跪拜 Guibai
← Back to the summary

The AI Buzzword Stack, from LLM to Agent

Foreword

I wonder if you've had this feeling:

Open any tech community or GitHub Trending now, and 7 out of 10 entries are basically AI-related:

Agent, skill, MCP, RAG, Vibe Coding, Context Engineering…

If you don't pay attention for a few days, you feel like you've already fallen behind...

What's worse is that many articles, when explaining one term, assume you already understand ten other terms.

To look up one word, you first have to look up five new words. If you can't understand one word, you can't continue reading.

So I've compiled a list of AI buzzwords, hoping to string together these high-frequency concepts in one article.

The order is also important: From the model itself → Model enhancement → Agent → Engineering → AGI

The concepts in the front are the foundation for the ones in the back.

To make it easier to understand, the whole article uses one analogy:

Think of a large model as a very smart intern.

He is very capable, but also has some limitations.

1. First, Get to Know AI's "Brain"

LLM (Large Language Model)

LLM, Large Language Model.

The core models behind AI products like ChatGPT, Claude, and DeepSeek all fall into this category.

Simply put: An LLM is a model trained on massive amounts of text data. Its most basic task is:

Predict the most likely information to appear next based on existing content.

For example:

The weather today is really...

The model will judge, based on its training experience, that the following might appear:

good, hot, cold...

Of course, today's large models have long surpassed simple text completion.

After large-scale training, it demonstrates:

So our AI intern:

Has read a vast amount of material, has a broad knowledge base, and can handle many problems.

Token: The Unit of Measurement in the Model's World

The model doesn't understand text directly.

Before text is input into the model, it is split into Tokens.

A Token can be understood as:

The basic unit of information processing for the model.

For example, the sentence:

I love AI

Might be split into:

I | love | AI

Three Tokens.

Chinese splitting is usually finer, for example:

人工智能很强 (Artificial intelligence is very strong)

Might be split into:

人工 | 智能 | 很 | 强 or

人 | 工 | 智 | 能 | 很 | 强

Why is Token important?

So a Token is like the "data traffic" in the model's world.

The more information you provide, the more Tokens are consumed.

Context Window: How Much the Model Can See at Once

Context Window, also known as the context window.

It represents:

The total amount of information a model can process in a single interaction.

Usually includes:

For example, if a model supports:

200K Context Window

It means it can process up to about 200,000 Tokens of context at a time.

Many people wonder: "Why does the AI forget things we talked about earlier after chatting for a long time?"

Actually, it's not that it forgot.

It's that: That content is no longer within the current context range.

It's like the intern only has one A4 sheet of paper on their desk.

Once it's full, old content must be erased before new content can be recorded.

Prompt: How to Assign Tasks to the Model

A Prompt is the instruction you send to the model.

For example: A basic Prompt:

Write a summary.

An advanced Prompt:

You are a technical lead. Please summarize the following proposal into 5 points for a CTO report.

Why is Prompt valuable?

Because the model needs to know:

2. Giving the Model Plug-ins

Although the model is smart, it has several problems:

The following technologies are designed to solve these problems.

RAG: Giving the Model an External Knowledge Base

Chinese: Retrieval-Augmented Generation.

Solves the problem:

The model doesn't know data outside its training.

For example:

The model didn't have this content during training.

The RAG process:

  1. Pre-process and chunk the materials
  2. Convert them into vectors
  3. Store them in a vector database
  4. Search for relevant content when a user asks a question
  5. Provide the results to the model for answering

This way, when the model answers, it no longer relies solely on training memory but can reference the latest materials.

Similar to:

Before answering a question, the intern can go to the company's document library to look up files.

Embedding: Turning Text into Numbers

Embedding is an important technology in RAG.

Simply put:

Embedding is converting text into numerical vectors that computers can understand.

For example, 苹果手机 (Apple phone) and iPhone.

Although the text is different, the meaning expressed is very close.

Through Embedding, the system can understand the semantic relationships between content and find information relevant to the user's question from a large amount of material.

However, actual RAG retrieval doesn't rely solely on vector similarity.

Engineering typically combines:

Simply put:

Embedding is responsible for understanding "is the meaning similar?", keyword search is responsible for matching "does this word exist?", and combining both allows for more stable retrieval of needed information.

Fine-tuning: Retraining Model Behavior

Fine-tuning.

If RAG is:

Giving the intern a document to look up.

Then Fine-tuning is:

Retraining this intern to change how they work.

For example, making the model:

Simple distinction:

RAG changes:

What the model knows.

Fine-tuning changes:

How the model does things.

Tool Calling: Teaching the Model to Use Tools

The model itself can only generate text.

It cannot:

The Tool Calling method:

Tell the model in advance:

You can use these tools.

For example:

getWeather(city)

The model judges:

"I need to check the weather."

Then generates a call request.

What actually executes is your program.

So the model is responsible for the decision, and the code is responsible for the execution.

Similar to:

It's like this intern knows "I need to check the weather," but he won't write code to call the weather API himself. Instead, he tells the system: "Please help me call the weather query tool." After the system executes it and gives him the result, he continues to complete the task.

MCP: A Unified Protocol for Connecting AI and Tools

The problem it solves:

Before:

Every AI application connecting to different tools required separate development.

For example:

Write one set of interfaces for Tool A.

Write another set for Tool B.

Costs were high.

MCP aims to establish a unified protocol:

Before:

Every device had a different interface.

Later:

Everyone unified around USB-C.

MCP's goal is to create a unified connection method between AI and tools.

Memory: Giving AI a Manageable Memory System

The problem Memory solves:

An Agent needs to continuously utilize past information across multiple tasks.

The model itself doesn't truly remember what happened before.

Memory's approach is:

Extract important information from the current context and save it to external storage.

For example:

When needed for subsequent tasks, the system retrieves the relevant information and provides it back to the model.

So Memory doesn't mean the model "has memory," but rather adds an information management mechanism to the Agent.

Similar to:

An intern won't automatically remember all project details just because a new day starts, but the company keeps project documents and work records. When they rejoin the project, they can first check the relevant materials before continuing work.

Skill: Giving AI a Set of Reusable Work Capabilities

A Skill can be understood as:

A set of task processing specifications and procedures prepared for the AI.

It usually includes:

For example:

A Skill doesn't require retraining the model.

It's more like adding a layer of "work method library" outside the model:

When an Agent encounters a certain type of task, it loads the corresponding Skill, allowing the model to complete the work according to the predefined method.

Similar to:

When an intern joins the company, the company doesn't retrain a person from scratch but gives them work manuals for different positions. When encountering a corresponding task, they first refer to the relevant process before starting to handle it.

3. From Chatbot to Agent

Agent: AI That Can Autonomously Achieve Goals

A regular chatbot: You ask a question, the AI answers.

An Agent: You give it a goal.

It will, on its own:

The core difference:

Who decides the next step?

If the process is hardcoded in advance: That's automation.

If the model decides dynamically: That's closer to an Agent.

ReAct: The Classic Working Mode of an Agent

ReAct: Reason + Act.

That is: Think → Act → Observe → Think again.

For example, fixing a bug:

  1. Analyze the error
  2. Modify the code
  3. Run tests
  4. Continue adjusting based on results

It doesn't generate an answer once, but continuously loops.

Workflow: AI with a Fixed Process

Workflow: A pre-defined process.

For example:

User question

↓

Classification

↓

Retrieve materials

↓

Generate answer

↓

Manual review

Each step's path is designed in advance.

Advantages:

Disadvantages:

4. The New Work of Engineers in the AI Era

Prompt Engineering: Making the Model Better Understand Tasks

Simply put, it's the study of:

How to design the instructions input to the model so that it produces more accurate and stable results.

In the early days of using large models, people discovered:

With the same model and the same question, just changing the way it's expressed could lead to vastly different results.

For example, a basic question:

Help me optimize this code.

The result might be quite generic.

A better Prompt:

You are a senior frontend engineer. Please analyze the issues in the following React code and provide optimization suggestions from three aspects: performance, maintainability, and code standards, and provide the modified code.

The more explicit the information the model receives, the more the output usually meets expectations.

Prompt Engineering mainly focuses on several aspects:

However, the problem Prompt Engineering mainly solves is:

How to better tell the model what to do.

When tasks become increasingly complex, just optimizing a single Prompt is no longer enough.

Because the model needs not just a question, but an entire working environment:

Thus, a larger concept emerged: Context Engineering

Context Engineering: Managing the Information the Model Sees

Essence:

What information should be provided to the model.

Includes:

In one sentence:

Prompt Engineering focuses on how to ask the model a better question.

Context Engineering focuses on how to design the entire information environment needed for the model to complete a task.

Harness Engineering: Building the AI's Working Environment

Harness can be understood as:

A complete operational framework built around the model, allowing AI to complete tasks safely and stably.

It usually includes:

Why is Harness needed?

Because the same model, placed in different Harnesses, can perform completely differently.

A bare model might only be able to answer questions;

But with tools, processes, memory, and security mechanisms added, it can become an Agent capable of executing complex tasks.

Similar to:

The engine determines the car's power, but the body, steering wheel, brakes, and control system determine whether the car can ultimately drive stably.

The model is the engine, and the Harness is the entire car.

Loop Engineering: Optimizing the Agent Loop

The core of an Agent is the continuous loop: Think → Act → Observe.

Loop Engineering focuses on:

Preventing the Agent from getting stuck in an infinite loop.

5. The End Goal: AGI

AGI (Artificial General Intelligence). Currently, there is no unified definition for AGI.

It is generally considered: It should possess human-like general abilities:

Today's large models are more like:

A super intern who is excellent in many fields.

But there is still a long way to go before reaching an intelligent system that can truly engage in long-term autonomous learning and adapt to the real world.

Final Words

After reading these concepts, you'll find that although there are more and more new terms in the AI field, the core changes behind them are not actually that complicated.

In the past, we were mostly using AI: Ask a question, and let the model generate an answer.

Now, AI is gradually transforming from a "chat tool" into an intelligent assistant that can understand tasks, call tools, process information, and assist in completing work.

For developers, what's important in the future is not just knowing how to write Prompts or call model APIs, but understanding the boundaries of a model's capabilities and learning how to design a better working environment for AI.

Technical terms will continue to emerge, but what's truly worth mastering is not memorizing every acronym, but understanding what problem they solve.

When you know why a technology emerged, it naturally stops feeling unfamiliar.