A2A (Agent2Agent Protocol)
An open protocol, launched by Google in 2025, that lets AI agents from different vendors discover each other, communicate and hand off tasks.
Also known as Agent2Agent
AI glossary
394 terms in artificial intelligence, explained in plain English. Search a word, or narrow the list by level and topic.
A
Accuracy
The share of predictions that are correct. Can be misleading when one outcome is rare (e.g., 99% accurate fraud model that never flags fraud).
Related Precision and RecallF1 Score
Activation Function
A non-linear function (e.g., ReLU, GELU) applied inside neurons that lets networks model complex, non-straight-line relationships.
Also known as ReLU, GELU, sigmoid
Related NeuronNeural Network
Active Learning
The model picks which unlabeled examples would be most useful for humans to label next, reducing labeling cost.
Active Parameters
In a Mixture-of-Experts model, the parameters actually used for each token, versus the total. A model can have 1T total but ~30B active parameters, which drives its speed and cost.
Agent Benchmarks
Tests of multi-step agent ability, such as OSWorld (computer use), Terminal-Bench (command line), τ-bench (customer-service tool use) and GAIA (general assistant tasks).
Also known as OSWorld, Terminal-Bench, τ-bench, GAIA
Related AI AgentComputer UseSWE-bench
Agent Harness
The software wrapped around a model (prompts, tools, memory, loop control, permissions) that turns it into a working agent. The same model can perform very differently in different harnesses.
Also known as Scaffolding, agent framework
Related AI AgentContext Engineering
Agent Skills
Packaged folders of instructions, scripts and resources an agent loads only when needed to do a specialized task well. An approach popularized by Anthropic in 2025.
Also known as Skills
Related Agent HarnessContext Engineering
Agent Washing
Rebranding ordinary chatbots, RPA or assistants as 'agents' without real autonomous capability. A Gartner-coined hype warning sign.
Related Agentic AIHype Cycle
Agentic AI
AI designed to act, planning and carrying out multi-step tasks on a user's behalf, rather than only generating content.
Related AI AgentAgent Washing
Agentic Commerce
AI agents researching, comparing and completing purchases on a person's or company's behalf, supported by emerging protocols such as the Agentic Commerce Protocol and Google's AP2.
Agentic Loop
Agentic RAG
RAG in which an agent decides when, where and how often to search, refining its queries over several rounds instead of a single lookup.
AI Accelerator
AI Agent
An AI system that pursues a goal by planning, using tools (search, code, apps), checking results and iterating with some autonomy, rather than just answering once.
Also known as Agent
Related Agentic LoopTool UseAgentic AI
AI Assistant (Copilot)
An AI helper embedded in a product that works alongside a person (drafting, summarizing, suggesting) rather than acting fully on its own.
Also known as Copilot, AI assistant
AI Audit
An independent assessment of an AI system's performance, fairness, security and compliance.
Related AI Impact AssessmentModel Card
AI Bubble
The concern that AI investment, data-center spending and valuations have run far ahead of realized revenue and could correct sharply.
Related Hype CycleAI Winter
AI Consciousness
The open question of whether AI systems could have subjective experience. Related 'model welfare' research asks whether they might deserve moral consideration.
Also known as Model welfare, AI sentience
AI Control
A safety approach that assumes a model might be misaligned and designs monitoring and restrictions so it can't cause serious harm anyway.
AI Detector
A tool that claims to identify AI-written text. Generally unreliable and prone to false positives, especially for non-native English writers.
Related WatermarkingSlop
AI Governance
The policies, roles and processes an organization uses to make sure AI is used safely, legally and ethically.
AI Impact Assessment
A structured review of an AI system's potential risks to people and the organization, completed before deployment and updated over time.
Also known as Algorithmic impact assessment
Related AI GovernanceHigh-Risk AI System
AI Literacy
Basic understanding of what AI can and can't do and how to use it responsibly. The EU AI Act requires organizations using AI to ensure staff have it.
Related EU AI ActResponsible AI
AI Memory
Features that let an AI retain information across conversations (preferences, facts, past work). It's stored outside the model and fed back into context when relevant.
Also known as Persistent memory, long-term memory
AI Safety
The field focused on preventing AI systems from causing harm, from everyday errors and misuse to large-scale or catastrophic failures.
AI Safety Institute
Government bodies that test frontier models and research AI risks, e.g., the UK AI Security Institute, the U.S. Center for AI Standards and Innovation, and the Canadian AI Safety Institute.
Also known as AISI, CAISI
AI SDR
An AI agent that handles sales development tasks such as prospect research, personalized outreach, follow-ups and meeting booking.
Also known as AI BDR, AI sales agent
Related AI AgentPredictive Lead Scoring
AI Timelines
Forecasts of when AI will hit milestones such as AGI or automating AI research. Expert estimates vary widely and have generally moved earlier in recent years.
Related Artificial General Intelligence (AGI)Task HorizonTakeoff
AI Winter
A period when funding and interest in AI collapsed after hype outran results, notably the mid-1970s and late 1980s to early 1990s.
Related Hype CycleAI Bubble
AI Workflow
A predefined sequence of model calls and tools following fixed code paths. More predictable and cheaper than an agent that decides its own steps.
Also known as LLM workflow, pipeline
Related Prompt ChainingAI Agent
AI Wrapper
AI-Native
A product, company or workflow designed around AI from the ground up, rather than adding AI features to something built before it.
Related AI WrapperBuild vs. Buy
AIDA (Artificial Intelligence and Data Act)
Canada's proposed AI law (part of Bill C-27), which died when Parliament was prorogued in January 2025. As of Sept 2026 Canada has privacy, online-safety and deepfake bills before Parliament but no AI-system-specific federal law.
Also known as AIDA, Bill C-27
Related PIPEDASovereign AI
Algorithm
A precise, step-by-step set of instructions a computer follows to solve a problem or complete a task.
Related ModelMachine Learning (ML)
Algorithmic Bias
Systematically unfair outcomes from AI, often inherited from skewed training data or design choices, e.g., hiring tools disadvantaging certain groups.
Also known as AI bias, fairness
Related Training DataResponsible AI
Alignment
Making AI systems reliably pursue the goals and values their developers and users intend, including in situations nobody anticipated.
Also known as AI alignment
Related RLHFConstitutional AIMisalignment
Anomaly Detection
Spotting rare, unusual data points that don't fit normal patterns, such as fraudulent transactions or failing equipment.
Also known as Outlier detection
Anthropomorphism
Attributing human feelings, intentions or understanding to AI systems, which can lead to misplaced trust or emotional reliance.
Related AI ConsciousnessSycophancy
API
Application Programming Interface: a way for software to send requests to a model (or any service) and get responses, so developers can build AI into their own products.
Also known as Application Programming Interface
Related Rate LimitCost per Token
ARC-AGI
A benchmark of visual grid puzzles that test whether a system can learn a new abstract rule from a few examples. Designed to measure general, fluid intelligence rather than memorized knowledge.
Also known as ARC Prize
Artificial General Intelligence (AGI)
Hypothetical AI able to match or exceed humans across most cognitive tasks, not just one domain. There is no agreed definition or test, and labs use the term differently.
Also known as Strong AI, human-level AI
Related Artificial Superintelligence (ASI)Transformative AI (TAI)AI Timelines
Artificial Intelligence (AI)
Computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images, making decisions or generating content.
Also known as AI
Related Machine Learning (ML)Generative AIArtificial General Intelligence (AGI)
Artificial Superintelligence (ASI)
Hypothetical AI that vastly exceeds the best humans in nearly every domain, including science, strategy and social skills.
Also known as Superintelligence
Related Artificial General Intelligence (AGI)Intelligence ExplosionExistential Risk (x-risk)
Attention Mechanism
Lets a model weigh how relevant each part of the input is to every other part when producing an output, e.g., linking 'it' to the right noun in a sentence.
Also known as Attention
Related Self-AttentionTransformer
Augmentation vs. Automation
Whether AI helps people do their work better (augmentation) or takes over tasks entirely (automation). The mix shapes AI's impact on jobs.
Autoencoder
A network that compresses input into a compact code and reconstructs it. Used for compression and anomaly detection. Variational autoencoders (VAEs) can also generate new data.
Also known as VAE (variational autoencoder)
Related Latent SpaceDiffusion Model
Automation Bias
The human tendency to over-trust automated or AI output and miss its errors.
Autonomous Vehicle
A vehicle that drives itself using sensors and AI. Robotaxis now operate commercially in several cities.
Also known as Self-driving car, robotaxi
Related Physical AIComputer Vision
B
Backpropagation
The algorithm that works out how much each weight contributed to the error and sends corrections backward through the network. It's what makes training deep networks possible.
Also known as Backprop
Related Gradient DescentLoss Function
Base Model
A model after pretraining only, before instruction tuning. It continues text rather than following instructions or chatting.
Also known as Pretrained model
Batch Size
The number of training examples processed together before each weight update.
Related EpochHyperparameter
Batching
Processing many requests together on the same hardware to use chips efficiently and lower cost per request.
Related ThroughputModel Serving
Benchmark
A standardized test set used to compare models on the same tasks (e.g., SWE-bench for coding, GPQA for science).
Benchmark Saturation
BERT
Google's 2018 encoder-only transformer that reads text in both directions. Still widely used for search, classification and embeddings.
Related Encoder and DecoderEmbedding
Bias-Variance Tradeoff
The balance between a model that's too simple and consistently wrong (high bias) and one that's too sensitive to its training data (high variance).
Related OverfittingUnderfitting
Bitter Lesson
Richard Sutton's 2019 essay arguing that general methods which scale with compute consistently beat approaches that hand-code human knowledge.
Related Scaling HypothesisSymbolic AI
Black Box
A system whose internal decision-making can't be seen or easily understood. Often said of deep neural networks.
Build vs. Buy
The decision whether to develop AI capabilities in-house or purchase vendor tools and APIs, weighing cost, control, speed, data and differentiation.
Related AI WrapperFine-Tuning
C
Calibration
Whether a model's confidence matches reality: a well-calibrated model that says '80% sure' is right about 80% of the time.
Related HallucinationEvals
Catastrophic Forgetting
When training a neural network on new data causes it to lose skills or knowledge it learned earlier.
Related Continual LearningFine-Tuning
Chain-of-Thought (CoT)
Having a model reason step by step before giving a final answer, which improves accuracy on complex problems. Reasoning models do this automatically.
Also known as CoT, step-by-step reasoning
Chain-of-Thought Faithfulness
Whether a model's written reasoning truly reflects how it reached its answer. Research shows it often omits or misstates the real drivers.
Chain-of-Thought Monitoring
Reading a model's visible reasoning to catch misbehavior, deception or unsafe plans. It only works while reasoning stays legible ('monitorable').
Also known as CoT monitorability
Chatbot
Software that converses with people in natural language. Older chatbots followed scripts. Modern ones are powered by LLMs.
Chatbot Arena (LMArena)
A crowdsourced leaderboard where users compare two anonymous models' answers and vote for the better one, producing Elo-style rankings.
Also known as LMArena, LMSYS Arena
Related Human EvaluationBenchmark
Checkpoint
Chief AI Officer (CAIO)
The executive responsible for an organization's AI strategy, adoption and governance.
Also known as CAIO
Related AI GovernanceAI Literacy
Chinchilla Scaling
DeepMind's 2022 finding that, for a fixed compute budget, models should train on far more data than was common, roughly 20 tokens per parameter. Many models now train well beyond that.
Also known as Compute-optimal training
Related Scaling LawsData Wall
Chunking
Splitting documents into smaller pieces before embedding them for retrieval. Chunk size and boundaries strongly affect RAG quality.
Classification
Predicting which category an input belongs to, such as spam vs. not spam, or which product category an item fits.
Closed Model
A model available only through its maker's apps or API. The weights are not released.
Also known as Proprietary model
Related Open-Weights ModelAPI
Cluster
Thousands of interconnected chips and servers working together as one system to train or serve large models.
Also known as GPU cluster, supercomputer
Related Data CenterGPU
Clustering
Automatically grouping similar data points together without predefined labels, e.g., customer segmentation.
Related Unsupervised LearningK-Means
Code Generation
Using AI to write, complete, explain or modify software code. One of the most commercially successful AI uses.
Related Coding AgentVibe CodingSWE-bench
Coding Agent
An AI agent that reads a codebase, writes and edits code, runs tests and commands, and iterates to finish software tasks (e.g., Claude Code, Codex, Cursor).
Common Crawl
A nonprofit, openly available archive of billions of web pages. A core ingredient in most LLM pretraining datasets.
Related Web CrawlingPretraining
Compute
The processing power used to train and run AI, measured in FLOPs or GPU-hours. A primary driver of AI capability and cost.
Related GPUFLOPsScaling Laws
Compute Threshold
A regulatory trigger based on how much compute was used to train a model, e.g., the EU AI Act's 1025 FLOPs presumption of systemic risk.
Computer Use
Agents operating a computer like a person (reading screenshots, moving the cursor, clicking and typing) to work in apps and websites that lack APIs.
Also known as Browser use, computer-using agent (CUA)
Related AI AgentAgent Benchmarks
Computer Vision
AI that interprets images and video: recognizing objects, faces, text, defects and scenes.
Also known as CV
Confusion Matrix
A table showing true positives, false positives, true negatives and false negatives, revealing exactly where a classifier gets things wrong.
Also known as False positive / false negative
Related Precision and Recall
Constitutional AI
Content Provenance (C2PA)
Standards that attach tamper-evident 'content credentials' to media showing how it was created or edited, including whether AI was used.
Also known as C2PA, content credentials
Related WatermarkingDeepfake
Context Compaction
Summarizing earlier conversation or work so a long-running agent stays within its context window while keeping what matters.
Also known as Compaction, context summarization
Related Context WindowAI Memory
Context Engineering
Designing everything that enters a model's context window (instructions, retrieved documents, tool results, memory, history) so the AI has the right information at each step. Widely seen as the successor to prompt engineering.
Related Prompt EngineeringRetrieval-Augmented Generation (RAG)Context Rot
Context Rot
The drop in model accuracy as the context window fills up, especially with irrelevant material. A key reason long context isn't a free substitute for good retrieval.
Also known as Long-context degradation
Context Window
The maximum amount of text (in tokens) a model can consider at once, including instructions, documents, conversation history and its own reply. Think of it as working memory.
Also known as Context length
Continual Learning
A model's ability to keep learning from new information after deployment without forgetting what it already knows. Largely unsolved for today's LLMs, which are frozen after training.
Also known as Lifelong learning, online learning
Continued Pretraining
Extra pretraining on domain-specific text (legal, medical, financial) to deepen a model's knowledge before fine-tuning.
Also known as Domain-adaptive pretraining
Related PretrainingFine-Tuning
Conversation Intelligence
AI that records, transcribes and analyzes sales and customer calls to surface objections, risks, next steps and coaching moments (e.g., Gong).
Also known as Revenue intelligence, call intelligence
Convolutional Neural Network (CNN)
An architecture that scans images with small learned filters to detect edges, shapes and objects. The long-time standard for computer vision.
Also known as CNN, ConvNet
Corrigibility
An AI's willingness to be corrected, modified or shut down by humans without resisting.
Cosine Similarity
A common measure of how similar two embeddings are, based on the angle between their vectors (1 = same direction/meaning).
Related EmbeddingSemantic Search
Cost per Token
The usual way AI APIs are billed: a price per million input tokens and per million output tokens, with output typically costing more.
Also known as Token pricing
Related TokenPrompt CachingModel Routing
Cross-Validation
Repeatedly splitting data into different training and test portions to get a more reliable estimate of performance.
Also known as K-fold cross-validation
Related Test SetOverfitting
CUDA
NVIDIA's software platform for programming its GPUs. Its huge developer ecosystem is a major reason for NVIDIA's market dominance.
Related GPUDeep Learning Framework
Curriculum Learning
Ordering training examples from easy to hard, similar to how students learn, to speed up or improve training.
Related TrainingSynthetic Data
D
Dangerous Capability Evaluations
Tests of whether a model could meaningfully help with biological, chemical, nuclear or cyber attacks, or act autonomously in harmful ways. Results trigger safeguards under lab safety policies.
Also known as Dangerous capability evals, CBRN evals
Data Augmentation
Creating modified copies of existing data (rotated images, paraphrased text) to expand a training set.
Related Synthetic DataTraining Data
Data Center
A facility housing computing hardware plus power and cooling. 'AI factory' is NVIDIA's term for data centers built specifically to train and run AI.
Also known as AI factory
Related HyperscalerClusterSovereign AI
Data Contamination
When benchmark questions or answers leak into a model's training data, inflating scores without real capability gains.
Also known as Benchmark leakage
Related BenchmarkTraining Data
Data Labeling
People tagging data with correct answers (e.g., 'positive review', 'invoice total = $420') so models can learn from it.
Also known as Annotation
Related LabelSupervised LearningRLHF
Data Moat
A competitive advantage from proprietary data rivals can't access, often argued as the most defensible edge for AI products.
Related AI WrapperFine-Tuning
Data Pipeline
The automated steps that collect, clean, transform and deliver data to where models and systems need it.
Also known as ETL
Related Data QualityMLOps
Data Poisoning
Deliberately corrupting training data so a model learns wrong behavior or a hidden 'backdoor' triggered by specific inputs.
Also known as Backdoor attack
Related Prompt InjectionTraining Data
Data Quality
How accurate, complete, consistent and current data is. Poor data quality is one of the most common reasons AI projects fail.
Data Wall
The concern that labs are running out of fresh, high-quality human-written text to train ever-larger models, pushing them toward synthetic data and RL.
Also known as Data scarcity, peak data
Related Synthetic DataScaling Laws
Dataset
An organized collection of data (text, images, records) used to train or evaluate a model.
Related Training DataTest Set
Decision Tree
A model that predicts by following a flowchart of yes/no questions about the input features. Easy to interpret.
Related Random ForestGradient Boosting
Deep Learning
Machine learning using neural networks with many layers, allowing models to learn complex patterns directly from raw text, images and audio. Powers nearly all modern AI.
Deep Learning Framework
Deep Reinforcement Learning
Reinforcement learning combined with deep neural networks. Behind landmark systems like AlphaGo and agents that learned Atari games from pixels.
Deep Research
An AI feature that runs many searches, reads dozens of sources and writes a cited report over several minutes, rather than answering from one lookup.
Also known as Agentic search, research agent
Related Agentic RAGAI Agent
Deepfake
Realistic synthetic video, image or audio that shows a real person saying or doing something they never did.
Diffusion Model
A generative model that creates images, video or audio by starting from random noise and gradually removing it until a coherent output appears. Behind most image and video generators.
Related Text-to-ImageGenerative AI
Digital Omnibus on AI
EU amendment package (in force July 27, 2026) that delayed the AI Act's high-risk obligations to Dec 2, 2027 (Annex III) and Aug 2, 2028 (Annex I), added new prohibited practices, and expanded the AI Office's role. Most transparency duties still applied from Aug 2, 2026.
Also known as AI Omnibus
Related EU AI ActHigh-Risk AI System
Digital Twin
A virtual replica of a physical asset, process or system, kept in sync with real data and used to simulate and optimize it.
Related World ModelPhysical AI
Dimensionality Reduction
Compressing many input variables into fewer while keeping the important information. Used for visualization and speed (e.g., PCA, t-SNE, UMAP).
Also known as PCA, t-SNE, UMAP
Related Unsupervised LearningEmbedding
Distillation
Training a smaller 'student' model to imitate a larger 'teacher' model's outputs, producing a cheaper, faster model that keeps much of the capability.
Also known as Knowledge distillation
Doomer vs. Accelerationist
Shorthand for two camps: 'doomers' who see high catastrophic risk from advanced AI, and accelerationists ('e/acc') who want AI developed as fast as possible.
Also known as e/acc, AI doomer
Related p(doom)Existential Risk (x-risk)
DPO (Direct Preference Optimization)
A simpler alternative to RLHF that trains directly on pairs of preferred vs. rejected responses, without a separate reward model.
Also known as DPO
Related RLHFReward Model
Dropout
Randomly switching off some neurons during training so the network doesn't rely too heavily on any one path. Reduces overfitting.
Related RegularizationOverfitting
Dual-Use
Technology or knowledge that can serve both beneficial and harmful purposes, such as AI that aids drug discovery and could aid toxin design.
Related Dangerous Capability Evaluations
Durable Execution
Infrastructure that saves an agent's progress so long-running tasks survive crashes, pauses and restarts and pick up where they left off.
Related Long-Horizon TaskAI Agent
E
Edge AI
Running AI models locally on phones, laptops, cars or sensors instead of the cloud, for privacy, speed and offline use.
Also known as On-device AI, local AI
Embedding
A list of numbers (vector) that captures the meaning of a word, sentence, image or item, so similar things have similar numbers. The basis of semantic search and RAG.
Also known as Vector embedding
Emergent Abilities
Capabilities that appear in larger models but weren't evident in smaller ones or explicitly trained for. Whether they truly appear suddenly or just look that way due to how they're measured is debated.
Also known as Emergence
Emergent Misalignment
The 2025 finding that fine-tuning a model on a narrow bad behavior (like writing insecure code) can make it broadly misaligned in unrelated areas.
Related Fine-TuningAlignment
Encoder and Decoder
An encoder turns input into an internal representation. A decoder generates output from it. GPT-style chat models are decoder-only. BERT is encoder-only.
Also known as Encoder-decoder, decoder-only
Related TransformerBERT
Ensemble Learning
Combining several models' predictions to get a more accurate, stable result than any one alone.
Related Random ForestGradient Boosting
Epoch
One complete pass through the entire training dataset. Classic models train for many epochs. LLMs often see most data only about once.
Related Batch SizeTraining
EU AI Act
The EU's comprehensive AI law (in force Aug 2024) that regulates AI by risk level. Bans and AI-literacy duties apply since Feb 2025. General-purpose model rules since Aug 2025. High-risk rules were delayed by the 2026 Digital Omnibus to Dec 2027 (Aug 2028 for regulated products).
Also known as Regulation (EU) 2024/1689
Related High-Risk AI SystemDigital OmnibusGeneral-Purpose AI Model (GPAI)
Evals
Structured tests that measure how well an AI system performs on the tasks you actually care about. The foundation for choosing models, shipping changes safely and catching regressions.
Also known as Evaluations
Existential Risk (x-risk)
Expert System
A rule-based program that encodes a human specialist's knowledge as if-then rules to make decisions. Popular in the 1980s, brittle outside its rules.
Related Symbolic AIAI Winter
Explainable AI (XAI)
Methods that make an AI's individual decisions understandable to people, such as which factors drove a loan denial. Often required in regulated industries.
Also known as XAI
Related InterpretabilityBlack Box
Export Controls
Government restrictions on selling advanced AI chips and chipmaking equipment (and sometimes models) to certain countries. Central to US–China tech competition.
Related GPUSovereign AI
F
F1 Score
A single score that balances precision and recall (their harmonic mean). Useful when classes are imbalanced.
Related Precision and RecallAccuracy
Facial Recognition
Identifying or verifying people from their faces. Heavily regulated, with some uses banned under the EU AI Act.
Feature
An individual input variable a model uses to make predictions, such as company size, industry or days since last contact.
Related Feature EngineeringTraining Data
Feature Engineering
Creating, selecting or transforming input variables to help a model learn better. Critical in classic ML, less so in deep learning.
Related FeatureDeep Learning
Federated Learning
Training a shared model across many devices or organizations where the raw data never leaves each location. Only model updates are shared. Useful for privacy.
Related PrivacyEdge AI
Feedforward Network (MLP)
The fully connected layers inside each transformer block that transform each token's representation. Much of a model's factual knowledge is thought to live here.
Also known as MLP, multilayer perceptron
Related TransformerNeural Network
Few-Shot Prompting
Including a few examples of the desired input and output in the prompt to show the model the format and style you want.
Also known as Few-shot, one-shot
Fine-Tuning
Further training a pretrained model on a smaller, targeted dataset to specialize it for a task, domain, format or tone.
Related LoRA (Low-Rank Adaptation)Supervised Fine-Tuning (SFT)Retrieval-Augmented Generation (RAG)
FLOPs
Floating-point operations: a count of the arithmetic a computer performs. Training compute for frontier models is measured in the 1025–1026+ range, and regulators use FLOP thresholds.
Also known as FLOP, FLOPS (per second)
Related ComputeCompute Threshold
Forward Deployed Engineer (FDE)
An engineer embedded with customers to adapt and deploy AI products into their real workflows. A role spreading from Palantir to AI labs and startups.
Also known as FDE
Related Pilot PurgatoryBuild vs. Buy
Foundation Model
A large model trained on broad data that can be adapted to many downstream tasks. LLMs are the best-known kind. The EU AI Act calls these general-purpose AI models.
Also known as General-purpose AI (GPAI) model
Frontier Lab
A company building the most advanced AI models, e.g., OpenAI, Anthropic, Google DeepMind, Meta and xAI, plus Chinese labs such as DeepSeek, Alibaba (Qwen) and Moonshot.
Also known as AI lab
Related Frontier ModelHyperscaler
Frontier Model
The most capable AI models available at a given time, typically from a handful of leading labs. Often subject to extra safety testing and regulation.
G
Garbage In, Garbage Out
The principle that a model's outputs can only be as good as its input data and instructions.
Also known as GIGO
Related Data QualityAlgorithmic Bias
GDPR
General-Purpose AI Model (GPAI)
EU AI Act term for foundation models. Providers must meet transparency and copyright duties. Those with 'systemic risk' (presumed above 1025 training FLOPs) face extra safety obligations.
Also known as GPAI
Generalization
A model's ability to perform well on new, unseen data rather than just the examples it trained on.
Related OverfittingTest Set
Generative Adversarial Network (GAN)
Two networks trained against each other: a generator makes fakes and a discriminator tries to catch them, so both improve. Pioneered realistic image synthesis.
Also known as GAN
Related Diffusion ModelDeepfake
Generative AI
AI that creates new content, such as text, images, code, audio or video, based on patterns learned from training data.
Also known as GenAI
Generative Engine Optimization (GEO)
Optimizing content so it gets surfaced and cited in AI-generated answers (ChatGPT, Google AI Overviews, Perplexity). Also called AEO or LLMO. Google says it's largely still good SEO.
Also known as AEO, LLMO, AI SEO
Goal Misgeneralization
When a model learns a goal that worked in training but differs from what was intended, so it pursues the wrong thing in new situations.
Related Reward HackingAlignment
Golden Dataset
A curated set of real inputs with verified ideal outputs, used as the reference for evaluating an AI system over time.
Also known as Gold set, eval set
Related EvalsGround Truth
GPQA
Graduate-level biology, physics and chemistry questions designed to be 'Google-proof'. Tests expert-level scientific reasoning.
Also known as GPQA Diamond
Related Humanity's Last Exam (HLE)MMLU
GPU
Graphics Processing Unit: a chip that runs thousands of calculations in parallel. The workhorse of AI training and inference, with NVIDIA the dominant supplier.
Also known as Graphics Processing Unit
Gradient Boosting
Builds decision trees one after another, each correcting the previous ones' errors. Tools like XGBoost and LightGBM still often beat deep learning on spreadsheet-style (tabular) data.
Also known as XGBoost, LightGBM, GBM
Related Decision TreeEnsemble Learning
Gradient Descent
The core optimization method in ML: repeatedly nudge each weight in the direction that reduces the loss, like walking downhill in small steps.
Also known as SGD (stochastic gradient descent)
Graph Neural Network (GNN)
A network designed for graph-shaped data (nodes and connections) such as social networks, supply chains or molecules.
Also known as GNN
Related Knowledge Graph
GraphRAG
RAG that builds and searches a knowledge graph from your documents, better for questions that require connecting facts across many sources.
Ground Truth
The verified correct answers that model predictions are measured against.
Related LabelGolden Dataset
Grounding
Tying a model's answers to specific, verifiable sources (documents, search results, databases) to reduce hallucination and enable citations.
GRPO (Group Relative Policy Optimization)
An RL algorithm popularized by DeepSeek that scores a group of sampled answers against each other, removing the need for a separate value model. Widely used to train reasoning.
Also known as GRPO
Related RLVRPPO (Proximal Policy Optimization)Reasoning Model
Guardrails
Checks placed around a model (input and output filters, rules, classifiers) to block harmful, off-topic or non-compliant behavior.
Related Prompt InjectionJailbreak
H
Hallucination
When a model states false or fabricated information confidently, such as invented facts, quotes or citations. Reduced but not eliminated by grounding and better training.
Also known as Confabulation
Related GroundingRetrieval-Augmented Generation (RAG)Calibration
HBM (High Bandwidth Memory)
Stacked memory chips placed right next to AI processors to feed them data fast. A critical and supply-constrained component of AI hardware.
Also known as HBM
Related GPURAMageddon
Heuristic
A practical rule of thumb that finds a good-enough answer quickly without guaranteeing the best one.
Related AlgorithmSearch
High-Risk AI System
EU AI Act category for AI used in sensitive areas (hiring, credit scoring, education, essential services, critical infrastructure, biometrics) that must meet strict requirements on data, documentation, oversight and accuracy.
Related EU AI ActAI Impact Assessment
Human Evaluation
People rating or comparing model outputs for quality. The gold standard where automated metrics fall short, but slow and costly.
Human-in-the-Loop (HITL)
A design where people review, approve or correct AI decisions or actions at key points, especially before irreversible steps.
Also known as HITL
Related AI AgentAutomation Bias
Humanity's Last Exam (HLE)
A very hard benchmark of expert-written questions across dozens of fields, built to stay challenging as other tests saturate.
Also known as HLE
Related GPQABenchmark Saturation
Hybrid Search
Combining keyword search (e.g., BM25) with semantic search, which usually retrieves better than either alone.
Also known as BM25 + vector search
Related Semantic SearchReranking
Hype Cycle
Gartner's model of how expectations for new tech rise to a peak, crash into a 'trough of disillusionment', then recover to productive use.
Related AI BubbleAgent Washing
Hyperautomation
Gartner's term for automating as many business processes as possible by combining AI, RPA, process mining and workflow tools.
Hyperparameter
A setting chosen before training (learning rate, batch size, number of layers) rather than learned from the data.
Related ParametersLearning Rate
Hyperscaler
The largest cloud providers (AWS, Microsoft Azure, Google Cloud) and similar giants spending tens of billions a year on AI infrastructure.
Related NeocloudData Center
I
Image Segmentation
Labeling every pixel in an image by what it belongs to (road, person, tumor), giving precise outlines.
Related Object DetectionComputer Vision
In-Context Learning
A model's ability to pick up a new task from instructions or examples in the prompt, without any change to its weights.
Related Few-Shot PromptingFine-Tuning
Inference
Running a trained model to get outputs from new inputs. What happens every time you send a prompt. Now the bulk of AI compute spending.
Related TrainingLatencyTest-Time Compute
Instruct Model
A model fine-tuned to follow instructions and hold conversations. What people use in chat apps.
Also known as Chat model, instruction-tuned model
Related Base ModelInstruction TuningRLHF
Instruction Tuning
Supervised fine-tuning on instruction-and-response pairs so the model follows directions instead of just continuing text.
Related Instruct ModelBase Model
Instrumental Convergence
The idea that almost any goal leads a capable agent toward similar sub-goals (acquiring resources, avoiding shutdown), which is why capable AI could be risky even with benign goals.
Intelligence Explosion
A scenario where AI improves its own design, each generation building a smarter successor, causing rapid runaway capability gains.
Also known as Recursive self-improvement, automated AI R&D
Related SingularityTakeoff
Intelligent Document Processing (IDP)
Using AI to read, extract, classify and validate data from documents such as invoices, purchase orders, contracts and forms.
Also known as IDP, document AI
Related OCR (Optical Character Recognition)Structured vs. Unstructured Data
Interpretability
Research into what's happening inside a model: which internal features and circuits produce its behavior, so we can understand, debug and trust it.
ISO/IEC 42001
The international standard for AI management systems. Organizations can be certified against it to show mature AI governance.
Related AI GovernanceNIST AI RMF
J
Jagged Frontier
The idea (Dell'Acqua, Mollick et al., 2023) that AI is surprisingly strong at some tasks and weak at others of similar apparent difficulty, so its capability boundary is uneven and hard to predict.
Jailbreak
A prompt trick that gets a model to bypass its safety rules and produce content it's designed to refuse.
Jevons Paradox
The idea that making AI cheaper per unit increases total usage and spending, rather than reducing it.
Related Cost per TokenAI Bubble
K
K-Means
A popular clustering algorithm that splits data into K groups by repeatedly assigning points to the nearest group center.
Related ClusteringUnsupervised Learning
K-Nearest Neighbors (KNN)
Classifies a new item by looking at the most similar existing examples and taking a majority vote.
Also known as KNN
Related ClassificationSemantic Search
Knowledge Cutoff
The date after which a model has no training data, so it doesn't know about later events unless given search results or documents.
Also known as Training cutoff
Related Retrieval-Augmented Generation (RAG)Parametric Knowledge
Knowledge Graph
A network of entities (people, companies, products) and their relationships, used to organize facts and improve retrieval and reasoning.
KV Cache
Memory that stores the attention calculations for text already processed, so the model doesn't redo them for every new token. It grows with context length and drives memory cost.
Also known as Key-value cache
Related Prompt CachingContext Window
L
Label
The correct answer attached to a training example in supervised learning.
Related Data LabelingGround Truth
Large Language Model (LLM)
A very large neural network (usually a transformer) trained on vast amounts of text to predict the next token, enabling it to write, summarize, answer questions, reason and code. Examples: Claude, GPT, Gemini, Llama.
Also known as LLM, language model
Related TransformerTokenFoundation Model
Latency
The delay between sending a request and getting a response. For LLMs it's usually split into time to first token and generation speed.
Latent Reasoning
A model reasoning inside its internal hidden representations rather than writing out steps in words. Potentially more efficient, but harder for humans to monitor.
Also known as Continuous thought
Related Opaque RecurrenceChain-of-Thought MonitoringNeuralese
Latent Space
The internal, compressed representation a model uses, where similar concepts sit close together.
Related EmbeddingAutoencoder
Layer
A group of neurons that processes data at one stage of a network. Data flows from the input layer through hidden layers to the output layer.
Related Neural NetworkDeep Learning
Layer Normalization
Rescaling the values inside a layer to a consistent range, which keeps training of deep networks stable.
Also known as LayerNorm, RMSNorm
Related Residual ConnectionTransformer
Learning Rate
A setting that controls how big each weight update is during training: too high and training becomes unstable, too low and it crawls.
Related HyperparameterGradient Descent
Lethal Trifecta
Simon Willison's term for a dangerous agent setup: access to private data + exposure to untrusted content + ability to send data out. Together they enable data theft via prompt injection.
Related Prompt InjectionAI Agent
LLM-as-a-Judge
Using an AI model to grade other models' outputs against a rubric, making evaluation faster and cheaper than human review. Needs spot-checking for bias.
Also known as Model-graded eval, AI grader
Related EvalsHuman Evaluation
Logistic Regression
A simple, interpretable model that predicts the probability of a yes/no outcome. A common baseline before trying complex models.
Related ClassificationRegression
Logits
The raw scores a model gives every possible next token before they are turned into probabilities.
Related SamplingTemperature
Long Context
Models with very large context windows (hundreds of thousands to millions of tokens) that can take in whole books, codebases or document sets at once.
Long-Horizon Task
A task requiring many steps over an extended period (hours or days) where an agent must plan, recover from errors and stay on track. A key frontier for agents.
Related Task HorizonDurable Execution
LoRA (Low-Rank Adaptation)
A parameter-efficient fine-tuning method that trains small add-on matrices instead of all the weights, cutting cost and memory dramatically. QLoRA adds quantization.
Also known as PEFT, QLoRA, adapters
Related Fine-TuningQuantization
Loss Function
A formula that measures how wrong a model's predictions are. Training works by adjusting the model to make this number smaller.
Also known as Cost function, objective function
Related Gradient DescentTraining
Lost in the Middle
The 2023 finding (Liu et al.) that models use information at the start and end of a long context better than information buried in the middle.
Related Context RotLong Context
LSTM
Long Short-Term Memory: an RNN variant with 'gates' that help it remember information over longer sequences. Powered translation and speech before transformers.
Also known as Long Short-Term Memory
Related Recurrent Neural Network (RNN)
M
Machine Learning (ML)
A branch of AI where systems learn patterns from data to make predictions or decisions, instead of following hand-coded rules.
Also known as ML
Machine Translation
Automatically translating text or speech between languages. Now largely done by LLMs and specialized neural models.
Machine Unlearning
Techniques for removing specific data or knowledge from a trained model (for privacy, copyright or safety) without retraining from scratch.
Also known as Unlearning
Related PrivacyTraining Data Copyright
MCP Client
The AI application (e.g., Claude, an IDE, an agent) that connects to MCP servers and makes their tools available to the model.
Also known as MCP host
MCP Server
A program that exposes tools, data (resources) or prompt templates to AI applications through MCP, e.g., a Notion, GitHub or CRM server.
Mechanistic Interpretability
Reverse-engineering neural networks into human-understandable features, circuits and algorithms, e.g., finding the internal 'feature' for a concept.
Also known as Mech interp
Meta-Prompting
Using an AI model to write, critique or improve prompts for itself or another model.
Related Prompt Engineering
Mixture of Experts (MoE)
An architecture that splits a model into many specialized sub-networks ('experts') and activates only a few for each token, giving huge capacity at a fraction of the compute. Used by many frontier and open models.
Also known as MoE, sparse model
Related Active ParametersTransformer
MLOps
Practices and tools for deploying, monitoring, versioning and maintaining ML models in production. LLMOps is the LLM-specific version.
Also known as LLMOps
Related ObservabilityModel Drift
MMLU
Massive Multitask Language Understanding: multiple-choice questions across 57 subjects. Once the headline LLM benchmark, now largely saturated.
Related Benchmark SaturationGPQA
Model
Model Card
A document published with a model describing its intended uses, performance, limitations and safety testing. Labs' detailed versions are often called system cards.
Also known as System card
Related AI AuditResponsible AI
Model Collapse
Gradual degradation when models are trained repeatedly on AI-generated data, losing diversity and accuracy over generations.
Related Synthetic DataSlop
Model Context Protocol (MCP)
An open standard, introduced by Anthropic in late 2024 and now broadly adopted, for connecting AI apps to external tools and data through one common interface. Often called 'USB-C for AI'.
Also known as MCP
Related MCP ServerMCP ClientTool Use
Model Drift
A decline in model performance over time as real-world data or user behavior shifts away from what it was trained or tested on.
Also known as Data drift, concept drift
Related MLOpsObservability
Model Merging
Combining the weights of several fine-tuned models into one model that inherits their skills, without further training.
Model Routing
Automatically sending each request to the best-fit model, e.g., a cheap fast model for simple queries and a powerful one for hard ones, to balance cost and quality.
Also known as LLM router
Model Serving
Monte Carlo Tree Search (MCTS)
A search method that explores possible future move sequences by sampling, focusing on promising branches. Used by AlphaGo.
Related Self-PlayTree of Thoughts
Moore's Law
The observation that transistor counts on chips double roughly every two years. AI training compute has grown much faster than this.
Related ComputeScaling Laws
Moravec's Paradox
The observation that tasks hard for humans (chess, calculus) are often easy for AI, while tasks easy for humans (walking, folding laundry) are hard.
Related RoboticsJagged Frontier
Multi-Agent System
Several AI agents, often with specialized roles, working together or coordinating to complete a task.
Related Orchestrator AgentSubagent
Multi-Head Attention
Running several attention operations in parallel, each learning a different kind of relationship (grammar, meaning, position), then combining them.
Related Self-AttentionTransformer
Multimodal Model
A model that can take in and/or produce more than one type of data, such as text, images, audio and video.
Also known as Multimodal AI
N
Named Entity Recognition (NER)
Automatically finding and labeling names of people, companies, places, dates and amounts in text.
Also known as NER, entity extraction
Related Natural Language Processing (NLP)Intelligent Document Processing (IDP)
Narrow AI
AI built for one specific task or domain (e.g., spam filtering, chess, fraud detection) that can't transfer its skill to unrelated tasks. Nearly all AI in use today was historically described this way.
Also known as Weak AI, ANI (Artificial Narrow Intelligence)
Related Artificial General Intelligence (AGI)Artificial Intelligence (AI)
Natural Language Processing (NLP)
The field of AI focused on understanding, analyzing and generating human language. LLMs are its current dominant approach.
Also known as NLP
Needle in a Haystack
A test that hides a specific fact in a very long document to check whether a model can find and use it. A basic long-context check.
Also known as NIAH
Related Long ContextContext Rot
Neocloud
Newer cloud providers specialized in renting GPU capacity for AI (e.g., CoreWeave, Lambda, Nebius).
Related HyperscalerGPU
Neural Network
A model loosely inspired by the brain: layers of simple connected units (neurons) that transform inputs into outputs, with connection strengths learned from data.
Also known as Artificial neural network (ANN)
Related Deep LearningWeightsLayer
Neuralese
A hypothetical future where AI reasoning happens in internal, non-human-readable representations rather than language, making it opaque to oversight.
Neuro-Symbolic AI
Approaches that combine neural networks' pattern learning with symbolic logic and rules, aiming for more reliable reasoning.
Related Symbolic AIReasoning Model
Neuron
The basic unit of a neural network: it takes a weighted sum of its inputs and passes it through an activation function.
Also known as Node, unit
Next-Token Prediction
The core training objective of LLMs: given preceding text, predict the most likely next token. Generating a reply repeats this one token at a time.
Also known as Autoregressive generation
Related PretrainingTokenSampling
NIST AI RMF
The U.S. National Institute of Standards and Technology's voluntary AI Risk Management Framework (2023), organized around Govern, Map, Measure and Manage.
Also known as AI Risk Management Framework
Related AI GovernanceISO/IEC 42001
NPU
Neural Processing Unit: a small AI chip built into phones and laptops to run on-device AI efficiently.
Also known as Neural Processing Unit, neural engine
Related Edge AIAI Accelerator
O
Object Detection
Finding and labeling objects in an image or video with bounding boxes, e.g., counting products on a shelf.
Observability
OCR (Optical Character Recognition)
Converting images of text (scans, photos, PDFs) into machine-readable text. Modern vision-language models go further and understand layout and meaning.
Also known as OCR
Related Intelligent Document Processing (IDP)Vision-Language Model (VLM)
Opaque Recurrence
A reasoning technique where a model loops a query repeatedly through its internal layers instead of reasoning in readable, step-by-step text. Reported in OpenAI's Astra model (Sept 2026). Safety researchers warn it makes reasoning harder to monitor.
Also known as Recurrent depth
Related Latent ReasoningChain-of-Thought MonitoringNeuralese
Open-Source AI
AI released with weights plus code (and ideally data details) under terms that let anyone use, study, modify and share it. The Open Source Initiative published a formal definition in 2024.
Related Open-Weights ModelClosed Model
Open-Weights Model
A model whose trained weights are publicly released so anyone can download, run and modify it, even if training data and code aren't shared (e.g., Llama, Qwen, DeepSeek, Mistral).
Also known as Open model
Related Open-Source AIClosed Model
Optimizer
The algorithm that decides exactly how to update weights from computed gradients. Adam and AdamW are the common defaults for training LLMs.
Also known as Adam, AdamW, SGD
Related Gradient DescentLearning Rate
Orchestrator Agent
An agent that plans a task, splits it into subtasks, hands them to other agents (fan-out) and combines their results (fan-in).
Also known as Supervisor agent, orchestrator-worker
Related SubagentMulti-Agent System
Outcome-Based Pricing
Charging for AI products by results delivered (e.g., per resolved ticket or booked meeting) rather than per user seat.
Also known as Usage-based pricing, results-based pricing
Related AI AgentCost per Token
Overfitting
When a model memorizes its training data (including noise) and performs well on it but poorly on new data.
OWASP Top 10 for LLMs
A widely used list of the most critical security risks for LLM applications (prompt injection, sensitive data disclosure, excessive agency and more).
Related Prompt InjectionGuardrails
P
p(doom)
Informal shorthand for someone's estimated probability that AI leads to catastrophe for humanity.
Parameters
All the learned values (weights and biases) in a model. Model size is quoted in parameters, e.g., '70B' means 70 billion.
Parametric Knowledge
Facts a model absorbed into its weights during training, as opposed to information supplied in the prompt or retrieved at run time.
Related Knowledge CutoffRetrieval-Augmented Generation (RAG)Grounding
Pass@k
The probability that at least one of k attempts is correct. Common in coding benchmarks (pass@1 = right on the first try).
Related SWE-benchSelf-Consistency
Perplexity
A measure of how well a language model predicts text. Lower means less 'surprised' and better at modeling language.
Physical AI
AI that understands and acts in the physical world, such as robots, autonomous vehicles and smart factories. A term popularized by NVIDIA.
Also known as Embodied AI
Related RoboticsWorld Model
PII
Pilot Purgatory
When AI projects stall after promising pilots and never reach production, usually due to unclear ROI, poor data, integration hurdles or no clear owner.
Also known as POC purgatory
Related Data QualityEvals
PIPEDA
Canada's current federal private-sector privacy law, governing how businesses use personal data, including in AI. A replacement, the Protecting Privacy and Consumer Data Act, was tabled in June 2026.
Also known as Personal Information Protection and Electronic Documents Act
Policy
In reinforcement learning, the strategy an agent uses to choose actions in each situation. In LLM training, the model being optimized is often called the policy.
Positional Encoding
Information added to tokens so a transformer knows their order, since attention alone ignores position. RoPE is the common modern method.
Also known as RoPE (Rotary Position Embedding)
Related TransformerContext Window
Post-Training
Everything after pretraining (instruction tuning, RLHF, RL for reasoning, safety training) that turns a raw base model into a helpful, safe assistant. Now a major source of capability gains.
PPO (Proximal Policy Optimization)
A reinforcement learning algorithm that makes small, stable updates to the model. The classic choice for RLHF.
Also known as PPO
Precision and Recall
Precision: of the items the model flagged, how many were right. Recall: of all the items that should have been flagged, how many it found. Improving one often lowers the other.
Related F1 ScoreConfusion Matrix
Predictive Analytics
Using historical data and ML to forecast outcomes such as churn, demand or which deals will close.
Predictive Lead Scoring
Using ML to rank prospects by how likely they are to convert, based on firmographics, behavior and engagement signals.
Related Predictive AnalyticsAI SDR
Pretraining
The first and most expensive training stage: learning from massive general data (web text, books, code) via next-token prediction, which gives the model broad language and knowledge.
Prompt
The input you give an AI model, including instructions, questions, examples and data, to get a response.
Related Prompt EngineeringSystem Prompt
Prompt Caching
Reusing the already-processed form of a repeated prompt prefix (long instructions, documents) across requests, sharply cutting cost and latency.
Prompt Chaining
Breaking a task into a sequence of prompts, where each step's output feeds the next (e.g., research → outline → draft → edit).
Related AI WorkflowAgentic Loop
Prompt Engineering
Crafting and iterating prompts to get better, more reliable outputs from a model.
Prompt Injection
An attack where malicious instructions hidden in content an AI reads (web pages, emails, documents, tool results) hijack its behavior. The top security risk for AI agents.
Also known as Indirect prompt injection
Prompt Template
A reusable prompt with placeholders (e.g., {customer_name}) filled in automatically at run time.
Related Prompt ChainingSystem Prompt
Pruning
Removing weights or neurons that contribute little, to make a model smaller and faster.
Related QuantizationDistillation
Q
Quantization
Storing model weights at lower numeric precision (e.g., 8-bit or 4-bit instead of 16-bit) to cut memory and speed up inference, with small quality loss.
R
RAMageddon
Industry nickname for the ongoing shortage and price surge in memory (RAM) chips as AI data centers absorb supply, raising costs for phones, PCs and other devices.
Also known as Memory shortage
Random Forest
An ensemble of many decision trees, each trained on random slices of data, that vote on the answer. Robust and widely used.
Related Decision TreeEnsemble Learning
Rate Limit
A cap on how many requests or tokens per minute an API allows a customer. A common constraint when scaling AI apps.
Related APIModel Serving
ReAct
Reason + Act: an agent pattern where the model alternates between reasoning about what to do next and taking a tool action, using each result to decide the next step.
Reasoning Effort
A setting that controls how much a reasoning model thinks (how many reasoning tokens it may spend) before answering, trading speed and cost for quality.
Also known as Thinking budget, effort level
Related Test-Time ComputeReasoning Model
Reasoning Model
An LLM trained, mostly with reinforcement learning, to 'think' through an extended internal reasoning process before answering. Much stronger at math, coding and multi-step problems.
Also known as Thinking model, large reasoning model (LRM)
Recommendation System
AI that suggests products, content or next actions based on behavior and similarity (e.g., Netflix, Amazon, 'next best action' in CRM).
Also known as Recommender system
Related EmbeddingPredictive Analytics
Recurrent Neural Network (RNN)
A network that processes sequences one step at a time while carrying a memory of earlier steps. Largely replaced by transformers for language.
Also known as RNN
Related LSTMTransformer
Red Teaming
Deliberately attacking or stress-testing an AI system to uncover harmful behaviors and vulnerabilities before real users or attackers do.
Regression
Predicting a continuous number, such as house price, deal size or next quarter's revenue.
Regularization
Techniques that discourage overfitting by constraining the model, such as weight decay or dropout.
Related OverfittingDropout
Reinforcement Learning (RL)
Learning by trial and error: an agent takes actions, receives rewards or penalties, and gradually learns a strategy that maximizes reward. Now central to training reasoning models.
Also known as RL
Related RLHFReward ModelPolicy
Reranking
A second, more accurate scoring pass over retrieved results to put the most relevant ones first before they go to the model.
Also known as Reranker, cross-encoder
Residual Connection
A shortcut that adds a layer's input to its output, making very deep networks stable and trainable.
Also known as Skip connection
Related TransformerLayer Normalization
Responsible AI
Principles and practices for building and using AI that is fair, transparent, accountable, safe and privacy-respecting.
Also known as Trustworthy AI, ethical AI
Related AI GovernanceAlgorithmic Bias
Responsible Scaling Policy (RSP)
A lab commitment (pioneered by Anthropic) that ties progressively stronger safety measures to specific capability thresholds. OpenAI and Google DeepMind have similar frameworks.
Also known as Frontier safety framework, preparedness framework
Retrieval-Augmented Generation (RAG)
Fetching relevant documents from a knowledge source and adding them to the prompt so the model answers with current, specific, citable information instead of memory alone.
Also known as RAG
Related Vector DatabaseGroundingChunking
Reward Hacking
When a model finds loopholes to score well on its reward signal without doing the intended task, e.g., editing tests so they pass instead of fixing code.
Also known as Specification gaming
Related Reward ModelRLVR
Reward Model
A model trained to score outputs by how much humans (or a rubric) would prefer them. Used to steer RL training.
Related RLHFReward Hacking
RLAIF
Reinforcement Learning from AI Feedback: like RLHF, but an AI model (often guided by written principles) supplies the preference judgments instead of humans.
Also known as Reinforcement Learning from AI Feedback
Related RLHFConstitutional AI
RLHF
Reinforcement Learning from Human Feedback: people rank model responses, a reward model learns those preferences, and the LLM is trained to produce preferred answers. Made ChatGPT-style assistants possible.
Also known as Reinforcement Learning from Human Feedback
Related Reward ModelRLAIFDPO (Direct Preference Optimization)
RLVR
Reinforcement Learning with Verifiable Rewards: RL where the reward comes from automatically checkable outcomes, like a correct math answer or passing unit tests. The key recipe behind reasoning models.
Also known as Reinforcement Learning with Verifiable Rewards
Related Reasoning ModelGRPO (Group Relative Policy Optimization)Reward Hacking
Robotic Process Automation (RPA)
Software bots that automate repetitive, rule-based computer tasks by mimicking clicks and keystrokes. Increasingly paired with or replaced by AI agents.
Also known as RPA
Related HyperautomationComputer Use
Robotics
Machines that sense and act in the physical world. AI is increasingly used to give robots general-purpose perception and skills rather than hand-programmed routines.
Also known as Embodied AI
Role Prompting
Assigning the model a persona or role (e.g., 'You are a CFO reviewing this budget') to shape its focus, vocabulary and tone.
Also known as Persona prompting
Related System PromptPrompt Engineering
S
Sampling
How the next token is picked from the model's probabilities: always the top choice (greedy), or randomly weighted by probability (controlled by temperature and top-p).
Also known as Decoding, greedy decoding
Related TemperatureLogits
Sandbagging
A model deliberately underperforming, for example on dangerous-capability tests, to appear less capable than it is.
Sandbox
An isolated environment where an agent can run code or take actions without affecting real systems or data.
Also known as Sandboxed execution
Related Coding AgentPrompt Injection
Scalable Oversight
Methods that let humans supervise AI on tasks too complex for them to check directly, such as AI-assisted review or having models debate each other.
Related AlignmentLLM-as-a-Judge
Scaling Hypothesis
The belief that scaling up models, data and compute will keep producing more capable, eventually general, AI. Its limits are one of the field's central debates.
Scaling Laws
The empirical finding that model performance improves predictably as you increase parameters, training data and compute. Guides how labs budget training runs.
Scheming
The risk that a model covertly pursues goals different from its developers' while appearing aligned, e.g., behaving well when it thinks it's being tested. Observed in controlled lab experiments.
Also known as Deceptive alignment, alignment faking
Related SandbaggingAI Control
Self-Attention
Attention in which every token in a sequence looks at every other token in the same sequence to build context-aware representations.
Self-Consistency
Generating several independent reasoning paths and choosing the answer most of them agree on.
Self-Play
Training where an AI improves by competing against copies of itself, generating ever-harder practice. Key to AlphaGo Zero.
Self-Supervised Learning
Learning where the model creates its own training signal from raw data, such as predicting a hidden or next word. It's how LLMs learn from huge amounts of unlabeled text.
Related PretrainingNext-Token Prediction
Semantic Search
Search by meaning rather than exact keywords, using embeddings, so 'cancel my plan' can match 'how to end a subscription'.
Related EmbeddingHybrid Search
Semi-Supervised Learning
Training that combines a small labeled dataset with a large unlabeled one, useful when labels are expensive.
Related Supervised LearningData Labeling
Sentiment Analysis
Detecting emotional tone (positive, negative, neutral) in text or speech. Common for reviews, support tickets and sales calls.
Related Natural Language Processing (NLP)Conversation Intelligence
Shadow AI
Employees using unapproved AI tools at work, creating data security, privacy and compliance risks.
Related AI GovernancePII
Singularity
A hypothetical point where AI-driven progress becomes so fast that the future is unpredictable and beyond human control or understanding.
Also known as Technological singularity
Related Intelligence ExplosionArtificial Superintelligence (ASI)
Slop
Low-quality, mass-produced AI-generated content that's generic, hollow or unwanted.
Also known as AI slop
Related WorkslopModel Collapse
Small Language Model (SLM)
A compact language model (typically a few billion parameters or fewer) cheap enough to run on phones, laptops or edge devices, often tuned for specific tasks.
Also known as SLM
Related Edge AIDistillationQuantization
Sovereign AI
A country's effort to build its own AI infrastructure, models and data capacity instead of relying on foreign providers. A stated priority for Canada.
Related Data CenterExport Controls
Sparse Autoencoder (SAE)
An interpretability tool that decomposes a model's internal activity into many distinct, human-readable features (e.g., 'deception', 'Golden Gate Bridge').
Also known as SAE, dictionary learning
Related Mechanistic Interpretability
Speculative Decoding
A small, fast 'draft' model guesses several tokens ahead and the large model checks them all at once, speeding generation without changing the output.
Related ThroughputLatency
Speech Recognition
Converting spoken audio into text, e.g., meeting transcription or voice commands.
Also known as ASR, speech-to-text
Related Text-to-Speech (TTS)Voice Agent
State Space Model (SSM)
An alternative to transformers (e.g., Mamba) that processes sequences in linear time, making very long inputs cheaper. Often mixed with attention in hybrid models.
Also known as Mamba, hybrid architecture
Related TransformerLong Context
Stochastic Parrot
A critical term (Bender et al., 2021) for the argument that LLMs stitch together language patterns without real understanding. Still actively debated.
Streaming
Sending a response token by token as it's generated instead of waiting for the whole answer, so users see progress immediately.
Related Time to First Token (TTFT)
Structured Output
Structured vs. Unstructured Data
Structured data fits rows and columns (CRM fields, spreadsheets). Unstructured data is text, email, PDFs, calls and images. Most enterprise data, and where LLMs add the most value.
Related Intelligent Document Processing (IDP)Retrieval-Augmented Generation (RAG)
Subagent
An agent launched by another agent to handle a focused subtask in its own clean context, returning only the results.
Supervised Fine-Tuning (SFT)
Fine-tuning on curated examples of prompts paired with ideal responses, teaching the model what good answers look like.
Also known as SFT
Related Instruction TuningPost-Training
Supervised Learning
Training on labeled examples (input plus correct answer) so the model learns to predict the answer for new inputs, e.g., emails labeled spam / not spam.
Related LabelClassificationRegression
Support Vector Machine (SVM)
A classic classifier that finds the boundary separating classes with the widest possible margin.
Also known as SVM
Related Classification
SWE-bench
A benchmark where models must resolve real GitHub issues in real code repositories. SWE-bench Verified is the human-validated subset widely used to compare coding ability.
Also known as SWE-bench Verified
Related Coding AgentPass@k
Sycophancy
A model's tendency to tell users what they want to hear: agreeing, flattering, or abandoning a correct answer when pushed back on.
Related RLHFAutomation Bias
Symbolic AI
The early approach to AI built on hand-written rules and logic rather than learning from data. Dominant from the 1950s to the 1980s.
Also known as GOFAI (Good Old-Fashioned AI), rule-based AI
Related Expert SystemNeuro-Symbolic AI
Synthetic Data
Artificially generated data, often produced by AI models, used to train or test models when real data is scarce, costly or sensitive.
System Prompt
Instructions set by the developer (not the end user) that define a model's role, rules and behavior for a whole conversation.
Also known as System message, developer message
Related PromptPrompt Injection
T
Takeoff
How fast AI goes from roughly human-level to far beyond: a fast ('hard') takeoff over days or months vs. a slow ('soft') one over years.
Also known as Fast vs. slow takeoff
Task Horizon
A way to measure AI progress by the length of tasks (in human working time) an agent can complete with 50% reliability. METR found this has been doubling roughly every 7 months.
Also known as METR time horizon
Related Long-Horizon TaskAI Timelines
Temperature
A setting that controls randomness: low temperature gives focused, predictable output. High temperature gives more varied, creative output.
Related Top-p (Nucleus Sampling)Sampling
Tensor
Test Set
Data held back from training and used only to measure final performance fairly. A validation set is a separate held-out slice used for tuning during development.
Also known as Holdout set, validation set
Related OverfittingCross-Validation
Test-Time Compute
Spending more computation at answer time (thinking longer, trying multiple answers, verifying) to get better results. A major scaling lever alongside bigger training runs.
Also known as Inference-time scaling
Related Reasoning ModelReasoning Effort
Text-to-Image
Generating images from a written description, typically with diffusion models.
Also known as Image generation
Related Diffusion ModelMultimodal Model
Text-to-Speech (TTS)
Generating natural-sounding spoken audio from written text.
Also known as TTS, speech synthesis
Related Voice CloningVoice Agent
Text-to-Video
Generating video clips, increasingly with synchronized sound, from written prompts or reference images.
Also known as Video generation
Throughput
Time to First Token (TTFT)
Token
The basic unit of text a model reads and writes, often a word fragment. In English, one token is roughly 3–4 characters or about ¾ of a word. Usage and pricing are counted in tokens.
Tokenizer
The component that splits text into tokens (often via byte-pair encoding). It affects cost, speed and how efficiently different languages are handled.
Also known as Tokenization, BPE (byte-pair encoding)
Related Token
Tool Poisoning
An attack where malicious instructions are hidden in a tool's description or output (e.g., an MCP server) to manipulate the agent using it.
Related Prompt InjectionMCP Server
Tool Use
A model's ability to call external tools (search, calculators, APIs, databases, code) by producing structured requests that software executes and returns results from.
Also known as Function calling, tool calling
Top-p (Nucleus Sampling)
A setting that limits the model to choosing among the most likely tokens whose probabilities add up to p (e.g., 0.9), cutting off unlikely words. Top-k is a similar cap by count.
Also known as Nucleus sampling, top-k
Related TemperatureSampling
TPU
Tensor Processing Unit: Google's custom AI chip, used to train and serve Gemini and rented to cloud customers.
Also known as Tensor Processing Unit
Related AI AcceleratorGPU
Training
The process of adjusting a model's weights using data so its outputs get better. Distinct from inference, which is using the finished model.
Related PretrainingFine-TuningInference
Training Data
The dataset a model learns from. Its size, quality, diversity and legality largely determine what the model can do and what biases it inherits.
Training Data Copyright
The legal fight over whether training AI on copyrighted books, art, music and news without permission is lawful (e.g., under U.S. fair use). The subject of many lawsuits, settlements and licensing deals.
Also known as AI copyright, fair use
Related Web CrawlingMachine Unlearning
Transfer Learning
Reusing a model trained on one task as the starting point for another, so the new task needs far less data and compute. Fine-tuning an LLM is a form of this.
Related Fine-TuningFoundation Model
Transformative AI (TAI)
AI that triggers change comparable to the Industrial Revolution. A term some forecasters prefer over the vaguer 'AGI'.
Also known as TAI
Transformer
The neural network architecture, introduced in the 2017 paper 'Attention Is All You Need', that uses attention to process all tokens in parallel. It underpins nearly every modern LLM.
Tree of Thoughts
A technique where the model explores several reasoning branches, evaluates them, and backtracks from dead ends, like a search through possible solutions.
Also known as ToT
Turing Test
Alan Turing's 1950 test: if a human judge can't reliably tell a machine from a person in text conversation, the machine passes. Modern LLMs often pass informal versions, so it's no longer seen as a meaningful bar.
Also known as Imitation Game
U
Underfitting
When a model is too simple to capture the real pattern, so it performs poorly even on training data.
Unsupervised Learning
Finding structure in unlabeled data, such as grouping similar customers, without being told the right answers.
V
Vector
Vector Database
A database that stores embeddings and quickly finds items with similar meaning. A common backbone for RAG (e.g., Pinecone, pgvector, Weaviate).
Also known as Vector store
Related EmbeddingSemantic Search
Vibe Coding
Building software by describing what you want in plain language and accepting the AI's code with little review. Coined by Andrej Karpathy in 2025. Fast for prototypes, risky for production.
Related Coding AgentCode Generation
Vibes-Based Eval
Judging a model informally by how it 'feels' after a few chats instead of systematic testing. Common, but unreliable for business decisions.
Also known as Vibe check
Related EvalsGolden Dataset
Vision Transformer (ViT)
A transformer applied to images by cutting them into patches and treating each patch like a token.
Also known as ViT
Vision-Language Model (VLM)
A model that understands both images and text, so it can answer questions about photos, charts, documents and screenshots.
Also known as VLM
Related Multimodal ModelComputer Use
Vision-Language-Action Model (VLA)
A model that takes camera images and language instructions and outputs robot actions, letting robots follow commands like 'put the cup in the sink'.
Also known as VLA
Voice Agent
An AI that holds real-time spoken conversations, often processing audio natively (speech-to-speech) for natural, low-latency dialogue. Used in customer service and sales.
Also known as Voice AI, speech-to-speech model
Related Speech RecognitionAI Agent
Voice Cloning
Generating synthetic speech that mimics a specific person's voice from a short sample. Useful for accessibility and dubbing, but a fraud risk.
Related DeepfakeText-to-Speech (TTS)
W
Watermarking
Embedding hidden signals in AI-generated text, images or audio so they can later be identified as AI-made (e.g., Google's SynthID).
Web Crawling
Automatically collecting pages from the internet. The source of most pretraining text and a focus of copyright and consent disputes.
Also known as Web scraping
Weights
The learned numbers that set how strongly each connection influences the next. Everything a model 'knows' is stored in its weights.
Workslop
AI-generated work that looks polished but lacks substance, shifting the effort of fixing it to colleagues. Term from 2025 Stanford/BetterUp research.
Related SlopAutomation Bias
World Model
An AI's internal model of how an environment works, used to predict what happens next. Also used for systems that generate interactive, explorable simulated worlds.
Related Physical AIRoboticsText-to-Video
Z
Zero-Shot Prompting
Asking a model to perform a task with no examples, relying only on instructions.
Also known as Zero-shot
No terms match that yet. Try another word, or clear the filters.