AI glossary

Plain-language explanations of the terms you meet in AI news, written for first-time readers. 261 terms so far.

Terms in this week's stories

1. Inference10
2. Attention4
3. MCP4
4. MoE4
5. Benchmark3
6. Decision model3
7. HBM3
8. World model3

Basics

First terms for reading AI news

• A/B testingBeginner A/B testing is an experimental method that randomly splits users into two groups, shows one group the existing version (A) and the other a new version (B), and then compares with real data which one produces better results.
• AGIBeginner AGI (artificial general intelligence) refers to AI that can do most of the intellectual work people do, not just one specific task, as well as or better than humans. There is still no agreed-upon precise definition or test for it.
• AI agentBeginner An AI agent is an AI that, given a goal, makes its own plan and uses tools to carry out multi-step tasks directly. It goes one step beyond a chatbot that only answers, to an AI that "gets work done."
• AI platformBeginner An AI platform is a shared foundation that brings together, in one place, the tools and infrastructure needed to build and serve AI, from data preparation to model training, deployment and operations. It is also called an ML platform.
• Artificial neural networkBeginner An artificial neural network is a computing structure that connects small units that take numbers and compute (neurons) in multiple layers, and learns rules by adjusting the strength of those connections on its own as it sees data. It is the basic backbone of today's deep learning and LLMs.
• Batch processingBeginner Batch processing is a method of collecting data over a period of time and processing it all at once at a set time. Totaling a day's sales every night is a typical example.
• Coding agentBeginner A coding agent is an AI program in which an LLM takes a programming task requested in plain language, plans it on its own, and carries it through to the end by reading, editing and running code.
• Context windowBeginner A context window is the maximum amount of text an AI model can read, keep in mind and work with at one time. It is usually expressed as a number of tokens.
• Cross-validationBeginner Cross-validation is a way to measure performance more reliably by splitting data into several parts and training and evaluating the model several times, taking turns using each part for evaluation.
• Data labelingBeginner Data labeling is the work of attaching correct-answer tags (labels) such as 'this is a cat' or 'this sentence is negative' to raw data like photos, text, audio and video. It is also called annotation, and it amounts to creating the teaching material that tells an AI what to learn.
• Deep learningBeginner Deep learning is a machine learning method that stacks many layers of artificial neural networks, modeled on the neural networks of the human brain, so that the model learns complex patterns from data on its own.
• DeepfakeBeginner A deepfake is an image, video or audio clip in which AI realistically synthesizes or swaps a real person's face, voice or actions. The word combines deep learning and fake.
• EU AI ActBeginner The EU AI Act is an artificial intelligence regulation created by the European Union (EU). It sets different obligations depending on how much risk an AI system could pose to people.
• Foundation modelBeginner A foundation model is a large AI model trained broadly in advance on a huge amount of data, so that with a little extra training or some instructions it can be used for many different tasks. Unlike a model that does only one job, it serves as the 'foundation' for many services.
• HallucinationBeginner Hallucination is when an AI confidently produces content that is untrue or unsupported as if it were fact.
• HumanoidBeginner A humanoid is a robot with a body structure similar to a person's, such as a head, torso, two arms and two legs. The aim is to use spaces and tools made for people as they are.
• JailbreakBeginner Jailbreaking is getting around the safety rules built into an AI model with cleverly worded questions or scenarios, so that it gives answers it would normally refuse.
• LLMBeginner An LLM (large language model) is an AI program that learns from a very large amount of text so it can understand and write text the way people do. It answers questions, summarizes, translates and writes.
• Machine learningBeginner Machine learning is a way of having a computer find rules (patterns) on its own by looking at lots of data and use them to make predictions or decisions, instead of people defining every rule by hand.
• MultimodalBeginner Multimodal describes the ability of AI to understand or generate several kinds of information together, such as images, sound and video, not just text.
• ParameterBeginner Parameters are the internal numbers an AI model adjusts on its own during training. The knowledge and way of judging that the model has learned are stored in these numbers.
• Pre-trainingBeginner Pre-training is the first training stage, in which an AI model reads a very large amount of data such as web text, books and code to learn basic knowledge about language and the world. The model that finishes this stage is then refined for various uses.
• PromptBeginner A prompt is the input that tells an AI what you want it to do. Questions, instructions, examples and reference material can all be part of a prompt.
• Prompt engineeringBeginner Prompt engineering is the work of designing and refining questions or instructions (prompts) so that an AI model produces the desired answer. It is a way to improve results by changing only the input text, without retraining the model.
• Recommender systemBeginner A recommender system is software that picks and shows products, videos, articles and other items a person is likely to enjoy, based on the user's past behavior and the characteristics of the content.
• Regression analysisBeginner Regression analysis is a method that finds the relationship between several input values and an outcome value, and then predicts a continuous number such as a house price, sales figure or temperature when a new input comes in. It is one of the most basic forms of supervised learning.
• Reinforcement learningBeginner Reinforcement learning is a way of training in which, instead of learning the correct answers in advance, an AI acts on its own and uses the rewards (scores) it receives for the results to find better actions.
• SLMBeginner An SLM (small language model) is a language model with relatively few parameters, so it is lightweight and runs fast. There is no fixed size threshold; "small" is a relative term compared with large LLMs.
• Sovereign AIBeginner Sovereign AI is a strategy in which a country (or company) aims to build up the core elements of AI — data, computing infrastructure, models and talent — so that it can control them itself without relying on other countries or outside companies.
• Speech recognitionBeginner Speech recognition (ASR, Automatic Speech Recognition, also called STT) is a technology that listens to what a person says and transcribes it into text.
• Structured dataBeginner Structured data is data organized in a table with defined rows and columns. Like an Excel sheet or a database table, what kind of value goes in each cell is decided in advance.
• Supervised learningBeginner Supervised learning is a machine learning approach that shows an AI data in which questions are paired with correct answers, so it learns on its own the rule 'for this kind of input, give this answer.'
• System promptBeginner A system prompt is a set of base instructions placed in an AI model before the user starts a conversation. It defines what role the model plays, what tone it answers in, and what it must not do.
• TokenBeginner A token is the smallest unit an AI language model uses when processing text. It may be a whole word, part of a word, or a few characters.
• TTSBeginner TTS (Text-to-Speech, speech synthesis) is technology that turns written sentences into speech that sounds like a human voice.
• Unsupervised learningBeginner Unsupervised learning is a machine learning approach in which the AI is given only data without answers attached and, on its own, groups similar items or finds hidden structure in it.
• Vibe codingBeginner Vibe coding is a way of programming in which, instead of writing code line by line, you describe what you want to build in words and have AI write the code. The person runs the result and keeps asking for changes in words until they are satisfied.

Models & generation

Language, generative, multimodal and speech models

• Activation functionBasic An activation function is a small transformation rule that decides how much of the signal each neuron in an artificial neural network passes to the next layer, and in what shape. This function is what allows a neural network to learn complex patterns beyond simple straight-line relationships.
• ANN searchAdvanced ANN search (Approximate Nearest Neighbor search) is a method for finding the items most similar to a query among a very large amount of data, not "exactly" but "almost exactly", in exchange for much higher speed. It is a core technology of vector databases and RAG retrieval.
• AttentionAdvanced Attention is a method by which an AI model, when processing input such as a sentence, gathers information by giving more weight to other words that are closely related to the word it is currently handling. It is a core component of the transformer architecture.
• Autoregressive modelAdvanced An autoregressive model produces its output by repeatedly looking at what it has generated so far, predicting the next piece, appending that piece to the input, and predicting the next one again. Most of today's LLMs write text this way.
• BackpropagationBasic Backpropagation is a method that passes the difference (error) between a neural network's answer and the correct answer backward from the output layer toward the input layer, calculating how much each weight is responsible for the error. The network is trained by adjusting the weights based on these values.
• ChunkingBasic Chunking is the process of splitting a long document into small pieces called 'chunks' so that AI can search and use it. It is a preparatory step mainly used when building systems such as RAG, in which AI finds external material to answer questions.
• Context engineeringBasic Context engineering is the work of designing all the information an AI model (LLM) sees when it produces an answer. It decides which of the instructions, conversation history, retrieved documents and tool results to include, how much, and in what order.
• CoTBasic CoT (Chain-of-Thought) is an approach in which an AI model does not answer immediately but reaches the answer by writing out its intermediate reasoning step by step.
• Decision modelBasic A decision model is an AI model that, instead of generating sentences, judges with probabilities which of a set of predefined options applies. For example, it decides whether a post violates a policy or not.
• Diffusion modelBasic A diffusion model is a generative AI model that starts from random noise and removes the noise little by little to create data such as images, video or sound. During training, it learns to identify the noise in noisy data.
• EmbeddingBasic An embedding is a representation that turns data such as words, sentences or images into a list of numbers (a vector) so that items with similar meanings end up close together. A computer measures the distance between these numbers to judge "how similar" things are.
• Fine-tuningBasic Fine-tuning is additional training that adapts an AI model, already trained broadly, to a specific field or task.
• GANBasic A GAN (generative adversarial network) is a method that pits two neural networks against each other, a generator that creates fake data and a discriminator that tells real from fake, so that the generator produces increasingly realistic images and other data.
• Gradient descentBasic Gradient descent is a way for AI models to learn by adjusting their parameters little by little in the direction that reduces how wrong they are (the loss). Most deep learning models today are trained with this method or variations of it.
• GraphRAGAdvanced GraphRAG is a retrieval-augmented generation (RAG) approach that organizes the entities in documents, such as people, companies and concepts, and their relationships into a 'knowledge graph', then uses this structure to help AI answer questions. It is especially strong at questions that connect scattered information or span an entire document collection.
• Hybrid searchAdvanced Hybrid search is a technique that runs two or more different search methods together and then merges the results into a single ranked list. It commonly combines 'keyword search,' which matches words exactly, and 'semantic search,' which finds documents with similar meaning.
• LoRAAdvanced LoRA (Low-Rank Adaptation) is a fine-tuning method that leaves a huge AI model's original parameters unchanged and trains only a very small add-on component alongside them to adapt the model to a new use.
• Loss functionBasic A loss function is a rule that calculates, as a single number, how far an AI model's prediction is from the correct answer. Training is the process of gradually adjusting the model to make this number (the loss) as small as possible.
• MoEAdvanced MoE (Mixture of Experts) is an architecture that places several small neural networks called "experts" inside a model and, for each input, picks only some of them to do the computation. This lets the model as a whole be large while reducing the amount of computation used at any one time.
• PerceptronBasic A perceptron is the simplest artificial neuron model: it multiplies each of several inputs by a weight, adds them up, and outputs 1 if the sum exceeds a threshold and 0 otherwise. It is the basic unit from which artificial neural networks started.
• Post-trainingBasic Post-training is the stage that refines a model that has finished pretraining so people can use it well. It additionally teaches the model things like following instructions, holding conversations, answering safely and reasoning.
• RAGBasic RAG (retrieval-augmented generation) is a method in which an AI first looks up relevant material in documents or databases before answering, and then builds its answer based on what it found. It lets the model use up-to-date information or internal company information it was never trained on.
• Reasoning modelBasic A reasoning model is a large language model (LLM) trained not to answer right away, but to first go through a long 'thinking process' that works through the problem step by step before giving a final answer.
• RerankerAdvanced A reranker is a model that takes the candidate documents first narrowed down by search, compares each one carefully with the question, and re-sorts them in order of true relevance. It mainly serves as a 'second screening' in systems such as RAG that pass search results to an AI.
• RNNBasic An RNN (recurrent neural network) is an artificial neural network that reads ordered data such as sentences or speech one piece at a time, carrying what it has seen so far forward as 'memory' for the next decision.
• Scaling lawsBasic Scaling laws are empirical rules stating that as you increase an AI model's size (number of parameters), the amount of training data, and the compute used for training, model performance improves in a fairly predictable way.
• Semantic searchBasic Semantic search is a search method that finds results by measuring how close the meaning of a query is to the meaning of a document, rather than checking whether the words in the query appear in the document as written.
• SFTBasic SFT (supervised fine-tuning) is additional training that shows a pre-trained language model examples made of "instruction and model answer" pairs so that it learns to respond in the desired way. It is the most basic post-training step for making a chatbot answer questions in a conversational format.
• State space modelAdvanced A state space model (SSM) is an AI model architecture that reads long sequential data, such as text or audio, one piece at a time from the start, continually overwriting a fixed-size summary called the 'state' with what it has seen so far. It is drawing attention as a way to replace or complement the Transformer.
• Structured outputBasic Structured output is a feature that makes an AI model answer with data that fits a fixed format (a schema), such as JSON, instead of free-form text. It lets programs read and use the model's answer directly.
• Test-time computeAdvanced Test-time compute refers to using more computation at the stage when an AI model, after training, actually answers questions (inference time) in order to improve answer quality.
• TransformerBasic The Transformer is a neural network architecture built around attention, a mechanism that weighs how related the words (tokens) in a sentence are to each other, all at once. Most LLMs today are based on this architecture.
• Vector DBBasic A vector DB is a database that converts data such as text and images into lists of numbers (vectors, embeddings), stores them, and quickly finds data that is 'close in meaning' to a query.
• VLMBasic A VLM (Vision-Language Model) is an AI model that can look at images or video and describe their content in text or answer questions about them. It can be seen as a multimodal model that gives a language model 'eyes.'

Perception & action

AI that sees, decides and moves

• A2AAdvanced A2A (Agent2Agent) is a common communication standard (protocol) that lets AI agents built by different companies or with different tools find each other, hand off tasks and exchange results.
• Agent memoryBasic Agent memory is a mechanism that lets an AI agent store information needed for the current conversation or task, along with what it learned in past conversations and tasks, and retrieve it when needed. The model itself remembers nothing once a conversation ends, so memory is managed separately outside the model.
• Agentic RAGAdvanced Agentic RAG is a form of RAG in which an AI agent plans and adjusts the search process itself. The AI decides what to look up again and when, so instead of searching once and stopping, it searches several times and refines the results before answering.
• Agentic workflowBasic An agentic workflow is a way of working in which an AI, instead of answering a question once and stopping, takes a goal, breaks it into steps on its own, uses tools, and repeatedly checks and fixes its results until the job is done.
• CNNBasic A CNN (convolutional neural network) is an artificial neural network that finds patterns in grid-shaped data such as images by sliding a small window (a filter) across it step by step. It has been widely used as the foundation of technology that recognizes objects in photos.
• Computer useBasic Computer use is the ability of an AI to look at a screen and operate computer programs directly, like a person, by clicking a mouse or typing on a keyboard. A key feature is that it can handle programs that have no dedicated connection interface (API).
• DPOAdvanced DPO (Direct Preference Optimization) is a method that trains an LLM directly on preference data, in which people chose the better of two answers, without a separate reward model or reinforcement learning stage.
• Function CallingBasic Function Calling is a capability in which an AI model, when it needs to use a feature (function) of an external program to answer a question, produces a request in a fixed format saying "run this function with these values." The actual execution is handled not by the model but by the app using the model.
• HarnessBasic A harness is all the code, configuration and execution machinery outside an AI model that surrounds it. It helps the model become an AI agent that actually does work, and is often summarized as 'agent = model + harness'.
• Human-in-the-loopBasic Human-in-the-loop (HITL) is a design approach in which a person steps in partway through an AI's work to check, approve or correct it. Instead of leaving everything to the AI, a person makes the call at each critical point.
• JEPAAdvanced JEPA (Joint-Embedding Predictive Architecture) is a training approach in which, instead of redrawing the hidden parts of an image or video pixel by pixel, the model predicts a 'summarized meaning' (embedding) of those parts, so that AI learns on its own how to understand the world.
• JevBasic Jev is a decision-only AI model released by TypeSafe AI on September 15, 2026. Instead of writing sentences, it directly returns which of a set of predefined options is correct, along with probabilities.
• MCPBasic MCP (Model Context Protocol) is an open protocol that lets AI apps connect to external resources such as files, databases and work tools in the same way. A connector built once for MCP can be used as-is in many AI apps that support MCP.
• Multi-agentBasic Multi-agent is an approach in which several AI agents split up roles and exchange results with one another to handle one large task together. Instead of one agent doing everything alone, they work like a team.
• RLHFAdvanced RLHF (reinforcement learning from human feedback) is a method that collects people's judgments about which of an AI's several answers is better, and uses reinforcement learning to refine the model so it answers in the ways people prefer.
• RLVRAdvanced RLVR (Reinforcement Learning with Verifiable Rewards) is a method of training AI models with reinforcement learning, using as the reward criteria that can automatically check right and wrong, such as matching against the correct answer or running tests.
• Sim-to-RealAdvanced Sim-to-Real is a technique for training a robot or AI in a virtual environment on a computer (a simulation) and then transferring those skills to the real world.
• SkillsBasic Skills (Agent Skills) are 'work manual bundles' that package instructions, scripts and reference materials into a single folder to help an AI agent do a specific task well. The agent pulls out and reads a bundle only when the task needs it.
• SubagentBasic A subagent is a helper agent to which the main AI agent hands off part of its work. It works in its own context and returns only a summary of the results to the main agent.
• Tool useBasic Tool use is a capability that lets an AI model call outside tools such as search, calculators, databases and external APIs instead of producing every answer on its own. It is also called function calling.
• VLAAdvanced A VLA (Vision-Language-Action model) is an AI model that understands both the scene seen through a camera and instructions given in words, and directly produces the actions a robot should take. A single model handles seeing, understanding and moving.
• World Action ModelBasic A world action model (WAM) is an AI model that combines a world model, which predicts how the surroundings will change, with the ability to decide the next action. The term began to be used in robotics and autonomous driving research around 2026, and its definition has not yet settled.
• World modelAdvanced A world model is a model that lets AI learn how the world works, like a simulator in its head, so it can predict in advance 'if I do this, what happens next?'

Foundations & applications

Efficiency, evaluation and science applications

• Active LearningAdvanced Active learning is a training approach in which the model picks out data it judges to be "confusing" and asks a person to label it with the correct answer. The goal is to improve performance efficiently while reducing labeling costs.
• BenchmarkBasic A benchmark is a standard set of test questions and a scoring method designed to compare AI models' abilities under the same conditions. Each model solves the same problems, and their skill is compared by score.
• Curriculum learningAdvanced Curriculum learning is a training method in which an AI model is given training data not in arbitrary order but in a planned sequence, from easy to hard.
• Data contaminationAdvanced Data contamination is when the test questions used to evaluate an AI model (benchmarks) end up in the model's training data, inflating its score beyond its real ability.
• Data filteringAdvanced Data filtering is the work of picking out and discarding low-quality, harmful or useless items from collected data before training an AI model. What you leave out shapes model performance as much as what you put in.
• Data mixAdvanced A data mix is the combination that sets in what proportions different kinds of data, such as web documents, code, books, math and multiple languages, are blended when training an AI model. Even with the same amount of data, a model's abilities change depending on the mixing proportions.
• Dataset cardBasic A dataset card is a document that describes where a dataset came from, what it contains, and what it may and may not be used for. It is a 'user manual and ingredients label' attached to data.
• DistillationAdvanced Distillation (knowledge distillation) is a method of training a smaller, faster student model to reach similar performance by using the answers of a large, high-performing teacher model as examples.
• EvalsBasic Evals (evaluations) are test sets, and the process of using them, that repeatedly check whether an AI model or AI feature works as intended, using predefined problems and scoring criteria. Unlike public benchmarks, they are usually built in-house to fit your own service and tasks.
• InferenceBasic Inference is the process of feeding new input into an AI model that has already finished training to get actual results (answers, images, classifications and so on). In other words, it is the stage of 'using' the model.
• KNNBasic KNN (k-Nearest Neighbors) is a method that, when new data comes in, finds the k most similar items among the data it already knows and makes a decision based on those neighbors' answers. It is one of the simplest machine learning methods: it has no separate complex training process and relies on the idea that 'similar things have similar answers'.
• Knowledge graphBasic A knowledge graph stores knowledge like a web, with 'entities' such as people, places, events and concepts as points and the relationships between them as lines. This lets a computer read and reason about 'what is connected to what, and how'.
• LLM-as-a-JudgeAdvanced LLM-as-a-Judge is an evaluation method in which, instead of people, an LLM (large language model) scores another AI's answers or picks the better of two answers.
• Model collapseAdvanced Model collapse is a phenomenon in which, as each new AI is repeatedly trained on data made by AI, the model gradually forgets the diverse reality of the original world and its outputs become monotonous or strange.
• OntologyBasic An ontology is a 'concept map' that defines the concepts in a field and sets out, in a form computers can read, what properties those concepts have and how they connect to one another.
• OverfittingBasic Overfitting is when an AI model fits its training data too closely, so it scores well on the data it learned from but performs much worse on data it has not seen before.
• StrataBasic Strata is an AI runtime application that lets you run large LLMs directly (locally) on a gaming PC. According to reports in early October 2026, it can quickly run a quantized version of Qwen3.8-Flash-Next, a model with 125 billion (125B) parameters, on a single gaming GPU.
• Synthetic dataBasic Synthetic data is data generated artificially by an AI model or program, rather than created or collected by people in the real world. It is mainly used to train or evaluate AI models.
• Weak SupervisionAdvanced Weak supervision is a method in which, instead of people labeling each data point one by one, many 'weak' signals of uneven accuracy, such as rules, keywords and existing models, are used to produce correct answers (labels) in bulk, and a model is then trained on them.

Security & safety

AI security, safety and alignment, privacy and regulation

• Adversarial exampleBasic An adversarial example is an input that has been deliberately altered by a small amount so that it looks almost unchanged to a human but leads an AI model to a completely different judgment. Think of it as an 'optical illusion' built to fool the model.
• AI BOMAdvanced An AI BOM (AI Bill of Materials) is a complete list of the models, datasets, libraries, settings and other components in an AI system, including their versions and sources.
• AI risk managementBasic AI risk management is the organization-wide practice of identifying, assessing and reducing in advance the harms an AI system can cause (discrimination, malfunctions, privacy violations, security incidents and so on). The AI RMF published by the US NIST is a leading reference framework.
• AI supply chain securityBasic AI supply chain security is the practice of checking and protecting the external components that go into building an AI system (datasets, pretrained models, libraries, plugins and so on) to ensure they have not been contaminated or tampered with.
• AI watermarkBasic AI watermarking is a technology that embeds hard-to-notice marks in AI-generated text, images, audio and video so that it can later be checked whether something was 'made by AI.'
• AlignmentBasic Alignment (AI alignment) is the work of making AI behave in line with the goals and values people actually want. The core is getting it to do assigned tasks well while not behaving harmfully or deceptively.
• Backdoor attackAdvanced A backdoor attack secretly plants a 'trigger' in an AI model so that it behaves normally most of the time but acts as the attacker wants only when a specific signal appears.
• Confidential computingAdvanced Confidential computing is a technology that keeps data encrypted and protected inside a hardware-created isolated area (a TEE, trusted execution environment) even while it is 'in use' during computation, so that not even the cloud operator or server administrator can see its contents.
• Content moderationBasic Content moderation is the work of checking whether posts, images and videos uploaded to an online service, or answers generated by AI, violate the service's policies, and taking action such as hiding or deleting them when there is a problem.
• Content provenanceBasic Content provenance means recording, for digital content such as photos, videos and text, "who made it, with what tool, and how it was edited," and making it possible to verify with a digital signature that this record has not been changed. The main standard is C2PA, and records attached under this standard are commonly called 'Content Credentials'.
• Data exfiltrationBasic Data exfiltration is when an unauthorized person or program secretly moves data from inside an organization to the outside. In the AI era, a new risk has emerged: AI agents being tricked into sending data out on their own.
• Data poisoningAdvanced Data poisoning is an attack that deliberately mixes wrong or manipulated data into the training data an AI learns from, so that the model makes wrong decisions or behaves the way the attacker wants.
• Differential privacyAdvanced Differential privacy is a mathematical privacy protection method that adds a calculated amount of 'noise' to data so that no one can tell from the analysis results alone whether a specific individual was included in the data.
• Federated learningBasic Federated learning is a method in which data is not gathered in one place; instead, each device or institution trains a model a little on its own data, and only those 'training results (updates)' are collected to build a single shared model.
• GuardrailsBasic Guardrails are safety mechanisms that check and block content before and after an AI model, so that the model does not take in or produce content that is dangerous or breaks the rules.
• Indirect prompt injectionAdvanced Indirect prompt injection is an attack in which an attacker secretly hides instructions in external material such as web pages, emails or documents, so that an AI follows those instructions when it reads the material. It happens even if the user never types a malicious command.
• InterpretabilityAdvanced Interpretability is the research field that works out, in a way people can understand, why an AI model gave a particular answer and what computation actually happens inside the model.
• Machine unlearningAdvanced Machine unlearning is a technique that tries to remove the influence of specific data (personal information, copyrighted works, dangerous knowledge, etc.) from an AI model that has already been trained, so that the model behaves as if it had been trained without that data from the start.
• Membership inference attackAdvanced A membership inference attack is an attack that tries to find out whether certain data was used to train an AI model, just by looking at the model's responses. If the training data is sensitive, the mere fact that something "was included" can reveal personal information.
• Model extractionAdvanced Model extraction is an attack in which an attacker sends a large number of queries to an AI model, collects the answers, and builds a copy model that behaves similarly to the original.
• OWASP LLM Top 10Basic The OWASP LLM Top 10 is a list, compiled by the nonprofit security community OWASP, of 10 security risks that come up especially often and cause serious problems in apps and services that use LLMs.
• Personal data de-identificationBasic Personal data de-identification means removing or replacing information that can identify a specific person (PII, personally identifiable information), such as names, phone numbers and resident registration numbers, so that no one can tell whose data it is. It is a basic protection step data goes through before being used for AI training or analysis.
• Principle of least privilegeBasic The principle of least privilege is a security principle of giving people or programs only as much permission as they strictly need to do their job. For AI agents, it means opening up only the tools and data they need.
• Prompt injectionBasic Prompt injection is an attack that secretly inserts instructions into text fed to an AI model, making the model follow the attacker's instructions instead of its original ones.
• Red teamingBasic Red teaming is a testing activity in which people deliberately take an attacker's perspective and try to trick and misuse an AI model in order to find its weaknesses.
• Reward hackingAdvanced Reward hacking is when an AI focuses only on raising the score (reward) it receives during training, exploiting loopholes in how it is scored instead of achieving the originally intended goal.
• SandboxBasic A sandbox is a separate execution environment that keeps a program or AI agent moving only within a set boundary. Whatever it does inside, it cannot freely touch important files or systems outside.
• Security AIBasic Security AI is technology that uses AI to detect and respond faster to security threats such as hacking attempts, malware and unusual access. It sometimes also refers to the trend of attackers using AI to automate attacks.
• SycophancyBasic Sycophancy is when an AI model puts what it thinks the user wants to hear ahead of facts or sound judgment in its answers.
• System prompt leakageBasic System prompt leakage is when the instructions an AI service has placed in advance for the model to follow behind the scenes (the system prompt) are exposed through user questions or attacks. If the instructions contain passwords or internal rules, those leak out too.
• Tool poisoningAdvanced Tool poisoning is an attack that hides malicious instructions in the description or metadata of a tool used by an AI agent so that the agent follows them. It draws particular attention in environments such as MCP, where external tools are easy to connect.
• Training data extractionAdvanced Training data extraction is an attack that feeds crafted inputs to an AI model to make it output, verbatim, original text it memorized during training (personal information, code, documents, etc.).
• Zero trustBasic Zero trust is a security approach that drops the idea of 'trust it because it is inside the company' and instead checks who is asking and whether they have permission every time a person, device or program accesses a resource.

Data, ops & platforms

Data engineering, MLOps, platforms and solutions, classic ML

• Anomaly detectionBasic Anomaly detection is a machine learning technique that learns the "normal pattern" shown by most of the data and automatically finds rare cases that deviate sharply from it.
• BM25Advanced BM25 is a keyword search ranking function that scores each document based on how often, and how rarely, the words in a search query appear in it. It has long been a standard method for search engines and is still used as a core component of RAG.
• CDCAdvanced CDC (Change Data Capture) is a technique that picks out only the data that has been newly added, modified or deleted in a database and sends it straight to other systems. Instead of copying the entire dataset every time, it lets you follow only 'what changed.'
• Class imbalanceBasic Class imbalance is a situation in a classification problem where some answer categories (classes) have a great deal of data and others have very little. A model trained on such data tends to get the common class right and miss the rare but important one.
• ClusteringBasic Clustering is the process of automatically grouping unlabeled data by similarity, dividing it into several groups (clusters). It is a representative method of unsupervised learning.
• Columnar storageAdvanced Columnar storage is a way of storing table data grouped by column rather than by row. It makes analytical work that reads only a few columns to compute sums or averages much faster and cheaper.
• Data augmentationBasic Data augmentation means creating more training examples by slightly changing or transforming training data you already have. It lets a model learn from a wider variety of cases without collecting new data.
• Data catalogBasic A data catalog is an organized list and search service that lets people look up, in one place, what data exists in a company, where it is, what it means, who manages it and whether it can be trusted.
• Data contractAdvanced A data contract is an agreement, set down in documents and code, between the side that produces data and the side that uses it, stating 'this data will be provided in this format, with this meaning and at this quality'. Changes that break the agreement are caught in advance.
• Data driftAdvanced Data drift is when the characteristics of the data a model actually receives gradually diverge from the data it was trained on. As a result, a model that was accurate at first quietly becomes less accurate over time.
• Data flywheelBasic A data flywheel is a virtuous cycle in which more people using a service means more data accumulates, that data makes the AI better, and that in turn brings in more users.
• Data governanceBasic Data governance is a framework that defines, and enforces, who manages and uses an organization's data, under what rules and how. It covers data owners, access rights, quality standards, retention and deletion rules, and more.
• Data lakeBasic A data lake is a large-capacity repository that cheaply stores all kinds of data in one place in its original form, from tabular data to logs, images and documents. Instead of defining a format in advance, data is stored first and read in whatever way is needed at the time of use.
• Data lakehouseAdvanced A data lakehouse is a data storage and analytics architecture that combines in one system the strengths of a 'data lake,' which cheaply stores anything in its original form, and a 'data warehouse,' which analyzes organized data quickly and reliably.
• Data leakageBasic Data leakage is the mistake of letting information that would not be available at real prediction time, especially test data or clues to the answer, slip into the training process. When this happens, evaluation scores come out high, but performance drops sharply in the real service.
• Data lineageBasic Data lineage is recording and showing where data came from, what processing it went through and where it flowed to. It lets you trace a single number back to its source and the path it took.
• Data meshAdvanced Data mesh is an organizational and operating approach in which, instead of a single central data team managing all company data, each business unit (domain) that produces data takes responsibility for its own data and offers it like a 'product'.
• Data observabilityAdvanced Data observability is continuously watching whether data is arriving on time and in the right volume and shape, alerting immediately when something goes wrong, and making it possible to trace the cause.
• Data pipelineBasic A data pipeline is a processing flow that automatically chains together collecting data scattered across many places, refining it for use and moving it to the storage where it is needed. Depending on the order of steps, it is divided into ETL (extract→transform→load) and ELT (extract→load→transform).
• Data qualityBasic Data quality is the degree to which data is accurate, complete, consistent and up to date enough to be fit for its purpose. Analyses or AI models built on low-quality data produce results that are hard to trust.
• Data versioningBasic Data versioning means recording the state of a dataset every time it changes, so you can retrieve the data as it was at any point in time and compare what changed.
• Data warehouseBasic A data warehouse (DW) is a store that gathers data from across a company in one place, organized in a form that is good for analysis. Unlike databases that handle day-to-day operations, it is designed for aggregating and querying large amounts of data at once.
• Decision treeBasic A decision tree is a prediction model that splits data by asking a series of yes/no questions and gives an answer at the final branch. Combining many decision trees and deciding the answer by vote is called a random forest.
• DeduplicationAdvanced Deduplication means keeping only one copy and filtering out the rest when identical or nearly identical documents, sentences or images appear multiple times in training data. It is an essential step, especially when cleaning LLM training data collected in bulk from the internet.
• Embedding pipelineAdvanced An embedding pipeline is an automated workflow that chains together collecting documents, splitting them into small pieces, converting them into embeddings (lists of numbers that capture meaning) and loading them into a vector DB. This preparation has to be in place before RAG or semantic search can work properly.
• Entity ResolutionAdvanced Entity resolution (finding and merging records of the same entity) is the task of determining whether records in different datasets, with slightly different names or spellings, actually refer to the same person, company or product, and combining them into one.
• ETL·ELTBasic ETL and ELT are ways of pulling data scattered across many places (Extract), reshaping it into a usable form (Transform) and moving it into one place (Load). If the transformation happens before loading, it is called ETL; if it happens after loading, it is called ELT.
• Experiment trackingBasic Experiment tracking is the practice of automatically recording which data, code and settings were used each time an AI model is trained, and what the results were, so runs can later be compared and reproduced exactly.
• Feature engineeringBasic Feature engineering is the work of selecting, processing and creating input values (features) from raw data so that a machine learning model can find patterns more easily.
• Feature storeAdvanced A feature store is a central repository where the input values (features) used by machine learning models are created and stored. It lets the same values be retrieved in exactly the same way both during training and when making predictions in a live service.
• Gradient boostingBasic Gradient boosting is a technique that builds a strong prediction model by adding small decision trees one after another, training each new tree to make up for what the previous trees got wrong. XGBoost, LightGBM and CatBoost are well-known implementations.
• LLM observabilityBasic LLM observability means recording and being able to inspect what an LLM-based service actually received as input, which steps it went through, and what it answered. It is used to find the cause when something goes wrong and to manage quality and cost.
• LLMOpsBasic LLMOps refers to the management practices and tools needed to build and operate services that use LLMs (large language models). It builds on MLOps and adds work that is new with LLMs, such as prompt management, answer quality evaluation and cost management.
• MLOpsBasic MLOps is a way of automating and systematically managing the whole process of building machine learning models, deploying them to services, and continuing to improve them while they run. The name means linking model development (ML) and operations (Ops) into one.
• Model registryAdvanced A model registry is a central repository that stores trained AI models by version and records and manages what data and settings each model was built with and what stage it is currently in (under validation, in production, or archived).
• OLAPBasic OLAP (online analytical processing) is a way of processing data that groups and aggregates stored data by multiple criteria such as period, region and product so it can be analyzed quickly. Rather than recording individual transactions, it is designed to answer aggregate questions such as 'What were sales by region last quarter?'
• Semantic layerAdvanced A semantic layer is an intermediate layer that defines in one place what business terms such as 'revenue' or 'active users' actually mean in the data and how they are calculated. Because dashboards, spreadsheets and AI all retrieve data through these definitions, the same question gets the same number.
• Stream processingBasic Stream processing is a way of handling data that continuously receives data the moment it is generated, computes on it right away and produces results. It contrasts with batch processing, which collects data and processes it all at once later.
• Text-to-SQLBasic Text-to-SQL is a technology that converts a question asked in everyday language into SQL, the query language databases understand. It lets you ask questions of data directly without knowing SQL.
• Time series forecastingBasic Time series forecasting analyzes past patterns in data collected in time order (daily sales, hourly electricity use, etc.) to anticipate future values.

Infrastructure & using models

AI infrastructure such as GPUs, serving and cost, plus terms you meet when choosing and reading models on Hugging Face

• Active parametersBasic Active parameters are the number of parameters a model actually uses in computation when processing a single token. MoE models select and use only a portion of their total parameters, so the total and active parameter counts differ greatly.
• AI data centerBeginner An AI data center is a facility that houses large numbers of accelerators such as GPUs for AI training and inference, along with the power, cooling and networking they require.
• Base and Instruct modelsBeginner A Base model has only completed pre-training and continues whatever text it is given, while an Instruct model has additional post-training that refines it to follow instructions and hold conversations.
• Chat templateBasic A chat template is a set of rules that converts a conversation made of system, user and assistant messages into exactly the string format (including special tokens) the model saw during training.
• CheckpointBasic A checkpoint is a file that saves a model's weights (and training state) at a specific point during or after training. It is used to resume training, or to load and compare the model as it was at that point.
• Continuous batchingAdvanced Continuous batching is a serving technique that groups multiple user requests to process them together on a GPU, removing finished requests right away and slotting in waiting requests right away so the GPU never sits idle.
• CUDABasic CUDA is a GPU programming platform made by NVIDIA. It is the software foundation that lets the GPU's many processing units be used for general computation such as AI training and inference.
• Distributed trainingAdvanced Distributed training is a way of splitting the work of training a single model across multiple GPUs or multiple servers so they process it together. It is used when the model or training data does not fit on a single GPU, or when one GPU would take too long.
• FlashAttentionAdvanced FlashAttention is a computation method that restructures the Transformer's attention calculation to fit the GPU's memory layout, making it faster and using less memory while leaving the results unchanged.
• FLOPSBasic FLOPS is the number of floating-point operations (operations on numbers with decimal points) a computer can perform per second, and it is the standard unit for describing the computing speed of AI chips.
• GGUFBasic GGUF is a model file format used to run LLMs in llama.cpp-family tools. It packs information such as the tokenizer and chat template into a single file along with the weights, and mainly stores quantized weights.
• GPTQ·AWQAdvanced GPTQ and AWQ are widely used quantization methods that shrink the weights of an already trained LLM to fewer bits, such as 4 bits. They make a model lighter without retraining, using only a small amount of sample data to limit quality loss.
• GPUBeginner A GPU (graphics processing unit) is a chip originally built to draw images on screen, and it excels at running thousands of simple calculations at the same time. This property has made it the core hardware for training and running today's AI models.
• GPU cloudBasic A GPU cloud is a service that lets you rent GPU servers from a cloud provider, for example by the hour, for AI training and inference instead of buying them yourself.
• HBMBasic HBM (High Bandwidth Memory) is memory built by stacking several memory chips on top of each other like an apartment building and placing them right next to a compute chip such as a GPU, so that very large amounts of data can be exchanged at once.
• Hugging FaceBeginner Hugging Face is an online repository for uploading and downloading AI models, datasets and demos, and also a company that builds open-source libraries that make them easy to use. It is often called 'the GitHub of AI models.'
• Inference engineBasic An inference engine is the runtime software that loads a trained AI model onto GPUs and lets it answer user requests quickly and efficiently. Widely used options include vLLM, SGLang and TensorRT-LLM for servers, and llama.cpp and Ollama for personal PCs.
• InterconnectAdvanced An interconnect is the high-speed communication network that links multiple GPUs and servers so they can exchange data. Typical examples are NVLink within a server and InfiniBand or high-speed Ethernet between servers.
• KubernetesBasic Kubernetes is an open-source management system that automatically spreads containers (a program and its runtime environment packed into one box) across multiple servers, restarts them when they stop, and scales them up or down to match usage. It is often abbreviated as K8s.
• KV cacheAdvanced A KV cache is a mechanism that stores in memory the intermediate results (Keys and Values) an LLM has already computed, so it can reuse them as it generates text one token at a time.
• LeaderboardBeginner A leaderboard evaluates multiple AI models using the same benchmark or user votes and shows the results as a ranking table. It serves as a map for narrowing down initial candidates when choosing a model.
• LLM gatewayAdvanced An LLM gateway is an intermediate layer that places a single entry point between multiple apps and the multiple AI models they use, routing every request through it. This entry point handles authentication, model selection (routing), cost management, logging and security checks in one place.
• Local LLMBeginner A local LLM is a language model you download and run directly on your own PC, laptop or company server instead of using a cloud API. Tools such as llama.cpp and Ollama make this process easier.
• Model cardBeginner A model card is a document that describes what an AI model was built for, what data it was trained on, how to use it and what it does poorly. On Hugging Face, the README on the first screen of a model page is the model card.
• Model licenseBeginner A model license is the set of usage rules that defines who may use, modify and redistribute an AI model's weights, for what purposes and under what conditions. Because it determines in particular whether commercial use is allowed, it should be the first thing you check when choosing a model.
• Model mergingAdvanced Model merging is a technique that mixes the weights of several fine-tuned models branched from the same base model to create a new model that combines the strengths of each, without additional training.
• Model parallelismAdvanced Model parallelism is a way to split a large model that does not fit on a single GPU across multiple GPUs and have them compute together. The main approaches are tensor parallelism (TP), which splits the computation within a layer, and pipeline parallelism (PP), which splits the layers in sequence.
• Model servingBasic Model serving is the work of putting a trained AI model on servers and running it so that it receives requests from apps or users, runs the model and returns the results.
• NPUBeginner An NPU (Neural Processing Unit) is a semiconductor chip designed specifically to run artificial neural network computations quickly while using little power. Think of it as a 'dedicated calculator' specialized for running AI models.
• On-device AIBeginner On-device AI is a way of running AI models directly inside devices such as smartphones, laptops and cars, rather than on servers across the internet. The device produces the answer without sending data outside.
• Open weightsBeginner Open weights is the practice of releasing the internal numbers (weights) an AI model learned during training so that anyone can download and use them. You can run the model directly on your own computer or server.
• PagedAttentionAdvanced PagedAttention is a memory management technique that splits the KV cache, which builds up while an LLM generates an answer, into small blocks (pages) and places only as many as needed into free spots in GPU memory. The inference engine vLLM made this technique widely known.
• Precision (FP32, BF16, FP8)Basic Precision refers to how many bits an AI model uses to store and compute a single number. Formats such as FP32, BF16 and FP8 use fewer bits as you go down, which reduces memory and computation but makes the representation coarser.
• Prompt cachingBasic Prompt caching is a technique that stores the computation results for the beginning of a prompt that is sent to an LLM identically every time, and reuses them in the next request. Because the same content is not recomputed from scratch, responses are faster and costs go down.
• QuantizationAdvanced Quantization is a technique that makes an AI model smaller and faster by representing the numbers inside it (parameters, etc.) with fewer bits. It slightly lowers precision in exchange for large reductions in memory and compute cost.
• RoPEAdvanced RoPE (Rotary Position Embedding) is a method that rotates vectors by a set angle according to position, so that a transformer can tell where each token sits in a sentence. Many LLMs today use it as their default way of handling position information.
• safetensorsBasic safetensors is a file format designed to store and load AI model weights (the numbers set through training) safely and quickly. Its main feature is that simply opening a file carries no risk of running hidden code.
• Speculative decodingAdvanced Speculative decoding is a technique that speeds up LLM answer generation by having a small model quickly guess several upcoming tokens in advance, and having a large model check those guesses all at once and accept only the parts that are correct.
• Throughput (tokens/s)Basic Throughput (tokens/second) is the number of tokens an inference server produces per second, a metric that shows how much work the same hardware can handle. You need to distinguish between the total amount the whole server processes and the speed a single user experiences.
• Token-based pricingBeginner Token-based pricing is a way of charging for AI models used through an API by counting the amount of text sent and received in tokens. Usually, input tokens fed into the model and output tokens the model writes have separate unit prices.
• TokenizerBasic A tokenizer is a converter that turns text into token pieces and numbers a model can read, and turns the numbers the model outputs back into text. Each model has its own fixed tokenizer.
• TPUBasic A TPU (Tensor Processing Unit) is an AI-specific chip that Google designed in-house to speed up deep learning computation, especially matrix multiplication.
• TTFTBasic TTFT (Time To First Token) is the time from when a user sends a request until the AI outputs the first token (a piece of text) of its answer. It is a key metric for making a chatbot feel 'responsive.'
• VRAMBeginner VRAM is the dedicated memory attached to a GPU; model weights and intermediate computation values must be loaded here for the GPU to compute quickly.

Industry AI

Terms you meet when AI is used in healthcare, finance, manufacturing, defense and other industries

• AI drug discoveryBasic AI drug discovery uses machine learning and generative AI to find targets that cause disease, design and screen substances that could become drugs, and predict their effects and side effects in advance, making drug development faster and more efficient.
• AI tutorBeginner An AI tutor is an AI private tutor that answers each student's questions and adapts explanations and problems to their level. It is usually built on an LLM and supports learning through conversation.
• AI-RANAdvanced AI-RAN is a telecom network design approach that deeply embeds AI into the radio access network (RAN), which connects mobile phones to base stations, to make network operations smarter with AI and to run AI services on the same base station computing resources.
• Algorithmic tradingBasic Algorithmic trading is a way of trading in which a computer automatically places orders, leaving decisions about when, what, at what price and how much to buy or sell to predefined rules or models.
• Alternative dataBasic Alternative data is information obtained outside traditional financial data such as financial statements, stock prices and credit records. It means using sources such as card payment flows, satellite images, web traffic and phone bill payment history for investment decisions or credit scoring.
• AMLBasic AML (Anti-Money Laundering) refers to the rules and technologies through which financial institutions verify customers, find suspicious transactions and report them to authorities, in order to prevent money earned from crime from being laundered to look like legitimate money.
• Autonomous weapon systemsBasic Lethal autonomous weapon systems (LAWS) are weapon systems that, once activated, can select targets and carry out attacks without a human approving each action. Whether and how to regulate them remains an ongoing debate in the international community.
• C4ISRAdvanced C4ISR is a combined term for Command, Control, Communications, Computers, Intelligence, Surveillance and Reconnaissance. It refers to the entire information and decision-making system a military uses to understand a situation, make decisions and pass them on to units.
• CDSSBasic A CDSS (Clinical Decision Support System) is software that shows alerts or recommendations based on patient information and medical knowledge when doctors and nurses make diagnostic or prescribing decisions.
• Credit scoring modelBasic A credit scoring model calculates from data how likely a person or business applying for a loan or card is to repay on time, and expresses it as a score or grade. Financial companies use this score to decide approval, credit limits and interest rates.
• Demand forecastingBasic Demand forecasting is estimating in advance which products will sell, when, where and in what quantities, based on past sales records and information such as weather, events and prices.
• Digital pathologyAdvanced Digital pathology is a way of doing pathology diagnosis in which tissue slides are scanned at high resolution into digital images that are then read, shared and analyzed on screen. It is also the technology that AI analysis builds on.
• Digital twinBasic A digital twin is a virtual model inside a computer that replicates a real object such as a factory, equipment, a city or a power grid. It continuously receives data from the real thing to track its state and is used to test things in advance.
• Drone swarmBasic A drone swarm is a way of operating multiple drones that exchange information with one another, move together like a single group and divide up tasks. Instead of a person flying each drone individually, the whole group is given a goal and the drones coordinate with each other as they move.
• Dual-use technologyBasic Dual-use technology is technology that can be used for peaceful civilian purposes as well as for military or harmful ones. Because most AI is general-purpose, this trait is especially strong in AI.
• EHR·FHIRBasic An EHR (electronic health record) is a system that stores and manages patients' medical records digitally, and FHIR (pronounced "fire") is an international standard that lets different EHRs and health apps exchange data in the same format.
• Fraud detectionBasic Fraud detection means automatically picking out suspicious, out-of-the-ordinary transactions such as card theft or scams among financial transactions like card payments and transfers, then blocking or verifying them. In the financial sector it is often called an FDS (Fraud Detection System).
• Grid AIBasic Power grid AI is AI that predicts how much electricity will be used and how much solar and wind power will be generated, and detects equipment failures in advance to help run the power grid reliably.
• Legal AIBasic Legal AI is technology that uses AI to help find, read, summarize and draft legal documents such as case law, statutes and contracts. Rather than replacing lawyers or legal teams, it is mainly used as a tool to reduce repetitive document work.
• Meaningful human controlBasic Meaningful Human Control is the principle that when AI or autonomous systems are involved in decisions to use force, people must keep a level of control at which they actually understand, judge and can take responsibility, rather than just pressing a formal approval button.
• Medical imaging AIBasic Medical imaging AI is technology in which AI analyzes medical images such as X-rays, CT, MRI and ultrasound to find lesions (areas suspected of disease) or assist in reading the images. The final diagnosis is usually made by a doctor, and the AI acts as a "second pair of eyes."
• Medical LLMBasic A medical LLM is a large language model adapted to medical data such as medical knowledge and clinical records, to help with answering questions, writing documents and summarizing information in clinical settings.
• Model risk managementAdvanced Model risk management is a framework that manages the entire life of a model, from development, validation and approval to monitoring and retirement, to prevent losses that occur when the credit scoring, pricing and AI models used by financial companies are wrong or misused.
• Predictive maintenanceBasic Predictive maintenance uses AI to analyze data from sensors attached to equipment, detect warning signs before a failure occurs, and carry out maintenance exactly when it is needed.
• Protein structure predictionAdvanced Protein structure prediction means working out computationally, from only the sequence of amino acids that make up a protein, what three-dimensional shape the protein actually folds into.
• Robo-advisorBeginner A robo-advisor is an online investment advisory and management service that uses a questionnaire to identify an investor's goals and risk tolerance, after which an algorithm automatically handles asset allocation and rebalancing (readjusting weights).
• Route optimizationBasic Route optimization is the task of finding the trip order and routes that minimize total travel distance, time or cost, while taking into account multiple delivery locations, vehicles and time constraints together.
• SaMDBasic SaMD (Software as a Medical Device) is a program that is treated as a medical device because it serves medical purposes such as diagnosing, treating, preventing or managing disease through software alone, without being part of a separate piece of medical equipment.
• Satellite imagery analysisBasic Satellite imagery analysis (GEOINT, geospatial intelligence) means analyzing imagery taken by satellites, aircraft and drones together with map data to find out what is happening on the ground. These days AI first finds objects and changes in vast amounts of imagery, and people then verify them.
• Smart factoryBeginner A smart factory is a factory that connects its equipment, sensors and production management systems through data, monitors production in real time, and runs more efficiently using AI and automation.
• Vision inspectionBasic Vision inspection is a quality inspection method in which AI looks at camera images of products and automatically finds defects such as scratches, foreign matter and assembly faults.