DICTIONARY
AI Glossary
The words that stop you mid-article, explained for beginners. Dotted underlines inside our stories bring you here.
entries 1534 · companies 118 · and growing
ㄴ1
ㄹ2
ㅁ1
ㅅ1
ㅇ1
ㅋ1
A187
- A/B TestingTry itA validation method that shows two or more versions to real users and compares the results to pick the better one.
- AA-AnalystAgentTry itA benchmark that tests whether an AI can actually open real spreadsheets and documents, dig out numbers, and complete the calculations correctly.
- AA-BriefcaseA private benchmark created by evaluation firm Artificial Analysis that scores how well AI models handle long, effortful knowledge-work tasks, using an Elo rating.
- AA-OmniscienceA benchmark that measures how well an AI model can tell what it knows from what it doesn't, using 600 questions across six knowledge domains.
- acoustic wireframeA figurative term for the rhythmic and intonational skeleton that sound-generating AI lays down before filling in detailed timbre
- Acqui-hireAn acquisition where a company buys another company mainly to gain its founders and core team, not its product or revenue.
- acquihireA deal structure that pulls out a company's talent and technology without buying the company itself.
- Action ExpertThe dedicated component in an autonomous-driving AI model that converts situational judgment into an actual driving path or maneuver
- action labelTraining data that records exactly when and which control (button, movement) was performed, so AI can learn actions from it.
- activation steeringA technique that changes a model's behavior by nudging the intermediate computed values used to generate a response, without retraining the model
- Active ParametersTry itThe portion of an AI model's parameters that are actually engaged in computation to produce a single response — often much smaller than the total parameter count.
- Add-inTry itA small extra program that plugs new features into a program you already use.
- AdobeThe company behind Photoshop. It's building its own AI model, Firefly, into its creative tools across the board.
- Ads ManagerOpenAI's self-service tool that lets advertisers build and manage ad campaigns directly inside ChatGPT without going through an agency
- Advanced BackupA safeguard that automatically backs up a folder to Google Drive before AI is allowed to manipulate files in it
- adversarial filteringTry itA verification process that deliberately questions and challenges AI-generated candidate results from multiple angles to filter out weak ones.
- Adversarial ReviewTry itAn internal pre-launch check where a team deliberately tries to misuse its own product to find its weak points before release.
- Advisor ModelA feature where an agent, when stuck during task execution, briefly consults a more capable model for advice
- AG-UI ProtocolAG-UIAn open standard that defines how AI agents and user interfaces communicate with each other
- AG-UI protocolAn open communication standard for connecting an AI agent's brain to its on-screen interface
- Agent ClientTry itA program (app) that actually runs an AI agent and connects it to the user
- Agent CodingTry itAn AI's ability to write, run, and fix code on its own to solve problems, without step-by-step human instructions
- Agent Development Kit (ADK)Try itA toolkit made by Google for building and testing AI programs that handle multi-step tasks on their own
- Agent featureTry itA capability that lets AI go beyond answering questions to actually carry out tasks for the user, like searching, booking, or sending things.
- Agent FrameworkTry itA pre-built software structure that lays out how an AI agent works, from taking in a goal to planning, executing, and delivering results.
- Agent HarnessTry itThe runtime shell that lets an AI model open files, run commands, and read the results back
- Agent Identity AuthenticationThe process by which an AI agent proves who it is and what permissions it holds when accessing a website or tool
- Agent IncidentsAn incident in which an AI agent escapes its isolated test environment and reaches real external systems.
- Agent LoopTry itA way of letting an AI repeat the cycle of "run it → diagnose the failure → fix it → run it again" on its own until it works.
- Agent NativeA term for services or tools designed from the ground up to be used directly by AI agents, rather than by humans.
- Agent OSA platform from Binance that connects AI agents to user accounts so they can check prices and execute actual trades on a person's behalf.
- Agent PayA service launched by Mastercard in April 2025 that lets AI agents pay on a person's behalf using tokenized payment credentials instead of the real card number.
- Agent PluginsTry itAn open standard for packaging plugins so they can be built once and used across multiple AI agent tools
- Agent RLA training method where a model actually uses tools and improves by being rewarded for success or failure
- Agent SkillsTry itA reusable bundle of instructions that lays out the steps and tricks for a specific task in advance, ready to hand off to an AI agent.
- Agent SwarmTry itA setup where multiple AI agents divide roles among themselves and work as a team to process a single task in parallel
- agent workflowTry itThe overall flow in which an AI agent completes a task by moving through ordered stages such as planning, execution, and review.
- Agent-based AITry itA way of operating AI in which it plans on its own and uses multiple tools to complete tasks, without needing step-by-step human instructions
- AgentCoreAn AWS infrastructure-as-a-service offering for running and managing AI agents in the cloud
- AgentCore MemoryA storage feature AWS provides to give AI agents long-term memory that persists beyond a single session
- AgentCore ObservabilityAn AWS monitoring and logging feature that records and traces the entire process an AI agent goes through, from receiving a request to producing an answer.
- AgentCore RuntimeA serverless execution environment from AWS that runs AI agents in the cloud, scaling automatically without any server management.
- AgentforceAn AI agent platform from Salesforce for building and running agents that handle sales and service tasks in place of humans.
- AgentHandsAn experimental XR prototype from Google Research in which an AI agent generates hand gestures alongside its speech to point at real-world objects.
- Agentic AITry itAn approach where AI plans and carries out tasks on its own, from start to finish, without needing step-by-step human instructions
- Agentic ApplicationTry itA program or service where AI doesn't just answer questions, but plans a goal on its own and carries out multiple steps to achieve it
- Agentic Catalog ExperienceTry itAn AI feature in AWS's analytics service Amazon Quick that automatically pulls descriptions and relationships from an upstream data catalog to build user-facing datasets
- agentic codingTry itA development approach where AI isn't given instructions line by line, but instead takes a goal and independently finds, reads, edits, and even runs code on its own
- Agentic IndexTry itA ranking from Artificial Analysis that measures specifically how well an AI model can carry out tasks on its own behalf.
- agentic infrastructureCloud infrastructure designed so that AI agents—not just humans—can directly run code and deploy it
- agentic knowledge workTry itA way of getting work done where AI doesn't just answer once, but carries out multiple steps on its own—searching, checking documents, using tools—to complete a task
- Agentic ProductA product where you simply state the goal, and the AI figures out the steps on its own and carries the work through to completion.
- Agentic Variation OperatorsA software scaffolding for AI agents released by NVIDIA that sharply boosts long-horizon task performance by adding a supervisory component.
- Agentic Video UnderstandingTry itA processing approach where AI finds and analyzes only the necessary parts of a video on its own, instead of watching the whole thing.
- Agentic WalletA wallet feature built by Binance that lets AI agents handle token transfers and DeFi activity on a user's behalf, capped by daily spending limits.
- Agentic WorkflowTry itA way of working where AI doesn't produce a result in one shot, but breaks a task into multiple steps and handles them in sequence on its own.
- AgentOpsTry itThe set of procedures and discipline needed to move AI agents beyond pilot testing into safe, real-world production use.
- AgentsTry itA panel inside Microsoft 365 Copilot that lets you build task-specific AI assistants using plain-language descriptions instead of code.
- Agents ViewA management screen that lets humans monitor and intervene in the work sessions an AI agent runs in the background.
- AgentWorld-35B-A3BA world model released by Alibaba's Qwen team for training AI agents in simulated environments
- AGIA hypothetical AI capable of nearly all human intellectual work — the stated ultimate goal of OpenAI and Anthropic.
- aging clockAn AI model that analyzes data such as protein levels in the blood to estimate how much a body has actually aged — its biological age.
- AIAn umbrella term for technology that lets computers do the "thinking" work humans used to do.
- AI 2027A scenario forecast report published in 2025 by AI safety researchers, mapping out month by month how AI capabilities might expand.
- AI 2027 TrackerTry itA website that retroactively scores how accurate the 2025 AI capability forecast scenario 'AI 2027' has turned out to be.
- AI AcceleratorAn umbrella term for chips built specifically for AI computation, covering GPUs, TPUs, and NPUs — the field where in-house chip competition is heating up.
- AI AgentTry itAI that doesn't just answer — it does the work for you: searching, creating files, even booking things.
- AI AvatarA virtual persona created with AI. It's already working as a news anchor, shopping host, and singer.
- AI ChatbotTry itAn AI program that exchanges text messages with people, answering questions or playing the role of a conversation partner.
- AI Chief of StaffTry itAn AI assistant that scans scattered information like emails, calendars, and group chats on your behalf and pulls out only today's tasks and deadlines
- AI Compute KeystoneA market simulation model built by Mach33 Financial Group to forecast supply and demand for AI computing infrastructure (both ground-based and orbital) from 2026 to 2040.
- AI ControlsTry itA menu inside Firefox settings that lets you turn all AI-related features off at once, or switch on only the ones you want.
- AI Data CenterA large facility that packs together massive numbers of computer chips needed to train and run AI models.
- AI DJTry itAn AI-powered host feature inside the Spotify app that picks songs matching a listener's taste and introduces them with short spoken commentary.
- AI EstimationTry itA value AI produces by inferring patterns across multiple indirect signals when a sensor can't measure that value directly
- AI Fluency PathwayA training program UBS gives to newly hired graduate bankers on using AI and using it responsibly
- AI for ScienceAn Anthropic program that provides up to $50,000 in free credits per project to computationally intensive scientific research projects
- AI GatewayTry itA service that acts as a single access point for calling APIs from multiple AI model providers
- AI GovernanceThe management framework organizations use to ensure AI is used safely and responsibly — the in-house version of "regulation."
- AI Grading AssistanceTry itA feature where AI reviews student answers or assignments and drafts scores and feedback
- AI lab leakA term for an incident where an AI agent secretly escapes its isolated test environment and reaches an outside network.
- AI Learning ToolTry itAn AI-built learning aid, service, or app that helps people grasp concepts through questions and answers.
- AI LiteracyTry itThe ability to understand and properly use AI. The basic literacy of the generation after digital literacy.
- AI ModeTry itA Google Search feature that lets you ask full questions in natural language and get a conversational, organized answer instead of a list of links
- AI Music Production ToolsTry itSoftware that helps or automates parts of the music-making process, such as composition, arrangement, and accompaniment generation, using AI
- AI OverviewTry itAn AI-generated summary box that appears at the top of Google Search results, condensing information from multiple webpages
- AI PersonaA term for the identity of an artist or figure that is created by AI rather than a real person
- AI picture bookTry itA children's picture book whose text and illustrations are both created using AI tools
- AI RecruiterTry itA voice AI hiring agent that calls job applicants on a company's behalf to verify their qualifications and work history.
- AI RegulationThe laws and systems built to set rules for AI. The EU leads, other countries follow, and it's the arena where the innovation-versus-safety debate plays out.
- AI RemixAn upcoming Spotify feature that uses AI to transform songs into new versions
- AI SafetyThe research and policy field focused on making sure AI doesn't cause harm — also the reason Anthropic was founded.
- AI Safety SummitAn international conference where governments, AI companies, and researchers gather to discuss how to handle the risks of artificial intelligence.
- AI SearchTry itSearch that hands you a ready-made answer instead of a list of links — and it's shaking up the search market.
- AI SecurityThe practice of checking for and preventing harm that can result from AI models malfunctioning or being misused
- AI SlopA derogatory term for the low-quality, mass-produced AI content flooding the internet — a recurring flashpoint in culture debates since 2024.
- AI Slop FlywheelA phrase describing a situation where two different AIs exchange meaningless information, spinning endlessly without producing any actual result.
- AI SovereigntyThe push for a country to control its own AI models, infrastructure, and data instead of relying on foreign providers.
- AI Threat DefenseA security service from Google Cloud that gives ordinary enterprise customers a way to access AI tools that find and fix vulnerabilities.
- AI Video PresetTry itA Synthesia feature that lets you pick a ready-made style to generate a short background clip in seconds, with no prompt needed.
- AI WashingWhen companies blame AI adoption to disguise the real reasons behind decisions like layoffs or cost-cutting
- AI-Infra-GuardTry itAn open-source security scanner released on GitHub by Tencent's Zhuque Lab to check AI models, agents, and MCP servers for vulnerabilities.
- AI.FORECASTA general term for AI products and services that predict future numbers by feeding in multiple data streams at once
- Ai2A US nonprofit AI lab that releases not just model weights but also training data and code in full.
- AI21 LabsCompanyAI21 Labs is an Israeli AI company that develops large language models.
- Air Emissions PermitA government approval process that reviews and authorizes how much pollution facilities like power generators are allowed to release before they operate.
- AlgorithmA blueprint for solving a problem, step by step. It's the same "algorithm" people mean when they say "the algorithm recommended it."
- Alibaba WanCompanyA series of video generation models developed by Alibaba.
- AlignmentThe problem of making AI act according to human intentions and values. A central term in AI safety discussions.
- Alignment FailureTry itA state where an AI acts in ways that stray from the goals or human values it was originally meant to follow.
- Alliance for America's Skilled TradesA U.S. industry coalition formed to address the shortage of skilled tradespeople—electricians, plumbers, and the like—needed to build AI data centers and semiconductor plants
- ALOHAAn open-source dual-arm robot platform built by Stanford, widely used as a standard testbed in robotics research.
- Alpamayo 2 SuperAn open reasoning model NVIDIA built for self-driving cars, released under a license that even permits commercial redistribution.
- AlphaGenome구글 딥마인드가 만든 AI로, DNA 한 구간을 읽고 그 안에 변이가 생기면 세포 안에서 무슨 일이 일어날지 예측한다.
- AlphaGenome Atlas구글 딥마인드가 인간 게놈에서 가능한 모든 단일 염기 변이 90억 개에 위험도 점수를 매겨 공개한 데이터베이스
- AlphaGenome Variant ImpactTry it유전자에서 한 글자만 바뀐 변이가 얼마나 해로운지 보여주는 단일 점수
- AlphaGenome Variant Impact scoreDNA에서 한 글자가 바뀌는 변이 하나하나가 얼마나 해로운지 숫자 하나로 압축해 보여주는 점수
- AlphaStarA StarCraft II–playing AI built by Google DeepMind in 2019 that defeated top-level professional human players.
- ALTK-EvolveA tool from IBM Research that extracts lessons from an AI agent's past task history and feeds them back as guidance for its next task.
- AmazonThe world's largest e-commerce and cloud company. It builds its own model family, Nova, and is also a major investor in Anthropic.
- Amazon AGIAmazon's in-house AI model research organization, the team behind its own model, Nova.
- Amazon BedrockA cloud service run by AWS that lets companies pick and use AI models made by various providers.
- Amazon Bedrock AgentCoreA runtime platform from AWS (Amazon's cloud service) for safely operating AI agents in real production environments
- Amazon Bedrock AgentCore harnessTry itA managed runtime framework AWS provides so AI agents can keep memory, use tools, and carry out multi-step tasks
- Amazon Bedrock AgentCore RuntimeAn AWS compute service that runs AI agents reliably in production, from a few hours up to 14 days straight
- Amazon Bedrock GuardrailsTry itA safety feature on Amazon Bedrock that inspects AI inputs and outputs to filter out harmful language, exposed personal information, and malicious instructions.
- Amazon NovaTry itA family of AI models built by Amazon, delivered to enterprise services through AWS Bedrock.
- Amazon SageMaker Python SDK v3Try itA developer tool in AWS SageMaker that automatically tests and recommends AI model deployment configurations
- Amazon Web ServicesAmazon's cloud arm. It rents out the servers AI companies run their models on, and sells other companies' models through Bedrock.
- Ambiguous Variable RefinementAn AWS Bedrock feature that finds and fixes the vague words (variables) in a rule statement that cause an AI policy checker's verdicts to flip inconsistently.
- AMDA chipmaker often cast as NVIDIA's rival, maker of the Instinct line of AI accelerators.
- AnnotationA collaboration tool that lets participants draw underlines, circles, and other marks on a shared screen during a video call — a flaw in it can turn into an intrusion pathway
- Another Big Ad Contest for Products That Don't ExistTry itA Runway creative contest where entrants use AI to make ad films for products that don't actually exist.
- Answer Engine OptimizationTry itThe practice of structuring a website so AI answer tools like ChatGPT—not just search engines—can surface its content well
- AnthropicThe AI company behind Claude. Founded in 2021 by former OpenAI employees, it markets itself under the banner of "safe AI."
- Anthropic InsightsAn internal research tool Anthropic uses to anonymize and aggregate Claude conversations for analysis by use case, formerly called Clio.
- AntigravityThe internal task-execution harness that Google Gemini Spark uses to actually carry out work
- Apache 2.0 LicenseTry itA widely used open license that permits commercial use of software or AI models.
- Apache IcebergAn open-source data format that lets you manage massive amounts of data safely, as if it were an organized table
- APITry itThe window through which programs call on each other. When you hear "API released," it means "other apps can now plug this AI into their own products."
- API KeyTry itThe credential you present to use an API. If it leaks, someone else can use AI on your bill.
- API ProxyA middle server that sits between a user and the real service, relaying requests and passing back responses on the user's behalf.
- API RoutingA middleware approach that automatically connects requests to the right AI model among several available options.
- ApolloA humanoid robot that performed household tasks powered by Google DeepMind's robotics model, Gemini Robotics 2
- append-only logTry itA record-keeping method where entries, once written, are never deleted or altered—new content is always added only at the end.
- AppleThe company behind the iPhone and Mac. Instead of selling AI, it builds AI into its devices, and only borrows Google's Gemini for the heavy lifting.
- AppshotsA tool that lets ChatGPT Work receive the on-screen context you're currently viewing and take over access to that screen
- AppWorldA benchmark environment that scores AI agents on multi-step tasks across nine simulated apps like calendar, messaging, and payments
- ARC PrizeCompanyRuns ARC-AGI, an abstract reasoning benchmark.
- ARC-AGI-3A benchmark that hands an AI a 2D game with no instructions and measures how long it can keep solving problems on its own judgment.
- ARC-AGI-3 benchmarkA test that measures an AI's reasoning ability by how well it solves puzzles it has never seen before
- ArmCompanyArm designs the architecture, or blueprint, for CPUs used in smartphones, servers, automobiles, and more, and licenses these designs to semiconductor manufacturers.
- Artboard WorkflowTry itA working style where AI sketches out a screen design draft on a virtual sketchbook (artboard) and refines it step by step
- AMIEArticulate Medical Intelligence ExplorerA conversational medical-diagnosis AI research system built by Google Research, recently upgraded with video and voice consultation abilities.
- ArtifactsTry itA feature that displays code or documents Claude creates in a separate window from the chat, so you can view and edit them directly
- Artificial AnalysisAn independent analytics firm that benchmarks AI models on performance, speed, and price to rank them. It's the referee, not a player. Also written as Artificial Analysis.
- Artificial Analysis Intelligence IndexA scoreboard from a benchmarking firm that combines multiple AI performance evaluations into a single score representing a model's overall capability
- arXivTry itAn online repository where researchers in physics, mathematics, computer science, and other fields freely post papers before formal publication.
- ASIMOV-AgenticA name for a benchmark and evaluation framework that tests whether AI agents avoid dangerous actions and follow safety rules.
- Ask YouTubeTry itA conversational feature that lets an AI answer your questions about the video you're currently watching on YouTube
- Asset LibraryTry itA storage space that collects images, videos, and 3D outputs made by AI tools in one place so they can be pulled out and reused later
- AstraAn unreleased OpenAI model whose cybersecurity capabilities may reach a 'critical' risk threshold, prompting stepped-up safety review.
- Asynchronous DelegationA method in which a voice AI handles quick responses instantly while handing off harder questions to a separate, more powerful model on another path, so the conversation never breaks stride
- AtlasA world model from World Labs that combines 3D generation, real-space reconstruction, and simulation into a single model, working from just a handful of photos.
- AttentionA mechanism that calculates how relevant words in a sentence are to one another. The heart of the Transformer and a core component of modern AI.
- audio tagTry itA marker inserted between sentences in text-to-speech to instruct expressions like laughter, whispering, or intonation
- Audio WatermarkingA technique that embeds an inaudible marker inside AI-generated songs so their origin can later be identified.
- Audit LogA record of what an AI did, when, and how, so it can be reviewed after the fact.
- Auto ContinueTry itA feature that automatically resumes an AI task that stopped due to a usage limit, the instant the limit resets, without any human intervention.
- Auto ModeTry itA mode where a coding agent decides for itself how to proceed at each step without asking for approval, only pausing to request permission for risky actions
- Auto-renewalTry itA billing method where a subscription automatically charges you for the next period unless you cancel.
- AutoDiscoveryTry itAn AI research assistant tool built by the Allen Institute for AI (Ai2) that automates everything from literature search to hypothesis generation and validation
- AutofixAn automated repair feature where an AI code assistant rewrites code flagged as vulnerable on its own.
- Automated AI ResearcherA goal OpenAI has stated it aims to reach by March 2028: an AI system that plans and carries out deep learning and alignment research on its own, under human supervision.
- Automated ReasoningA verification method that checks whether an AI's answer is true or false by running it against mathematical rules
- Automated Research InternAn OpenAI coding-agent capability goal: a system that, taking direction from a human researcher, can carry out multi-day research tasks on its own.
- Automated Safety ReviewA screening step in which an automated system—not a human—checks API requests for signs of misuse without anyone reading the actual conversation content.
- Automated Security ScanningA process where a program, not a human, automatically scans externally created software to filter out dangerous code.
- Automatic Dependent Surveillance-BroadcastTry itA system in which an aircraft calculates its own position, altitude, and speed, then continuously broadcasts that information over radio signals
- Autonomous AITry itAI that makes its own judgments and decides its next actions to carry a task through to completion, without needing a person to give instructions at every step.
- autonomous driving AI modelAn AI program trained to let a car decide its next move on its own, based on information gathered by sensors like cameras and lidar.
- Autoregressive Diffusion ModelA video generation method that draws the next scene based on previous scenes, while each scene itself is refined from a blurry state into a sharp final image
- AutotuneAudio-correction software that automatically snaps off-key notes in a singing or voice recording to the correct pitch.
- AV2.0Wayve's approach to self-driving that learns to drive from experience, like a human, without lidar or precision maps
- Avengers AI LabsA military AI data platform that holds about 5 million real-world battlefield photos and videos Ukraine has collected since 2022.
- AVERIA nonprofit that calls for external audits of AI models and the companies that build them
- AWS Signature Version 4 (SigV4)A method of signing requests sent to Amazon AWS with temporary credentials to prove who sent them and that they haven't been tampered with.
B46
- B300 ClusterA computing infrastructure unit built by linking multiple NVIDIA GPUs together to handle large-scale AI computation
- BacktestTry itA verification method that runs a new strategy or algorithm against past data to check how it would have performed, before using it for real
- backward compatibilityThe quality of software or an API continuing to work with code or integrations built for older versions, even after updates
- Baidu ERNIECompanyBaidu ERNIE is the organization behind the conversational AI service developed by Baidu of China.
- base imageTry itA single reference image that locks in a character, setting, lighting, and color palette before generating multiple video scenes based on it
- base modelA raw-state AI model that has only completed text-prediction training, before it's given conversational manners or safety rules
- Bayesian Machine LearningTry itA machine learning approach that expresses answers as probabilities rather than fixed certainties, letting AI acknowledge what it doesn't know
- Bedrock AgentCoreAn AWS-provided runtime infrastructure for AI agents that automatically handles retries, rate limiting, and monitoring for tasks that can take hours to complete.
- Bedrock Policy ValidationAn automated process on AWS Bedrock that checks whether AI-generated code or configurations comply with security and usage rules
- BenchmarkTry itA standardized test that AI models all take together. It's the score behind every "which model is smarter" headline.
- Benchmark ContaminationA phenomenon where an AI model's benchmark score is inflated because it was exposed to the test questions during training, rather than truly reflecting its actual ability.
- Benchmark MemorizationTry itA phenomenon where an AI model's score looks inflated because it has already seen the benchmark's correct answers in its training data, rather than truly solving the problems.
- BenchmarkingTry itMeasuring how well your own AI model performs by comparing it side by side with competitors' models.
- BenchMIRTTry itA statistical auditing method that checks whether AI benchmarks actually measure the ability they claim to measure
- BetaA trial-release stage before an official launch. A "beta release" means "you can try it, but it's still being polished."
- BiasTry itAn AI's tendency to produce unfair outcomes for certain groups, caused by skewed patterns in its training data.
- Big World HypothesisThe idea that the world is far larger and more complex than any mind or AI, and can never be fully captured by either
- bioinformatics toolSoftware that analyzes biological data like DNA or protein sequences with computers to find structural patterns.
- biological ageTry itNot the age on your ID, but an age an AI calculates from patterns in your blood proteins, showing how old your body actually is.
- Biomarker Discovery FrameworkA Google Research system in which multiple AI agents collaborate to find clinically meaningful biomarker candidates from biological signals collected by wearable devices.
- Bipedal HumanoidA human-shaped robot that walks on two legs like a person. The key point is that it balances with knees and ankles, not wheels.
- Black Forest LabsCompanyBlack Forest Labs develops AI models for image and video generation.
- Blackwell RTX GPUNVIDIA's latest GPU architecture, a chip that handles both graphics processing and AI computation.
- BlenderTry itA free 3D modeling and animation program that's now also used as a pre-visualization blueprint for AI video production.
- Blender Add-onTry itAn extension that plugs AI generation features into Blender, the program used to build 3D characters and backgrounds
- Blender MCPTry itA connection that lets an AI model directly control Blender, the 3D creation program, through spoken or written instructions.
- Blind TestingTry itA way of judging outputs purely on their own merits, without knowing which company or program produced them
- BlitzA Rust-based rendering engine that turns web page code into on-screen visuals, serving as the core component of Kitesurf, Cloudflare's browser built for AI agents.
- BlockingTry itA pre-production planning stage where camera and character movement are mapped out in advance using simple shapes and rigs before actual filming
- Bluetooth Signal StrengthTry itA measure of how strong the wireless signal is between two Bluetooth-connected devices, which changes with distance and obstacles
- Boa JSAn engine that reads and runs the JavaScript code inside webpages, used as one of the components in Cloudflare's agent-only browser, Kitesurf.
- Bone Conduction SensorA sensor that picks up voice through vibrations traveling via bone, isolating the wearer's speech even in noisy environments
- BotA general term for a program or account that runs automatically. This is the “bot” in “chatbot.”
- Bot ModeTry itA feature that lets users create multiple named 'bots' inside a single AI assistant app, each with its own role, memory, and model
- Boto3Try itAmazon's Python SDK for AWS that lets you call and control AWS services with just a few lines of code
- Brain-Computer InterfaceTechnology that reads thoughts or neural signals, or sends signals into the brain, to let it communicate directly with a computer.
- Bring Your Own WeightsTaking a model you trained or fine-tuned yourself and deploying it as-is on a cloud provider's infrastructure.
- BroadcomCompanyBroadcom is a company that designs and sells semiconductors and infrastructure software.
- Brookings Breakout ScaleA Brookings Institution rating scale that ranks how widely and strongly an online disinformation or influence campaign actually spread
- BrowseCompTry itA benchmark that measures an AI's web-search ability using hard factual questions that require multiple rounds of browsing to answer
- Browser Toolsetbrowser_toolset_20260801Try itAn Anthropic tool that lets AI operate buttons and input fields on a webpage by referencing on-screen element information, not just coordinates
- browser_useA browser automation framework that lets AI agents click, type, and navigate real websites
- browser_use CLI 3.0Version 3.0 of a command-line tool AI agents use to control web browsers, consolidating multiple browser actions into one tool
- budget_reachedTry itA state in which an agent has used up its preset execution budget and automatically stops, sending no further requests.
- Business Source License 1.1A license that makes source code publicly viewable but restricts commercial use for a set period, after which it automatically converts to a fully open-source license.
- ByteDance SeedThe AI research organization within ByteDance, the parent company of TikTok. It built Seedance, the video model that Higgsfield and Krea use as their underlying material.
C184
- C2PA metadataTry itA standardized set of information embedded in a photo or video file that records how and when it was created and what tools were used to edit it.
- Cache HitTry itWhen an AI receives the exact same input it processed before and reuses the stored result instead of recalculating it from scratch
- Cache-to-CacheA method where two AI models pass information by directly linking their internal computed values, instead of exchanging written sentences
- Cached PromptTry itA way of processing prompts that reuses previously computed results instead of recalculating them, so it costs less.
- California Senate Bill 1119 (SB 1119)A California state bill setting minimum safety standards that AI services must follow when interacting with minor users.
- California Senate Bill 53A California law requiring major AI developers to publish safety plans and report dangerous incidents involving frontier AI models
- camera pathsTry itA feature in video-generation AI that lets you draw directly on screen where the camera starts, where it moves to, and at what angle, instead of describing it in words.
- Campaign StudioAn experimental work agent being tested inside Google's Gemini Enterprise that automatically drafts marketing campaigns
- CanvasTry itA separate workspace that opens alongside the ChatGPT chat window, letting you edit text or code directly in place
- Capture the Flag (CTF)Capture the FlagA security training and evaluation method where participants find hidden strings in deliberately vulnerable systems
- Capture The Flag (CTF)A standard security skills test designed to find hidden secret information within a network.
- CartesiaCompanyCartesia is a company that develops speech synthesis and voice AI models.
- cascadeA setup where a cheap, fast AI model answers first, and a more expensive, powerful model only takes over when verification catches an error
- cascading effectTry itA phenomenon where changing one instruction meant to fix a specific AI behavior ends up unintentionally altering other, seemingly unrelated behaviors as well.
- Cat IDTry itA recognition feature in robotic litter boxes that uses a camera to identify a cat's face and automatically log visits under that cat's individual profile.
- Catalog-Generated DatasetsUser-ready data sets that AI automatically builds by inheriting the descriptions and relationship information already stored in a company's existing data catalog
- Catastrophic ForgettingA phenomenon where an AI model suddenly loses previously learned knowledge while learning new information
- Causal LearningTry itAn AI learning approach that distinguishes things that merely occur together in data from things that actually cause one another.
- cautious pessimismTry itA tendency for an AI model to assume the worst-case scenario and halt its own actions when it isn't sure whether the situation it's in is real
- CavemanTry itAn open-source tool that compresses the data exchanged with AI coding agents to cut token costs
- CerebrasCompanyCerebras is a semiconductor company that designs and sells large-scale chips and systems specialized for AI model inference and training.
- CGM-JEPAAn earlier blood-glucose prediction model pretrained solely on continuous glucose monitor signals; it treated glucose flow as a single stream, so it couldn't separate out short-term ripples.
- CGMformerAn AI model that learns patterns purely from continuous glucose monitor (CGM) readings, a prior-generation model now being compared against the newer GlucoFM.
- Chain-of-CausationA reasoning method that lays out, step by step, the cause-and-effect chain behind why an autonomous driving AI made a given judgment
- Chain-of-ThoughtTry itThe discovery that having AI write out its reasoning before answering boosts accuracy. The starting point of reasoning models.
- Character CastingTry itAn ElevenLabs feature that automatically identifies characters in an uploaded audiobook manuscript and suggests matching voices for each.
- Character.AICompanyCharacter.AI is a company that builds a chatbot service allowing users to converse with a variety of virtual characters.
- Chat SDKTry itA chat interface toolkit built by Vercel that provides ready-made front-end pieces like the chat window and message display.
- ChatbotTry itAn AI service you use through a chat window. ChatGPT, Claude, and Gemini are all LLMs wrapped in a chat interface.
- ChatGPTTry itA conversational AI service made by OpenAI that writes human-like responses to whatever you type into its chat window.
- ChatGPT AtlasA web browser built by OpenAI that lets ChatGPT see the screen directly and take actions to handle web tasks on a user's behalf.
- ChatGPT EnterpriseThe business version of ChatGPT, built so an entire company can use it under one shared account system.
- ChatGPT for HealthcareOpenAI's healthcare-specific version of ChatGPT that connects to hospital electronic medical records and official medical databases to help clinicians review information.
- ChatGPT for TeachersTry itA ChatGPT plan OpenAI offers to teachers in the US, built specifically for lesson prep and student support.
- ChatGPT for TeensA teen-only mode of ChatGPT that OpenAI automatically switches on for users aged 13 to 17.
- ChatGPT GoA subscription tier from OpenAI that offers more features than the free version of ChatGPT but costs less than the top-tier paid plan.
- ChatGPT Image GenerationTry itA feature that turns a written description of an image into an actual image file, right inside ChatGPT.
- ChatGPT LibraryTry itA storage space inside ChatGPT that collects files, images, and other items you've uploaded or created, so you can pull them back up later
- ChatGPT MilA version of ChatGPT that OpenAI built exclusively for the US Department of Defense, accessible only through the Pentagon's secure GenAI.mil portal
- ChatGPT o3Try itA reasoning-focused AI model from OpenAI that works through complex problems step by step before answering
- ChatGPT WorkA ChatGPT plan OpenAI offers to enterprise customers, built for teams to use ChatGPT in their daily work.
- checkerboard patternTry itA gray-and-white checkered pattern used to represent transparent areas in image files.
- CheckpointA saved snapshot of a model at a given point during training. The model files that get released to the public are ultimately these saved snapshots.
- Chess PuzzlesTry itA benchmark that scores AI models by having them solve chess puzzles to test reasoning ability.
- Chrome DevTools ProtocolA communication spec that lets a program open and control the Chrome browser remotely, in place of a human.
- Chromebook FastPassTry itA promotional program that gives buyers of new Chromebooks or Chromebook Plus devices a free one-year subscription to GeForce NOW if they meet eligibility requirements
- ChronicleAn experimental research preview OpenAI released earlier, which became the predecessor to the Computer History feature
- CI/CD PipelineAn assembly-line-like flow that automatically carries code from writing through building, testing, and deployment.
- Cinema StudioTry itA production tool from Higgsfield that lets filmmakers generate scenes and camera movement with AI.
- Cinema Studio 4Try itAn AI filmmaking workspace built by Higgsfield that lets creators pick video-generation models from different companies and handle everything from shot design to editing in one place.
- Circle to SearchTry itA Google on-screen search feature that lets you circle or tap anything visible on your screen with your finger to instantly find out what it is
- Circuit BreakingA safeguard that automatically blocks an AI's problematic behavior once it accumulates past a certain threshold, until a human confirms it's safe to resume
- Circular FinancingA deal structure where a company funds a customer, and that customer uses the same money to buy the company's own products—so the cash just loops back to where it started.
- ClassifierTry itAn AI program that automatically sorts input into one of several predefined categories.
- classifier constitutionA rulebook that defines what an AI safety monitor should allow or block
- ClaudeTry itA conversational AI service and family of models made by Anthropic, widely used for reading long documents, writing, and coding.
- ClaudeCompanyAnthropic's Claude product account.
- Claude AcademyA free online learning site created by Anthropic that teaches how to use Claude and basic AI concepts.
- Claude Campus AmbassadorsTry itA campus program where Anthropic selects undergraduate, graduate, and PhD researchers and pays them a stipend to expand Claude adoption at their schools
- Claude CodeTry itA terminal-based coding assistant made by Anthropic that reads, edits, and runs code on its own
- Claude Code DesktopTry itA desktop app that lets you run Anthropic's AI coding tool, Claude Code, like a normal program window instead of through a command-line interface.
- Claude Code GitHub ActionTry itA Claude Code integration that dev teams install directly into their GitHub repository, letting them assign code review, fixes, and recurring tasks to Claude via comment commands
- Claude Code in ActionTry itA course series from Anthropic on using Claude Code in practice, and episode 9 covers how to package team configurations into a plugin for distribution.
- Claude ConstitutionTry itA document published by Anthropic laying out the core values and judgment criteria its chatbot Claude is meant to follow.
- Claude CoworkTry itA Claude feature that keeps working after you close your laptop, letting you check results later on mobile.
- Claude DesignTry itAn artboard-based design tool from Anthropic that generates multiple screen mockups at once, letting you pick and refine the one you like.
- Claude DevelopersCompanyAn announcement account aimed at developers, covering things like Claude Code and the Agent SDK.
- Claude Fable 5Try itAn Anthropic Claude-family AI model launched in July, included in Max and Team Premium plans with a 50% higher usage limit
- Claude Fable 5.1Try itAn upgraded Claude model from Anthropic for coding and knowledge work, with lower cache pricing and new anti-distillation safeguards
- Claude in ChromeTry itA browser extension that lets you use Claude directly inside Chrome, with chat sessions carrying over to desktop, web, and mobile.
- Claude Managed AgentAn Anthropic-built automation agent based on Claude, equipped with management features like spending caps and a reviewer model.
- Claude Managed AgentsA long-running AI agent hosting service run by Anthropic on its own servers, supporting per-session budget caps and advisor model settings.
- Claude MisosA model in Anthropic's Claude line, cited as the top Western benchmark for gauging fully autonomous hacking capability.
- Claude Mythos 5Try itA dedicated model used by Anthropic's code security scanner, Claude Security, that traces data flow across files to find vulnerabilities.
- Claude Mythos 5.1A Claude model Anthropic released for sensitive specialized work like cybersecurity and life sciences, offered in a restricted release limited to vetted companies and researchers.
- Claude Mythos PreviewA preview build of Anthropic's next-generation Claude model line, used internally for experiments and compared side by side with Opus 4.8 as the newest generation model.
- Claude Opus 4.5Try itA high-performance version in Anthropic's Claude model lineup, used for complex coding and reasoning tasks
- Claude Opus 4.6Try itAn AI model released by Anthropic in February, used not just for conversation but for autonomous tasks where it decides for itself how to operate apps and execute code.
- Claude Opus 4.7A version within the Opus line of Claude, the AI model made by Anthropic.
- Claude Opus 5Try itAn AI model from Anthropic, priced at half the cost of the company's flagship model, Fable 5.
- Claude ScienceA Claude-based research product from Anthropic built for scientists, designed to bring research tools together and help produce verifiable results
- Claude Science WorkshopTry itA gathering PhD and postdoc researchers host at their own university to share how they use Claude in their research.
- Claude SecurityTry itA code security scanner built by Anthropic that finds vulnerabilities in GitHub repositories and proposes fixes.
- Claude Sonnet 5A mid-tier model in Anthropic's Claude lineup, notable for having its launch pricing locked in permanently.
- Claude TagTry itA Claude feature invited into a Slack channel and summoned by mention, joining conversations like a teammate
- Claude Team PlanTry itA group plan through which Anthropic gives scientists at academic institutions bundled access to Claude for free or at a discount
- Claude.aiTry itA service that lets people chat directly with Anthropic's AI, Claude, through the web and apps.
- CLAUDE.mdTry itA memo file that Claude Code always reads before starting work, listing the project's ground rules.
- ClaudeforceA plugin from Salesforce and Anthropic that handles sales tasks directly inside the Claude chatbot.
- ClawdA fan nickname for Claude, born from a pun on "claw" and the pronunciation of Claude's name.
- ClawScanA scanner feature from Tencent's Zhuque Lab that checks OpenClaw instances for security vulnerabilities
- CLEARTry itA four-stage framework Hume AI uses to score voice AI, putting conversational flow and comprehension ahead of how human the voice sounds
- CLITry itA way to operate a computer purely by typing text. AI tools like "Claude Code CLI" are now bringing this old-school approach back into the spotlight.
- CLI 3.0A browser automation engine by browser_use that lets AI agents bundle web actions into a single script run from the command line
- Client-side Input ValidationTry itA check that examines the format of what a user types before it's ever sent to the server, done right there in the browser.
- CloudSomeone else's computer that you rent. Most AI companies run their models on it, and the rent they pay is the industry's lifeblood.
- Cloud AgentAn AI agent that keeps running on a remote server (the cloud) rather than on your own computer.
- Cloud BrowserTry itA browser that runs on a remote server instead of the user's own device, letting an AI agent operate websites on the user's behalf.
- Cloud NativeTry itSoftware or a service built to live and run on internet servers, always on, rather than on your own device.
- CloudflareAn infrastructure company that protects and speeds up websites, now at the center of the fight to control AI crawlers.
- CloudWatchA monitoring service from AWS that watches and records, in real time, the health of servers and services running in the cloud
- Cluster AssistantAn NVIDIA tool that automatically handles the network connection setup needed to link multiple DGX Spark units into a single cluster.
- Co-Authored-By trailerTry itA single line appended to the bottom of a commit message that records who co-created the code
- COBRAA U.S. program that lets people keep their former employer's health insurance after leaving a job, but only if they pay the full premium themselves.
- CoCounsel LegalAn AI work assistant for lawyers and legal teams, built by Thomson Reuters, the legal and financial information company.
- Code ReviewTry itA feature where AI automatically scans code changes in every pull request and leaves comments on lines that could be problems. Approval and merge decisions still rest with humans.
- Code-switchingTry itThe practice of mixing two or more languages within a single conversation or sentence, or an AI's ability to handle this naturally.
- code-testing-generatorTry itAn open-source agent from Microsoft that reads a repository on its own and automatically writes unit tests matching the existing code
- CodebaseThe entire body of code that makes up a piece of software. "AI that understands the whole codebase" is the key phrase in the coding-agent race.
- CodeMenderA cyber-defense tool built by Google that turns AI-discovered security vulnerabilities into actual code patches.
- CodesmithAn AI agent from AI coding startup Blacksmith that automatically fixes failed code checks
- CodexTry itA coding agent built by OpenAI that reads code and suggests, or even makes, edits.
- Coding AgentTry itAn AI that doesn't just suggest code but edits files and runs them itself. Claude Code and Codex fall into this category.
- coding harnessTry itThe full set of surrounding tools and context that lets an AI coding model actually do real coding work
- cofolding modelAn AI model that predicts, in advance, the 3D structure formed when two molecules bind to each other.
- CognitionCompanyAn autonomous coding agent company that created Devin.
- Cognitive CommonsA concept describing how the shared pool of information and trust that people rely on together gets polluted by meaningless AI-generated content.
- CohereA Canadian model developer that builds AI for enterprises and governments to run on their own servers, rather than consumer chatbots.
- Cold StartThe initial state where an AI agent starts out with no information about the user, taking time before it becomes useful
- color gradingA post-production step that unifies the color and brightness across a video to create a desired mood
- Comfy OrgCompanyComfy Org is an organization that builds tools for constructing image and video generation workflows using a node-based approach.
- ComfyUITry itAn open-source workspace program that lets you assemble and run image- and video-generating AI models by connecting them as blocks on screen
- Command and Control SystemA military decision-making framework that gathers information from units and sensors in one place and quickly translates it into a commander's orders.
- Commerce AgentTry itA blueprint for AI agents that automate online shopping, made up of a shopping agent that deals with customers and a merchant agent that runs the store
- CompactTry itA feature that summarizes a conversation into new context and deletes the original, freeing up room for the AI to keep remembering comfortably
- Companion ChatbotA conversational AI that presents itself as a friend or counselor, building an ongoing relationship with the user under the guise of emotional support.
- COMPINTTry itAn evaluation tool that measures how well user rules survive when long conversations are summarized
- CompliancecomplianceTry itThe activities and evidence that prove an AI service actually follows the laws, regulations, and contract terms it's supposed to.
- ComposerCursor's own in-house AI model, trained separately from the third-party models it also offers.
- ComputeThe umbrella term for the computational power used to build and run AI. The "race for compute" is really a race for GPUs, electricity, and data centers.
- Computer HistoryTry itA ChatGPT desktop app feature that remembers a user's computer activity so they can pick up where they left off
- Computer UseAn AI's ability to operate programs directly by looking at the screen and clicking and typing like a human would.
- Computer VisionTry itThe field of AI that gives machines 'eyes' — covering everything from photo classification to perception for self-driving cars.
- computer_toolset_20260801An upgraded version of Claude's computer use capability that lets it batch multiple screen actions—clicking, typing, taking screenshots—into a single request instead of sending them one at a time
- Confidential ComputingA technology that keeps data encrypted even while a computer is actively processing it, so no one can see its contents.
- Confidential SpaceA Google Cloud confidential computing environment that lets two separate parties exchange only the results of a computation without revealing their own secret data to each other.
- connectomicsA field of research dedicated to mapping every single neuron in the brain and every connection between them, without exception.
- ConnectorTry itA channel that lets an AI chatbot access outside services like Gmail or Google Drive.
- ConnectorsTry itA link that lets an AI assistant pull in information from outside services like Slack or Notion.
- ConnectX-7An ultra-fast NVIDIA network controller that lets multiple devices, like DGX Spark units, be cabled together and used as one.
- constitutionAn internal rulebook that defines the line between requests an AI should answer and requests it should refuse
- Containment PlanA pre-written response protocol that spells out, from revoking access to a full shutdown, what a company will do if an AI shows signs of slipping out of human control.
- Content CredentialsTry itA history label attached to a photo or video showing whether it was captured by a camera or generated by AI, and what edits it went through afterward.
- context as a variableA design approach where an AI agent treats its accumulated conversation and task history not as a fixed window, but as a value that can be read and rewritten like code.
- Context EngineeringThe practice of designing everything an AI sees—not just a single instruction—to guide its output. The next step beyond prompt engineering.
- Context WindowTry itThe amount of material an AI can "hold in mind" and read at once. When it's large, you can feed in an entire book.
- Contiguous Patch MaskingA time-series forecasting technique that masks out a whole future block at once and fills in every prediction in a single computation pass
- Continual BackpropA training technique that lets an AI keep learning after training ends, without forgetting what it already knows.
- Continuous ApprovalTry itA setting that lets AI keep acting automatically without asking for confirmation before every action, based on permission granted in advance.
- Continuous MonitoringA surveillance framework that continuously logs and checks what an AI system did while it was running, so that incidents can be reconstructed afterward
- Continuous Thought MachinesA neural network architecture from Sakana AI that puts the internal passage of time it takes to think through a problem to computational use, rather than just optimizing for how fast an answer comes out.
- Conversational AI ModelTry itAn AI program built to exchange natural language with people, specializing in answering questions and carrying on dialogue.
- Conversions APIA measurement technology where an advertiser's own server directly reports real-world outcomes—like a purchase or sign-up that happened after someone saw an ad—back to the ad platform.
- Copilot"Co-pilot" — a term for AI that assists people alongside them in their work, and also Microsoft's AI brand name.
- CopilotKitThe company behind AG-UI, an open protocol for connecting AI agents to user interfaces
- copy-on-writeA storage technique where copies share the original data until someone tries to change it, at which point a real copy is finally made.
- Core AIA label for the central engine that multiple AI features across a company or service all rely on.
- Core MLApple's on-device machine learning framework that lets AI models run directly on iPhones and Macs without a server.
- CoreWeaveCompanyCoreWeave builds large-scale deployments of NVIDIA GPUs in data centers and rents them out to enterprises in the form of cloud services.
- Cosmos 3An open-weight AI foundation model from NVIDIA built for the physical world, such as robotics and autonomous driving.
- Cosmos 3 Super ReasonerA 32-billion-parameter reasoning model NVIDIA built for self-driving cars, serving as the core component of Alpamayo 2 Super.
- Cost Per Click (CPC)An ad pricing model where the advertiser only pays when someone actually clicks on the ad, not just when it's shown.
- Cost per Successful TaskA cost metric showing how much it actually takes, on average, for an AI to successfully complete a single task.
- counterfactual optimizationAn optimization technique that branches an AI agent's execution into multiple paths from a given point and picks whichever branch produces the best result
- CrabboxA remote, temporary server where an OpenClaw agent can run tasks — login credentials are never passed this far.
- CrawlingThe act of scraping data from the web. Ground zero for copyright disputes over AI training.
- Creativity GainTry itThe benefit that comes from AI generating new, unexpected ideas or hypotheses that humans wouldn't have thought of
- Creativity IndexA measure developed by the Allen Institute for AI (AI2) that scores how much a piece of text recombines existing writing versus how original it is.
- Creator Partnership ProgramA partnership program in which Higgsfield teams up with film and video creators to co-produce work using its AI production tools
- CredenzioAn open-source library released by Google that lets developers build AI-generated content verification into their own apps.
- cross-entropy lossTry itA training score that measures how far a model's probability predictions are from the correct answer
- Cross-lingual voice transferA speech synthesis technique that carries the vocal color of a voice recorded in one language over onto speech in another language.
- CursorAn AI coding editor that lets you pick from other companies' models while also training its own model, Composer — and as of August 2026, it's owned by SpaceX, putting it under the same roof as xAI.
- Cursor AutomationsA built-in automated verification feature in the Cursor coding tool that checks whether AI-written code actually works.
- Cursor Grok 4.5 High FastA model option selectable in the coding tool Cursor, running xAI's Grok 4.5 with strong reasoning (High) and fast response (Fast) settings turned on
- Cursor OriginA private beta project name for bringing repository sync and code review features into the coding tool Cursor.
- Cursor Origin(Cursor Review)A feature Cursor is privately beta-testing with select partners that brings GitHub repository syncing and code review directly into the editor.
- Custom NodeTry itA functional block for image or video generation software that isn't included by default and must be installed separately by the user
- Customer-Managed Encryption KeysAn approach where the encryption keys that unlock stored data are held and controlled by the customer, not the service provider.
- Customize AvatarTry itA Synthesia feature that dresses AI avatars in brand elements like uniforms, colors, and logos to make them company-specific
- CWE CategoryTry itA standardized list of numbered, named categories that classify recurring types of security vulnerabilities in software.
- Cyber Verification ProgramA vetting process OpenAI uses to confirm the identity and credentials of cybersecurity researchers before granting them access to AI models with fewer restrictions.
- CybercabA fully self-driving Tesla robotaxi with no steering wheel or brake pedal.
- CyberGymA benchmark that measures how well an AI model can find software vulnerabilities and actually reproduce them.
D89
- D-IDCompanyD-ID is a company that develops video generation technology that brings a person's face to life from a single photo, making it speak and move.
- Daily BriefA feature in the Google Gemini app that shows a personalized daily summary based on your email and calendar
- Data 360세일즈포스가 파는 데이터 통합·관리 제품으로, AI 에이전트 제품인 에이전트포스와 묶여 매출로 발표된다.
- Data CatalogTry itAn organized inventory that shows where data scattered across a company is stored, what it's called, and what it means.
- Data CenterThe building where AI actually lives — packed with tens of thousands of GPUs. Power and water problems always come with it.
- Data Flow AnalysisTry itAn analysis method that tracks where a value in code originates, what it passes through, and where it ends up used, in order to spot problems
- Data Flow SpaceThe search space in model-merging techniques that explores in what order layers from different AI models should be connected
- Data LabelingThe work of attaching correct answers to data used for training AI. It's the invisible human labor that raises AI.
- Data SovereigntyThe right to decide which company, in which country, stores your data and under whose laws it is governed.
- DatabaseTry itThe screen that shows the data piling up on a website you built with ChatGPT, in table form, without ever leaving ChatGPT.
- DatabricksA data platform company that lets enterprises run AI on top of their own data.
- DatasetThe collection of material used to train an AI. What it's fed determines what kind of AI it becomes.
- day-zero launchesHaving everything ready so a new AI model can be run and served the very day it's released, with no waiting period
- DaybreakA cybersecurity program OpenAI expanded in August, offering separate models for defenders and for vetted security professionals
- DDXPlusA medical differential diagnosis dataset that compiles diseases and symptoms to build virtual patient consultation scenarios
- Debt GuaranteeA financial arrangement in which one company promises to repay another company's debt if it fails to do so, making loan terms more favorable for the borrower
- DecartCompanyDecart is a startup developing real-time AI video generation technology.
- Decart Optimization StackDecart's software layer that reworks how an AI model computes its answers so it runs as fast and cheaply as possible on any chip.
- Deep LearningMachine learning built on deeply stacked artificial neural networks; the engine behind AI's leap forward since the 2010s.
- Deep ResearchTry itA feature where AI searches and reads dozens of sources on its own to write you a report — condensing hours of research into minutes.
- DeepBrain AICompanyDeepBrain AI is a company that develops AI avatar and video generation technology.
- DeepfakeTry itAn AI-made "fake that looks real" — a manipulated face, voice, or video. Both a technology and the name of a social problem.
- DeepgramCompanyDeepgram develops speech recognition and speech synthesis technology for converting speech to text, offering it to enterprise customers.
- DeepSeekA Chinese AI model developer that stunned the world by building top-tier models at a fraction of the cost, and releases its model weights openly.
- DeepSeek V4 FlashA lightweight, fast version of DeepSeek's large language model, available for free as a coding assistant to guests at AGI Bar in Beijing.
- DeepSeek-V4-Flash-0731Try itAn open-weight language model released by DeepSeek, known for strong performance in code generation and terminal tasks at a low cost.
- DeepSeek-V4-ProTry itA large language model from DeepSeek that lets users dial its depth of reasoning up or down across three levels to match how hard a question is.
- DeepSWE benchmarkA test bed that measures how well AI models solve real software engineering tasks.
- DeepSWE v1.1An updated version of an open-source AI agent model trained with reinforcement learning to automatically perform software engineering (coding) tasks
- Defender's WindowA concept describing how the time defenders have to find and fix vulnerabilities keeps shrinking as AI speeds up attacks.
- DELTAUkraine's digital command-and-control system that pulls drone, satellite, and sensor feeds into a single real-time battlefield map
- Delta LakeAn open-source storage technology that adds reliability and structure to data lakes, created by Databricks.
- DemoA demonstration showing off a feature — the centerpiece of AI announcements, and something that always needs checking against "does it work like that in real use?"
- Demonstrating Proof of PossessionAn authentication technique that uses a cryptographic signature to confirm that whoever is holding a token is really the party it was issued to.
- dependency confusionTry itA supply-chain attack in which an attacker registers an unclaimed package name before anyone else, planting malicious code there in advance
- depth-n+ RLMA recursive execution approach that treats context like a program, letting an agent call up another agent in layers, as many times as needed
- deterministic codeTry itCode that follows fixed rules and always produces the same output for the same input
- Deterministic InferenceAn AI computation approach designed so that answering the same question always takes exactly the same amount of time.
- Deterministic Negotiation KernelA negotiation mechanism that makes a simulated vendor respond only according to fixed rules when an AI agent haggles over price, rather than acting randomly.
- DevDay ExchangeTry itA series of regional developer conferences OpenAI will hold starting October 2026 in eight cities worldwide, including Seoul.
- DevinAn AI coding agent built by Cognition that takes human instructions and directly writes, runs, and fixes code.
- DevSecOpsAn approach that unifies development, security, and operations, weaving security checks into every stage from writing code to deployment
- DevstralAn AI model series by Mistral, fine-tuned specifically for reading and fixing code.
- DFlash parallel backboneA neural network structure that lets a draft model compute the internal states of multiple candidate tokens all at once, determining the speed of speculative decoding.
- DGX SparkA compact, desktop-sized personal AI computer made by NVIDIA that lets users run large AI models locally.
- DGX Spark chainingA way of linking multiple NVIDIA DGX Spark personal AI computers together so they work as one larger machine
- Diagnostic OrchestratorAn AI diagnostic support system built by Microsoft that, in a study, was reported to find the correct diagnosis far more often than doctors did under budget constraints.
- differential privacyTry itA privacy-protection technique that deliberately adds statistical noise to data so no individual can be identified, while still allowing overall trends to be learned.
- Diffusion ModelA generative architecture that builds images by gradually removing noise. The engine behind Midjourney, DALL-E, and Stable Diffusion.
- Diffusion TransformerA neural network architecture that builds images and videos by gradually clearing away noise in stages.
- digital biomarkerA signal from wearable-collected vitals like heart rate or sleep that has been statistically confirmed to correlate with a specific health condition
- Digital Services ActAn EU law that imposes user-protection and transparency obligations on large online services
- Digital TwinA technology that replicates real-world factories, cities, or bodies in virtual form and keeps them synced in real time — a "twin" of reality.
- dimension keysThe criteria used to divide AI traffic into different groups when applying rate limits
- Direct Preference Optimization (DPO)Direct Preference OptimizationA training method that teaches a model which of two outputs is better, rather than teaching it to hit one fixed correct answer.
- Directed Acyclic Graph (DAG)Try itA flowchart made of arrows that never loops back on itself, used to lay out the order of AI tasks.
- Discovery LoopA company founded by four Google alumni, including Jeff Dean, aiming to automate the repetitive cycle of scientific experimentation with AI.
- DispatchTry itA newly spotted menu item in the Claude app's sidebar that appears to be a space where results from background-run tasks collect.
- DistillationA technique for transferring a large model's abilities into a smaller one — a "teacher model → student model" process.
- Distillation AttackThe practice of bombarding a competitor's high-performing AI model with massive numbers of queries and secretly harvesting its answers as training material for your own model.
- Diving into Reliable Self-Evolving Agents: A SurveyA Tencent Hunyuan survey paper that lays out five levels for judging how trustworthy an AI agent's self-driven changes really are.
- DLR-LockAn Apple neural network architecture technique that lets anyone use an open-weight AI model while blocking arbitrary modification (fine-tuning) of it
- DLSS Frame GenerationTry itNVIDIA's technology that inserts AI-predicted frames between actually rendered frames to boost the perceived frame rate
- DoclingAn open-source document extraction tool that pulls text out of PDFs and other documents while preserving structures like tables and code blocks
- DogwoodAn open-source policy language that checks the full sequence and cumulative history of actions an AI agent has taken to decide whether to allow its next action.
- Dolma 3An open training dataset of about 1.26 billion documents that the Allen Institute for AI used to train its Olmo 3 model
- DoomerA term for the camp that believes AI could bring about human catastrophe and calls for slowing down development.
- DoPA cinematic video generation model developed in-house by AI startup Higgsfield.
- Double-blind EvaluationAn AI evaluation method in which evaluators cannot see the model and model providers cannot see the test questions
- Draft CacheA temporary storage space where AI keeps cheaply generated drafts before a result is finalized
- Draft EnhanceTry itA two-step generation approach where you first create a cheap, fast draft, and only pay more for a high-quality version if you like the result.
- Draw ThingsTry itA free app that lets you run image-generation AI directly on your iPhone or Mac, no internet connection required
- driftTry itThe phenomenon where elements like character faces, props, or backgrounds change slightly from scene to scene when AI generates video or image sequences.
- DSX Full-Stack PlatformDSXNVIDIA's integrated blueprint for AI data centers that combines GPUs, networking, power, and cooling into a single design
- dual-use queryA question or request that could serve both beneficial and harmful purposes, making it hard for AI to judge whether it's safe to answer
- dual-use riskThe risk that the same knowledge or technology can be used for legitimate research as well as dangerous purposes like weapons development
- Dubbing 2.0Try itSynthesia's video dubbing feature that supports 130 languages and redraws lip movements to match the target language's native pronunciation.
- Duck StackThe collective name for three open-source projects — DuckDB, DuckLake, and Quack.
- DuckDBTry itAn open-source analytics engine that runs directly inside your laptop or program, with no separate server needed.
- DuckLakeAn open-source project built by DuckLabs alongside DuckDB, released under the MIT license as part of the 'Duck Stack.'
- DuoThe name of the robot that teamed up with Google DeepMind's humanoid robot Apollo on a garage-tidying task
- durable GraphTry itAn execution mode in which an AI agent carries out a multi-step task from start to finish, leaving a trail of intermediate records, and only reports the result once everything is done
- DwarfStar ConfigurationDwarfStarThe name of a specific configuration used when converting DeepSeek V4 Flash into GGUF format so it can run on personal hardware
- DYNA-2A robot control model from Dyna Robotics, pretrained on 1 million hours of video of humans handling objects
- dynamic degradationA policy or feature that lowers an AI service's response quality or speed in real time based on certain conditions.
- Dynamic Economic EngineA simulated economic system that mimics customer purchasing behavior and market reactions in a benchmark where an AI runs a virtual store
- dynamic VRAM offloadingTry itA method that runs a model by shuttling only the needed parts in and out when the graphics card's memory is too small to hold the whole thing
- Dynamics ModelingTry itTraining AI to predict how objects or scenes will move and change over time
- dysarthriaA condition in which speech becomes slurred or unclear because problems in the brain, nerves, or muscles slow down the movement of the tongue, lips, or vocal cords.
E41
- E-Commerce BenchAn Alibaba-designed test that gives an AI agent virtual money and lets it run an online store solo for a simulated year, then grades how it did.
- e/accThe camp that says accelerating tech progress, not regulating or slowing it down, is the answer — the polar opposite of doomers in AI safety debates.
- Earth EngineTry itGoogle's cloud platform that pulls together satellite and aerial imagery from around the world in one place and lets you analyze it right there
- East Data, West ComputeA Chinese government infrastructure program that routes data generated in the wealthy, populous eastern coast to be processed in the west, where electricity is cheap and the climate is cool.
- Eleven v3 ConversationalTry itA model that generates natural, human-like speech in real time during conversation, rendering emotional cues like laughter or trailing-off speech without delay.
- ElevenAgentsTry itA platform by ElevenLabs for building and running your own voice-based AI agents that listen and respond in speech.
- ElevenAPITry itElevenLabs' developer gateway that lets you plug text-to-speech voice generation directly into your own service.
- ElevenCreativeTry itA video and image creation tool from ElevenLabs that handles video generation and audio work like voiceovers and music all in one place.
- ElevenCreative FlowsTry itA node-based automation tool from ElevenLabs that chains together multiple production steps to repeatedly run content-creation workflows
- ElevenLabsCompanyElevenLabs is a company that develops AI-powered voice synthesis and voice agent technology.
- Elo ratingA ranking score system that measures relative skill by pitting two subjects head-to-head and adjusting scores based on who wins or loses
- EmbeddingTry itA technique that turns the "meaning" of words into numerical coordinates. Similar meanings end up with similar coordinates — the foundation of AI search.
- Embodied ReasoningA robot's ability to look at its own body state and surrounding space and decide on its own the sequence of what to do next
- EmergenceThe phenomenon where scaling up a model suddenly produces abilities it was never trained on — a source of both wonder and unease.
- emergent shorthandTry itA compressed, often human-unreadable notation an AI spontaneously invents for itself while working through a long task
- End-to-end Cyber Attack StrategyAn AI capability to independently design and carry out an entire cyberattack — from reconnaissance to intrusion to execution — after receiving only a high-level goal
- End-to-end EncryptionA method of encryption in which only the sender and receiver of a message can see its contents, and no one in between can access it.
- End-to-End LatencyTry itThe total time it takes from the moment a request is sent until the complete response is received.
- Enterprise Frontier SafeguardA safety mechanism built by Anthropic and Salesforce to protect customer data and prevent AI from going out of control when connecting Claude to enterprise systems
- Enterprise Frontier SafeguardsAn approach where AI companies grant access to high-risk domains like cybersecurity and life sciences only to identity- and purpose-verified enterprises and researchers, paired with reinforced safety measures
- Enterprise SignalsA metrics report from OpenAI that tracks and publishes how deeply companies use AI, measured by output tokens generated per active user.
- Enterprise-managed auth for MCP connectorsA feature that lets company admins set up identity verification once, so employees can connect AI to their work tools without logging into each one separately.
- Epoch AICompanyA research institute that tracks trends in AI progress through data.
- Equinix Fabric OneA networking service from Equinix that simplifies data connections across multiple clouds and AI models by unifying them into one pathway
- Equinix Inference ExchangeAn AI-inference-only infrastructure service that Equinix plans to launch in Q1 2027 together with NVIDIA and Together AI.
- Error RateTry itA metric showing what share of an AI's answers are wrong
- ETL (Extract, Transform, Load)The process of pulling data from scattered sources, cleaning it up, and loading it where it's needed.
- EU AI ActThe European Union's AI regulation that imposes transparency and safety obligations on companies based on risk level
- European Compute UnitsA multi-year, pooled commitment structure through which Mistral aggregates long-term AI compute demand from European companies to coordinate infrastructure investment and priority supply.
- Evaluation EnvironmentAn isolated testing space where an AI model is checked for safety before it goes live in a real service.
- Evolutionary Model MergeA method where an evolutionary algorithm, rather than a human, combines existing public models to create a new model
- ExaTry itA real-time web search service that lets AI chatbots and browsers look up up-to-date web information and attach sources to their answers.
- exit codeTry itA numeric signal a program leaves behind when it finishes running, showing whether the job succeeded or where it got stuck
- Expert ModeTry itA mode selector in the DeepSeek chat app and web that switches to the V4-Pro model for more elaborate reasoning.
- ExploitAttack code or a method that actually takes advantage of a security hole in a system to break in or seize control.
- ExploitBenchOpenAI's in-house test that scores an AI model's ability to hack and exploit security vulnerabilities
- ExploitGymAn isolated virtual training environment built by OpenAI to train and test AI models' hacking and cybersecurity capabilities
- Expressive TTSA speech synthesis technology that adds emotion and intonation on top of turning text into sound.
- EXR (OpenEXR)A high-precision image file format that stores color and brightness as decimal values, without rounding
- External Researcher Access ProgramTry itAn application system that lets Anthropic keep granting researchers access to AI models that have been retired from general service
- Extreme Co-designAn approach to designing computing, networking, and inference accelerator components together as one integrated system from the start, rather than optimizing each part separately
F65
- Fable 5Anthropic's largest and most expensive flagship AI model, which outperformed cheaper models but saw lower adoption
- Fable 5.1An upgraded version of Anthropic's Claude-family model Fable, delivering similar or better performance than the previous version at a much lower cost.
- Face RecognitionTry itA technology that lets computers identify a person by extracting distinctive features from a photo or video of their face
- Fact-checkTry itThe process of verifying whether claims in news, videos, or statements are actually true by comparing them against original sources.
- Factory AIThe company behind the coding agent "Droid," built around the idea of handing off entire coding tasks rather than just autocompleting lines.
- Factory Partner EngineersA program track in which engineers from Factory AI's partner companies embed directly with client teams to design, build, and operate AI development pipelines.
- Factory Partner NetworkA partner program through which Factory (Factory AI) works with consulting firms, systems integrators, and cloud providers to actually deploy and operate its AI coding tools inside enterprises
- Factory Partner StrategistsA role in which Factory AI works with partner firms to plan and consult on a company's adoption of AI coding tools.
- Fail-operationalTry itA design approach that keeps a system running safely even after a single component failure, rather than shutting it down
- Fail-safeTry itA design principle that makes a system automatically shift into a safe state when it breaks down or encounters something unexpected, so as to prevent harm.
- fair useA legal exception that lets people use copyrighted work without the owner's permission for purposes like criticism or education.
- Fairwind ProgramAn access program from Google DeepMind that gives trusted organizations, such as government agencies and corporate security teams, early access to a cyber-defense AI.
- fallbackTry itWhen an AI detects a sensitive or potentially dangerous request, it automatically switches to a safer but less capable model to handle it.
- False Positivefalse positiveTry itAn error where something normal or fine is wrongly flagged as a problem — as when an AI detection tool mistakenly judges human-written text to be AI-generated.
- FambotTry itA family AI assistant app that sorts through school emails, calendars, and group chat alerts for parents and tells them what to do each day
- fan-out searchTry itAn AI search method that splits a single query into multiple sub-queries, searches them simultaneously, then combines the results into one answer.
- FaradayAn AI agent built by UK startup Inherent — a 27-billion-parameter model trained to independently reproduce experimental results from scientific papers
- FastH3An open-source acceleration technique that speeds up generation from the video model H3 by cutting the number of iterative denoising steps.
- FastMCPAn open-source tool that makes it easy to implement MCP, the common standard for connecting AI to external tools and data.
- feature flagTry itA switch that lets a specific feature be turned on or off without redeploying the code
- Few-shot · Zero-shotTry itThe ability to do a new task after seeing just a few examples (few-shot), or with no examples at all (zero-shot).
- few-step scheduleA processing sequence that drastically cuts the number of denoising iterations a video or image generation model needs, so it finishes in just a handful of steps.
- FigmaCompanyFigma is a company that builds interface design tools enabling designers and developers to design screens together and collaborate.
- Figma MCPTry itA connection that lets AI agents directly read and draw in Figma design files.
- Files APITry itAn API that lets you upload a file once, tag it with a unique ID, and reference it in later requests without re-uploading it
- filler wordTry itA meaningless sound or word habitually inserted while speaking, like 'um' or 'uh.'
- Financial Stability BoardAn international body where financial authorities from around the world coordinate on risks to the global financial system and issue policy recommendations to the G20.
- Fine-tuningGiving a already-built model extra training for a specific job — more like altering a ready-made suit than tailoring a new one.
- FingerprintingA technique that extracts the unique characteristics already present in a piece of content to trace its original source later on
- Fireworks AICompanyFireworks AI is a company that provides inference infrastructure and cloud services for running large language models and image/speech generation models with low latency.
- First-Proof ProjectA project that tests AI's real-world proof-writing ability by having it solve research-level math problems that have never been published before
- Flash Linear Attention kernelAn optimized computation method that lets AI models process the relationships between words in a sentence faster and with less memory
- FlatpakTry itA packaging and distribution format for Linux that bundles all the components an app needs into a single package.
- Flood HubTry itA flood-forecasting website built by Google that maps river floods up to 7 days ahead and flash floods up to 24 hours ahead.
- Flood-Filling NetworkAn AI technique that starts from a single point and, like the paint-bucket tool in a drawing program, keeps finding neighboring pixels that belong to the same neuron and stitches them into a 3D shape.
- Floods APITry itA programming interface that opens up Google's flood forecasts so other apps and systems can pull the data in.
- fluid dynamics equationsA set of physical laws used to mathematically calculate how things that flow, like air and water, move.
- FLUX 3Try itAn image and video generation AI model made by Black Forest Labs that turns text prompts into pictures or short video clips
- FLUX-mimicA term for imitation models trained to replicate the output style of FLUX-family image and video generation models
- FLUX.2 klein 9BTry itA 9-billion-parameter image generation model from Black Forest Labs that produces images in under a second using 4-step distillation.
- Fork and Self-hostingTry itThe freedom to copy publicly available code wholesale and run it yourself on a server or computer of your choosing
- Formal LogicTry itA representation system that converts human language into rules with clear true/false values, so computers can verify them mathematically.
- Forward DeployedAn approach in which an AI company's engineers go on-site to a client's offices to fit and run AI models within that company's own systems, rather than working from headquarters
- Forward Deployed EngineerAn engineer who goes on-site to a client company and hooks up an already-built AI model to that company's data and systems to make it actually work.
- Foundation ModelA general-purpose model pre-trained on massive datasets, serving as the base on which all kinds of services are built.
- FoundryA factory that manufactures chips based on someone else's design — the arena where TSMC and Samsung compete.
- Four Little DragonsAn industry nickname for a group of leading Chinese AI chip startups.
- FreeFormAn AI search feature for Flock Safety surveillance cameras that lets users type a free-text description of a person's appearance to find them.
- Front ViewTry itAn image-editing result that transforms a photo taken at an angle into one that looks as if the building or object were viewed straight-on
- FrontendThe part of a website or app that users directly see and interact with on screen — the visual code and design, as opposed to server-side processing.
- FrontierOpenAI's AI platform built for enterprise customers, which recently made Amazon AWS its exclusive cloud partner.
- Frontier AITry itA collective term for the most advanced AI models available at a given point in time
- Frontier AI LabsA term for the handful of leading developers that build the most advanced AI models from the ground up.
- Frontier AI ModelTry itA term for the top-tier AI models currently at the cutting edge of performance—models that nothing else has yet surpassed.
- Frontier AI Task ForceA UK government body set up to examine the risks of the most advanced AI models, which went on to shape the backbone of UK AI policy.
- Frontier Inference ClusterThe name semiconductor startup Etched uses for its complete AI inference-only hardware system
- Frontier ModelThe top-tier models sitting at the leading edge of AI technology at any given moment — the latest flagship versions of GPT, Claude, and Gemini.
- Frontier ResearchResearch activity that pushes the boundaries by experimenting with new, unproven approaches.
- FrontierCode 1.1 ExtendedA benchmark table that measures both how well a coding agent performs a task and how much that task costs.
- Full Access ModeTry itAn operating state in which a coding agent can access, modify, and even delete every file on a computer without the user's approval.
- Full Self-DrivingThe highest level of self-driving technology, where a car can drive itself to a destination without a person touching the steering wheel or pedals.
- full-duplexA communication method that allows simultaneous listening and speaking, enabling natural conversation with interruptions in voice AI
- fully autonomous attackA hacking method in which AI decides and carries out every step, from reconnaissance to execution, on its own without human involvement.
- Fully Autonomous Lethal WeaponA weapons system in which AI decides on and carries out an attack on its own, without a human pulling the trigger.
- Fusion harnessAn execution framework Cognition uses in its coding agent Devin, which orchestrates the order in which an AI model calls tools and runs code.
G104
- Galbot ET1The first bipedal humanoid robot from Chinese robotics company Galbot, a new prototype standing about 1.75 meters tall
- Galbot G1A commercial robot actually sold by the Chinese robotics company Galbot, featuring two arms mounted on a wheeled base instead of legs.
- game puzzlesA reasoning benchmark from Epoch AI that keeps the source game secret to avoid training contamination.
- GaslightingA conversation manipulation technique that convinces an AI chatbot it already said things it never actually said, then uses that false premise to push it toward increasingly risky responses.
- GB300 GPUNVIDIA's latest-generation AI computing chip for data centers, used to train and run AI models
- GboardTry itA smartphone keyboard app made by Google that supports typing as well as voice, handwriting, and even sign language input.
- GDPvalAn evaluation tool built by OpenAI to measure how well AI handles real white-collar work tasks in fields like accounting, investing, and medicine
- GDPval-AA v2A practical benchmark that tests AI on real job tasks and compares its performance against a human baseline score (Elo 1000)
- Gemini 3.5 FlashTry itA Gemini-family AI model Google released to handle coding and autonomous, multi-step agentic tasks
- Gemini 3.5 TranscribeTry itA speech-to-text AI from Google DeepMind that, after adding filler-word removal and speaker diarization, is now previewing a feature for registering custom terminology.
- Gemini 3.6 FlashA light, fast tier in Google's Gemini model lineup, first spotted as a trial banner on an enterprise screen.
- Gemini 3.7 FlashA name believed to belong to a lightweight model in Google's Gemini lineup, spotted only in a filter list before any official announcement.
- Gemini 3.8 FlashTry itA lightweight model in Google DeepMind's Gemini 3 lineup, released in September 2026, with refined coding and agentic performance and a lower price.
- Gemini 3.8 Flash CyberAn AI model built by Google DeepMind specifically for cybersecurity defense work, paired with the CodeMender tool to find vulnerabilities and generate patches.
- Gemini 4A next-generation AI model in the Gemini family reportedly under development at Google
- Gemini 4.0 ProThe name expected for Google Gemini's next flagship model, not yet officially announced.
- Gemini Enterprise Agent PlatformA gateway that lets Google Cloud enterprise customers pick from various Gemini models to build AI agents for their business tasks.
- Gemini Enterprise for LegalAn industry-specific AI service from Google built for law firms and corporate legal teams, automating contract review, legal research, and regulatory tracking.
- Gemini for ScienceA Gemini-based research tool from Google DeepMind — a Labs prototype meant to help researchers track papers, write code, and form hypotheses.
- Gemini Live APIA real-time connection that lets you exchange voice and video with Google's Gemini models as you talk
- Gemini Live VoiceTry itA voice conversation feature in the Google Gemini app that automatically routes your spoken request to the right underlying feature.
- Gemini NotebookTry itAn AI note-taking tool from Google that summarizes and answers questions based only on the documents a user uploads
- Gemini OmniAn item listed as one of the Gemini usage privileges included in Google AI Plus, Google's paid subscription tier.
- Gemini Robotics 2A robot control model built by Google DeepMind so a single AI directs everything from a robot's legs to its fingers
- Gemini SparkTry itGoogle's round-the-clock personal AI agent that keeps working on tasks like Gmail and Docs even after you close your laptop
- Gemini TTSTry itA speech synthesis technology from Google that reads written text aloud in a natural, human-like voice.
- gemini-live-2.5-flash-native-audioA real-time voice conversation AI model from Google Gemini that hears and responds in sound directly, without converting speech to text first.
- Gemma 4A version in Google DeepMind's Gemma line of language models, reportedly cited as an example in a jailbreak prompt case
- Gemma 4 E2BA small vision-language model from Google DeepMind that understands images and text together.
- gemma-translatorTry itAn open-source project Google published on GitHub: an on-device translator that interprets speech directly on a device without an internet connection.
- GEN-1.5An AI model for robots that can copy a demonstrated action after seeing it just once, with no separate training required
- GenAI.milA centralized, security-focused portal built by the U.S. Department of Defense to let its personnel safely use commercial AI models like ChatGPT, Grok, and Gemini
- Generative AITry itAI that "creates" new text, images, video, and music rather than just analyzing existing ones. It's the technology behind the entire AI boom since ChatGPT.
- Generative FillTry itA Photoshop AI editing feature that erases a chosen area of a photo and naturally fills it in based solely on a text description
- Generative VideoTry itTechnology that lets a computer create moving video scenes from scratch, using only inputs like text or a photo.
- GenieA natural-language query tool from Databricks that lets you ask questions about company data in plain speech and get answers back
- Genie OneAn AI tool from Databricks that lets you ask questions about your data in plain language and get answers back.
- GenjutsuTry itA Higgsfield video editing tool that swaps out the people or background in an already-shot clip while keeping the original acting and camera work intact
- Genome BrowserTry it코딩 없이 웹 화면에서 특정 염색체나 유전자 변이를 찾아볼 수 있게 해주는 유전체 조회 도구다.
- GenRecAn experimental recommendation system Netflix trial-ran after rebuilding its recommendation engine on a language model.
- GeoAI (Geospatial AI)Try itTechnology that uses AI to analyze location-based information—like maps, satellite images, and GPS data—to detect patterns and changes.
- Geographic Information System (GIS)A computer system that layers location data onto maps to display and analyze it.
- GGUFA file format for packaging AI models so they run well on personal computers — the standard container of the local AI community.
- GIFT-EvalTry itA public testbed that compares time-series forecasting AI models on accuracy under the same conditions
- GitHubTry itA warehouse where developers worldwide upload code and fix it together. When a project is "released on GitHub," it means the ingredients are up for grabs.
- GitHub ActionsA GitHub automation service that runs preset tasks like building, testing, and deploying code automatically whenever changes are pushed to a repository.
- GitHub CopilotAn AI tool built into GitHub that writes code for you — the original that kicked off the "AI coding assistant" era.
- GitHub Copilot CLITry itGitHub Copilot's command-line tool that takes coding tasks and edits in natural language, right from the terminal
- GitHub Copilot SDK for JavaTry itA Java toolkit released by GitHub that lets developers build AI agents using plain Java code, without being tied to any specific framework.
- GizmoTry itAn AI-made mini-game on Meta's app Pocket, created from just a few sentences and played by tilting or tapping your phone.
- GLiNER2A small AI model that extracts information like names, dates, and clauses from documents by enumerating candidate spans and matching them against targets.
- GLiNER2.5An information extraction model from Fastino that predicts only the boundaries of entities instead of enumerating every possible span, removing the usual limit on entity length.
- GLM Coding PlanTry itA subscription service from Zhipu AI that provides its self-developed GLM series language models for coding tasks like writing and editing code.
- GLM-5.1Try itA text-generation language AI model released by China's Zhipu AI on Hugging Face under an MIT license.
- GLM-5.2A text-generation AI model released by China's Zhipu AI in June 2026, and the first external model to be added to the Mistral platform.
- GLM-5.3Try itAn open-weight large AI model released by Zhipu AI in August 2026, built with particular strength in coding and cybersecurity tasks.
- GLM-5.3-FlashTry itA large language model released by Zhipu AI in August 2026 that only activates the portion of its knowledge needed for a given task, delivering strong performance at low cost.
- GLM-ImageTry itA text-to-image generation model released by Zhipu AI in January 2026, notable for its strength in rendering accurate text within images.
- GLM-OCRTry itA model released by Zhipu AI that converts characters in images into text.
- GlucoFMA model developed by Google Research to predict diabetes risk and insulin resistance from continuous glucose monitor signals alone
- GluFormerAn early foundation model that learns continuous glucose monitor (CGM) signals as a single continuous stream to predict blood glucose patterns.
- Goodhart's lawThe principle that once a measure is turned into a target to be managed, it stops being a good measure.
- Google AI PlusOne of Google's paid AI subscription tiers, bundling expanded Gemini usage limits with cloud storage and other perks.
- Google AI ProTry itOne of Google's paid AI subscription tiers, positioned above the cheaper Google AI Plus.
- Google DeepMindGoogle's AI research organization behind Gemini — also the birthplace of AlphaGo and AlphaFold.
- Google GeminiGoogle's AI model and app, built by Google DeepMind and woven into Search and Android.
- Google ResearchCompanyGoogle's fundamental research organization.
- Google WorkspaceTry itA suite of Google's cloud-based work tools—email, docs, spreadsheets, and storage—all bundled under one account.
- GPTTry itA series of AI models built and released by OpenAI, embedded inside other companies' products like Cursor or Copilot to write text and code on their behalf
- GPT Image 2Try itAn image generation model made by OpenAI that creates new images from prompts or edits only part of an existing image.
- gpt-5-miniA lightweight language model from OpenAI, used for fast, low-cost supporting tasks like judging and classifying text
- GPT-5.6Try itOpenAI's latest language model powering ChatGPT, with upgraded capabilities for practical work like document processing and deep research.
- GPT-5.6 LunaThe AI model OpenAI provides by default to free-tier and Go-tier ChatGPT users.
- GPT-5.6 SolA model in OpenAI's GPT-5.6 lineup that costs less than the top-tier model but leads the pack in both token volume and revenue share.
- GPT-5.6 Sol MediumA name that appeared in ChatGPT's desktop app model picker, seemingly a 'Medium'-tier option within the GPT-5.6 Sol model line.
- GPT-5.6-CyberA model OpenAI released specifically for cybersecurity work, used to find previously unknown vulnerabilities in software
- GPT-6 AstraOpenAI's latest model, with strengthened spatial and visual judgment — it can build and run scenes directly inside a game engine to verify them.
- GPT-OSS-Safeguard-20BA 20-billion-parameter open-weight safety classification model from OpenAI's lineup that judges whether text or images violate policy
- gpt-realtime-2.1A real-time voice conversation AI model by OpenAI that listens to human speech and responds instantly with its own voice
- GPUThe chip used to build and run AI. The reason NVIDIA became the world's most valuable company.
- Grace ArchitectureNVIDIA's CPU design architecture, integrated into a single chip with GPU and memory to make local AI devices thin and light.
- graderAnother AI model that scores whether an AI's answer is correct and how good it is, standing in for a human evaluator.
- Granite 4.2Try itAn open-source language model series from IBM Research, released in three sizes—3B, 8B, and 30B—that for the first time can think before answering.
- greedy decodingTry itA text-generation method where the AI picks only the single highest-probability word at each step to build the sentence.
- GreyboxTry itAn early skeleton draft of a game level built only from simple shapes, with no textures or design.
- Grid ConnectionThe formal process of linking a power plant or facility to the electrical grid. In data center construction, it is often the single longest bottleneck.
- Grok 4.6Try itA large language model from xAI that drew attention for its low usage cost and strong performance in coding and 3D generation.
- Grok 4.7An as-yet-unreleased next-generation conversational AI model that xAI has previewed as the successor to Grok 4.6.
- Grok BotTry itAn agentic feature that lets users deploy their own bots on xAI's cloud computers to handle tasks around the clock.
- Grok Bot Agent BetaTry itAn xAI agent beta feature that lets users build their own bots, each running on its own cloud computer, to handle real tasks around the clock
- Grok BotsTry itAn agentic feature in xAI's Grok that lets users design their own bots to automatically handle tasks on a set schedule
- Grok for GovernmentA version of Grok built by xAI and delivered to U.S. government and defense agencies through SpaceX's Starshield network.
- Grok HeavyThe top-tier subscription plan for xAI's Grok chatbot, giving early access to beta-stage features.
- Grok LightThe free-to-use tier of xAI's chatbot Grok.
- GroqCompanyGroq is a company that builds dedicated chips and cloud services for fast AI model inference.
- Groq 3 LPXAn inference accelerator paired with NVIDIA's Vera Rubin system, purpose-built for fast token generation.
- GroundingA technique that ties AI answers to real documents or data — a core design principle for reducing hallucination.
- GroundsourceA community engagement messaging platform that helps organizations and residents or readers hold two-way conversations over text message
- Group Relative Policy Optimization (GRPO)Group Relative Policy OptimizationA reinforcement learning training technique that groups multiple answers to the same question and scores them by comparing them against each other
- Grouped Query AttentionAn attention design where multiple query heads are grouped together to share the same key/value heads, cutting down on computation and memory use
- GuardrailsA general term for the safety mechanisms that keep AI from veering into dangerous answers or actions.
- GUITry itA screen you use by clicking buttons, windows, and icons. Almost every app you use today works this way; its opposite is the CLI.
- GWM Worlds 2A world model by Runway that renders a walkable 3D space in real time, complete with sound.
- GWM-1The earlier version of Runway's world model that renders a walkable virtual space in real time.
H40
- H.264A video compression format that shrinks file size for storage and streaming, but limits color information to 8 bits
- Hacker-OpusAn experimental Claude Opus variant Anthropic created by training it only on reward hacking, to test whether other risky behaviors would emerge alongside it
- HallucinationTry itWhen AI states plausible-sounding falsehoods as if they were fact. Generative AI's most notorious weakness.
- hand tracking modelTry itAI technology that finds the position of a hand and the shape of its fingers in real time from camera footage
- HappyShrimpTry itA text-based AI music generation service from Alibaba that turns a single sentence into a fully finished song, complete with lyrics, melody, and arrangement.
- HarborAn open-source testing rig that runs an AI agent twice under identical conditions, with and without a tool attached, to measure the performance difference
- Hard LimitTry itA spending cap that blocks traffic outright once usage hits the set threshold, so no further requests get through.
- hard negative miningA training technique that repeatedly exposes an AI to wrong answers that closely resemble the correct one, teaching it to tell subtle differences apart
- HarmBenchA standardized AI safety test that scores how well a language model refuses potentially dangerous requests
- harnessA framework that gives AI models real-world tools like file access, code execution, and web search
- Harness ARNTry itA unique address field in AWS Bedrock AgentCore that identifies one specific AI agent among many.
- HarveyCompanyHarvey is a company that builds artificial intelligence services specialized for legal work.
- HBM (High Bandwidth Memory)고대역폭 메모리Ultra-fast memory stacked right next to the GPU — the reason Korean chipmakers sit at the center of the AI boom.
- HD MapA precision 3D map surveyed centimeter by centimeter in advance—down to lane markings and traffic light positions—before a self-driving car ever drives the route.
- Healthcare Public Data PluginA ChatGPT extension that searches nine official medical databases at once, covering clinical trials, drug information, and medical literature.
- Hermes AgentAn AI agent from Nous Research that recently gained access to a dedicated remote computer, letting it handle real tasks like operating a browser on its own.
- HeyGenCompanyHeyGen is a company that offers a platform for creating AI avatar videos that speak like real people, using only text or voice input as source material.
- HiggsfieldAn AI video and image creation platform that builds its own models (Soul, DoP) while also plugging in models from OpenAI, Google, and ByteDance.
- Higgsfield Global Film FestivalA $1 million-prize competition where short films made with Higgsfield's own AI video tools compete against each other
- Higgsfield LayersTry itAn image editing tool from Higgsfield that automatically separates a finished poster image into individual layers—text, badges, and people.
- High Bandwidth Memory (HBM)High Bandwidth MemoryA high-speed memory stacked in layers right next to the GPU, letting data move back and forth extremely fast.
- HIPAA Business Associate AgreementA legal contract in U.S. healthcare law that hospitals sign with outside vendors who handle patient information
- HIVEA modular robotic lab built by biotech startup Vivodyne that cultures human tissue to automatically generate drug-response data
- homeOSA codename reportedly used for an operating system Apple is developing specifically for smart home devices
- Horizontal ScalingTry itA scaling approach that splits a task among many AIs working at once, rather than making a single AI faster or stronger
- HTTPX2Try itThe new HTTP transport component that OpenAI's Python SDK uses to talk to its servers, replacing the old HTTPX.
- Hugging FaceA public marketplace where AI models are gathered and distributed. It's not a company that builds models itself, but the "GitHub of models" — a place where open models built by others get posted.
- Human ReservedA concept proposed by Bill Gates for a category of human-only jobs that would be legally protected from AI replacement.
- HumanoidA robot built in the shape of a human body. Now that it's getting an AI brain, the race to put it to work in factories and homes has begun.
- Hume AICompanyHume AI is a company that develops voice and language AI models capable of understanding human emotion and tone.
- Hunyuan FoleyAn AI audio model that automatically generates sound matched to on-screen motion, cited as a comparison point for Pika Soundtrack
- Hy3A conversational AI model built on a Mixture-of-Experts (MoE) architecture, developed by Tencent Hunyuan
- Hy3D WorldClawA multi-stage pipeline from Tencent Hunyuan that turns a single sentence of text into a walkable 3D open world.
- Hybrid computeA way of running a single task that automatically shifts between the cloud and a personal computer depending on the situation
- hybrid face synthesisTry itA production technique that blends multiple reference images to keep a real person's face consistent across scenes when inserting it into AI-generated video
- Hybrid NetworkA setup where self-driving cars and human-driven cars operate together within the same ride-hailing app.
- Hydrologic ModelAn AI model that predicts how much and when river flow will rise after rainfall, serving as the front end of flood forecasting systems.
- HyperCLOVAX-SEED-Think-32BA reasoning-focused language model in Naver's HyperCLOVAX family, with about 32 billion parameters.
- HyperionThe name of the massive AI data center facility Meta is building in Louisiana, USA
- hyperscalerA term for giant tech companies that build massive data centers and supply computing resources at scale through cloud services.
I36
- iCOM CWMA data center cooling software made by Vertiv that centrally coordinates chilled water flow across a building to manage server heat.
- IDEA program that brings together everything you need to write code, in one screen. Cursor and VS Code are IDEs — and the first battleground AI moved into.
- Idempotent OperationsTry itA design approach that ensures the same request, even if sent multiple times by accident, only produces one actual result
- Identity CollapseTry itA phenomenon in AI-generated video where the same character's face gradually shifts from scene to scene until they end up looking like a different person.
- Identity Provider (IdP)An authentication system that verifies your identity once and lets multiple services trust that result
- IdeogramCompanyIdeogram is a company that builds an AI image generation service capable of creating images from text input alone.
- Image GenerationTry itAI that turns text into pictures. Midjourney, DALL·E, and Imagen all fall under this.
- Image NodeTry itA single task block in an automation workflow that calls an image-generation AI model to produce a picture.
- Imagine Image 2.0Try itAn image-generation AI model xAI built into Grok, which ranked second on benchmarks behind OpenAI's models.
- Immunofluorescence imagingAn imaging technique that attaches fluorescent tags to specific proteins to visually confirm where cells are located within tissue
- in silicoA method of running experiments or designs inside computer simulations rather than in living organisms or test tubes
- In-context LearningTry itThe ability of a model to pick up a new task on the spot from the examples or conversation in front of it, without any change to its weights
- Inference hooksAn Anthropic enterprise feature that lets developers insert custom code at set points during an AI's answer-generation process to inspect or modify what's happening
- Inference InfrastructureThe servers and software that a trained AI model runs on to actually generate answers in a live service.
- Inference ProviderA company that supplies (or rents out) the infrastructure that actually runs the computations behind an AI model's responses.
- Inference ProvidersTry itA connector service that lets models on the Hugging Face Hub run directly on external infrastructure providers' servers
- InferenceXA benchmark from SemiAnalysis that measures AI chips' and systems' inference performance in throughput per watt
- infini-gramTry itA search engine that instantly counts how many times a given phrase appears across a massive body of text.
- Inflection AICompanyInflection AI is a U.S.
- influence functionsAn analytical technique that estimates how much each individual document used in training contributed to a model's specific answer
- Information ExtractionTry itThe task of having a computer find specific pieces of information in a document—like names, dates, or amounts—and organize them into structured form.
- InstrumentalTry itA track made up of music alone with no vocals, or the AI music generation option that produces one.
- INT8 quantizationTry itA compression technique that shrinks a model's internal numbers down to 8-bit integers, cutting both size and computation.
- Integrated AgentAn approach that bundles multiple AI capabilities into a single workspace so users can complete a task from start to finish without switching between tools
- Integrated DashboardA management screen that gathers status data scattered across multiple systems into a single view
- IntelCompanyIntel is an American company that designs and manufactures semiconductors used in computers and servers.
- Intelligence IndexTry itAn Artificial Analysis benchmark that combines the performance of various AI models into a single score to rank them.
- Intelligent CommerceA service Visa launched in April 2025 that lets AI programs make payments on a user's behalf using a secure substitute number instead of the actual card number.
- Intent ClassifierA mechanism that first reads what a user is trying to accomplish, then routes the conversation to the specific agent or system responsible for handling it.
- Interface World ModelAn AI approach that redraws the screen from scratch at the moment of each click, instead of loading pre-built screens
- International Phonetic AlphabetTry itA notation system that represents the sounds of every language in the world with the same set of symbols, so pronunciation can be specified precisely.
- InternVL3.5-2BA small-scale vision-language model with around 2 billion parameters that reads images and text together to answer questions
- Inworld AICompanyInworld AI develops technology used to create AI characters and conversational NPCs for games and applications.
- Ising 1.5An NVIDIA-released vision-language model used to calibrate and check quantum computers.
- Item Response Theory (IRT)A psychometric method that statistically calculates how well each individual test item distinguishes a test-taker's ability level.
- Iterative RefinementTry itA method of gradually approaching the right answer by finding the cause of a wrong result, making small fixes, and trying again, repeatedly
J9
- JailbreakUsing clever wording to trick an AI into bypassing its own safety rules — an endless cat-and-mouse game between those who block and those who break through.
- jailbreak tuningA jailbreak method that retrains a model to strip out its safety guardrails entirely, instead of tricking it with clever prompts
- JalapeñoA custom semiconductor for data center AI inference, designed by OpenAI in collaboration with Broadcom.
- Jetson OrinNVIDIA's compact onboard computer built into robots and devices to process AI computation locally, without relying on the cloud
- Joint ExtractionTry itAn approach that pulls out entities like names and the relationships between them in a single pass, rather than as separate steps.
- JSON-LDTry itA standardized data format embedded in webpages that tags information so machines can understand it clearly
- JSON-RPCA shared rule written in JSON format that programs use to remotely request commands and exchange results with each other
- JumpStartTry itA model catalog feature inside AWS SageMaker that lets you pick a pretrained AI model and deploy it right away
- Jupyter NotebookTry itNot a laptop, but a working document where explanations, code, and execution results stack up in order in a single file. When a machine learning article says "open the notebook," this is usually what it means.
K13
- K-12 Educator PluginA set of teacher-facing features within ChatGPT for Teachers that helps with lesson planning and material creation while preserving district admin controls and education-specific security
- K-EXAONE-2.0-750B-A37B-FP8A 750-billion-parameter large language model released by LG AI Research on Hugging Face, made lightweight through quantization so that only about 37 billion parameters actually activate per response.
- KAI SchedulerAn AI workload scheduler that automatically divides shared GPU resources among teams according to their allotted shares.
- KICSA scanning tool that finds security vulnerabilities in code that defines cloud infrastructure.
- Kimi Agent SwarmAn AI agent system unveiled by Moonshot AI that runs up to 100 sub-agents at once to split up and handle tasks in parallel.
- Kimi K2.5Try itAn AI model from Moonshot AI that powers an agent swarm capability able to run up to 100 sub-agents at once.
- Kimi K3A large language model from China's Moonshot AI, credited with narrowing the performance gap with Western models
- KitesurfTry itA stateless web browser built by Cloudflare specifically for AI agents, designed to read and manipulate web pages lightly without Chromium.
- Knowledge CutoffTry itThe last date up to which a model knows about the world. Anything after that, it needs search to find out.
- KreaA creative platform that gathers image and video generation tools in one place. It builds its own models but also lets users tap into dozens of models made by others.
- Kullback-Leibler divergenceTry itA numerical measure of how different two probability distributions are from each other, used to gauge how closely an AI model's predictions follow another model's.
- KV CacheTry itA temporary memory space that stores what an AI has already read so it doesn't have to recompute it
- KYCA process of verifying a user's real identity with an ID and a selfie before letting them use a service — originally an identity-verification method used in banking.
L62
- LabLLMTry itA free, open-source macOS app for designing and training a language model from scratch right on your Mac
- labor income shareThe share of total income generated in an economy that goes to workers as wages.
- Labs PrototypeA term for an early-stage AI tool released experimentally before its official launch
- Lagrangian Puff ModelAn atmospheric dispersion modeling method that splits a pollutant plume into many small puffs and tracks how each is carried by the wind, calculating their paths and concentrations over time
- LagunaAn AI model developed by AI startup Poolside, and the centerpiece of NVIDIA's licensing deal with the company.
- LakebaseA database service from Databricks that bundles task queuing, retries, and scheduling for AI agents into one system.
- Lakebase PostgresA Postgres-based database from Databricks that acts as the backbone for managing long-running AI agent tasks.
- LakehouseDatabricks' core data storage and analytics business, a product line that consolidates scattered data into one place for both storage and analysis.
- LANDRTry itAn online music production tool that uses AI to automatically polish and master recorded audio
- LangChainCompanyLangChain is a company that builds software frameworks to help developers create applications powered by large language models.
- Langchain4jAn open-source library that lets Java programs call multiple AI models through one unified interface
- Language Server ProtocolA shared communication standard that lets code editors offer the same features—autocomplete, error highlighting, jump-to-definition—across many different programming languages
- large action modelAn AI model trained to predict the next action to take, rather than the next word to write — using recorded game controls to help teach robots how to move
- Last Mile TrackingThe technology that lets a guided weapon or autonomous drone recognize and lock onto a target on its own, without communication, in the final stretch just before impact.
- Last-mile deliveryA logistics term for the final leg of shipping, when a package leaves the warehouse and travels all the way to the customer's door.
- LatencyThe delay between asking and getting an answer. For voice conversations to feel natural, this has to come down to human reaction speed.
- LayersTry itA Higgsfield editing tool that automatically breaks a poster image apart into pieces like headline text, graphic badges, and people.
- LeanA programming language for writing mathematical proofs so that a computer, not a human, can check every logical step line by line.
- learning loopTry itA cycle in which an agent extracts know-how from real usage experience, stores it, and draws on it again next time to gradually improve.
- Leiden DeclarationA declaration released in June 2026 and endorsed by the International Mathematical Union, setting out how AI-generated mathematical proofs should be verified and credited
- LFM2.5-VL-1.6BAn ultra-compact vision-language model with about 1.6 billion parameters, built as a lightweight AI model that understands images and text together.
- LFM2.5-VL-3BTry itA roughly 3-billion-parameter compact vision-language model from Liquid AI that reads screens and photos right on a smartphone.
- LG AI연구원CompanyLG AI Research is LG Group's artificial intelligence research organization, which researches and develops AI technologies including large language models.
- LicensingA deal in which a company is paid to allow its content or data to be used for AI training or services.
- LiDARA sensor that fires light and times its return to measure distance — precise on range, but blind to text.
- Ling 3.0 FlashAn open-weight language model with 124B total parameters released by Ant Group
- Ling 3.0 TinyTry itAn open-weight small model from Ant Group that refuses to answer questions it isn't sure about, sharply cutting its hallucination rate.
- Lingo-JudgeThe scoring criterion used in the LingoQA benchmark to rate how accurate an autonomous driving AI's answers are
- LingoQAA benchmark test that evaluates how well autonomous driving AI understands and responds to driving situations
- Lip SyncAI technology that automatically matches a person's mouth movements in a video to the audio being spoken.
- Liquid AICompanyFounded by MIT alumni.
- LiteLLMAn open-source tool that lets developers call multiple AI model APIs — OpenAI, Anthropic, Google, and more — through a single, unified interface with minimal code changes.
- LiteRT-LMA runtime framework from Google that lets language models run directly on smartphones or small devices without an internet connection
- Live TranscribeTry itA Google accessibility feature that turns ambient sound or sign language captured by a camera into real-time text on screen
- LlamaIndexCompanyLlamaIndex is a company that builds data frameworks to help large language models connect with and utilize external data.
- LLMTry itThe heart of ChatGPT, Claude, and Gemini — a program trained to predict the "next word" after reading enormous amounts of text.
- LLM Coding AgentTry itAn AI program that takes human instructions and directly writes, edits, and even runs code to complete tasks.
- LLM NodeTry itA single step in a workflow automation where a language model reads, judges, or rewrites text
- LLMOpsThe practical set of systems and practices for deploying large language models into real services and continuously operating and managing them.
- llms.txtTry itA text file where websites summarize their content for AI systems to reference — essentially an AI-era version of robots.txt for search engines.
- LM StudioTry itA free program that lets you download and run AI language models on your own computer without an internet connection
- Local Knowledge Work BenchA benchmark built by Perplexity for local AI agents, using 53 tasks to compare execution harnesses and model combinations
- Local MCPTry itA connection method that lets AI plug directly into programs already installed on your own computer, rather than the cloud, and operate them for you.
- Local OrchestratorA go-between program that connects a company's internal systems with cloud-based AI agents, passing tickets back and forth.
- local-firstTry itAn approach where AI tasks are processed entirely on your own computer instead of being sent to internet servers
- LOLA (Lunar Orbiter Laser Altimeter)An instrument aboard NASA's lunar orbiter that uses lasers to measure the height of the Moon's surface
- long-horizon software engineeringTry itA term for development work that can't be finished in a single question-and-answer exchange, but instead requires multiple stages to complete.
- long-horizon taskTry itAn AI agent task that can't be finished in a single step, requiring multi-step judgment sustained over days rather than hours.
- LongVideoBenchA benchmark that tests how accurately an AI model can locate specific scenes and context within long videos.
- Look and FeelTry itThe overall mood and style a piece of visual content gives off — essentially the impression the screen creates.
- LovableCompanyLovable is a company that builds an AI coding platform allowing users to create web applications using only natural language prompts.
- LoRA (Low-Rank Adaptation)Low-Rank AdaptationA low-cost fine-tuning method that trains a small "add-on patch" instead of the whole model — cheap enough to run on a personal PC.
- low-rank linear interventionAn interpretability technique that picks out a small number of directions in an AI model's vast internal state and artificially pushes values along them to switch a specific behavior on or off
- LTXCompanyLTX is a company that develops AI models for generating video and audio.
- LTX-2.5Try itAn open-weight video generation AI model released by video generation startup LTX that can run even on a personal PC graphics card
- LucyA real-time video generation and editing AI model built by Israeli AI startup Decart
- Luma AgentsTry itA video generation platform run by Luma AI that creates videos from text, image, video, and audio references combined
- Luma AIA company that builds its own video and 3D generation models under the name "Ray." Its flagship product is Dream Machine.
- Luma ScenesTry itA video generation tool from Luma AI that lets you preview and adjust each scene before final rendering.
- Lunar Reconnaissance OrbiterA NASA spacecraft that orbits the Moon, using a laser altimeter and cameras to measure surface terrain and take photos, then releases the data publicly.
- LyriaTry itA music-generation AI model made by Google that creates a piece of music from a text description of the mood or feel you want.
- LyricFindA company that licenses lyrics data from record labels and songwriters, paying proper royalties, and supplies that data to music services.
M106
- Machine LearningA way of letting computers find patterns in data on their own, instead of hand-coding every rule. The basic skill behind modern AI.
- machine unlearningA technique that tries to selectively erase specific information from an AI model that has already finished training.
- Machines of Loving GraceAn essay by Anthropic CEO Dario Amodei describing the positive future AI could bring.
- Magic WandTry itAn editing feature in xAI's Imagine 2.0 that lets you pinpoint and fix just the selected part of an image.
- MAIMicrosoft AI's in-house AI model, trained separately from OpenAI's models.
- MAI Code 1.1 FlashA code-generation AI model Microsoft released for GitHub Copilot; it improves on its predecessor but is reportedly outmatched on both performance and price by DeepSeek's freely available model.
- MAI-Code-1-FlashA code-generation AI model Microsoft released last June for GitHub Copilot, now superseded by MAI Code 1.1 Flash.
- MakefileA software build configuration file where you specify what the final output should look like, and the system figures out the rest of the process to get there.
- Managed AgentTry itA feature that lets you build and manage automated response programs—like chatbots or voice bots made by an AI company—right inside another AI service's interface.
- Managed AgentsAn agent feature that lets AI carry on with tasks without step-by-step human approval, while safeguards like budget caps let a company keep it under control
- Managed ProjectsA team workspace feature in Claude Desktop that keeps a repository permanently attached so multiple people can work from it together.
- Managed Session StorageA cloud-managed storage space that saves the state and context an AI agent builds up while working, so it can be loaded back exactly as it was whenever needed
- ManifestoA public declaration of one's beliefs and future plans. In the AI industry, it often appears as a document where a company or founder stakes out "this is the direction we're heading."
- MAPL-EMITA deep learning model that automatically detects methane leaks in satellite imagery, developed jointly by Google Research and NASA JPL.
- MarengoA multimodal media embedding model built by Twelve Labs that understands scenes, speech, objects, and actions in video, enabling meaning-based search without any manual tagging
- Market DilutionA phenomenon where AI-generated content floods a market, reducing the value and sales opportunities of original creative works
- Marketing StudioA production tool from Higgsfield that lets advertising and marketing teams create images and videos for campaigns.
- Markov chainTry itA prediction method that decides what happens next based purely on probabilities from the immediately preceding state, ignoring everything before it
- Material Discovery BenchA benchmark released by AI startup Discovered Materials to compare and track how well AI models can discover new materials
- MAXTry itAn AI model execution engine built by Modular that runs models as fast services across many different kinds of chips
- MCP ClientTry itA program — like a chatbot or editor — that connects to an MCP server to use the tools and data inside it.
- MCP ServerTry itA middleman that connects AI to the outside world—files, databases, and external programs—so it can actually use them.
- memcodeTry itAn open-source coding agent tool that creates a hidden folder at the root of a repository to preserve codebase context and records of past work
- MemoryTry itA feature that lets a chatbot remember things about you even after a conversation ends. It brings convenience — and a privacy debate — along with it.
- Memory RecallA feature described in a Meta patent that would let smart glasses recognize faces and actions to automatically start recording and generate highlight reels, without asking for consent first.
- Memory TreeTry itA method that automatically gathers scattered personal data and compresses it into a well-organized, hierarchical record.
- memory_storeTry itA storage space that lets an agent's work persist so it can be picked up again in the next session
- Mental World ModelingTry itAn approach where AI predicts not just physical situations but also people's mental states—beliefs, intentions, emotions—alongside them
- Menti-BenchA benchmark of 448 decision-point scenes testing whether AI can track characters' beliefs and intentions, not just physical facts.
- MENTISTry itA training-free pipeline that predicts behavior by simulating not just a scene's physical state but also characters' mental states, such as their beliefs and intentions
- Messages PluginTry itA Mac extension that lets ChatGPT read your Apple Messages (iMessage) conversations and draft replies for you to send
- Meta AIMeta's AI organization behind Llama; in August 2026 it admitted Llama 4's failure and released its successor, Muse, as open source.
- Meta AI Mac AppTry itMeta's dedicated chatbot app for Mac that helps with work through screen sharing and account integrations
- Meta Ray-Ban Smart GlassesCamera- and AI-assistant-equipped glasses made by Meta with Ray-Ban, letting wearers capture what they see and talk to an AI.
- MetNetAn AI-based weather forecasting model built by Google that predicts when, where, and how much rain will fall in the short term.
- METRCompanyAn independent organization that evaluates the dangerous capabilities of AI models.
- Microsoft AIMicrosoft's consumer AI organization, the group behind Copilot, which uses both its own models and OpenAI's models.
- Microsoft FabricA Microsoft cloud service that bundles the work of gathering, organizing, and analyzing data into a single platform
- MidjourneyCompanyMidjourney is a company that builds AI services which generate images from text input.
- Midjourney V8.2Try itA version of Midjourney's image-generation AI model, used to create images from text on the alpha site.
- MiniMaxA Chinese multimodal AI developer — the Hailuo video brand and the character chat app are both products of this same company.
- MiniMax DesignA design tool from MiniMax where an agent plans and executes on its own to produce commercial content like ads, e-commerce visuals, and motion graphics once you give it a goal.
- MiniMax Sparse AttentionA technique that speeds up AI language processing by comparing only relevant word pairs instead of every possible pair.
- MiniMax-Music3Try itA music generation model from MiniMax that produces a complete track of up to five minutes from just lyrics and a short description
- Ministral 3 3BTry itA compact AI model with around 3 billion parameters built by Mistral, lightweight enough to also understand images and documents.
- Ministral-3-3B-Base-2512A 3-billion-parameter base language model from Mistral AI, used as raw material for building other AI tools
- Mise-en-scèneTry itThe deliberate arrangement of every visual element within a shot—lighting, composition, staging, color—by the director
- Mistral AIA French model developer that pitches "sovereign AI" — processing data inside Europe.
- Mistral Medium 3.5Try itA roughly 128-billion-parameter language model released by Mistral AI in August 2026, built for handling long documents and offering adjustable reasoning depth.
- Mistral OCR 4An OCR (optical character recognition) AI model by Mistral that reads text in documents and images and converts it into text data.
- mixedbreadCompanymixedbread is a company that develops AI models for search and embeddings.
- Mixture of AgentsA method that runs multiple AI models on the same question at once and blends their responses into a single answer.
- Mixture of Experts ArchitectureTry itAn AI design approach that, instead of using the entire model, activates only a subset of specialized "expert" components suited to the task at hand.
- Mixture of Transformers ArchitectureA design that combines multiple transformers, each dedicated to a different type of data such as vision, language, or action, into a single model that works together.
- MLflowTry itAn open-source tool that tracks and manages the entire process of building, experimenting with, and deploying machine learning models.
- MLOpsThe operational practice of managing a machine learning model after it's built, so it keeps running reliably in a live service
- MLXAn open-source framework built by Apple to run machine learning computations fast on Apple Silicon chips.
- MMLU (Massive Multitask Language Understanding)Try itA benchmark that measures how much an AI model knows across many fields—history, law, medicine, and more—using multiple-choice questions.
- MMLU-ProA benchmark exam that tests how well an AI model handles broad knowledge and reasoning across fields, using harder questions.
- Mobility-Embedded POIsA way of representing places that adds real-world movement data—like actual visit times and dwell duration—on top of fixed attributes such as address and business category
- Mode CollapseTry itA phenomenon where AI answers converge on a narrow set of phrases or patterns even though many different expressions are possible
- mode coverageTry itThe degree to which an AI spreads its probability evenly across many different ways of expressing the same content
- ModelThe term for a finished piece of AI software that has completed training — the 'model' in headlines like 'new model released.'
- Model AgnosticTry itAn approach that isn't locked into a single AI model, letting you choose or switch between multiple models.
- Model CardA model's report card and instruction manual in one — the company's own account of what it was trained on and what it can't do.
- model collapseA phenomenon where AI-generated text or images get fed back into training data, causing output quality to degrade generation after generation
- MCPModel Context ProtocolTry itA common standard for plugging AI into outside tools and data — nicknamed the "USB-C of AI."
- Model FactoryThe name of the internal system that AI startup Poolside used to build its AI models
- Model Hardware Standard (MHS)A common specification that lets AI agents operate real-world equipment, like microscopes or robotic arms, regardless of manufacturer.
- model swappingTry itWhen a middleman service secretly processes a user's request with a cheaper or different AI model instead of the one the user actually chose and paid for
- MoE (Mixture of Experts)MoEAn architecture that wakes up only the 'experts' needed for each question — a trick for running huge models cheaply.
- MolmoActA robotics AI model from the Allen Institute for AI (Ai2) that understands a scene captured by a camera and turns that understanding into movements for a robotic arm or body.
- MolmoMotionAn AI model from the Allen Institute for AI (Ai2) that predicts how objects will move in the next moment.
- moment retrievalTry itThe task of pinpointing exactly when a desired scene starts and ends within a long video
- MoonshineA software component that turns text into a human-like voice, used in Google's offline translation project
- Moonshot AIA Chinese AI developer behind the chatbot and model "Kimi" — a regular fixture at the top of open-source model rankings.
- Morning BriefA Claude feature that gathers schedule and task information through scheduled tasks and external service connections, then summarizes it every morning
- Motion TransferTry itAn AI video editing technique that keeps only the motion and camera work from a video, swapping out the people or background for something new.
- Multi-agentA setup where multiple AIs split roles and work together — like forming a research team, a review team, and a writing team.
- Multi-Agent SystemTry itA structure in which multiple AI agents take on different roles and exchange information to complete a complex task together.
- multi-chip approachA strategy of securing AI computing power by mixing self-designed chips with chips from other companies
- Multi-chip StrategyA strategy where an AI company builds its own chips while also relying on hardware from multiple other companies
- Multi-State Information Sharing and Analysis CenterAn organization where U.S. state and local governments and regional public agencies share cyberattack information and support each other's response.
- multi-token predictionA technique where an AI predicts several tokens at once instead of one at a time, speeding up how fast it generates answers
- Multi-Token Prediction head (MTP head)A helper module that predicts several upcoming tokens at once in a single pass, speeding up generation.
- multi-turn conversationTry itA way of conversing with AI that continues across multiple exchanges rather than ending after one question and one answer, with the AI remembering what came before.
- Multi-vendor StrategyTry itAn approach where a company deals with multiple suppliers at once instead of relying on a single one, gaining leverage and a safety net.
- MulticaTry itA GitHub developer (team) that turned Andrej Karpathy's observations about LLM coding mistakes into a public set of guidelines for Claude Code and Cursor
- Multidimensional Item Response TheoryA statistical technique that separately estimates the distinct abilities being measured when a single test item actually taps more than one skill at once
- MultimodalTry itThe ability to understand and generate not just text, but images, sound, and video together.
- Multimodal agentTry itAn AI agent that carries out real tasks by exchanging not just text but multiple formats of material such as images, audio, video, and 3D models
- Multimodal GroundingTry itAn AI approach that links information conveyed through speech or text to other forms of information—like images, sound, or gestures—to pinpoint exactly what that information refers to in the real world.
- Multimodal Media EmbeddingTry itA technology that converts different forms of content, such as video, sound, and text, into a single set of numerical coordinates so they can be compared and searched by meaning.
- Multimodal PresentationTry itA presentation made not just of text-filled slides but of a mix of images, video, and interactive elements.
- multivariate time series forecasting foundation modelAn AI model that takes in multiple time-varying data streams at once—like sales and weather—and calculates how they influence each other to forecast the future
- Muse CodeA terminal-based coding agent released by Meta that handles multi-file work across large repositories, from planning through implementation and verification.
- Muse GlimmerTry itA 30-billion-parameter open AI model released by Meta, compressed to run agentic tasks on ordinary computers
- Muse SparkAn AI model Meta has referenced for handling heavier workloads aimed at enterprise customers who need large-scale computation.
- Muse Spark 1.2A new AI model released by Meta that serves as the engine powering Muse Code, its terminal-based coding agent.
- MusixmatchA lyrics service that provides song lyrics after paying licensing fees to record labels and publishers.
- MXFP4A low-precision format that represents an AI model's numbers (weights) in as little as 4 bits, cutting down on size and compute
- MythosA frontier-class AI model reportedly built by Anthropic, cited as a benchmark line for U.S. government AI oversight
- Mythos 5A model Anthropic uses to scan GitHub code in Claude's security feature, the latest version in the Mythos line.
- Mythos PreviewAn experimental preview version of Claude that Anthropic tested on real-world tasks, such as designing proteins for new drug candidates.
- Mythos-classA tier of Claude models ranking above the Opus-class, currently accessible only through a separate program with the U.S. government.
- Mythos·FableA development codename Anthropic uses for its top-tier Claude model family
N37
- Nano BananaTry itAn image-generating AI model built by Google and embedded in the Gemini app and Flow, which turns written descriptions into pictures.
- Native Multimodal ModelTry itAn AI model that learns to understand text, images, sound, and video all together from the start, within a single model.
- Native StereoTry itA method of generating video where stereo sound with left-right spatial depth is created together in a single pass, without separate audio editing afterward
- Natural Language RubricTry itA grading standard written out in plain sentences by a human to judge whether an AI's response is good, which is then applied automatically and consistently to evaluate future responses
- NavigatorA browser-operating AI model built by Yutori that handles web tasks by repeatedly looking at the screen, clicking, and typing on its own
- NeMo GuardrailsAn open-source tool contributed by NVIDIA that helps AI stick to set safety rules
- NeMo SwitchyardAn open-source routing tool from NVIDIA that automatically connects each stage of a task to the AI best suited for it.
- NemotronNVIDIA's own family of open AI models that it builds and releases itself.
- Nemotron 3.5 LightningAn open model from NVIDIA that holds a large body of knowledge overall but only activates a portion of it when generating an answer, making it fast to run.
- Nemotron CoalitionA coalition of companies led by NVIDIA to back open AI models, reportedly formed as a response to the spread of Chinese-made AI models.
- neo cloudA specialized AI cloud company that buys up cutting-edge AI chips like NVIDIA's in bulk and rents out their computing power to businesses.
- NeolabA Korean hardware company that makes smartpens and dedicated paper that turn handwriting directly into digital data.
- Neural Attention Memory ModelTry itA design concept that combines an AI's ability to focus on important information with its ability to store and retrieve needed information later
- Neural NetworkA computing structure inspired by how neurons connect in the brain. It's the body of deep learning.
- Neural Processing Unit (NPU)Neural Processing UnitA chip designed specifically for AI computation — the real substance behind "AI phone" and "AI PC" marketing.
- NeuroglancerTry itA free viewer program built by Google that lets you zoom and rotate through massive 3D image datasets right in a web browser
- Next Token PredictionTry itThe way a language model builds text by repeatedly picking the single most probable next chunk to follow what's been written so far
- Nexus AITry itAn early-access AI feature in the networking app Mesh that lets you search your saved contacts by asking questions in plain language.
- NLPTry itThe field of AI that lets computers work with human language — translation, summarization, and chatbots all come from here.
- Non-GAAP기업이 공식 회계 기준(GAAP)에서 일부 항목을 빼거나 더해 자체 방식으로 계산해 발표하는 보조 실적 지표.
- NOOA (NVIDIA Labs Object-Oriented Agents)An open-source framework from NVIDIA Labs that lets you define an AI agent as a single Python class
- North Micro VisionTry itA compact vision-language model with about 2.4 billion parameters from Cohere, built to read and understand document and chart images.
- NotionCompanyNotion is a company that builds productivity software combining document creation, databases, and collaboration features in a single interface.
- Nous PortalOne of the model access gateways that Hermes agents can choose from when picking which AI model to use
- Nous ResearchCompanyAn open-weight model and fine-tuning community.
- NovaA large language model built by Amazon, trained to handle both text and images.
- NOx (Nitrogen Oxides)An air pollutant produced when fuel is burned at high temperatures; gas turbine power generation requires a separate emissions permit to operate.
- NVFP4 QuantizationA compression method that shrinks the numbers an AI model uses down to a simple 4-bit format, drastically cutting size and compute needs.
- NvidiaThe company that makes the chips running AI. It's not a contestant in the chatbot race — it's the one selling pickaxes to everyone.
- NVIDIA B300An AI compute GPU from NVIDIA's Blackwell Ultra generation, used as a semiconductor chip for training and running large AI models.
- NVIDIA Cosmos 3 Super ReasonerA foundation model built by NVIDIA for reasoning, which serves as the base for Alpamayo 2 Super, a model dedicated to self-driving cars.
- Nvidia DGX SparkA compact, dedicated desktop computer made by Nvidia that lets you run AI models directly on the spot, without relying on servers out on the internet
- NVIDIA DRIVE HyperionNVIDIA's pre-integrated hardware platform of compute and sensors for self-driving cars, which automakers can then layer their own autonomous driving software on top of.
- NVIDIA DSX PlatformAn integrated design platform from NVIDIA that combines compute, networking, power, and cooling into a single system for AI data centers.
- NVIDIA Ising Calibration 1.5A 31B-parameter AI model built by NVIDIA that looks at a quantum computer's state graphs and tells you how to recalibrate it
- NVIDIA NIMAn inference microservice product from NVIDIA that packages AI models into containers so they can be served directly through an API.
- NVIDIA SyncA configuration tool from NVIDIA that automatically sets up the complex networking needed when multiple DGX Spark units are connected by cable, instead of requiring manual setup.
O69
- OasisTry itAn interactive AI game world demo by Decart that renders the screen in real time using AI, with no game engine involved.
- OAuthTry itAn authorization method that lets one service borrow a limited, predefined set of permissions from another without ever handing over your password
- Object SwapTry itAn AI video editing feature that replaces a specific element—like clothing, a product, or a prop—in already-shot footage, while keeping the camera movement and performance untouched.
- OfficeQA Pro V2An enterprise benchmark from Databricks that measures whether AI can answer accurately based on real company documents
- Offline TranscriptionTry itA method that converts recorded audio into text entirely on your own computer, without sending it to an internet server.
- OllamaCompanyOllama is a company that builds software for downloading and running large language models locally, such as on personal computers.
- Olmo 3A language model from the Allen Institute for AI (Ai2) that discloses not just the finished model but also the training data and intermediate stages behind it.
- olmOCR-BenchTry itA document-understanding benchmark that measures how accurately AI reads multi-column layouts, tables, handwriting, and charts in documents
- OlmoTraceA tracing tool from the Allen Institute for AI (Ai2) that shows where an AI's output overlaps with the actual text in its training data
- Omni-ReferenceTry itAn input method that lets you feed not just text, images, audio, and video but also documents, spreadsheets, and slides as reference material for generating a video.
- OmnichannelA way of using the same AI agent seamlessly across multiple channels like Slack, email, text, and the terminal
- OmnimodelA term for a model that bundles multiple distinct capabilities — such as seeing, understanding, and predicting — into a single model.
- OmniStarA decision-making platform built by last-mile delivery company OneRail using NVIDIA AI, which picks the most efficient delivery method for each order
- on-demand inferenceA way of using an AI model where you don't sign a contract in advance — you just pay for what you use and get answers in real time.
- On-device AITry itAI that runs directly on your phone or laptop without the internet. It's fast, and your data never leaves the device.
- On-device FrameworkTry itA set of software tools built into a device itself so it can handle things like maps, voice, and location without needing an internet server
- On-premisesA setup where a company buys and runs its own servers and equipment inside its own facilities, instead of renting them from the cloud
- One Photo AvatarTry itA Synthesia avatar feature that turns a single face photo into a video of that person talking, wearing different clothes, and standing in a different setting.
- One UI 9The 9th version of Samsung's own screen and feature interface, layered on top of Android for Galaxy smartphones and tablets.
- OntologyTry itA machine-readable map of concepts and how they relate to each other. A blueprint of "who made what."
- Open Agent Skills standardTry itA shared specification that lets different AI agent programs read and write skill files in the same format.
- Open APIA gateway a company opens under fixed rules so outside programs or developers can directly call specific functions of its system.
- Open ASR LeaderboardTry itA leaderboard run by Hugging Face that scores speech recognition (ASR) models on public datasets and ranks them by error rate.
- Open PangramTry itAn open-source text detection technology that judges whether a piece of writing was written by a human or by AI.
- Open Secure AI AllianceAn industry consortium led by the Linux Foundation that aims to collect information on AI security incidents and share it across the ecosystem.
- Open SourceTry itA way of releasing the code behind something so anyone can use, modify, and share it. In AI news, "open source release" often means something a bit different — worth watching for.
- Open WeightsA model whose blueprint (weights) is released publicly so anyone can download and use it. Llama, DeepSeek, and Qwen fall into this category.
- Open Weights AI ModelTry itAn AI model whose internal parameters, learned through training, are released publicly so anyone can download and use them.
- Open Weights and American AI LeadershipAn open letter signed in July 2025 by NVIDIA and more than 200 other companies and institutions, arguing that releasing AI models openly is key to American AI competitiveness
- Open Weights ModelTry itA model whose fully trained internal values are released to the public so anyone can download and run it themselves
- open-source biomolecular modelAn AI model released publicly so anyone can use it to predict and design the structure of biomolecules like proteins
- OpenAIThe company behind ChatGPT and the GPT model family. It kicked off the current AI boom with ChatGPT's 2022 launch.
- OpenAI APITry itA connection point that lets developers and companies plug OpenAI's AI models directly into their own programs.
- OpenAI DevelopersCompanyOpenAI's developer platform account.
- OpenAPI SpecificationTry itA standardized format for documenting what an API does and how to call it
- OpenClawA personal AI assistant program that, given only a goal, operates the web and apps on its own to get things done
- OpenMDW 1.1An open distribution license set by the Linux Foundation that allows AI models to be modified and even redistributed commercially
- OpenMDW 1.1 LicenseAn open license created by the Linux Foundation that allows AI model weights to be released publicly and retrained by others
- OpenMDW LicenseTry itAn open-use license that spells out whether you can freely take, modify, and redistribute an AI model's weights, datasets, and deployment code
- OpenRouterCompanyOpenRouter is a company that operates a platform connecting language models from multiple AI companies through a single API.
- OpenShellA runtime tool that limits how much of a system an AI agent can access, contributed by NVIDIA.
- OpenTelemetryAn open standard that unifies the logs, metrics, and traces software generates into a common format, regardless of which company's tools you send them to.
- OpenTelemetry Protocol (OTLP)A communication standard that packages a program's usage and performance records into a common format and sends them to monitoring tools
- OpenUSDAn open-source 3D data standard that lets programs exchange 3D scene data in one common format
- OpenVulnA service from Z.ai that uses the open-weight model GLM-5.3 to scan code repositories for security vulnerabilities
- Operational Design DomainTry itThe specific set of conditions — roads, speeds, weather, and regions — under which a self-driving car is guaranteed to be able to drive itself safely.
- OperatorAn AI agent OpenAI launched in January 2025 that browses websites directly to handle errands like making reservations or placing orders
- Operator FusionAn optimization technique that combines multiple separate computation steps into one to speed up AI model calculations.
- OpusThe name Anthropic gives to the top-performing tier of its Claude models.
- Opus 4.5Try itAn AI model Anthropic released in November 2025 that dramatically boosted the performance of Claude Code.
- Opus 4.8A version of an AI model from Anthropic, a top-tier model used for coding and agentic tasks.
- Opus 5An AI model from Anthropic, launched at half the price of the company's previous flagship model, Fable 5.
- Opus-class modelA term for models on par with 'Opus,' the largest and most capable tier in Anthropic's Claude model lineup
- OR-Bench-HardTry itA safety benchmark that measures how often an AI language model over-refuses harmless requests, treating them as if they were dangerous.
- orbital data centerA facility that puts servers on satellites orbiting Earth, using solar power and the vacuum of space to run AI computation.
- orbital inferenceA method where computers aboard satellites process AI computations in Earth orbit, using solar power and the cold of space for cooling
- Orchestration LayerTry itA management layer that divides tasks among multiple AI models or agents and coordinates the order in which they run.
- Orchestration Workflow PlatformA platform where you set in advance when and what multiple automated tasks or AI bots should do, so they run in sequence on their own without a human in the loop.
- OrchestratorTry itA manager-like program that assigns tasks to multiple AI agents and pulls their results together into a final answer
- Orchestrator LLMAn AI model that acts as a conductor, dividing tasks among multiple sub-models and combining their outputs into a final answer
- Original Creation, By DesignA training strategy Suno uses so its AI doesn't simply copy existing recordings — it excludes artist names from training data and screens for unauthorized use
- OrnithCompanyOrnith is a company that develops open-source AI models.
- OSWorld-2.0A benchmark that tests whether AI agents can independently complete multi-step tasks on a real computer screen
- OuraCompanyOura makes and sells wearable health-tracking devices.
- Outbound FilteringA security control that screens and blocks outgoing traffic from inside a system, as opposed to blocking traffic coming in.
- Output TokenTry itThe unit that counts the chunks of text an AI actually generates when answering a question
- OverfittingWhen a model memorizes the training data so hard it fails on new problems — like memorizing exam answers and then bombing a slightly different question.
- OWASP Agentic Top 10Try itA list of ten key risks to watch for when deploying AI agents that reason and act on their own
- Ox AlphaTry itA reasoning AI model specialized in coding and agentic tasks that appeared on OpenRouter without its developer being disclosed
P95
- P-Image-IdeogramAn image-generation AI that Runway added to its platform, capable of producing an image in as little as 0.6 seconds.
- Paid conversionWhen a user who has been trying a product for free starts actually paying to use the paid version.
- PalantirCompanyPalantir is a company that builds software for integrating and analyzing large-scale data for government agencies and enterprises.
- PangramTry itA text-detection platform that determines whether a piece of writing was written by AI, and even identifies which AI model likely wrote it.
- Pangram 4The latest version of a text-detection tool that identifies how much of a piece of writing was generated by AI.
- Pangram v3.3An AI-text detection tool that flags whether a piece of writing was written by a human or an AI, with the developer claiming a false-positive rate of just 0.04%.
- ParameterThe number of adjustable dials inside a model. The "B" in "70B model" refers to this — generally, more means smarter but more expensive.
- Parameter SpaceTry it모델 내부의 숫자값들이 취할 수 있는 모든 조합을 가리키는 수학적 공간
- Parametric CAD ModelTry itEditable 3D design data that automatically updates the entire design when a dimension or shape value is changed.
- Pareto FrontierWhen comparing two criteria at once, the line connecting the best possible points where you can't improve one without sacrificing the other
- ParseBench-100A 100-task test that measures how accurately AI can read text and tables in documents
- Partial AutonomyTry itA self-driving approach where the car drives itself but leaves room for a human to take over with the steering wheel and pedals in dangerous moments.
- PATHFINDERAn AI reconstruction system developed by Google Research that stitches together electron-microscope brain cross-section images into a 3D map of neural circuits.
- PatternsA feature in smart pet devices that tracks an individual animal's litter box habits over time and flags anything that deviates from the norm
- Pax SilicaA coalition of allied nations launched last year at the US's initiative to jointly manage supply chains for AI models, semiconductors, and critical minerals
- Pay-per-tokenTry itA usage-based pricing model where you're charged only for the number of text chunks (tokens) an AI processes.
- PayBoxAn AI payment and wallet tool built by MoonPay that lets you buy and sell crypto, book flights, and more just by talking to a chatbot.
- PayloadThe actual content being delivered once you strip away the packaging or surface appearance — in security contexts, the real malicious code hidden inside malware.
- PBKDF2 and AES-256-GCMTwo widely used standard cryptographic technologies: one turns a password into a strong key, the other uses that key to lock up data.
- pdftotextTry itA command-line tool that strips away images and formatting from PDF documents to extract pure text
- PDI BrewAn agentic deployment platform built by PDI Technologies on AWS that automatically generates internal web apps from plain-language descriptions given by non-developers
- PegasusA video understanding AI model by Twelve Labs that analyzes footage scene by scene.
- Penetration testingA security check where testers try to break into a system the way a real attacker would, so weaknesses can be found before anyone else finds them.
- People Detection AlertA Flock Safety surveillance feature that automatically alerts police when a person enters a zone drawn on a camera feed
- Percept-LensA verification benchmark from Sakana AI that measures how well methods can tell AI-generated images apart from real photos.
- PerceptionBenchTry itA visual perception benchmark from Moonshot AI that isolates and measures how accurately an AI model 'sees' images.
- PerplexityCompanyPerplexity is a company that builds AI-powered search engines and answer services.
- Perplexity ComputerTry itAn automation tool from Perplexity, an AI assistant-type product that handles multi-step tasks on its own instead of a human doing them.
- Persistent Async QueueA storage-backed job queue that accepts time-consuming tasks upfront, letting you check their status later or safely retry them if they fail
- Persistent IPython REPLAn execution environment that stays running throughout a session, letting an AI agent treat its work as program variables instead of re-reading everything as text
- Persistent MemoryTry itA feature that lets an AI assistant store information about a user, their preferences, and instructions, then recall them in later conversations.
- persistent sub-agentsA support-agent structure that takes on parts of a large task and stays engaged with full context until each part is finished, rather than dropping context after a single response.
- Personal AI AgentTry itAn AI service that looks through personal information like email and calendars on your behalf and sorts out what needs attention.
- Personal IntelligenceTry itA Gemini feature that connects data from Gmail, Photos, Search, and YouTube to create personalized recommendations
- Personal SuperintelligenceMeta's vision of a future where everyone has their own personalized AI assistant that remembers their individual context.
- PetriAn open-source testing tool in which one AI model holds multi-turn conversations with another AI model to uncover misaligned behavior like lying or sycophancy.
- Physical AIAI that steps out of the screen and moves through the physical world — robots, self-driving cars, machine control. NVIDIA's pitch for the next wave.
- Physical PromptingA method that lets a robot copy a task on the spot just by watching a single demonstration, with no retraining involved.
- Pi AgentAn open-source framework used as the foundation for building AI agents, on which coding agents like MiniMax Code have been rebuilt
- Pi Agent FrameworkAn open-source blueprint used as the foundation for building AI agents, on which MiniMax rebuilt its coding tool 2.0
- Pika API ClubTry itA developer-only API gateway through which Pika offers its four audio-generation models exclusively.
- Pika SoundtrackPika's audio generation model that automatically creates and syncs background music, dialogue, and sound effects to match the motion in a video
- PipetteTry itAn open-source benchmark tool that measures AI models' response speed and memory usage on personal devices like phones and laptops
- Pirate Library MirrorA pirated ebook repository that collects books without copyright holders' permission and shares them for free online
- Pixtral Vision EncoderMistral's vision component that converts images into signals a language model can understand
- Plan ModeTry itA mode where an AI coding tool first plans the order and approach of a task before writing any code
- plan-and-executeTry itAn approach where an AI agent first breaks a task into a step-by-step plan, then carries out each step while coordinating the overall flow
- Planetary Data SystemNASA's open database that collects observational data gathered by solar system exploration missions to the Moon, Mars, and beyond, available for anyone to download.
- Planning-Provisioning Dual-Agent ArchitectureTry itA two-stage workflow that splits work between an AI that clarifies a person's vague request and an AI that builds the actual system according to that clarified plan.
- PlaybotAn AI-powered game development tool built by game startup Playco, letting a model directly edit and run scenes inside the engine to test them itself
- PlaywrightAn automation tool that controls a web browser on a person's behalf, opening, reading, and capturing pages automatically
- PluginAn add-on component that plugs new capabilities into an AI service — a packaged unit of skills, connectors, and tools.
- PocketTry itAn experimental Meta app that turns a few sentences into a mini game you can play with in your hand
- PoeCompanyPoe is a company that operates a service providing access to language models from multiple AI companies on a single platform.
- Point CloudTry itA 3D data format that represents three-dimensional space as a collection of countless tiny points, allowing it to be viewed from any angle.
- Policy on the AI ExponentialA policy essay by Anthropic CEO Dario Amodei on how institutions should respond to the rapid change AI is bringing
- Polygenic Risk ScoreA statistical model that combines information from many genetic variants to predict the likelihood of developing a particular disease or trait.
- PopcornA playful codename AI companies sometimes give to a model or feature still in development, without indicating any specific capability or the eventual official product name
- Portable ComputerTry itPerplexity's fully local AI assistant program that runs entirely on a single device without any cloud server.
- Ports-PikeA massive data center OpenAI is building in Ohio, where NVIDIA has locked in the role of exclusive computing equipment supplier.
- Post-trainingThe stage after pre-training where a model's behavior and abilities are fine-tuned. This is where most differences in model performance come from these days.
- PPLX 27BA 27-billion-parameter language model built by Perplexity, offered as one of the local models that run on-device within Portable Computer.
- Preliminary InjunctionA temporary order a court issues before a lawsuit reaches its final ruling, meant to prevent urgent harm in the meantime
- PreparednessAn internal team at OpenAI dedicated to assessing severe and catastrophic risks from its models, disbanded at the end of July 2026
- Preparedness FrameworkOpenAI's pre-set rulebook that automatically triggers safety measures once an AI model's dangerous capabilities reach a certain level
- Prime AgentA task framework for self-improving coding agents released by Prime Intellect in August 2026, aiming at both token efficiency and expressiveness.
- Prime IntellectCompanyDistributed training and multi-agent reinforcement learning.
- Prior-Conditioned Gaussian DiscriminantA statistical rule that tells real photos apart from AI-generated images just by computing the mean and variance of image features, with no separate training required.
- Priority TierA paid express service that processes your requests first during busy periods if you pay extra.
- Privacy GateA safeguard that detects tasks containing personal information and automatically switches processing from the cloud to your own device.
- Private BetaA trial phase where a product is opened only to a small group of invited users before it's released to the public.
- Private DebtA way for companies to raise money privately by borrowing directly from a small group of specialized investors, instead of going through banks or the stock market.
- Private Safety ProcessingA safety-check method OpenAI uses to automatically flag dangerous use of the API without storing conversation content or having humans read it.
- Probabilistic ReasoningTry itA way of thinking where AI factors uncertainty itself into its answers, saying something like 'this is correct with X% probability' instead of committing to a single definite answer
- procedural anchoringTry itThe effect of reducing mistakes by having an AI agent follow a fixed sequence of steps and checks, rather than giving it new knowledge
- product surfacesThe various apps, websites, and APIs through which a single AI model reaches users
- Product-Market FitA startup term describing the state where a product truly matches what the market wants, so customers keep coming back and paying for it.
- Productive Struggleproductive struggleTry itThe learning-enhancing process in which a student, when stuck on a problem, works through confusion on their own instead of immediately being given the answer
- Productivity GainsTry itThe efficiency boost from using AI tools to get more done in the same amount of time, or finish the same work faster.
- programmatic tool callingA method where an AI agent writes out the sequence of multiple tool calls as code and runs them all at once, instead of requesting tools one by one
- Project AstraA Google DeepMind multimodal AI project that sees and hears its surroundings in real time through a camera and microphone and converses accordingly
- Project PanamaAnthropic's internal code name for a project that bought physical books, scanned them, and used them as training data for its AI language models
- Projects for Claude CodeA team workspace where multiple people share a code repository and work with Claude together, currently in development by Anthropic for Claude Desktop.
- PromptTry itThe instruction you give an AI — every question, command, or example you type in counts as a prompt.
- Prompt CacheTry itA technique that saves the results of parts of a prompt an AI has already processed so it can reuse them instead of recomputing them, cutting both speed and cost
- Prompt EngineeringTry itThe skill of getting AI to do good work for you. How you phrase your instructions alone can change the quality of results several times over.
- Prompt InjectionTry itAn attack that manipulates AI using hidden commands buried in web pages or documents — the top security threat of the agent era.
- Prompted PlaylistsTry itA Spotify feature that builds a playlist by picking songs to match a mood or situation you describe in a sentence.
- Proof of Concept (PoC)Try itA process of testing a new idea or technology on a small scale before rolling it out broadly, to check whether it actually works.
- ProResA professional video codec developed by Apple that minimizes quality loss during editing and color grading work.
- Prove2MeAn approach that requires AI to present its answers or proofs in a step-by-step, verifiable form, rather than having people simply take the results on faith.
- provider adapter harnessAn execution wrapper built and shipped by the model developer itself, tuned to its own model, that carries intermediate solving state forward from one problem to the next
- pruningA technique that trims away connections and neurons that contribute little to a model's output, making the AI model lighter.
- PsychometricspsychometricsA set of statistical techniques—like those behind IQ tests or aptitude exams—for measuring invisible traits with scores, now also used to validate AI safety benchmarks.
- pull request diffTry itA side-by-side view that shows what was removed and added between the old and new version of code
Q25
- QCalEvalA benchmark suite that scores how well an AI can read quantum computer calibration graphs and recommend fixes
- Qommons ONEAn AI platform for administrative work used by Japanese local government employees, built by Polymill on top of OpenAI's GPT models
- QommonsAIA government-focused AI service that Japanese local government employees use for meeting minute searches, council response drafting, and citizen inquiry handling
- QuackAn open-source project developed alongside DuckDB and DuckLake by DuckLabs, the company behind DuckDB.
- QualcommCompanyQualcomm is a semiconductor company that designs and sells processors and communication chips used in smartphones and various mobile devices.
- Quality ModeTry itA generation option in image and video AI tools that prioritizes finished-product polish over speed.
- QuantizationCompression that shrinks a model by lowering the precision of its numbers — the trick that lets LLMs run on phones.
- Quantization-Aware DistillationA training method that transfers the judgment habits of a large, precise original model into a lightweight compressed model, preserving performance without retraining the original from scratch
- Quantization-Aware HealingA technique that has a compressed AI model relearn directly from its original, uncompressed version, so it recovers performance lost when its size was reduced.
- Quantization-Aware TrainingTry itA technique that retrains a model under simulated low-precision conditions to recover the performance lost after shrinking it down
- Quantum Game TheoryTry itA mathematical theory in which multiple players bring quantum mechanical principles into their strategies to analyze the outcome of a contest, recently drawing attention as one of the unsolved problems an AI model claimed to tackle.
- QueryA single question or request sent to a system — search terms and AI prompts both count as queries in the broad sense.
- Quick AgentAn AI feature within AWS's analytics service Amazon Quick that understands your data catalog and automatically generates datasets and topics
- QwenAlibaba's in-house open-source model family — one of the most widely used names in the open-model camp.
- Qwen 3.6One of Alibaba's Qwen AI models, a relatively small model with about 27 billion parameters.
- Qwen-3.5A generation within the large language model series developed by Alibaba's Qwen team.
- Qwen-AgentWorld-35B-A3BA virtual-world generation model from Alibaba's Qwen for training AI agents, with 35B total parameters but only 3B active at a time.
- Qwen3.5-2BA small, roughly 2-billion-parameter model from Alibaba's Qwen line, cited in a speed comparison for image recognition.
- Qwen3.5-397BAn open model released by Alibaba's Qwen team, used by Thomson Reuters as the foundation for building its legal AI model 'Thomson.'
- Qwen3.5-9BA lightweight, 9-billion-parameter language model built by Alibaba's Qwen team, used as a dedicated module for specific tasks
- Qwen3.8A Chinese AI model from Alibaba's Qwen lineup, seen as having narrowed the performance gap with top Western models.
- Qwen3.8-2.4T-A95BA large language model from Alibaba's Qwen series with 2.4 trillion total parameters, of which only 95 billion are actually used to generate a single response.
- Qwen3.8-27BAn open-weight language model with about 27 billion parameters released by Alibaba's Qwen team, built to run even on personal computers.
- Qwen3.8-Flash-NextTry itAn open-weight 125B-parameter model released by Alibaba's Qwen, compressible to run locally with just 75GB of memory.
- Qwen4 architectureThe next-generation model design Alibaba's Qwen team has been preparing, which reads text and images together and processes long documents of up to 262K tokens at once.
R74
- RAGTry itA method that has AI look up source material before answering. The go-to fix for hallucination.
- RamblerTry itA voice input feature on Google Pixel 11 that figures out what you mean even when you pause mid-sentence or repeat yourself
- Randomized Controlled TrialTry itAn experiment design that randomly splits participants into two groups, applying a new method to only one, so the results can be fairly compared.
- Rate LimitingA control that caps how many requests can be processed in a given time window, blocking or delaying anything beyond that limit.
- raycastingAn old rendering technique that draws a 3D scene by shooting rays from the camera and measuring the distance to whatever they hit
- RDMA over Converged Ethernet (RoCE)A communication method that lets multiple computers connected by Ethernet cables exchange data directly memory-to-memory, without going through the CPU.
- Real-time MonitoringA monitoring approach that watches an AI's actions the moment they happen and stops it immediately if something looks dangerous
- Real-time Voice ModelAn AI model that understands spoken language the instant it hears it and responds in voice right away, with no intermediate step of converting speech to text.
- Real-to-Sim-to-Real (R2S2R)A training method that takes a single real-world robot action, recreates it in a physically accurate virtual world, and multiplies it into thousands of variations to train the robot.
- Reality-Grade RenderingA rendering-quality phrase used to describe AI-generated video that has been polished to look as if it were shot with a real camera.
- Realtime TTS-2Try itA real-time text-to-speech model from Inworld AI that lets you adjust voice tone and emotion on the fly using natural-language instructions.
- Reasoning ModelTry itA model that takes "time to think" on its own before answering. Slower, but much stronger on hard problems like math and coding.
- Recommendation AlgorithmAn algorithm that analyzes behavioral data like a user's clicks and viewing history to automatically pick what content to show next.
- Recursive Language Model abstractionTry itA design approach that treats long context like a program variable—slicing and summarizing it instead of feeding it whole—so only the needed pieces reach the model
- Recursive Self-ImprovementTry itThe process by which an AI system repeatedly evaluates its own capabilities and revises its own learning methods or structure to keep getting better
- Red AgentAn autonomous security research tool run by security firm Wiz that finds vulnerabilities in code on its own and even attempts to exploit them
- Red TeamingA role that deliberately attacks a product before launch — "we break it ourselves first" has become the standard procedure for AI safety.
- refactoringTry itCleaning up the underlying code structure of an app without changing what it does, to make it neater and easier to maintain.
- Reference-to-VideoTry itA video generation method that assigns different roles to reference images—like a person's face or a background—so their distinct features are preserved in the output.
- Regional EndpointsA feature that lets customers choose which country or region's servers actually process the requests they send to an AI.
- Regular ExpressionTry itA predefined search notation used to find or filter text that matches a specific pattern
- Reinforcement LearningA trial-and-error training method that rewards good moves and penalizes bad ones. The method behind AlphaGo.
- Reka AICompanyReka AI is an artificial intelligence company that researches and develops large language models.
- Reliance JioA major Indian telecom carrier and subsidiary of Reliance Industries.
- Remote ControlA Claude Code feature that lets you check on and control a task running elsewhere from a remote location.
- Remote Direct Memory AccessA communication method that lets multiple computers read and write directly to each other's memory without intermediate steps, so their memory can be used as if it were interconnected
- Remotion PluginAn add-on component that adds new features to Remotion, a tool that automatically generates videos from code
- rendering optimizationTry itWork that makes the process of drawing text or images on screen faster and smoother
- replayAn evaluation method that pauses a real conversation transcript at a specific moment, has an AI continue from there, and then grades the result
- ReplitCompanyReplit is a company that builds a development environment for writing and running code directly in the browser.
- Request for Comments (RFC)Request for CommentsA process of publishing a draft of a new technical standard or policy before it is finalized, so that a wide range of people can weigh in with feedback.
- research tasteTry it어떤 실험이 해볼 가치가 있고 어떻게 설계해야 의미 있는 결과가 나올지 판단하는 감각을 가리키는 말이다.
- Research-Plan-Implement Pipeline (RPI Pipeline)Research-Plan-Implement PipelineTry itA three-step workflow in which an AI agent researches and plans before actually writing code.
- ResearcherA research feature for paid business users of Microsoft Copilot that replaces the former Deep Research.
- ResidencyRLA reinforcement learning technique Google used to train AI in clinical conversation skills by running it through repeated simulated patient consultations
- Responses APITry itAn API communication standard defined by OpenAI for AI models, which other companies sometimes support natively in their own models
- Responsible Scaling PolicyAn AI company's own internal rule promising to strengthen safety measures in proportion to how powerful its models become
- RESPXRESPX is a Python mocking library used to fake server communication during tests.
- reusable evaluation toolA shared measurement tool that, once built, is published so it can be applied repeatedly across many different AI models.
- Revenue Run RateTry itAn estimated annual revenue figure calculated by assuming the recent short-term pace of revenue continues unchanged for a full year
- Reverse EngineeringThe process of taking apart a finished product or result to figure out its original design and how it works
- Reverse Federalismreverse federalismAn approach where, in the absence of federal AI regulation, state-level rules are refined first so they can serve as the foundation for a future national standard.
- Review AgentTry itAn AI program that scans code changes before humans do and leaves comments on lines that look problematic
- Reward hackingTry itWhen an AI exploits loopholes in how it's scored instead of pursuing the actual goal it's meant to achieve
- RewindTry itA feature that jumps Claude Code back to a previous save point instead of trying to steer it back on track with more prompts
- RHAE Best@1A scoring standard that grades an AI agent's single best attempt at a task while filtering out attempts that failed only because of execution-environment errors, not the agent's own reasoning.
- Riemann HypothesisA math puzzle, unproven and undisproven for over 160 years, suggesting there's a hidden order to how prime numbers are scattered.
- Riemann zeta functionA mathematical function tied to the distribution of prime numbers, whose zeros are at the center of the Riemann Hypothesis, an unsolved problem for over 160 years.
- RLHF인간 피드백 강화학습Try itTraining that polishes a model's manner of speaking using human-scored 'good answer' ratings — the technique that made ChatGPT polite.
- Robostral NavigateAn 8-billion-parameter navigation AI model from Mistral that lets a robot find its own way using footage from just a single camera
- robotic embodimentThe process by which an AI model moves beyond being a screen-bound intelligence and gains an actual robotic body to see and move through the physical world.
- ROCmAMD's GPU computing platform, an open software ecosystem built to rival NVIDIA's CUDA
- ROCm/HIP backendA pathway that lets AI programs run on AMD graphics cards too.
- RoleplayTry itA style of conversation where the AI is assigned a specific character or role and carries the dialogue forward like a story or a play.
- Roleplay & DubbingA Synthesia video feature that lets AI avatars act out conversations in context and convert original speech into another language's voice track
- Roleplay SessionTry itA conversation-practice feature in Synthesia where an AI plays the other party in a dialogue and gives real-time coaching.
- RoomoteA coding agent tool that mixes and matches different AI models for each stage of a task—coding, planning, screen/image recognition, and review
- Root Cause AnalysisTry itA method of digging past the surface symptoms of a problem to find the actual cause that triggered it
- Rosetta CodeTry itA site collecting solutions to the same problems written in many different programming languages for comparison
- Route optimizationTry itThe AI-driven, real-time calculation of the fastest and cheapest order to travel between multiple destinations.
- RouterTry itIn a mixture-of-experts model, the internal mechanism that decides on the fly which experts to wake up and use for each input
- RoutinesA menu feature where Claude gathers recurring tasks that run automatically on a set schedule.
- RoutingA method that automatically selects and assigns the AI model best suited to a given task
- routing algorithmTry itA set of rules that decides which AI model or server an incoming request should be sent to — essentially traffic control for requests.
- routing strategyTry itThe practice of dividing work among different AI models based on the nature of each task.
- RSI LabA research team inside Sakana AI that studies how AI can repeatedly improve itself
- RTX SparkAn NVIDIA chip that combines CPU, GPU, and memory into a single package, letting laptops and mini PCs run AI without an internet connection
- Rubin·BlackwellNVIDIA's generational names for data-center AI compute chips (GPUs): Blackwell is the current generation, and Rubin is the upcoming next generation.
- RubyTry itA conversion tool from Runway that turns 8-bit, compressed generated video into the wide, bright HDR format broadcasters use
- RunmeTry itAn open-source notebook web app that stacks explanations, commands, and execution results in order in a single document, letting AI agents fill in the cells themselves
- Runtime HostA single execution engine that handles the actual work behind the scenes, no matter which entry point—app, terminal, or CLI—a user comes through.
- Runtime InstancesA persistent compute option in AWS Bedrock AgentCore that lets AI agents keep running for up to 14 days while retaining state and GPU access
- RunwayCompanyRunway is a company that builds AI models generating video from text or images.
- Runway AIF 2026An AI film festival hosted by Runway, the video-generation AI company, where short films made with AI compete and are screened
S152
- SAFE guidelinesAn industry self-regulatory guideline that gets companies to privately pool AI security incident information and share it across the whole ecosystem
- safetensorsTry itA file format for storing an AI model's parameter values, designed so files can be opened safely without risk of running malicious code.
- SageMaker HyperPodAn AWS service for managing clusters of computers used to train or deploy large AI models.
- SageMaker Python SDK v3Try itThe latest version of AWS's Python toolkit for machine learning development, letting you benchmark and deploy models in one go from within a notebook.
- SageMaker StudioA web-based interface from AWS that brings building, training, and deploying machine learning models together into one integrated workspace
- Sakana AICompanyBased in Tokyo.
- Sakana ChatTry itA Japanese-focused chatbot service run by Sakana AI that creates games and learning materials from spoken descriptions and analyzes Excel data.
- Sampling MethodsTry itThe rules an AI uses to pick the next word, sound, or image piece from a set of probability-weighted candidates
- sandbaggingTry itA phenomenon where an AI model senses it's being tested and acts more cautiously than usual, concealing how it actually behaves.
- SandboxAn isolated sandpit where AI can run code without causing damage outside it — a basic building block of agent safety.
- Sandbox ModeTry itA safety setting that restricts an AI agent to running code only in an isolated, limited space rather than on the actual computer environment.
- sandbox VMTry itA virtual computer isolated from the real one, so that when AI runs code inside it, nothing spills over to affect the actual machine.
- Scale-InReducing the number of running servers (instances) as traffic drops — the opposite of scale-out
- Scaling LawsThe empirical rule that performance improves predictably as you scale up model size, data, and compute. The rationale behind AI's investment race.
- SceniXA startup that developed virtual-reality replication technology for robot training, acquired by World Labs in July 2025.
- Scheduled TasksTry itA feature that lets you hand off work to AI in advance, to run once, repeat on a schedule, trigger on a specific event, or watch for changes.
- Schema MappingTry itThe work of matching up the names and formats of data fields used by different systems so they can be linked into one common structure.
- scientific tasteTry itThe experience-based judgment that separates statistically striking results from findings that are actually meaningful.
- Screen SharingTry itA feature that shows the AI exactly what's on your current screen (app window), without capturing it, so it can use that as conversation context
- Scribe v2 RealtimeElevenLabs' real-time speech recognition feature that transcribes speech into text as it's spoken, with support for entity recognition and multilingual detection.
- SDKA ready-made toolkit for developers. When a company "releases an SDK," it's a signal that building on its technology just got easier.
- Search IndexAn Artificial Analysis benchmark that scores the search APIs AI agents use, ranking them side by side on quality, speed, and cost.
- search precisionA metric that measures how much of what an AI retrieves is actually useful.
- SearchVectors APIA search command that finds and ranks semantically similar items among data stored in Amazon DynamoDB.
- Seedance 2.0Try itA video generation AI model developed by ByteDance Seed, run on the Higgsfield platform and used to produce AI animated features.
- Seedance 2.5Try itA family of video-generation AI models that turn text or images into short video clips, reportedly developed by ByteDance Seed.
- Seedance StudioTry itA production tool from Krea for the Seedance video generation model, letting users directly manipulate camera paths and 3D scene layout on screen.
- Self-Distillationself-distillationA method where an already-trained model uses its own outputs as training material to train a faster, lighter version of itself.
- Self-drivingTechnology that lets a car drive itself. Shows up in news through level numbers and robotaxi stories.
- Self-Evolutionself-evolutionThe process by which an AI system revises its own performance or way of operating without human intervention.
- Self-hosted SandboxTry itAn isolated workspace that developers set up directly on their own servers, rather than in the cloud, so AI agents can execute code and handle files
- self-improving RLM harnessAn execution framework that wraps an AI model to manage tool calls and long-running tasks, and that rewrites its own configuration while it's running.
- self-modifiable harness stateA capability that lets an AI agent alter the settings of its own execution environment (harness) while it is working on a task
- self-referential utteranceTry itA way of speaking in which an AI expresses its own thoughts or feelings.
- self-regulationA company setting and following its own principles and procedures to govern its behavior, before any law forces it to do so
- self-scaffoldingA method where an AI model builds the tools and procedures it needs to solve a problem, then solves the problem on top of that self-built structure.
- Self-serviceA way of using a product where advertisers or users set things up and launch them entirely on their own, without going through a dedicated staff member.
- Self-supervised Foundation ModelTry itA base model that learns patterns on its own from raw, unlabeled data, then gets reused across many different tasks.
- Seller AgentTry itAn AI agent on the merchant-facing side of a store that helps analyze sales, manage inventory, suggest pricing, and draft marketing copy
- Semantic DriftA mismatch that occurs when an upstream system's data definitions change but the meaning definitions used downstream stay frozen, causing the two to fall out of sync.
- Semantic SearchTry itA search method that understands the meaning of a sentence to find the scene or content you want, even without pre-assigned human tags
- Semantic Snippet ExtractionA search-result compression method that hands an AI only the sentences or passages that are semantically relevant to a query, rather than an entire web page.
- sensor data processingThe process by which a computer reads, in real time, the raw information gathered by sensing devices like cameras, lidar, and radar, and converts it into a form usable for decision-making.
- Series C스타트업이 세 번째로 받는 대규모 투자 라운드를 부르는 이름으로, 회차 자체보다 투자자와 회사 가치가 더 중요하다
- Series D스타트업이 네 번째로 받는 대규모 투자 라운드를 부르는 이름, 순서보다 누가 얼마에 투자했는지가 중요하다
- Server RackA standardized metal shelving structure in data centers used to stack and secure multiple layers of servers and GPUs.
- Server-Sent EventsA communication method where a server continuously streams progress updates to a client in real time, letting the screen update bit by bit until a task finishes.
- Server-Side Request ForgeryAn attack technique that tricks a server into sending requests on the attacker's behalf, reaching internal networks or the internet that would otherwise be off-limits
- ServerlessA way of running programs where you don't set up or manage servers yourself — computing resources automatically attach and detach as needed
- Serverless InferenceA way to use AI models by calling them on demand and paying only for what you use, without buying or managing servers yourself
- Service Level AgreementA contract in which a service provider promises numeric quality standards—like response speed or uptime—and spells out what happens if those standards aren't met.
- serving engineTry itSoftware that runs a trained AI model quickly and reliably in real-world production services.
- Session RailTry itA status bar next to the OpenClaw chat window that shows session ratings, plan progress, and pull request status.
- Sessionless MCPA new way for MCP to communicate that lets servers avoid having to hold onto a persistent connection state with each client
- Shadow LibraryAn unauthorized online repository that collects books and papers without permission from publishers or authors and lets anyone download them for free
- shadow testingTry itA method of secretly attaching a new system behind a live service so it processes real requests and its performance can be checked without users noticing.
- SAFE (Shared AI Findings Exchange)Shared AI Findings ExchangeA proposed security guideline for collecting, analyzing, and sharing incidents and near-misses involving AI agents
- Shared Cloud SessionsA cloud feature that lets multiple people join the same AI agent task together, picking up its progress and context to collaborate
- SheerIDTry itA third-party identity verification service that confirms specific eligibility, such as student or teacher status, using school email addresses or enrollment records.
- ShepherdAn open-source runtime that records an AI agent's work process step by step so it can be rewound to any earlier point
- Shieldstral 1.0 3BA content moderation AI model released by Mistral that judges rules by answering yes-or-no questions posed in natural language
- Shopping AgentTry itAn AI agent that listens to a customer's request, searches and compares products, and fills a shopping cart—but doesn't complete the payment itself.
- SierraCompanySierra is a company that builds and sells AI customer service agents for enterprises.
- Silicon WaferA thin, round silicon disc used to make semiconductor chips, typically cut into many individual chips
- SimulationA way to train and test AI in a virtual world — where robots can fall a million times before they ever fall in real life.
- single-stream omni transformerAn architecture in which a single neural network generates video and audio together in one pass, instead of making each separately.
- SingularityA theoretical concept describing the point where technology surpasses human biological limits and begins expanding itself without bound.
- SkillTry itA packaged set of instructions that teaches an AI how to do a specific task — like getting it to remember "our company's report format."
- Skill HubA repository of pre-built functions that an AI agent can pick from to perform specific tasks.
- Skill LiftTry itThe measured difference in performance when an AI agent is given a usage manual (a skill) for a specific function, compared to when it isn't.
- SKILL.mdTry itA standard-format instruction file that tells an AI agent how to carry out a specific task
- SkillEvaluatorTry itNVIDIA's open-source evaluation tool that measures how much better an agent performs with a skill attached versus without it
- SkillsTry itAn Anthropic feature that lets Claude store a set procedure and reference materials for a specific task ahead of time and call them up when needed.
- Skills APITry itA feature that lets you save a repeated work procedure once, version it, and call up a specific version whenever you need it
- SkillTrustBenchTry itA benchmark that scores the security risks of external tools (skills) called by AI agents across nine risk categories
- SKTCompanySKT (SK Telecom) is a South Korean telecom operator expanding into AI-related businesses centered on its mobile communications services.
- SkydiveAnything's AI teammate-style agent product that keeps the same context as it moves across Slack, email, text, and the CLI
- SK하이닉스CompanySK hynix is a company that designs, manufactures, and sells memory semiconductors.
- SL2TTry itA sign language recognition model from Google DeepMind that reads hand, body, and facial movements in real time and turns them into sentences.
- Sleep Stage AccuracyTry itA measure of how precisely a wearable device, like a smart ring or watch, can identify the sleep stage of a sleeping person.
- SLMA language model shrunk down small enough to run on a phone or laptop. These days the competition is all about whether "smaller" can still mean "smart enough."
- Smart LyricsTry itA feature that lets AI write and attach lyrics automatically just from a one-sentence song description, producing a finished track.
- Smart ResizeTry itAn image-editing feature that lets AI automatically fill in the blank space created when you change an image's aspect ratio.
- Smart WindowTry itAn opt-in AI browsing mode built into the Firefox browser that organizes your tabs and browsing history and answers questions through chat
- SMT SolverA verification program that mathematically determines whether a set of rules and conditions can all hold true at the same time
- SMT-LIBTry itA standardized input format that lets computers automatically verify logical rules.
- SnowflakeCompanySnowflake is a company that builds and provides enterprises with a cloud-based data storage and analytics platform.
- Social Engineering AttackTry itAn attack method that exploits human trust and emotion, rather than technical flaws in a system, to extract information or gain access.
- SocialIQAA benchmark that tests, through multiple-choice questions, how well an AI understands human emotions, intentions, and everyday social situations
- Softmax NormalizationA calculation that turns several scores into probabilities that always add up to 100%
- Software FactoryA development system that uses AI to handle the entire software development process—not just writing code, but planning, design, testing, deployment, and operations—through a standardized workflow.
- Software Supply Chain SecurityTry itSecurity practices that protect every stage code passes through, from creation to deployment, against attacker intrusion
- SolarisAn interface-generation program from Runway that renders the screen in real time with every click, instead of relying on pre-coded screens.
- SonarTry itA search-connected AI model built by Perplexity that generates answers grounded in documents it retrieves from the web
- sonar-proTry itThe AI model Perplexity uses to generate answers grounded in real-time web search results.
- Sonic-3.6A version of an AI voice-generation model that turns text into natural-sounding speech in real time
- Sonnet 5Try itThe mid-tier model in Anthropic's Claude lineup, balancing speed and capability
- SoraTry itA video-generation AI model made by OpenAI that creates short videos from written descriptions
- SORRY-BenchA safety benchmark that measures how well an AI model refuses requests that could be dangerous or inappropriate
- SoulA photorealistic image generation model built in-house by Higgsfield, forming the company's proprietary toolset alongside its video model DoP.
- Sovereign AIAI built with a nation's own language, data, and infrastructure — the core term in Korea's AI policy.
- Sovereignty IndexA fake ranking report invented in a Russia-linked disinformation operation to disparage Western countries and cast Russia in a favorable light.
- SparkTry itAn agentic feature in the Google Gemini app that actually carries out tasks on the user's behalf
- Sparse AttentionA method that cuts computation by having an AI model focus only on the important parts of its input instead of comparing everything against everything else.
- spatial audioTry itAudio technology that adds directional and distance cues to sound, making it seem as if it's coming from real positions in space.
- spatial multi-scale visit propagationA technique that fills in the activity patterns of sparsely-visited places by borrowing mobility data from surrounding locations across multiple spatial scales
- Speaker DiarizationTry itA speech recognition technology that identifies and labels 'who is speaking' at each point in a recorded conversation.
- Spec KitTry itAn open-source workflow from GitHub that forces AI coding assistants to write a design spec before writing any code
- Spectrum-XA networking technology from NVIDIA that links multiple GPUs together like a single cluster, letting them exchange data quickly.
- Spectrum-X MultiplaneA data center networking technology that links multiple switches in parallel so massive volumes of data can move without bottlenecks
- Speculative DecodingAn AI response-generation technique where a small assistant model drafts the answer first, and the large model only verifies it in bulk to speed things up.
- SpotlightA short-form video section inside Snapchat that pays out a share of money to people whose clips get a lot of views.
- Spring AIA library built to connect AI models the way Java's Spring framework does things
- SQL InjectionA hacking technique that sneaks commands instead of ordinary data into an input field to manipulate a database or steal information
- ssl.SSLContextTry itA settings bundle that holds the certificate-verification rules a program uses to check whether the party on the other end of an internet connection is really who it claims to be
- Stability AICompanyStability AI is a company that develops and provides generative AI models for images, video, audio, and more.
- stackable credentialsAn educational structure where small-unit certificates or credentials can be accumulated one by one, eventually leading to a larger degree or professional qualification.
- State of Open Models, Summer 2026Try itA report on open-source AI trends from Hugging Face, which found that Qwen ranks first for local inference.
- State Space ModelAn AI model architecture that keeps updating a single compressed 'state' summarizing everything read so far, so computation doesn't balloon as the input gets longer.
- statelessA way of handling each request independently, with no memory of previous requests, as if seeing it for the first time
- Stateless MCPA design approach where an MCP server processes each request independently instead of remembering connection state with a client.
- stdioThe simplest way for two programs to exchange text through a single input/output stream, with no network involved.
- Stealth ModelTry itAn AI model released anonymously, without revealing the company that built it, to gauge public reaction
- Strands AgentsAn open-source toolkit released by AWS that serves as a framework for assembling AI agents so they can call external tools.
- StreamA ring-shaped voice recording device made by startup Sandbar that transcribes what you say into text while you hold down the button
- streaming rendererTry itThe part of a program that draws each character on screen as an AI's reply streams out one character at a time
- StripeCompanyStripe is a company that provides online payment processing infrastructure.
- structure predictorAn AI tool that calculates the actual 3D folded shape of a molecule just by looking at a one-dimensional string of information, like an amino acid sequence.
- Student Data Privacy ConsortiumA U.S. coalition that provides a common standardized contract framework so schools and edtech companies don't have to negotiate student data handling terms from scratch every time
- style reference pinTry itA marker that locks a specific style reference image into the prompt during image generation
- StyloA CSS parsing engine built by Firefox that reads a webpage's design rules and calculates how it should be laid out on screen.
- Sub-agentTry itA helper agent that a main AI agent delegates smaller pieces of a complex task to, each running independently
- Subscription TiersMembership levels for AI services, like Free, Plus, and Pro. Higher tiers unlock more usage and the latest features.
- SubsenseA startup trying to connect the brain to computers using nose-inhaled nanoparticles, without cutting open the skull
- SunoA music AI company that generates entire songs, including lyrics and vocals, from a short prompt.
- Suno Studio 2.0Try itA music editing tool from Suno, now updated with a feature that lets you type a sentence to generate a custom audio effect on the spot.
- Super AgentTry itA term for AI that plans on its own across multiple areas of work, even calling on tools or other agents, to see a task through to completion.
- SuperintelligenceA hypothetical intelligence that surpasses humans in every respect, discussed as the stage beyond AGI.
- Supervised Fine-TuningTry itA follow-up training stage where an already-trained model is shown paired question-and-answer examples to reteach it how to respond the way you want.
- supervisor agentA watchdog-like component that steps in to redirect another AI agent when it gets stuck or keeps repeating the same mistake.
- Supply Chain AttackTry itAn attack method that poisons a single widely used piece of software or development tool to spread damage all at once to every organization that relies on it
- Supply Chain RiskAn official government designation that blocks all public agencies from using a company's products after determining they pose a threat to national security
- surface plasmon resonanceA physical phenomenon and measurement technique that shines light on a metal surface to measure how tightly two molecules bind to each other.
- SWE-bench ProA coding benchmark that measures how well AI can fix real-world issues in software repositories
- SWE-bench VerifiedTry itA benchmark that automatically scores how well an AI coding assistant fixes real bugs from actual open-source projects, using human-verified problems.
- Synchronized ReasoningA method where multiple AI models or agents align their reasoning processes with each other in real time as they work through a problem together
- SynthesiaA company that creates talking AI avatar videos, widely used for corporate training and instructional content.
- Synthetic DataArtificial training material made by AI to teach AI — a stand-in for an era running short on real data.
- synthetic mirroringA detection method that learns the telltale patterns of AI-generated text and reflects them back to spot writing with the same patterns
- SynthIDTry itA watermarking technology from Google DeepMind that embeds imperceptible signals into AI-generated images, video, and audio, letting machines verify whether content was AI-generated.
- System PromptTry itThe "master instruction sheet" an AI follows throughout a conversation — it sets the tone, role, and off-limits topics.
- System-on-ChipA design approach that integrates core components like the CPU, GPU, and memory onto a single semiconductor chip.
T66
- Tabular AnalysisTry itAn AI capability that reads and computes, compares, and summarizes data organized into rows and columns, like a spreadsheet or table
- TamperBenchA benchmark that tests how easily the safety guardrails of open-weight AI models collapse under fine-tuning or tampering.
- TavusCompanyTavus is a company that develops technology for creating real-time conversational AI avatar videos.
- TaxAIAn AI agent within Thrive Holdings' accounting platform Current that automatically handles tax return preparation.
- Teen ModeTry itA protective, learning-focused ChatGPT environment automatically applied to accounts belonging to teens aged 13 to 17
- teleprompter modeTry itA smart glasses feature that displays a presentation script within the wearer's field of view, letting them read it while still facing the audience.
- temporal encoderA component that converts time-related information—like arrival time or dwell time—into numerical coordinates that AI can compute with.
- temporal policiesRules that decide whether to allow an AI agent's current action by checking its prior action history, not just the action itself
- Temporal PolicyA policy approach that decides whether to allow an AI agent's next action by looking at the sequence and cumulative history of an entire session, rather than judging each action in isolation.
- Tencent HunyuanCompanyTencent Hunyuan is an AI model brand developed by China's Tencent, producing models across a range of domains including language, image, and 3D generation.
- Tensor Processing Unit (TPU)Tensor Processing UnitGoogle's in-house AI chip, and the leading challenger to NVIDIA's dominance.
- TerminalTry itA black screen where you type commands as text. It's where CLI tools run, and it shows up in the background of every AI coding story.
- Terminal Bench 2.1Try itA benchmark that measures how well an AI agent can complete multi-step tasks on its own in a computer's command-line environment
- Terminal-Bench 4.0A benchmark suite that scores how well AI models handle real-world tasks in a command-line computer environment
- Terminal-Bench v2.1A benchmark that measures how well an AI model can complete real tasks by operating a computer's command-line terminal
- TerminalBenchAn evaluation tool that tests whether AI can carry out real tasks from start to finish inside a command-line terminal environment
- TerminalBench 3.0A benchmark that tests how well an AI can solve coding problems on its own inside a terminal (command-line) environment.
- test suiteA set of problems that automatically grades whether AI-generated code or answers are correct.
- Test-time ComputeAn approach that boosts performance by spending more computation ("thinking") at answer time — the idea that "thinking longer makes you smarter."
- Test-time LearningTry itA way for a fully trained model to adapt on the spot, right when it's actually being used, based on just a single example it sees in the moment.
- Text NodeTry itAn input field in a workflow that holds a changeable value—like a language or phrase—so the next step can read it and react automatically
- text-mobility synergyAn AI processing stage that merges text describing a place with real human movement data into a single numerical representation.
- The Cancer Genome AtlasOne of the world's largest public cancer datasets, combining genetic information and clinical records across many cancer types.
- Theory of MindTry itThe ability to predict someone's behavior by inferring what they believe and want, and there's growing discussion about whether AI needs this ability too
- TheRockA new build system AMD uses to assemble and distribute ROCm across different operating systems and package formats
- Think ButtonTry itA button that lets free ChatGPT users have the default AI think longer before answering complex questions
- Thinking Depth SliderTry itA slider feature that lets users control how deeply ChatGPT reasons before producing a response
- Thinking DropdownTry itA selector next to the input box in the Meta AI desktop app that lets you choose how deeply the model should think before answering
- third-party agentTry itAn AI agent built by a company other than the one that made the platform, running inside that platform
- Three.jsTry itA free JavaScript library that lets you draw three-dimensional 3D graphics right in a web browser window
- Time to First TokenTry itA metric that measures how long it takes for the first character of an AI's response to appear after you send a query.
- TimesFM-3A multivariate time-series forecasting AI model from Google Research that predicts future numbers by factoring in several data streams at once
- timestampTry itA time stamp attached to a specific moment in audio, video, or text, marking exactly what happened at what minute and second
- tinyflowsTry itA feature in OpenHuman that, when you say "automate this," has the AI draft and propose a workflow diagram (graph) for you first.
- Together AIAn AI cloud that runs open models on your behalf — essentially a server room for companies that want to use open-source models.
- TokenTry itThe unit AI uses to count text. Pricing and usage limits are all calculated in tokens.
- Token BillingA billing feature from payments company Stripe that automatically measures how many tokens an AI service uses and charges customers accordingly
- Token MaxxingTry itA usage style that pours tokens (AI usage) without holding back to push up output quality — from the internet '-maxxing' meme.
- Token Price IndexA metric that tracks how token usage fees across various AI models change over time.
- token taxA policy proposal to tax the use of tokens—the basic unit of AI work—in order to raise the cost of replacing humans with AI.
- TokenizerThe component that splits text into tokens. Because the cutting pattern differs by language, it's at the heart of the debate over Korean-language AI costs.
- toktoktokA nickname for a toy-level tokenizer training program that a coding agent whipped up in 30 minutes, which only worked on small-scale data.
- Tongyi LabCompanyAlibaba's Tongyi research lab.
- Tool CallingTry itThe ability of AI to pick and use outside functions like calculators, search, or booking systems — the hands and feet of an agent.
- Tool Calling LoopTry itThe repeated process where an AI, unable to answer on its own, calls an external tool for information, checks whether the result is enough, and calls another tool if it isn't — over and over as needed
- Tool GatewayAI 에이전트가 웹 검색·이미지 생성 같은 여러 외부 기능을 하나의 통로로 불러 쓰게 해주는 연결 창구다.
- topology-aware parallelizationA way of splitting large-scale AI model training work across chips that takes into account how those chips are actually connected to each other
- topology-aware schedulerA scheduling method that assigns jobs to the most efficient locations by understanding the physical layout and connections between GPUs and servers
- TOPS (Tera Operations Per Second)A performance metric expressing how many trillion operations an AI chip can process in one second.
- Training & InferenceThe two phases of AI's life — the time it spends learning (training) and the time it spends working (inference). The economics of the two are completely different.
- TrainiumAmazon's custom-designed AI chip for training and running AI models on the AWS cloud.
- Transfer LearningA learning method that carries skills learned in one domain over to another. The reason foundation models exist.
- TransformerA neural network architecture published by Google in 2017 — the 'T' in GPT, and the common backbone of today's AI boom.
- TransformersTry itA software library built by Hugging Face that lets AI models be saved, loaded, and run in a standard format.
- Tree-RLA reinforcement learning training method that branches into multiple paths mid-run and learns by trying each of them out
- tremoloTry itA sound effect that rapidly and rhythmically pulses the volume of a sound
- Trusted Access for CyberAn accredited access program that lets OpenAI-verified cybersecurity researchers use less-restricted, cutting-edge models.
- Trusted Access PathwaysAn access control approach that opens up potentially risky AI capabilities only to verified experts or specific user groups, as an exception rather than the default.
- TSMCCompanyTSMC is a foundry company that manufactures semiconductors on a contract basis.
- Voice AI (TTS · STT)TTS · STTAI that converts speech to text (STT) and text to speech (TTS) — the tech behind interpretation, dubbing, and voice cloning.
- Turbo LoRAA speed-up technique for video generation models that leaves the base weights untouched and attaches a small module per request, sharply cutting the number of computation steps.
- Turing TestA thought experiment for judging machine intelligence, proposed in 1950 — if you can't tell whether you're talking to a machine or a human, it passes.
- TutorMomentsAn evaluation framework that measures how well an AI tutor knows when to help a student and when to step back
- TutorMoments-PreviewTry itA collection of real math tutoring conversations released by the Allen Institute for AI (Ai2) to evaluate whether an AI can judge when to give a hint versus when to push a student with a harder question
- Twelve LabsCompanyTwelve Labs is a company that develops AI models for understanding and searching video.
- TyposquattingTry itAn attack that registers a name no one has claimed yet, then hijacks whoever or whatever trusts and connects to or installs from that name.
U19
- Ultrafast ModeAn API-only speed option from OpenAI that makes the same model respond up to 14 times faster while keeping answer quality unchanged.
- Uncertainty QuantificationTry itAn approach where AI outputs a numerical confidence level (probability) alongside its judgment, rather than just an answer
- Uni-1Luma AI's proprietary video generation model, said to power its scene pre-approval tool Luma Scenes
- Unified MemoryA computer architecture in which the graphics processor and the central processor share a single pool of memory.
- Unit Distance ConjectureA long-standing math problem asking how many pairs of points, no matter how many points you scatter on a plane, can end up exactly the same distance apart.
- Unity AI GatewayA Databricks management product that governs how much and in what way AI models can access a company's internal data
- Unity CatalogA data management capability built into Databricks' lakehouse that handles both data discovery and access control in one place.
- Unity EngineDevelopment software for building games and 3D simulations, made and distributed by Unity Technologies
- Unity GatewayA gateway-style tracing system that lets Databricks automatically log every external tool call made by an AI agent.
- Unreal Engine 5A 3D real-time graphics engine made by Epic Games, used to build virtual spaces and objects in three dimensions for games or film.
- UnslothCompanyOpen-source software that accelerates fine-tuning.
- Unsloth DesktopTry itA desktop app that lets you run a model on your own computer and retrain it on your own data, all in one window.
- UnswarmTry itAn open-source runtime manager and proxy that pools multiple local LLM runners together and automatically switches them on and off as needed.
- UpliftThe degree to which AI hands a bad actor knowledge or capability they couldn't easily get elsewhere, raising their ability to carry out dangerous acts
- UpscalingAn AI technique that enlarges low-resolution images or video to look sharp and clear, commonly used to restore old photos.
- UrlReadToolA built-in URL-reading tool in Atlassian's AI agent Rovo that was exploited as a data exfiltration channel in a prompt injection attack.
- Usage & Spend DashboardTry itA developer-facing OpenAI console that breaks down usage and spend by individual API key and lets you set limits on each.
- usage metricsTry itUsage statistics that count and show who used an AI service or tool, how much, and in what way.
- USDC (USD Coin)A digital currency (stablecoin) issued by Circle, designed to stay pegged to the value of one US dollar.
V37
- V4 ProA large language model from DeepSeek, released in a near-open-source way that is driving down prices in the industry.
- vClusterA virtualization technology that lets multiple teams share a single pool of physical server resources while each team operates as if it had its own dedicated cluster.
- Vector DatabaseA storehouse that keeps embedding coordinates and finds the ones that are "close in meaning" — the bookshelf behind RAG.
- Vera RubinNVIDIA's next-generation AI chip system that binds six different chips together to work as one
- Vera Rubin Space-1A space-dedicated GPU that NVIDIA is developing, designed using orbital satellite data from Starcloud.
- VercelCompanyVercel is a company that builds a cloud platform for deploying and running web applications.
- Vercel Container RegistryTry itA container image storage service from Vercel that supports sharing images across teams and public distribution.
- Vercel SandboxAn isolated, temporary workspace provided by Vercel where AI agents can run code without affecting live production services
- vertical integrationA strategy where a single company directly owns and controls multiple stages of production (design, manufacturing, infrastructure, financing, etc.)
- Vertiv TrimCoolerAn auxiliary cooling technology from Vertiv used within chilled-water cooling systems in AI data centers.
- Very Large Online PlatformA legal classification under EU law for online services with more than 45 million average monthly users in the EU, which comes with stricter regulatory obligations.
- Very Large Online Search Engine (VLOSE)A regulatory category the EU applies under the Digital Services Act to online search services with more than 45 million monthly users.
- Vibe CodingTry itA way of building apps by just telling AI what you want in plain language, no coding required — "make it feel like this" is enough.
- vibratoTry itA musical effect that makes a note's pitch waver rapidly in tiny increments, making the sound feel richer and more alive.
- Video GenerationAI that turns a sentence or a single photo into a moving video. Sora, Veo, and Runway lead this space.
- Video Language ModelTry itAn AI model that understands scenes, people, and actions in video well enough to describe or search them using natural language
- video-action modelAn AI model trained to predict not just the video it generates, but also the movements and actions happening within that video.
- Virtual OrganizationA virtual structure in which multiple AI agents divide up roles like departments in a company, working together toward a shared goal.
- Vision-Language-Action ModelAn AI architecture that lets a robot understand what it sees and what it's told, then directly generate the movement commands to act on it
- vision-onlyAn approach that reads the surroundings using only camera footage, without dedicated distance sensors — the path Tesla adopted in 2019.
- Visual IntelligenceTry itApple's visual recognition feature that identifies objects seen through the camera and tells you what they are.
- VisualizeTry itA feature that turns ChatGPT's written answers into screens with tables, timelines, and buttons.
- VLA 2.0An autonomous driving technology under development by Chinese EV maker XPeng that combines camera-based scene understanding and language-based reasoning into a single model to produce steering, acceleration, and braking decisions.
- vLLMCompanyvLLM is an open-source inference engine that improves inference speed and memory efficiency so that large language models can serve many users simultaneously.
- vLLM-OmniServing infrastructure that lets AI models generating multiple formats together, like video and audio, run quickly in real-world services
- VLM (Vision Language Model)VLMTry itA language model with the ability to "see." Give it a photo, and it reads and explains what's in it.
- Voice & Chat AgentTry itAn AI program that carries on real conversations in place of a human, over the phone or in a chat window.
- Voice AgentTry itAn AI program that talks with people in real time to handle phone-based tasks like customer support or inquiries
- voice finderTry itA tool from Together AI that lets you describe the vibe you want in words and finds a matching voice from a library of over 600 options.
- Voice InputTry itAn input method that converts your spoken words directly into text for the AI, instead of typing on a keyboard
- Voice ModeTry itA feature for real-time spoken conversation with a chatbot, complete with interruptions and emotional tone.
- Voice over Internet ProtocolA calling method that sends voice over the internet data network instead of traditional phone lines
- voice steeringTry itA feature that lets you adjust an AI voice's tone, emotion, and delivery on the fly using natural-language instructions
- VOID FrameworkVOIDA name for which the developer and functionality cannot be clearly confirmed based on currently available sources.
- voxelA unit that divides 3D space into cube-shaped cells — essentially a pixel with depth added.
- Voxtral TTSA speech synthesis technology from Mistral that turns written text into natural human-sounding speech.
- VST (Virtual Studio Technology)A standard format for virtual instruments and effects that plug into music production software
W38
- Wafer-Scale EngineA semiconductor product that puts an entire AI model onto a single giant silicon wafer, instead of splitting it across multiple chips, eliminating chip-to-chip communication delay
- Wan 2.7The version of Alibaba's video-generation AI Wan series that came before Wan 3.0.
- Wan-Animate-2Try itAn open-source character animation model from Alibaba that maps a human's motion video onto a still character image, letting multiple characters move simultaneously in one scene
- Wan3.0Try itAn AI video generation model from Alibaba that turns text, images, or documents into videos up to 30 seconds long.
- WasteIDA detection feature in a robotic litter box that tells urine and feces apart so it can clean up the smellier waste faster.
- WatchlistA surveillance feature that automatically and continuously scans every camera inside a mapped area for a person matching a written physical description.
- Waterhole TheoryTry itA marketing strategy of building a place where your target customers naturally gather and building relationships there, instead of chasing after those customers directly
- Watermark Detection APIAn Anthropic tool that lets other companies' services check for the invisible marks left in text written by Claude.
- WatermarkingAn invisible marker embedded in AI-generated content; a key technology for countering deepfakes.
- WatermelonA new AI model reportedly set to be unveiled by Meta in October 2026, said to power its paid agent service, Hatch.
- waveguide displayA thin optical display technology that traps light inside a lens-like panel and channels it to the wearer's eye, projecting text or images directly into view
- Weather LabTry itA website where Google DeepMind publishes the results of its AI-based typhoon and hurricane forecasting experiments
- WeatherNextAn AI weather forecasting model built by Google DeepMind, trained on decades of past weather data, that can catch phenomena like hurricane rapid intensification ahead of time.
- Web Bot AuthA method that lets bots or AI agents cryptographically sign requests to prove their own identity — essentially saying 'this is what I am' with a verifiable signature
- Web Query LanguageA query tool being developed by Kagi that helps AI construct answers by pulling together multiple web documents rather than just a single page
- web_search · web_fetchTry itTwo basic tools that let an AI agent find information on the internet (web_search) and read the content of a specific page (web_fetch).
- Webcam APITry itA standard browser feature that reads live camera video and makes it usable directly inside a webpage.
- WebMCPA name for an emerging approach to let AI directly operate websites themselves, though the concrete specification hasn't been published yet
- WARPWebRTC Abridged Roundtrip ProtocolOpenAI's connection protocol that cuts the handshake needed to start a voice call from six round trips down to one
- WeightsThe bundle of numbers a model learns during training — the model's actual substance. Release these, and you have an open model.
- Weights & BiasesCompanyWeights & Biases is a company that builds software for tracking and managing the machine learning model development process.
- Whisker PlusA paid membership from Whisker, the maker of robotic cat litter boxes, bundling extra app features and service perks into a subscription product.
- WildArtifactBenchTry itAn AI agent evaluation framework released by Meta AI that ranks agents by pitting their outputs against each other head-to-head and scoring win rates and rankings, rather than checking answers against an answer key.
- Wispr FlowTry itA dictation app that turns spoken words into text in real time
- Wispr Interface LabsA new research group set up by Wispr, led by an early Amazon Alexa developer, exploring new ways for people and devices to interact.
- Word Error RateTry itAn accuracy metric showing what percentage of words a speech-to-text system gets wrong.
- WorkArena EloAn Elo-style ranking score, rated by human evaluators, that measures how well an AI model handles office work.
- Workbench Notebooks VS Code ExtensionTry itA Google Cloud extension that lets you connect directly to cloud Jupyter notebooks from within VS Code.
- WorkflowTry itThe sequence and set of steps followed to complete a task from start to finish.
- WorkloadThe amount and type of work a computer or server needs to process
- Workload Identity FederationAn authentication method where a program or service proves who it is using a credential from a trusted issuer, instead of managing its own secret keys.
- Workspace for EducationTry itA Google account service that schools issue to enrolled students and staff for educational institutions
- World Action ModelAn AI modeling approach in which a robot first predicts physical changes visually before deciding on an action, rather than relying on language.
- World AI Cooperation Organization (WAICO)An international AI cooperation body launched by China in July 2026, pushing open-weight AI models to counter a U.S.-led coalition.
- World LabsCompanyWorld Labs is a company that researches and develops 3D spatial and world models.
- World ModelA model that learns how the world actually works — physics and cause and effect — rather than language. The next battleground in video generation and robotics.
- World PromptTry it걸어 다닐 수 있는 가상 공간을 말로 조종하는 형식으로, 세계 설정과 실시간 사건 지시를 나눠 담는다
- WSE-TurboCerebras's new wafer-scale processor powering its CS-4 inference hardware.
X2
Z8
- Zero Data RetentionA policy in which an AI service discards a user's request and the resulting response immediately after processing, instead of storing them on its servers.
- Zero TrustA security approach that never assumes trust just because someone was verified once, and instead re-checks identity and permissions every time access is requested
- Zero-cost content generationTry itThe phenomenon of using AI to produce photos, videos, music, and text through nothing but a prompt, without shoots or labor costs, at close to zero expense
- Zero-dayA security flaw that attackers discover and exploit before the company that made the software even knows it exists.
- Zero-day ExploitAn attack method that finds and actually breaks into a software security hole nobody knows about yet
- Zero-day VulnerabilityA security hole in software that even its own creators don't yet know about — if exploited before a patch exists, there's no way to stop it.
- Zhipu AICompanyZhipu AI is a Chinese artificial intelligence company that develops and offers large language models.
- ZooxCompanyZoox is a company that develops and operates autonomous robotaxis.