AI Glossary

A Convolutional Neural Network (CNN) is a deep learning architecture that applies learned weight-sharing filters over local patches of input data — most commonly images — to detect hierarchical spatial features such as…

A Decoder-Only architecture is a transformer variant that uses a single self-attention stack with causal (left-to-right) masking to predict each next token from its preceding context, without a separate encoder, and is…

Deep learning is a subfield of machine learning that trains neural networks with many layers to learn hierarchical representations directly from raw data such as pixels, audio waveforms, or text tokens, without…

A diffusion model is a generative AI system that produces images, audio, or video by learning to reverse an iterative noise-addition process, starting from random noise and progressively denoising it through learned…

An embedding model converts text, images, or other data into fixed-length numerical vectors in a high-dimensional space, where semantically similar items are geometrically close to each other.

An Encoder–Decoder architecture is a neural network design in which an encoder maps an input into a latent representation and a separate decoder generates the output sequence from that representation, making it…

A foundation model is a large-scale AI system pre-trained on broad, diverse data—text, images, code—to serve as a general-purpose base adaptable to many downstream tasks. The term was introduced by Stanford HAI in 2021…

A frontier model is an AI system operating at the current limits of machine learning capability, typically trained with the largest known compute budgets. The term is used in AI policy to identify systems whose…

A Generative Adversarial Network (GAN) is a dual-network architecture where a generator produces synthetic data and a discriminator attempts to classify it as real or fake; adversarial training between the two drives…

Generative AI refers to machine learning systems that produce new content—text, images, audio, video, or code—by learning statistical patterns from large training datasets. Unlike discriminative models that classify…

A large language model (LLM) is a transformer-based neural network with billions of parameters trained on internet-scale text to model and generate human language, capable of performing diverse tasks through…

Long context refers to the capacity of a language model to process and reason coherently over very large sequences of tokens in a single input. Context window sizes have grown from hundreds of tokens in early models to…

Machine learning is a branch of artificial intelligence in which algorithms automatically improve their performance on a task by learning patterns from data, without being explicitly programmed with task-specific rules…

Mixture of Experts (MoE) is a neural network architecture where a learned routing mechanism activates only a small subset of specialized sub-networks (experts) for each input token, allowing large total parameter counts…

A multimodal model is an AI system that processes and generates data across more than one modality — such as text, images, audio, or video — within a single unified architecture.

A neural network is a computational model composed of interconnected layers of numerical units (neurons) that learn to map inputs to outputs by adjusting connection weights through exposure to training data.

An open-weights model is an AI system whose trained parameter weights are publicly released, allowing anyone to download, run, and modify the model without accessing the original developer's infrastructure.

A reasoning model is an AI system designed to solve complex, multi-step problems by generating explicit intermediate reasoning steps — often called a chain of thought — before producing a final answer.

A Recurrent Neural Network (RNN) is a neural network that processes sequential data by maintaining a hidden state vector updated at each time step, allowing information from earlier inputs to influence later predictions.

A small language model (SLM) is a language model with a compact parameter count—typically 1 to 13 billion parameters—optimized for efficient inference on consumer devices, edge hardware, or single-GPU servers rather…

Speech recognition (ASR) is a technology that converts spoken audio into written text, using machine learning models trained on large corpora of speech to accurately transcribe words and sentences in real time or from…

State Space Model (SSM) is a class of sequence-processing architectures derived from control theory that represent data streams through a latent state vector updated by linear recurrences, enabling efficient processing…

A text-to-image model is a generative AI system that produces raster images from natural-language text prompts, synthesizing visual content that matches the described scene, style, or subject.

Text-to-speech (TTS) is a technology that converts written text into synthesized spoken audio, using AI models trained on human speech recordings to produce natural-sounding voice output.

A text-to-video model is a generative AI system that synthesizes video clips from natural-language text prompts, producing temporally coherent sequences of frames that match the described motion, scene, or narrative.

A Transformer is a neural network architecture centered on self-attention, which lets every position in a sequence directly attend to every other position simultaneously, enabling fully parallel training and effective…

A Vision-Language Model (VLM) is an AI model that jointly processes visual inputs (images or video) and natural language text, enabling tasks such as image captioning, visual question answering, and document…

A world model is an internal representation that an AI system learns of its environment's dynamics, enabling it to predict the consequences of actions and simulate future states without directly interacting with the…

Backpropagation is an algorithm for computing the gradient of a neural network's loss with respect to its weights by propagating error signals backward through the network layers, enabling gradient-based optimization.

Catastrophic forgetting is the tendency of a neural network to abruptly lose performance on previously learned tasks when trained sequentially on new data, because weight updates for the new task overwrite…

Continual learning is a machine learning paradigm in which a model learns from a continuous stream of tasks or data over time while retaining performance on previously acquired knowledge, without full retraining from…

Data augmentation is the practice of artificially expanding a training dataset by applying label-preserving transformations to existing examples — such as image flipping, cropping, or noise injection — to improve model…

Direct Preference Optimization (DPO) is a training algorithm that fine-tunes language models to align with human preferences by reformulating the RLHF objective as a binary classification loss over preference pairs…

An epoch is one complete pass of the entire training dataset through a machine learning model. Models are typically trained for multiple epochs, each pass refining the model's weights through backpropagation applied to…

Federated learning is a machine learning technique that trains a shared model across many decentralized devices or servers without centralizing raw data, transmitting only model parameter updates to a coordinating…

Fine-tuning is the process of further training a pre-trained AI model on a smaller, task-specific dataset so it performs better on that task. Instead of building a model from scratch, you adapt an existing one to your…

Gradient descent is an iterative optimization algorithm that trains machine learning models by repeatedly adjusting parameters in the direction that most reduces a loss function, using partial derivatives computed via…

Grokking is a delayed generalization phenomenon in neural network training where a model first memorizes training data, then — after prolonged training well past apparent overfitting — abruptly learns the underlying…

Instruction tuning is a supervised fine-tuning technique that adapts a pre-trained language model on instruction-response pairs, teaching it to follow natural-language directives rather than merely predict the next…

Knowledge distillation is a compression technique in which a small student model is trained to match the output distribution of a larger teacher model, producing a compact model that retains much of the teacher's…

LoRA is a parameter-efficient fine-tuning technique that adds trainable low-rank matrix pairs to frozen pre-trained model layers, enabling large-model adaptation with a fraction of the original parameter count.

A loss function is a mathematical function that measures the discrepancy between a model's predictions and the true target values, producing a scalar score that optimization algorithms minimize during training.

A model checkpoint is a saved snapshot of a neural network's weights and optimizer state at a specific point during training, enabling resumption after hardware failures and selection of the best-performing version…

Overfitting occurs when a machine learning model learns the training data too closely—including its noise and idiosyncrasies—resulting in high accuracy on training examples but poor generalization to unseen data.

Pre-training is the initial large-scale training phase in which a neural network learns general representations from a massive corpus using self-supervised objectives, before any task-specific fine-tuning.

QLoRA is a fine-tuning method that quantizes a frozen base model to 4-bit precision while training LoRA adapters at higher precision, enabling large language models to be fine-tuned on a single consumer or professional…

Reinforcement learning is a machine learning paradigm in which an agent learns a decision-making policy by interacting with an environment and receiving scalar reward signals, optimizing for maximum cumulative reward…

Reinforcement Learning from AI Feedback (RLAIF) is a variant of RLHF in which an AI model generates the preference labels used to train the reward model, reducing dependence on costly and hard-to-scale human annotation.

Reinforcement Learning from Human Feedback (RLHF) is a training pipeline that collects human preference judgments between model outputs, trains a reward model on those judgments, and uses reinforcement learning to…

Reinforcement Learning with Verifiable Rewards (RLVR) is a training approach in which RL reward signals come from objective, programmatically checkable criteria — such as numerical correctness of a math answer or code…

Scaling laws are empirical power-law relationships showing that language model performance improves predictably as model parameters, training data volume, and compute budget increase, enabling researchers to forecast…

Self-supervised learning is a training paradigm in which a model generates its own supervisory signal from unlabeled data by solving pretext tasks, eliminating the need for costly human annotations.

Supervised learning is a machine learning paradigm in which a model is trained on a labeled dataset of input-output pairs to learn a mapping function, then applied to predict outputs for unseen inputs; it underlies most…

Synthetic data is artificially generated data — produced by algorithms, simulations, or generative models rather than collected from real-world events — used to train, validate, or test machine learning systems while…

Training data is the labeled or unlabeled dataset fed to a machine learning model during the optimization process, allowing it to adjust internal parameters by minimizing prediction error; its quality, scale, and…

Transfer learning is a technique in which a model pre-trained on one large dataset or task is adapted to a different but related task, substantially reducing the need for labeled data and training compute.

Unsupervised learning is a machine learning paradigm in which models find patterns, structure, or compact representations in data without labeled examples, using techniques such as clustering, dimensionality reduction…

Batching in model inference is the practice of grouping multiple input requests and processing them together in a single forward pass through the model, improving GPU utilization and throughput. Continuous batching…

Chain-of-Thought (CoT) is a prompting technique in which a language model generates explicit intermediate reasoning steps before producing a final answer, substantially improving accuracy on arithmetic, logic, and…

A context window is the maximum number of tokens a language model can process in a single inference call, covering both the input prompt and the generated output. Exceeding it causes input truncation or an API error…

Inference is the process of applying a trained machine learning model to new input data to produce predictions or outputs. It is the deployment-time operation, distinct from training, in which no model parameters are…

KV-Cache (Key-Value Cache) is a memory buffer that stores the key and value tensors produced by a transformer's attention layers for already-processed tokens, eliminating redundant recomputation during autoregressive…

Latency in AI inference is the elapsed time between submitting a request to a model and receiving its response, typically measured in milliseconds. In large language models it is subdivided into time-to-first-token…

Logits are the raw, unnormalized real-valued scores output by the final linear layer of a neural network before softmax normalization; in language models, one logit per vocabulary token represents that token's relative…

Model serving is the infrastructure layer that deploys a trained ML model to handle real-time or batch prediction requests, managing scaling, load balancing, versioning, and reliability in production. It bridges the gap…

On-device AI is the execution of machine learning model inference directly on a user's local hardware—smartphone, laptop, or embedded chip—without transmitting data to a remote cloud server, enabling lower latency…

Perplexity is a language model evaluation metric defined as the exponentiated average negative log-likelihood per token on a text corpus; lower perplexity means the model assigns higher probability to the observed token…

Prompt caching is an API and serving technique that stores the computed KV-cache state for a shared prompt prefix — such as a system prompt or a large document — and reuses it across multiple separate requests…

Quantization is the technique of representing a neural network's weights — and optionally its activations — in lower-precision numeric formats such as INT8 or INT4 instead of the default FP16 or BF16, reducing memory…

Reasoning effort is an inference-time parameter that controls how many internal computation steps or tokens a model allocates to reasoning before producing an answer, allowing users to trade response latency and cost…

Speculative decoding is an inference technique that uses a small draft model to propose multiple tokens in parallel, then verifies them with the large target model in a single forward pass, reducing latency by 2–4×…

Streaming in AI inference is the delivery of model output token by token to the client as each token is generated, rather than waiting for the full response before transmitting anything. It reduces perceived latency to…

Temperature is a scalar hyperparameter that divides a language model's logits before the softmax step, controlling output randomness: values below 1.0 sharpen the distribution toward high-probability tokens; values…

Test-time compute (TTC) is the computational budget—processing cycles, memory, and time—that a model expends during inference rather than training, allowing it to spend more effort on harder problems without any change…

Throughput in AI inference is the volume of work a model serving system processes per unit of time, commonly expressed as output tokens per second or completed requests per second across all concurrent users. It…

A token is the basic unit of text that a language model processes, typically a word, subword fragment, or punctuation mark. In common English prose, one word corresponds to approximately 1.3 tokens under widely used…

Tokenization is the process of splitting raw text into discrete units called tokens — typically subword fragments — that a language model numerically encodes and processes. A token averages roughly 4 characters in…

Top-p (nucleus) sampling is a decoding strategy that restricts token selection to the smallest set of tokens whose cumulative probability meets a threshold p, dynamically adapting the candidate pool size to the model's…

An agent harness is a software framework that wraps a language model with the infrastructure required for autonomous operation—tool invocation, context management, error handling, and execution loops—transforming a…

Agent memory refers to the mechanisms an AI agent uses to store, retrieve, and update information across steps, sessions, or tasks — spanning in-context working memory, external vector databases, key-value stores, and…

Agent orchestration is the coordination of multiple AI agents by a supervisory system that routes tasks, manages dependencies and information flow, and aggregates results to accomplish a complex multi-step goal.

Agent planning is the process by which an AI agent decomposes a complex goal into an ordered sequence of subtasks or actions, selecting and scheduling steps to achieve an objective that cannot be completed in a single…

Agent2Agent Protocol (A2A) is an open specification published by Google in April 2025 that defines how autonomous AI agents discover each other, exchange tasks, and coordinate across different frameworks and vendors.

An agentic workflow is a structured pipeline in which one or more AI agents execute multi-step tasks autonomously—using tools, memory, and decision logic—to complete a goal that requires more than a single model call.

An AI agent is a system where a language model does not just answer, but plans and executes multi-step tasks: it calls tools and APIs, reads the results and decides the next action toward a goal. Unlike a chatbot, an…

An AI agent swarm is a multi-agent architecture where a complex project is decomposed across specialized parallel agents (researcher, implementer, security auditor, independent verifier) coordinated by a central…

An autonomous agent is an AI system that perceives its environment, makes decisions, executes actions, and iterates toward a defined goal with minimal or no human intervention between steps, operating over extended time…

A Browser Agent is an AI system that autonomously controls a web browser—navigating pages, clicking links, filling forms, and extracting information—to complete web-based tasks on a user's behalf without step-by-step…

Claude Code is Anthropic's terminal-native agentic coding tool that operates directly inside a project's filesystem: reading and editing files, running shell commands, spawning subagents, and connecting to external…

A Coding Agent is an AI system that autonomously writes, edits, tests, and debugs software across multiple files and execution environments, translating high-level task descriptions into working, verified code changes.

Computer Use is an AI capability, first released publicly by Anthropic in October 2024, that allows a language model to control a computer's graphical interface—moving the cursor, clicking buttons, typing text, and…

Deep Research is an agentic AI capability in which a model autonomously plans and executes a multi-step research process—issuing search queries, reading web pages, and synthesizing sources—to produce a comprehensive…

Function calling is a language model API feature, first formalized by OpenAI in June 2023, that enables a model to output structured JSON specifying which predefined function to invoke and with what arguments, enabling…

Human-in-the-Loop (HITL) is a system design pattern in which human oversight is structurally embedded at one or more decision points in an automated AI pipeline, requiring a person to approve, correct, or redirect the…

Mixture of Agents (MoA) is an ensemble framework in which multiple large language models independently generate responses to a query, and one or more aggregator models synthesize those outputs into a single…

Model Context Protocol (MCP) is an open standard published by Anthropic in November 2024 that defines a universal interface for connecting AI assistants to external tools, databases, and services, replacing ad-hoc…

Model Context Protocol (MCP) is an open standard that connects AI models and autonomous agents to external data sources, databases, browsers, and APIs through a unified interface ('USB-C for AI').

A multi-agent system (MAS) is an architecture in which multiple autonomous AI agents collaborate, each with distinct roles and capabilities, to solve tasks too complex or large for a single agent to handle alone.

ReAct (Reasoning + Acting) is a prompting framework for AI agents that interleaves chain-of-thought reasoning steps with discrete tool-use actions, allowing the model to observe each action's result and update its…

A sandbox is an isolated execution environment that restricts an AI agent's access to real systems, networks, or data, allowing it to run untrusted or potentially harmful code without affecting the host environment.

A subagent is an AI agent invoked by a parent or orchestrator agent to execute a specific bounded subtask autonomously and return results, without direct interaction with the end user during its execution.

Tool use is the capability of an AI language model to invoke external functions, APIs, or services—such as web search, code execution, or database queries—during inference, allowing it to retrieve information and take…

AI alignment is the field of research and engineering concerned with ensuring that AI systems pursue goals and exhibit behaviors consistent with human intentions, values, and interests, particularly as systems become…

AI bias refers to systematic, skewed patterns in an AI system's outputs that arise from biased training data, flawed problem framing, or model design choices, causing the system to treat individuals or groups…

AI safety is the interdisciplinary field aimed at ensuring that artificial intelligence systems operate reliably, predictably, and without causing unintended harm, spanning technical research, policy, and governance.

AI Slop is low-quality, mass-produced content generated by AI systems—characterized by generic phrasing, factual errors, and lack of original insight—that floods websites, social media, and search results at scale.

AI watermarking is a technique for embedding imperceptible signals into AI-generated content — text, images, audio, or video — to enable later identification of its machine-made origin. It supports provenance…

A benchmark is a standardized test set used to measure and compare AI model performance on defined tasks. Benchmark scores provide a common scale for tracking progress across models and communicating capability claims…

Constitutional AI is a training method developed by Anthropic in which an AI model critiques and revises its own outputs according to a written set of principles, reducing reliance on large-scale human annotation to…

Content moderation is the process of reviewing, filtering, and actioning user-generated content on digital platforms to enforce community guidelines and legal requirements, increasingly performed or assisted by AI…

Context rot is the degradation of a language model's effective reasoning and retrieval accuracy as its context window fills with text, causing it to underweight or lose track of information—especially content placed in…

Data privacy in AI refers to the practices, technical controls, and legal requirements governing how personal information is collected, stored, used, and protected throughout AI system development and deployment. It…

A deepfake is AI-generated or AI-manipulated media—video, audio, or images—in which a person's likeness or voice is fabricated or replaced to create realistic but false depictions, typically using deep learning…

Guardrails are safety mechanisms—rules, filters, classifiers, or policy constraints—applied to AI systems to prevent them from producing harmful, inappropriate, or policy-violating outputs.

A hallucination is a confident, fluent output from an AI language model that is factually incorrect, fabricated, or not grounded in its input or knowledge, produced without any indication of uncertainty.

Interpretability is a field of AI research aimed at understanding the internal computations of neural networks — how they represent knowledge, form decisions, and produce outputs — in order to verify their safety and…

A jailbreak is a technique used to circumvent an AI model's built-in safety guidelines, causing it to produce content or perform actions that its developers explicitly designed it to refuse.

Mechanistic interpretability is a subfield of AI research that attempts to reverse-engineer the internal computations of neural networks—identifying specific circuits, features, and algorithms that produce model…

Model collapse is the progressive degradation of an AI model's output quality and diversity when trained on AI-generated data, causing successive model generations to lose coverage of rare but important patterns present…

Model evaluation (evals) is the systematic process of measuring an AI model's performance, capabilities, and failure modes using structured tests. Evals guide development decisions and are increasingly required by…

Prompt injection is an attack in which malicious instructions embedded in an AI model's input override the system's original directives, causing it to produce unintended or harmful outputs.

Red teaming is a structured adversarial testing process in which a dedicated team attempts to find failures, harmful outputs, or security vulnerabilities in an AI system before it is deployed.

A refusal is an AI model's deliberate decision to decline a user's request, typically because the system's safety training determined the request could lead to harmful, illegal, or policy-violating outputs.

Reward hacking is a failure mode in reinforcement learning where an agent discovers unintended strategies that maximize its numerical reward signal without accomplishing the task the designers actually intended.

A sparse autoencoder (SAE) is a neural network trained to reconstruct inputs via a large overcomplete hidden layer—most units inactive for any given input—used in mechanistic interpretability to decompose superimposed…

An AI accelerator is any processor or hardware block—GPU, TPU, NPU, FPGA, or custom ASIC—engineered to speed up the matrix math and data-parallel operations central to training and running artificial intelligence models.

An AI data center is a facility purpose-built or heavily retrofitted to house GPU and accelerator clusters, designed specifically to handle the extreme power density, cooling demands, and high-bandwidth networking that…

An ASIC (Application-Specific Integrated Circuit) is a chip designed and fabricated to perform one specific function, delivering higher performance and energy efficiency for that task than a general-purpose processor…

CUDA (Compute Unified Device Architecture) is NVIDIA's proprietary parallel computing platform and programming model that lets developers write C/C++ code executed directly on NVIDIA GPUs, turning them into…

FLOPS (Floating-Point Operations Per Second) measures how many floating-point arithmetic operations a processor executes each second. It is the primary benchmark for comparing AI hardware, with modern high-end GPUs…

A GPU is a processor originally designed for computer graphics rendering, built around thousands of small parallel compute cores. Since the early 2010s, GPUs have become the dominant hardware for AI model training and…

A GPU cluster is a collection of servers, each containing multiple GPUs, interconnected by high-speed networking to act as a single parallel compute resource for training or serving AI models that exceed the capacity of…

High-Bandwidth Memory (HBM) is a stacked DRAM technology that bonds multiple memory dies vertically and connects them to a GPU via a very wide internal bus, achieving memory bandwidths of several terabytes per second…

An interconnect is the high-speed communication link that transfers data between GPUs within a server (intra-node, e.g., NVLink) or between servers in a cluster (inter-node, e.g., InfiniBand); its bandwidth and latency…

An NPU is a dedicated hardware block integrated into a system-on-chip (SoC) that accelerates neural network inference locally on a device, enabling AI features—like voice recognition or image enhancement—without…

A TPU is a custom ASIC designed by Google specifically to accelerate tensor operations—the matrix multiplications at the core of neural network training and inference—offering higher throughput per watt than…

VRAM (Video Random Access Memory) is the dedicated high-speed memory on a GPU used to store model weights, activations, and KV cache during AI training and inference. Its capacity is the primary constraint on which…

An AI API is a standardized web interface that lets developers send data to a hosted AI model and receive generated text, predictions, or other outputs in return, without deploying their own model infrastructure. Major…

An AI lab is an organization — inside a company, university, or nonprofit — dedicated to researching and building artificial intelligence systems, from foundational models and safety techniques to applied tools and APIs.

An AI moat is a durable competitive advantage in artificial intelligence that is difficult for rivals to replicate, such as proprietary training data, distribution scale, specialized talent, regulatory approval, or deep…

AI regulation refers to laws, rules, and standards established by governments and regulatory bodies to govern the development, deployment, and use of artificial intelligence systems, addressing risks such as…

An AI wrapper is a product or service built by calling a third-party AI model API — such as OpenAI's GPT or Anthropic's Claude — and adding a user interface, prompt templates, or workflow logic, without training or…

ARC-AGI (Abstraction and Reasoning Corpus for AGI) is a benchmark of visual analogy puzzles requiring systems to infer transformation rules from a few input-output grid examples and apply them to new grids — tasks…

Artificial General Intelligence (AGI) is a hypothetical AI system capable of performing any intellectual task a human can, matching or exceeding human cognitive flexibility across all domains without task-specific…

Artificial Superintelligence (ASI) is a hypothetical AI system whose cognitive capabilities across every domain substantially exceed those of the most capable humans. It remains a theoretical concept as of 2026, well…

A chatbot is a software program that conducts text or voice conversations with users, either by following scripted decision trees or, in modern implementations, by generating free-form responses using a large language…

Compute, in AI contexts, refers to the computational resources — primarily GPU or TPU processing power, memory, and supporting infrastructure — required to train and run AI models. It is the primary limiting resource…

A copilot is an AI assistant embedded within a software application that provides real-time suggestions, automates subtasks, and answers natural-language questions, keeping the human user as the primary decision-maker…

Enterprise AI refers to the deployment of artificial intelligence systems within large organizations to automate processes, augment decision-making, and generate business value at scale, integrated with enterprise…

The EU AI Act (Regulation (EU) 2024/1689) is the world's first comprehensive horizontal AI regulation, adopted by the European Union in 2024, which classifies AI systems by risk tier and imposes requirements ranging…

GPQA (Graduate-Level Google-Proof Q&A) is a benchmark of 448 expert-written multiple-choice questions in biology, chemistry, and physics, designed so that non-specialists cannot answer them correctly even with…

A GPU cloud, often called a neocloud, is a cloud provider that specializes in renting high-density GPU clusters for AI training and inference, rather than offering the broad portfolio of managed services that…

HumanEval is a code-generation benchmark of 164 hand-written Python programming problems created by OpenAI in 2021, used to measure a language model's ability to produce functionally correct code, scored via the pass@k…

A hyperscaler is one of a small group of cloud providers — primarily Amazon Web Services, Microsoft Azure, and Google Cloud — operating data center infrastructure at planetary scale with comprehensive portfolios of…

LM Arena is a crowd-sourced AI evaluation platform where users compare two anonymous model responses side-by-side and vote on the better one, generating an Elo-based leaderboard of language model quality across…

MMLU (Massive Multitask Language Understanding) is a benchmark evaluating language models on approximately 14,000 multiple-choice questions across 57 academic subjects — from calculus and medicine to law and ethics…

Needle in a Haystack is a long-context evaluation test that measures whether a language model can accurately retrieve a specific planted fact (the "needle") from a large irrelevant text corpus (the "haystack") at…

Open-source AI refers to AI models, frameworks, or systems whose weights, code, or architecture are publicly released — allowing anyone to download, inspect, modify, and deploy them — in contrast to closed proprietary…

Sovereign AI refers to national or regional efforts to develop, own, and control AI infrastructure, models, and data pipelines domestically — rather than depending on foreign commercial providers — to preserve…

SWE-bench is a benchmark that evaluates AI coding systems on real software engineering tasks by requiring them to resolve genuine GitHub issues in open-source Python repositories, with success defined as producing a…

Token economics is the pricing model governing commercial access to AI language models via API, where users are charged per token — roughly 3–4 characters of text — with separate rates for input (prompt) and output…

Agentic RAG is an architecture in which a language model acts as an autonomous agent that iteratively decides when and what to retrieve from external sources, enabling multi-step reasoning over complex queries that…

The attention mechanism is a neural network component that allows a model to dynamically weight the relevance of different input positions when computing each output, enabling context-sensitive processing over sequences…

Chunking is the process of splitting source documents into smaller text segments before embedding and storing them in a vector database, enabling efficient and precise retrieval in RAG and search systems.

Context engineering is the discipline of designing the dynamic information environment fed to an LLM or agent — system rules, layered memory retrieval, tool schemas, and verification constraints — superseding simple…

An embedding is a dense, fixed-length numerical vector that represents data — such as text, an image, or audio — in a high-dimensional space where semantically similar items are located geometrically close to each other.

Few-shot learning is a machine learning paradigm in which a model adapts to a new task using only a small number of labeled examples—typically between one and twenty—rather than the large labeled datasets required by…

Flash Attention is an IO-aware, exact attention algorithm that computes scaled dot-product attention using GPU on-chip SRAM tiling to dramatically reduce high-bandwidth memory access. It achieves 2–4x faster training…

GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) are the practices of structuring content, schema markup, and entity authority so AI search engines (ChatGPT Search, Perplexity, Claude, Gemini…

GraphRAG is a retrieval-augmented generation approach that structures knowledge as a graph of entities and relationships rather than flat text chunks, enabling multi-hop reasoning and synthesis across interconnected…

Grounding is the practice of connecting an LLM's output to external, verifiable information sources—such as retrieved documents, live web results, or structured databases—to reduce hallucination and improve factual…

In-context learning is the ability of a large language model to perform new tasks by reading examples or instructions placed directly in the input prompt, adapting its behavior at inference time without any update to…

A knowledge graph is a structured database representing real-world entities as nodes and their semantic relationships as labeled edges, enabling machines to traverse factual networks and support multi-hop reasoning…

Model routing is the practice of automatically directing each AI inference request to the most appropriate model from a fleet — balancing cost, latency, and quality — rather than sending all queries to a single model.

OCR (Optical Character Recognition) is a technology that converts images containing printed or handwritten text into machine-encoded, editable text. It enables computers to read and process documents that exist only in…

Prompt engineering is the practice of designing and refining text inputs to guide large language models toward accurate, consistent, or specifically formatted outputs, using techniques such as instruction specification…

RAG (retrieval-augmented generation) is a technique that lets a language model pull relevant documents from an external knowledge base before answering and ground its response in them. It reduces hallucinations and…

Reranking is a two-stage retrieval technique that first retrieves a broad candidate set using a fast bi-encoder or keyword method, then re-scores those candidates with a slower, more accurate cross-encoder model to…

Semantic search is a retrieval method that matches queries to documents based on conceptual meaning rather than keyword overlap, using embedding vectors and similarity metrics to surface contextually relevant results…

Structured output is a technique for constraining a language model's generation to conform to a predefined schema — such as JSON or XML — so that downstream systems can reliably parse and consume the result without…

A system prompt is a set of instructions provided to a language model before any user input, defining its persona, constraints, output format, and behavioral rules that persist throughout the entire conversation.

Tree of Thought (ToT) is a prompting framework for large language models that explores multiple candidate reasoning paths simultaneously, enabling deliberate backtracking and evaluation of intermediate steps rather than…

A vector database is a data store purpose-built for storing high-dimensional numerical vectors and performing fast approximate nearest-neighbor (ANN) search across millions or billions of them, enabling semantic…

Vibe coding is the practice of building production web apps, Telegram bots, and autonomous AI agents by directing agentic coding tools (like Claude Code and Cursor) in natural language rather than writing boilerplate…

Zero-shot learning is the ability of a machine learning model to correctly handle tasks or classify inputs from categories it was never shown during training, by leveraging learned semantic relationships and broad…