Crypto Cheat Sheet AI: Explaining 30 Common Slang Terms in One Shot

Original Title: "AI Insider Jargon Dictionary (March 2026 Edition), Recommended for Bookmarking"

Original Author: Golem, Odaily Planet Daily

Now, if you're in the crypto world and not paying attention to AI, you're easily subject to ridicule (yes, my friend, think about why you clicked in).

Are you completely clueless about the basic concepts of AI, asking the soy sauce bean for the meaning of every acronym in a sentence? Are you lost in a sea of proprietary terms at AI events, pretending you're not disconnected?

While it's not realistic to dive into the AI industry in a short amount of time, knowing a summary of high-frequency AI industry basics is worthwhile. Luckily, this article is prepared for you below. Sincerely advise you to read through and bookmark.

Basic Vocabulary (12)

· LLM (Large Language Model)

The core of LLM is a deep learning model trained on massive amounts of data, proficient in understanding and generating language. It can process text and increasingly handle other types of content.

In contrast is the SLM (Small Language Model) - usually emphasizing a language model with lower costs, lighter deployment, and more convenient localization.

· AI Agent

AI Agent refers not only to a "chatting model" but a system capable of understanding goals, invoking tools, step-by-step task execution, planning, and validation when necessary. Google defines an agent as software that can reason based on multimodal input and act on behalf of the user.

· Multimodal

Its AI model is not only text-based but can simultaneously process various input-output forms like text, images, audio, videos, etc. Google specifically defines multimodal as the ability to process and generate different types of content.

· Prompt

The user's input command to the model, the most basic form of human-machine interaction.

· Generative AI (Generative AI / AIGC)

Emphasizing AI "generation" rather than just classification or prediction, generative models can produce text, code, images, emojis, videos, etc., based on a prompt.

· Token

This is one of the concepts in the AI field most similar to the "Gas Unit." Models do not understand content based on "words," but rather process input and output based on tokens, with billing, context length, and response speed usually highly correlated to tokens.

· Context Window

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

· 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

In contrast to training, refers to the process where a model, once deployed, receives input and generates output. In the industry, it is often said that "training is expensive, but inference is even costlier" because many costs in the real commercialization phase occur during inference. The distinction between training and inference is also the foundational framework for discussions of deployment costs in mainstream vendors.

· Tool Use / Tool Calling

Means that a model not only outputs text but can also call tools such as search, code execution, databases, external APIs, etc. This has already been regarded as a key capability of agents.

· API

Infrastructure for AI products, applications, and agents when interacting with third-party services.

Advanced Vocabulary (18)

· Transformer

A model architecture that makes AI better at understanding 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 in the entire piece of content.

· Attention

The central mechanism in Transformers, its role is to enable the model to automatically determine "which words are most worthy of attention" when reading a sentence.

· Agentic / Agentic Workflow

This is a recently popular term, which means a system is no longer just "question and answer," but has a certain degree of autonomy to break down tasks, decide on the next steps, and invoke external capabilities. Many vendors see it as a sign of "moving from Chatbot to executable system."

· Subagents

An Agent further breaks down into multiple dedicated sub-agents to handle subtasks.

· Skills

With the rise of OpenClaw, this term has become more common. It refers to installable, reusable, and combinable capability units/instructions for an AI Agent, but also warns of tool misuse and data exposure risks.

· Hallucination

It refers to a model confidently generating erroneous or absurd output by "perceiving non-existent patterns," presenting a seemingly reasonable but actually incorrect overconfident output.

· Latency

The time it takes for a model to process a request and produce an output, is one of the most common engineering jargon, frequently encountered in discussions on deployment and productization.

· Guardrails

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

· Vibe Coding

This term is also one of the hottest AI slang terminologies today, meaning users express their needs directly through conversation, and AI writes the code, without the user needing to understand how to code specifically.

· Parameters

Numerical scales used internally in a model to store capabilities and knowledge, often used to roughly measure the scale of a model. Phrases like "hundreds of billions of parameters" are common bragging statements in the AI community.

· Reasoning Model

It usually refers to models that are better at multi-step reasoning, planning, validation, and complex task execution.

· MCP (Model Context Protocol)

This is a very hot new buzzword in the past year, serving as a common interface between models and external tools/data sources.

· Fine-tuning

Continuing training on a base model to make it more suitable for a specific task, style, or domain. Google's terminology directly considers tuning and fine-tuning as related concepts.

· Distillation

Transferring the capabilities of a large model to a smaller model, like having the "teacher" instruct the "student."

· RAG (Retrieval-Augmented Generation)

This has almost become a standard configuration in enterprise AI. Microsoft defines it as a "search + LLM" pattern, using external data to ground the answers, addressing issues such as outdated training data and lack of understanding of private knowledge bases. The goal is to base the answers on real documents and private knowledge rather than solely on the model's own recall.

· Grounding

Often associated with RAG, it means ensuring that the model's answers are based on external sources such as documents, databases, web pages, rather than relying only on parameter memorization. Microsoft explicitly identifies grounding as a core value in the RAG documentation.

· Embedding (Vector Embedding / Semantic Vector)

Encoding textual, image, audio, and other content into high-dimensional numerical vectors for semantic similarity calculations.

· Benchmark

An evaluation method that uses a standardized set of criteria to test a model's capabilities, often used by various models to "prove their strength" through leaderboard rankings.

Original Article Link