AI Jargon Dictionary (March 2026 Edition), Recommended to Bookmark

golem

Odaily资深作者

@web3_golem

2026-03-11 12:48

This article is about 7386 words, reading the full article takes about 11 minutes

30 new terms to leave an impression, so novices no longer have to awkwardly nod and smile during gatherings.

AI Summary

Expand

Core Viewpoint: This article aims to provide blockchain/cryptocurrency practitioners with a basic and advanced vocabulary guide for the AI field, helping them understand the core concepts and technical frameworks of the AI industry and adapt to the trend of AI-Web3 convergence.
Key Elements:
1. LLM and SLM: Large Language Models (LLMs) are deep learning models that process multimodal content, while Small Language Models (SLMs) emphasize low cost and localized deployment.
2. AI Agent: Refers to an intelligent system capable of understanding goals, calling tools, and executing tasks step-by-step, representing an executable system beyond chatbots.
3. RAG (Retrieval-Augmented Generation): Combines external data retrieval with large model generation to address issues of outdated model knowledge and private data access, serving as a foundational configuration for enterprise AI applications.
4. Inference vs. Training Costs: Industry consensus is that "training is expensive, but inference costs even more"; during the commercialization phase, costs primarily occur during the model inference execution process.
5. Token and Context Window: Tokens are the basic unit for AI model processing and billing, while the context window determines the total number of tokens a model can "remember" in a single processing session.
6. Agentic Workflow: Refers to a workflow with autonomous task decomposition and execution capabilities, seen as a key milestone in transitioning from chatbots to executable systems.
7. Tool Calling and Risks: A model's ability to call external tools (APIs, databases) is a core capability of an Agent, but it comes with risks of tool misuse and data exposure.

Original | Odaily (@OdailyChina)

Author｜Golem (@web3_golem)

Nowadays, if you're in the crypto space and not paying attention to AI, you're practically asking to be mocked (yes, my friend, think about why you clicked on this).

Are you completely clueless about the basic concepts of AI, asking Doubao what every abbreviation in a sentence means? Or are you at AI offline events, utterly confused by all the jargon, trying to pretend you're still in the loop?

While diving headfirst into the AI industry overnight isn't realistic, familiarizing yourself with high-frequency basic vocabulary is always a good investment. Lucky for you, this article is exactly what you need↓ Sincerely recommend you read it thoroughly and bookmark it.

Basic Vocabulary

LLM (Large Language Model)

At its core, an LLM is a deep learning model trained on massive datasets, excelling at understanding and generating language. It processes text and is increasingly capable of handling other content types.

Its counterpart is the SLM (Small Language Model)—typically emphasizing lower cost, lighter deployment, and easier localization.

AI Agent

An AI Agent refers not just to a "chatty model" but to a system that can understand goals, call tools, execute tasks step-by-step, and even plan and verify when necessary. Google defines an agent as software that can reason based on multimodal inputs and perform actions on behalf of the user.

Multimodal

This refers to AI models that don't just read text but can simultaneously process and generate multiple input/output forms like text, images, audio, and video. Google explicitly defines multimodality as the ability to process and generate different types of content.

Prompt

The instruction a user gives to a model, the most fundamental form of human-computer interaction.

Generative AI / AIGC

Emphasizes AI's ability to "generate" rather than just classify or predict. Generative models can create text, code, images, memes, videos, and more based on prompts.

Token

One of the concepts in the AI world most akin to a "Gas unit." Models don't understand content by "word count" but process input and output by tokens. Billing, context length, and response speed are all strongly tied to tokens.

Context Window / Context Length

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

Memory

Allows a model or Agent to retain user preferences, task context, and historical states.

Training

The process by which a model learns parameters from data.

Inference

Opposite of training, refers to the process where a deployed model receives input and generates output. The industry often says, "Training is expensive, inference costs even more," because many costs in real commercialization occur during inference. The training/inference distinction is also a fundamental framework for mainstream vendors discussing deployment costs.

Tool Use / Tool Calling

Means the model doesn't just output text but can call tools like search, code execution, databases, and external APIs. This is now considered a key capability of Agents.

API (Application Programming Interface)

The infrastructure for AI products, applications, and Agents to connect to third-party services.

Advanced Vocabulary (18)

Transformer

A model architecture that allows AI to better understand contextual relationships, serving as the technical foundation for most large language models today. Its key feature is the ability to simultaneously consider the relationship between each word and all other words in a segment of text.

Attention (Mechanism)

The most critical core mechanism of the Transformer. Its function is to let the model automatically determine "which words are most worth focusing on" when reading a sentence.

Agentic / Agentic Workflow

A very hot term recently. It refers to a system that is no longer just "Q&A" but has a degree of autonomy in decomposing tasks, deciding next steps, and calling external capabilities. Many vendors see it as a marker of moving "from Chatbot to executable system."

Subagents

An Agent further decomposed into multiple specialized smaller Agents to handle subtasks.

Skills (Reusable Capability Modules)

With the explosion of OpenClaw, this term has become noticeably more common. These are installable, reusable, composable capability units/instruction manuals for AI Agents. However, there are specific warnings about potential tool misuse and data exposure risks.

Hallucination

Refers to the model confidently generating nonsense—"perceiving patterns that don't exist"—resulting in erroneous or absurd outputs. It's the model's seemingly plausible but actually incorrect overconfident output.

Latency

The time it takes for a model to output results after receiving a request. One of the most common engineering jargon terms, frequently appearing in discussions about implementation and productization.

Guardrails

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

Vibe Coding

Another of today's hottest AI buzzwords. It means users directly express requirements through conversation, and the AI writes the code, with the user not needing to specifically understand how to code.

Parameters

The numerical scale of numbers inside a model used to store capabilities and knowledge. Often used as a crude measure of model size. "Billions of parameters" and "trillions of parameters" are among the most common intimidating phrases in the AI circle.

Reasoning Model

Typically refers to models more adept at multi-step reasoning, planning, verification, and complex task execution.

MCP (Model Context Protocol)

A very hot new buzzword in the past year. Its function is similar to establishing a universal interface between models and external tools/data sources.

Fine-tuning / Tuning

Continuing to train a base model to make it more suitable for specific tasks, styles, or domains. Google's terminology list directly lists tuning and fine-tuning as related concepts.

Distillation

Compressing the capabilities of a large model into a smaller one as much as possible, akin to having a "teacher" instruct a "student."

RAG (Retrieval-Augmented Generation)

This has almost become a basic configuration for enterprise AI. Microsoft defines it as a "search + LLM" 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, not just the model's own memory.

Grounding

Often appears alongside RAG. It means ensuring model answers are based on external sources like documents, databases, web pages, etc., rather than relying solely on parametric memory for "free improvisation." Microsoft explicitly lists grounding as a core value in its RAG documentation.

Embedding (Vector Embedding / Semantic Vector)

The process of encoding content like text, images, or audio into high-dimensional numerical vectors to facilitate semantic similarity calculations.

Benchmark

A standardized testing method using a unified set of criteria to evaluate model capabilities. Also the ranking language favorite among various model vendors to "prove their strength."