Glossary

GLOSSARY

Plain-language definitions of the terms that appear across our AI news and editorials. Dotted words in articles open their meaning on tap.

Fundamentals

LLMLarge Language Model: An AI model trained on vast amounts of text that generates language by predicting the most likely next word. It underpins systems like ChatGPT, Claude, and Gemini.
Token: The smallest unit of text an AI processes, formed by splitting text into chunks of words or characters. The amount an AI can handle at once (its context) is capped.
ContextContext window: The span of information an AI can take in and consider at once. Cramming in too much irrelevant material can actually lower the quality of its output.
Prompt: The instruction or question given to an AI. How it is written can substantially change the quality of the result.
Inference: The step where a trained AI actually produces an answer to an input. It is the counterpart to training.
Multimodal: The ability of an AI to handle several kinds of input together—text, images, audio, and video—rather than text alone.
Open-weight: An AI model whose weights (its trained parameters) are published, so anyone can download, run, and modify it.
Open source: Released under terms that let anyone use, study, modify, and share it. For AI, the Open Source Initiative's definition (OSAID 1.0) requires more than the weights: information about the training data and the training code must be open as well. Weights can be downloadable yet carry conditions on commercial use, so an open-weight model is not necessarily open source.
GPTGenerative Pre-trained Transformer: OpenAI's family of large language models, and the name of the underlying approach: a pre-trained Transformer that generates text.
AGIArtificial General Intelligence: A hypothetical AI able to perform the full breadth of intellectual tasks at a human level, not just narrow ones. Its definition and timeline remain contested.
ModalityModality: A type or channel of information an AI works with—text, image, audio, or video, for example. Handling several at once is called multimodal.
Frontier modelFrontier model: The largest, most capable AI models at the leading edge of what is currently possible—such as the latest generations of GPT, Claude, and Gemini.
Generative AIGenerative AI: A broad term for AI that can produce new text, images, audio, or code. Trained on large amounts of data, it generates human-like outputs in response to instructions (prompts).
Parameter: The internal numbers a model settles on through training, commonly cited as a measure of its size. Totals like one trillion are quoted for scale, while mixture-of-experts models also report how many actually fire on a single inference (active parameters).
LTMLarge Tabular Model: A foundation model built for tabular, structured data. Where an LLM works over words, an LTM predicts from rows and columns such as transaction histories or customer records.

Connectivity

MCPModel Context Protocol: An open standard for connecting AI to external data and tools. Proposed by Anthropic in 2024 and since adopted across the industry, it is often called the USB-C of AI.
APIApplication Programming Interface: An interface that lets software exchange functions and data directly, program-to-program, without going through a screen.
CLICommand-Line Interface: A way of operating software by typing text commands. It suits automation and repetitive tasks better than clicking through a graphical screen.
Context engineeringContext engineering: The skill of designing what an AI reads in and what it does not. As the number of connectable sources grows, accuracy hinges less on connecting everything than on offering only what is needed, neatly arranged.
SDKSoftware Development Kit: A bundled set of tools for building software on top of a service or capability. It serves as the scaffolding for embedding AI into apps and agents.

Risks

Context pollutionContext confusion: A phenomenon in which irrelevant, duplicated material piles up in an AI's context and clouds its reasoning. When many similar tools line up, it may pick the wrong one or invent tool names that do not exist, lowering accuracy.
Hallucination: When an AI produces convincing-sounding output that is actually false. It makes fact-checking essential.
Shadow AIShadow AI: Generative AI tools or services that employees use for work without official approval or oversight. This 'shadow' use that IT cannot see spreads easily because it is convenient, yet it carries risks of data leakage and weak control.
AI governanceAI governance: The frameworks and rules an organization puts in place to manage the risks of using AI and to ensure safe, appropriate use. It turns legal compliance, information management, and accountability into concrete design.
GuardrailsGuardrails: Mechanisms that set bounds on an AI's outputs and actions to stop dangerous deviations or runaway behavior. They are the safety rails for running agents, paired with the harness that supplies roles and procedures.
Recursive self-improvementRSI / Recursive self-improvement: A hypothesized loop in which an AI autonomously designs and trains a more capable successor, repeating the cycle to grow rapidly. If progress outpaces our ability to control it, humans could lose the initiative—making it a central concern in AI safety.
Sandbox: An execution environment walled off from everything else, used to keep whatever an AI runs or touches contained inside it. It is the premise for letting agents act safely—though cases of models slipping out through gaps such as network access have been reported.
Prompt injection: An attack that hides malicious instructions inside text an AI reads, hijacking the instructions it was actually given. When the payload arrives through outside material such as a web page or document, it is called indirect prompt injection.
Jailbreak: A technique that talks an AI around its built-in safety controls, usually through crafted phrasing. Because it can pull out responses the model should refuse, it has become a factor in whether a model is released at all.
DMCADigital Millennium Copyright Act: A 1998 U.S. amendment to copyright law. It sets out the notice-and-takedown safe harbor that shields platforms which remove flagged material, and separately bans circumventing technical protection measures—the latter being what AI-scraping suits have turned on. Japanese copyright law has no statute of the same name.
DMADigital Markets Act: The EU's Digital Markets Act, adopted in 2022. It places obligations—such as bans on self-preferencing and duties to grant rivals interoperability—on designated large platforms known as gatekeepers, with the European Commission enforcing them case by case through formal decisions.
CSAMChild Sexual Abuse Material: The internationally used term for imagery depicting the sexual abuse of children. In the United States, providers that detect it on their services must report it to the National Center for Missing & Exploited Children (NCMEC), and how AI-generated imagery fits that framework is now being litigated.

Training

Fine-tuning: Additional training that adapts an already general-purpose model to a specific task or dataset.
Distillation: A technique that trains a smaller model to imitate a larger one, keeping much of the performance while reducing size and cost.
Dataset: A curated body of data used to train or evaluate a model. What goes into it shapes performance—and, where copyrighted works are included, it has become a point of legal dispute.
Scaling laws: The empirical rule that model performance improves as compute, data, and model size grow. Lately the constraints have shifted to power and construction lead times, and talk has turned to a ceiling on scaling.

Applications

RAGRetrieval-Augmented Generation: A technique where an AI first retrieves external documents, then generates its answer grounded in them—making it easier to reflect fresh or proprietary information.
AI agent: An AI that takes a goal and works through multiple steps on its own, calling the tools and data it needs along the way.
HarnessHarness: The surrounding machinery that lets a model use tools, take steps, and loop—things a bare model cannot do on its own. It is the scaffolding that wires a model to actions like file edits and commands and drives the plan-and-execute cycle.
GroundingGrounding: Anchoring an AI's answers to trustworthy sources—internal documents or search results—so the output stays tied to facts. The aim is to curb made-up answers (hallucination).
Physical AI: A term for AI that acts on the physical world—robots, vehicles, factory equipment—as opposed to generative AI that stays on the screen. Development is converging on single models that handle vision, language, and action together.
Coding agent: An AI agent that does more than write code: it runs, tests, and fixes it on its own. Unlike line-by-line completion, you hand it a task and it works the whole loop through.
OCROptical Character Recognition: Technology that reads characters out of images or PDFs and turns them into usable text. Recent AI models go further, classifying regions such as tables, formulas, and signatures so the document's structure comes out with it.
Chatbot: Software that responds in the form of a conversation. Since generative AI, most are built on LLMs, and the word now usually points to the chat window itself as the way in—ChatGPT being the archetype.

Infrastructure

NPUNeural Processing Unit: A processor specialized for AI computation. Built into phones and PCs, it enables AI to run on the device itself (on-device AI).
GPUGraphics Processing Unit: A processor built for massive parallel computation. It has become the core compute resource behind both training and running AI.
CUDACUDA: NVIDIA's software platform for general-purpose computing on its GPUs. Because much AI development is built on this environment and its libraries, it works as a moat that locks in developers.
DRAMDynamic Random Access Memory: The volatile memory used as a computer's main working memory. The high-bandwidth memory (HBM) for AI is built by stacking DRAM, and the market has consolidated into a few suppliers.
HBMHigh Bandwidth Memory: High-bandwidth memory stacked right beside the compute chip. AI workloads move huge amounts of data fast, so its supply shapes both the performance and the volume of AI chips.
CoWoSChip on Wafer on Substrate: An advanced packaging technique that mounts a compute chip and high-bandwidth memory densely on one substrate and wires them together. It is the core step that turns AI silicon into a usable product, and lately the main supply bottleneck.
FoundryFoundry: A chip plant that manufactures other companies' designs. The industry runs on a split between design-only firms (fabless) and manufacturing-only firms (foundries).
EUV lithographyExtreme Ultraviolet: A lithography technique that uses extreme ultraviolet light to print the finest circuits on leading-edge chips. Only one vendor can supply the machines commercially, making it a single point in the chip supply chain.
Process nodeProcess node: The generation label for how fine a chip's circuitry is—smaller numbers like 3nm mean denser, higher-performance chips. Each step finer sharply raises manufacturing difficulty and cost, so few plants can handle the leading edge.
Data center: A facility that houses and runs servers and network equipment at scale. Because training and running AI keeps large fleets of GPUs busy for long stretches, securing power and cooling becomes the binding constraint on where and how fast one can be built.
Gigawatt: A unit of electrical capacity; one gigawatt equals one million kilowatts. AI data centers are now routinely sized in gigawatts—roughly one large power plant per unit—making it the currency in which build-outs are announced.

Evaluation

Benchmark: A shared set of tasks used to score and compare AI model performance, such as GPQA or LiveCodeBench.
SWE-bench: A benchmark measuring how well an AI can resolve real-world software issues on its own. It is widely cited as a yardstick for AI coding ability.
GPQAGraduate-Level Google-Proof Q&A: A benchmark of graduate-level science questions, designed to be hard to answer by web search alone, used to gauge an AI's expert-level reasoning.
AIMEAmerican Invitational Mathematics Examination: A benchmark drawn from a challenging U.S. mathematics competition. Its multi-step problems test an AI's mathematical reasoning.