Recently, I've been meeting with former colleagues and friends offline. Knowing that I'm working on AI content, they are all very interested, and our conversations often naturally revolve around AI.
But as we talked, I noticed a problem.
I would naturally mention terms like Agent, Skill, and MCP, assuming everyone had heard of them. However, the person across from me would pause and ask: "What exactly is an Agent?" "What does Skill mean?"
At that moment, I realized I had fallen into the "curse of knowledge." When you see and use a word too much, you forget that you were also confused the first time you saw it.
A few months ago, I was also overwhelmed by these English abbreviations. During that time, I posted "Daily AI Basics for Beginners," learning one word a day for nearly 30 days, combing through a batch of commonly used AI vocabulary one by one.
With more and more people asking lately, I decided to make this even more complete.
I have compiled 66 AI terms that I consider high-frequency and explained them in plain language. For each word, I'll tell you: what it is, where you'll encounter it, and what real-life scenario makes it easiest to remember.
You don't need to memorize them all at once. Just read through them once so that the next time you see LLM, RAG, Agent, MCP, API, or CLI, you at least know roughly where they stand in the AI world.
I. Understanding AI and Large Models First
This group answers two questions: What is AI, and what are Large Language Models processing when they speak? Once you understand this map, the following terms will be much easier.
AI is about making machines perform tasks that usually require human understanding, judgment, prediction, or creation. Face unlock on phones, traffic prediction on maps, video recommendations on platforms, and chatbots answering questions all belong to AI. It is a broad umbrella term; chat tools are just one type of application.
An algorithm is a clear set of steps to solve a problem, much like a recipe. A recipe dictates cutting vegetables first, then heating the pan, and finally seasoning; an algorithm dictates what the computer reads first, how it compares, and finally what it outputs. Many traditional algorithms cannot chat at all; AI algorithms are just one part of the algorithm family.
NLP is the technical field that allows computers to process human language. Translation, summarization, search, speech-to-text, and chatting are all related to it. When you input a colloquial question and the AI understands the general meaning and answers, it is using natural language processing capabilities.
LLMs are language models trained with massive amounts of text and computation, specializing in understanding and generating text. Common models or model families include OpenAI's GPT, Anthropic's Claude, Google's Gemini, as well as DeepSeek and Qwen. It's important to distinguish between the product and the model: ChatGPT is the product you open and use, while GPT is the model family it might call behind the scenes.
SLMs are language models with smaller sizes and lower computational requirements. Currently, there is no globally unified parameter boundary. Common examples include Microsoft's Phi series and the small-parameter versions of model families like Gemma and Qwen. Their knowledge range might be narrower, but for fixed tasks like summarization, classification, or on-device assistants, they can be faster, cheaper, and easier to run on phones or standard computers.
Tokens are the small chunks models use when reading and writing text. They are not exactly equal to one Chinese character, one English word, or one punctuation mark. The model breaks a sentence into many tokens before calculating. Billing and context capacity are also often counted by tokens. Just think of them as the "building blocks" AI uses to process text.
Context is all the information the AI can currently refer to, including your question, previous dialogue, uploaded files, system rules, and tool results. Like a doctor's consultation, the same phrase "I feel unwell" will lead to different judgments depending on whether there are age, symptoms, and test reports. The clearer the context provided to the AI, the more grounded its answer will be.
The context window is the maximum amount of information a model can process at once, usually expressed in tokens. It's like a desk of limited size: there's a limit to how many documents can be spread out at once. When content exceeds this, the system must truncate, summarize, process in batches, or let earlier information exit the window.
Context length often has two meanings: the maximum tokens a model allows, or the actual tokens used in a specific request. The former is like the maximum area of the desk; the latter is like the documents you actually have spread out this time. When products advertise "ultra-long context," they are usually referring to the maximum capacity.
MLLMs can process information beyond text, such as images, audio, or video. If you give it a photo of a fridge, it can see the ingredients and recommend dinner based on text requirements; if you upload a meeting recording, it might organize the key points after listening. "Multimodal" here refers to multiple forms of information.
VLM specifically connects vision and language, primarily handling tasks like "looking at a picture and understanding or answering with text." It can write captions for photos, identify chart content, or find buttons based on screenshots. VLM emphasizes the coordination of vision and language, and its scope is usually more specific than the general term multimodal model.
MoE arranges multiple sets of "experts" inside the model, letting only a portion of them participate in calculations each time. It's like a hospital triage: skin problems go to dermatology, bone injuries go to orthopedics, rather than having all doctors consult at once. This expands model capability while controlling the computation needed for each run.
After reading this group, you just need to distinguish: LLM mainly processes language, VLM connects images and language, and MLLM can handle multiple types of information. Tokens are text blocks, and Context is the material the model can refer to this time.
II. Most Common Terms When Chatting with AI
This group is used every day. Prompts are responsible for assigning tasks, context provides materials, and AI generates answers based on this information.
A prompt is the entire input you give to the AI, which can be a sentence, a question, a document, or an image. "Summarize this contract into five risks" is a prompt. Prompts don't need to be written like magic spells; as long as the goal, materials, constraints, and output requirements are clear, the AI is more likely to get things right.
Prompt engineering is the method of designing, testing, and improving prompts. It's much like assigning work to a colleague: if the first result is wrong, you check if the task was clear, if the materials were complete, and if the format was specified, then adjust the instructions. Its value lies in reducing communication errors and rework.
A system prompt is a high-priority rule set behind the scenes in an AI application to define identity, tone, capability range, and safety boundaries. It's like a company's employee handbook; users usually don't see all of it. If your request conflicts with system rules, the model will generally prioritize the system prompt.
A user prompt is the current requirement entered by the user—the content you send in the chat box. The system prompt is like the store's operating rules, while the user prompt is like "Please give me a hot drink with less sugar." Both affect the result, which is why the same sentence in different AI products might yield different answers.
An instruction is a clear action you ask the AI to perform within a prompt, such as summarizing, translating, comparing, classifying, or checking. A complete prompt can also include background information and constraints. For example, "Read the email below, extract the date, and do not add information not in the original text" contains two instructions.
Markdown is a format for typesetting text using simple symbols.
- # can represent a header,
- - can represent a list,
- \\text\\ can represent bold.
Many AI outputs look clearly layered because they use Markdown; you'll often encounter it when writing articles, documents, or project descriptions.
JSON is a structured data format commonly used by programs, usually consisting of curly braces, field names, and corresponding content. It's like an electronic spreadsheet with fixed fields that humans can read and programs can easily process. Asking AI to "output in JSON" is often to let another software use the results automatically.
Conversation history refers to the messages that appeared previously in the current chat. AI can continue a topic from five minutes ago usually because the application resends the previous chat records to the model. After a chat gets too long, early content might be truncated or summarized, so AI sometimes forgets what you said at the very beginning.
Hallucination is when AI generates content that sounds reasonable but is actually wrong or non-existent, such as fabricating book titles, links, data, or personal experiences. It's like a person with great language skills who refuses to admit they "don't know" and fills in the blanks. Always check original sources for dates, numbers, citations, and important decisions.
The most important things to remember from this group: Prompt determines how you ask, Context determines what the AI can refer to, and Hallucination reminds you that answers still need verification.
III. Letting AI Read Data: RAG and Search
When you see "letting AI read company files" or "building a personal knowledge base," you'll likely encounter this group of words. The entire path is just one sentence: find relevant materials first, then let the model answer.
RAG allows AI to search for relevant content from specified materials before answering, then hands the found snippets to the model to organize the answer. It's like an open-book exam: the student flips through the textbook first, then answers based on it. RAG can reduce fabrication, but it might also search for the wrong materials, so citations still need human checking.
A knowledge base is a collection of materials centrally stored, organized, and available for system queries. A company's product manuals, policy documents, and FAQs put together can become a customer service knowledge base; a person's notes, articles, and reading records can also form a knowledge base. Whether the materials are accurate and updated directly affects answer quality.
Retrieval is finding the most relevant content from a large batch of materials. A librarian listening to your question and bringing the three most suitable books from tens of thousands is retrieval. In RAG, the model usually doesn't read the entire database first but retrieves a few snippets to answer.
Keyword search mainly looks for whether certain words appear directly. If you search for "refund deadline," it will prioritize finding content containing "refund" and "deadline." This method is fast and good for checking product numbers, names, and fixed terms; it might miss answers if the questioning style and original wording differ significantly.
Semantic search focuses more on whether the meaning of two passages is close. If you search "I don't want this anymore, how to return it," it might also find a document titled "Return Policy." It's suitable for natural language questions, but understanding can deviate, so important scenarios often combine it with keyword search.
Hybrid search uses keyword search and semantic search together. ID numbers and contract numbers are suitable for exact matching; "how to apply for reimbursement" needs understanding of meaning. Results from both are merged and ranked, accounting for both literal consistency and similar meaning, a method many RAG systems adopt.
IV. Agent, Tools, and MCP
This is the most popular and often most confusing group of words recently. Agent, Skill, and MCP, which I mentioned most when chatting with friends, are all in this chapter.
First, look at the shortest path:
1You propose a goal2→ Agent decides the next step3→ Calls tools to read data or execute actions4→ Continues processing based on results5→ Human checks key steps
A chatbot is software that primarily answers questions through dialogue. Early chatbots might only reply based on fixed keywords; now many chatbots connect to large models for more natural answers. Its focus remains "you ask, it answers," and it usually doesn't independently complete a long string of external operations.
An AI Assistant or Copilot is an AI product that helps people complete tasks, such as assisting with writing, organizing meetings, making spreadsheets, or programming. The "copilot" metaphor is easy to remember: it can observe, remind, and help, but the steering wheel is still in the human's hands, and the final result needs human checking.
A workflow is a pre-arranged set of fixed steps. After receiving an invoice, identifying the amount first, then writing it into a table, and finally notifying finance is a workflow. Each step is usually determined in advance, making it stable and easy to check; it won't flexibly decide the next step like an Agent when encountering unplanned situations.
An AI Agent is a system that can decide the next step around a goal, call tools, observe results, and continue acting. A normal chat will tell you how to buy a ticket; an Agent can open a booking service, query flights, and fill in information within authorized limits. It can act, so permissions, errors, and human confirmation are very important.
Agentic AI refers to a class of systems or design approaches that emphasize AI's ability to continuously plan, use tools, and execute multi-step tasks. This term's scope is broad, and different companies use different names. When you see a product labeled "Agentic," you can continue to ask three things: what tools can it use, how many steps can it take on its own, and where must a human confirm.
Autonomy refers to the degree to which a system can decide and act on its own without step-by-step human direction. Automatically organizing a copy of a file has low autonomy and risk; automatically sending emails, making payments, or deleting data significantly increases risk. The more important the task, the more permissions should be narrowed and confirmation points added.
Planning is breaking a goal into steps that can be executed in sequence. Like preparing for a trip by confirming dates first, then checking transport, booking accommodation, and making a packing list. Agents also plan when encountering complex tasks; a beautiful plan doesn't guarantee successful execution, and each completed step requires checking actual results.
A tool is an external capability an AI can call, such as web search, calculator, database, calendar, and email. The model itself is more like the brain responsible for judgment; tools give it hands and eyes. Without tools, the model can only answer based on existing context; after connecting tools, it has the chance to query and act.
Tool calling is when the model selects a tool according to specified formats and fills in the required parameters, which the program then executes and returns the result to the model. It's like a manager filling out a work order for a staff member to operate. The model is responsible for judging "what to call," and specific code is responsible for completing the query, calculation, or writing.
A plugin is an extension package installed into a software to add functionality. A browser can translate web pages after installing a translation plugin; an AI application might connect to files, calendars, or third-party services after installing a plugin. Plugins usually depend on specific products, and what they can do depends on the interfaces and permissions provided by the product.
A skill is a set of reusable instructions, scripts, and resources that teach an Agent to stably complete a certain type of task. It's like a standard operating manual for an employee: what materials to read first, what steps to follow, and how to check results are all clearly written. A prompt often solves a one-time task, while a skill is more suitable for repeatedly executing the same type of work.
MCP is an open protocol that allows AI applications to connect to external data and tools. The best metaphor is the USB-C of the AI world: in the past, every tool had to be individually adapted; now they can provide capabilities according to the same connection standard. It solves "how to connect," but doesn't decide task goals for the Agent.
An MCP Host is the AI application the user actually uses, such as a desktop assistant, code editor, or Agent platform. It is responsible for managing users, models, security permissions, and multiple connections. Think of the Host as a computer: it is the entire usage environment and where the MCP Client resides.
The MCP Client is located inside the Host and is responsible for establishing and maintaining communication with a specific MCP Server. It discovers what tools and materials the Server provides and sends call requests over. Like a computer connecting to a printer, a dedicated communication channel is needed internally to manage this connection.
An MCP Server is a program that provides tools, materials, or prompt templates to AI applications according to MCP specifications. Having "Server" in the name doesn't mean it must be far away; it can also run on your computer. For example, a local file Server can let AI read files in authorized directories.
An MCP Tool is a specific action the Server provides for the AI to request execution, such as querying weather, creating a schedule, reading a database, or sending a message. Each Tool specifies what parameters it needs. Query actions have lower risk, while sending, deleting, or paying should have clear permissions and confirmation.
An MCP Resource is content provided through MCP for the AI to read, such as files, database records, project structures, or application states. It's more like materials placed on a desk, while a Tool is more like a button that can be pressed. Whether a resource can be read still depends on the access permissions set by the Host and Server.
An MCP Prompt is a reusable prompt template provided by the Server to help users initiate tasks in a fixed way. It's like a pre-installed table template, arranging the fields and task structure to be filled in. It doesn't automatically execute actions; it still needs to be selected by the user and processed by the model.
Memory is information the system saves and reuses later, which might include current chat summaries, user preferences, history tasks, and long-term data. When an AI "remembers you," it usually means this information is stored and provided to the model again in subsequent dialogues. Memory affects experience and also requires attention to privacy and deletion methods.
Human in the loop means key steps must be checked, selected, or approved by a person. AI can organize a refund list and fill out drafts, but the actual transfer is confirmed by finance. This utilizes AI's speed while avoiding irreparable consequences caused by incomplete information or wrong judgment.
Multi-agent is letting multiple Agents collaborate, for example, one retrieving data, one drafting an article, and one checking citations. It's like a project team; when division of labor is reasonable, it can handle complex tasks in parallel. However, more members also increase communication, redundant work, and error propagation, so more Agents aren't always better.
A2A refers to the communication method between Agents to exchange tasks, status, and results. MCP focuses more on how Agents connect to tools and data, while A2A focuses more on how one Agent hands a task to another. Think of the former as "connecting devices" and the latter as "handing over work between colleagues."
Computer Use is letting AI operate software by observing the screen, moving the mouse, clicking, and typing. It's like a remote operator who can use web and desktop applications without dedicated APIs. Interface changes might cause it to click the wrong place; human confirmation is mandatory for payments, deletions, and outgoing messages.
A sandbox is an isolated environment that limits a program's activity range. It's like a fenced-in sandpit for a child to play in; they can try things out, and if something goes wrong, it won't touch important files or the system. When letting an Agent run code, download files, or modify projects, a sandbox reduces the impact range but cannot replace all security checks.
Just remember one sentence for this chapter: Agent is responsible for deciding and acting, Tool provides specific capabilities, Skill teaches it how to do it, and MCP is responsible for connecting AI with external capabilities.
V. Computer Terms You'll Encounter in AI Tutorials
These words weren't all invented by the AI field, but they often appear when learning about Agents, automation, and local models. If you can understand them, many tutorials won't feel like a foreign language.
An API is an entry point for software to communicate according to an agreement. You don't have to break into the kitchen in a restaurant; you just order through a waiter. An API is like the waiter of the software world, taking your request to the back end and bringing back the result. Weather apps, payment services, and large models can all be called by other programs through APIs.
An API Key is a secret credential used when a program accesses a service, used to identify who is calling and often connected to usage limits and billing. It's like a bank card PIN; anyone who gets it can consume your quota. Do not put full Keys in public code, screenshots, tutorials, or group chats.
An SDK is a collection of code, instructions, examples, and tools commonly used when developing applications for a platform. While an API tells you "how to order," an SDK prepares the menu, utensils, and demonstration dishes, saving developers from writing a lot of basic code. Different programming languages usually have their own SDKs.
CLI is an interface for operating a computer by typing text commands. A graphical interface lets you click "New Folder," while a CLI lets you type a line of command to do the same. It looks like a black window, but it's actually just another way to operate; the installation and running of AI open-source tools often use CLI.
A terminal is the window where commands are displayed and entered, such as "Terminal" on Mac or Windows Terminal. Think of it as a chat window between you and the computer, where CLI commands are the messages sent to the computer. Terminal is the window, CLI is the method of operating via text; the two often appear together.
A repository, often shortened to Repo, is a place where project files and modification history are centrally saved. It's like a project folder with version records, where code, instructions, images, and configurations can all be placed. When someone says "clone the Repo," they usually mean copying this project and its history to your own computer.
Git is a tool for recording changes in file versions. It can tell you what content was modified and by whom, and it can keep versions from different stages. It's like saving every revision when writing an article, but more systematic. Git works locally and doesn't need to be online all the time, and it is not the same as GitHub.
GitHub is a website for hosting Git repositories and supporting multi-person collaboration. Developers publish code, write instructions, report problems, and jointly modify projects on it. Git is the version management tool, and GitHub is the platform providing online hosting and collaboration; when learning AI open-source tools, you'll often start with a GitHub link.
Don't be nervous the first time you see CLI. Before copying a command, check if it will read, modify, delete, or upload anything, and practice with a copy when involving important files.
VI. Common Terms for Local Models and Hardware
You only need to look closely at this group of words when you are ready to run models on your own computer. If you usually use online AI products, knowing the general meaning is enough.
Local deployment is running a model or program on your own computer, phone, or company machine. Data goes through fewer external services and can be used offline; however, installation, updates, performance, and security must be handled by yourself. It's suitable for those who value data control or need a fixed environment, but it doesn't mean it's naturally absolutely safe.
A cloud API is when the model runs on a service provider's remote server, and users send requests and get results over the network. It's like going to a restaurant to order; you don't need to buy the stove and ingredients yourself, making it easy to start; at the same time, it is affected by network, price, call limits, platform rules, and data policies.
Open source usually means the source code is public under a certain license, allowing others to view, modify, and distribute it within the scope allowed by the license. Public download doesn't mean it can be used commercially at will; different licenses have very different requirements. When you see "Open Source AI," you still need to see which parts of the code, data, and weights are actually open.
Open weights mean the parameters after model training can be downloaded, giving users the chance to run or further adjust them on their own devices. Training data, full training code, and commercial rights are not necessarily open at the same time. Therefore, open weights and full open source are two different things; how they can specifically be used depends on the model license.
GPUs were originally good at processing graphics and are also very suitable for completing large amounts of similar calculations simultaneously, so they are widely used to train and run AI models. It's like a large group of people who can solve problems in parallel. When buying a computer to run local models, the GPU model and video memory size are usually more important than in normal office scenarios.
VRAM is high-speed memory used by the GPU itself, where model weights and intermediate data need to be placed during operation. The model is like a large set of building blocks, and VRAM is like the tabletop; if the tabletop is too small, the blocks won't fit, and the model might fail to load or require slower, more space-saving methods.
The 7B or 70B in model names usually represents approximately 7 billion or 70 billion parameters. Parameters can be understood as the massive amount of internal numbers the model learned after training. Larger numbers usually mean more memory and calculation resources; it doesn't directly mean stronger capability, as training quality and specific tasks are equally important.
VII. If You Only Remember 10 Words
You don't need to memorize 66 words. After reading for the first time, remember these 10 first:
- AI: Making machines handle tasks that require understanding, judgment, prediction, or creation.
- LLM: Large models responsible for understanding and generating language.
- Token: Small chunks used by models when reading and writing text.
- Context: Information the model can refer to this time.
- Prompt: Requirements and materials you give to the AI.
- Hallucination: AI making wrong content sound very real.
- RAG: Retrieving from specified materials first, then organizing the answer.
- Agent: A system that can continuously decide and act around a goal.
- MCP: An open protocol connecting AI with external data and tools.
- CLI: An interface for operating a computer via text commands.
A few months ago, I also started by understanding one word a day. You don't need to remember all 66 words today; as long as a general picture appears in your mind the next time you see Agent, Skill, or MCP, this dictionary has already served its purpose.
In the future, when you encounter a strange abbreviation, you can first ask which category it belongs to: Model, Dialogue, Data, Agent, Interface, or Hardware? If you can put the word back into these six drawers, most AI articles will be much easier to read.
About the Author
