AI 行业 Top 50 术语词典 + Agentic AI 深度问答
AI 行业 Top 50 术语词典 + Agentic AI 深度问答
Topic: AI Industry Top 50 Glossary & Agentic AI Q&A
Category: AI Fundamentals
Level: 入门 → 中级
Last Updated: 2026-03-18
中文版
第一部分:AI 行业 Top 50 核心术语
以下术语按从基础到进阶的顺序排列,涵盖 AI 行业最常用的 50 个概念。
🔵 基础概念(1-10)
| # | 英文术语 | 中文术语 | 解释 |
|---|---|---|---|
| 1 | Artificial Intelligence (AI) | 人工智能 | 让机器模拟人类智能(如学习、推理、决策)的技术总称。 |
| 2 | Machine Learning (ML) | 机器学习 | AI 的子领域,让机器通过数据自动学习规律,而不是靠人写规则。 |
| 3 | Deep Learning (DL) | 深度学习 | 机器学习的子领域,使用多层神经网络来处理复杂模式(如图像、语音)。 |
| 4 | Neural Network (NN) | 神经网络 | 模仿人脑神经元结构的计算模型,由多层节点组成,层层传递和处理信息。 |
| 5 | Algorithm | 算法 | 解决特定问题的一组明确步骤和规则,是 AI 模型的数学基础。 |
| 6 | Training | 训练 | 用大量数据让模型学习规律和模式的过程,类比人类”刷题”。 |
| 7 | Inference | 推理 | 训练好的模型在新数据上做预测/生成的过程,类比”考试”。 |
| 8 | Dataset | 数据集 | 用于训练和评估模型的结构化数据集合。 |
| 9 | Model | 模型 | 从数据中学到的”知识”的数学表示,能对新输入做出预测。 |
| 10 | Parameter | 参数 | 模型内部的可学习变量,参数越多模型越大(如 GPT-4 有万亿级参数)。 |
🟢 模型与架构(11-20)
| # | 英文术语 | 中文术语 | 解释 |
|---|---|---|---|
| 11 | Large Language Model (LLM) | 大语言模型 | 在海量文本上训练的超大规模语言模型,如 GPT-4、Claude、Gemini。 |
| 12 | Transformer | Transformer 架构 | 2017 年 Google 提出的神经网络架构,是 GPT/BERT/LLM 的核心基础。 |
| 13 | Attention Mechanism | 注意力机制 | Transformer 的核心组件,让模型能”关注”输入中最重要的部分。 |
| 14 | Foundation Model | 基础模型 | 在大规模数据上预训练的通用大模型,可适配各种下游任务。 |
| 15 | Pre-training | 预训练 | 在大量无标注数据上训练模型学习通用知识的第一阶段。 |
| 16 | Fine-tuning | 微调 | 在预训练模型基础上,用特定领域数据进一步训练以适应特定任务。 |
| 17 | Transfer Learning | 迁移学习 | 将一个任务上学到的知识应用到另一个相关任务,减少训练成本。 |
| 18 | Embedding | 嵌入/向量表示 | 将文本/图像等转换为高维数字向量,使计算机能理解语义相似性。 |
| 19 | Tokenization | 分词/Token 化 | 将文本切分成模型能处理的最小单元(token),如”Hello”→[“Hel”,”lo”]。 |
| 20 | Diffusion Model | 扩散模型 | 通过逐步去噪生成图像/视频的模型架构,如 Stable Diffusion、DALL-E。 |
🟡 生成式 AI 与交互(21-30)
| # | 英文术语 | 中文术语 | 解释 |
|---|---|---|---|
| 21 | Generative AI (GenAI) | 生成式 AI | 能创造新内容(文本、图像、代码、音频)的 AI 系统。 |
| 22 | Prompt | 提示词 | 用户给 AI 的输入指令,AI 根据 prompt 生成响应。 |
| 23 | Prompt Engineering | 提示工程 | 设计和优化 prompt 以获得更好 AI 输出的技术和方法论。 |
| 24 | Context Window | 上下文窗口 | 模型一次能处理的最大 token 数量(如 128K tokens)。 |
| 25 | Token | Token | LLM 处理文本的基本单位,约等于 0.75 个英文单词或 0.5 个中文字。 |
| 26 | Temperature | 温度 | 控制 AI 输出随机性的参数:0=确定性高,1=创造性强。 |
| 27 | Hallucination | 幻觉 | AI 生成看似合理但实际上错误或虚构的内容,是 LLM 的核心挑战。 |
| 28 | Grounding | 接地/落地 | 将 AI 输出与可靠数据源关联,减少幻觉的技术。 |
| 29 | Chain-of-Thought (CoT) | 思维链 | 让 AI 逐步推理而非直接给答案的提示技术,提高复杂问题的准确性。 |
| 30 | Few-shot / Zero-shot Learning | 少样本/零样本学习 | 模型只需少量甚至零个示例就能完成新任务的能力。 |
🟠 Agent 与工具(31-40)
| # | 英文术语 | 中文术语 | 解释 |
|---|---|---|---|
| 31 | Agentic AI | 智能体 AI | 能自主规划、决策、使用工具并迭代完成复杂任务的 AI 系统。 |
| 32 | AI Agent | AI 智能体 | 具备感知-思考-行动循环能力的自主 AI 程序。 |
| 33 | Function Calling | 函数调用 | LLM 输出结构化的工具调用指令,由外部代码执行真实操作。 |
| 34 | Tool Use | 工具使用 | AI Agent 调用外部工具(搜索、API、数据库等)来扩展能力。 |
| 35 | ReAct (Reasoning + Acting) | 推理+行动 | Agent 的核心模式:思考→行动→观察→再思考,循环直到完成。 |
| 36 | Multi-Agent System | 多智能体系统 | 多个专业 Agent 分工协作完成复杂任务的系统架构。 |
| 37 | Orchestration | 编排 | 协调多个 AI 组件/Agent 协同工作的管理层。 |
| 38 | MCP (Model Context Protocol) | 模型上下文协议 | Anthropic 提出的开放协议,标准化 AI 与外部工具/数据的连接方式。 |
| 39 | Plugin / Skill | 插件/技能 | 为 AI 添加特定能力的可扩展模块(如搜索网页、操作数据库)。 |
| 40 | Human-in-the-Loop | 人机协同 | AI 在关键决策点请求人类审核和确认的安全机制。 |
🔴 高级概念与应用(41-50)
| # | 英文术语 | 中文术语 | 解释 |
|---|---|---|---|
| 41 | RAG (Retrieval-Augmented Generation) | 检索增强生成 | 先从知识库检索相关信息,再让 LLM 基于检索结果生成回答,减少幻觉。 |
| 42 | Vector Database | 向量数据库 | 专门存储和检索 Embedding 向量的数据库,是 RAG 的核心组件。 |
| 43 | RLHF (Reinforcement Learning from Human Feedback) | 人类反馈强化学习 | 用人类偏好反馈来优化模型输出的训练方法,让 AI 更符合人类期望。 |
| 44 | Alignment | 对齐 | 确保 AI 的行为和目标与人类价值观一致的研究方向。 |
| 45 | Responsible AI | 负责任的 AI | 确保 AI 系统公平、透明、安全、可解释的设计原则和实践。 |
| 46 | AGI (Artificial General Intelligence) | 通用人工智能 | 具备人类级别通用智能的 AI,能处理任何智力任务(尚未实现)。 |
| 47 | Edge AI | 边缘 AI | 在本地设备(手机、IoT)上运行的 AI,无需云端,低延迟高隐私。 |
| 48 | Multimodal AI | 多模态 AI | 能同时处理多种输入类型(文本+图像+语音+视频)的 AI 模型。 |
| 49 | Synthetic Data | 合成数据 | 由 AI 生成的用于训练其他模型的人造数据,解决数据稀缺和隐私问题。 |
| 50 | AI Governance | AI 治理 | 管理 AI 开发和使用的政策、流程和法规框架。 |
第二部分:Agentic AI 深度问答(Session 记录)
以下是一次关于 Agentic AI 的完整问答记录,从概念理解到实现原理再到实践路径。
Q1: 什么是 Agentic AI?
Agentic AI(智能体 AI)是指能够自主感知环境、制定计划、做出决策并采取行动来完成目标的 AI 系统。
与传统 AI(被动响应单次提示)不同,Agentic AI 的核心特征包括:
- 自主性:能独立分解复杂任务并逐步执行
- 规划能力:制定多步骤策略来达成目标
- 工具使用:调用外部工具(搜索、代码执行、API 等)
- 反思与迭代:评估结果,遇到失败时调整策略重试
- 持续交互:在多轮循环中与环境互动,而非一次性输出
简单类比:传统 AI 像一个回答问题的助手,Agentic AI 像一个能独立完成项目的员工——你给目标,它自己想办法做到。
Q2: Agentic AI 具体是怎么实现的?
核心是一个 循环架构(Agent Loop),基于 ReAct 模式:
用户目标 → 思考(Reason) → 行动(Act) → 观察结果(Observe) → 再思考 → ... → 完成
关键组件:
| 组件 | 作用 | 实现方式 |
|---|---|---|
| LLM 大脑 | 理解、推理、决策 | GPT/Claude 等大模型 |
| 工具系统 | 与外界交互 | Function Calling / Tool Use API |
| 记忆 | 维持上下文 | 对话历史 + 向量数据库 |
| 规划器 | 分解复杂任务 | Prompt 引导 / CoT / 任务图 |
| 反思机制 | 自我纠错 | 检查输出、重试策略 |
技术要点:
- 工具调用:LLM 输出结构化的函数调用指令,运行时执行后将结果返回给 LLM
- 多 Agent 协作:多个专业 Agent 分工,通过消息传递协同
- 编排框架:LangChain、AutoGen、CrewAI 等封装了上述模式
Q3: Agentic AI 是一个有代码写出来的软件吗?举个实现例子?
是的,本质就是用代码写出来的软件。 核心是一个循环程序 + LLM API 调用:
def agent_loop(user_request):
messages = [{"role": "user", "content": user_request}]
while True:
# 1. 调用 LLM API
response = call_llm(messages, tools=available_tools)
# 2. 任务完成 → 返回答案
if response.is_final_answer:
return response.text
# 3. 需要调用工具 → 执行工具
for tool_call in response.tool_calls:
result = execute_tool(tool_call)
messages.append({"role": "tool", "content": result})
# 4. 工具结果喂回 LLM → 回到第1步
具体例子——用户说”帮我写一篇 post”时的执行流程:
第1轮: LLM 决定调用 skill → 启动 knowledge-deep-dive
第2轮: LLM 调用 glob 了解文件结构 → 返回文件列表
第3轮: LLM 生成内容 → 调用 create 创建 markdown 文件
第4轮: LLM 调用 powershell → git add && commit && push
第5轮: LLM 判断完成 → 返回最终答案,循环结束
Agent = 普通代码(while 循环 + 工具)+ LLM API(大脑),没有魔法。
Q4: 我如果自己想创建一个 Agentic AI,可行吗?
完全可行! 门槛比想象的低。你需要准备:
| 需要 | 说明 | 成本 |
|---|---|---|
| LLM API | OpenAI / Azure OpenAI / Claude API | 按量付费 |
| 定义工具 | JSON Schema 描述可用函数 | 写配置 |
| 工具逻辑 | 真正执行操作的代码 | Python/JS 函数 |
三种路径:
- 🟢 入门(1天):直接用 API 的 Function Calling + while 循环
- 🟡 进阶(几天):用框架 LangChain / Semantic Kernel / AutoGen
- 🔵 高级(持续迭代):多 Agent 协作系统
Q5: 不会代码,还能怎么实现?
零代码也完全可以!
| 平台 | 特点 | 难度 |
|---|---|---|
| Microsoft Copilot Studio | 微软官方,集成 M365/Azure | ⭐ 最简单 |
| Dify.ai | 开源,可视化拖拽编排 | ⭐⭐ |
| Coze(扣子) | 字节跳动,中文友好 | ⭐ |
| GPTs(ChatGPT) | ChatGPT 内配置自定义 Agent | ⭐ 最简单 |
| Power Automate | 低代码自动化流程 + AI | ⭐⭐ |
以 Copilot Studio 为例,全程不写一行代码:
- 打开 Copilot Studio(浏览器)
- 新建 Agent → 自然语言描述角色
- 添加知识源(文档/SharePoint)
- 添加动作(发邮件/创建工单)
- 测试 → 发布
第三部分:关键要点总结
Agentic AI 的本质: LLM(大脑) + 工具(手脚) + 循环(行为模式)
┌─────────────────────────────────────────────────┐
│ Agentic AI 核心公式 │
│ │
│ Agent = while 循环 + LLM API + 工具函数 │
│ │
│ 每一轮: │
│ 用户目标 → LLM思考 → 调用工具 → 观察结果 → 继续 │
│ │
│ 实现门槛: │
│ 会代码 → 几十行 Python 即可 │
│ 不会代码 → Copilot Studio / Dify / GPTs │
└─────────────────────────────────────────────────┘
English Version
Part 1: Top 50 AI Industry Terms
The following terms are arranged from foundational to advanced, covering the 50 most essential concepts in the AI industry.
🔵 Foundational Concepts (1-10)
| # | Term | Definition |
|---|---|---|
| 1 | Artificial Intelligence (AI) | The broad field of creating machines that simulate human intelligence — learning, reasoning, and decision-making. |
| 2 | Machine Learning (ML) | A subset of AI where machines learn patterns from data automatically, rather than following hand-coded rules. |
| 3 | Deep Learning (DL) | A subset of ML using multi-layered neural networks to handle complex patterns like images and speech. |
| 4 | Neural Network (NN) | A computing model inspired by the human brain, consisting of interconnected layers of nodes that process information. |
| 5 | Algorithm | A set of well-defined steps and rules for solving a specific problem — the mathematical foundation of AI models. |
| 6 | Training | The process of feeding large amounts of data to a model so it learns patterns — like a student studying with practice problems. |
| 7 | Inference | Using a trained model to make predictions on new data — like taking an exam after studying. |
| 8 | Dataset | A structured collection of data used to train and evaluate models. |
| 9 | Model | A mathematical representation of learned “knowledge” from data, capable of making predictions on new inputs. |
| 10 | Parameter | Learnable variables inside a model. More parameters = larger model (e.g., GPT-4 has trillions of parameters). |
🟢 Models & Architecture (11-20)
| # | Term | Definition |
|---|---|---|
| 11 | Large Language Model (LLM) | Massive language models trained on enormous text corpora, such as GPT-4, Claude, and Gemini. |
| 12 | Transformer | A neural network architecture proposed by Google in 2017 — the core foundation of GPT, BERT, and all modern LLMs. |
| 13 | Attention Mechanism | The core component of Transformers that allows the model to “focus” on the most relevant parts of the input. |
| 14 | Foundation Model | A large model pre-trained on massive data that can be adapted for various downstream tasks. |
| 15 | Pre-training | The first phase of training where the model learns general knowledge from large-scale unlabeled data. |
| 16 | Fine-tuning | Further training a pre-trained model on domain-specific data to adapt it for a particular task. |
| 17 | Transfer Learning | Applying knowledge learned from one task to a related task, reducing training costs significantly. |
| 18 | Embedding | Converting text, images, etc. into high-dimensional numeric vectors so computers can understand semantic similarity. |
| 19 | Tokenization | Splitting text into minimal units (tokens) that a model can process, e.g., “Hello” → [“Hel”, “lo”]. |
| 20 | Diffusion Model | A model architecture that generates images/videos through iterative denoising, e.g., Stable Diffusion, DALL-E. |
🟡 Generative AI & Interaction (21-30)
| # | Term | Definition |
|---|---|---|
| 21 | Generative AI (GenAI) | AI systems that can create new content — text, images, code, audio, and video. |
| 22 | Prompt | The input instruction given to an AI model, which generates a response based on the prompt. |
| 23 | Prompt Engineering | The art and science of designing and optimizing prompts to get better AI outputs. |
| 24 | Context Window | The maximum number of tokens a model can process at once (e.g., 128K tokens). |
| 25 | Token | The basic unit LLMs use to process text — approximately 0.75 English words or 0.5 Chinese characters. |
| 26 | Temperature | A parameter controlling output randomness: 0 = deterministic, 1 = creative. |
| 27 | Hallucination | When AI generates plausible-sounding but factually incorrect or fabricated content — a core LLM challenge. |
| 28 | Grounding | Connecting AI outputs to reliable data sources to reduce hallucinations. |
| 29 | Chain-of-Thought (CoT) | A prompting technique that makes AI reason step-by-step rather than jumping to answers, improving accuracy. |
| 30 | Few-shot / Zero-shot Learning | The model’s ability to complete new tasks with few or zero examples. |
🟠 Agents & Tools (31-40)
| # | Term | Definition |
|---|---|---|
| 31 | Agentic AI | AI systems capable of autonomously planning, deciding, using tools, and iterating to complete complex tasks. |
| 32 | AI Agent | An autonomous AI program with a perceive-think-act loop capability. |
| 33 | Function Calling | LLM outputs structured tool invocation instructions that external code executes for real-world actions. |
| 34 | Tool Use | AI Agents calling external tools (search, APIs, databases) to extend their capabilities. |
| 35 | ReAct (Reasoning + Acting) | The core Agent pattern: Think → Act → Observe → Think again, looping until task completion. |
| 36 | Multi-Agent System | An architecture where multiple specialized Agents collaborate to complete complex tasks. |
| 37 | Orchestration | The management layer coordinating multiple AI components/Agents to work together. |
| 38 | MCP (Model Context Protocol) | An open protocol by Anthropic that standardizes how AI connects with external tools and data. |
| 39 | Plugin / Skill | Extensible modules that add specific capabilities to AI (e.g., web search, database operations). |
| 40 | Human-in-the-Loop | A safety mechanism where AI requests human review and confirmation at critical decision points. |
🔴 Advanced Concepts & Applications (41-50)
| # | Term | Definition |
|---|---|---|
| 41 | RAG (Retrieval-Augmented Generation) | Retrieves relevant information from a knowledge base first, then has the LLM generate answers based on retrieved results. |
| 42 | Vector Database | A specialized database for storing and searching Embedding vectors — a core RAG component. |
| 43 | RLHF (Reinforcement Learning from Human Feedback) | A training method that uses human preference feedback to optimize model outputs for better alignment. |
| 44 | Alignment | Research ensuring AI behavior and goals are consistent with human values and intentions. |
| 45 | Responsible AI | Design principles and practices ensuring AI systems are fair, transparent, safe, and explainable. |
| 46 | AGI (Artificial General Intelligence) | Human-level general intelligence capable of any intellectual task — not yet achieved. |
| 47 | Edge AI | AI running on local devices (phones, IoT) without cloud dependency — low latency, high privacy. |
| 48 | Multimodal AI | AI models that process multiple input types simultaneously (text + images + audio + video). |
| 49 | Synthetic Data | AI-generated artificial data used to train other models, solving data scarcity and privacy issues. |
| 50 | AI Governance | Policy, process, and regulatory frameworks for managing AI development and deployment. |
Part 2: Agentic AI Deep-Dive Q&A (Session Record)
A complete Q&A session covering Agentic AI from concept to implementation to practical paths.
Q1: What is Agentic AI?
Agentic AI refers to AI systems that can autonomously perceive their environment, make plans, make decisions, and take actions to achieve goals.
Unlike traditional AI (passively responding to single prompts), Agentic AI features:
- Autonomy: Independently decompose complex tasks and execute step by step
- Planning: Create multi-step strategies to achieve goals
- Tool Use: Call external tools (search, code execution, APIs, etc.)
- Reflection & Iteration: Evaluate results, adjust strategy when failures occur
- Continuous Interaction: Interact with the environment in multi-turn loops
Analogy: Traditional AI is like an assistant who answers questions. Agentic AI is like an employee who can independently complete projects — you give the goal, it figures out how.
Q2: How is Agentic AI implemented technically?
The core is a loop architecture (Agent Loop) based on the ReAct pattern:
User Goal → Reason → Act → Observe → Reason again → ... → Complete
Key Components:
| Component | Role | Implementation |
|---|---|---|
| LLM Brain | Understanding, reasoning, decisions | GPT/Claude models |
| Tool System | Interact with the real world | Function Calling / Tool Use API |
| Memory | Maintain context | Conversation history + Vector DB |
| Planner | Decompose complex tasks | Prompt guidance / CoT / Task graphs |
| Reflection | Self-correction | Output checking, retry strategies |
Technical highlights:
- Tool Calling: LLM outputs structured function call instructions; runtime executes them and returns results
- Multi-Agent Collaboration: Specialized Agents divide work, communicate via messages
- Orchestration Frameworks: LangChain, AutoGen, CrewAI encapsulate these patterns
Q3: Is Agentic AI software written in code? Can you give a concrete example?
Yes, it’s fundamentally software written in code. The core is a loop program + LLM API calls:
def agent_loop(user_request):
messages = [{"role": "user", "content": user_request}]
while True:
response = call_llm(messages, tools=available_tools)
if response.is_final_answer:
return response.text
for tool_call in response.tool_calls:
result = execute_tool(tool_call)
messages.append({"role": "tool", "content": result})
Concrete example — when a user says “write a post”:
Round 1: LLM decides to invoke skill → starts knowledge-deep-dive
Round 2: LLM calls glob to understand file structure → returns file list
Round 3: LLM generates content → calls create to make markdown files
Round 4: LLM calls powershell → git add && commit && push
Round 5: LLM determines completion → returns final answer, loop ends
Agent = Normal code (while loop + tools) + LLM API (brain). No magic.
Q4: Can I create my own Agentic AI?
Absolutely! The barrier is lower than you think.
| What You Need | Description | Cost |
|---|---|---|
| LLM API | OpenAI / Azure OpenAI / Claude API | Pay-per-use |
| Tool Definitions | JSON Schema describing available functions | Configuration |
| Tool Logic | Code that actually performs operations | Python/JS functions |
Three paths:
- 🟢 Beginner (1 day): API Function Calling + while loop
- 🟡 Intermediate (a few days): Frameworks like LangChain / Semantic Kernel / AutoGen
- 🔵 Advanced (ongoing): Multi-Agent collaboration systems
Q5: What if I don’t know how to code?
Zero-code options are absolutely available!
| Platform | Features | Difficulty |
|---|---|---|
| Microsoft Copilot Studio | Microsoft official, integrates with M365/Azure | ⭐ Easiest |
| Dify.ai | Open source, visual drag-and-drop Agent builder | ⭐⭐ |
| Coze | ByteDance, generous free tier | ⭐ |
| GPTs (ChatGPT) | Configure custom Agents inside ChatGPT | ⭐ Easiest |
| Power Automate | Low-code automation workflows + AI | ⭐⭐ |
Example with Copilot Studio — zero code required:
- Open Copilot Studio (browser)
- Create Agent → describe role in natural language
- Add knowledge sources (documents/SharePoint)
- Add actions (send email/create ticket)
- Test → Publish
Part 3: Key Takeaways
The essence of Agentic AI: LLM (brain) + Tools (hands) + Loop (behavior pattern)
┌─────────────────────────────────────────────────────┐
│ Agentic AI Core Formula │
│ │
│ Agent = while loop + LLM API + Tool functions │
│ │
│ Each round: │
│ Goal → LLM thinks → Call tools → Observe → Continue│
│ │
│ Implementation barrier: │
│ Can code → A few dozen lines of Python │
│ Can't code → Copilot Studio / Dify / GPTs │
└─────────────────────────────────────────────────────┘