IN THE MEANTIME, I MET WITH SOME OF MY OLD COLLEAGUES AND FRIENDS. YOU KNOW I'M INTERESTED IN DOING WHAT I'M DOING ABOUT AI, AND CHATTING IS OFTEN UNWITTINGLY BYPASSING AI。
But talking, I found a problem。
I'll say it naturally, Agent, Skill, MCP, and I think you've all heard it. And the person sitting across from me would stop and ask me, "What the hell is Agent?" "Skill, what do you mean?"
It was at that moment that I realized that I had fallen into the curse of knowledge. When a word is read too much and used too well, it is also a fog when it is first seen。
A FEW MONTHS AGO, I WAS ALSO OVERWHELMED BY THESE ACRONYMS. I SENT OUT "AI ZERO BASE DAILY" AT THAT TIMEyoung whiteTEACHING, LEARNING A WORD EVERY DAY, HOLDING ON FOR ALMOST 30 DAYS, COMBING OUT A COMMON SET OF AI VOCABULARY。
I've been asking more and more people lately, and I'm just going to do it a little more。
I'VE ORGANIZED 66 OF THE HIGH-FREQUENCY AI TERMS I THINK, TO EXPLAIN THEM IN HUMAN TERMS。Every word tells you what it is, where it happens, and what it's best to remember。
You don't have to carry it down once. Read it again, next time I see LLM, RAG, Agent, MCP, API, CLI, at least know where it's probably in the AI world。
I. FIRST KNOW AI AND THE BIG MODEL
THIS GROUP ANSWERS TWO QUESTIONS: WHAT IS AI AND WHAT IS BEING DEALT WITH WHEN THE BIG LANGUAGE MODEL SPEAKS. IT'LL BE EASIER TO READ THE MAP FIRST。
One of them is called Artificial IntelligenceAI）
AI IS TO ENABLE MACHINES TO PERFORM TASKS THAT IN THE PAST USUALLY REQUIRE PEOPLE TO UNDERSTAND, JUDGE, PREDICT OR CREATE. MOBILE PHONE HUMAN FACE UNLOCKED, MAP FORECAST TRAFFIC BLOCKED, PLATFORM RECOMMENDED VIDEO, CHAT ROBOTS ANSWERED QUESTIONS, ALL OF WHICH BELONG TO AI. IT'S A BIG OVERALL NAME, AND CHAT TOOLS ARE JUST ONE OF THOSE APPLICATIONS。
Algorithm
ALGORITHM IS A CLEAR SET OF STEPS TO SOLVE THE PROBLEM, MUCH LIKE A RECIPE. THE RECIPE PROVIDES FOR FIRST CUTTING, THEN HOT POTS, AND LAST DRESSING, AND ALGORITHMS FOR FIRST READING, HOW TO COMPARE, AND LAST OUTPUT. A LOT OF TRADITIONAL ALGORITHMS DON'T TALK AT ALL, AND AI ALGORITHMS ARE JUST PART OF THE GREAT FAMILY OF ALGORITHMS。
3 NLP (Natural Language Processing, Natural Language)
NLP IS A TECHNICAL AREA FOR COMPUTER PROCESSING OF HUMAN LANGUAGES, WHERE TRANSLATION, SUMMARY, SEARCH, VOICE-TO-TEXT AND CHAT ARE ALL RELEVANT. YOU ENTER AN ORAL QUESTION THAT AI UNDERSTANDS AND ANSWERS, AND USES NATURAL LANGUAGE PROCESSING SKILLS BEHIND IT。
4 LLM (Large Language Model, Large Language Model)
LLM is a language model that is trained in a great deal of text and a great deal of computing, and is good at understanding and generating text. Common models or model families include OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek, Qwen, etc. And here's the product and the model: ChatGPT is the product you opened up for, and GPT is the model family that might be called on behind it。
5 |SLM (Small Language Model, Small Language Model)
SLM is a smaller and less demanding language model and does not currently have a globally harmonized parameter line. Common examples include the Microsoft Phi series and small-parameter versions of model families such as Gemma and Qwen. It may have a narrower scope of knowledge, but can be faster, cheaper, and more easily run on mobile phones or ordinary computers when doing fixed tasks such as abstracting, classification, equipment end-of-pipe assistants。
6|Token
Token is a small block that is used to read and write a text in a model and is not exactly equivalent to a Chinese word, a word or a pointer. Models break down sentences into a lot of Tokens before they are calculated, and fees and context capacity are often measured at Token. It's enough to write it down as an AI block used to process text。
7 |Context
Context is the full range of information currently available to AI, including your questions, previous conversations, uploading of files, system rules and tools. The same sentence, “I am not well”, as the doctor does, makes a completely different determination with regard to age, symptoms and examination reports. The clearer the context for AI, the better the answer。
8|Context Windows
The context window is the maximum amount of information a model can process at a time, usually expressed as Token. It's like a limited desk: there's an upper limit to the amount of information that can be distributed at the same time. When content cannot be released, the system can only delete, summarize, batch or get early information out of the window。
9 |Context Length
The length of the context is often used in two ways: the model allows a maximum of token or how much token is actually used in a request. The former is like the largest area on the desk, and the latter is like your actual distribution. When it is seen that the product promotes “excessive context”, the general reference is to maximum capacity。
10 |MLLM (Multimodal Large Language Model, Large Model)
MLLM CAN CONTINUE TO PROCESS INFORMATION SUCH AS PICTURES, SOUND OR VIDEO OUTSIDE THE TEXT. YOU GIVE IT A PICTURE OF THE REFRIGERATOR, IT SEES THE FOOD, THEN YOU ASK FOR A RECOMMENDATION FOR DINNER, AND YOU UPLOAD THE RECORDING OF THE MEETING, AND IT CAN SORT OUT THE FOCUS. THE “MULTIMODEL” HERE IS A VARIETY OF FORMS OF INFORMATION。
11 |VLM (Vision-Language Model, Visual Language Model)
VLM IS DEDICATED TO VISUAL AND LANGUAGE CONNECTIONS AND DEALS MAINLY WITH TASKS SUCH AS " VIEW AND UNDERSTAND OR ANSWER " . IT CAN DESCRIBE PHOTOGRAPHS, IDENTIFY CHART CONTENTS, AND LOOK FOR BUTTONS BASED ON SCREENSHOTS. VLM EMPHASIZES VISUAL-LINGUISTIC CO-OPERATION, WHICH IS USUALLY MORE SPECIFIC THAN A GENERIC MULTI-MODULAR MODEL。
12MoMoE (Mixture of Experts, Mixed Expert Model)
MoE will arrange for multiple groups of “experts” within the model, only part of which will be involved in the calculation each time. Like the hospital, skin problems go to dermatology, bone injuries go to osteoporosis, and all doctors are not required to consult at the same time. This expands model capabilities while controlling the amount of computing required for each operation。
After reading this group, you need to be clear:LLM mainly handles languages, VLM connects pictures and languages, and MLLM processes multiple information. Token is a text building block, Context is a reference for models this time。
II. MOST COMMON WORDS FOR CHATTING WITH AI
THIS GROUP IS AVAILABLE EVERY DAY. THE REMINDER IS FOR THE TASK AND PROVIDES INFORMATION IN THE CONTEXT, AND AI GENERATES ANSWERS BASED ON THAT INFORMATION。
13|Prompt
Prompt is all the input you gave to AI, either a sentence, a question, a file or a picture. The phrase “to summarize this contract as five risks” is a hint. Prompt doesn't have to write like a spell. Targets, materials, restrictions and output requirements are clear. It's easier for AI to get things right。
14|Prompt Engineering
Phrasing is the method of designing, testing and improving the phrasing. It is very much like a job assigned to a colleague: the first time the results were incorrect, the task was checked for clarity, the information was complete, the format was clear and the instructions were adjusted. Its value lies in reducing communication errors and repeating work。
15 |System Prompt (system hint)
System Prompt is AI that applies the high-priority rules set out behind it to define identity, tone, range of capabilities and safe boundaries. It is like an employee ' s manual for a company, and users usually do not see all the content. If your demands conflict with the system rules, the model will normally follow the system alert first。
16|User Prompt (User Hint)
User Prompt is the current user input requirement that you send in the chat box. The system alerts are like the rules of operation in a store, and the user alerts are like, "Please give me a cup of less sugar." Both will influence the outcome together, so the same sentence will be placed in different AI products and the answer may be different。
17|Instruation
Instract is an action in a message that explicitly requires AI to perform, such as summary, translation, comparison, classification or inspection. A complete reminder may also contain background information and qualifications. For example, “read the mail below, pick up the date, do not add information not in the original language”, which contains two instructions。
18° Markdown
Markdown is a format that uses simple symbols for layout text。
- ONE TP5T CAN REPRESENT THE TITLE
- – It's a list
- ** The text** may indicate a thicker。
A lot of AI outputs seem to be leveled out because of the use of Marktown; you often encounter it when you write a public number, document or project description。
19 JSON
JSON IS THE STRUCTURED DATA FORMAT COMMONLY USED BY THE PROGRAM, USUALLY CONSISTING OF BRACKETS, FIELD NAMES AND CORRESPONDING CONTENT. IT IS LIKE A FIELD-BASED SPREADSHEET THAT CAN BE READABLE AND THE PROGRAM CAN EASILY CONTINUE. ALLOW AI TO PRESS JSON OUTPUT, OFTEN TO GIVE THE RESULT TO ANOTHER SOFTWARE FOR AUTOMATIC USE。
20|Conversion History of Dialogue
THE HISTORY OF THE CONVERSATION IS AN EARLIER STORY IN THE CURRENT CONVERSATION. AI WAS ABLE TO FOLLOW UP ON THE TOPIC FIVE MINUTES AGO, USUALLY BECAUSE THE APPLICATION SENT THE PREVIOUS CHAT RECORD BACK TO THE MODEL. AFTER TOO LONG A CONVERSATION, EARLY CONTENT COULD BE CUT OFF OR SUMMED UP, SO AI SOMETIMES FORGETS WHAT YOU FIRST SAID。
21% Hallucination
THE ILLUSION IS THAT AI GENERATES WHAT SOUNDS REASONABLE, ACTUALLY WRONG OR NON-EXISTENT, SUCH AS A BOOK NAME, LINKS, DATA AND PERSONAL EXPERIENCE. IT FILLS THE GAP LIKE A PERSON WITH GOOD LANGUAGE SKILLS WHO DOES NOT WANT TO ADMIT THAT HE DOES NOT KNOW. THE ORIGINAL SOURCE MUST BE CHECKED WHEN IT COMES TO DATES, FIGURES, CITATIONS AND IMPORTANT DECISIONS。
The group should bear in mind:Prompt decides how you ask, and Context decides what AI can refer to, and Hallucination reminds you that your answer is still to be verified。
III. ENABLING AI TO READ: RAG AND SEARCH
WHEN YOU SEE "LET AI READ THE COMPANY'S FILE" "CREATE A PERSONAL KNOWLEDGE BASE," YOU'LL PROBABLY ENCOUNTER THIS PHRASE. THE ENTIRE ROUTE HAS ONLY ONE SENTENCE: FIND THE RELEVANT INFORMATION AND LET THE MODEL ANSWER。
22RARAG (Retrieval-Augmented General, Retrieval Enhancement)
RAG WILL ALLOW AI TO SEARCH FOR THE INFORMATION FROM THE SPECIFIED INFORMATION BEFORE ANSWERING, AND THEN GIVE THE SAMPLE TO THE MODEL ORGANIZATION. IT'S LIKE A ROLL-OPENING TEST: STUDENTS GO THROUGH TEXTBOOKS AND THEN ANSWER ON THEM. THE RAG CAN REDUCE THE NUMBER OF FABRICATIONS AND MAY SEARCH FOR FAULTY INFORMATION, SO MANUAL TESTING IS STILL REQUIRED. [2]
23|Knowledge Base (knowledge base)
The knowledge base is a collection of information that is centrally stored, collated and made available to the system. The company ' s product manuals, system documents and frequently asked questions come together and can serve as a guest knowledge base; personal notes, articles and books can also form a knowledge base. The accuracy and timeliness of the information directly affect the quality of responses。
24|Retrieval
THE SEARCH WAS TO IDENTIFY THE MOST RELEVANT ELEMENTS OF THE PROBLEM FROM A LARGE BODY OF INFORMATION. THE LIBRARIAN HEARD YOUR QUESTIONS AND BROUGHT THREE OF THE MOST APPROPRIATE BOOKS FROM TENS OF THOUSANDS OF BOOKS. THE RAG MODEL DOES NOT NORMALLY READ A COMPLETE DATABASE, BUT RETRIEVES A SMALL AMOUNT OF FOOTAGE BEFORE ANSWERING。
25|KeywordSearch
Keyword searches are mainly based on whether certain words appear directly. You search for “terms of refund”, which gives priority to what contains “refunds” and “durations”. This method is fast, checking product numbers, names and fixesTerminologyUseful; answers may be omitted when the question is very different from the original。
26|Semantić
Semantic search is more concerned about the proximity of two phrases. You search for something that doesn't want to go back, and it may find a document entitled "Return Policy". It is appropriate to deal with the issue of natural languages, but there may also be deviations in understanding, so important scenes are often accompanied by keyword searches。
27|Hybrid Seach (mixed search)
MIXED SEARCH USES KEYWORD SEARCH AND SEMANTIC SEARCH TOGETHER. IDENTITY CARD NUMBERS, CONTRACT NUMBERS ARE APPROPRIATE FOR PRECISE MATCHING; THE MEANING OF “HOW TO CLAIM REIMBURSEMENT” IS EVEN MORE IMPORTANT. THE TWO RESULTS ARE MERGED AND THEN SEQUENCED, TAKING CARE OF THE CONSISTENCY OF THE LETTER AND THE SIMILARITY OF THE MEANING, AS MANY RAG SYSTEMS DO。
IV. Agent, Tools and MCP
It's the hottest and easiest set of words to turn people around. Agent, Skill, MCP, which I talked to most of my friends, are all in this chapter。
Look at the shortest route:
You propose a target
Agent decides the next step
• Call a tool to read information or execute actions
Continue with the results
• Critical steps for human examination
28|Chatbot
Chatbot is a software that answers questions mainly through dialogue. Early chat robots may only respond by fixed keywords, and many Chatbots now access large models and answer more naturally. It continues to focus on “You ask, it answers”, and usually does not perform a long list of external operations on its own。
29|AI Asistant / Copilot
AI Assistant or Copilot is an AI product that helps people perform their tasks, such as supporting writing, organizing meetings, preparing tables or programming. The metaphor of “co-pilot” is good to remember: it can observe, warn and help, and the wheel remains in hand and the end result needs to be checked。
30|WorkFlow
Workflow is a pre-arranged and fixed step. When invoices are received, the amount is identified, then the form is written and the financial information is finally communicated, which is a workflow. What is done at each step is usually determined in advance, stable and well-checked; in cases outside the plan, it will not be as flexible as Agent to decide on the next step。
31 |AI Agent (AI smart body)
AI Agent is a system that can determine the next step, call tools, observe results and continue to act around a goal. Regular chat shows you how to buy tickets, and Agent can open the booking service, search the classes and fill in the information within the authorized range. It is actionable, so authority, error and manual confirmation are important. [4]
32|Agentic AI
Agenic AI is a type of system or design that emphasizes the ability of AI to plan, use tools and perform multi-step tasks on a continuous basis. The term is broad in scope and is not consistent with the different corporate names. When you see the product written “Agentic”, you can continue to ask three questions: what tools can be used, how many steps can you walk by yourself, and where manual identification is required。
33|Autonomy (autonomy)
Autonomy refers to the extent to which a system can decide and act on its own without a gradual command. Automatically organize a copy of a document, with lower autonomy and risk; automatic mail, payment or removal of information would significantly increase the risk. The more important the task, the more it becomes, the smaller the mandate and the more the confirmation point。
34 |Planning
Planning is a step to break down a target into one that can be implemented in a sequential manner. Checking traffic, booking of accommodation and making of baggage lists is done on the basis of confirmation of the date of preparation for travel. Agent also plans a complex mission; a well-written plan does not mean that the implementation will be successful, and the actual results will need to be seen for each step to be completed。
35|Tool (tool)
Tool is an AI-accessible external capability, such as web search, calculator, database, calendar and mail. The model itself is more like the brain responsible for judging, and the tools add hands and eyes to it. In the absence of a tool, the model can only respond in the context where it already exists; access to the tool gives it an opportunity to query and act。
36|Tool Calling / Function Calling
Tool call is a model that selects a tool in the prescribed format and fills in the required parameters, the program is then executed and the results are returned to the model. It is like a copy of a letter from the manager, and the staff members follow it. The model is responsible for judging what to call and the specific code is responsible for completing queries, calculations or writing. [3]
37 |Plugin (plugin)
Plugin is an extension package installed in a software to add functionality. The browser will be able to translate the web page with the installation of a translation plugin and AI will be able to connect to a file, calendar or third-party service with the application. Plugins are usually dependent on specific products and what can be done depends on the interfaces and privileges provided by the product。
38 |Skill
Skill is a reusable set of instructions, scripts and resources for Agent to stabilize certain types of tasks. It is like a standard operating manual available to staff: what information is read first, what steps are taken and how the results are checked. Prompt often solves a task, and Skill is better suited to repeat the same type of work。
39 |MCP (Model subject Protocol, model context protocol)
MCP IS AN OPEN-ENDED PROTOCOL FOR AI APPLICATIONS TO CONNECT EXTERNAL INFORMATION AND TOOLS. THE BEST THING TO REMEMBER IS.. AI WORLD USB-CIn the past, each tool had to fit alone, and capacity could now be provided according to the same connecting norm. It solves how to connect, and it doesn't determine mission targets for Agent. [5]
40|MCP Host (MCP Host)
MCP Host is an AI application that is actually used by a user, such as a desktop assistant, code editor or Agent platform. It manages users, models, security privileges and multiple connections. You can think of Host as a computer: it's the whole of the environment, and it's where MCP Clinic is。
41|MCP Clinic (MCP client)
MCP Clinic is located inside Host and is responsible for establishing and maintaining communications with a MCP Server. It will find out what tools and information Server has provided before sending calls. When a computer connects a printer, an internal communication channel is required to manage this connection。
42|MCP Server (MCP server)
MCP Server is the procedure for providing tools, information or tips to AI applications in accordance with MCP specifications. It's got a name. It doesn't mean it must be far away. It can run on your computer. For example, a local file, Server, allows AI to read the file in the authorized directory。
43|MCP Tool (MCP tool)
MCP Tool is a specific action provided by Server to AI to execute the request, such as searching for weather, creating a calendar, reading a database or sending messages. Each Tool will explain what parameters it needs. Query type actions are less risky, and actions such as sending, deleting, paying should be clearly authorized and confirmed。
44 |MCP Resources
MCP Resources is the content available to AI through MCP, such as documents, database records, project structure or application status. It's more like information on the table, Tool, more like a button that can be pressed. The readability of the resource is still dependent on access rights set by Host and Server。
45|MCP Prompt
MCP Prompt is a reusable reminder template provided by Server to help users start a mission in a fixed manner. It is a pre-packaged table template, pre-arranged field and task structure. It does not automatically execute the action and ultimately remains to be selected by the user and handed over to the model for processing。
46|Memory
Memoory is information that is maintained by the system and is reused at a later date, which may include current chat summaries, user preferences, historical tasks and long-term information. AI “Remember you” usually means that this information is stored and is again made available to models in subsequent dialogues. Memory affects the experience and also requires attention to privacy and the means of removal。
47|human in the Loop
A PERSON MUST BE CHECKED, SELECTED OR APPROVED AS TO THE KEY STEPS OF RETURN. AI COULD COMPILE A REFUND LIST AND COMPLETE A DRAFT, WHICH WOULD BE FINANCIALLY CONFIRMED PRIOR TO THE ACTUAL TRANSFER. THIS TAKES ADVANTAGE OF THE SPEED OF AI AND AVOIDS THE IMMEDIATE IRREVERSIBLE CONSEQUENCES OF THE SYSTEM WHEN INFORMATION IS INCOMPLETE OR MISCALCULATED。
48|Multi-agent (multi-intellectual)
Multi-intellectual bodies are involved in the division of labour between various Agents, such as a search, a draft article and a check reference. It is like a project team, which can handle complex tasks in parallel when the division of labour is reasonable; the more members, the more communication, the more duplication of work and the more mistransmissions, the better it is not Agent。
49 |A2A (Agent-to-Agent)
A2A means the mode of communication between Agent for the exchange of tasks, status and results. MCP is more concerned about how Agent connects tools and information, and A2A is more concerned about how one Agent gives the task to another Agent. The former could be described as “wielding equipment” and the latter as “crossing between colleagues”。
50 |Computer Use
Computer Use is for AI to operate the software by observing the screen, moving the mouse, clicking and entering. It's like a remote agent that can use web pages and desktop applications without a dedicated API. The interface changes make it misdirected and must be confirmed in case of payments, deletions and outgoing messages。
51 | Sandbox (sandbox)
Sandbox is an isolated environment that limits the scope of procedural activities. It is like drawing a fenced sand pit for the child, where it can be tried and where mistakes can be made and important documents and systems cannot be touched as much as possible. When Agent runs codes, downloads files or changes items, the sandbox reduces the impact, but is not a substitute for all security checks。
The chapter reads only one sentence:Agent is responsible for decisions and actions, Tool provides specific capabilities, Skill teaches it how to do it, MCP is responsible for connecting AI and external capabilities。
V. LOOK AT THE COMPUTER LANGUAGE THAT I'M GONNA FIND IN THE AI TEACH
These words are not invented exclusively by the AI field, but often appear when learning Agent, automation and local models. A lot of courses don't sound like foreign languages。
52 |API (Application programming interface)
API IS THE ENTRY POINT FOR EXCHANGE BETWEEN SOFTWARE AS AGREED. YOU DON'T BREAK INTO THE KITCHEN IN THE DINING ROOM, JUST THROUGH THE WAITER; API, LIKE THE SOFTWARE WORLD WAITER, TAKES YOUR REQUEST TO THE BACKSTAGE AND BRINGS BACK THE RESULTS. WEATHER APPLICATIONS, PAYMENT SERVICES AND LARGE MODELS CAN BE CALLED THROUGH OTHER PROGRAMS THROUGH API。
53|API Key
API Key is a secret voucher used by the program to access a service to identify who is on call and is often linked to the amount of use and billing. It's very much like a bank card code, and the people who get it could consume you. Do not put the whole Key in an open code, screenshot, tutorial or group chat。
54 SDK (Software Development Kit, Software Development Toolkit)
SDK IS THE CODE, DESCRIPTION, EXAMPLES AND COLLECTION OF TOOLS COMMONLY USED IN DEVELOPING A PLATFORM APPLICATION. API TELLS YOU, "HOW CAN YOU ORDER A LIST?" SDK DIRECTLY PREPARES MENUS, UTENSILS, AND DEMONSTRATIONS, SO THAT DEVELOPERS DON'T WRITE MUCH BASIC CODE. THE DIFFERENT PROGRAMMING LANGUAGES USUALLY HAVE THEIR OWN SDK。
55 |CLI (Command-Line Interface, command line interface)
CLI OPERATES THE COMPUTER INTERFACE BY ENTERING A TEXT COMMAND. GUI ALLOWS YOU TO CLICK NEW FOLDER, CLI ALLOWS YOU TO ENTER A LINE OF COMMANDS TO DO THE SAME THING. IT LOOKS LIKE A BLACK WINDOW, WHICH IS ACTUALLY JUST ANOTHER WAY OF DOING IT; THE INSTALLATION AND OPERATION OF THE AI OPEN SOURCE TOOL IS OFTEN USING CLI。
56|Terminal
Terminal is the window where commands are displayed and entered, and the "end" and Windows Telminal on Mac belong to this category. You can think of it as a chat window between you and the computer, and the CLI command is a message to the computer. Terminal is the window, the CLI is by word, and the two often appear together。
57 |Repository / Repo
Repository Common Repo is a central repository of project documents and history. It's like a project folder with a version of the record, with codes, instructions, pictures and configurations. It is often the project and its history that is copied to its computer。
58|Git
Git is a tool to document changes in document versions. It can tell you which elements have been modified and who has modified them, and it can retain versions at different stages. Saves every revision while writing an article, but functions more systematically. Git works locally, does not need to be connected and does not equal GitHub。
59 | GitHub
GitHub is a website that hosts Git warehouses and supports multi-person collaboration. The developers publish codes, write notes, report problems and co-modify projects. Git is a version management tool, GitHub is a platform for online hosting and collaboration; when learning the AI Open Source tool, a GitHub link is often obtained first. [6]
FIRST TIME I SAW CLI DON'T GET NERVOUS. BEFORE COPYING THE COMMAND, SEE WHAT IT READS, CHANGES, DELETES OR UPLOADS, AND PRACTICES WITH A COPY WHEN IT COMES TO IMPORTANT DOCUMENTS。
Common language for local models and hardware
IT'S ONLY WHEN YOU'RE READY TO RUN A MODEL ON YOUR COMPUTER THAT YOU NEED TO LOOK AT IT. IT'S GOOD TO KNOW WHAT IT MEANS TO USE ONLINE AI PRODUCTS。
60|Local Defense (local deployment)
Local deployment allows models or programs to operate on their own computers, mobile phones or corporate machines. The information can be used with less access to external services and when offline; installation, updating, performance and safety are handled on their own. It is appropriate for people who value data control or need a fixed environment and does not represent absolute natural security。
61 |Claud API
THE CLOUD API IS A MODEL THAT OPERATES ON A REMOTE SERVER FOR THE SERVICE PROVIDER, WHERE THE USER SENDS THE REQUEST OVER THE NETWORK AND GETS RESULTS. IT IS LIKE GOING TO A RESTAURANT TO ORDER FOOD, AND DOES NOT NEED TO BUY STOVES AND FOOD FOR ITSELF, AND IS EASILY ACCESSIBLE; IT IS ALSO INFLUENCED BY NETWORKS, PRICES, CALL LINES, PLATFORM RULES AND DATA POLICIES。
62|Open Source
OPEN SOURCES USUALLY INDICATE THAT THE SOURCE CODE IS PUBLICLY AVAILABLE UNDER A CERTAIN LICENCE AND THAT OTHERS CAN VIEW, MODIFY AND DISTRIBUTE IT TO THE EXTENT PERMITTED BY THE LICENCE. PUBLIC DOWNLOADS DO NOT AMOUNT TO FREE TRADE, AND LICENSING REQUIREMENTS VARY WIDELY. WHEN YOU SEE OPEN SOURCE AI, YOU NEED TO CONTINUE TO SEE WHICH PARTS OF THE CODE, DATA AND WEIGHT ARE OPEN。
63% Open Weights
Open weights are the parameters that can be downloaded after model training has been completed and users have the opportunity to run or continue to adjust their own equipment. Training data, complete training codes and commercial rights are not always open at the same time. Open weights and complete open sources are therefore two things, depending on how they are used。
64|GPU (Graphics Processing Unit, Graphic Processor)
GPU WAS USED EXTENSIVELY FOR TRAINING AND RUNNING AI MODELS, AS IT USED TO HANDLE GRAPHICS AND TO PERFORM MANY SIMILAR CALCULATIONS AT THE SAME TIME. IT'S LIKE A BUNCH OF PEOPLE WHO CAN DO THINGS IN PARALLEL. THE GPU MODEL AND THE VISIBLE SIZE ARE USUALLY MORE VISIBLE THAN NORMAL OFFICE SCENES WHEN PURCHASING COMPUTERS TO RUN LOCAL MODELS。
65 | VRAM (VISIBLE)
VRAM IS THE HIGH-SPEED MEMORY USED BY GPU ITSELF, WHERE THE WEIGHTS AND INTERMEDIATE DATA OF THE MODEL WHEN RUNNING NEED TO BE PLACED. MODELS ARE LIKE A LARGE SET OF BUILDING BLOCKS, WITH VISIBLE MEMORY DESKTOPS; TOO SMALL A TABLE IS TOO SMALL TO RELEASE THE BUILDING BLOCKS, AND THE MODEL MAY NOT BE LOADED, OR A SLOWER, MORE SPACE-EFFICIENT APPROACH IS NEEDED。
66 |B (Billion, billions)
7B, 70B IN MODEL NAMES, USUALLY REPRESENTING ABOUT 7 BILLION, 70 BILLION PARAMETERS. PARAMETERS CAN BE UNDERSTOOD AS A LARGE NUMBER OF INTERNAL FIGURES LEARNED AFTER MODEL TRAINING. THE LARGER THE NUMBERS, THE GREATER THE RAM AND COMPUTING RESOURCES; IT DOES NOT DIRECTLY AMOUNT TO GREATER CAPACITY, AND THE QUALITY OF TRAINING AND SPECIFIC TASKS ARE EQUALLY IMPORTANT。
VII. If only 10 words are remembered
66 words don't need a back. After the first reading, remember the following 10:
- AI: the task of making machine processing require understanding, judgement, prediction or creation。
- LLM: Large models for understanding and generating languages。
- Token: a block used by the model to read and write text。
- Context: information the model can refer to this time。
- Prompt: YOU GAVE AI THE REQUEST AND INFORMATION。
- Hallucination:AI SAYS THE WRONG THING LIKE IT'S TRUE。
- RAG: Retrieving from specified information before organizing answers。
- Agent_: systems that allow for continuous decision-making and action around objectives。
- MCP: OPEN AGREEMENT TO CONNECT AI TO EXTERNAL INFORMATION, TOOLS。
- CLI: Operate the computer interface with a text command。
A few months ago, I started with a word I understood every day. You don't need to remember all the 66 words today, as long as the next time you see Agent, Skill or MCP, there's a general picture in your head, and the dictionary is working。
In the future, if you encounter a strange abbreviation, you can ask what it belongs to:Models, conversations, information, Agent, interfaces, hardware MOST OF THE AI ARTICLES WOULD BE MUCH BETTER READ。