing & AI Gateways
RDF
A standard for describing facts on the web as subject, predicate and object. Knowledge Graphs
ReAct
A pattern in which a model alternates between reasoning and taking actions with tools. Agentic Workflows & Patterns
Read-only connection
A database role that can select but never insert, update or delete, so a bad query cannot change data. Text-to-SQL and Natural Language Querying
Reanalysis
A best estimate of past global weather built by blending observations with a physics model; ERA5 is the usual training set. AI for Climate, Weather and Materials
Recall
Of everything truly positive, the share the model found. Evaluation Metrics & Cross-Validation
Receptive field
The part of the input that a given unit can see. Convolutional Neural Network (CNN)
Reciprocal rank fusion
A simple way to merge ranked lists from different searches. Hybrid Search & Reranking
Recomputation
Running part of the forward pass a second time during backpropagation to rebuild activations that were not kept. Gradient Checkpointing and Memory-Efficient Training
Recursive reward modeling
Using AI assistants to help humans evaluate outputs that are too hard to judge unaided. Scalable Oversight
Refusal
A model declining a request because the spec or policy says it should not comply. Model Specs and Behavior Policies
ReLU
An activation that passes positive values and outputs zero otherwise; the common default. Activation Functions
Remasking
Re-hiding low-confidence tokens between steps so they can be predicted again with more context. Diffusion Language Models
Repetition penalty
A setting that makes a model less likely to repeat tokens it has already used. Text Generation & Sampling
Representativeness
How well the data covers the real cases, groups and conditions the model will face. Data Quality for AI
Reranker
A slower cross-encoder that rescores top candidates and is often used to label hard negatives. Hard Negative Mining
Residual compression
ColBERTv2's trick of storing each token vector as a centroid id plus a small quantized residual to cut index size. ColBERT and Late Interaction Retrieval
Reskilling
Training workers for different tasks or roles when their current ones change. AI and the Future of Work
Retrieval grader
A check, often a model call, that scores whether each retrieved passage is relevant before it is used. Agentic RAG
Retrieval-augmented research
A legal AI tool that looks up real sources before answering, which reduces but does not remove hallucinations. AI in Law and Legal Work
Return on investment (ROI)
The gain from an investment divided by its cost, usually shown as a percentage over a set period. Measuring AI Value and ROI
Revocation
The act of cancelling an earlier executive order, as EO 14179 did to EO 14110. US AI Policy and Executive Orders
Reward hacking
When a policy finds outputs that score well on the reward model without actually being better. Reward Models and Process Reward Models
Reward model
A model that takes a prompt and a response and returns a single score estimating how much a human would prefer it. Reward Models and Process Reward Models
Rewinding
Resetting surviving weights to their values from an early training step instead of step zero, needed for deep networks. The Lottery Ticket Hypothesis
Risk taxonomy
A structured list of risk types used to classify incidents consistently. AI Incidents and Incident Databases
ROC curve
A plot of true positive rate against false positive rate across all thresholds. Evaluation Metrics & Cross-Validation
Rogue AI
An AI system that pursues goals its developers did not intend and resists correction. Existential Risk from AI
Role prompting
Asking a model to act as a particular expert or persona. Prompt Engineering
Root mean squared error (RMSE)
The square root of MSE, in the same units as the target. Loss Function
Router
A step where the model chooses which index, database or tool should answer a given question. Agentic RAG
Router agent
An agent or model that decides which specialist or tool should handle a request. Agentic Workflows & Patterns
Rule-based system
Software that decides using hand-written if-then rules. Symbolic AI & Expert Systems

S

Saliency
An estimate of how much the loss would rise if a given weight were removed; pruning drops the least salient weights. Pruning and Sparsity
Saliency map
An image highlighting which pixels most influenced a prediction. Explainable AI (XAI)
Sample-wise non-monotonicity
The case where adding training data temporarily makes test error worse, because it moves the interpolation threshold. Double Descent
Sanctions
Penalties a court imposes on lawyers for misconduct, including filing fabricated citations. AI in Law and Legal Work
Sandbagging
A model deliberately underperforming on a test. Deceptive Alignment & Scheming
Sandbox
An isolated environment that limits what agent-run code can touch. AI Guardrails
Sandwiching
An experiment design where non-expert people working with a model are compared against both experts and the model alone. Scalable Oversight
SARSA
A reinforcement learning method that learns from the action actually taken next. Q-Learning & Deep Q-Networks (DQN)
Scalar
A single number. Linear Algebra for ML
Scaling
Moving from a few successful pilots to AI used consistently across many teams, with shared platforms and governance. AI Maturity Models
Schema drift
A change in the structure or meaning of incoming data that silently breaks a trained model. Data Quality for AI
Schema linking
Matching the words in a question to the tables and columns that hold the answer. Text-to-SQL and Natural Language Querying
Scientific foundation model
A large model pretrained on scientific data such as sequences, spectra or simulations, then adapted to many tasks. AI for Science
scikit-learn
The standard Python library for classic machine learning. ML Frameworks (PyTorch, TensorFlow, JAX)
Segment Anything (SAM)
Meta's model family that can outline objects in images and video from simple prompts. Semantic Segmentation
Selection bias
Errors that arise when the data does not represent the people a model will serve. AI Bias & Fairness
Selective checkpointing
Keeping cheap-to-store or expensive-to-recompute activations and discarding the rest, rather than checkpointing whole layers. Gradient Checkpointing and Memory-Efficient Training
Self-consistency
Sampling several reasoned answers and taking the most common one. Chain-of-Thought Reasoning
Self-driving lab
Robots plus machine learning that choose, run and learn from experiments with little human input. AI for Science
Self-paced learning
A curriculum where examples with low loss are admitted first and the threshold rises over time. Curriculum Learning
Self-play
Training an agent by having it play against copies of itself. Multi-Agent Reinforcement Learning
Self-RAG
A method where the model emits reflection tokens that say whether to retrieve and whether its draft is supported. Agentic RAG
Semantic cache
A cache keyed by the meaning of a request, found through embedding similarity, rather than its exact text. Semantic Caching
Semantic layer
Shared business definitions that sit over data, so that AI and analysts mean the same thing. Enterprise AI
Semantic memory
General facts an agent has stored, separate from when it learned them. Agent Memory, Planning & Reflection
Semantic network
A graph of concepts joined by labeled relationships. Knowledge Representation & Reasoning
Semi-private evaluation set
Tasks withheld from the public so that a score cannot come from memorizing the answers. ARC-AGI and Reasoning Benchmarks
Seq2seq
A model that turns one sequence into another, such as a sentence into its translation. Encoder-Decoder Models
Service-level agreement
The measurable promises a vendor makes on uptime, latency, support and accuracy, with remedies if missed. Evaluating AI Vendors and Procurement
Serving engine
Software such as vLLM that runs a model efficiently for many users at once. Inference Optimization
Shadow AI
Staff using AI tools without the organization's approval or knowledge. AI Governance
Shadow deployment
Running a new model alongside the old one without users seeing its output. Model Deployment & Serving
SHAP
A method that assigns each feature a share of the credit for a prediction. Explainable AI (XAI)
Sharding
Splitting a large model or dataset into pieces spread across machines. Distributed Training
Short-term memory
What an agent keeps in its current context while working on a task. Agent Memory, Planning & Reflection
Sigmoid
An S-shaped function that squeezes any number into the range zero to one. Activation Functions
Sign election
The TIES step that picks, for each weight, the sign most of the merged models agree on, and keeps only deltas with that sign. Model Merging
Similarity threshold
The minimum similarity score a stored question must reach before its answer is reused. Semantic Caching
Simulated annealing
An optimization method that accepts some worse moves early on to escape local optima. Evolutionary & Swarm Optimization
SLERP
Spherical linear interpolation, a way to blend two weight vectors along an arc so their magnitude is preserved. Model Merging
Sliding window attention
Attention restricted to a fixed number of nearby tokens, so cost stays constant as the sequence grows. GQA, MQA, Multi-head Latent and Sliding Window Attention
Softmax
A function that turns raw scores into probabilities that add up to one. Activation Functions
Sovereign AI
A nation building AI on its own infrastructure, data and workforce. AI Data Centers & the AI Factory
SPARQL
The query language for RDF knowledge graphs. Knowledge Graphs
Sparse (inducing-point) GP
An approximation that summarizes the data with a small set of points to escape the cubic cost. Gaussian Processes
Sparse autoencoder
A tool that splits a model's internal activity into more readable features. Mechanistic Interpretability
Sparse vector
A long vector in which most values are zero, as in keyword search. Hybrid Search & Reranking
Sparsity
The fraction of weights removed; a 90 percent sparse ticket keeps one weight in ten. The Lottery Ticket Hypothesis
Speaker diarization
Working out who spoke when in a recording. Speech Recognition (ASR) & Whisper
Special tokens
Reserved tokens that mark things such as the start of text or a change of speaker. Tokenization
Specification gaming
An AI meeting the literal goal it was given while missing what was intended. Reward Design & Reward Hacking
Speculative decoding
Using a small model to draft tokens that a large model then checks, to speed up generation. Inference Optimization
Speech-to-speech model
A single model that takes audio in and produces audio out, with no text transcript step in between. Voice Agents
SRAM
The small, very fast on-chip memory next to the GPU compute units where FlashAttention keeps each tile. FlashAttention and IO-Aware Attention
Stacking
Training a model to combine the predictions of several other models. Ensemble Learning
Standard deviation
A measure of how spread out values are around the mean. Statistics for ML
Step-back prompting
Asking a model to state the general principle before it answers the specific question. Prompt Engineering
Stockout
Running out of a product customers want to buy; the main cost of under-forecasting. AI in Retail, Supply Chain and Manufacturing
Stop sequence
A string that ends generation as soon as the model produces it. Text Generation & Sampling
Stopping criteria
The rules that tell an agent when to finish or give up. Agent Harnesses & Scaffolding
Streaming
Sending a model's output piece by piece as it is generated. Model Deployment & Serving
Streamlit
A Python library for building simple data apps. Model Deployment & Serving
Stride
How many pixels a filter moves at each step. Convolutional Neural Network (CNN)
Structured pruning
Removing whole channels, heads or layers so ordinary dense hardware sees a smaller model. Pruning and Sparsity
Sub-agent
A helper agent given a focused task and its own context by a lead agent. Multi-Agent Systems & Orchestration
Sublinear memory
Activation memory that grows with the square root of depth instead of linearly, the result Chen et al. proved. Gradient Checkpointing and Memory-Efficient Training
Suffix array
A sorted index of every suffix of a corpus, used to find exact repeated substrings across billions of tokens. Data Deduplication and Pruning
Surrogate model
A fast learned stand-in for an expensive simulation or experiment. AI for Science
Survival function
The probability that the event has not happened by time t. Survival Analysis
System card
A lab's report on how a released model was tested for safety and what was found. Frontier Safety Frameworks
System prompt
Standing instructions that set a model's role and rules for a whole conversation. Prompt Engineering

T

t-SNE
A method for drawing high-dimensional data in two dimensions, keeping neighbors close. Dimensionality Reduction
T5
Google's text-to-text transformer, an influential encoder-decoder model. Encoder-Decoder Models
Tanh
An S-shaped activation that outputs values from minus one to one. Activation Functions
Task vector
The difference between a fine-tuned model's weights and its base model's weights, treated as a direction you can add or subtract. Model Merging
Task-based analysis
Breaking a job into tasks and judging each one, because AI affects tasks, not whole jobs, first. AI and the Future of Work
Teacher forcing
Feeding the true previous token during training rather than the model's own prediction. Encoder-Decoder Models
Temperature
A setting that controls randomness: low stays predictable, high varies more. Text Generation & Sampling
Temporal difference learning
Updating value estimates using the next estimate, without waiting for the end of an episode. Q-Learning & Deep Q-Networks (DQN)
Tensor
A grid of numbers with any number of dimensions: the basic data structure of deep learning. Linear Algebra for ML
Test set
Data kept untouched until the end to estimate how a model performs on new cases. Evaluation Metrics & Cross-Validation
Test-time compute
Extra computation spent while answering, for example sampling many candidate solutions and checking them. ARC-AGI and Reasoning Benchmarks
Text and data mining exception
An EU rule that lets copies be made for analysis, which rights holders can opt out of for commercial uses. AI, Copyright and Training Data
Thinker-Talker
Qwen's split between a module that understands inputs and writes text and a module that turns that into streaming speech. Any-to-Any (Omni) Models
Thompson sampling
Choosing options in proportion to the chance that each one is the best. Reinforcement Learning (RL)
Threat modeling
Working out in advance who might attack a system, how, and what to do about it. Red Teaming AI Systems
Throughput
How many requests or tokens a system can handle in a given time. Inference Optimization
Ticket deflection
Resolving a customer's question through self-service or a bot so no human ticket is created. AI in Sales and Customer Support
Tiling
Splitting the query, key and value matrices into blocks small enough to process entirely in fast memory. FlashAttention and IO-Aware Attention
Time to first audio
The delay from the end of the caller's speech to the first sound of the agent's reply. Voice Agents
Time to first token
How long a person waits before a streamed answer begins. Inference Optimization
Time to live (TTL)
How long an entry stays valid before the cache drops it and fetches a fresh answer. Semantic Caching
TMRoPE
A position encoding that lines up audio and video frames in time so the model knows what was said when. Any-to-Any (Omni) Models
Token
The unit of text a language model reads and writes, often a word piece. Tokenization
Token budget
The total number of training tokens a run can afford, which the mixture must be sized to. Pretraining Data and Data Mixtures
Tokenizer tax
The extra tokens, and so extra cost and context use, that a tokenizer imposes on some languages and scripts. Token Density and Information Density
Tokens per second
How quickly a model generates text once it starts. Inference Optimization
Tool schema
The description of a tool's name and inputs that lets a model call it correctly. Tool Use & Function Calling
Top-k sampling
Sampling only from the k most likely next tokens. Text Generation & Sampling
Top-p (nucleus) sampling
Sampling only from the smallest set of tokens whose probabilities add up to p. Text Generation & Sampling
Topic cluster
A group of pages on one subject linked to a central page, a common content planning pattern. Topic Density and Topical Coverage
Topic distribution
The share of each topic in a document, as a topic model estimates it. Topic Density and Topical Coverage
Topical authority
A practitioner term for a site's depth and consistency of coverage on one subject; Google names a topic authority system only for news. Topic Density and Topical Coverage
Total cost of ownership
The full cost of an AI capability over its life, including licenses, compute, people, maintenance and retraining. Build, Buy or Partner
Tracing
Recording each step of a request, such as every model and tool call in an agent. LLMOps
Training data
The text, images, or other works a model learns from, which may be licensed, public, or scraped. AI, Copyright and Training Data
Transformative use
A use that gives the original a new purpose or meaning, which weighs in favor of fair use. AI, Copyright and Training Data
Tree of thoughts
Having a model explore and compare several lines of reasoning before answering. Chain-of-Thought Reasoning
Triplet loss
A loss that pulls an anchor toward a positive and pushes it away from a negative by at least a margin. Hard Negative Mining
True positive
A case the model flagged that really was positive. Evaluation Metrics & Cross-Validation
Turing test
Alan Turing's 1950 proposal: can a machine's conversation pass as human? History of AI
Turn detection
Deciding when the caller has finished speaking so the agent can reply without talking over them. Voice Agents

U

U-Net
A network shaped like a U that is widely used for segmentation and inside diffusion models. Semantic Segmentation
UMAP
A faster method for mapping high-dimensional data into a few dimensions for viewing. Dimensionality Reduction
Undersampling
Dropping examples of a common class to balance training data. Data Augmentation
Unstructured pruning
Zeroing individual weights anywhere, which gives the most compression but needs special kernels to run faster. Pruning and Sparsity
Usage policy
The contract-level rules for what customers may and may not do with a model. Model Specs and Behavior Policies
Usage rights
The legal and contractual permissions that say whether data may be used to train, fine-tune or ground a model. Data Strategy for AI
Use case
A specific business problem an AI system is meant to solve, described with its users, data and expected outcome. AI Strategy

V

Validation set
Data held back during training to tune settings and detect overfitting. Evaluation Metrics & Cross-Validation
Value function
An estimate of how much future reward a state or action will bring. Reinforcement Learning (RL)
Value head
The network output estimating who is winning from a position, replacing a hand-written evaluation function. AlphaZero and MuZero
Vanishing gradient
When gradients shrink through many layers until early layers almost stop learning. Recurrent Neural Network (RNN)
Variance
The average squared distance of values from their mean. Statistics for ML
Vector
An ordered list of numbers. Linear Algebra for ML
Vendor lock-in
Depending on one provider so heavily that switching becomes costly. Model Routing & AI Gateways
Visual inspection
Cameras and image models checking products on a line for defects faster and more consistently than people. AI in Retail, Supply Chain and Manufacturing
VRAM
Memory on a graphics card, which often limits the size of model you can run. GPU & TPU Computing

W

Weak supervisor
A human or smaller model whose judgments are noisy or incomplete relative to the system being trained. Scalable Oversight
Weak-to-strong generalization
A strong model trained on labels from a weaker supervisor ends up performing better than that supervisor. Scalable Oversight
Weaviate
An open source vector database with built-in hybrid search. Vector Databases
Weight decay
A penalty that pulls weights toward zero during training, which is a common form of regularization. Regularization
Weights
The numbers inside a model that training adjusts. Neural Network
Weights & Biases
A popular platform for tracking experiments and comparing training runs. Experiment Tracking
Whisper
OpenAI's open speech-to-text model. Speech Recognition (ASR) & Whisper
White-box testing
Testing with full access to a model's internals. Red Teaming AI Systems
Winning ticket
A sparse subnetwork plus its original initial weights that trains to the full network's accuracy on its own. The Lottery Ticket Hypothesis

Z

Z-score normalization
Rescaling a value by how many standard deviations it sits from the mean. Feature Engineering
Zero-shot prompting
Asking for a task with no examples at all. Prompt Engineering