AI Detection
Tools and techniques for identifying whether content was generated by AI rather than humans. Includes statistical analysis, watermark detection, and classifier-based approaches. Accuracy remains an active research challenge.
AI Legislation
Laws and regulations governing AI development and use (e.g., EU AI Act, data protection laws, sector rules).
AI Literacy
The skills, knowledge, and understanding people need to use, operate, or oversee AI systems responsibly in their context. Under the EU AI Act, providers and deployers must take measures to ensure sufficient AI literacy for staff and others using AI systems on their behalf.
AI Management System (AIMS)
An organizational management system for establishing policies, objectives, controls, and processes for the responsible development, provision, or use of AI systems. ISO/IEC 42001 defines requirements for creating, maintaining, and improving an AIMS.
AI Orchestration
Coordinating multiple AI components (models, retrievers, tools, guardrails) into a coherent pipeline. Includes routing, fallbacks, error handling, and ensuring components work together reliably.
AI Risk Management Framework (AI RMF)
NIST framework for managing AI risk through four functions: Govern, Map, Measure, and Manage, across the full AI lifecycle.
AI Slop
Low-quality AI-generated content that prioritizes volume over usefulness, accuracy, originality, or editorial care. The term is often used for mass-produced pages, images, posts, or summaries that add little value and may pollute search results or user feeds.
AI System
A machine-based system that infers from input data how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. AI systems can differ in autonomy and may adapt after deployment.
AI Watermarking
Alias for watermarking signals embedded in AI‑generated outputs to indicate AI origin.
AI‑as‑a‑Service
Cloud delivery model where AI capabilities (APIs, hosted models) are provided on demand without managing infrastructure.
Alan Turing
Pioneer of computer science and AI; proposed the Turing Test and foundational ideas in computation.
Alert Fatigue
Operational overload caused by excessive or low-quality alerts, leading to slower responses and missed critical incidents.
Algorithm
A set of rules or instructions given to an AI or computer system to help it learn, solve problems, or make decisions. It is a foundational concept in computer science.
ALiBi (Attention with Linear Biases)
A positional bias method enabling better extrapolation to longer contexts by adding linear biases to attention scores.
Alignment (AI Alignment)
Designing AI systems so their behavior is consistent with human values, safety constraints, and intended goals. In practice, this includes policy design, reinforcement learning from human feedback (RLHF), and rigorous evaluations.
Allowlist
A list of explicitly permitted items (e.g., allowed tools, domains, actions). All other items are denied by default. Often used to reduce tool misuse and data exfiltration.
Andrew Ng
AI researcher and educator known for online ML courses and industrial AI adoption frameworks.
Anomaly Detection (Security)
Identifying outliers that may indicate intrusions, fraud, or system compromise using statistical or ML methods.
API (Application Programming Interface)
A way for different software programs to communicate with each other. Many AI tools offer APIs so developers can integrate AI capabilities into their own applications.
API Key
A unique authentication token that identifies and authorizes your application to use an AI service's API. API keys should be kept secret and never exposed in client-side code. They're used to track usage, enforce rate limits, and bill for API calls.
Approval Workflow
Defined sequence of review and sign-off steps required before deploying high-impact model or infrastructure changes.
Approximate Nearest Neighbor (ANN)
A family of algorithms and indexes (e.g., HNSW, IVF) used to rapidly find vectors that are most similar to a query vector in high‑dimensional spaces. Core to fast vector search in embedding‑based applications.
Artificial General Intelligence (AGI)
A hypothetical type of AI that possesses the ability to understand, learn, and apply knowledge across a wide range of tasks at a human-like level. This is distinct from current AI, which is typically specialized (Narrow AI).
Artificial Intelligence (AI)
The field of building machine-based systems that can infer how to generate outputs such as predictions, content, recommendations, or decisions. Modern AI systems vary in autonomy, adaptiveness, and risk depending on where and how they are deployed.
Attention Mask
A mask applied during attention to block or down-weight certain positions (e.g., padding tokens, future tokens in causal masking).
Attention Mechanism
A key component in Transformer models (like those used in LLMs) that allows the model to weigh the importance of different parts of the input sequence when processing information, crucial for understanding context and long-range dependencies.
Audit Logging
Tamper-resistant logs capturing who did what and when (requests, tool calls, data access). Critical for compliance, incident response, and forensic analysis.
Audit Trail
Tamper-resistant record of system and user actions used for security investigations, compliance evidence, and operational forensics.
Automatic Speech Recognition (ASR)
Technology that enables computers to convert spoken language into written text. It's the core of voice assistants and dictation software.
AWQ (Activation-aware Weight Quantization)
A quantization technique that preserves model quality by considering activation outliers when quantizing weights.
Backpressure
Flow‑control strategy that slows producers when consumers or downstream systems are saturated to prevent overload.
Backpropagation
Algorithm for computing gradients in neural networks by propagating errors backward from the output layer to input layer, enabling weight updates.
Batch Normalization
Technique that normalizes layer inputs during training, reducing internal covariate shift and improving training stability and speed.
Batch Processing
Processing multiple requests together in a single API call or job, typically with lower priority but reduced cost. Useful for non-time-sensitive tasks like bulk data analysis, content moderation, or embedding generation. Can be 50% cheaper than real-time processing.
Beam Search
A decoding strategy that keeps the top k partial sequences at each step to approximate the most likely output.
Benchmark (AI Benchmark)
A standardized test or dataset used to evaluate model quality, robustness, and performance. Examples include MMLU, HELM, and custom task‑specific evals.
Benchmark Drift
Shifts in measured performance due to dataset changes, model updates, or prompt/pipeline modifications.
BERT
Bidirectional Encoder Representations from Transformers - a pre-trained language model that understands context from both directions (left-to-right and right-to-left), excellent for comprehension tasks.
BF16 (BFloat16)
16-bit floating point format with a wider exponent range than FP16, often improving training stability while retaining performance benefits.
Bi‑encoder
A dual‑tower architecture that encodes query and document separately to enable fast vector similarity search.
Bias (in AI)
Systematic errors in AI output that result from prejudices in the training data or algorithmic design. AI bias can lead to unfair or discriminatory outcomes against certain groups.
BLEU Score
Bilingual Evaluation Understudy - a metric for evaluating machine translation quality by comparing generated text to reference translations. Measures n-gram overlap. Scores range from 0 to 1 (or 0-100). Higher is better, but BLEU has limitations for creative or diverse outputs.
Blue‑Green Deployment
Two parallel environments (blue/green); traffic switches to green after verification, enabling quick rollback.
BM25
A classical lexical ranking function used in information retrieval that scores documents based on term frequency and inverse document frequency. Often combined with embeddings for hybrid search.
Bulkhead Pattern
Isolation pattern that partitions resources (threads, pools, services) so failures in one domain do not exhaust the entire system.
Burn Rate (SLO)
Speed at which a service consumes its error budget relative to SLO targets; high burn rate signals urgent reliability risk.
Business Continuity Plan (BCP)
Documented strategy for maintaining critical business operations during and after disruptions, including roles, communication plans, and recovery priorities.
Byte Pair Encoding (BPE)
A subword tokenization algorithm that builds a vocabulary by iteratively merging common character pairs. BPE tokenizers balance vocabulary size with the ability to represent rare words.
Calibration (Model Confidence)
How well a model’s confidence aligns with actual correctness. A well-calibrated model assigns higher probability to correct answers more often.
Canary Release
Gradual rollout to a subset of users to validate stability and performance before full deployment.
Capacity Planning
Process of forecasting resource demand and provisioning compute, storage, and concurrency to meet performance and reliability targets.
Causal Masking
An attention constraint used in decoder-only transformers so each token can only attend to earlier tokens (not future tokens). This enables autoregressive generation.
Chain of Thought (CoT)
A research prompting technique that elicits intermediate reasoning for complex tasks. For user-facing products, prefer concise rationale, verifiable steps, or final-answer explanations instead of exposing full hidden reasoning traces.
Chain-of-Thought Prompting
A research technique that elicits intermediate reasoning before a final answer. In deployed products, use concise explanations, checks, or rationale summaries instead of exposing full hidden reasoning.
Champion-Challenger
Evaluation pattern where a production champion model is continuously compared against challenger models to determine if a safer or better alternative should be promoted.
Change Failure Rate
Percentage of production changes that result in incidents, rollbacks, hotfixes, or degraded service.
Change Management
Structured process for assessing, approving, implementing, and reviewing production changes to reduce risk and service disruption.
Chaos Engineering
Practice of injecting failures in production‑like environments to validate resilience, alerts, and recovery procedures.
Chatbot
An AI program designed to simulate human conversation through text or voice. Used for customer service, information retrieval, and companionship.
Checkpoint
A saved snapshot of a model’s weights (and sometimes optimizer state) during or after training, enabling resume, fine-tuning, and deployment.
Circuit Breaker
Pattern that trips to fail fast when downstreams are unhealthy, preventing resource exhaustion and cascading failures.
Citations
References to sources used in research or content creation, often formatted according to a specific style guide.
Classification (in ML)
A supervised machine learning task where the model learns to assign a predefined category or label to a given input. For example, classifying an email as 'spam' or 'not spam', or identifying an image as containing a 'cat' or a 'dog'.
Classifier-Free Guidance (CFG)
A conditioning technique for diffusion models that increases prompt adherence by combining conditional and unconditional predictions during sampling. Higher guidance increases alignment but may reduce diversity.
CLIP
Contrastive Language-Image Pretraining - a multimodal model that understands both images and text, enabling powerful vision-language applications.
Clustering
An unsupervised machine learning task where the model groups a set of unlabeled data points in such a way that objects in the same group (called a cluster) are more similar to each other than to those in other clusters. It's used to discover hidden patterns or structures in data.
ColBERT
A late‑interaction retrieval model that allows fine‑grained token‑level matching while remaining efficient.
Cold Start
Initial request penalty while provisioning resources (functions, models, containers). Mitigate via warm pools, prewarming, and smaller artifacts.
Compliance Drift
Gradual divergence between required controls and actual system state caused by untracked changes, environment sprawl, or weak governance automation.
Compound AI Systems
AI applications that combine multiple components (LLMs, retrievers, tools, classifiers, guardrails) working together rather than relying on a single model. Most production AI systems are compound systems.
Compute Budget
The total amount of computational resources (GPU hours, FLOPs) allocated for training or inference. Compute budget decisions are guided by scaling laws to optimize performance-cost tradeoffs.
Computer Vision
A field of AI that enables computers to 'see' and interpret visual information from the world, such as images and videos, allowing them to identify objects, faces, and scenes.
Concept Drift
Changes in the relationship between inputs and target outputs over time (the underlying concept changes). Even if input distributions look similar, correctness can degrade.
Confusion Matrix
Table showing the performance of a classification model by displaying true positives, true negatives, false positives, and false negatives.
Consent Management
Processes and systems used to collect, record, update, and enforce user consent preferences for personal data processing.
Constitutional AI
An approach to AI alignment where models are trained using a 'constitution'—a set of principles and rules that guide behavior. Instead of relying solely on human feedback, the model critiques and refines its own outputs against these principles. Developed by Anthropic for Claude.
Content Authenticity
Practices and technical signals used to show where digital content came from, how it was created, and whether it was edited. In AI workflows this can include provenance metadata, watermarking, human review notes, and source citations.
Content Moderation
Filtering unsafe or policy‑violating content using classifiers, rules, or human review.
Context Caching
A technique that reuses previously computed attention/key‑value states for repeated prefixes, reducing latency and cost in long or iterative prompts. Particularly useful for multi-turn conversations and repeated system prompts.
Context Compression
Techniques that reduce prompt length (summarization, distillation, selection) to fit context limits and lower cost.
Context Engineering
The discipline of designing what context an AI system receives at inference time: system prompts, retrieved documents, tool outputs, memory, and user constraints. Strong context engineering improves relevance, factuality, and consistency without retraining the base model.
Context Overlap
The amount of text shared between consecutive chunks to preserve continuity across boundaries during retrieval.
Context Window
The maximum amount of input and output text, measured in tokens, that a model can handle in one request. A larger context window can help with long documents and multi-turn tasks, but retrieval quality, attention behavior, latency, and cost still matter.
Continuous Batching
Serving technique where new requests can be added to an ongoing batch between decoding steps, improving throughput and GPU utilization.
Contrastive Learning
A self-supervised learning approach that learns representations by contrasting similar (positive) and dissimilar (negative) pairs. The model learns to pull similar items closer and push dissimilar items apart in embedding space.
Control Objective
A specific security, reliability, or compliance outcome that controls are designed to achieve and verify.
ControlNet
A technique/module that adds additional conditioning inputs (edges, depth, pose, segmentation) to control image generation in diffusion models.
Coreference Resolution
NLP task that identifies when different words refer to the same entity (e.g., 'John' and 'he' refer to the same person).
Cosine Similarity
Measure of similarity between two vectors by calculating the cosine of the angle between them, commonly used for text embeddings.
Cost per Token
Granular cost metric for inference; optimize via prompt compression, caching, batching, and model choice.
Cross-encoder
A model that scores a query–document pair jointly by encoding them together, often used for high-precision reranking.
Cross‑attention
Attention mechanism where one sequence attends to another, used in encoders‑decoders and multimodal models.
Cross‑Entropy Loss
The standard loss function for language modeling and classification that measures how well predicted probabilities match the true labels. Minimizing cross‑entropy is equivalent to maximizing log-likelihood.
CUDA
NVIDIA’s GPU computing platform and programming model. Many ML libraries and kernels (attention, GEMM) are optimized for CUDA GPUs.
Curriculum Learning
A training strategy that presents examples in a meaningful order (e.g., easy to hard, short to long) rather than randomly. Can improve training efficiency and final model quality.
DALL-E
OpenAI's text-to-image generation model that creates realistic images from textual descriptions using diffusion models.
Data Augmentation
A technique used to increase the diversity and size of a training dataset by creating modified copies of existing data or generating new synthetic data. For images, this might include rotating, cropping, or changing the brightness. This helps improve model performance and reduce overfitting.
Data Classification
Process of labeling data by sensitivity and handling requirements to apply appropriate security, privacy, and access controls.
Data Deletion
Deleting stored user or customer data (including logs and derived artifacts) on request or after retention periods. Requires identifying downstream copies and backups.
Data Drift
Changes in the input data distribution over time (e.g., new topics, user behavior shifts). Data drift can degrade retrieval quality and model performance.
Data Exfiltration (LLM Security)
An attack goal where a model is manipulated into leaking sensitive information (secrets, system prompts, private documents) via tool use or outputs.
Data Flywheel
A virtuous cycle where a deployed AI system generates data (user interactions, feedback, corrections) that improves future model versions, which in turn attracts more users and data.
Data Governance
Framework of roles, policies, standards, and controls that ensures data is managed consistently for quality, security, compliance, and business value.
Data Minimization
Privacy principle to collect, process, and retain only data necessary for a specific purpose.
Data Parallelism
Training approach where model replicas process different data shards in parallel and synchronize gradients.
Data Processing Agreement (DPA)
A vendor contract that defines personal data processing terms; often used interchangeably with a Data Processing Addendum (DPA). Check scope, sub‑processors, transfers, and security controls under GDPR.
Data Protection Impact Assessment (DPIA)
Structured assessment required for high-risk personal-data processing to evaluate necessity, proportionality, and mitigation of risks to individuals.
Data Residency
The requirement to store and process data within specific geographic regions for compliance, contracts, or latency.
Data Retention
Policies defining how long data (prompts, logs, embeddings, outputs) is stored. Shorter retention reduces privacy risk but may limit debugging and compliance needs.
Data Retention Policy
Rules defining how long data is stored, when it is archived, and when it is securely deleted based on legal and business requirements.
Data Set
A collection of data (e.g., images, text, numbers) used to train or evaluate an AI model. The quality, size, and diversity of the data set are critical for model performance.
Data Subject Access Request (DSAR)
Request from an individual to access their personal data and related processing information held by an organization.
Data Versioning
Tracking versions of datasets and data snapshots used for training/evaluation so results are reproducible and changes are auditable.
DDIM (Denoising Diffusion Implicit Models)
A sampling method for diffusion models that can reduce the number of denoising steps and enable faster generation while maintaining quality.
DDPM (Denoising Diffusion Probabilistic Models)
A foundational diffusion formulation that learns to reverse a forward noising process step-by-step. DDPM defines the training objective and sampling process used by many diffusion models.
Deep Learning (DL)
A subfield of Machine Learning that uses artificial neural networks with multiple layers ('deep' architectures) to analyze complex patterns in large datasets. It's particularly effective for tasks like image recognition and natural language processing.
Deepfake
AI-generated synthetic media (video, audio, images) that convincingly depicts people saying or doing things they never did. Created using deep learning techniques like GANs or diffusion models. Raises significant ethical and security concerns.
Dense Model
A model architecture where all parameters are used for every input. Traditional neural networks are dense. Contrasts with sparse/MoE architectures where only subsets of parameters are active.
Denylist
A list of explicitly forbidden items (e.g., blocked domains, commands, tool actions). Used alongside allowlists for layered safety controls.
Dependency Confusion
Supply-chain attack where malicious public packages mimic private dependency names so build systems install attacker-controlled code.
Dependency Parsing
NLP technique that analyzes grammatical structure of sentences by identifying relationships between words (subject-verb, modifier-head, etc.).
Deterministic Mode
A mode of operation where the output is entirely determined by the input, without any randomness or variability.
Deterministic vs Non‑deterministic Outputs
Deterministic outputs are reproducible given the same prompt and parameters (e.g., temperature=0). Non‑deterministic outputs vary due to sampling. Teams choose based on creativity vs. repeatability.
Diarization (Speaker Diarization)
Partitioning audio into speaker‑homogeneous segments (who spoke when).
Differential Privacy
A mathematical framework for privacy-preserving data analysis that adds calibrated noise to prevent individual data points from being identified.
Diffusion Model
A type of generative AI model that creates data (often images) by learning to reverse a gradual noising process. It starts with random noise and refines it step-by-step into a coherent output, often guided by a text prompt (e.g., Stable Diffusion, GPT-Image).
Direct Preference Optimization (DPO)
A fine-tuning method that learns directly from preference pairs (chosen vs rejected) without an explicit reward model or reinforcement learning loop. DPO is often simpler than RLHF while achieving strong alignment results for instruction-following behavior.
Disaster Recovery (DR)
Practices and tooling to restore service after major failures (region loss, data corruption). Involves backups, replication, and failover.
DLP (Data Loss Prevention)
Policies and systems that detect and prevent sensitive data from leaving allowed boundaries (e.g., blocking secrets/PII from being sent to external tools or vendors).
DORA Metrics
A set of software delivery and reliability metrics commonly used to assess engineering performance: deployment frequency, lead time for changes, change failure rate, and MTTR.
DP
Alias for Differential Privacy.
DPR (Dense Passage Retrieval)
A bi‑encoder approach that learns dense embeddings for questions and passages to improve open‑domain QA retrieval.
Drift Detection
Techniques used to identify statistically significant shifts in data, labels, or model behavior that may reduce production reliability or accuracy.
Early Stopping
Technique to prevent overfitting by stopping training when validation performance stops improving.
ECE (Expected Calibration Error)
A metric that measures miscalibration by comparing predicted confidence to empirical accuracy across probability bins.
Edge AI
Running AI models directly on local devices (smartphones, IoT devices, edge servers) rather than in the cloud. Offers benefits like lower latency, better privacy, offline capability, and reduced bandwidth costs. Requires optimized models through quantization and distillation.
Embeddings
Numerical representations (vectors) of words, sentences, or other data types in a multi-dimensional space. AI models use embeddings to understand semantic relationships and similarities between data points, enabling tasks like semantic search or text classification.
Emergent Abilities
Capabilities that appear in large models but are absent in smaller ones, seemingly arising unpredictably at certain scales. Examples include in-context learning, chain-of-thought reasoning, and instruction following.
Encoder-Decoder Architecture
Neural network architecture with two main components: encoder processes input into a compressed representation, decoder generates output from that representation. Used in translation and summarization.
Ensemble Learning
Method that combines multiple machine learning models to improve overall performance and robustness.
Entity Linking
Process of connecting named entities in text to unique identifiers in a knowledge base (e.g., linking 'Paris' to the city vs. the person).
Episodic Memory (Agents)
Memory of past interactions/events (what happened in previous sessions). Used to personalize behavior and avoid repeating mistakes.
Error Budget
Allowable unreliability derived from SLOs. Consumed by incidents; gates release velocity and risk.
Error Rate
Proportion of failed requests or operations over total requests in a time window, used as a primary service health indicator.
Escalation Policy
Predefined routing and ownership rules that determine who is paged, when incidents are escalated, and how handoffs occur across severity levels.
Ethical AI
A branch of ethics focused on the moral implications of AI. It addresses issues like fairness, accountability, transparency, privacy, and the societal impact of AI technologies to ensure responsible development and deployment.
EU AI Act
A European Union regulation establishing a risk-based framework for AI systems, with obligations for high-risk uses (e.g., documentation, risk management, monitoring) and rules for certain prohibited practices.
Evals (Model Evaluation)
Task‑specific tests that measure quality, robustness, and safety of model outputs on real workloads. Strong evals guide model, prompt, and guardrail choices.
Evaluation Harness
Automated tests and datasets to assess quality, safety, latency, and cost across model versions.
Evasion Attacks
Adversarial attacks that manipulate input data during inference to avoid detection or change model predictions.
Exception Management
Governance process for granting, documenting, approving, and expiring temporary policy exceptions with compensating controls.
Excessive Agency
A vulnerability where an LLM-powered system is allowed to take damaging actions because its tools, permissions, autonomy, or approval gates are too broad. The risk increases when agents can call external systems, write data, spend money, or trigger workflows without strict constraints.
Experiment Tracking
Logging runs, hyperparameters, datasets, metrics, and artifacts to reproduce results and compare experiments across model versions.
Explainable AI (XAI)
A set of methods and techniques in AI aimed at making the decisions and predictions made by AI models, especially complex ones like deep neural networks, understandable and interpretable to humans. This helps build trust and allows for debugging.
Exponential Backoff
Retry delays grow exponentially between attempts, often with jitter, to reduce load during failures.
F1-Score
Harmonic mean of precision and recall, providing a single metric that balances both measures.
FAISS
Facebook AI Similarity Search — a library for efficient vector similarity search and clustering at scale.
Faithfulness (RAG)
An evaluation dimension measuring whether generated statements are supported by the provided context. Faithfulness focuses on avoiding unsupported claims.
Fallback
A reliability pattern where the system degrades gracefully by switching to a simpler or more stable behavior (e.g., a cheaper model, cached answer, or human handoff) when errors occur.
Feature Flags
Runtime toggles to enable, disable, or target features safely without redeploying.
Feature Store
A system for managing, serving, and versioning ML features consistently between training and inference, improving reliability and reducing training-serving skew.
Federated Learning
A machine learning approach where models are trained across multiple decentralized devices or servers holding local data samples, without exchanging the raw data. Enables privacy-preserving AI by keeping sensitive data on-device while still benefiting from collaborative learning.
Few-shot Learning/Prompting
An approach where an AI model is given a few examples (shots) of a task to learn from before it attempts the task on new input. This helps guide the model for specific outputs or styles, improving performance with limited examples.
Fine-tuning
The process of taking a pre-trained AI model (like a general LLM) and further training it on a smaller, domain-specific dataset. This adapts the model's knowledge and behavior for a particular task, style, or industry, making it more specialized.
FlashAttention
A family of GPU-optimized attention implementations that compute attention faster and with lower memory by using tiling and IO-aware kernels. It enables longer contexts and higher throughput for transformer inference and training.
Foundation Model
A large-scale AI model (often an LLM or vision model) trained on vast amounts of broad, unlabeled data. These models are designed to be adaptable (e.g., through fine-tuning or prompting) to a wide range of downstream tasks and applications.
FP16 (Half Precision)
16-bit floating point format used to speed up training and inference with lower memory usage. Often paired with loss scaling to maintain numerical stability.
Fusion‑in‑Decoder (FiD)
A retrieval architecture that encodes passages independently then fuses them within the decoder for generation.
GDPR (General Data Protection Regulation)
A comprehensive data privacy law in the European Union that sets rules for collecting, processing, and storing personal information from individuals within the EU.
GEMM (General Matrix Multiply)
A core linear algebra operation (matrix multiplication) that dominates compute cost in transformers. Highly optimized kernels are key for fast training and inference.
General-Purpose AI Model (GPAI)
A general-purpose AI model is trained for broad use and can be integrated into many downstream AI systems or applications. Under the EU AI Act, providers of GPAI models have transparency and documentation obligations, with additional duties for models classified as having systemic risk.
Generative Adversarial Network (GAN)
An AI architecture consisting of two neural networks—a generator and a discriminator—that compete against each other. The generator creates synthetic data, and the discriminator tries to distinguish it from real data, leading to increasingly realistic outputs, especially images.
Generative AI
A category of AI focused on creating new, original content, such as text (articles, poems), images, audio (music, speech), video, or code. It learns patterns from training data and generates novel outputs based on user prompts.
GGUF
A model file format optimized for efficient CPU/GPU inference in the llama.cpp ecosystem.
GPT
Generative Pre-trained Transformer - decoder-only architecture optimized for text generation, foundation of models like ChatGPT.
GPTQ
Alias for a post‑training quantization method that approximates weight updates for efficient inference.
GPU (Graphics Processing Unit)
Specialized hardware optimized for parallel computation. GPUs are the primary accelerator for training and serving modern deep learning models.
Graceful Degradation
System behavior that preserves essential user functionality when dependencies fail or resources are constrained, instead of full outage.
Gradient Accumulation
Technique to simulate a larger batch size by accumulating gradients over multiple forward/backward passes before applying an optimizer step. Useful when GPU memory limits batch size.
Gradient Clipping
Technique to prevent exploding gradients by capping gradient values during backpropagation.
Graph Neural Networks
Neural networks designed to work with graph-structured data, learning representations of nodes, edges, and entire graphs.
Graph‑of‑Thought (GoT)
Reasoning approach that structures intermediate steps as a graph of interconnected thoughts.
Gray Failure
Partial, hard‑to‑detect degradation where systems appear up but misbehave for a subset of users or traffic.
Grounded Generation
A generation approach where model outputs are explicitly anchored to retrieved or verified sources instead of relying only on parametric memory. Grounded generation reduces hallucinations and improves trust in production systems.
Groundedness
How well a response is anchored in verifiable sources or retrieved context. Grounded outputs can be traced to citations or documents.
Grounding (Knowledge Grounding)
Linking model outputs to verifiable sources or enterprise knowledge to improve factuality and trust. Typically implemented via RAG with citations.
Group‑Query Attention (GQA)
An attention variant that shares key/value across groups of heads to reduce memory while retaining quality.
Guardrails (AI)
Technical and policy controls that ensure AI systems operate within defined safety, quality, and compliance boundaries. Includes input/output filtering, policy enforcement, content moderation, and automated validation. Essential for enterprise AI deployment.
Hallucination (AI)
A phenomenon where a generative AI model, particularly an LLM, produces outputs that sound plausible and confident but are factually incorrect, nonsensical, or not based on the provided input. Critical evaluation of AI output is necessary due to this.
HBM (High Bandwidth Memory)
High-performance memory used on many data-center GPUs (e.g., H100). Higher bandwidth improves training and inference throughput for large models.
Health Checks
Endpoints or probes indicating service liveness/readiness for orchestrators and load balancers.
High-Risk AI System
An AI system category under the EU AI Act for uses that may significantly affect health, safety, or fundamental rights. High-risk systems face obligations around risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy, robustness, and cybersecurity.
HIPAA
US regulation governing protected health information (PHI). For AI systems, HIPAA drives requirements like access controls, audit trails, BAAs, and data minimization.
Hit@k
Binary metric indicating whether at least one relevant item appears in the top k results. Useful when a single good context chunk is sufficient.
HNSW (Hierarchical Navigable Small World)
A popular ANN index enabling fast approximate nearest neighbor search in high dimensions.
Homomorphic Encryption
Cryptographic technique allowing computations on encrypted data without decryption, enabling privacy-preserving AI.
Hugging Face
An ecosystem for model sharing, datasets, and tooling (Transformers, tokenizers, model hubs). Widely used for open-weight model distribution and deployment.
Human Evaluation
Manual evaluation by people using rubrics (correctness, helpfulness, safety, style). Often used to validate automatic metrics and detect edge-case failures.
Human-in-the-loop (HITL)
Operational setup where humans review, correct, or approve model decisions for sensitive, low-confidence, or high-impact cases.
Hybrid Search
Combining lexical search (e.g., BM25) and semantic vector search to improve relevance and recall. Useful when queries include rare terms, synonyms, or mixed intents.
HyDE (Hypothetical Document Embeddings)
Retrieval technique that generates a hypothetical answer/document for a query and embeds it to retrieve more relevant passages. Often improves recall for complex questions.
Hyperparameter Tuning
Process of finding optimal hyperparameters for machine learning models using techniques like grid search or Bayesian optimization.
Idempotency
Designing operations that produce the same result when retried, preventing duplicates during network failures or webhook retries.
Imbalanced Data
Situation where classes in a dataset have unequal representation, requiring special handling techniques.
In-context Learning
The ability of a large language model to learn and perform a new task based solely on the examples and instructions provided within the prompt, without needing to be retrained or fine-tuned. This is the mechanism behind few-shot and zero-shot prompting.
Incident Commander
Designated lead responsible for coordinating incident response, assigning roles, driving decisions, and maintaining shared situational awareness.
Incident Response
Coordinated process for detecting, triaging, containing, communicating, and resolving incidents while minimizing user and business impact.
Incident Severity (SEV Levels)
Categorization of incident impact (e.g., SEV‑1 critical). Determines response urgency, comms, and escalation.
Indirect Prompt Injection
A prompt injection attack where malicious instructions are embedded in external content (web pages, documents, emails) that the model later ingests during browsing or retrieval.
Inference
The process of using a trained AI model to make predictions, generate content, classify data, or perform its designated task on new, previously unseen data. This is the 'live' operational phase of an AI model.
Inference Cost
The computational and financial cost of running a trained model to generate predictions or outputs. Factors include model size, token count, hardware requirements, and API pricing. Optimizing inference cost is crucial for production AI applications.
Inference Time
The duration from sending a request to receiving the complete response. Includes network latency, queue time, and actual model computation. Critical for user experience in real-time applications. Measured in milliseconds for simple tasks, seconds for complex ones.
Inpainting
A generative technique that fills in missing or masked regions of an image while matching surrounding context. Used for edits, object removal, and iterative refinement.
Insecure Output Handling
Security weakness where LLM outputs are trusted and executed without validation, sanitization, or policy checks.
Instruction Following
A model's ability to understand and execute user instructions accurately. Instruction-tuned models are trained on instruction-response pairs to improve this capability. Core to assistant-style AI systems.
Instruction Tuning
Fine‑tuning models on instruction‑following datasets to improve helpfulness and adherence to prompts.
INT4
4-bit quantized precision used for very memory-efficient inference. Can significantly reduce VRAM usage but may require careful calibration to preserve quality.
INT8
8-bit integer precision commonly used for quantized inference to reduce memory and improve throughput with minimal quality loss.
ISO/IEC 42001
An international standard for AI management systems (AIMS). It provides requirements for establishing, implementing, maintaining, and continually improving AI governance and controls.
IVF (Inverted File Index)
A vector index that partitions vectors into coarse clusters (lists) and searches a subset for speed.
Jailbreaking (AI)
Techniques to bypass an AI model's safety guardrails and content policies, often through carefully crafted prompts that trick the model into generating prohibited content. AI companies continuously work to patch jailbreaks, but it remains an ongoing challenge.
JSON Mode
A generation mode or constraint where the model is instructed (or enforced) to output valid JSON. Often paired with schema validation to improve reliability for downstream automation.
JSON Schema
A specification for defining the structure and validation of JSON data, often used for data exchange and API documentation.
K-Fold Cross-Validation
Technique that divides data into k subsets, using each as validation set once while others serve as training data.
Kernel Fusion
Combining multiple GPU operations into a single kernel to reduce memory bandwidth overhead and improve throughput.
Knowledge Cutoff
The most recent date of data used to train a model. Facts after this date may be unknown to the model unless provided via retrieval or browsing.
Knowledge Distillation
Training a smaller "student" model to replicate the behavior of a larger "teacher" model, improving efficiency by reducing size and inference cost while retaining performance.
Knowledge Graph
Structured representation of knowledge as entities and relationships that supports semantic search, reasoning, and retrieval. Used for question answering, entity linking, and enterprise knowledge systems.
KV Cache (Key‑Value Cache)
Cached attention key/value tensors reused across decoding steps to avoid recomputation and reduce latency and cost.
KV Cache Eviction
Policies for discarding old key/value attention states in long sessions to manage memory.
Large Language Model (LLM)
A language model trained at large scale to predict and generate text or code. Most modern LLMs use Transformer-style architectures and can support tasks such as summarization, drafting, coding, question answering, extraction, and tool-using agents.
Latency (AI Systems)
The time between sending a request and receiving a response. Optimizations span model choice, prompt size, retrieval efficiency, batching, and parallelization.
Latent Diffusion
A diffusion approach that performs the denoising process in a compressed latent space (rather than pixel space), improving efficiency while preserving quality. Stable Diffusion is a prominent example.
Latent Space
A compressed, abstract, multi-dimensional representation of data learned by an AI model. In this space, similar data points are closer together. Generative models like GANs and Diffusion Models operate in this latent space to create new data by manipulating these compressed representations.
Lawful Basis
Legal justification required under GDPR Article 6 for processing personal data, such as consent, contract, legal obligation, or legitimate interests.
Layer Normalization (LayerNorm)
A normalization technique that stabilizes training by normalizing activations within a layer. Commonly used throughout transformer blocks.
Leaderboard
A ranking of AI models based on standardized benchmark performance. Popular leaderboards include Chatbot Arena (LMSYS), HuggingFace Open LLM Leaderboard, and MMLU. Helps compare models objectively, though real-world performance may vary.
Learning Rate Scheduling
Techniques to dynamically adjust the learning rate during training, often decreasing it over time.
Learning Rate Warmup
Training schedule where the learning rate ramps up from a small value over the first steps/epochs to stabilize optimization, especially for transformers.
Least Privilege
Security principle that users, services, and processes should only have the minimum permissions required for their tasks.
Lip‑sync (AI)
Aligning mouth movements in video with target speech for dubbing or avatars.
LLM-as-a-Judge
Using a language model to evaluate other model outputs (or its own outputs) against criteria like correctness, safety, relevance, and style. It’s widely used for scalable evals but requires careful prompt design, calibration, and spot-checking to avoid bias and false confidence.
LLMOps
Operational practices for deploying, monitoring, evaluating, and iterating on LLM‑powered systems.
Load Shedding
Resilience strategy that intentionally drops or rejects low-priority requests during overload to protect core service functionality.
Local Differential Privacy
Differential privacy applied at the individual level before data sharing, providing stronger privacy guarantees.
Logit Bias
A technique to bias token probabilities during decoding by adding positive or negative offsets to specific token logits. Used to encourage or forbid certain tokens.
Logprobs
Log probabilities of tokens under the model during generation. Logprobs enable confidence scoring, debugging, and more controlled decoding.
Long Short-Term Memory (LSTM)
Type of recurrent neural network cell designed to remember information for long periods, solving the vanishing gradient problem.
Long-Context Models
Models designed to process unusually long input sequences. They help with long documents and multi-document tasks, but still require careful context selection because important information can be missed or diluted inside very long prompts.
LoRA (Low‑Rank Adaptation)
A parameter‑efficient fine‑tuning method that injects low‑rank adapters into a frozen model, drastically reducing compute and cost while achieving strong task performance.
Low-Rank Factorization
Matrix decomposition technique to reduce model parameters by approximating weight matrices with lower-rank representations.
Machine Learning (ML)
A subset of AI where systems learn from data to improve their performance on a specific task over time, without being explicitly programmed for every single scenario. It relies on algorithms and statistical models to find patterns in data.
Map‑Reduce RAG
A pipeline pattern that answers sub‑questions per chunk (map) then aggregates into a final response (reduce).
Mean Average Precision (mAP)
Metric for evaluating object detection models, averaging precision across different recall levels.
Membership Inference Attacks
Privacy attacks that determine whether a particular data sample was used in training a model.
MITRE ATLAS
Adversarial ML knowledge base and tactic/technique framework used to model attacks and defenses across the machine learning lifecycle.
Mixture of Experts (MoE)
A model architecture with multiple expert sub-networks and a router that selects which experts to activate for each input. MoE can increase model capacity while keeping per-token compute lower than activating every parameter.
MLOps (Machine Learning Operations)
A set of practices that aims to deploy and maintain machine learning models in production reliably and efficiently. It combines ML, DevOps, and Data Engineering principles to manage the ML lifecycle.
MMR (Maximal Marginal Relevance)
Diversification technique that balances relevance and novelty to reduce redundancy in retrieved passages.
Model Card
A documentation standard that provides essential information about an AI model, including its intended use, training data, limitations, biases, performance metrics, and ethical considerations. Helps users make informed decisions about model selection and deployment.
Model Collapse
A potential long-term problem where generative AI models, trained on data that itself was generated by other AIs, begin to lose information, forget less common patterns, and produce less diverse and more homogenous outputs over successive generations. It's a form of degenerative feedback loop.
Model Context Protocol (MCP)
An open protocol for connecting LLM applications to tools, data sources, and "context" providers in a standardized way. MCP helps make tool integration more portable across models and runtimes, similar to how HTTP standardizes web communication.
Model Denial of Service
An attack or failure mode where expensive prompts, large context requests, repeated tool calls, or resource-heavy workflows overload an AI system, increasing latency, cost, or outages. Mitigations include rate limits, token budgets, timeouts, quotas, and workload isolation.
Model Distillation (Knowledge Distillation)
A technique where a smaller 'student' model is trained to mimic the behavior of a larger 'teacher' model. This creates faster, cheaper models that retain much of the teacher's performance. Used to create efficient models for edge devices or cost-sensitive applications.
Model Drift
Degradation or changes in model behavior/performance over time due to drift, data pipeline changes, prompt updates, or system modifications. Requires monitoring and periodic re-evaluation.
Model Endpoint
A URL or API address where a deployed AI model can be accessed for inference. Endpoints can be hosted by AI providers (like OpenAI's API), on your own infrastructure, or on specialized platforms. Each endpoint has specific authentication, rate limits, and pricing.
Model Inversion
Attack technique that reconstructs sensitive training data from model outputs or parameters.
Model Merging
Combining weights from multiple fine-tuned models into a single model without additional training. Techniques like SLERP, TIES, and DARE can blend capabilities from different specializations.
Model Monitoring
Monitoring deployed models and pipelines for quality, safety, latency, cost, and drift. Often includes alerting, evaluation on live traffic, and rollback strategies.
Model Registry
A system that stores, versions, and governs models and their artifacts (weights, configs, lineage, evaluation results) for deployment and auditing.
Model Risk Management
Governance and control practices that reduce business, compliance, and operational risks from incorrect, biased, or misused model outputs.
Model Routing
Selecting which model (or pipeline) should handle a request based on intent, complexity, cost, latency targets, or risk level. Routing can combine fast/cheap models with fallback to stronger ones.
Model Serving
The infrastructure and runtime layer used to host models and return predictions in real time or batch mode. Model serving includes autoscaling, caching, observability, and rollout strategies.
Model Validation
Independent verification that a model is accurate, robust, fair, and fit for intended use before and after deployment.
Model Versioning
Tracking and managing model versions (weights, prompts, retrieval configs, eval results) so changes are auditable, reproducible, and safely deployable.
Model Weights
The learned parameters of a trained model. Weights are typically stored as checkpoint files (e.g., safetensors) and determine the model’s behavior during inference.
Moderation Classifier
A model that detects policy‑violating content (e.g., hate, self‑harm, sexual content) in inputs or outputs to enforce safety policies.
MRR (Mean Reciprocal Rank)
Average reciprocal rank of the first relevant result; emphasizes early precision.
MTBF (Mean Time Between Failures)
Average time between service failures. Improved by testing, redundancy, and robust change management.
MTEB (Massive Text Embedding Benchmark)
A benchmark suite for evaluating embedding model quality across many tasks.
MTTA (Mean Time to Acknowledge)
Average time between incident alert creation and first human acknowledgment. Lower MTTA indicates faster operational response.
MTTR (Mean Time to Recovery)
Average time to restore service after an incident. Lower via clear ownership, runbooks, and rollbacks.
Multi-Agent Systems
Architectures where multiple AI agents collaborate, compete, or specialize on different subtasks. Agents may have different roles (planner, executor, critic) or work on parallel subproblems.
Multi-Head Attention
Attention mechanism that runs multiple attention operations in parallel, allowing models to focus on different aspects simultaneously.
Multi‑Query Attention (MQA)
An attention optimization that shares key/value across heads for lower memory and faster decoding.
Multimodal AI
AI systems that can process or generate more than one data modality, such as text, images, audio, video, tables, or code. Multimodal systems are useful when the task depends on combining signals, for example analyzing a screenshot and explaining it in natural language.
Multimodal Learning
AI systems that process and understand multiple types of data simultaneously (text, images, audio, video).
Named Entity Recognition (NER)
NLP task that identifies and classifies named entities in text (persons, organizations, locations, etc.).
Natural Language Processing (NLP)
A subfield of AI focused on enabling computers to understand, interpret, process, and generate human language (both written and spoken). LLMs are a key technology driving advancements in NLP.
Natural Language Understanding (NLU)
A component of NLP focused on the more challenging task of enabling machines to comprehend the meaning, intent, sentiment, and context of human language, beyond just syntactic parsing.
NCCL
NVIDIA Collective Communications Library used for fast multi-GPU communication (all-reduce, broadcast), critical for distributed training.
NDCG (Normalized Discounted Cumulative Gain)
A ranking metric that accounts for graded relevance and position in the list.
Needle in a Haystack Test
An evaluation method that tests a model's ability to retrieve specific information (the 'needle') placed at various positions within a long context (the 'haystack'). Used to assess long-context retrieval capabilities.
Negative Prompt
A text prompt describing what should be avoided in generated content (common in text-to-image). It steers sampling away from unwanted artifacts or styles.
Negative Sampling
Efficient training technique for large vocabularies by sampling negative examples instead of computing all possibilities.
NeRF (Neural Radiance Fields)
A 3D representation learned from images to render novel views, used in graphics and video.
Neural Architecture Search
Automated process of designing optimal neural network architectures using machine learning techniques.
Neural Network
A computational model inspired by the structure and function of the human brain, composed of interconnected processing units called 'neurons' organized in layers. Neural networks are the core of Deep Learning.
Neuro-Symbolic AI
Hybrid approach combining neural networks (for pattern recognition) with symbolic reasoning (for logic and rules).
NIST AI Risk Management Framework (AI Risk Management Framework)
A framework from NIST to help organizations manage AI risks across governance, mapping, measuring, and managing. Often used to structure safety, privacy, and reliability programs.
NLG
Alias for Natural Language Generation (NLG), creating text from structured or unstructured inputs.
Observability
End‑to‑end tracing, logging, and metrics to understand behavior, debug incidents, and improve quality.
OCR
Alias for Optical Character Recognition.
Off-policy
In reinforcement learning, learning from data collected by a different (behavior) policy, such as replay buffers. Often more sample-efficient but can be less stable.
On-call Rotation
Scheduled responsibility model where engineers take turns responding to production alerts and incidents.
On-policy
In reinforcement learning, learning from data collected by the current policy being optimized. Often more stable but sample-inefficient.
On‑Prem vs Cloud
Hosting models: on‑premises provides isolation and control; cloud offers elasticity and managed services. Many enterprises use hybrid patterns.
One-Hot Encoding
Method to convert categorical variables into binary vectors, where each category becomes a separate dimension.
ONNX
Open Neural Network Exchange — a model format for exporting and running models across frameworks and runtimes.
ONNX Runtime
A high-performance runtime for executing ONNX models across CPU and GPU backends.
Open Source AI
AI models, datasets, and tools whose source code, design, or data is made publicly available, often under a license that permits use, modification, and distribution. This fosters collaboration and innovation.
OPQ (Optimized Product Quantization)
A rotation applied before PQ to reduce quantization error and improve search accuracy.
Optimizer
Algorithms that adjust model parameters (weights) during training to minimize a loss function. Common optimizers include SGD, Adam, AdaGrad, and AdaDelta.
Out-of-Distribution (OOD)
Inputs that differ significantly from the training distribution. OOD detection and robustness testing are important for safe deployment.
Output Validation
Validating model outputs against expected formats or rules (e.g., JSON Schema, regex, allowlists). Output validation reduces silent failures and blocks unsafe or malformed responses.
Overfitting
A common problem in machine learning where a model learns the training data too well, including its noise and random fluctuations. An overfit model performs exceptionally well on the data it was trained on but fails to generalize and make accurate predictions on new, unseen data.
Overfitting vs. Underfitting
Overfitting occurs when a model learns training data too well but fails on new data; underfitting when it fails to capture patterns.
Overoptimization
When optimizing too strongly against a proxy metric (like a reward model) degrades actual performance. Common in RLHF where excessive optimization against the reward model can reduce true helpfulness.
PagedAttention
A KV-cache management approach (popularized by vLLM) that allocates KV blocks like a paging system, reducing fragmentation and improving memory efficiency for many concurrent sequences.
Pairwise Preference Evaluation
An evaluation method where raters choose between two outputs (A vs B). This supports win-rate metrics and is commonly used to build preference datasets.
Parameter (in AI models)
Internal variables or 'weights' within an AI model, especially neural networks, that are learned and adjusted during the training process. Models with more parameters (e.g., billions in LLMs) can often learn more complex patterns and store more information.
PEFT (Parameter‑Efficient Fine‑Tuning)
Methods like LoRA, prefix‑tuning, and adapters that update a small subset of parameters to adapt models.
Penalties (Frequency/Presence)
Decoding parameters that discourage repetition by lowering probabilities of previously generated tokens.
Percentiles (p95/p99)
Latency distribution cutoffs indicating worst‑case experience. Track p95/p99 for endpoints, TTFB, and decode latency to detect tail issues.
Perplexity
A metric measuring how well a language model predicts text. Lower perplexity indicates better prediction. Calculated as the exponential of the average negative log-likelihood. While useful for comparing models, it doesn't always correlate with human-perceived quality.
Pipeline Parallelism
Distributing different layers/stages of a model across devices and streaming micro-batches through the pipeline.
Plan-and-Execute
Agent pattern that first creates a high-level plan (steps/subtasks) and then executes the steps, often re-planning when new information arrives.
Playground
An interactive web interface provided by AI companies (like OpenAI, Anthropic) where you can test models, adjust parameters (temperature, max tokens), and experiment with prompts without writing code. Useful for prototyping and learning how models behave.
Poisoning Attacks
Adversarial attacks that corrupt training data to manipulate model behavior during the learning phase.
Policy as Code
Practice of defining and enforcing governance and security rules in machine-readable code so policies are versioned, tested, and automatically applied.
Policy Optimization
Algorithms that improve reinforcement learning policies by maximizing expected rewards.
Positional Encoding
Techniques to encode token positions for attention (sinusoidal, learned, RoPE, ALiBi).
Postmortem
Structured review after an incident to document timeline, impact, contributing factors, and preventive actions. Effective postmortems are blameless and action-oriented.
PPO
Proximal Policy Optimization - a policy gradient method for reinforcement learning that ensures stable training.
PQ (Product Quantization)
A compression technique that splits vectors into subvectors and quantizes them to reduce memory.
Precision vs. Recall
Precision measures accuracy of positive predictions; recall measures how many actual positives were found.
Precision@k
Fraction of retrieved items in the top k that are relevant. Higher precision@k means fewer irrelevant results among top hits.
Preference Learning
A family of training methods where models learn from human or AI preferences (comparisons, rankings) rather than explicit labels. Includes RLHF, DPO, and RLAIF. Critical for alignment and instruction-following.
Prefill vs Decode
Two phases of transformer inference: prefill computes attention over the prompt (often heavy compute), while decode generates tokens step-by-step (often memory/KV-cache bound). Optimizing both phases is key for low latency and high throughput.
Pretraining
The initial training phase where a model learns general representations from large-scale data before being fine-tuned for specific tasks. Pretraining captures broad knowledge that transfers to downstream applications.
Privacy by Design
Approach that embeds privacy controls into system architecture and processes from the start, rather than as an afterthought.
Program‑of‑Thought (PoT)
Reasoning technique where the model writes small programs (often Python) to compute accurate answers.
Prompt
The instruction, question, text, or other input provided by a user to an AI model (especially a generative AI model) to guide its response or content generation. Crafting effective prompts is a skill known as 'Prompt Engineering'.
Prompt Caching
Reusing precomputed prompt or prefix representations to reduce latency and cost for repeated or shared context.
Prompt Chaining
A pipeline pattern where multiple prompts are executed in sequence, with each step transforming or validating the previous output. Prompt chains improve reliability for complex tasks like extraction, summarization, and multi-step reasoning.
Prompt Engineering
The practice of carefully designing and refining the input text (prompts) given to an AI model to elicit the desired output. Effective prompt engineering involves understanding the model's capabilities, using clear instructions, providing examples, and iterating on the prompt structure.
Prompt Injection
Adversarial inputs that try to override system instructions or misuse tools/APIs. Defend with input sanitization, strict tool scopes, allow/deny-lists, and output validation.
Prompt Leakage
Unintended disclosure of hidden instructions (system prompts), policies, or private context. Can occur via prompt injection, tool misuse, or insufficient isolation.
Prompt Library
A collection of pre-defined prompts or templates used for generating text or other content.
Prompt Linting
The process of analyzing and improving the quality and effectiveness of prompts used in language models.
Prompt Logging
Storing prompts, retrieved context, tool calls, and model outputs for debugging and evaluation. Must be paired with redaction, retention controls, and access restrictions.
Prompt Template
A reusable prompt pattern with placeholders for variables; standardizes instructions and improves consistency.
Provenance
The origin, history, and ownership of data or content, often tracked for transparency and accountability.
Pruning
Technique to reduce model size by removing unnecessary parameters or connections while maintaining performance.
Purpose Limitation
Privacy principle that personal data should be collected for specified, explicit purposes and not reused incompatibly.
PyTorch
Popular open-source deep learning framework developed by Facebook, known for dynamic computation graphs and Python integration.
QLoRA (Quantized Low-Rank Adaptation)
An extremely memory-efficient fine-tuning method that combines quantization with LoRA, enabling fine-tuning of very large models on a single GPU (the QLoRA paper tuned a 65B-parameter model on one 48 GB GPU) instead of a multi-GPU cluster.
Query Expansion
Retrieval technique that adds related terms/synonyms to a query to improve recall, often used in lexical or hybrid search.
Query Rewriting
Transforming a user query into a better retrieval query (e.g., adding context, disambiguation, or keywords) to improve search relevance in RAG.
RAG Fusion
A retrieval pattern that issues multiple query variants (e.g., rewritten queries), retrieves results for each, and fuses/merges them (often with rank fusion) to improve recall and robustness.
RAGAS
An evaluation framework for Retrieval-Augmented Generation that measures dimensions like faithfulness and answer relevance using LLM-assisted scoring. Useful for scalable RAG quality monitoring.
Random Seed
A value used to initialize a random number generator, allowing for reproducibility and consistency in randomized processes.
Rate Limiting
Restrictions on how many API requests you can make within a time period (e.g., 60 requests per minute). Rate limits prevent abuse and ensure fair resource allocation. Exceeding limits typically results in HTTP 429 errors. Important for planning application architecture.
RBAC (Role‑Based Access Control)
Authorization model mapping roles to permissions, reducing ad‑hoc grants. Use least privilege for APIs, webhooks, and admin tools.
ReAct (Reasoning + Acting)
A prompting pattern where a model alternates between brief reasoning summaries and actions, such as tool calls or searches, then uses observations to continue the task. In production, expose concise rationale and tool traces rather than hidden chain-of-thought.
React Prompting
Common shorthand for ReAct prompting, not the React JavaScript framework. ReAct combines concise reasoning summaries with actions such as tool calls, searches, or observations.
Reader
A component that reads retrieved passages and extracts or synthesizes an answer. In modern RAG, the generator model often also acts as the reader.
Reasoning Models
AI models optimized for tasks that benefit from more inference-time computation, planning, and verification. They are useful for math, coding, structured analysis, and multi-step decisions; users should ask for concise rationale and verifiable outputs rather than hidden chain-of-thought.
Recall@k
Fraction of relevant items retrieved within the top k results; key metric for retrieval quality.
Receiver Operating Characteristic (ROC)
Curve plotting true positive rate against false positive rate to evaluate binary classification models.
Records of Processing Activities (ROPA)
Mandatory internal documentation of personal-data processing activities, including purposes, categories, recipients, transfers, retention, and safeguards.
Red Teaming (AI)
Structured adversarial testing of AI systems to uncover vulnerabilities (prompt injection, jailbreaks, unsafe tool actions) and improve defenses.
Redaction
Removing or masking sensitive information (PII, secrets) from logs, prompts, or outputs. Redaction reduces leakage risk but must preserve enough context for debugging.
Reflexion
An agent technique where the model reflects on failures, critiques its behavior, and updates future actions based on feedback (self-improvement loop).
Regression (in ML)
A supervised machine learning task where the model learns to predict a continuous numerical value. For example, predicting the price of a house based on its features (size, location) or forecasting future sales based on historical data.
Regularization
Techniques to prevent overfitting by adding penalties to model complexity (L1, L2 regularization).
Reinforcement Learning
A type of Machine Learning where an AI agent learns by interacting with an environment. It receives rewards for desirable actions and penalties for undesirable ones, gradually optimizing its strategy or 'policy' to maximize cumulative reward.
Release Train
Fixed-cadence release model where changes ship on scheduled windows, improving predictability and coordination across teams.
Reranker
A model that re-scores and reorders retrieved documents based on their relevance to the query. Applied after initial retrieval to improve precision by using more sophisticated relevance judgments.
Reranking
A second‑stage scoring step that re‑orders initial search results using a more powerful model (e.g., cross‑encoder), improving relevance in RAG pipelines.
Residual Networks (ResNets)
Deep neural network architecture using skip connections to ease gradient flow and enable very deep networks.
Retries with Jitter
Retry strategy that randomizes delays to avoid thundering herds and contention during partial outages.
Retrieval Augmented Generation (RAG)
A technique that enhances Large Language Models by allowing them to retrieve relevant information from external knowledge sources (like databases or documents) before generating a response. This helps ground the model's output in factual, up-to-date information, reducing hallucinations and improving accuracy.
Retriever
A component in a RAG system that searches a knowledge base (lexical, vector, or hybrid) and returns candidate passages/documents for a query.
Reward Hacking
When an RL agent finds unintended ways to maximize its reward signal without achieving the desired behavior. A key challenge in RLHF where models may game the reward model rather than genuinely improve.
Reward Model
Component in RLHF that learns to predict human preferences and assigns scores to different AI outputs.
ReWOO
Reasoning Without Observation — an agent pattern that plans tool calls and reasoning steps while minimizing intermediate observations, aiming to reduce error accumulation and cost.
Risk Register
Centralized list of identified risks with probability, impact, owner, mitigation status, and review cadence.
RLAIF (Reinforcement Learning from AI Feedback)
A variant of RLHF where feedback/rankings are generated by AI models instead of (or in addition to) humans. It can scale preference data collection but needs robust safety and bias controls.
RLHF (Reinforcement Learning from Human Feedback)
A training technique where human evaluators rank or rate model outputs, then those preferences are used to improve model behavior. RLHF is commonly associated with instruction following, helpfulness, and reducing unwanted outputs, but it is only one part of modern alignment pipelines.
RMSNorm
Root Mean Square Normalization — a variant of LayerNorm that normalizes by RMS without mean-centering. Used in several modern LLM architectures for efficiency.
Robustness
A system’s ability to maintain performance under distribution shifts, adversarial inputs, noise, and edge cases.
ROC-AUC Score
Area under the ROC curve, measuring a model's ability to distinguish between classes.
ROCm
AMD’s open software stack for GPU computing. Enables training and inference on AMD GPUs.
Rollback
Reverting to a previous stable version when a release degrades SLOs or causes incidents.
RoPE (Rotary Positional Embeddings)
A positional method that rotates queries/keys in complex space to encode relative positions, aiding long context.
ROUGE Score
Recall-Oriented Understudy for Gisting Evaluation - metrics for evaluating text summarization by measuring overlap between generated and reference summaries. Includes ROUGE-N (n-grams), ROUGE-L (longest common subsequence), and ROUGE-W (weighted sequences).
RPO (Recovery Point Objective)
Maximum acceptable data loss measured in time (how far back you can restore). Influences backup cadence and replication.
RTO (Recovery Time Objective)
Maximum acceptable downtime after an incident before service must be restored. Plan via DR strategies, runbooks, and canary rollouts.
Runbook
Step‑by‑step operational guide to diagnose and remediate incidents. Include rollback steps and verification checks.
Runbook Automation
Automating common incident diagnostics and remediation steps to reduce manual toil, variance, and recovery time.
safetensors
A safe and fast tensor serialization format (commonly used with Hugging Face) designed to avoid arbitrary code execution risks present in some pickle-based formats.
Sandboxing
Isolating untrusted code, tools, or browsing in restricted environments to limit damage and prevent access to secrets or sensitive resources.
SBOM (Software Bill of Materials)
A machine-readable inventory of components and dependencies used to build software (and sometimes ML systems). SBOMs help assess vulnerabilities and supply-chain risk.
Scaling Laws
Empirical relationships showing how model performance improves predictably with increased compute, data, and parameters. Scaling laws guide decisions about model size and training budget allocation.
ScaNN
A Google ANN library for efficient vector similarity search with quantization and reordering.
SDK (Software Development Kit)
A collection of software development tools, libraries, and documentation provided by hardware or software vendors to help developers build applications for a specific platform or service (e.g., an AI tool might offer an SDK for easier API integration).
Secrets Management
Practices and tooling for storing and accessing secrets (API keys, tokens, credentials) securely. Prevents leaks to logs, prompts, client-side code, or third parties.
Security Baseline
Minimum required security controls and configuration standards that all systems must meet before production use.
Segregation of Duties
Control principle that separates critical responsibilities across multiple people or roles to reduce fraud, abuse, and high-risk unilateral changes.
Self-Attention
Attention mechanism that computes relationships between elements within the same sequence, enabling contextual understanding.
Self-Supervised Learning
A learning paradigm where models learn from unlabeled data by creating supervisory signals from the data itself (e.g., predicting masked words, next tokens, or image patches). Foundation for modern LLMs and vision models.
Self‑Consistency
A reasoning technique where the model samples multiple solution paths and selects the most consistent answer (e.g., via majority vote). Often improves accuracy on reasoning tasks.
Semantic Chunking
The process of breaking down text or content into smaller, meaningful chunks or segments, often used for information retrieval or question answering.
Semantic Memory (Agents)
Long-term factual knowledge a system stores (documents, notes, structured facts). Often implemented via vector databases and retrieval.
Semantic Role Labeling
NLP task that identifies roles played by entities in relation to predicates (who did what to whom, when, where, etc.).
Semantic Search
A search technique that aims to understand the user's intent and the contextual meaning of query terms, rather than just matching keywords. It returns more relevant results by understanding relationships between concepts. This is often powered by embeddings and vector databases.
SentencePiece
A tokenization library and algorithm family (often BPE or unigram) that operates directly on raw text (including whitespace), commonly used in many open-weight LLMs.
Service Credits
Remedies offered under an SLA when uptime or SLOs are missed, typically as bill credits. Review exclusions, calculation, and cap.
Service Dependency Map
Representation of upstream and downstream service relationships used to understand blast radius, ownership, and incident impact paths.
Shadow Deployment
Release strategy where a new model receives mirrored production traffic for evaluation while the current model continues serving official outputs.
SLA (Service Level Agreement)
Contractual guarantees on uptime, response times, and remedies. Review scope, exclusions, and credit calculations.
SLI (Service Level Indicator)
Measured metric of service quality (e.g., availability, latency percentiles, TTFB). Drives SLOs and error budgets.
SLO (Service Level Objective)
Target reliability/latency goals for a service. Paired with SLIs and enforced via error budgets and incident response.
Small Language Model (SLM)
A compact language model optimized for low latency, lower cost, and on-device or edge deployment. SLMs are often used for classification, extraction, routing, and lightweight assistants where speed and privacy are more important than maximum reasoning depth.
SOC 2
An independent audit framework assessing controls for Security, Availability, Processing Integrity, Confidentiality, and Privacy.
Source Attribution
The practice of acknowledging and crediting the original sources of information, data, or content.
Sparse Models
Neural networks with mostly zero weights, reducing computation and memory requirements while maintaining performance.
Speculative Decoding
An inference optimization technique where a smaller draft model proposes tokens and a larger target model verifies them. It can reduce latency when the draft model predicts well, but actual speedups depend on model pairing, hardware, batch size, and serving implementation.
SRE (Google SRE Book: Service Level Objectives)
Engineering discipline combining software and systems thinking to achieve reliable, scalable services using SLOs, error budgets, automation, and blameless postmortems.
SSE
Alias for Server‑Sent Events streaming.
Stable Diffusion
A widely used text-to-image model family based on latent diffusion. It generates images by denoising in a latent space conditioned on text embeddings, and is popular for customization via fine-tuning and control modules.
Stochastic Gradient Descent (SGD)
Optimization algorithm that updates parameters using gradients from single training examples or small batches.
Stop Sequences
Decoding control where generation halts when the model outputs any of the specified stop strings. Used to prevent the model from continuing into unwanted sections.
Stratified Sampling
Sampling technique that maintains class proportions from the original dataset in training/validation splits.
Streaming Response
An API response mode where the model's output is sent incrementally as it's generated, rather than waiting for the complete response. This provides a better user experience by showing progress in real-time, similar to how ChatGPT displays responses word-by-word.
STRIDE
Threat-modeling method categorizing risks into Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, and Elevation of Privilege.
Structured Output
Model output in a defined format (JSON, XML, tables) rather than free-form text. Enables reliable integration with downstream systems. Often enforced through JSON mode, function calling, or grammar constraints.
STT (Speech‑to‑Text)
Alias of ASR — converting speech to text.
Supervised Fine-Tuning (SFT)
Initial training phase in RLHF where model learns from labeled examples before preference learning.
Supervised Learning
A type of Machine Learning where the model learns from labeled data. This means each input data point in the training set is paired with a known correct output or 'label,' allowing the model to learn the mapping between inputs and outputs.
Supply Chain Security
Practices that reduce risks from third-party code, model artifacts, and dependencies (e.g., poisoned packages, compromised weights). Includes provenance, signature verification, and controlled publishing.
Synthetic Data
Artificially generated data that mimics real-world data patterns. Used for training when real data is scarce, sensitive, or expensive. Can be generated by rule-based systems, simulations, or generative AI models.
Synthetic Monitoring
Proactive monitoring approach that runs scripted checks against critical user journeys to detect issues before users report them.
System Message
Alias for the system prompt that sets behavior, tone, and constraints for the model.
System Prompt
A hidden instruction that sets the model’s behavior, tone, and constraints. Keeping it stable improves consistency and makes changes auditable.
Systemic Risk (AI)
A risk that can have broad effects beyond one user or one application, such as large-scale misuse, serious incidents, cybersecurity failures, or impacts across many downstream systems. Under the EU AI Act, certain GPAI models with systemic risk face additional obligations.
TCO (Total Cost of Ownership)
True cost over time including subscription, infrastructure, integration, operations, and support.
Temperature (Sampling)
A decoding parameter that controls randomness. Higher values increase diversity and creativity; lower values improve determinism and repeatability.
Tensor Cores
Specialized GPU units that accelerate mixed-precision matrix operations (e.g., FP16/BF16) used heavily in deep learning.
Tensor Parallelism
Splitting a single model’s layers/weights across multiple devices to fit large models and increase compute throughput.
TensorFlow
Google's open-source machine learning framework, known for production deployment and graph-based computation.
TensorRT
NVIDIA’s inference optimization SDK that compiles and optimizes neural networks for fast GPU execution.
Test-Time Compute
Computational resources spent during inference to improve output quality, such as using a stronger reasoning model, sampling multiple candidate answers, self-consistency voting, or running extra verification steps.
Text-to-Image
A type of generative AI that creates images from textual descriptions (prompts).
Text-to-Video
A type of generative AI that creates video clips from textual descriptions or image inputs.
Threat Modeling
Structured process to identify, prioritize, and mitigate security threats during system design and architecture decisions.
Throughput (TPS)
The number of requests or tokens processed per second. Critical for scaling and cost efficiency in production systems.
Token (in LLMs)
In Large Language Models, text is often broken down into smaller units called tokens for processing. Tokens can be whole words, parts of words (subwords), or even individual characters and punctuation. Model context windows and pricing are often measured in tokens.
Token Pricing
The cost structure for using AI APIs, typically charged per 1,000 tokens (1K tokens). Pricing often differs between input tokens (prompt) and output tokens (completion). Understanding token pricing is crucial for budgeting AI applications.
Tokenization
The process of splitting text into tokens (subwords/characters). Tokenization affects context limits, latency, and cost calculations.
Tokenizer
A component that converts text into tokens (IDs) for a model and back into text. Tokenizer choice (BPE, SentencePiece, etc.) affects token counts, costs, and how well a model handles languages, whitespace, and rare words.
Top‑k
Sampling from the top k most probable tokens at each step to control randomness and quality.
Top‑k Retrieval
Selecting the k most similar passages to a query based on vector similarity or hybrid scoring.
Top‑p (Nucleus Sampling)
A decoding method that samples from the smallest set of tokens whose cumulative probability is at least p, balancing quality and diversity.
TPU (Tensor Processing Unit)
Google’s accelerator for matrix-heavy ML workloads. TPUs are commonly used for training and serving large models in Google’s ecosystem.
Training (AI Model)
The process of exposing an AI model to a dataset, allowing it to learn patterns, relationships, and features from the data by adjusting its internal parameters (weights). The goal is to enable the model to perform a specific task accurately on new, unseen data.
Training Data Poisoning
A security risk where an attacker manipulates training, fine-tuning, or retrieval data so an AI system learns harmful behavior, leaks information, or gives unreliable outputs. In generative AI systems, poisoning can target datasets, feedback loops, or knowledge bases used by RAG.
Training-serving Skew
Mismatch between training-time and serving-time data, features, or preprocessing logic that causes production performance to diverge from offline evaluation.
Transfer Learning
A machine learning method where a model developed for a specific task is reused as the starting point for a model on a second, different but related task. This is the core principle behind using pre-trained foundation models and then fine-tuning them, saving significant time and computational resources.
Tree of Thoughts (ToT)
An advanced prompting technique where the model explores multiple reasoning paths simultaneously, like a tree search. Each branch represents a different approach to solving the problem. The model evaluates and selects the most promising paths, enabling more complex problem-solving.
TTFB (Time to First Byte)
Latency from request start until the first byte is received. Impacted by network, cold starts, and model prefill.
TTI (Time to Interactive)
Time until a page or app becomes reliably interactive for users. Optimize assets, hydration, and streaming.
TTS (Text‑to‑Speech)
Synthesis of natural‑sounding speech from text.
U-Net
A neural network architecture commonly used as the denoiser in diffusion models. It predicts noise (or related targets) at each step, conditioned on text or other signals.
Unsupervised Learning
A type of Machine Learning where the model learns from unlabeled data, identifying hidden patterns, structures, or relationships within the data without predefined correct answers (e.g., clustering similar customers, anomaly detection).
VAD (Voice Activity Detection)
Detecting speech segments in audio streams to segment or trigger recognition.
VAE (Variational Autoencoder)
A generative model that learns a probabilistic latent representation. In latent diffusion pipelines, a VAE encodes images into a latent space and decodes latents back into pixels.
Vanishing Gradient Problem
Issue in deep networks where gradients become extremely small, preventing effective weight updates in early layers.
Vector Database
A specialized database designed to efficiently store and query high-dimensional vectors, which are numerical representations of data like text, images, or audio (known as embeddings). They are essential for applications like semantic search, recommendation systems, and Retrieval Augmented Generation (RAG).
Vector Index
A data structure optimized for fast approximate nearest neighbor search in high-dimensional vector spaces. Examples include HNSW, IVF, and ScaNN. Essential for efficient semantic search and RAG at scale.
Vector Store
Alias used in practice for a vector database used to store embeddings and run vector similarity search.
Vendor Lock‑in
High switching costs due to proprietary APIs, models, or data formats. Mitigate with standards and abstraction layers.
Vision-Language Model (VLM)
A multimodal model that processes both images and text. VLMs support tasks such as image captioning, visual question answering, screenshot understanding, chart interpretation, and visual reasoning.
vLLM
An open-source LLM inference engine optimized for high throughput and efficient KV cache management. Commonly used for serving open-weight models.
Voice Cloning
AI technology that creates a synthetic replica of a specific person's voice from audio samples. Can be used for personalized TTS, dubbing, or accessibility but raises ethical concerns around consent and deepfakes.
VQA (Visual Question Answering)
Answering natural‑language questions about images or video.
VRAM
Video RAM on a GPU. VRAM capacity is a key constraint for model size, batch size, KV cache, and context length during inference.
War Room
Dedicated communication channel and coordination process used during critical incidents for rapid decision-making and synchronized execution.
Warm Path vs Cold Path
Warm path serves cached/ready resources for low latency; cold path performs full initialization. Architect for warm hits and graceful cold behavior.
Watermarking (AI Outputs)
Techniques to embed invisible markers in AI-generated content to prove authenticity and detect manipulation.
Web Browsing
The act of navigating and interacting with web pages, often used in the context of web scraping or automated browsing.
Webhooks
HTTP callbacks that deliver events to subscribers. Validate signatures, implement retries idempotently, and design for back‑pressure.
Weight Decay
A regularization technique (often L2-like) that penalizes large weights during training. In AdamW, weight decay is applied decoupled from the gradient update.
Word Embeddings
Dense vector representations of words that capture semantic meaning and relationships between words.
Working Memory (Agents)
Short-lived context used for the current task (current goals, temporary variables). Often implemented as in-context state within the prompt.
World Models
Internal representations that allow AI systems to simulate and predict how the environment will change in response to actions. Key for planning, reasoning, and agentic behavior.
Zero-shot Learning/Prompting
An AI model's ability to perform a task it hasn't been explicitly trained on, by leveraging its general knowledge and understanding the instructions provided in the prompt, without any specific examples of that task.