Original | Odaily Planet Daily (@OdailyChina)

Author | Golem (@web 3_golem)

Now, if people in the crypto circle do not pay attention to AI, they are easily ridiculed (Yes, my friend, think about why you clicked in here).

Do you have no understanding of the basic concepts of AI, asking what every acronym means? Are you also confused by various jargon at offline AI events, pretending not to be lost?

While it is unrealistic to jump into the AI industry in a short period, knowing the high-frequency basic vocabulary of the AI industry is certainly worthwhile. Fortunately, the following article is prepared for you ↓ I sincerely recommend you to read and save it.

Basic Vocabulary (12)

LLM (Large Language Model)

LLM is a deep learning model trained on massive data that excels in understanding and generating language; it can handle text and is increasingly able to process other types of content.

In contrast, there is SLM (Small Language Model) — which usually emphasizes lower cost, lighter deployment, and easier localization.

AI Agent (AI Intelligent Entity)

An AI Agent refers not only to "chatting models" but to systems capable of understanding objectives, utilizing tools, executing tasks step by step, and when necessary, planning and validating. Google defines an agent as software that can reason based on multimodal input and perform actions on behalf of users.

Multimodal

Its AI model does not only read text but can simultaneously process various input and output formats such as text, images, audio, and video. Google explicitly defines multimodal as the ability to handle and generate different types of content.

Prompt

The instructions given to the model by the user, which is the most basic form of human-computer interaction.

Generative AI (Generative AI / AIGC)

Emphasizes AI "generation" rather than mere classification or prediction; generative models can generate text, code, images, memes, videos, and other content based on prompts.

Token

This is one of the concepts in the AI circle that resembles "Gas units." Models do not understand content by "word count" but process inputs and outputs by tokens, with billing, context length, and response speed usually strongly related to tokens.

Context Window (Context Window / Context Length)

Refers to the total amount of tokens the model can "see" and utilize at once, also called the number of tokens the model can consider or "remember" during a single processing session.

Memory

Allows the model or Agent to retain user preferences, task contexts, and historical states.

Training

The process by which models learn parameters from data.

Inference

In contrast to training, inference refers to the process where the model receives inputs and generates outputs after going live. The industry often says "training is expensive, inference costs even more," because many costs occur during the inference phase in real commercialization. This distinction between training and inference is also the fundamental framework discussed by mainstream vendors when discussing deployment costs.

Tool Use / Tool Calling

Means that the model not only outputs text but can call external tools such as search, code execution, databases, external APIs, etc., which has been considered one of the key capabilities of an Agent.

API

The infrastructure used when AI products, applications, or Agents connect with third-party services.

Advanced Vocabulary (18)

Transformer

A model architecture that makes AI better at understanding contextual relationships, and is the technical foundation of most large language models today, characterized by its ability to simultaneously observe the relationships between every word and other words in a passage.

Attention

This is the most critical mechanism of the Transformer, designed to allow the model to automatically determine "which words are the most worth focusing on" while reading a sentence.

Agentic / Agentic Workflow

This is a popular term recently, meaning a system that is no longer just "question and answer," but can autonomously break down tasks, decide on the next steps, and call external capabilities. Many vendors consider it a sign of "moving from chatbot to executable system."

Subagents

An Agent that splits into multiple dedicated small Agents to handle sub-tasks.

Skills

With the rise of OpenClaw, this term has clearly become more common, referring to installable, reusable, and combinable capability units/user manuals for AI Agents, while also particularly reminding of the risks of tool misuse and data exposure.

Hallucination

Refers to the model making statements that are nonsensical but presented seriously, generating erroneous or absurd outputs by "perceiving nonexistent patterns," which is an overconfident output that seems reasonable but is actually false.

Latency

The time it takes for the model to produce results from receiving a request, one of the most common engineering jargon, frequently appearing when discussing deployment and productization.

Guardrails

Used to limit what the model/Agent can do, when to stop, and what content cannot be output.

Vibe Coding

This term has also become one of the hottest AI buzzwords today, meaning that users express their needs directly through conversation, while AI writes code, without users needing to understand how to write code specifically.

Parameters

The scale of numbers used internally by the model to store capabilities and knowledge, often used to crudely measure model size, with terms like "billions of parameters" and "hundreds of billions of parameters" being the most common way to impress in the AI circle.

Reasoning Model

It usually refers to models that excel in multistep reasoning, planning, validation, and executing complex tasks.

MCP (Model Context Protocol)

This has become a very popular new buzzword in the past year, serving a similar purpose to establish a common interface between models and external tools/data sources.

Fine-tuning / Tuning

Further training on the base model to better adapt it to specific tasks, styles, or domains. Google's glossary directly considers tuning and fine-tuning as related concepts.

Distillation

Compressing the capabilities of a large model to share with a smaller model, akin to teaching a "student" by a "teacher."

RAG (Retrieval-Augmented Generation)

This has almost become a basic configuration for enterprise AI. Microsoft defines it as a model that combines "search + LLM," using external data to ground answers and solve issues like outdated model training data or not being aware of private knowledge bases. The goal is to base answers on real documents and private knowledge rather than relying solely on the model's memory.

Grounding

Often mentioned alongside RAG, it means ensuring that the model's responses are based on external references such as documents, databases, or web pages rather than just relying on the parameters' memory for “creative” outputs. Microsoft clearly identifies grounding as a core value in its RAG documentation.

Embedding

This refers to encoding text, images, audio, and other content into high-dimensional numerical vectors for semantic similarity calculations.

Benchmark

A way to assess model capabilities using a set of standardized criteria, also a language that various models love to use to "prove their strength" on leaderboards.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。