AI terms, in plain words.
Find a clear definition. Stay for the explanation.
890 terms
| Term | In one line | Read |
|---|---|---|
| 3D Gaussian splatting | 3D Gaussian splatting rebuilds a real scene from photos as a large set of small, fuzzy 3D blobs, so you can view it from new angles in real time. | 2 min quick read |
| 3D generation models | 3D generation models create 3D objects, such as meshes or point clouds, from a text prompt or an image. | 2 min quick read |
| A/B testing LLMs | A/B testing an LLM app means showing two versions, such as two prompts or models, to different real users and comparing which one performs better. | 2 min quick read |
| A2A protocol | A2A is an open protocol for AI agents: one agent can find another, hand it a task and track the result, even when different companies built them. | 4 min explainer |
| Accuracy | Accuracy is the share of a model's predictions that were right, the correct answers divided by all answers. | 4 min explainer |
| Activation function | An activation function is the non-linear step a neuron applies to its weighted sum, so a network of layers can learn curved patterns, not only straight lines. | 4 min explainer |
| Activepieces | Activepieces is an open-source automation tool for building workflows and AI agents that connect apps, and its integrations double as MCP tools for AI assistants. | 1 min quick read |
| Adam and AdamW | Adam is a popular training method that adjusts each weight's step size automatically; AdamW is a fixed-up version that handles weight decay properly. | 2 min quick read |
| Adobe Firefly | Adobe Firefly is Adobe's family of generative AI models, and the app built on them, for making images, video, audio and vector graphics from text, aimed at commercial use. | 2 min quick read |
| Adversarial examples | An adversarial example is an input changed very slightly, often invisibly to people, so that an AI model gets it confidently wrong. | 2 min quick read |
| Agent evaluation | Agent evaluation tests an AI agent on set tasks, checking both its final result and the steps it took, over several tries. | 4 min explainer |
| Agent Lightning | Agent Lightning is Microsoft Research's open-source framework for training AI agents with reinforcement learning while they keep running in their usual code. | 2 min quick read |
| Agent memory | Agent memory is how an AI agent saves useful information outside the model and loads the right pieces back into its prompt later. | 5 min explainer |
| Agent orchestration | Agent orchestration is how an app with several AI agents decides which agent works, in what order, and who picks what happens next. | 4 min explainer |
| Agent planning | Agent planning is how an AI agent splits a big task into smaller steps first, then carries out those steps one by one. | 4 min explainer |
| Agent skills | Agent skills are reusable packages of instructions, knowledge, or tool procedures that help an AI agent perform a defined class of tasks consistently. | Definition |
| Agent state | Agent state is the running record an AI agent keeps while it works, so a task can pause, resume, or recover from a crash. | 5 min explainer |
| Agent Substrate | Agent Substrate is open-source software that runs many AI agents on Kubernetes by saving idle agents as snapshots in storage and resuming them when a request arrives. | 5 min explainer |
| agent-browser | agent-browser is an open-source command-line tool from Vercel Labs that lets an AI agent control a web browser, one short command at a time. | 2 min quick read |
| Agentic models | An agent plans, acts, observes the result, adjusts its approach, and repeats until the task is complete. | 3 min explainer |
| Agentic RAG | Agentic RAG puts an AI agent in charge of retrieval, so the model decides when to search, what to look up and whether to search again before answering. | 2 min quick read |
| Agentic workflows | An agentic workflow runs a task as several language model and tool steps, joined by a path that your code fixes in advance. | 4 min explainer |
| AGENTS.md | AGENTS.md is a plain Markdown file in a code project that tells AI coding agents how to work on it: setup commands, tests and code style. | 1 min quick read |
| AGI | AGI is a proposed category for AI at least as capable as a person across most thinking tasks. Definitions differ; a 2023 Google DeepMind framework marked the level matching many definitions as not yet achieved. | 4 min explainer |
| Agno | Agno is open-source software for building AI agents and running them as a live service. You can build single agents, teams of agents and step-by-step workflows. | 4 min explainer |
| AI accelerators | AI accelerators are chips built to speed up the maths inside AI models, which is largely multiply-add sums. | 5 min explainer |
| AI agent with tools | An AI agent with tools is a program where a language model decides when to call functions you provide, such as search or a calculator, and uses the results to finish a task. | 2 min quick read |
| AI agents | An AI agent is a language model that chooses its own next steps and tools while it works through a task. | 4 min explainer |
| AI art | AI art is artwork made with the help of AI models that generate or transform images, sound or video, usually guided by an artist's choices and data. | 2 min quick read |
| AI engineering | AI engineering is the work of building useful applications on top of ready-made foundation models, rather than training models from scratch. | 2 min quick read |
| AI ethics | AI ethics is about the values AI should respect, such as human rights, fairness, privacy and accountability, and how to put them into practice. | 2 min quick read |
| AI governance | AI governance is the set of rules, processes and oversight that organisations and governments use to manage the risks of AI while still getting its benefits. | 2 min quick read |
| AI in education | AI in education covers tools that tutor students, help teachers plan and give feedback, and the rules schools need to use them safely. | 2 min quick read |
| AI in finance | AI in finance is the use of machine learning and language models by financial firms, central banks and supervisors, for work such as data analysis, forecasting, payments and compliance. | 2 min quick read |
| AI in healthcare | AI in healthcare uses machine learning to help diagnose illness, read medical images, speed up research and support clinicians, under safety rules for medical devices. | 2 min quick read |
| AI incidents | An AI incident is a real event where an AI system caused or nearly caused harm, and incident databases collect them so others can learn. | 2 min quick read |
| AI model | An AI model is the part of an AI system that has learned from data. It is a structure plus a set of learned numbers that turns an input into a prediction or new content. | 4 min explainer |
| AI safety | AI safety is the field working to stop AI systems from causing accidents or harm, from everyday mistakes today to risks from much more powerful future systems. | 2 min quick read |
| AI search | AI search answers your question directly in written form, drawing on many web pages and linking to them, instead of only returning a list of links. | 2 min quick read |
| AI weather forecasting | AI weather forecasting uses models trained on decades of past weather to predict the weather days ahead, often faster and sometimes more accurately than traditional simulations. | 2 min quick read |
| AI winter | An AI winter is a period when disappointment with artificial intelligence leads to reduced funding, investment, public interest, and research activity. | Definition |
| ai-memory | ai-memory is an open-source tool that gives AI coding agents a shared long-term memory of a project, so a new session or a different agent can carry on where the last one stopped. | 2 min quick read |
| Aider | Aider is an open-source AI pair programmer that runs in your terminal and edits the code in your project using the language model of your choice. | 2 min quick read |
| Alexa+ | Alexa+ is Amazon's rebuilt Alexa voice assistant, powered by generative AI, that holds more natural conversations and carries out tasks for you. | 1 min quick read |
| AlexNet | AlexNet is the deep convolutional neural network that won the 2012 ImageNet image recognition challenge by a wide margin, showing that deep networks trained on raw pixels could beat older methods. | 2 min quick read |
| Algorithm | An algorithm is a precise, step-by-step procedure for getting a result. In AI, learning algorithms are the procedures that turn data into a trained model. | 4 min explainer |
| Alignment | Alignment is the work of making AI systems do what people actually intend, and behave in line with human values, rather than just what they were literally told. | 2 min quick read |
| Allen Institute for AI | The Allen Institute for AI (Ai2) is a non-profit research institute in Seattle, founded by Paul Allen in 2014, that makes fully open AI models such as Olmo. | 2 min quick read |
| AlphaCode | AlphaCode is a DeepMind system from 2022 that writes computer programs to solve competitive programming problems. | 2 min quick read |
| AlphaFold | AlphaFold is a Google DeepMind AI system that predicts the 3D shape of a protein from its sequence of amino acids. | 2 min quick read |
| AlphaGenome | AlphaGenome is a Google DeepMind AI model that predicts how a stretch of DNA, and small changes to it, affect what happens inside human cells. | 1 min quick read |
| AlphaGo | AlphaGo is a DeepMind program that became the first to beat a professional human, and then a world champion, at the board game Go. | 2 min quick read |
| AlphaZero | AlphaZero is a DeepMind algorithm that taught itself chess, shogi and Go from scratch, given only the rules, and beat top programs in all three. | 2 min quick read |
| Amazon Bedrock | Amazon Bedrock is an AWS service that lets developers call AI models from several companies through one set of APIs, with extra tools for search and safety. | 5 min explainer |
| Amazon Bedrock AgentCore | A set of AWS services, called Amazon Bedrock AgentCore, for running AI agents in the cloud, with hosting, memory, sign-in, tool access and monitoring. | 5 min explainer |
| Amazon Nova | Amazon Nova is Amazon's family of foundation models and related services, used by developers through Amazon Bedrock on AWS. | 2 min quick read |
| Amazon Polly | Amazon Polly is an AWS cloud service that turns written text into natural-sounding speech for apps such as contact centres, AI voice agents and e-learning. | 1 min quick read |
| Amazon Q Developer | Amazon Q Developer is an AI assistant from AWS that helps people write code and understand, build and run applications on AWS. | 2 min quick read |
| Amazon SageMaker AI | SageMaker AI, from AWS, lets you build, train and run machine learning models without managing your own servers. | 4 min explainer |
| AMD Instinct | AMD Instinct is AMD's line of data centre GPUs for AI training, inference and supercomputing, programmed with AMD's open ROCm software. | 2 min quick read |
| Anomaly detection | Anomaly detection finds the few items or events that do not fit the usual pattern in data, such as a fraudulent card payment or an odd burst of network traffic. | 4 min explainer |
| Anthropic API | The Anthropic API, also called the Claude API, lets developers call Claude models from their own code over the web. | 2 min quick read |
| AnythingLLM | AnythingLLM is an open-source, all-in-one AI app for chatting with your own documents, building agents and connecting any local or cloud language model. | 2 min quick read |
| Apple Foundation Models | Apple Foundation Models are the language models Apple built to power Apple Intelligence: a small model that runs on the device and a larger one on Apple's servers. | 2 min quick read |
| Apple Intelligence | Apple Intelligence is the set of AI features built into Apple's iPhone, iPad, Mac, Watch and Vision Pro software, running on the device where it can and on Apple's private servers when needed. | 2 min quick read |
| Apple M-series chips | Apple M-series chips are the processors Apple designs for Mac and iPad, combining CPU, GPU, Neural Engine and shared memory on one chip. | 2 min quick read |
| Apple Neural Engine | The Apple Neural Engine is a part of Apple's chips built to run machine learning models quickly and with little power, on iPhone, iPad and Mac. | 2 min quick read |
| Approximate nearest neighbor search | Approximate nearest neighbor search finds stored vectors that are very close to a query quickly, by checking only a promising part of the collection and accepting a few misses. | 4 min explainer |
| ARC-AGI | ARC-AGI presents demonstration grid pairs to a system, then asks it to construct outputs for new test grids. | 3 min explainer |
| Arize Phoenix | Arize Phoenix is an open-source tool for tracing and evaluating AI apps, built by Arize AI, that you can run on your own machine or servers. | 2 min quick read |
| Artificial intelligence | An AI system infers from received inputs how to generate predictions, content, recommendations or decisions for explicit or implicit objectives. | 5 min explainer |
| AssemblyAI | AssemblyAI is a company whose developer APIs transcribe recorded and live speech and then analyse what was said. | 2 min quick read |
| Atria Dawn | Atria Dawn is an open agentic model from the Shanghai Artificial Intelligence Laboratory, built for research and engineering jobs that take many steps and tools. | 1 min quick read |
| Attention | Attention lets a model build an output by assigning weights to available pieces of information and mixing their values. | 3 min explainer |
| Autoencoders | An autoencoder learns an encoder that maps an input to a code and a decoder that reconstructs the input from that code. | 3 min explainer |
| AutoGen | AutoGen is an open-source Microsoft framework for building apps where several AI agents talk to each other, and sometimes to people, to finish a task. | 4 min explainer |
| AutoGPT | AutoGPT began as an early experiment in letting GPT-4 act on its own toward a goal, and is now a platform where a coordinator agent hands jobs to AI specialists. | 2 min quick read |
| AutoML | AutoML automates the steps of building a machine learning model, such as choosing the model and its settings, so it takes less expert time. | 2 min quick read |
| Autonomous vehicles | Autonomous vehicles are cars that use sensors and AI to drive themselves; robotaxis such as Waymo's now carry riders with nobody at the wheel. | 2 min quick read |
| Autoregressive models | An autoregressive model factorizes a sequence probability into conditional probabilities. | 3 min explainer |
| AWS Inferentia | AWS Inferentia is a family of computer chips that Amazon designed to run trained AI models cheaply and quickly in its cloud. | 4 min explainer |
| AWS Trainium | AWS Trainium is a computer chip Amazon designed for training and running AI models, rented through special Amazon EC2 cloud servers. | 4 min explainer |
| Axolotl | Axolotl is a free, open-source tool for fine-tuning and post-training open language models, driven by a single configuration file. | 2 min quick read |
| Azure Machine Learning | Azure Machine Learning is Microsoft's cloud service for building, training, deploying and managing machine learning models across their whole life cycle. | 2 min quick read |
| Azure OpenAI | Azure OpenAI lets companies use OpenAI's models through Microsoft's Azure cloud, with Azure billing, safety filters and data controls. | 5 min explainer |
| Backpropagation | Backpropagation computes each parameter's gradient by applying the chain rule from the last layer back to the first. | 4 min explainer |
| Base models | A base language model is a pretrained model before further instruction or conversational fine-tuning. | 3 min explainer |
| Batch inference | Batch inference sends many AI requests as one job that runs in the background, trading an instant reply for a lower price. | 3 min explainer |
| Batch normalization | Batch normalization rescales the values flowing between layers of a neural network during training, which can allow higher learning rates and faster training. | 2 min quick read |
| Batch size | Batch size is how many training examples are grouped before the weights change once. | 3 min explainer |
| Bayesian inference | Bayesian inference treats an unknown number as uncertain, starts from a belief about it, and uses Bayes' theorem to update that belief as data arrives. | 5 min explainer |
| Beam search | Beam search keeps a fixed number of high-scoring partial sequences, expands them, and retains the best candidates at each step. | 3 min explainer |
| Benchmark contamination | Benchmark contamination happens when test questions, or close copies of them, end up in a model's training data, so its score can look better than its real skill. | 4 min explainer |
| Benchmarks | A benchmark is a fixed set of test tasks with a scoring rule, so different AI models can be compared on the same exam. | 3 min explainer |
| BERT | BERT is a 2018 Google language model that reads text in both directions at once, so it understands each word using the words before and after it. | 2 min quick read |
| bfloat16 | bfloat16 is a 16-bit number format built for machine learning that keeps the same range as 32-bit floats while using half the memory. | 2 min quick read |
| BGE | BGE, short for BAAI General Embedding, is a family of open embedding and reranking models from the Beijing Academy of Artificial Intelligence, built mainly for retrieval in RAG systems. | 2 min quick read |
| Bias in AI | Bias in AI is when a system treats people or groups unfairly, usually because of patterns in its training data or choices made while building it. | 2 min quick read |
| Bias–variance tradeoff | Bias is average error across training sets, while variance indicates sensitivity to varying training sets. | 4 min explainer |
| BIG-bench | BIG-bench is a large, community-built benchmark of more than 200 language tasks, designed to probe what language models can and cannot yet do. | 2 min quick read |
| bitsandbytes | bitsandbytes is an open-source PyTorch library that shrinks large language models with 8-bit and 4-bit quantisation so they need far less memory. | 2 min quick read |
| BLEU and ROUGE | BLEU and ROUGE are automatic scores that judge machine-written text by how much its words and phrases overlap with human-written reference text. | 2 min quick read |
| BM25 | BM25 is a classic formula search engines use to rank documents by how well their words match a query, and it is still the default in tools like Elasticsearch. | 2 min quick read |
| Bolt.new | Bolt.new is an AI tool that turns a written request into a working website, web app or mobile app, all inside your web browser. | 2 min quick read |
| Braintrust | Braintrust is a platform for evaluating and observing AI apps and agents, used to trace production runs and score outputs before and after release. | 2 min quick read |
| Browser agents | A browser agent is an AI system that uses a web browser for you, opening pages, clicking, typing and reading, to complete a task you describe. | 2 min quick read |
| Browser automation agent | A browser automation agent is an AI system that operates a web browser for you, clicking, typing and reading pages to finish a task you describe in words. | 2 min quick read |
| Browser Use | Browser Use is an open-source library that lets an AI agent operate a real web browser to finish a task you describe in words. | 1 min quick read |
| Byte-pair encoding | Byte-pair encoding builds a subword vocabulary by repeatedly merging the most frequent adjacent symbol pair. | 3 min explainer |
| C4 | C4, the Colossal Clean Crawled Corpus, is a large, cleaned-up set of English web text made from Common Crawl, first built to train Google's T5 model. | 2 min quick read |
| Canva AI | Canva AI is the AI assistant built into the Canva design tool: you describe what you need, and it produces designs, text, images and more that you can then edit. | 2 min quick read |
| Catastrophic forgetting | Catastrophic forgetting is when a neural network trained on a new task loses much of what it learned on an earlier one. | 2 min quick read |
| CatBoost | CatBoost is an open-source gradient boosting library from Yandex that handles text-like category columns, such as city or colour, without you first converting them to numbers. | 2 min quick read |
| Cerebras | Cerebras is a chip company that builds one very large AI processor from a whole silicon wafer and offers fast AI inference as a cloud service. | 2 min quick read |
| Chain of thought | Chain-of-thought prompting asks for or demonstrates intermediate reasoning text before a final answer, but that text is not a guaranteed view of hidden model internals. | 4 min explainer |
| Character.AI | Character.AI is an entertainment app where people chat and tell stories with AI characters, many of them made by other users. | 1 min quick read |
| Chat templates | A chat template converts a structured list of messages into a formatted sequence. | 3 min explainer |
| Chat with PDF | A chat-with-PDF app lets you upload a PDF and ask questions about it in plain language, with answers drawn from the file. | 2 min quick read |
| Chat with your database | A chat-with-your-database app lets people ask questions in plain English and gets answers by having a language model write and run database queries. | 2 min quick read |
| Chatbots | A chatbot is software you can talk to in text or voice; modern AI chatbots use large language models to understand requests and write natural replies. | 2 min quick read |
| ChatGPT | ChatGPT is OpenAI’s conversational product for asking questions, creating things and completing work with models and tools. | 5 min explainer |
| ChatGPT Work | ChatGPT Work is an OpenAI product in which an agent takes your team's real materials and produces a first draft of a finished piece of work. | 2 min quick read |
| Checkpoints | A checkpoint is a saved snapshot of a model's learned values, so training can stop and pick up later, or the model can be loaded and used. | 1 min quick read |
| Chinchilla scaling | Chinchilla scaling allocates a fixed training-compute budget by growing model parameters and training tokens in roughly equal proportions. | 3 min explainer |
| Chroma | Chroma is an open-source database for AI apps that stores documents with their embeddings and metadata, and finds the ones most similar to a query. | 1 min quick read |
| Chunking | Chunking splits a long document into smaller pieces that can be embedded, searched, or fitted into a model prompt. | 3 min explainer |
| Classification | Classification predicts which discrete category or categories an input belongs to, such as identifying an email as spam or legitimate. | Definition |
| Classifier-free guidance | Classifier-free guidance is a trick in image generators that makes pictures follow the prompt more closely, at the cost of some variety. | 2 min quick read |
| Claude | Claude is Anthropic's family of AI models and the assistant built on them, used for writing, analysis, coding and agent work. | 2 min quick read |
| Claude Agent SDK | Anthropic's Claude Agent SDK is a Python and TypeScript library for building agents that run on the same loop, tools and context handling as Claude Code. | 5 min explainer |
| Claude app | Claude is Anthropic’s AI product for conversation, analysis and completing tasks with files, connected tools and editable outputs. | 5 min explainer |
| Claude Code | Claude Code is Anthropic's coding assistant that reads a project, edits files and runs commands for you, asking before risky steps. | 4 min explainer |
| Claude Code Game Studios | Claude Code Game Studios is an open-source Claude Code template that turns a described game idea into a short brief and a playable browser game, with an optional 49-agent studio mode. | 2 min quick read |
| Claude Duet | Claude Duet is Anthropic's app in which Claude acts as a proactive agent: it points out what needs you, works on tasks while you are away and asks before acting for you. | 1 min quick read |
| Claude Fable 5.1 | Claude Fable 5.1 is an Anthropic model, introduced on September 1, 2026, aimed at hard coding and knowledge work, including problems that take a long time to solve. | 2 min quick read |
| Claude Fable 5.2 | Claude Fable 5.2 is Anthropic's top Claude model, released in November 2026 for hard reasoning, coding and long-running agent work. | 2 min quick read |
| Claude Haiku | Claude Haiku is the quickest, lowest-cost tier of Anthropic's Claude models; the current version, Claude Haiku 5.5, came out in October 2026. | 1 min quick read |
| Claude Meetings | Claude Meetings is an Anthropic feature, now in beta, that brings Claude into a video call as a visible participant who helps while people talk. | 2 min quick read |
| Claude Opus 5.5 | Claude Opus 5.5 is an Anthropic model released in September 2026 for long-running coding, agent and knowledge work, and the first of its Claude 5.5 family. | 2 min quick read |
| Claude Sonnet 5.5 | Claude Sonnet 5.5 is Anthropic's faster, cheaper partner to Claude Opus 5.5, aimed at everyday tasks such as fixing bugs and making documents, slides and spreadsheets. | 2 min quick read |
| Claude-Mem | Claude-Mem is an open-source plugin that gives AI coding agents such as Claude Code a memory that lasts from one working session to the next. | 2 min quick read |
| Clef | Clef is a set of open decision models from Cloudflare: instead of writing text, they answer fixed questions with a probability for each allowed choice. | 2 min quick read |
| CLI-Anything | CLI-Anything is an open-source project from HKUDS that gives ordinary software a command-line interface, so AI agents such as Claude Code, Cursor and OpenClaw can operate it by typing commands. | 2 min quick read |
| Cline | Cline is an open-source AI coding agent that works in your editor, terminal or a desktop app, and asks your approval before it acts. | 2 min quick read |
| CLIP | CLIP is a 2021 OpenAI model that learns to match images with text, so it can label pictures it was never specifically trained to sort. | 2 min quick read |
| Closed models | A fully closed model keeps its weights and code proprietary for internal use. | 3 min explainer |
| Cloudflare OS | Cloudflare OS is an open-source AI workspace from Cloudflare where people ask agents to do work and build small personal apps that run in sandboxes. | 2 min quick read |
| Cloudflare Workers AI | Cloudflare Workers AI lets developers run open-source AI models on GPUs in Cloudflare's network, paying only for what they use. | 2 min quick read |
| Clustering | Clustering sorts data that has no labels into groups of items that resemble each other. | 4 min explainer |
| COCO | COCO (Common Objects in Context), first published as Microsoft COCO, is a large image dataset used to train and test models that find, outline and describe objects in photos. | 2 min quick read |
| Code generation | Code generation is AI writing computer code from a description in plain language, such as a comment or a request. | 2 min quick read |
| Code interpreters | A code interpreter lets an AI model write a small program, run it in a sealed-off sandbox, and use the result in its answer. | 4 min explainer |
| Code models | Code Llama is a family of language models for code with infilling and instruction-following capabilities. | 3 min explainer |
| Code review bot | A code review bot is an AI tool that reads a pull request and leaves comments pointing out bugs, risks and possible fixes before a person merges it. | 2 min quick read |
| Codestral | Codestral is Mistral AI's model for writing and completing code, first released in May 2024 and trained on more than 80 programming languages. | 2 min quick read |
| Coding assistant | A coding assistant is an AI tool that helps you write, change and understand code, from suggesting lines to editing many files and running commands. | 2 min quick read |
| Cohere Embed | Cohere Embed is Cohere's family of embedding models, offered through an API, that turn text and images into vectors for search, classification and RAG. | 2 min quick read |
| Cohere Platform | The Cohere platform is Cohere's API for its language models, including Command for text, Embed for search and Rerank for ordering results. | 2 min quick read |
| Cohere Rerank | Cohere Rerank is a model, offered through Cohere's API, that takes a search query plus a set of candidate texts and sorts those texts so the best matches come first. | 2 min quick read |
| ColBERT | ColBERT is a search model that compares queries and documents word by word using many small embeddings, rather than squeezing each into one vector. | 2 min quick read |
| Collaborative filtering | Collaborative filtering recommends items by learning from what many people rated, clicked or watched, then predicting the gaps in your own history. | 4 min explainer |
| ColPali | ColPali is an openly released model that searches documents by looking at images of their pages directly, instead of first extracting the text. | 2 min quick read |
| ComfyUI | ComfyUI is a free, open-source app where you build AI image, video and audio generators by wiring boxes called nodes together. | 4 min explainer |
| Command A | Command A is a large language model from Cohere built for business use, especially for agents, tool use and answering questions from company documents. | 2 min quick read |
| Common Crawl | Common Crawl is a non-profit that regularly crawls the web and publishes the pages as a free, open archive, and cleaned-up versions of it, such as C4, are used as text datasets for AI. | 2 min quick read |
| Computer use | Computer use lets an AI interpret a graphical interface and take actions such as clicking, typing, scrolling, or reading screen content. | Definition |
| Computer vision | Computer vision is the part of AI that lets machines understand images and video, from recognising objects to reading text and tracking movement. | 2 min quick read |
| Computer-use models | A computer-use model reads a screen and emits mouse or keyboard actions that an application can execute. | 4 min explainer |
| Confusion matrix | A confusion matrix is a table that counts a classifier's predictions by what it predicted and what was really true, so you see which mistakes it makes. | 5 min explainer |
| Constitutional AI | Constitutional AI trains and evaluates model behavior against an explicit written set of principles. | 4 min explainer |
| Content moderation pipeline | A content moderation pipeline automatically checks user posts, and sometimes AI replies, for harmful content and decides whether to allow, block or send them to a human. | 2 min quick read |
| Content provenance | Content provenance is a verifiable record of where a photo, video or file came from and how it was changed, including whether AI was used. | 2 min quick read |
| Context compaction | Context compaction swaps the older part of a long AI conversation for a short summary, so the chat can keep going inside the model's memory limit. | 4 min explainer |
| Context engineering | Context engineering is choosing, trimming and updating the tokens a model sees at each step, so an AI agent works from a small, useful context. | 4 min explainer |
| Context window | A context window is all the text a language model can reference while generating a response, including the response itself. | 3 min explainer |
| Contextual retrieval | Contextual retrieval has a model write a short note that places each chunk in its document, then indexes the note with the chunk so search can find it. | 5 min explainer |
| Continual learning | Continual learning trains a model on a stream of changing data while measuring whether later learning damages or helps earlier tasks. | 4 min explainer |
| Continuous batching | Continuous batching is a way of serving language models where finished requests leave the batch and new ones join at every step, instead of waiting for the whole batch. | 2 min quick read |
| ControlNet | ControlNet is an add-on network for image diffusion models that steers the picture with an extra guide image, such as an edge outline, depth map or human pose. | 2 min quick read |
| Convolutional neural networks | A convolutional neural network uses sparse convolutions that reuse the same weights at multiple locations in an ordered grid. | 3 min explainer |
| Copilot+ PCs | Copilot+ PCs are a class of Windows 11 laptops and PCs with a built-in AI chip, an NPU, fast enough to run AI features directly on the device. | 2 min quick read |
| Copyright and AI | Copyright and AI covers two big legal questions: can AI-made work be protected by copyright, and is it legal to train AI on copyrighted material? | 2 min quick read |
| Coqui TTS | Coqui TTS is a Python library for turning text into speech, with pretrained voices in many languages and tools to train or fine-tune your own. | 1 min quick read |
| Cosine similarity | Cosine similarity scores how closely two vectors point the same way, ignoring how long they are. The score runs from −1 to 1. | 4 min explainer |
| Cost optimization | Cost optimization for AI apps means cutting what you pay for model calls, using fewer tokens, smaller models, caching and cheaper batch processing, without losing too much quality. | 2 min quick read |
| CPUs | A CPU is the chip that runs a computer's software, carrying out instructions with just a few cores backed by lots of cache memory. | 5 min explainer |
| Crawl4AI | Crawl4AI is an open-source tool that crawls and scrapes websites for AI, turning any site into clean Markdown ready for language models, RAG and agents. | 1 min quick read |
| CrewAI | CrewAI is an open-source toolkit, written in Python, for building teams of AI agents, each with a role, that work through a list of tasks together. | 4 min explainer |
| Cross-validation | Cross-validation checks how well a model will do on new data by testing it on parts of the data it did not train on, several times over. | 2 min quick read |
| Curriculum learning | Curriculum learning trains a model on easier examples first and harder ones later, the way a school course builds up. | 2 min quick read |
| Cursor | Cursor is a coding tool with an AI agent that can read your project, edit files and run commands for you. | 4 min explainer |
| Cursor Agent SDK | Cursor's Agent SDK is a framework for writing AI agents as code, testing them and deploying them. Cursor builds its own agents, such as Bugbot, with it. | 2 min quick read |
| Custom GPTs | Custom GPTs are versions of ChatGPT set up for one purpose, with their own instructions, files and tools. OpenAI now plans to retire them in favour of Plugins. | 1 min quick read |
| Customer support bot | A customer support bot is a chat assistant that answers customer questions, looks things up and takes simple actions, handing hard cases to a person. | 2 min quick read |
| Data analysis agent | A data analysis agent answers questions about a spreadsheet or dataset by writing and running code, then reporting numbers and charts. | 2 min quick read |
| Data and concept drift | Data drift is when the data a deployed model sees changes from what it was trained on; concept drift is when the link between inputs and right answers changes. | 2 min quick read |
| Data augmentation | Data augmentation makes extra training examples by changing existing ones in small ways that keep the answer the same, such as flipping or cropping a photo. | 4 min explainer |
| Data curation | Data curation is choosing, cleaning and filtering training data so a model learns from better examples; a model's quality depends heavily on its data. | 2 min quick read |
| Data labeling | Data labeling attaches the answer a supervised model should learn to raw examples such as images, text, audio or table rows. | 3 min explainer |
| Data leakage | Data leakage is when a model learns from information it will not have when making real predictions, so its test score looks better than it really is. | 4 min explainer |
| Data parallelism | Data parallelism runs replicas of one model on different slices of a batch, then combines their gradients before a synchronized update. | 4 min explainer |
| Data poisoning | Data poisoning is an attack where someone slips harmful examples into a model's training data so the model learns hidden, unwanted behaviour. | 2 min quick read |
| Data privacy | Data privacy is about what an AI company may do with your chats, such as training models, and the settings that let you limit it. | 5 min explainer |
| Databricks Mosaic AI | Databricks Mosaic AI brings model development, evaluation, retrieval, agents and governance into the Databricks data and lakehouse environment. | Definition |
| Decision trees | A decision tree makes a prediction by asking a series of yes-or-no style questions about the data, like a flowchart, until it reaches an answer. | 2 min quick read |
| Decoder-only models | A decoder-only model predicts each next token from the tokens that came before it. | 3 min explainer |
| Deep learning | Deep learning is machine learning with neural networks that stack many layers, so each layer can build a more abstract picture of the data than the one before. | 4 min explainer |
| Deep research agents | A deep research agent is an AI tool that spends minutes searching and reading many sources on its own, then writes a long, cited report on your question. | 2 min quick read |
| DeepEval | DeepEval is an open-source Python framework for unit testing LLM apps, a bit like Pytest, with ready-made metrics such as answer relevancy and hallucination. | 2 min quick read |
| Deepfakes | A deepfake is AI-generated or AI-altered image, audio or video that looks or sounds like a real person, place or event but is not authentic. | 2 min quick read |
| Deepgram | Deepgram is a voice AI company whose developer APIs turn speech into text, turn text into speech, and run full voice agents. | 2 min quick read |
| DeepInfer | DeepInfer is DeepSeek's open-source inference framework, the software that runs its models to answer requests for its large cloud services. | 2 min quick read |
| DeepSeek API | The DeepSeek API lets developers call DeepSeek's language models from code using the same request formats as OpenAI's or Anthropic's APIs. | 2 min quick read |
| DeepSeek Harness | DeepSeek Harness is an open-source agent app from DeepSeek for everyday tasks and coding, which you extend with add-ons called plugins. | 2 min quick read |
| DeepSeek Search API | The DeepSeek Search API lets developers search the web with DeepSeek's own search engine, and its companion Fetch API turns a single web page into Markdown for an AI model to read. | 2 min quick read |
| DeepSeek-R1 | DeepSeek-R1 is an open reasoning model from the AI lab DeepSeek that learned to think through problems step by step mostly by reinforcement learning. | 2 min quick read |
| DeepSeek-V4.1-Flash | DeepSeek-V4.1-Flash is an open-weight model from DeepSeek that reads text and images, can take in up to a million tokens, and is built to make long agent tasks cheaper. | 2 min quick read |
| DeepSpeed | DeepSpeed is an open-source library, once a key piece of Microsoft's AI at Scale initiative, that makes training and running very large models across many GPUs practical. | 2 min quick read |
| Descript | Descript is a video and audio editor where you change a recording by editing its written transcript, with AI tools to help. | 1 min quick read |
| Devin | Devin is an AI software engineer made by Cognition that can take on coding tasks by itself, writing, running and testing code. | 2 min quick read |
| Devstral | Devstral is Mistral AI's family of coding models built for software engineering agents that fix real issues across whole codebases. | 2 min quick read |
| Differential privacy | Differential privacy is a way to learn from data about many people while making sure the result reveals very little about any one person. | 2 min quick read |
| Diffusers | Diffusers is Hugging Face's open-source Python library for running and training diffusion models that generate images, video and audio. | 2 min quick read |
| Diffusion models | A diffusion model learns to reverse a process that gradually adds noise to data. | 3 min explainer |
| Dify | Dify is a platform for building AI apps such as chatbots, agents and RAG pipelines on a visual canvas, with a self-hostable community edition and a managed cloud. | 2 min quick read |
| Dimensionality reduction | Dimensionality reduction represents data using fewer variables while preserving useful structure, aiding visualization, compression, denoising, or downstream modeling. | Definition |
| DINOv2 | DINOv2 is a Meta vision model from 2023 that learns general-purpose image features from pictures alone, with no labels or captions. | 2 min quick read |
| Discord bot | An AI Discord bot is a bot account in a Discord server that reads messages or slash commands and replies using a language model. | 2 min quick read |
| Distillation | Distillation trains a small student model to copy the full answer probabilities of a large teacher model, so you can run something smaller and cheaper. | 5 min explainer |
| Docker Agent | Docker Agent is Docker's open-source framework for putting together teams of AI agents that are described in a YAML file and run from the terminal. | 2 min quick read |
| Docling | Docling is an open-source, MIT-licensed toolkit that converts PDFs and many other document formats into clean Markdown or JSON, keeping layout, reading order and tables. | 1 min quick read |
| Document parsing | Document parsing turns files such as PDFs, slides and spreadsheets into clean, structured text that AI systems can search and read. | 2 min quick read |
| Document Q&A | Document Q&A is a system that answers natural-language questions using the contents of your documents, such as forms, reports or scanned pages. | 2 min quick read |
| Dolma | Dolma is an open, three-trillion-token English text dataset from the Allen Institute for AI, used to pretrain its OLMo language models. | 2 min quick read |
| Dot product | The dot product multiplies two equal-length lists of numbers position by position and adds the results into one number set by the angle between them and their lengths. | 4 min explainer |
| DPO | DPO teaches a language model to prefer the better of two answers by learning from the comparison itself, with no separate reward model trained first. | 4 min explainer |
| Dropout | During training, dropout randomly replaces some input tensor elements with zero. | 3 min explainer |
| Drug discovery | AI in drug discovery uses models to predict how molecules and proteins behave, helping scientists find and design promising medicines faster. | 2 min quick read |
| DSPy | DSPy is a Python framework where you describe an AI job by naming its inputs and outputs, and it tunes the prompts for you. | 4 min explainer |
| DVC | DVC (Data Version Control) is an open-source command-line tool that versions datasets and models alongside your code in Git, so machine learning projects can be reproduced. | 2 min quick read |
| E5 | E5 is a family of open text embedding models that turn text into vectors for search, clustering and classification, trained on a large set of text pairs. | 2 min quick read |
| Edge0-145B | Edge0-145B is an open-source language model from Edge0 with 145 billion parameters, built to run on a personal laptop rather than in a data centre. | 2 min quick read |
| Elasticsearch | Elasticsearch is a distributed engine from Elastic for searching and analysing data that also works as a vector database, combining keyword search, vector search and filters in one system. | 2 min quick read |
| EleutherAI | EleutherAI is a non-profit institute for AI research that trains and releases open models, datasets and tools, and studies how to make open models safe. | 2 min quick read |
| Eleven v4 | Eleven v4 is a text-to-speech model from ElevenLabs built to sound emotional and natural, with a faster Turbo version for voice agents that must reply quickly. | 2 min quick read |
| ElevenLabs | ElevenLabs is an AI audio company whose tools turn text into lifelike speech, transcribe speech, clone voices, and generate music and sound effects. | 1 min quick read |
| Email assistant | An AI email assistant reads incoming mail, decides what needs attention and drafts replies for you to check and send. | 2 min quick read |
| Embedding layer | An embedding layer turns integer indexes into dense vectors of fixed size. | 3 min explainer |
| Embedding models | An embedding model maps an input to a numeric vector whose distance from another vector can measure relatedness. | 3 min explainer |
| EmbeddingGemma | EmbeddingGemma is Google's family of small open embedding models, built on Gemma, that turn content into vectors for search and retrieval on your own device. | 2 min quick read |
| Embeddings | An embedding maps a discrete item such as a token ID to a dense vector of numbers. | 3 min explainer |
| Emergent abilities | Emergent abilities are task capabilities that appear absent in smaller models but present at a larger scale. | 3 min explainer |
| Encoder-decoder | An encoder-decoder model combines an encoder with an autoregressive decoder for sequence generation. | 3 min explainer |
| Encoder-only models | An encoder-only model uses bidirectional self-attention to represent a supplied sequence. | 3 min explainer |
| Ensemble learning | Ensemble learning combines the predictions of several models, usually by a vote or an average, so that their separate mistakes partly cancel out. | 4 min explainer |
| Epoch | An epoch is one complete pass through every example used to train a model. | 4 min explainer |
| EU AI Act | The EU AI Act is the European Union's law on artificial intelligence, which sorts AI uses by risk and sets stricter rules for the riskier ones. | 2 min quick read |
| Evals | An eval checks an AI system's work. You feed it an input and use a set of rules to score what comes back. | 5 min explainer |
| Evaluation harness | An evaluation harness is a tool that runs a model through many test tasks in the same way every time and scores the results, so models can be compared fairly. | 2 min quick read |
| ExecuTorch | ExecuTorch is PyTorch's open-source toolkit for running AI models on the device itself, such as a phone, headset, laptop, browser or tiny chip, instead of on a server. | 1 min quick read |
| Existential risk | Existential risk from AI is the concern that very advanced AI could cause a catastrophe on the scale of human extinction, and that reducing that risk should be a global priority. | 2 min quick read |
| Expert systems | An expert system stores a specialist's know-how as if-then rules and uses a separate inference engine to apply those rules to a new case and explain its advice. | 5 min explainer |
| F1 score | The F1 score rolls a classifier's precision and recall into one number from 0 to 1, and it stays low if either of them is low. | 4 min explainer |
| FAISS | FAISS is an open-source library from Meta for quickly finding the most similar vectors in large collections, a core job of vector databases. | 2 min quick read |
| faster-whisper | faster-whisper runs OpenAI's Whisper speech-to-text models through the CTranslate2 inference engine, so audio is transcribed faster and with less memory. | 1 min quick read |
| FastMCP | FastMCP is an open-source Python framework for building servers, clients and apps that use the Model Context Protocol (MCP). | 2 min quick read |
| Feature engineering | Feature engineering turns raw data, like prices and colour names, into the lists of numbers a model can learn from. | 4 min explainer |
| Feature importance | Feature importance gives each input column a score for how much a trained model relies on it, so you can see what drives its predictions. | 4 min explainer |
| Feature scaling | Feature scaling puts numeric variables on comparable ranges, limiting disproportionate influence from large-magnitude features in methods sensitive to distances or gradient sizes. | Definition |
| Features | Features are the input facts a machine learning model reads about each example, such as a car's mileage or colour, turned into a list of numbers. | 5 min explainer |
| Federated learning | Federated learning trains a shared model across many devices, such as phones, without collecting their data in one place. | 2 min quick read |
| Few-shot learning | Few-shot learning is the ability to perform a task from only a small number of labelled examples. | 3 min explainer |
| Few-shot prompting | Few-shot prompting places a small set of example inputs and outputs in the prompt so the model can imitate the task on a new input. | 3 min explainer |
| Fine-tuning | Fine-tuning takes a model that is already trained and keeps training it on a smaller set of examples for one task. | 4 min explainer |
| Fine-tuning pipeline | A fine-tuning pipeline is the repeatable set of steps that prepares example data, trains an existing model on it, and checks whether the new version is actually better. | 2 min quick read |
| FineWeb | FineWeb is a large, openly released collection of cleaned English web text, made for pretraining language models. | 2 min quick read |
| Firecrawl | Firecrawl is a service that gets web content for AI agents: it searches the web, scrapes pages and returns them as clean Markdown or structured data. It is open source and also hosted. | 1 min quick read |
| Fireworks AI | Fireworks AI is a cloud platform for running and fine-tuning AI models, either pay-per-token or on dedicated GPUs. | 2 min quick read |
| FlashAttention | FlashAttention computes exact attention with tiling that reduces reads and writes between GPU memory levels. | 3 min explainer |
| FlashProve | FlashProve is Anthropic's open-source research prototype for showing, with a proof, which AI model produced a reply, without exposing that model's weights. | 2 min quick read |
| FLOPs | FLOPs count the small arithmetic steps, like one multiply or one add on decimal numbers, that a computer does to train an AI model. | 5 min explainer |
| Flow matching | Flow matching is a way to train generative models to turn random noise into data along smooth paths, often simpler and faster than classic diffusion training. | 2 min quick read |
| FLUX | FLUX is a family of image generation models from Black Forest Labs, offered both through a paid API and as open-weight models you can download. | 2 min quick read |
| Foundation models | A foundation model is trained on broad data and can be adapted for many downstream tasks. | 3 min explainer |
| FP8 | FP8 is an 8-bit floating-point number format that lets AI chips train and run models faster, using a quarter of the memory of 32-bit numbers. | 2 min quick read |
| Fraud detection | AI fraud detection scores transactions and accounts in real time for signs of fraud, so suspicious ones can be blocked or reviewed before money is lost. | 2 min quick read |
| Frequency and presence penalties | Frequency and presence penalties are settings that make a language model less likely to repeat words it has already used in its answer. | 1 min quick read |
| Frontier models | Frontier models are highly capable general-purpose AI models at or beyond the capabilities of the most advanced current models. | 3 min explainer |
| Full Self-Driving (Supervised) | Full Self-Driving (Supervised) is Tesla’s driver-assistance software for navigation and vehicle control; the human driver must remain attentive and responsible. | Definition |
| Function calling | Function calling lets a model request a named function by returning structured arguments that an application, or the provider, then runs. | 4 min explainer |
| Gamma | Gamma is an AI tool that drafts presentations, documents, webpages and social posts from a topic, your own text, or an existing file. | 1 min quick read |
| GANs | A GAN trains a generator to make samples and a discriminator to distinguish generated samples from training data. | 3 min explainer |
| Gemini | Gemini is Google's family of AI models, built from the start to handle text, code, images, audio and video, and the name of Google's AI assistant. | 2 min quick read |
| Gemini 3.5 | Gemini 3.5 is a generation of Google's Gemini models: 3.5 Flash-Lite, the now-deprecated 3.5 Flash, and speech models for transcription and live translation. | 2 min quick read |
| Gemini 4 Argon | Gemini 4 Argon is Google's next frontier AI model, built for deep reasoning over long, complex tasks, and released first to a small group of trusted cyber defenders. | 2 min quick read |
| Gemini 4 Carbon | Gemini 4 Carbon is a frontier AI model from Google DeepMind that Google calls its strongest, built for long, many-step tasks and for finding and fixing security flaws. | 2 min quick read |
| Gemini 4 Hydrogen | Gemini 4 Hydrogen is a Google model built to carry out agent tasks quickly, and the model that brought Gemini 4 to all Google AI Studio tiers, the Gemini app and Google Search. | 2 min quick read |
| Gemini API | The Gemini API is Google's service for calling Gemini and other Google models, such as Nano Banana and Lyria, from your own code. | 2 min quick read |
| Gemini app | Gemini is Google’s AI app: you ask questions, upload files for answers and summaries, and can have it research a topic across many sources. | 5 min explainer |
| Gemini CLI | Gemini CLI is an open-source tool from Google: an AI agent that lets you use Gemini models from your terminal for coding and other tasks. | 2 min quick read |
| Gemini Decisions API | The Gemini Decisions API is a Google endpoint that answers closed questions, such as yes or no, which option fits or which rubric level applies, instead of writing free text. | 1 min quick read |
| Gemini Enterprise | Gemini Enterprise is Google Cloud's AI product for companies: a work assistant that searches company data, answers questions and runs AI agents. | 2 min quick read |
| Gemini Nano | Gemini Nano is the most efficient model in Google's Gemini family, designed to run directly on phones and other devices instead of in the cloud. | 1 min quick read |
| Gemini Notebook | Gemini Notebook, formerly NotebookLM, is Google's AI research assistant that answers questions using only the sources you give it, with citations. | 1 min quick read |
| Gemini Omni Flash | Gemini Omni Flash is Google's fast model for making and editing short videos with sound from text, images and video in the Gemini API. | 1 min quick read |
| Gemini Robotics | Gemini Robotics is Google DeepMind's family of AI models, built on Gemini, that let robots understand what they see and control their movements. | 2 min quick read |
| Gemma | Gemma is Google's family of lightweight open models, built from the same research as Gemini, that anyone can download and run. | 2 min quick read |
| Gemma 4 | Gemma 4 is Google's open model family released in April 2026 under the Apache 2.0 licence, with sizes from phones up to workstations. | 2 min quick read |
| Generative AI | Generative AI learns patterns from existing data, then uses those patterns to create new text, images, audio, video, code or other content. | 5 min explainer |
| Generative engine optimization | Generative engine optimization (GEO) is making web content more likely to be found, used and cited by AI search tools that write answers from many sources. | 2 min quick read |
| Genetic algorithms | A genetic algorithm searches for good answers the way breeding does. It keeps a population of candidates, mixes the fitter ones and repeats over many generations. | 4 min explainer |
| Genie 4 | Genie 4 is Google DeepMind's world model: an AI that simulates a place you can move through in real time, makes its sounds, and changes it when you type. | 2 min quick read |
| Genkit | Genkit is Google's open-source toolkit for making AI-powered and agent apps, with one interface for models from many providers. | 2 min quick read |
| GGUF | GGUF is a single-file format for storing AI models so they can be loaded quickly and run by llama.cpp and other GGML-based tools. | 2 min quick read |
| GitHub Copilot | GitHub Copilot is an AI helper for programmers that offers the next lines of code while you write, explains code when you ask, and can take on jobs you give it. | 4 min explainer |
| Glean | Glean is an AI platform for companies that searches across all of a team's work apps and uses what it finds to answer questions and run agents. | 2 min quick read |
| Glimpse | Glimpse is Elorian's first AI model, a "visual thinking" model released on 16 October 2026 that reasons about pictures and video by going back to look at them while it thinks. | 2 min quick read |
| GLM-4.5 | GLM-4.5 is an open mixture-of-experts language model from Z.ai, released in August 2025 and built for agents, reasoning and coding in one model. | 2 min quick read |
| GloVe | GloVe (Global Vectors) is a method from Stanford for learning word embeddings from counts of how often words appear near each other across a whole body of text. | 2 min quick read |
| Google ADK | Google ADK is an open-source toolkit from Google for writing AI agents in code, giving them tools, a record of each chat and helper agents. | 5 min explainer |
| Google AI Overviews | Google AI Overviews are AI-written summaries that appear at the top of some Google Search results, with links to the web pages behind them. | 1 min quick read |
| Google AI Studio | Google AI Studio is a website where you can try Google's Gemini models, test prompts and settings, and then get code for the Gemini API. | 2 min quick read |
| Google Colab | Google Colab is a free, hosted notebook service where you write and run Python in the browser, with access to GPUs and TPUs. | 2 min quick read |
| Google Flow | Google Flow is Google's AI filmmaking tool for creating video clips and scenes for your stories with its generative models, such as Veo. | 2 min quick read |
| Google TPU | A Google TPU (Tensor Processing Unit) is a chip Google designed specifically to speed up machine learning, available to others through Google Cloud. | 2 min quick read |
| Google Workspace with Gemini | Google Workspace with Gemini, formerly Gemini for Google Workspace, puts Google's Gemini AI inside Gmail, Docs, Sheets, Meet and other work apps. | 2 min quick read |
| GPQA | GPQA is a 448-question multiple-choice benchmark written by experts in biology, physics and chemistry. | 3 min explainer |
| GPT | GPT is OpenAI's family of language models: transformers first trained on huge amounts of text, then adapted for specific tasks. | 2 min quick read |
| GPT Image | GPT Image is OpenAI's current family of image models, which create new pictures from text and edit existing ones. | 2 min quick read |
| GPT-2 | GPT-2 is a 2019 OpenAI model, 1.5 billion parameters in size, trained with one goal: guess the next word of web text. | 1 min quick read |
| GPT-3 | GPT-3 is a 2020 OpenAI model, 175 billion parameters in size, which showed a big model could do new tasks from a few examples in the prompt. | 2 min quick read |
| GPT-4 | GPT-4 is a large OpenAI model from 2023 that reads both text and images and answers in text. | 2 min quick read |
| GPT-5 | GPT-5 is an OpenAI model family released on 7 August 2025 that routes each question to either a fast model or a deeper reasoning model. | 2 min quick read |
| GPT-6 Astra | GPT-6 Astra is OpenAI's top model in its API, meant for the hardest jobs such as complex reasoning, coding and research. | 2 min quick read |
| GPT-6 Cyber | GPT-6 Cyber is an OpenAI model for security work: it helps approved teams look into software flaws, recreate them safely and check that fixes hold. | 1 min quick read |
| GPT-6.1 Terra | GPT-6.1 Terra is an OpenAI API model built to be fast and efficient while it understands images, operates app screens and writes clear text. | 1 min quick read |
| gpt-oss | gpt-oss is a pair of open-weight reasoning models from OpenAI, released in August 2025 under the Apache 2.0 licence. | 2 min quick read |
| GPT-Rosalind | GPT-Rosalind is OpenAI's series of reasoning models for biology, drug discovery and translational medicine (turning lab findings into treatments), offered only to researchers OpenAI has vetted. | 2 min quick read |
| GPUs | A GPU is a chip built to run many similar calculations at the same time, which suits the matrix maths inside neural networks. | 5 min explainer |
| Gradient boosting | Gradient boosting builds one strong model from many small decision trees, added one at a time, each trained to fix the errors the trees before it still make. | 4 min explainer |
| Gradient checkpointing | Gradient checkpointing stores fewer activations by recomputing them when needed, exchanging extra computation for lower memory use. | 4 min explainer |
| Gradient descent | Gradient descent repeatedly measures how loss changes with each parameter and moves the parameters a small step toward lower loss. | 3 min explainer |
| Gradio | Gradio is an open-source Python library for turning a machine learning model, or any Python function, into a simple web app you can share with a link. | 1 min quick read |
| Grammarly | Grammarly is an AI writing assistant that checks and improves what you type across many apps and websites. | 2 min quick read |
| Granite | Granite is IBM's family of open AI models for businesses, released under the Apache 2.0 licence, covering language, code, vision, speech, embeddings, safety and time series. | 2 min quick read |
| Graph neural networks | A graph neural network updates each node's hidden state based on messages. | 3 min explainer |
| Graphify | Graphify is an open-source tool that maps a codebase, with its docs, database schemas, settings and PDFs, into a knowledge graph that AI coding assistants can query. | 1 min quick read |
| GraphRAG | GraphRAG turns your documents into a map of people, places and links, groups and summarises it, then answers questions from that map. | 4 min explainer |
| Greedy decoding | Greedy decoding generates a sequence by choosing the most likely next token at every step. | 3 min explainer |
| Grok | Grok is xAI's family of AI models and the chatbot of the same name; its site now uses the name SpaceXSI. It can chat, search the web and X, and create images and video. | 2 min quick read |
| Grok 4.7 | Grok 4.7 is a model from SpaceXSI, the company behind Grok, released in September 2026 for coding and knowledge work. | 2 min quick read |
| Grok app | Grok is the assistant from SpaceXAI (formerly xAI) for conversation, web and X search, file analysis, voice and media creation. | 5 min explainer |
| Grok Voice | Grok Voice is SpaceXSI's speech-to-speech model: it listens to what you say and replies out loud in real time, so apps can hold spoken conversations. Version 3.0 arrived on 14 October 2026. | 2 min quick read |
| Groq | Groq is a company that runs AI models on its own chip, the LPU, and sells fast inference through the GroqCloud API. | 2 min quick read |
| Ground truth | Ground truth is the best available account of what really happened, the reality that a model's predictions are checked against. | 5 min explainer |
| Grounding | Grounding gives a model relevant source material and asks it to base its response on that evidence. | 3 min explainer |
| Grouped-query attention | Grouped-query attention lets each group of query heads share one key head and one value head. | 3 min explainer |
| GRPO | Group relative policy optimization trains a policy by comparing rewards among multiple responses to the same prompt, avoiding a separate learned value model. | Definition |
| GRU | A GRU is a recurrent unit with reset and update gates. | 3 min explainer |
| GSM8K | GSM8K is a benchmark of about 8,500 grade-school maths word problems used to test whether a language model can reason through several simple steps. | 2 min quick read |
| Guardrails | Guardrails are checks placed around an AI model that inspect what goes in and what comes out, and can block, change or check inputs and replies that are unsafe, off-topic or break the app's rules. | 5 min explainer |
| Guardrails AI | Guardrails AI is an open-source Python framework that checks what goes into and comes out of an LLM, and helps get structured data back from it. | 2 min quick read |
| Guidance | Guidance is an open-source Python library for controlling language models, mixing ordinary code with generation and constraining output with rules such as regular expressions and grammars. | 2 min quick read |
| Hailuo | Hailuo is MiniMax's AI video generation product and model line, which turns text and images into short videos. | 2 min quick read |
| Hallucination | A hallucination is generated content that sounds plausible but is false, unsupported by evidence, or inconsistent with the given input. | 3 min explainer |
| Handy | Handy is a free, open-source desktop app that turns what you say into typed text, doing all of the work on your own computer with no internet connection needed. | 2 min quick read |
| Harvey | Harvey is an AI platform built for legal work, used by law firms and company legal teams to research, draft and analyse documents. | 2 min quick read |
| Haystack | An open-source Python toolkit from deepset that builds AI apps (search, RAG, agents) by linking small parts into a pipeline. | 4 min explainer |
| HBM memory | HBM (high bandwidth memory) is memory made by stacking DRAM chips vertically and placing them right beside a processor, giving AI chips very fast access to data. | 2 min quick read |
| Headroom | Headroom is an open-source tool that shrinks the logs, files and tool results an AI agent reads before they reach the model, so the agent spends fewer tokens. | 2 min quick read |
| Helicone | Helicone is an open-source AI gateway and observability platform: you route model requests through it, and it logs them while giving access to many providers through one API. | 2 min quick read |
| HellaSwag | HellaSwag is a benchmark that asks a model to pick the most sensible ending to a short everyday scene, testing commonsense reasoning. | 2 min quick read |
| HeyGen | HeyGen is an AI video platform that makes videos of talking digital presenters, including avatars of real people, and translates videos into other languages. | 2 min quick read |
| Hidden Markov models | A hidden Markov model infers a sequence of states you cannot see from a sequence of observations you can, using the odds of each state following another and of each state producing each observation. | 4 min explainer |
| HNSW | HNSW finds the stored vectors closest to a query quickly by searching a stack of linked graphs, from a sparse top layer down to a dense bottom one. | 4 min explainer |
| Hugging Face | Hugging Face runs the Hub, a website where people share AI models, datasets and small demo apps so others can find, download and reuse them. | 4 min explainer |
| Hugging Face Accelerate | Hugging Face Accelerate is an open-source library that lets ordinary PyTorch training code run on several GPUs, TPUs or machines after adding a few lines. | 2 min quick read |
| Hugging Face Datasets | Hugging Face Datasets is an open-source Python library for loading, processing and sharing datasets for text, image and audio AI, often in one line of code. | 2 min quick read |
| Hugging Face Transformers | Transformers is a free, open-source Python library that lets you download a ready-trained AI model by name and run or train it with a short piece of code. | 4 min explainer |
| Human evaluation | Human evaluation means people, not software, judge how good an AI model's answers are. | 2 min quick read |
| Human in the loop | With this setup, a person checks an AI system's work at chosen points and can approve it, change it, reject it or step in. | 5 min explainer |
| HumanEval | HumanEval measures functional correctness for programs synthesized from docstrings. | 3 min explainer |
| Humanity's Last Exam | Humanity's Last Exam, or HLE, is a benchmark of 2,500 very hard expert-written questions, built because top AI models had maxed out older tests. | 2 min quick read |
| HunyuanVideo | HunyuanVideo is an open-source video generation model from Tencent's Hunyuan team, released to give the public a video model close to the best closed ones. | 2 min quick read |
| Hy-Memory-30B-A3B | Hy-Memory-30B-A3B is an open-weight model from Tencent Hunyuan that does one job for AI agents: deciding what to write into their long-term memory. | 2 min quick read |
| Hy-Nomas | Hy-Nomas is an open video world model from Tencent: it builds a scene one short video piece at a time, and you can steer each new piece with text or camera moves. | 1 min quick read |
| Hy4 | Hy4 is an AI model from Tencent. It is the first in its Hy line that can understand images and operate a computer, in both command-line and on-screen apps. | 1 min quick read |
| Hybrid search | Hybrid search runs a keyword ranker and a vector ranker on one query, then merges the two lists into a single order. | 4 min explainer |
| HyDE | HyDE has a language model write a made-up answer to your question, then searches for real documents that look like that answer. | 4 min explainer |
| Hyperparameters | Hyperparameters are the settings people choose to control training, such as the learning rate, rather than values the model learns itself. | 5 min explainer |
| Ideogram | Ideogram is an AI tool that makes images from a text description, and it is especially good at putting readable words inside pictures. | 1 min quick read |
| Image captioning app | An image captioning app looks at a picture and writes a short sentence describing what is in it. | 2 min quick read |
| Image classification | Image classification is teaching a model to look at a picture and say which category it belongs to, such as cat or dog. | 1 min quick read |
| Image editing models | An image editing model changes a picture you already have. | 4 min explainer |
| Image generator app | An image generator app creates pictures from a written description, called a prompt, using a text-to-image model. | 2 min quick read |
| Image segmentation | Image segmentation labels an image pixel by pixel, tracing each object's precise outline instead of just drawing a box around it. | 2 min quick read |
| Imagen | Imagen is Google's earlier family of text-to-image models, first described in 2022 and now shut down in the Gemini API in favour of Nano Banana. | 2 min quick read |
| ImageNet | ImageNet is a huge collection of human-labelled photos, organised by word meanings (mostly nouns) from WordNet, that also became a well-known benchmark for image recognition AI. | 2 min quick read |
| Imbalanced data | Imbalanced data is a dataset where one class is far rarer than another, such as a few fraud cases among many ordinary card payments. | 4 min explainer |
| Imitation learning | Imitation learning teaches an AI to act by copying examples of an expert doing the task, instead of learning only from rewards. | 2 min quick read |
| In-context learning | In-context learning is when a language model picks up how to do a task from examples or instructions in the prompt, without being retrained. | 2 min quick read |
| Indirect prompt injection | Indirect prompt injection is when an attacker hides instructions in content an AI app reads later, such as an email or web page, instead of typing them in. | 4 min explainer |
| Inference | Inference is using a trained AI model on new input to get an answer, such as a label, a number or a written reply, without changing what the model learned. | 5 min explainer |
| Inference optimization | Inference optimization is a set of methods, like caching, quantization and speculative decoding, that let a trained language model answer with less work or less memory. | 3 min explainer |
| Inkling 2 Small | Inkling 2 Small is a big AI model from Thinking Machines Lab that reads text and images and answers in text. It was released on 5 November 2026. | 1 min quick read |
| Inspect AI | Inspect is an open-source Python framework, made by the UK AI Security Institute, for building and running evaluations of large language models. | 2 min quick read |
| Instruct models | An instruct model is a pretrained base version further fine-tuned on instructions and conversational data. | 3 min explainer |
| Instruction following | Instruction following is how well a model does what a written request asks, meeting each rule it sets, such as rules on content, style, format or numbers. | 3 min explainer |
| Instruction tuning | Instruction tuning fine-tunes a pretrained model on many tasks written as instructions paired with desired responses. | 3 min explainer |
| Instructor | Instructor is an open-source library that gets structured data out of language models by filling in a Pydantic model you define, with validation and automatic retries. | 2 min quick read |
| Interpretability | Interpretability is how well people can understand why a machine learning model gave the answer it did. | 2 min quick read |
| Invoice processing | AI invoice processing reads invoices, whether scanned, photographed or PDF, and pulls out key details such as supplier, dates, totals and line items as structured data. | 2 min quick read |
| Jailbreaks | A jailbreak is a prompt written to trick an AI model into ignoring its safety training and producing something it would normally refuse. | 4 min explainer |
| Jan | Jan is an open-source desktop app for chatting with AI models that run on your own computer, positioned as an open alternative to ChatGPT. | 2 min quick read |
| JAX | JAX is a numerical computing library for Python that looks like NumPy but can compile code and compute gradients automatically on GPUs and TPUs. | 2 min quick read |
| Jev | Jev is a decision model from TypeSafe AI: instead of writing text, it answers typed questions with structured values and probabilities that software can use directly. | 2 min quick read |
| JSON mode | JSON mode is an API setting that asks a language model to reply in JSON, text a program can parse (read into data), and on some providers, such as OpenAI, enforces valid syntax, without fixing which keys or types appear. | 4 min explainer |
| JSON Schema | JSON Schema is a standard way to describe what a piece of JSON data must look like, which fields it has, their types and their rules, so it can be checked automatically. | 2 min quick read |
| Jules | Jules is Google's coding agent: it works on code in your GitHub repositories in the background, fixing bugs, writing tests and building features. | 2 min quick read |
| Jupyter | Jupyter is a free, open project whose notebooks let you mix runnable code, notes, charts and results in one shareable document, used widely for data science and machine learning. | 2 min quick read |
| K-means clustering | K-means splits unlabelled data into k groups by putting each point with its nearest centre, then moving each centre to the average of its group, over and over. | 4 min explainer |
| K-nearest neighbors | k-nearest neighbours labels a new example by finding the k most similar stored examples and taking their majority vote, or their average for numbers. | 4 min explainer |
| Kaggle | Kaggle is a website for data science and machine learning where people enter competitions, share and download datasets, and run notebooks. | 2 min quick read |
| Keras | Keras is a high-level Python library for making, training and deploying deep learning models, which runs on top of JAX, TensorFlow or PyTorch. | 2 min quick read |
| Kimi K2 | Kimi K2 is an open-weight mixture-of-experts language model from Moonshot AI, with 1 trillion parameters in total, built especially for agent tasks like tool use and coding. | 2 min quick read |
| Kimi K3.1 Pro | Kimi K3.1 Pro is the flagship model behind Kimi, made by Moonshot AI: a 3-trillion-parameter-class model that reads up to 1 million tokens and is tuned for long coding jobs and knowledge work. | 1 min quick read |
| Kiro | Kiro is an AI coding agent, available as an editor, a command-line tool, a web app and a mobile app, that plans work as written specs before building it. | 2 min quick read |
| Kling | Kling is an AI video and image generator from the Kling AI team that turns text prompts or pictures into short, film-like clips. | 2 min quick read |
| Knowledge cutoff | A knowledge cutoff is the latest period a model can reliably know from its training alone. | 3 min explainer |
| Knowledge graph builder | A knowledge graph builder uses a language model to read documents, pick out the people, places and things in them, and link them into a graph you can query. | 2 min quick read |
| Knowledge graphs | A knowledge graph stores facts as labelled links between real-world things, so software can follow the links to answer questions. | 4 min explainer |
| Kokoro | Kokoro is a small open-weight text-to-speech model, with just 82 million parameters, that turns written text into natural-sounding speech quickly and cheaply. | 1 min quick read |
| Kolibri | Kolibri is an open-weight language model from Aleph Alpha for English and German, built for regulated work such as public administration and aerospace. | 2 min quick read |
| Kronos | Kronos is a family of open AI models trained on candlestick price data from markets, built to forecast prices and volatility and to generate realistic artificial price series. | 2 min quick read |
| KV cache | A KV cache lets a language model save work from earlier tokens, so each new token does not redo the same maths. | 3 min explainer |
| Labels | A label is the answer attached to a training example, such as a category, a number or a ranking, that a supervised model learns to predict. | 5 min explainer |
| LAION-5B | LAION-5B is an open dataset of about 5.85 billion image links paired with their text captions, collected from the web and used to train image models such as Stable Diffusion. | 2 min quick read |
| LanceDB | LanceDB is an open-source, embedded database for AI data that stores vectors alongside text, images, video and other media, and searches them by vector, keywords or SQL. | 1 min quick read |
| LangChain | LangChain is an open-source framework for building apps and agents on top of large language models, with one standard way to talk to many AI providers. | 5 min explainer |
| Langflow | Langflow is an open-source, Python-based tool for building AI agents and workflows in a visual editor, then serving them as an API or an MCP server. | 2 min quick read |
| Langfuse | Langfuse is an open-source platform for tracing, evaluating and managing prompts in LLM apps, which teams can use in the cloud or host themselves. | 2 min quick read |
| LangGraph | LangGraph is an open-source framework for building AI agents as graphs, where steps share one state that can be saved, paused and resumed. | 5 min explainer |
| LangSmith | LangSmith is LangChain's platform for tracing, testing and monitoring LLM apps and agents, so teams can see what happened in each run and measure quality. | 2 min quick read |
| Language learning tutor | An AI language tutor lets you practise a new language by chatting or talking with an AI character, then gives feedback on what you said. | 2 min quick read |
| Latency | Request latency is the elapsed time between sending an inference request and receiving its final response. | 3 min explainer |
| Laya | Laya is an open-source family of small encoder models from Convai Innovations that answers typed questions about a text, such as yes or no, pick one, or rate on a scale, without generating any words. | 2 min quick read |
| Layer normalization | Layer normalization normalizes selected activations independently for each example in a batch. | 3 min explainer |
| Lead qualification agent | A lead qualification agent researches new sales enquiries, sorts them by how promising they are and drafts a follow-up for a person to approve. | 2 min quick read |
| Learning rate | The learning rate multiplies a gradient, and that product is how far the weight moves. | 3 min explainer |
| Learning-rate schedules | A learning-rate schedule changes how big the training steps are as training goes on, for example starting gently, then shrinking steps towards the end. | 2 min quick read |
| Letta | Letta is an open-source platform for stateful AI agents, agents with long-term memory that they can edit themselves, which grew out of the MemGPT research project. | 2 min quick read |
| LFM (Liquid Foundation Models) | LFMs, short for Liquid Foundation Models, are Liquid AI's family of small, efficient open-weight models made to run on phones, laptops and other devices. | 2 min quick read |
| LibreChat | LibreChat is an open-source, self-hosted chat web app with a ChatGPT-like interface that works with models from many AI providers. | 2 min quick read |
| LightGBM | LightGBM is an open-source gradient boosting library that builds decision trees quickly and with little memory, for classification, ranking and other tasks. | 2 min quick read |
| LightOnOCR | LightOnOCR is LightOn's family of small, open OCR models that turn pictures of document pages into text, with version 3 also mapping where each part of the page sits. | 2 min quick read |
| Lightpanda | Lightpanda is a headless browser, written from the ground up in Zig, that AI agents and automation scripts use to load and read web pages. | 2 min quick read |
| Linear algebra for ML | The small part of linear algebra that models use every day, vectors, matrices, their products and their sizes, to hold data, run layers and learn. | 4 min explainer |
| Linear attention | Linear attention rewrites attention with feature maps so key-value summaries are formed before they are combined with queries. | 3 min explainer |
| Linear regression | Linear regression predicts a number by multiplying each feature by a learned weight, adding the results and adding a bias. | 5 min explainer |
| LiteLLM | LiteLLM is an open-source Python library and gateway that lets you call models from over 100 providers through one OpenAI-style interface. | 2 min quick read |
| Llama | Llama is Meta's family of large language models whose weights can be downloaded, so developers can run and adapt them on their own hardware. | 2 min quick read |
| Llama 4 | Llama 4 is Meta's April 2025 generation of downloadable models, which use a mixture-of-experts design and understand both text and images. | 2 min quick read |
| Llama Guard | Llama Guard is a family of open models from Meta that check the prompts going into, and the answers coming out of, an AI system for unsafe content. | 2 min quick read |
| LLaMA-Factory | LLaMA-Factory (now spelled LlamaFactory) is an open-source toolkit for fine-tuning many open language and vision models through a command line or a web interface. | 2 min quick read |
| llama.cpp | llama.cpp is an open-source program that runs large language models on your own computer. | 4 min explainer |
| llamafile | llamafile packs a language model and the software to run it into one executable file that works on most operating systems without installing anything. | 2 min quick read |
| LlamaIndex | LlamaIndex is an open-source toolkit that brings your own files and data to a language model when you ask a question, so apps and agents can answer from them. | 4 min explainer |
| LlamaParse | LlamaParse is LlamaIndex's hosted document parser that turns PDFs, scans, tables and charts into clean text, markdown or JSON for LLM apps. | 2 min quick read |
| LLM | An LLM predicts a token or sequence of tokens, sometimes many paragraphs long. | 3 min explainer |
| LLM evaluation | LLM evaluation means testing an AI feature with structured tests, so you can check how accurate and reliable it is even though its answers vary. | 5 min explainer |
| LLM gateways | An LLM gateway is one server that sits between your apps and AI model providers, adding limits, caching, fallbacks and logs to model calls. | 5 min explainer |
| LLM observability | LLM observability means recording what an AI app does on each request, such as the prompt, each step, the time taken and the tokens used. | 4 min explainer |
| LLM-as-a-judge | LLM-as-a-judge means asking a strong language model to grade answers to open-ended questions. | 3 min explainer |
| LLMOps | The day-to-day work of running an app built on a language model, from managing prompts and testing answers to watching speed and cost. | 5 min explainer |
| llms.txt | llms.txt is a proposed Markdown file a website can publish to give AI tools a short, clean guide to its content, with links to the pages that matter. | 1 min quick read |
| LM Studio | LM Studio is a desktop app for downloading open language models and running them on your own computer, with a chat window and a local API. | 2 min quick read |
| LMArena | LMArena (first called Chatbot Arena) ranks AI chat models by letting people compare two anonymous answers side by side and vote for the better one. | 2 min quick read |
| LobeChat | LobeChat is an AI chat and agent app with public code on GitHub; its repository now presents the project as LobeHub, a workspace for building and working with AI agents. | 2 min quick read |
| Local LLM chat app | A local LLM chat app runs a language model on your own computer, so you can chat with it without sending messages to a cloud service. | 2 min quick read |
| LocalAI | LocalAI is an open-source server that runs AI models on your own hardware and offers them through APIs that copy OpenAI's and Anthropic's. | 2 min quick read |
| Logistic regression | Logistic regression weighs and adds up an example's features, squeezes the total into a probability between 0 and 1, and uses a threshold to pick a class. | 4 min explainer |
| Logits | Logits are a model's raw, unnormalized prediction scores. | 3 min explainer |
| Long context | Transformer-XL reuses hidden states from previous segments instead of computing them from scratch for each new segment. | 4 min explainer |
| LoRA | LoRA adapts a large model to a new task by freezing its weights and training two thin matrices whose product is added to them. | 5 min explainer |
| Loss function | A loss function turns the difference between a model's prediction and its target into a number that training tries to reduce. | 3 min explainer |
| Lost in the middle | Language models often use facts near the opening or closing of a long input better than facts buried in the middle. | 5 min explainer |
| Lovable | Lovable is an AI tool that builds full web applications from a description written in everyday language, with real code you can edit. | 2 min quick read |
| LSTM | An LSTM is a recurrent unit whose multiplicative gates regulate access to its memory path. | 3 min explainer |
| Luma Dream Machine | Dream Machine is Luma AI's app for turning text prompts into images and short animated videos; Luma's newer app is built around AI agents. | 2 min quick read |
| Lyria | Lyria is Google DeepMind's family of music generation models, which create songs and instrumental tracks from text descriptions or images. | 2 min quick read |
| Machine learning | Machine learning trains software on data so it can find patterns and make useful predictions or generate content for new inputs. | 5 min explainer |
| Machine translation | Machine translation is software, usually a machine learning model, that turns text in one human language into another. | 1 min quick read |
| Magistral | Magistral was Mistral AI's first family of reasoning models, released in June 2025, which think step by step before answering. | 2 min quick read |
| Mamba | Mamba is a sequence architecture whose state space parameters depend on the input, letting each token control what information is kept or forgotten. | 4 min explainer |
| Manus | Manus is a general-purpose AI agent that carries out whole tasks for you on its own virtual computer, rather than only answering questions. | 2 min quick read |
| Marker | Marker is an open-source tool from Datalab that converts PDFs and other documents into Markdown, JSON or HTML, keeping tables, equations and code readable for AI apps. | 1 min quick read |
| Markov chains | A Markov chain models transitions among states where the next state's probability depends on the current state under the standard Markov assumption. | Definition |
| Markov decision process | A Markov decision process is the standard maths framework for reinforcement learning: states, actions, rewards, and the chances of moving from one state to the next. | 2 min quick read |
| Masked language modeling | Masked language modeling trains a model by hiding some words in a sentence and asking it to guess them from the words on both sides. | 2 min quick read |
| Mastra | Mastra is an open-source TypeScript toolkit for making AI agents, step-by-step workflows and the apps that use them. | 3 min explainer |
| MATH benchmark | MATH is a benchmark of 12,500 competition maths problems, each with a full worked solution, used to test harder mathematical reasoning in AI models. | 2 min quick read |
| Matrices | A matrix is a rectangular grid of numbers with a fixed number of rows and columns. Machine learning keeps data and layer weights in matrices and multiplies them. | 4 min explainer |
| Max tokens | Max tokens sets an upper bound on how many tokens a model may generate for one response. | 3 min explainer |
| MCP | MCP standardizes how AI applications connect to external context and tools. | 4 min explainer |
| MCP server | An MCP server is a small program that offers tools, data or prompt templates to AI apps through the Model Context Protocol, a shared standard for connecting them. | 2 min quick read |
| Mechanistic interpretability | Mechanistic interpretability tries to reverse-engineer a neural network, finding the internal features and circuits that produce its behaviour. | 2 min quick read |
| Meeting summarizer | A meeting summarizer records or receives a meeting's audio, transcribes it, and produces a short summary of what was decided and who needs to do what. | 2 min quick read |
| Mem0 | Mem0 is a memory layer for AI agents and assistants that pulls out important facts from conversations, stores them, and brings them back later so replies can be personal. | 1 min quick read |
| Mercury 3 | Mercury 3 is Inception's newest language model. Instead of writing its answer one token after another, it improves rough text step by step into a final reply, a method called diffusion. | 2 min quick read |
| Meta AI | Meta AI is Meta’s assistant across its apps, the web and a standalone app for conversation, search, voice and image creation. | 5 min explainer |
| Metadata filtering | Metadata filtering limits a vector search to records whose labels, such as category or year, match rules you set. | 4 min explainer |
| Microsoft 365 Copilot | Microsoft 365 Copilot is Microsoft's AI assistant for work, built into apps such as Word, Excel, PowerPoint, Outlook and Teams. | 2 min quick read |
| Microsoft Agent Framework | Microsoft Agent Framework is Microsoft's open-source framework for building AI agents and multi-agent workflows in Python and .NET, succeeding Semantic Kernel and AutoGen. | 2 min quick read |
| Microsoft Copilot | Microsoft Copilot is an AI helper from Microsoft that you can type to, talk to or show a picture, and it will reply, write something new or carry out a job for you. | 5 min explainer |
| Microsoft Copilot Studio | Microsoft Copilot Studio is a low-code tool for building AI agents and automated workflows that connect to a company's data and run in apps like Teams. | 2 min quick read |
| Microsoft Execution Containers | Microsoft Execution Containers (MXC) is Microsoft's open-source tool for running AI agents, or the code they create, in a fenced-off space that only reaches what it is allowed to. | 2 min quick read |
| Microsoft Foundry | Microsoft Foundry is Microsoft's Azure platform for building AI apps and agents, with a large model catalogue, tools, testing and safety controls in one place. | 5 min explainer |
| Microsoft-Decision-1 | Microsoft-Decision-1 is Microsoft's decision model: given a fixed set of possible answers, it gives each one a probability instead of writing text. It is offered in Microsoft Foundry and on OpenRouter. | 2 min quick read |
| Midjourney | Midjourney is an AI service that makes images, and short videos, from text descriptions called prompts. | 2 min quick read |
| Milvus | Milvus is an open-source vector database built to store and search very large numbers of embeddings, for apps such as semantic search, RAG and recommendations. | 2 min quick read |
| MiniCPM-V | MiniCPM-V is OpenBMB's series of open models that understand images and video and are built to run on phones and other local devices. | 2 min quick read |
| MiniMax | MiniMax is an AI company whose language models, such as MiniMax-Text-01 and the open-weight MiniMax-M1, are known for very long context windows of around a million tokens. | 2 min quick read |
| Ministral | Ministral is Mistral AI's family of small language models built to run on devices and local machines rather than in big data centres. | 2 min quick read |
| Missing data | Missing data are empty cells in a dataset. You can drop them, fill them with reasoned guesses or use a model that accepts gaps, and why they are missing decides which is safe. | 5 min explainer |
| Mistral Large | Mistral Large is the top tier of Mistral AI's general models; the newest, Mistral Large 4, is a multimodal model with 1 trillion parameters and open weights. | 2 min quick read |
| Mistral Medium | Mistral Medium is the middle-sized line of language models from the French company Mistral AI, aimed at businesses that want strong results at a lower price. | 2 min quick read |
| Mistral Small | Mistral Small is Mistral AI's efficient model line; the current Mistral Small 4 combines chat, reasoning, coding and image input in one open model. | 2 min quick read |
| Mistral Studio | Mistral Studio, formerly La Plateforme, is Mistral AI's developer console and API for using its models. | 2 min quick read |
| Mixed precision | Mixed precision trains with both 16-bit and 32-bit numbers so the run is faster and uses less memory. | 4 min explainer |
| Mixtral | Mixtral is a Mistral AI open-weight model from December 2023 that uses a mixture of experts, so only part of the network works on each token. | 2 min quick read |
| Mixture of experts | A mixture-of-experts layer uses a learned gate to select a sparse combination of expert subnetworks for each input. | 3 min explainer |
| MLC LLM | MLC LLM is an open-source project that compiles language models so they run efficiently on many kinds of hardware, from GPUs to phones and web browsers. | 2 min quick read |
| MLflow | MLflow is an open-source platform for keeping track of machine learning and AI work, from logging experiments to versioning models and evaluating LLM apps. | 2 min quick read |
| MLOps | MLOps is a way of working that makes building and releasing machine learning models simpler and more automatic. | 5 min explainer |
| MLX | MLX is Apple's open-source machine learning framework for Apple silicon Macs, with a NumPy-like Python API. | 2 min quick read |
| MMLU | MMLU is a 57-task test of a text model's multitask accuracy. | 3 min explainer |
| MNIST | MNIST is a classic dataset of small images of handwritten digits, 0 to 9, used to train and test image classifiers. | 1 min quick read |
| Model cards | A model card is a short document published with an AI model that explains what it is for, how it was trained and tested, and where it falls short. | 2 min quick read |
| Model collapse | Model collapse is the gradual damage that happens when AI models are trained on content made by earlier AI models instead of by people. | 2 min quick read |
| Model merging | Model merging combines the weights of several fine-tuned models into one model, without training it again, to get the strengths of each. | 2 min quick read |
| Model parallelism | Model parallelism splits one model across several chips, so a model too big for a single chip can still be trained or run. | 2 min quick read |
| Model routing | Model routing sends each prompt to the model that suits it, so simple requests go to cheaper models and hard ones to stronger models. | 4 min explainer |
| Model serving | Model serving means running a trained model on a server so apps can send it requests over the network and get answers back. | 5 min explainer |
| Monte Carlo methods | Monte Carlo methods estimate a number by running many random simulations and averaging what comes out, instead of working the answer out exactly. | 4 min explainer |
| Moshi | Moshi is a speech model from the AI lab Kyutai that listens and talks at the same time, generating speech directly for real-time spoken conversations instead of chaining separate steps. | 2 min quick read |
| Movie Gen | Movie Gen is a set of Meta research models, described in 2024, that generate HD video with matching sound from text and can edit or personalise videos. | 2 min quick read |
| Multi-agent systems | A multi-agent system is a group of AI agents, each a language model using tools in its own loop, that work together on one job. | 5 min explainer |
| Multi-agent team | A multi-agent team is a setup where several AI agents, each with its own role, split up a job and pass work between them. | 2 min quick read |
| Multi-armed bandits | A multi-armed bandit applies an action, observes its reward, then continues the process with another action. | 3 min explainer |
| Multi-head attention | Multi-head attention consists of several attention layers running in parallel. | 3 min explainer |
| Multilayer perceptron | A multilayer perceptron can fit a nonlinear model to training data. | 3 min explainer |
| Multimodal models | A multimodal model can process more than one type of input, such as text, images, audio or video. | 4 min explainer |
| Muse | Muse is Meta's personal AI agent, launched in September 2026, that you message with a goal and that then works on it for you across your apps. | 2 min quick read |
| Muse Ember | Muse Ember is Meta's smallest and most efficient Muse model, released in November 2026 for the many small, repeated steps AI agents take. | 2 min quick read |
| Muse Spark 1.1 | Muse Spark 1.1 is a Meta large language model, announced in July 2026, that powers the Meta AI assistant and is built to plan and act on tasks. | 2 min quick read |
| Muse Spark Open | Muse Spark Open is Meta's open-weight model for coding and agent work, released on 15 October 2026, so developers can download it and run it on their own hardware. | 2 min quick read |
| Muse Wonder | Muse Wonder is a Meta AI model series, announced in October 2026, made for agents that carry out personal and work tasks. It has two models: Muse Wonder and Muse Wonder Full. | 2 min quick read |
| Music generation models | A music generation model can generate music in the raw-audio domain. | 3 min explainer |
| n8n | n8n is a workflow automation tool that links apps, code and AI models into automated flows, and can run on n8n's cloud or on your own servers. | 2 min quick read |
| Naive Bayes | Naive Bayes sorts things into classes by multiplying how common each class is by how likely each feature is in that class, then picking the top score. | 4 min explainer |
| Named entity recognition | Named entity recognition (NER) finds the names of real-world things in text, such as people, companies, places and products, and labels what kind each one is. | 2 min quick read |
| Nano Banana | Nano Banana is the nickname for Gemini's built-in image models, which create and edit pictures from text, images or a mix of both. | 2 min quick read |
| Narrow AI | Narrow AI is designed for a limited task or domain, even when its performance there appears sophisticated or exceeds human ability. | Definition |
| Natural language processing | Natural language processing, or NLP, is the field of getting computers to work with human language, such as text people type or say. | 1 min quick read |
| NeMo Guardrails | NeMo Guardrails is NVIDIA's open-source toolkit for adding programmable rules, called rails, around an LLM chatbot to keep it on topic and safe. | 2 min quick read |
| Nemotron | Nemotron is NVIDIA's family of open AI models, released with their weights, training data and recipes, aimed at building specialised AI agents. | 2 min quick read |
| Nemotron 3.5 Transcribe | Nemotron 3.5 Transcribe is NVIDIA's open-weight speech model that turns recordings into text and can also translate, tell speakers apart and answer questions about the audio. | 2 min quick read |
| Neural networks | A neural network is a model built from layers of simple units that each weigh their inputs, add them up and bend the result, so that together they can learn curved, complex patterns. | 4 min explainer |
| Neural radiance fields | A neural radiance field maps a 3D position and viewing direction to density and view-dependent color. | 3 min explainer |
| Next-token prediction | Next-token prediction estimates which token should follow the tokens already present in a sequence. | 3 min explainer |
| Next.js AI chatbot | The Next.js AI chatbot is Vercel's open-source starter for building an AI chat web app with Next.js and the AI SDK. | 2 min quick read |
| NLLB | NLLB (No Language Left Behind) is a Meta translation model that translates directly between 200 languages, including many with little online text. | 2 min quick read |
| Nomic Embed | Nomic Embed is a family of open text embedding models from Nomic AI, released with its weights, training code and training data so anyone can reproduce it. | 2 min quick read |
| Notion AI | Notion AI is the AI assistant built into Notion, the notes and docs app, that finds information, writes content and carries out tasks inside your workspace. | 2 min quick read |
| Nous Research | Nous Research is an AI research group that trains and releases open models, whose models include the Hermes series, and which builds the open-source Hermes Agent. | 2 min quick read |
| NPUs | An NPU is a part of a chip built to speed up AI models at low power. | 5 min explainer |
| NVIDIA Blackwell | Blackwell is an NVIDIA GPU architecture built for training and running very large AI models, including reasoning and mixture-of-experts models. | 2 min quick read |
| NVIDIA Cosmos | NVIDIA Cosmos is NVIDIA's open set of world models, AI that simulates how scenes look and change, used to help robots and self-driving cars learn. | 1 min quick read |
| NVIDIA CUDA | CUDA is NVIDIA's software platform that lets programmers use NVIDIA GPUs for general computing, including training and running AI models. | 2 min quick read |
| NVIDIA DGX Spark | NVIDIA DGX Spark is a small desktop computer from NVIDIA for building and running AI models and agents on your own desk instead of in the cloud. | 2 min quick read |
| NVIDIA H100 | The NVIDIA H100 is a data centre GPU, built on NVIDIA's Hopper architecture, designed for large-scale AI and high-performance computing. | 2 min quick read |
| NVIDIA NIM | NVIDIA NIM packages AI models as ready-to-run containers, called microservices, that serve the model through a standard API on NVIDIA GPUs. | 2 min quick read |
| NVIDIA RTX Spark | NVIDIA RTX Spark is a chip for Windows laptops and small desktops that joins an NVIDIA graphics processor and main processor, so AI models and agents can run on the PC itself. | 2 min quick read |
| NVIDIA Run:ai | NVIDIA Run:ai is open-source software that decides how AI jobs share graphics chips (GPUs) across Kubernetes clusters, the systems that run programs over many computers. | 2 min quick read |
| o3 | o3 is an OpenAI reasoning model from April 2025 that thinks for longer before it answers and can decide when to use tools. | 2 min quick read |
| Object detection | Object detection is the computer vision task of finding each object in an image, drawing a box around it and naming what it is. | 2 min quick read |
| OCR | OCR (optical character recognition) is technology that reads the text in images and scans and turns it into text a computer can search and edit. | 2 min quick read |
| Ollama | Ollama is an open-source program that downloads open AI models and runs them on your own computer, with a simple command and a local API. | 4 min explainer |
| OLMo | OLMo is a family of fully open language models from Ai2 (the Allen Institute for AI), released with its training data, code and checkpoints, not just its weights. | 2 min quick read |
| OlmoEarth | OlmoEarth is Ai2's family of open models for satellite images; version 1.3 turns a year of images of each spot on Earth into a short list of numbers, called an embedding. | 1 min quick read |
| Omni models | An omni model processes several input and output modalities within one multimodal model rather than routing everything through a text-only core. | 4 min explainer |
| On-device AI | On-device AI runs the model on your own phone, laptop or browser, so your data does not have to go to a server. | 4 min explainer |
| On-device assistant | An on-device assistant runs its AI model on your own phone, laptop or browser instead of a cloud server, so it can work offline and keep your data local. | 2 min quick read |
| One-hot encoding | One-hot encoding turns a category, such as a colour or a type of transport, into a list of 0s with a single 1 marking which category it is. | 4 min explainer |
| ONNX | ONNX (Open Neural Network Exchange) is an open file format for machine learning models, so a model trained in one framework can run in many other tools and runtimes. | 2 min quick read |
| ONNX Runtime | ONNX Runtime is an open-source engine that runs machine learning models saved in the ONNX format quickly on many kinds of hardware and operating systems. | 2 min quick read |
| Open WebUI | Open WebUI, a free chat app you host yourself, lets you talk to AI models running on your own computer or in the cloud, from one web page. | 4 min explainer |
| Open-weight models | An open-weight model makes its trained weights publicly available for download. | 3 min explainer |
| OpenAI Agents SDK | OpenAI's Agents SDK is a small open-source toolkit from OpenAI for building AI agents that use tools, pass work to each other and can record each run as a trace. | 5 min explainer |
| OpenAI API | The OpenAI API lets developers send requests from their own code to OpenAI's models and get back text, images, audio or structured data. | 2 min quick read |
| OpenAI Codex | Codex is OpenAI's coding agent: an AI that helps people write, review and ship software, working from a terminal, a code editor or the cloud. | 2 min quick read |
| OpenAI Decisions API | The Decisions API is an OpenAI API that answers fixed-form questions about text, images or JSON, such as whether something is true, which category fits or how severe an issue is. | 1 min quick read |
| OpenAI embeddings | OpenAI embeddings are models that turn text into lists of numbers, so a program can measure how closely two pieces of text are related. | 2 min quick read |
| OpenAI-compatible APIs | An OpenAI-compatible API accepts the same request format as OpenAI's API, so apps written for OpenAI can talk to other models by changing a few settings. | 2 min quick read |
| OpenAPI | OpenAPI is a standard way to describe a web API in one document, listing its endpoints, inputs and outputs, so people and programs can understand and call it. | 2 min quick read |
| OpenCode | OpenCode is an open-source AI coding agent: an assistant that writes and edits code and runs commands for you, from a terminal, a desktop app or a web app. | 2 min quick read |
| OpenCV | OpenCV is a free, open-source library of computer vision and machine learning tools for working with images and video. | 2 min quick read |
| OpenRouter | OpenRouter is a service that gives developers one API for many AI models from different companies, with automatic fallback when a model or provider fails. | 2 min quick read |
| OpenSearch | OpenSearch is an open-source search and analytics suite, licensed under Apache 2.0, that can also store embeddings and run vector, semantic and hybrid search for AI apps. | 2 min quick read |
| OpenShell | OpenShell is NVIDIA's open-source runtime that keeps each AI agent inside a locked-down sandbox, so it can use files, tools and keys without free rein over your data or network. | 1 min quick read |
| OpenVINO | OpenVINO is an open-source set of tools for converting, shrinking and running trained AI models efficiently, built to get the most from Intel CPUs, GPUs and NPUs. | 2 min quick read |
| Opik | Opik is an open-source platform for tracing, testing and monitoring apps built on language models, including RAG systems and AI agents. | 2 min quick read |
| Optuna | Optuna is an open-source Python framework that automatically searches for good hyperparameters, the settings you choose before training a machine learning model. | 2 min quick read |
| Orca | Orca is an open-source desktop app for running several AI coding agents at once, such as Claude Code and Codex, each on its own copy of your project. | 2 min quick read |
| Orkas | Orkas is a free, open-source desktop app for macOS and Windows in which a lead AI agent, the Commander, splits a goal into tasks and shares them out to a team of specialist agents. | 2 min quick read |
| Otter.ai | Otter.ai is an AI tool that joins your meetings, records and transcribes them as they happen, then writes summaries you can search and ask questions about. | 1 min quick read |
| Outliers | An outlier is a data point that sits an unusually long way from the rest. It may be a mistake to fix or a rare real event worth keeping. | 4 min explainer |
| Outlines | Outlines is an open-source Python library from .txt that makes a language model's output follow a set structure, such as a JSON schema or a regular expression, while it is being generated. | 2 min quick read |
| Overfitting | Overfitting happens when a model learns its training examples so specifically that it performs poorly on new examples. | 3 min explainer |
| PagedAttention | PagedAttention is a way of storing a language model's memory of past tokens in small fixed-size blocks, so a server wastes less memory and can handle more requests at once. | 2 min quick read |
| PageIndex | PageIndex is an open-source retrieval engine that answers questions about long documents by having a language model reason through a tree of their sections, with no vector database. | 2 min quick read |
| Paperclip | Paperclip is an open-source app for running a team of AI agents like a small company: you give them goals, roles and budgets, then follow their work and spending in one place. | 2 min quick read |
| Parameter-efficient fine-tuning | PEFT adapts a pretrained model by training a small set of added or selected parameters while keeping most base weights frozen. | 4 min explainer |
| Parameters | Parameters are the model values that training can change to improve how inputs map to outputs. | 3 min explainer |
| PEFT | PEFT is a free Hugging Face library that teaches a big AI model a new task by training a small set of extra weights instead of the whole model. | 4 min explainer |
| Perceptron | A perceptron is a linear classifier. | 3 min explainer |
| Perplexity | Perplexity is an AI search product that searches sources, synthesizes an answer and links the evidence used. | 5 min explainer |
| Perplexity (metric) | Perplexity is the exponentiated average negative log-likelihood that a language model assigns to a token sequence. | 4 min explainer |
| Perplexity Comet | Perplexity Comet is a web browser from Perplexity with a built-in AI assistant that can answer questions about pages and carry out browsing tasks for you. | 1 min quick read |
| Personal knowledge base | An AI personal knowledge base is a private assistant that searches and answers questions over your own notes and files. | 2 min quick read |
| Personalization | Personalization is using AI to tailor what each person sees, such as videos, products or news, based on what they and similar people have liked before. | 2 min quick read |
| pgvector | pgvector is an open-source extension that adds vector similarity search to PostgreSQL, so you can keep embeddings in the same database as the rest of your data. | 1 min quick read |
| Phi | Phi is Microsoft's family of small language models, designed to give strong results while being cheap and light enough to run on phones, laptops and edge devices. | 2 min quick read |
| Phi-4 | Phi-4 is a 14-billion-parameter open language model from Microsoft Research that reaches strong reasoning scores for its size by training heavily on high-quality synthetic data. | 2 min quick read |
| Pinecone | Pinecone is a vector database service for storing embeddings and finding the most similar items fast, often used to give AI apps the right documents. | 2 min quick read |
| Pipeline parallelism | Pipeline parallelism splits a model's layers into stages on different GPUs and streams small slices of each batch through them like an assembly line. | 2 min quick read |
| Pixtral | Pixtral was Mistral AI's line of models that understand images as well as text; Mistral has since retired it in favour of newer multimodal models. | 2 min quick read |
| Podcast transcriber | A podcast transcriber turns an episode's audio into a searchable transcript, often with timestamps and labels showing who is speaking. | 2 min quick read |
| Positional encoding | Positional encoding adds token-order information to representations before attention reads them. | 3 min explainer |
| PPO | Proximal policy optimization is a reinforcement-learning algorithm that limits the size of policy updates to improve training stability. | Definition |
| Praxis-1 | Praxis-1 is Runway's open-weight robot model: it predicts video and the movements a robot should make, both at once. | 2 min quick read |
| Precision and recall | Precision asks how many of the things a model flagged were right. Recall asks how many of the real cases it managed to find. | 5 min explainer |
| Predictive maintenance | Predictive maintenance uses sensor data and machine learning to predict when a machine is likely to fail, so it can be fixed first. | 2 min quick read |
| Prefill and decode | Prefill processes all prompt tokens in parallel to initialize model state; decode then generates new tokens sequentially using that cached state. | Definition |
| Pretraining | In one transfer-learning setup, a model is pretrained on a data-rich task before fine-tuning on a downstream task. | 3 min explainer |
| Prime Agent | Prime Agent is Prime Intellect's open-source AI agent for coding, research and other jobs that take a long time. | 2 min quick read |
| Principal component analysis | PCA finds the directions in which data spreads out most, then keeps only the first few, so many features become a few new ones. | 4 min explainer |
| Probability | A probability is a number from 0 to 1 that says how likely something is. Many machine learning problems need one as the answer. | 4 min explainer |
| Prompt caching | Prompt caching saves the work a model did on the unchanged start of a prompt, so the next request with the same start is faster and cheaper. | 4 min explainer |
| Prompt chaining | In prompt chaining, you give a model one small job at a time in a set order, and each answer becomes the starting material for the next job. | 5 min explainer |
| Prompt engineering | Prompt engineering is writing and testing the instructions and examples you give a model so its answer meets a goal you can check. | 4 min explainer |
| Prompt injection | Prompt injection is text that sneaks new instructions into what an AI model reads, so the model behaves in unintended ways. | 5 min explainer |
| Prompt playground | A prompt playground is a web page or app where you can try prompts on a model, change its settings and see the results straight away, before writing any code. | 2 min quick read |
| Prompt templates | Prompt templates are reusable prompt structures with variable fields, helping applications apply consistent instructions and context across many requests. | Definition |
| Prompt versioning | Prompt versioning means saving every change to an app's prompts as a numbered version, so teams can compare, test and roll back prompts like code. | 2 min quick read |
| promptfoo | promptfoo is an open-source command-line tool for testing prompts, models, agents and RAG apps, and for red-teaming them to find security weaknesses. | 2 min quick read |
| Protein structure prediction | Protein structure prediction is working out the 3D shape a protein folds into from its sequence of amino acids, a problem AI models such as AlphaFold largely cracked. | 2 min quick read |
| Pruning | Pruning removes weights, units, or connections judged less important, shrinking computation or storage while aiming to preserve model quality. | Definition |
| Pydantic AI | Pydantic AI is a Python framework for building AI agents whose tool inputs and final answers are checked against types you define. | 4 min explainer |
| PyTorch | PyTorch is a free Python library for building and training neural networks, with fast maths on GPUs and automatic gradients. | 4 min explainer |
| PyTorch Geometric | PyTorch Geometric (PyG) is an open-source library built on PyTorch for writing and training graph neural networks, which learn from data made of connected points. | 2 min quick read |
| PyTorch Lightning | PyTorch Lightning is an open-source library that organises PyTorch training code, handling the repetitive engineering so the same model can scale from a laptop to many GPUs. | 2 min quick read |
| Q-learning | Q-learning improves an estimate of each state-action value from sampled rewards and the best estimated value at the next state. | 3 min explainer |
| Qdrant | Qdrant is an open-source vector database and search engine, written in Rust, that stores embeddings with extra data and finds the closest matches with strong filtering. | 1 min quick read |
| QLoRA | QLoRA fine-tunes a large language model by storing its frozen weights in 4 bits and training only a small add-on that sits beside them. | 5 min explainer |
| Qualcomm Snapdragon X | Qualcomm Snapdragon X is a family of Arm-based laptop processors with a built-in NPU for AI, used in Windows Copilot+ PCs. | 2 min quick read |
| Quantization | Quantization stores a model's numbers with fewer bits, such as 8-bit integers instead of 32-bit decimals, so it needs less memory. | 4 min explainer |
| Query rewriting | Query rewriting transforms a user’s request into one or more clearer search queries, improving retrieval when the original wording is vague or conversational. | Definition |
| Question answering | Question answering is the AI task of giving a direct answer to a question in plain language, either by finding it in a text or by generating it. | 2 min quick read |
| Qwen | Qwen is the family of large language models built by Alibaba Cloud, with many sizes and versions released with open weights. | 2 min quick read |
| Qwen-Image | Qwen-Image is the Qwen team's family of open image models that both create pictures from a text prompt and edit existing ones. | 2 min quick read |
| Qwen3 | Qwen3 is the 2025 generation of Alibaba's Qwen language models, whose models can either reason step by step or reply quickly, in sizes from 0.6 to 235 billion parameters. | 2 min quick read |
| Qwen3-Coder | Qwen3-Coder is the Qwen team's family of open coding models, built to act as agents that read, write and test code across whole projects. | 2 min quick read |
| RAG | RAG lets an AI model look up relevant documents first, then answer from what it found. | 4 min explainer |
| RAG chatbot | A RAG chatbot is a starter project where a chatbot answers questions from your own documents by first searching them, then giving the best passages to a language model. | 2 min quick read |
| Ragas | Ragas is an open-source Python library for evaluating LLM applications, best known for metrics that score RAG systems without needing human-written reference answers. | 2 min quick read |
| RAGFlow | RAGFlow is an open-source retrieval-augmented generation (RAG) engine that turns messy documents into a searchable knowledge base for chat and agents. | 2 min quick read |
| Random forest | A random forest combines many decision trees, each trained with some randomness, and uses their joint answer as the prediction. | 2 min quick read |
| Rate limits | A rate limit caps how much you can use an AI service in a set stretch of time, such as how many requests you send each minute. | 4 min explainer |
| Ray | Ray is an open-source framework for spreading Python and AI work across many machines, so the same code can grow from one laptop to a big cluster. | 2 min quick read |
| Ray-Ban Meta glasses | Ray-Ban Meta glasses are smart glasses made by Meta with EssilorLuxottica that have a camera, open-ear speakers and the Meta AI voice assistant built in. | 2 min quick read |
| ReAct | ReAct is a prompting method in which a language model alternates written thoughts with tool actions, reading each result before it picks the next step. | 5 min explainer |
| Reasoning models | Reasoning models use intermediate reasoning tokens before producing a final response. | 4 min explainer |
| Recommendation engine | A recommendation engine suggests items, such as videos, products or apps, that a person is likely to want, based on what they and similar people have liked. | 2 min quick read |
| Recommender systems | A recommender system picks the few items, out of a catalogue too big to browse, that each person sees first, by predicting what that person is likely to want. | 4 min explainer |
| Recurrent neural networks | At each time step, a recurrent neural network updates its hidden state from the previous hidden state and the current input. | 3 min explainer |
| Recursive Flint | Flint is Recursive's open-weight coding model, a mixture-of-experts design built to run coding agents quickly on a laptop rather than in the cloud. | 1 min quick read |
| Red teaming | Red teaming means deliberately testing an AI system the way an attacker would, to find its weak spots. | 4 min explainer |
| Reflection Beam | Reflection Beam is Reflection's first open-weight model, a 501-billion-parameter mixture-of-experts model aimed at coding, reasoning and agent work. | 2 min quick read |
| Regression | Regression predicts a number, such as a house price, a tree's life span or a rainfall total, from an example's features. | 3 min explainer |
| Regularization | Regularization is any change to training that limits how closely a model can fit its training examples, so it does better on new data. | 5 min explainer |
| Reinforcement learning | Reinforcement learning trains a program, called an agent, to choose actions by trial and error, guided by a number called a reward that says how well it is doing. | 5 min explainer |
| Reinforcement learning with verifiable rewards | RLVR, short for reinforcement learning with verifiable rewards, trains a language model by rewarding answers that a program can check as correct, such as a maths result. | 2 min quick read |
| ReLU | ReLU is an activation function that replaces every negative input with zero and leaves every positive input unchanged. | 3 min explainer |
| Replicate | Replicate is a cloud service that runs machine learning models behind a simple API, so you can use them without managing your own servers. | 2 min quick read |
| Replit Agent | Replit Agent is an AI agent inside Replit that turns a plain-language description into a working app, and can also make designs and slides. | 2 min quick read |
| Reranker models | A reranker sorts existing text candidates by semantic relevance to a specified query. | 3 min explainer |
| Reranking | Reranking takes the short list a first search returned and scores each item again with a slower, more careful model. | 4 min explainer |
| Research agent | A research agent is an AI system that plans a research task, searches many sources on its own and writes a report with citations. | 2 min quick read |
| Residual connections | A residual connection combines a residual mapping F(x) with the block input x as F(x) + x. | 4 min explainer |
| Residual networks | A residual block adds a learned residual function F(x) to a shortcut carrying the block input x. | 3 min explainer |
| Resume screener | A resume screener is an AI tool that reads job applications, pulls out details such as skills and experience, and helps rank or sort candidates for a role. | 2 min quick read |
| Reward hacking | Reward hacking is when an AI trained by rewards finds a loophole that earns a high score without actually doing the task people wanted. | 2 min quick read |
| Reward models | Reward models assign scores to candidate behaviors or outputs, approximating preferences or objectives that reinforcement learning can optimize. | Definition |
| Ripple-65B | Ripple-65B Preview is an open-weight language model from Ricursive Intelligence, aimed at agentic coding (AI that plans, writes and runs code over many steps) and small enough for consumer hardware such as NVIDIA's DGX Spark. | 2 min quick read |
| RLAIF | RLAIF runs the usual RLHF training loop, but another AI model supplies some of the preference labels that people would normally give. | 4 min explainer |
| RLHF | RLHF uses human comparisons to learn a reward signal, then optimizes a model to produce responses that score better under that signal. | 4 min explainer |
| RoBERTa | RoBERTa is a 2019 language model from Facebook AI that showed BERT had been undertrained, and got better results by training the same design longer on more data. | 2 min quick read |
| Robotics | AI in robotics gives machines the ability to see, plan and act in the physical world, increasingly using models that link vision and language to movement. | 2 min quick read |
| Robotics foundation models | A robotics foundation model is pretrained across many tasks, environments or robot bodies so it can serve as a starting point for new robot policies. | 4 min explainer |
| ROC curve | An ROC curve shows, for every threshold a classifier could use, how many real positives it catches against how many false alarms it raises. | 4 min explainer |
| Rotary position embeddings | RoPE encodes absolute position with a rotation matrix and adds explicit relative-position dependence to self-attention. | 3 min explainer |
| Runway | Runway is an AI company that offers an online creative suite for generating and editing video, images and audio, plus an API for developers. | 1 min quick read |
| RWKV | RWKV is a language model design that trains in parallel like a transformer but runs like a recurrent network, so its memory use stays constant however long the text gets. | 2 min quick read |
| Safetensors | Safetensors is a file format from Hugging Face for storing model weights safely and loading them fast, as a replacement for Python's pickle format. | 2 min quick read |
| Sales email writer | An AI sales email writer drafts personalised outreach or follow-up emails from details about a prospect, for a salesperson to review before sending. | 2 min quick read |
| Salesforce Agentforce | Agentforce is Salesforce's platform for building AI agents that work with a company's customer data to handle tasks such as support questions and sales leads. | 2 min quick read |
| SAM 2 | SAM 2 is Meta's 2024 follow-up to Segment Anything that outlines and tracks objects through videos as well as in still images. | 2 min quick read |
| Sandboxing agents | Sandboxing an agent means running its commands inside a fence, set up by the operating system, that limits which files and websites they can reach. | 4 min explainer |
| Sarvam Prism | Sarvam Prism is an open-weight decision model from Sarvam AI: you give it a question and a set of options, and it gives back a probability for every option. | 2 min quick read |
| Scalable oversight | Scalable oversight is the problem of how people can reliably supervise AI systems that may become better than them at the task being checked. | 2 min quick read |
| Scaling laws | Scaling laws are empirical relationships describing how model performance changes predictably with factors such as parameter count, training data, and computation. | Definition |
| scikit-learn | scikit-learn is an open-source Python library for classic machine learning, such as classification, regression and clustering, with one consistent way to train models. | 2 min quick read |
| Seedance | Seedance is ByteDance Seed's family of video generation models, which create short videos from text or images, including stories told over several shots. | 2 min quick read |
| Seedream | Seedream is ByteDance Seed's family of image generation models, which create and edit pictures from text in English and Chinese. | 2 min quick read |
| Segment Anything | Segment Anything (SAM) is a Meta model from 2023 that outlines any object in an image when you click on it or draw a box around it. | 2 min quick read |
| Self-attention | Self-attention updates each position by comparing it with positions in the same sequence and mixing their information. | 3 min explainer |
| Self-consistency | Self-consistency generates several independent reasoning paths for the same problem and selects the answer that appears most consistently across them. | Definition |
| Self-hosting LLMs | Self-hosting means running a language model on a computer you control, so your prompts go to your own machine instead of a company's service. | 4 min explainer |
| Self-reflection | Self-reflection is when an AI model checks its own first answer, writes feedback on it, and then tries again using that feedback. | 3 min explainer |
| Self-supervised learning | Self-supervised learning trains a model on raw data by hiding part of each example and asking the model to predict it, so the data supplies its own labels. | 5 min explainer |
| Semantic caching | Semantic caching saves a model's past answers and reuses one when a new prompt means the same thing, even if it is worded differently. | 2 min quick read |
| Semantic Kernel | Semantic Kernel is an open-source Microsoft toolkit that connects AI models to your own code, so a model can ask your functions to do real work. | 4 min explainer |
| Semantic search | Semantic search finds text by meaning, so a passage can match a question even when the two share no words. | 3 min explainer |
| Semantic search engine | A semantic search engine finds results by meaning rather than exact words, so a query can match text that uses different wording. | 2 min quick read |
| Semi-supervised learning | Semi-supervised learning trains a model on a few labelled examples plus many unlabelled ones, so the unlabelled data can help when answers are scarce. | 5 min explainer |
| Sentence Transformers | Sentence Transformers is an open-source Python library for turning text, and other media, into embeddings: lists of numbers where similar meanings end up close together. | 2 min quick read |
| Sentence-BERT | Sentence-BERT (SBERT) is a version of BERT trained to turn whole sentences into embeddings, so similar sentences can be found by comparing vectors instead of rerunning the model on every pair. | 2 min quick read |
| Sentiment analysis | Sentiment analysis uses software to work out whether a piece of text sounds positive, negative or somewhere in between. | 1 min quick read |
| Sentiment dashboard | A sentiment dashboard collects text such as reviews or social posts, scores each one as positive, negative or neutral, and charts how opinion changes. | 2 min quick read |
| SEO content assistant | An SEO content assistant is an AI tool that helps plan, draft and improve web pages so they are useful to readers and easier to find in search engines. | 2 min quick read |
| SGLang | SGLang is open-source server software that runs large language models on GPUs and answers many requests quickly, partly by reusing work shared between prompts. | 4 min explainer |
| SigLIP | SigLIP is a Google image-text model, similar to CLIP, trained with a simpler sigmoid loss so that it learns well and efficiently when matching pictures to captions. | 2 min quick read |
| Siri | Siri is Apple's voice assistant; its newest version, Siri AI, is a conversational assistant powered by Apple Intelligence that you can talk or type to. | 1 min quick read |
| SkillSpector | SkillSpector is NVIDIA's open-source scanner that checks skills for AI agents, such as those for Claude Code or Codex, for security risks before you install them. | 1 min quick read |
| Slack bot | An AI Slack bot is an app inside a Slack workspace that answers questions, runs tasks and replies in threads, using a language model behind the scenes. | 2 min quick read |
| Small language models | Small-model research includes sub-billion-parameter language models for mobile deployment. | 3 min explainer |
| Smolagents | Smolagents is a small open-source Python library from Hugging Face for building AI agents, including ones that act by writing short pieces of Python code. | 4 min explainer |
| Social media assistant | A social media assistant is an AI tool that helps come up with post ideas, write and rewrite posts, and adapt them to each platform's style. | 2 min quick read |
| Softmax | Softmax determines a probability for each possible class, and those probabilities add up to exactly 1. | 3 min explainer |
| SpaceXSI API | The SpaceXSI API, formerly the SpaceXAI API and before that the xAI API, lets developers use Grok models from their own code for text, images, video and voice. | 2 min quick read |
| spaCy | spaCy is an open-source Python library for natural language processing, built to pull structure such as words, grammar and names out of large amounts of text. | 2 min quick read |
| Sparse attention | Sparse attention computes selected query-key connections instead of filling the entire attention matrix. | 3 min explainer |
| Sparsity | Sparsity means that many possible model values or computation paths are zero, absent, or inactive for a given input. | 4 min explainer |
| Spec-driven development | Spec-driven development means writing down what to build and why, as a specification, before an AI coding agent writes any code, then building against that spec. | 2 min quick read |
| Speculative decoding | Speculative decoding uses a target model to evaluate guesses in parallel. | 4 min explainer |
| Speech-to-text app | A speech-to-text app turns recorded or live speech into written text using a speech recognition model. | 2 min quick read |
| Speech-to-text models | A speech-to-text model can transcribe a speech utterance into written characters. | 3 min explainer |
| Spreadsheet assistant | A spreadsheet assistant is an AI helper inside or alongside a spreadsheet that builds formulas, charts and summaries from plain-language requests. | 2 min quick read |
| Spring AI | Spring AI is an open-source framework that brings language models, embeddings and vector stores into Java apps built with Spring. | 2 min quick read |
| SQuAD | SQuAD, the Stanford Question Answering Dataset, tests whether a model can read a Wikipedia passage and answer questions about it. | 2 min quick read |
| Stable Audio | Stable Audio is Stability AI's family of models that generate music and sound effects from text descriptions, including open-weight versions. | 2 min quick read |
| Stable Diffusion | Stable Diffusion is an open text-to-image model that creates pictures from written prompts, and was light enough to run on a single graphics card with at least 10 GB of memory. | 2 min quick read |
| Stable Diffusion 3 | Stable Diffusion 3 is Stability AI's third major text-to-image model family, built on a new transformer design that is much better at following prompts and spelling text. | 2 min quick read |
| Stagehand | Stagehand is an open-source SDK from Browserbase for building AI agents that control a web browser using plain-English instructions plus normal code. | 2 min quick read |
| StarCoder | StarCoder is an open code-generation model from the BigCode community, trained on permissively licensed code from GitHub in more than 80 programming languages. | 2 min quick read |
| State space models | A continuous-time linear state space model uses x-dot = Ax + Bu for state evolution and y = Cx + Du for readout. | 4 min explainer |
| Statistics for ML | Statistics describes data and measures how far to trust a model's score, since every test set is only a sample of the cases the model will meet. | 4 min explainer |
| Stochastic gradient descent | Stochastic gradient descent steps the weights using a slope estimated from one randomly picked example. | 3 min explainer |
| Stop sequences | A stop sequence is a configured character string that stops output generation. | 3 min explainer |
| Strands Agents | Strands Agents is a free AWS toolkit, released as open source under Apache 2.0, for making AI agents with Python or TypeScript, where the model itself plans the steps and picks the tools. | 5 min explainer |
| Strands Decider | Strands Decider is an open-source decision model for AI agents: instead of writing text, it chooses among options or places something on a scale, and says how confident it is in each answer. | 2 min quick read |
| Streaming responses | Streaming responses deliver a model’s output incrementally as it is generated, improving perceived speed and enabling responsive conversational interfaces. | Definition |
| Streamlit | Streamlit is a free, open-source Python framework that turns ordinary Python scripts into interactive web apps, such as dashboards and chat apps, without web development. | 1 min quick read |
| Streamlit chatbot | A Streamlit chatbot is a chat web app written entirely in Python using Streamlit's built-in chat elements. | 2 min quick read |
| Structured data extraction | Structured data extraction uses a language model to pull specific fields out of messy text and return them in a fixed format, such as JSON, that software can use. | 2 min quick read |
| Structured outputs | Structured outputs constrain a model response to a declared schema so software receives expected fields and types instead of free-form prose. | 3 min explainer |
| Study tutor | An AI study tutor is a chatbot set up to help you learn a topic by asking guiding questions and checking your understanding, rather than just handing over answers. | 2 min quick read |
| Sub-agents | A sub-agent is a helper AI that a main agent sends off to do one side task in its own workspace, then report back a short result. | 3 min explainer |
| Summarization | Summarization is when a model turns a long text into a shorter version that keeps the important information. | 3 min explainer |
| Suno | Suno is an AI music app that creates whole songs from a short text description or from lyrics you write yourself. | 1 min quick read |
| Superintelligence | Superintelligence is a hypothetical form of intelligence that substantially exceeds the best human abilities across most cognitively demanding fields. | Definition |
| Superpowers | Superpowers is an open-source set of skills that gives coding agents such as Claude Code and Codex a fixed way of working: brainstorm, plan, then build with tests. | 2 min quick read |
| Supervised fine-tuning | Supervised fine-tuning keeps training an already-trained model on example questions paired with good answers, so it learns to answer that way. | 4 min explainer |
| Supervised learning | Supervised learning trains a model on examples that already carry the right answer, called a label, so it can predict that answer for new examples that do not. | 4 min explainer |
| Support vector machines | A support vector machine separates two classes with the boundary that leaves the widest possible gap to the nearest training points of each class. | 5 min explainer |
| SWE-bench | SWE-bench asks a system to generate a repository patch for a real GitHub issue. | 3 min explainer |
| Sycophancy | Sycophancy is when an assistant favors matching a user's beliefs over a truthful response. | 3 min explainer |
| Symbolic AI | Symbolic AI writes knowledge down as readable symbols and if-then rules, then reaches answers by applying logic and search to them. | 4 min explainer |
| Synthesia | Synthesia is an AI video platform that lets businesses make presenter-style videos with digital avatars, without cameras, actors or a studio. | 1 min quick read |
| Synthetic data | Synthetic data is artificial data built from seed data so some patterns of that seed can be reused. | 4 min explainer |
| Synthetic data generator | A synthetic data generator uses a language model to create new training or test examples, such as questions and answers, instead of collecting them all from people. | 2 min quick read |
| SynthID | SynthID is Google DeepMind's tool for putting a hidden watermark into AI-made images, video, audio and text, and for finding that mark later. | 2 min quick read |
| System prompt | A system prompt is high-priority context that sets an AI assistant's role, boundaries, style, and response rules before the user's request is handled. | 3 min explainer |
| t-SNE | t-SNE is a nonlinear visualization method that places similar high-dimensional examples nearby in two or three dimensions, but global distances may be misleading. | Definition |
| T5 | T5 is a 2019 Google model that treats every language task as turning one piece of text into another, from translation to summarising. | 2 min quick read |
| Tavus Griffin | Griffin is a video model from Tavus that holds live, face-to-face conversations as a human-like person on screen. Tavus calls it a Human Interaction Model. | 2 min quick read |
| Temperature | Temperature enters softmax by dividing each logit by T. | 3 min explainer |
| Tensor parallelism | Tensor parallelism splits the big weight matrices inside each layer of a model across several GPUs, so a layer too large for one chip can still run. | 2 min quick read |
| TensorFlow | TensorFlow is Google's open-source platform for building, training and deploying machine learning models, from phones to large clusters of machines. | 2 min quick read |
| TensorRT-LLM | TensorRT-LLM is NVIDIA's open-source library for running large language models quickly and efficiently on NVIDIA GPUs. | 2 min quick read |
| Test set | A test set is a reserved group of labeled examples used for a final check of a model on data it did not train on. | 3 min explainer |
| Test-time compute | Test-time compute lets a language model improve its output by using more computation at test time. | 3 min explainer |
| Text classification | Text classification is sorting pieces of text into set categories, such as spam or not spam, positive or negative, or which topic a message is about. | 2 min quick read |
| Text Generation Inference | Text Generation Inference (TGI) is Hugging Face's open-source server for running language models in production; it is now in maintenance mode. | 2 min quick read |
| Text-to-image models | A text-to-image model can use text as a conditioning input to an image generator. | 3 min explainer |
| Text-to-speech models | A text-to-speech model synthesizes speech directly from text. | 3 min explainer |
| Text-to-SQL | Text-to-SQL turns a question written in everyday language into a SQL query that a database can run, so people can ask for data without writing code. | 2 min quick read |
| Text-to-video models | Given a text prompt, a text-to-video model generates a video. | 3 min explainer |
| The Stack | The Stack is an open dataset of source code from the BigCode project, built for training language models that write code. | 2 min quick read |
| ThinkingBox | ThinkingBox is an open-source test bed from Microsoft that grades AI agents on what they actually change in a system, not on what they say they did. | 2 min quick read |
| Throughput | Throughput is the amount of completed inference work divided by the time spent measuring it. | 3 min explainer |
| Time series forecasting | Time series forecasting predicts future values of something measured over time, such as sales or electricity demand, from the patterns in its past. | 5 min explainer |
| Time to first token | Time to first token (TTFT) is how long you wait after sending a prompt before the first piece of the AI model's answer arrives. | 1 min quick read |
| Timestral | Timestral is Mistral AI's first set of foundation models for time series, released on 15 October 2026, which forecast future values of data recorded over time. | 2 min quick read |
| timm | timm (PyTorch Image Models) is an open-source library of ready-made computer vision models, most with pretrained weights, plus the tools to train and test them. | 1 min quick read |
| Together AI | Together AI is a cloud platform for running, fine-tuning and training open-source AI models, reached through an API that works like OpenAI's. | 2 min quick read |
| Token costs | AI model APIs charge by the token, with separate prices for the text you send in and the text the model writes back. | 4 min explainer |
| Tokenization | Tokenization turns text into a sequence of vocabulary IDs that a model can process, then reverses generated IDs back into readable text. | 3 min explainer |
| Tokens | Tokens are chunks of text processed by text-generation and embedding models. | 3 min explainer |
| Tokens per second | Tokens per second reports how many tokens a server returns in a set amount of time. | 3 min explainer |
| Tool use | Tool use is a loop where a model proposes a call, a program outside the model runs it, and the result is written back. | 5 min explainer |
| Top-k sampling | Top-k sampling keeps the k highest-probability next-token candidates, rescales their probabilities and samples one of them. | 3 min explainer |
| Top-p sampling | Top-p keeps the smallest set of most probable tokens whose probabilities add up to at least p. | 3 min explainer |
| TorchTPU | TorchTPU is open-source software from Google and Meta's PyTorch team that lets PyTorch programs run directly on Google's TPU chips. | 2 min quick read |
| TPUs | TPUs are chips Google designed to speed up the maths of machine learning, rented out through Google Cloud. | 5 min explainer |
| TradingAgents | TradingAgents is an open-source research framework in which AI agents play the staff of a trading firm and work together on stock trading decisions. | 1 min quick read |
| Training | Training is the repeated loop that nudges a model's internal numbers so its predictions get closer to the right answers in its examples. | 5 min explainer |
| Training data | Training data is the set of examples a machine learning model studies to learn its patterns, kept apart from the examples used to test it. | 5 min explainer |
| Transfer learning | Transfer learning reuses representations learned on a source task to improve learning on a different target task. | 3 min explainer |
| Transformers | A transformer is a neural network that uses attention to decide which parts of a sequence matter to one another. | 3 min explainer |
| Translation app | An AI translation app converts text, or sometimes speech, from one language into another using a machine translation model. | 2 min quick read |
| Tree of thoughts | Tree of Thoughts is a prompting and search approach that explores several candidate reasoning paths, evaluates them, and selects promising branches. | Definition |
| Triton Inference Server | Triton Inference Server is NVIDIA's open-source server software for running trained AI models in production, so apps can send it requests and get predictions back. | 2 min quick read |
| TRL | TRL is Hugging Face's open-source library for post-training language models, with ready-made trainers for methods such as SFT, DPO and GRPO. | 2 min quick read |
| TruLens | TruLens is an open-source library for tracing and evaluating LLM apps and agents, with RAG triad checks and metrics that score each run. | 2 min quick read |
| TruthfulQA | TruthfulQA is a benchmark of 817 questions designed to catch AI models repeating common myths and misconceptions instead of true answers. | 2 min quick read |
| Turing test | The Turing test asks whether a judge, chatting by text with a hidden person and a hidden machine, can reliably tell which one is the machine. | 5 min explainer |
| txtai | txtai is an open-source, all-in-one Python framework from NeuML for semantic search, RAG and language model workflows. | 2 min quick read |
| U-Net | U-Net combines upsampled decoder features with matching high-resolution features from a contracting path. | 4 min explainer |
| Udio | Udio is an AI music service that generates songs from a short text description, and it is being rebuilt under licensing deals with major record labels. | 2 min quick read |
| Ultralytics | Ultralytics builds a line of fast YOLO computer vision models, plus the open-source Python package and command-line tool used to train and run them. | 2 min quick read |
| UMAP | UMAP reduces high-dimensional data using a graph of local relationships, often producing useful visualizations whose apparent spacing should not be treated as exact geometry. | Definition |
| Underfitting | Underfitting is when a model is too simple, or trained too little, to learn the real pattern, so it does badly even on its own training data. | 5 min explainer |
| Unlimited OCR 2 | Unlimited OCR 2 is Baidu's model for parsing very long documents, reading every page in a single request so it can use context from across pages. | 2 min quick read |
| Unsloth | Unsloth is an open-source tool for training and running AI language models on your own computer, built to fine-tune them faster and with less graphics-card memory. | 4 min explainer |
| Unstructured | Unstructured is an open-source Python library that turns messy documents, such as PDFs, web pages, Word files and images, into clean, structured pieces ready for language models. | 1 min quick read |
| Unsupervised learning | Unsupervised learning trains a model on examples that have no answers attached, so it finds structure on its own, such as groups of similar items or points that do not fit. | 5 min explainer |
| V-JEPA | V-JEPA is a Meta model that learns about the world from video by predicting the hidden parts of a clip as abstract features, not pixels. | 2 min quick read |
| v0 | v0 is an AI tool from Vercel that builds user interfaces and full-stack web apps from a description in plain language. | 2 min quick read |
| Validation set | A validation set is held-out data used during development to compare model choices without training on the same examples. | 3 min explainer |
| Vanishing and exploding gradients | Vanishing and exploding gradients are training failures in deep networks where the learning signal shrinks to almost nothing or grows out of control as it passes through layers. | 2 min quick read |
| Variational autoencoders | A variational autoencoder pairs a deep latent-variable model with a corresponding approximate inference model. | 4 min explainer |
| Vector databases | A vector database stores vectors and returns the saved records closest to a query vector, often with a score and any extra fields you ask for. | 4 min explainer |
| Vectors | A vector is an ordered list of numbers that you can also picture as an arrow in space. Machine learning stores examples, words and images this way. | 4 min explainer |
| Veo | Veo is Google's family of video generation models, which turn text or images into short video clips; newer versions also generate sound. | 2 min quick read |
| Veo 3 | Veo 3 is a Google video generation model from 2025 that creates sound along with the picture from a text prompt. | 1 min quick read |
| Vercel AI SDK | The Vercel AI SDK is an open-source TypeScript toolkit, free to use, that lets apps talk to AI models from many companies through the same code. | 4 min explainer |
| Vertex AI | Vertex AI is Google Cloud’s platform for choosing AI models, tuning them on your own examples and running models and agents for real users. | 5 min explainer |
| Vespa | Vespa is an open-source platform for search and recommendation that stores text, structured and vector data and ranks results with machine-learned models as each query runs. | 2 min quick read |
| Vibe coding | Vibe coding means building software mainly by describing what you want to an AI assistant in plain language and letting it write the code. | 2 min quick read |
| Video summarizer | A video summarizer watches or listens to a video and gives you a short written summary, often with timestamps pointing to key moments. | 2 min quick read |
| Virtual assistants | A virtual assistant is an AI helper, usually built into a phone, speaker or computer, that answers questions and carries out tasks when you speak or type to it. | 2 min quick read |
| Vision transformers | A vision transformer represents an image as a sequence of patches and processes that sequence with a transformer. | 3 min explainer |
| Vision-language models | A vision-language model can take visual data and text as input and produce text as output. | 4 min explainer |
| Vision-language-action models | A vision-language-action model maps visual observations and a language instruction to robot actions. | 4 min explainer |
| vLLM | vLLM is open-source software that runs large language models on a server and answers many requests at once, using GPU memory carefully so fewer bytes go to waste. | 4 min explainer |
| Vocabulary | A model vocabulary supplies the mapping used to convert text into an ID sequence and back. | 3 min explainer |
| Voice agents | A voice agent is an AI app you talk to out loud; it listens, works out what you want, can use tools, and answers in speech. | 5 min explainer |
| Voice assistant | An AI voice assistant is an app you talk to out loud: it turns your speech into meaning, decides what to do, and answers in a spoken voice. | 2 min quick read |
| Voice cloning | Voice cloning is AI that learns how a particular person sounds from a recording, then speaks any new text in that voice. | 2 min quick read |
| VoxCPM | VoxCPM is OpenBMB's open text-to-speech model; version 2 speaks 30 languages and can invent a voice from a written description or copy one from a short clip. | 1 min quick read |
| Voxtral | Voxtral is Mistral AI's family of speech models: they turn spoken audio into text, answer questions about it, and in newer versions transcribe live and speak. | 2 min quick read |
| Voyage embeddings | Voyage embeddings are text embedding models from Voyage AI, offered through an API, with general-purpose versions and special ones for code, finance and law. | 2 min quick read |
| Wan | Wan is an open family of video generation models that turn text or images into short video clips, with code and weights free to download. | 2 min quick read |
| Watermarking | AI watermarking hides a signal inside AI-generated images, audio, video or text so software can later tell that the content came from an AI system. | 2 min quick read |
| Waymo | Waymo runs a fully autonomous ride-hailing service: you book a ride in an app and a car with no human driver takes you there. | 2 min quick read |
| Weaviate | Weaviate is an open-source database for AI apps that keeps data objects together with their embeddings, and supports meaning-based, keyword and hybrid search. | 1 min quick read |
| Web scraping agent | A web scraping agent collects content from websites and uses a language model to pull out the facts you asked for as clean, structured data. | 2 min quick read |
| WebGPU | WebGPU is a web standard that lets web pages use the computer's graphics chip (GPU) for fast graphics and general computing, including running AI models in the browser. | 2 min quick read |
| Weights | Trainable weights are model parameters that can be learned from data. | 3 min explainer |
| Weights & Biases | Weights & Biases (W&B), now part of CoreWeave's Forge platform, is a tool for tracking machine learning experiments and for evaluating AI applications. | 2 min quick read |
| Whisper | Whisper is an open-source OpenAI speech recognition model, released in 2022, that turns speech into text and can translate it into English. | 2 min quick read |
| Word2Vec | Word2Vec is a 2013 method for learning word embeddings: lists of numbers for each word, so that words used in similar ways get similar numbers. | 2 min quick read |
| Workflow automation with LLMs | Workflow automation with LLMs means adding language-model steps to an automated process, such as sorting, summarising or drafting, inside a fixed series of steps. | 2 min quick read |
| World models | A world model predicts how an environment may change after an action, so an agent can evaluate possible futures before acting. | 4 min explainer |
| XGBoost | XGBoost is an open-source library that makes predictions by adding up the answers of many small decision trees, trained one after another with gradient boosting. | 2 min quick read |
| YOLO | YOLO, short for You Only Look Once, is a family of object detection models that find and label objects in an image in a single fast pass, fast enough for live video. | 2 min quick read |
| Zapier Agents | Zapier Agents lets you set up AI assistants that carry out tasks across thousands of connected apps, following instructions you write in plain language. | 2 min quick read |
| Zero-shot learning | Zero-shot learning is when a model handles a task it was never specifically trained on, with no examples of that task at all. | 2 min quick read |