You keep hearing these AI terms. Here's what they actually mean.
A massive AI model trained on huge amounts of text. It learns patterns in language so it can generate human-like text, answer questions, write code, and more. Well-known LLMs include Claude (Anthropic), Gemini (Google), GPT (OpenAI), and Llama (Meta).
A compact language model designed to be fast, efficient, and run on smaller hardware, even your phone. SLMs trade some general knowledge for speed and privacy, ideal when you need a model to work locally or on a specific task.
The architecture created by OpenAI that kicked off the generative AI revolution. "Generative" means it creates new text, "Pre-trained" means it learned from huge datasets beforehand, and "Transformer" is the neural network design that makes it work. The Transformer architecture is also used by Gemini, Claude, and Llama.
The basic unit of text that LLMs work with. Usually a word or word-piece. "explaining" becomes "explain" + "ing". There are two types: prompt tokens (your input: the question, system instructions, and any context) and completion tokens (the model's output). Pricing, speed, and context limits are all measured in tokens.
An LLM's ability to think step-by-step before answering, rather than just pattern-matching. Some models (GPT-5.4 Thinking, Claude extended thinking) have reasoning built in. But agent frameworks can also add reasoning to any model through structured chain-of-thought prompting, no special model needed.
A model that understands and generates more than just text: images, audio, video, and code too. Instead of being limited to one "mode," it handles several at once, like a person who can read, listen, and see.
A model that understands and responds in multiple human languages. Modern LLMs like Claude, Gemini, and GPT are trained on text from dozens of languages, so they can translate, answer, and converse without needing separate models for each language.
A system built around an LLM that can plan and take action, not just answer questions. The LLM is the brain; the agent is the whole package: it calls tools, browses the web, writes files, and loops until a task is done. Autonomy is a spectrum: some agents run end-to-end on their own, others pause at key moments for human approval.
An open standard (created by Anthropic) that gives LLMs and agents a universal way to connect to tools, apps, and data sources. Think USB-C for AI: one protocol that works everywhere. MCP servers expose tools for the model to call, resources for it to read, and prompts for it to use.
An open protocol (created by Google, now under the Linux Foundation) that lets AI agents communicate with each other, even across different companies. Where MCP connects an agent to tools, A2A connects agents to other agents.
A packaged, reusable set of instructions that teaches an agent how to do something specific. Instead of re-explaining a task every time, you give the agent a skill file and it knows how to perform it whenever needed.
A pre-built extension that adds capabilities to an LLM application. Anthropic's Claude uses MCP-based integrations (connecting to GitHub, Slack, databases, etc.) while OpenAI offers custom GPTs with "Actions" and a plugin marketplace. The industry is converging on open protocols like MCP rather than vendor-locked plugin systems.
A design pattern where a human reviews, approves, or corrects an agent's actions before they take effect. Instead of letting an agent act fully autonomously, HITL adds checkpoints, critical for high-stakes decisions like deployments, financial transactions, or sending emails on your behalf.
A technique where an LLM retrieves relevant documents from a knowledge base before generating a response. Instead of relying solely on what it memorized during training, it fetches real, up-to-date information first, dramatically reducing hallucinations.
Specialized models that convert text, images, or other data into vectors (lists of numbers that capture meaning). They power semantic search and RAG by making it possible to compare content by meaning rather than keywords. Modern embedding models can be multilingual and multimodal.
semantic_text field, turning your documents into searchable vectors automatically.A list of numbers that represents the meaning of text, images, or other data. Similar meanings produce similar vectors. Embedding models convert content into vectors so computers can compare meaning mathematically.
Search by meaning, not keywords. Instead of matching the exact words you typed, it understands your intent and finds content that means the same thing, even if the wording is completely different. Powered by search engines like Elasticsearch.
semantic_text field type. It automatically generates vectors and handles chunking. Searching "how to fix a slow computer" then finds "Speed Up Your PC". No keywords match, but the meaning is close. This is the retrieval step in RAG.