Author|Golem(@web 3_golem)
Now, if people in the crypto circle don't pay attention to AI, they are easily ridiculed (yes, my friend, think about why you clicked in).
Are you completely clueless about the basic concepts of AI, asking Douban what every abbreviation in a sentence means? Or are you at AI offline events, confused by all the jargon, yet pretending you're still online?
Although it's unrealistic to jump into the AI industry overnight, it's always beneficial to know the high-frequency basic vocabulary of the AI industry. Fortunately, the following article is prepared for you↓ Sincerely recommend you to read it thoroughly and bookmark it.
Basic Vocabulary (12)
LLM (Large Language Model)
The core of an LLM is a deep learning model trained on massive amounts of data, excelling at understanding and generating language. It processes text and is increasingly capable of handling other types of content.
In contrast, there is the SLM (Small Language Model)—typically emphasizing lower cost, lighter deployment, and greater convenience for localization.
AI Agent (AI Intelligent Agent)
An AI Agent refers not just to a "chat model," but to a system that can understand goals, call tools, execute tasks step-by-step, and even perform planning and verification when necessary. Google defines an agent as software that can reason based on multimodal inputs and perform actions on behalf of the user.
Multimodal (Multimodal)
Such AI models don't just process text but can simultaneously handle multiple forms of input and output, such as text, images, audio, and video. Google explicitly defines multimodality as the ability to process and generate different types of content.
Prompt (Prompt)
The instruction input by the user to the model, which is the most basic form of human-machine interaction.
Generative AI (Generative AI / AIGC)
Emphasizes AI's ability to "generate" rather than merely classify or predict. Generative models can generate text, code, images, memes, videos, and other content based on prompts.
Token (Token)
This is one of the concepts in the AI world most similar to a "Gas unit." Models don't process content by "word count" but by tokens. Billing, context length, and response speed are often strongly related to tokens.
Context Window (Context Window / Context Length)
Refers to the total number of tokens a model can "see" and utilize at one time. It can also be described as the number of tokens the model can consider or "remember" during a single processing instance.
Memory (Memory)
Allows the model or Agent to retain user preferences, task context, and historical states.
Training (Training)
The process by which a model learns parameters from data.
Inference (Inference Execution)
Opposite of training, it refers to the process where the model receives input and generates output after deployment. The industry often says, "Training is expensive, but inference is even more costly," because many costs during the actual commercialization phase occur during inference. The distinction between training and inference is also a fundamental framework in discussions about deployment costs among mainstream vendors.
Tool Use / Tool Calling (Tool Calling)
Means the model doesn't just output text but can also call tools such as search, code execution, databases, and external APIs. This is already considered one of the key capabilities of an Agent.
API (Interface)
The infrastructure used when AI products, applications, or Agents connect to third-party services.
Advanced Vocabulary (18)
Transformer (Transformer Architecture)
A model architecture that makes AI better at understanding contextual relationships. It is the technical foundation for most large language models today. Its biggest feature is the ability to simultaneously consider the relationships between every word in an entire segment of content.
Attention (Attention Mechanism)
It is the most critical core mechanism of the Transformer. Its role is to allow the model to automatically determine "which words are most worth focusing on" when reading a sentence.
Agentic / Agentic Workflow (Agentic / Agentic Workflow)
This is a very hot term recently. It means a system is no longer just "question-and-answer" but has a certain degree of autonomy to break down tasks, decide the next steps, and call external capabilities. Many vendors see it as a sign of "moving from Chatbot to an executable system."
Subagents (Sub-agents)
An Agent further breaks down into multiple specialized smaller Agents to handle sub-tasks.
Skills (Reusable Capability Modules)
With the explosion of OpenClaw, this term has become noticeably more common. These are installable, reusable, and composable capability units/operating instructions for AI Agents, but there are also specific warnings about tool misuse and data exposure risks.
Hallucination (Machine Hallucination)
Refers to the model confidently spouting nonsense, "perceiving patterns that do not exist," thereby generating incorrect or absurd outputs. This is the model's seemingly reasonable but actually erroneous overconfident output.
Latency (Latency)
The time it takes for the model to output results after receiving a request. It is one of the most common engineering jargon terms and frequently appears in discussions about implementation and productization.
Guardrails (Guardrails)
Used to restrict what the model/Agent can do, when to stop, and what content cannot be output.
Vibe Coding (Vibe Coding)
This term is also one of the hottest AI jargons today. It means users directly express requirements through conversation, and the AI writes the code, while the user doesn't need to specifically understand how to code.
Parameters (Parameters)
The internal numerical scale of the model used to store capabilities and knowledge. Often used as a rough measure of the model's size. "Billions of parameters" and "trillions of parameters" are among the most common intimidating phrases in the AI circle.
Reasoning Model (Strong Reasoning Model)
It typically refers to models that are better at multi-step reasoning, planning, verification, and executing complex tasks.
MCP (Model Context Protocol)
This is a very hot new jargon in the past year. Its role is similar to establishing a universal interface between the model and external tools/data sources.
Fine-tuning / Tuning (Fine-tuning)
Continuing training on a base model to make it more adapted to specific tasks, styles, or domains. Google's glossary directly lists tuning and fine-tuning as related concepts.
Distillation (Distillation)
Compressing the capabilities of a large model into a smaller model as much as possible, akin to a "teacher" teaching a "student."
RAG (Retrieval-Augmented Generation)
This has almost become a basic configuration for enterprise AI. Microsoft defines it as a "search + LLM"模式 (pattern), using external data to ground answers, solving problems like outdated training data and lack of knowledge about private knowledge bases. The goal is to base answers on real documents and private knowledge, rather than relying solely on the model's own memory.
Grounding (Fact Alignment)
Often appears together with RAG. It means making the model's answers based on external sources like documents, databases, web pages, etc., rather than relying solely on parameter memory for "free creation." Microsoft explicitly lists grounding as a core value in its RAG documentation.
Embedding (Vector Embedding / Semantic Vector)
Encoding content like text, images, and audio into high-dimensional numerical vectors to perform semantic similarity calculations.
Benchmark (Benchmark Test)
An evaluation method that uses a unified set of standards to test model capabilities. It is also the ranking language most loved by various companies to "prove they are strong."
Recommended Reading
Lobster's Key 11 Questions: The Most Easy-to-Understand Breakdown of OpenClaw Principles
