DICTIONARY
K-Culture Glossary
The words that stop you mid-article, explained for beginners. Dotted underlines inside our stories bring you here.
entries 174 · companies 35 · and growing
A26
- AdobeThe company behind Photoshop. It's building its own AI model, Firefly, into its creative tools across the board.
- 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.
- AGIA hypothetical AI capable of nearly all human intellectual work — the stated ultimate goal of OpenAI and Anthropic.
- AIAn umbrella term for technology that lets computers do the "thinking" work humans used to do.
- 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 GovernanceThe management framework organizations use to ensure AI is used safely and responsibly — the in-house version of "regulation."
- AI LiteracyTry itThe ability to understand and properly use AI. The basic literacy of the generation after digital literacy.
- 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 SafetyThe research and policy field focused on making sure AI doesn't cause harm — also the reason Anthropic was founded.
- 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 SlopA derogatory term for the low-quality, mass-produced AI content flooding the internet — a recurring flashpoint in culture debates since 2024.
- Ai2A US nonprofit AI lab that releases not just model weights but also training data and code in full.
- AlgorithmA blueprint for solving a problem, step by step. It's the same "algorithm" people mean when they say "the algorithm recommended it."
- AlignmentThe problem of making AI act according to human intentions and values. A central term in AI safety discussions.
- 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 Web ServicesAmazon's cloud arm. It rents out the servers AI companies run their models on, and sells other companies' models through Bedrock.
- AMDA chipmaker often cast as NVIDIA's rival, maker of the Instinct line of AI accelerators.
- AnthropicThe AI company behind Claude. Founded in 2021 by former OpenAI employees, it markets itself under the banner of "safe AI."
- 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.
- 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.
- 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.
- 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.
B5
- BenchmarkTry itA standardized test that AI models all take together. It's the basis for those "which model is smarter" articles.
- 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.
- BotA general term for a program or account that runs automatically. This is the “bot” in “chatbot.”
- 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.
C18
- Chain-of-ThoughtTry itThe discovery that having AI write out its reasoning before answering boosts accuracy. The starting point of reasoning models.
- ChatbotTry itAn AI service you use through a chat window. ChatGPT, Claude, and Gemini are all LLMs wrapped in a chat interface.
- 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.
- ClaudeTry itA conversational AI service and family of models made by Anthropic, widely used for reading long documents, writing, and coding.
- 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.
- 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.
- CloudflareAn infrastructure company that protects and speeds up websites, now at the center of the fight to control AI crawlers.
- 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.
- 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.
- CohereA Canadian model developer that builds AI for enterprises and governments to run on their own servers, rather than consumer chatbots.
- 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 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.
- 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.
- Copilot"Co-pilot" — a term for AI that assists people alongside them in their work, and also Microsoft's AI brand name.
- CrawlingThe act of scraping data from the web. Ground zero for copyright disputes over AI training.
- 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.
D13
- Data CenterThe building where AI actually lives — packed with tens of thousands of GPUs. Power and water problems always come with it.
- Data LabelingThe work of attaching correct answers to data used for training AI. It's the invisible human labor that raises AI.
- 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.
- 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.
- DeepfakeTry itAn AI-made "fake that looks real" — a manipulated face, voice, or video. Both a technology and the name of a social problem.
- 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.
- DemoA demonstration showcasing a feature. The centerpiece of AI announcements, and also the thing that gets checked against "demo vs. real-world use."
- Diffusion ModelA generative architecture that builds images by gradually removing noise. The engine behind Midjourney, DALL-E, and Stable Diffusion.
- 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.
- DistillationA technique for transferring a large model's abilities into a smaller one — a "teacher model → student model" process.
- DoomerA term for the camp that believes AI could bring about human catastrophe and calls for slowing down development.
E3
- 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.
- EmbeddingTry itA technique that turns the "meaning" of words into numerical coordinates. Similar meanings land close together — the foundation of AI search.
- EmergenceThe phenomenon where scaling up a model suddenly produces abilities it was never trained on — a source of both wonder and unease.
F6
- Factory AIThe company behind the coding agent "Droid," built around the idea of handing off entire coding tasks rather than just autocompleting lines.
- 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).
- Fine-tuningGiving a already-built model extra training for a specific job — more like altering a ready-made suit than tailoring a new one.
- 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.
- 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.
G10
- 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.
- GGUFA file format for packaging AI models so they run well on personal computers — the standard container of the local AI community.
- 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 CopilotAn AI tool built into GitHub that writes code for you — the original that kicked off the "AI coding assistant" era.
- 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.
- GPUThe chip used to build and run AI. The reason NVIDIA became the world's most valuable company.
- GroundingA technique that ties AI answers to real documents or data — a core design principle for reducing hallucination.
- 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.
H5
- HallucinationTry itWhen AI states plausible-sounding falsehoods as if they were fact. Generative AI's most notorious weakness.
- 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.
- HiggsfieldAn AI video and image creation platform that builds its own models (Soul, DoP) while also plugging in models from OpenAI, Google, and ByteDance.
- 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.
- 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.
I2
J1
K2
L5
- LatencyThe delay between asking and getting an answer. For voice conversations to feel natural, this has to come down to human reaction speed.
- LiDARA sensor that fires light and times its return to measure distance — precise on range, but blind to text.
- LLMTry itThe heart of ChatGPT, Claude, and Gemini — a program trained to predict the "next word" after reading enormous amounts of text.
- 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.
- Luma AIA company that builds its own video and 3D generation models under the name "Ray." Its flagship product is Dream Machine.
M14
- 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.
- 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."
- 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.
- Meta AIMeta's AI organization behind Llama; in August 2026 it admitted Llama 4's failure and released its successor, Muse, as open source.
- Microsoft AIMicrosoft's consumer AI organization, the group behind Copilot, which uses both its own models and OpenAI's models.
- MiniMaxA Chinese multimodal AI developer — the Hailuo video brand and the character chat app are both products of this same company.
- Mistral AIA French model developer that pitches "sovereign AI" — processing data inside Europe.
- ModelThe term for a finished piece of AI software that has completed training — the 'model' in headlines like 'new model released.'
- 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.
- MCPModel Context ProtocolTry itA common standard for plugging AI into outside tools and data — nicknamed the "USB-C of AI."
- MoE (Mixture of Experts)MoEAn architecture that wakes up only the 'experts' needed for each question — a trick for running huge models cheaply.
- Moonshot AIA Chinese AI developer behind the chatbot and model "Kimi" — a regular fixture at the top of open-source model rankings.
- Multi-agentA setup where multiple AIs split roles and work together — like forming a research team, a review team, and a writing team.
- MultimodalTry itThe ability to understand and generate not just text, but images, sound, and video together.
N4
- 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.
- NLPTry itThe field of AI that lets computers work with human language — translation, summarization, and chatbots all come from here.
- 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.
O6
- 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.
- OntologyTry itA machine-readable map of concepts and how they relate to each other — a blueprint for "who made what."
- 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.
- OpenAIThe company behind ChatGPT and the GPT model family. It kicked off the current AI boom with ChatGPT's 2022 launch.
- 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.
P7
- ParameterThe number of adjustable dials inside a model. The "B" in "70B model" refers to this — generally, more means smarter but more expensive.
- 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.
- PluginAn add-on component that plugs new capabilities into an AI service — a packaged unit of skills, connectors, and tools.
- 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.
- PromptTry itThe instruction you give an AI — every question, command, or example you type in counts as a prompt.
- 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.
Q3
- QuantizationCompression that shrinks a model by lowering the precision of its numbers — the trick that lets LLMs run on phones.
- QueryA single question or request sent to a system — search terms and AI prompts both count as queries in the broad sense.
- QwenAlibaba's in-house open-source model family — one of the most widely used names in the open-model camp.
R5
- RAGTry itA method that has AI look up source material before answering. The go-to fix for hallucination.
- 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.
- Red TeamingA role that deliberately attacks a product before launch — "we break it ourselves first" has become the standard procedure for AI safety.
- Reinforcement LearningA trial-and-error training method that rewards good moves and penalizes bad ones. The method behind AlphaGo.
- RLHF인간 피드백 강화학습Try itTraining that polishes a model's manner of speaking using human-scored 'good answer' ratings — the technique that made ChatGPT polite.
S14
- SandboxAn isolated sandpit where AI can run code without causing damage outside it — a basic building block of agent safety.
- Scaling LawsThe empirical rule that performance improves predictably as you scale up model size, data, and compute. The rationale behind AI's investment race.
- SDKA ready-made toolkit for developers. When a company "releases an SDK," it's a signal that building on its technology just got easier.
- Self-drivingTechnology that lets a car drive itself. Shows up in news through level numbers and robotaxi stories.
- 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.
- 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."
- 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."
- Sovereign AIAI built with a nation's own language, data, and infrastructure — the core term in Korea's AI policy.
- Subscription TiersMembership levels for AI services, like Free, Plus, and Pro. Higher tiers unlock more usage and the latest features.
- SunoA music AI company that generates entire songs, including lyrics and vocals, from a short prompt.
- SuperintelligenceA hypothetical intelligence that surpasses humans in every respect, discussed as the stage beyond AGI.
- 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.
- System PromptTry itThe "master instruction sheet" an AI follows throughout a conversation — it sets the tone, role, and off-limits topics.
T13
- 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.
- Test-time ComputeAn approach that boosts performance by spending more computation ("thinking") at answer time — the idea that "thinking longer makes you smarter."
- 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 MaxxingTry itA usage style that pours tokens (AI usage) without holding back to push up output quality — from the internet '-maxxing' meme.
- 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.
- 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.
- 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.
- 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.
- 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.
- 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.
U1
V7
- Vector DatabaseA storehouse that keeps embedding coordinates and finds the ones that are "close in meaning" — the bookshelf behind RAG.
- 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.
- Video GenerationAI that turns a sentence or a single photo into a moving video. Sora, Veo, and Runway lead this space.
- vision-onlyAn approach that reads the surroundings using only camera footage, without dedicated distance sensors — the path Tesla adopted in 2019.
- 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 ModeTry itA feature for real-time spoken conversation with a chatbot, complete with interruptions and emotional tone.
- voxelA unit that divides 3D space into cube-shaped cells — essentially a pixel with depth added.
W3
- WatermarkingAn invisible marker embedded in AI-generated content; a key technology for countering deepfakes.
- WeightsThe bundle of numbers a model learns during training — the model's actual substance. Release these, and you have an open model.
- 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.