New words arrive fast in AI: today it's large models, tomorrow it's Agents, and a few days later up pop MCP, A2A, RAG, Embedding, and Knowledge Graph. Once you actually start using AI tools, you'll also run into context window, model parameters, inference, GEO, edge AI, Vibe Coding, Harness Engineering, API Key, relay, AI gateway, model routing, and VPN—terms tied to model capability, content visibility, access, networking, and security. 2026 brought another batch: Loop Engineering, context engineering, context compaction, evals, Agent Skills, subagents, MoE, test-time compute, world models, guardrails, and prompt injection. Many articles assume the reader already understands half of it, and the more you read the more it feels like memorizing acronyms.
This series aims to do something more down-to-earth: explain these words by putting them back into real usage scenarios. It assumes you don't need to write code, and you don't need to understand servers, algorithms, or math formulas. If you've used search engines, chat apps, online shopping, or online documents, you can keep up.
Each article tries to answer three questions: what everyday thing it's actually like, why people keep bringing it up, and how far you need to learn before it's enough. The articles aim to read like explanations you can go through directly, not a glossary.
Where to Start Reading
If you just want to build a map first, you can start with AI / Artificial Intelligence, ML / Machine Learning, DL / Deep Learning, and LLM / Large Language Model. These pieces will help you distinguish that "AI is the broad concept, machine learning is a class of methods, and large language models are just the most visible kind of model today." Next, read GenAI / Generative AI, ChatGPT Images / Image Features, and Multimodal AI, and you'll understand why today's AI does more than classify and predict—it can write text, generate images, and understand many kinds of content.
If you're already using ChatGPT, Claude, or other AI tools, we suggest reading Prompt, Token, Context Window, Inference, and Hallucination first. They explain the phenomena you most often encounter in daily AI use: why phrasing the same question differently gives different results, why long conversations forget things, and why a model can be confidently wrong.
If you want AI to answer based on your own materials, or to put AI into a workflow, you can continue with Embedding, Vector Database, Knowledge Graph, RAG / Retrieval-Augmented Generation, GEO / Generative Engine Optimization, Agent, Tool Calling, AI Workflow, Harness Engineering, MCP / Model Context Protocol, and A2A / Agent2Agent. Reading straight through, you'll see how AI goes from a "chat window" to a "system that gets things done—one that can look things up, call tools, and collaborate with other assistants."
If you care about how AI services are accessed, priced, deployed, and secured, then read API / Application Programming Interface, API Key, API Relay, AI Gateway, Model Routing, VPN / Virtual Private Network, Local Model, and Edge AI. These words are more like utilities and access-control systems: you don't necessarily see them every day, but they determine whether AI can run reliably, securely, and at low cost.
If you're already putting AI into real projects, the next layer is engineering practice: Context Engineering covers what belongs in a finite window, Context Compaction covers what to do when a long session fills up, Subagent covers isolating verbose work, Agent Skills covers fixing procedural knowledge in place, Evals covers telling whether a change made things better or worse, and Loop Engineering strings these into a loop that runs itself.
As soon as AI touches external content or triggers real actions, read Prompt Injection and Guardrails alongside it—they explain why a rule written in a prompt isn't a constraint. To understand why models keep getting stronger and cheaper, read MoE / Mixture of Experts and Test-time Compute; if robotics and embodied AI interest you, add World Model.
If you want to know how these models actually work inside, read Transformer Architecture and Attention Mechanism — they explain why context windows have a ceiling and why long inputs cost more. Then Chain of Thought explains what "thinking…" is really doing, Diffusion Model and AI Video Generation cover the image and video side, and Scaling Law and AGI / Artificial General Intelligence cover the basis for — and the dispute over — models continuing to improve. If robotics interests you, follow with Embodied AI.
If you're already wiring AI into a team's daily work, a few entries are firmly practical: Agent Memory on remembering the right things across sessions, GraphRAG on questions that span documents, AGENTS.md and Spec-Driven Development on making project rules durable, AI Observability on debugging when something goes wrong, and Shadow AI on governing AI use that bypasses approval. On the cost side, see Quantization, SLM / Small Language Model, Open Weights, and LoRA / Low-Rank Adaptation.
Finally, if you want to understand how models are adapted to fit a particular task, read Model Parameters, Fine-tuning, Reinforcement Learning, Distillation, and Synthetic Data. And if you're interested in "writing code with AI," save Vibe Coding and Harness Engineering for last—they reconnect the earlier prompts, tool calling, context, and workflows to a real usage scenario.
Topic Roadmap
An AI Short Story: Stringing the Terms Together
Story One: First, Open a Shop That Learns
You can start by picturing the AI world as a "smart content shop" that's being upgraded. AI / Artificial Intelligence is the owner's overall goal: to let machines do some things that once required human judgment, such as recognizing what a customer wants, recommending products, writing copy, reading images, and organizing materials. ML / Machine Learning is like letting the staff draw lessons from past orders instead of just following rigid rules; DL / Deep Learning is like giving the staff a more complex brain that can learn finer patterns from large volumes of images, audio, and text.
Later the shop brings in a head assistant who is especially good at reading, writing, and conversation—this is the LLM / Large Language Model. It doesn't just look things up in a dictionary; it can organize language based on context, explain problems, and draft content. Model Parameters can be understood as the "traces of experience" this assistant retains after long training, which determine roughly how it will understand and express things.
Story Two: The Head Assistant Starts Writing Copy and Making Images
When this head assistant does more than analyze information—when it starts writing articles, making summaries, and generating images—it enters the realm of GenAI / Generative AI. ChatGPT Images / Image Features is like the shop's design assistant: you describe a poster, an illustration, or an edit in natural language, and it tries to generate the visual content.
The instructions you give the AI are the Prompt. It's like telling a staff member, "Make this poster a bit warmer, aimed at beginners, and don't make the headline too stiff." But there's a limit to how much this assistant can hold in mind at once—that's the Context Window. And in its eyes, text isn't processed whole article by whole article; it gets cut into small pieces called Tokens.
Story Three: Connecting the Assistant to a Materials Warehouse
Relying only on the head assistant's own memory, the shop makes mistakes easily. It might mix up inventory, prices, and policies, or even produce a Hallucination: fluent, but factually wrong. So the shop connects it to a database—that's RAG / Retrieval-Augmented Generation: first go find materials on the shelves and in the file cabinets, then answer based on those materials.
To let the machine know which materials are "similar in meaning," you use Embedding to turn text into comparable numbers; these vectors usually go into a Vector Database, like a materials warehouse arranged by meaning. If you also want the assistant to know "who is related to whom and how," you use a Knowledge Graph to organize the connections among people, products, concepts, and processes into a web of relationships. GEO / Generative Engine Optimization is like reorganizing the shop's signage and product descriptions so the AI assistant can more easily understand, cite, and recommend your content.
Story Four: Turning the Assistant from "Answering" to "Doing"
When AI does more than answer—when it starts breaking down tasks for you, looking things up, calling tools, checking results, and deciding the next step—it becomes more like an Agent. Tool Calling is like it knowing when to grab a calculator, check the inventory system, send an email, or generate a spreadsheet; AI Workflow strings "take the order, look things up, write a draft, review, publish" into a fixed process. Harness Engineering goes further and cares about this AI assistant's working environment: how context, tools, permissions, testing, logging, and human confirmation are designed so that it not only can do things but does them in a controllable way.
If the shop has several AI assistants—each handling customer service, copywriting, data analysis, and technical support—they need a language for collaboration, and that's A2A / Agent2Agent. And MCP / Model Context Protocol is more like a unified plug standard, making it easier for different AI assistants to connect to tools, files, and databases.
Story Five: Connecting the Shop to the Real World
For the shop to actually operate, it can't do without access and security. API / Application Programming Interface is like a service window between software programs; external systems place orders and retrieve results through it. API Key is like a pass, used to prove "who is calling the service." If you don't queue up at every model provider directly but instead hand off requests through a single counter, that's an API Relay.
As requests grow, the AI Gateway acts like a front-desk manager, uniformly handling permissions, logging, cost, rate limiting, and security; Model Routing is like assigning different staff based on the type of question: simple questions go to cheap, fast models, and complex tasks go to more capable ones. A VPN / Virtual Private Network is like building an encrypted tunnel for remote employees so they can access internal systems more securely.
Story Six: Making the Assistant Cheaper, Smoother, and Closer to the Scene
Finally, some words are about "how a model becomes better suited to real-world use." Fine-tuning is like having the head assistant specifically learn your shop's phrasing, formats, and business habits; Reinforcement Learning is like it receiving feedback after doing a task and gradually learning which approaches work better; Distillation is like distilling a senior employee's experience into a lighter training manual, so a small model can also learn part of the capability.
Inference is the process where, after training is complete, the model actually gets to work answering questions, generating images, or executing tasks. If you run the model on your own computer, that's a Local Model; if the model runs near a phone, camera, car, or factory equipment, that's Edge AI. When AI understands text, images, speech, and video all at once, it's Multimodal AI. And Vibe Coding can be pictured as no longer tightening every screw by hand—instead, you tell the AI "I want a certain kind of software," then try it out and have it make changes as you go.
Read these stories together, and the terms are no longer isolated acronyms. They're actually telling the same story: how we train an assistant that understands and generates content, how we give it materials and tools, how we connect it to real systems, how we control cost, security, and errors, and finally how we fit it into daily work.
All Articles
Basic Concepts
- AI / Artificial Intelligence: The starting point for every concept.
- ML / Machine Learning: How models learn from data.
- DL / Deep Learning: Why neural networks transformed image, speech, and language processing.
- LLM / Large Language Model: The core capability behind products like ChatGPT.
- GenAI / Generative AI: From "analyzing information" to "creating content."
- ChatGPT Images / Image Features: Generating and editing images with natural language.
- Prompt: The basic skill of communicating with a model.
- Token: The small pieces AI cuts text into as it reads.
- Context Window: How much information a model can reference at once.
- Model Parameters: The internal values a model learns after training.
- Inference: The process where a trained model actually produces results.
- Multimodal AI: Understanding text, images, sound, and video at the same time.
- Reasoning Model: Models like o1, o3, and DeepSeek R1 that "think slowly."
- Agentic AI: The line between producing content and producing action.
- MoE / Mixture of Experts: A very large model that only uses a small slice each time.
- Test-time Compute: Spending more compute at the moment of asking to buy accuracy.
- World Model: A simulator that predicts how an environment changes under actions.
- Prompt Caching: Reusing the same prefix to greatly reduce API call costs.
- AI Search: Products like Perplexity and SearchGPT that generate answers while searching.
- Transformer Architecture: The skeleton shared by nearly every large model today.
- Attention Mechanism: The component that decides where a model looks at each step.
- Diffusion Model: Denoising a field of static step by step into images and video.
- Chain of Thought: Letting a model show its working before answering.
- Scaling Law: Quality improves along a predictable curve as parameters, data, and compute grow.
- AGI / Artificial General Intelligence: Still undefined, so "when it arrives" is mostly an argument about definitions.
- Embodied AI: AI with a body that acts and learns from consequences.
- AI Video Generation: Denoising across space and time to produce coherent footage.
Knowledge and Retrieval
- RAG / Retrieval-Augmented Generation: Letting a model answer based on external materials.
- Embedding: Turning "similar in meaning" into numbers a machine can compare.
- Vector Database: Storing and searching materials by "similarity in meaning."
- Knowledge Graph: Organizing knowledge with entities and relationships.
- GEO / Generative Engine Optimization: Making content easier for AI answers to understand and cite.
- GraphRAG: Retrieving along entity relationships to answer questions that span documents.
Agents and Protocols
- Agent: From chatting to executing tasks.
- MCP / Model Context Protocol: The "plug standard" that lets AI find tools and materials.
- A2A / Agent2Agent: The "communication rules" that let multiple AI assistants work together.
- Tool Calling: Letting a model request external tools to get things done.
- Computer Use: AI directly operating the mouse and keyboard, clicking web pages and app interfaces.
- Subagent: A stand-in sent to look one thing up, returning only the conclusion.
- Agent Skills: Packaging procedural knowledge for an AI with SKILL.md.
- Agent Memory: Persisting information outside the window and retrieving it on demand.
- Voice Agent: Real-time voice AI that listens, speaks, and actually gets things done.
Engineering Practice
- Harness Engineering: Designing a controllable, verifiable working environment for an AI Agent.
- Loop Engineering: Moving from prompting an Agent to designing loops that run themselves.
- Context Engineering: Deciding what goes into a finite window each turn.
- Context Compaction: Swapping history for a summary when a long session fills up.
- Evals: Turning "it feels worse" into a comparable score.
- AI Workflow: Stringing multiple AI steps into an automated process.
- Vibe Coding: Directing AI in natural language to quickly generate and modify code.
- Spec-Driven Development: Treating versioned specifications as the single source of truth.
- AGENTS.md: The project file written for coding agents.
- AI Observability: Recording model calls and agent decisions as queryable traces.
Access and Deployment
- API / Application Programming Interface: The "service window" through which software programs talk.
- API Relay: The "transfer counter" that unifies access to multiple model services.
- API Key: The common "pass" used when calling AI services.
- Model Routing: Assigning requests to the right model based on the task.
- AI Gateway: A single entry point that manages model access, permissions, cost, and security.
- Local Model: Running a model on your own computer.
- VPN / Virtual Private Network: Understanding network access, secure tunnels, and remote connections.
- Edge AI: Running models on devices close to the data.
- Quantization: Trading a little precision for memory and speed.
- SLM / Small Language Model: Specialized models that run on consumer hardware.
- Open Weights: Downloadable, self-hostable models — not the same thing as open source.
Training and Optimization
- Fine-tuning: Making a model more accustomed to a certain way of speaking or a task format.
- Reinforcement Learning: Learning strategies through feedback from actions.
- Distillation: Distilling a large model's capabilities into a small model.
- LoRA / Low-Rank Adaptation: Freezing the base model and training a small set of side parameters.
- Synthetic Data: Manufactured training material for where real data falls short.
Safety and Governance
- Alignment: How to make AI behavior consistent with human intent and values.
- Hallucination: Why models can be confidently wrong.
- Prompt Injection: The top risk, where commands are disguised as content.
- Guardrails: Validation layers running outside the model — prompts don't count.
- Shadow AI: Employees using AI on company data without approval — bans rarely work.