AI & Technology Glossary 📖

A2A (Agent-to-Agent) is an open protocol driven by Google that lets AI agents from different companies discover, communicate, and collaborate regardless of who built them. If MCP is the 'USB' connecting an agent to tools, A2A is the 'common language' that lets two agents talk to each other. Together, MCP and A2A form the infrastructure for an interoperable multi-agent ecosystem.

AEO is the evolution of SEO for the AI era. While SEO aims to rank on Google, AEO aims to get ChatGPT, Claude, Perplexity, and AI-powered search engines to cite you directly when someone asks a relevant question. It requires clear, structured content with precise data and topic authority.

Unlike a chatbot that only answers questions, an agent acts. It works in a loop: observe the environment → decide what to do → execute an action using a tool (search the web, run code, send emails, read files) → observe the result → repeat until the goal is complete. The LLM is its 'brain' and the tools are its 'hands'. Claude Code, Manus, and Devin are examples of ready-to-use agents. There are also multi-agents (several specialized agents working as a team) and sub-agents (an orchestrator agent that delegates subtasks to other agents and consolidates results).

Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and others who left OpenAI concerned about AI safety. They created Claude, considered one of the safest, most honest, and most capable models on the market. Their focus on 'constitutional AI' (teaching values to the model) sets them apart. Amazon has invested billions in Anthropic.

API stands for 'Application Programming Interface'. It's basically a contract between two systems: 'if you send me this request in this format, I'll respond with this data'. You don't need to know how the other system works — just how to talk to it. Almost everything on the internet runs on APIs: the weather on your phone, card payments, maps in delivery apps. In AI, it's how developers connect their apps to models like GPT or Claude.

Like university rankings, benchmarks are tests that let you compare models objectively (or at least try to). MMLU, HumanEval, SWE-bench, and GPQA are some of the most popular. Each company showcases their models with their best scores, which sometimes makes apples-to-apples comparisons tricky.

CLI (Command Line Interface) is any program you use from a terminal by typing commands — no buttons, no windows. Most developer tools are CLIs: Git, Docker, npm, the AWS CLI, Vercel CLI, etc. They're faster to use than a visual interface once you know them, and they can be easily chained and automated in scripts. The opposite is a GUI (Graphical User Interface) — what most people use day to day.

An embedding converts text into a vector of numbers (a list of coordinates) that captures semantic meaning. Two texts that are similar in meaning will have close vectors, even if they use different words. It's the foundation of semantic search, recommendations, and RAG.

Fine-tuning is like hiring a generalist and paying for an intensive course in your industry. You take a base model and train it with specific examples from your use case (company emails, legal documents, medical transcriptions) so it performs better in that context.

GEO and AEO are used almost interchangeably. The idea is the same: adapt your content strategy to be relevant in a world where people ask AIs instead of searching Google. This includes using natural language, answering direct questions, and structuring content well with verifiable data.

The GPU was created for video games — processing millions of pixels in parallel. It turns out that training AI models requires exactly the same kind of massive parallel math. That's why NVIDIA dominates the AI market: their GPUs (H100, A100) became the 'gold' of the AI era. Without enough GPUs, you can't train large models.

Hallucinations are the Achilles' heel of LLMs. The model doesn't 'know' when it doesn't know something — it simply generates the most probable text, which sometimes is wrong but sounds convincing. Made-up citations, nonexistent statistics, wrong dates... all presented with the confidence of an expert.

Training a model can take months and cost millions. Inference is the moment that model answers your question — which takes seconds. When you pay for tokens via an API, you're paying for inference, not training. Companies invest heavily in making inference faster and cheaper.

AI is not a robot with feelings, and it's not going to take over the world (yet). It's software trained on enormous amounts of data to do tasks that only humans could do before: read, write, analyze, translate, generate images. There's narrow AI (does one thing well) and general AI (doesn't truly exist yet).

Linux is an operating system (like Windows or macOS) created in 1991 by Linus Torvalds and maintained by thousands of volunteers. It's free, open source, and brutally efficient. It's not very popular on personal desktops, but it dominates servers, smartphones (Android is Linux), and supercomputers. If you use the internet, you depend on Linux whether you know it or not.

LLM stands for 'Large Language Model'. It's the engine behind ChatGPT, Claude, Gemini, and the like. It's trained by reading a huge fraction of the internet, books, and texts, learning to predict what word comes next. Simple in theory, overwhelming in practice.

ReAct (Reason + Act) is the fundamental pattern of AI agents: the model reasons about what to do (Reason), executes an action like searching the web or reading a file (Act), observes the result, and repeats the cycle until the goal is complete. It's literally how an agent 'thinks'. Popular variants include Chain-of-Thought (reasoning step by step in a straight line), Tree-of-Thought (exploring multiple reasoning paths like branches of a tree), and Reflexion (the agent evaluates its own mistakes to improve).

MCP (Model Context Protocol) is like a universal USB for AI — a standard protocol that lets any AI model connect to any tool or data source in a uniform way. Before MCP, every integration was custom work. With MCP, if you build an 'MCP server' for your database or CRM, any compatible model can use it automatically. It was created by Anthropic and quickly adopted by OpenAI, Google, and most of the ecosystem.

A model is the result of training an AI with specific data and parameters. Two models from the same company can be very different in intelligence, speed, and price. Choosing the right model for your task is the art of not spending more than you need to.

A multi-agent system is like a company where each employee has a specific role, but instead of people they're AI agents. One can research, another write, another review quality, another publish — working in parallel or sequence. It's more powerful than a single agent because each one specializes and the final result is better than what any one could achieve alone. An orchestrator agent usually coordinates the team, delegating subtasks and consolidating results.

The first LLMs only understood text. Multimodal models can also see images, listen to audio, and even analyze video. GPT-4o, Gemini 2.5 Pro, and Claude 3.5 Sonnet are multimodal. This makes them far more versatile for real-world tasks.

In AI, open source means the model's weights (what it learned during training) are public and downloadable. You can run them on your own server, modify them, build on top of them. Meta's Llama and Mistral are the best-known examples. The counterpart is proprietary models like GPT or Claude, which can only be accessed via API.

OpenAI was founded in 2015 as a nonprofit by Sam Altman, Elon Musk, and others. It later became a capped-profit company. It created the GPT models, the ChatGPT assistant, and DALL-E for images. In November 2022, it launched ChatGPT and forever changed the public perception of AI. Today it's the most influential company in the AI ecosystem, with Microsoft as its main investor.

OpenClaw is the system that powers this assistant. It's a personal AI agent installed on your own server that connects to your messaging channels (WhatsApp, Telegram, Discord), has persistent memory, can execute automated tasks, manage your email, monitor systems, and much more. Think of it as an assistant that never sleeps and never forgets — unless you configure it to.

In a multi-agent system, the orchestrator is the agent with the complete view of the task. It receives the goal, breaks it down into subtasks, decides which sub-agent handles each one, sends them instructions, monitors their progress, and consolidates the final results. It doesn't do the heavy lifting directly — its value lies in intelligent coordination. It's the key piece of the multi-agent pattern.

Perplexity AI is a search engine that instead of giving you a list of links, gives you a direct AI-generated answer with cited sources alongside it. It combines real-time web search with language models. It's especially useful for questions that require synthesizing information from multiple sources. Many use it as a Google replacement for quick research.

The prompt is simply what you tell the model. It can be a short question ('what is a token?') or a long block of text with detailed instructions, context, examples, and constraints. The quality of the prompt enormously determines the quality of the response — hence the profession of 'prompt engineer' was born.

RAG (Retrieval Augmented Generation) is the solution to one of the biggest problems with LLMs: they only know what they learned during training. With RAG, before responding, the model searches a custom document base (your manual, your knowledge base, your policies) and uses that information to give precise, up-to-date answers.

RAM is where the computer stores everything it's currently using: open apps, loaded files, program states. It's different from the hard drive (permanent storage): RAM is fast but temporary — when you shut down, it's wiped. In AI, RAM (and especially GPU VRAM) determines how large the models you can run locally are.

A server is simply a computer configured to receive and respond to requests from other devices. It can be physical (a machine in a datacenter) or virtual (a portion of a larger machine). Web servers serve pages, database servers store information, mail servers send emails. When you open cual.ai, a server somewhere in the world responds to your request in milliseconds.

When you use ChatGPT, Claude, or any AI app, there are instructions you can't see that tell the model how to behave: its personality, what it can and can't do, what language to respond in, etc. Companies use this to create specialized versions of a generic model.

In the context of AI agents like OpenClaw, a skill is a packaged set of instructions and tools that teaches the agent how to perform a specific task: check the weather, manage emails, search GitHub, transcribe audio, etc. It's like installing a new ability in the agent. Skills can be shared, updated, and combined — turning a generic agent into a specialist for your workflow.

In a multi-agent system, sub-agents are the ones doing the fieldwork. Each one specializes in something: searching for information, writing text, analyzing data, generating code, etc. They receive instructions from the orchestrator, execute their task, and return the result. They operate under supervision — they don't decide what to do, but how to do well what they were asked. They're like the musicians in an orchestra: each one masters their instrument.

Temperature ranges from 0 to 2 (depending on the model). Temperature 0 = very deterministic, consistent responses — always the same answer to the same question. High temperature = more variety, creativity, and also more risk of saying weird things. For technical or precision tasks, low temperature. For creativity, high temperature.

The terminal (also called console, shell, or command line) is the way to talk directly to the operating system in plain text. It looks intimidating, but it's simply a different language. Developers use it because it's faster and more powerful than clicking through menus. On Linux and Mac it's called Terminal; on Windows, PowerShell or CMD.

Models don't read whole words but fragments called tokens. A short word can be 1 token, a long one can be 2 or 3. Spaces and punctuation count too. That's why API prices are measured in 'per million tokens'. This definition is roughly 50 tokens.

Without tools, an AI model can only generate text. With tools, it can act: search Google, execute code, read PDFs, send emails, query databases, make API calls, create files, post on social media — basically anything that can be done through software. Tools are what turn a chatbot into an agent. The MCP protocol standardizes how these tools connect to models, and each tool defines what parameters it needs and what it returns.

The context window is the model's working memory — everything it can read and consider at any given time. If your conversation exceeds that limit, the model literally forgets what you said at the beginning. Modern models have windows of 128K, 200K, or even 1 million tokens, equivalent to several complete books.

A VPN (Virtual Private Network) encrypts all your internet traffic and routes it through a server elsewhere. This does two things: protects your data on public WiFi networks (coffee shops, airports), and hides your real location (you can 'appear' in another country). People use them for privacy, corporate security, and yes — to watch Netflix from other countries.

A VPS (Virtual Private Server) is a virtual fraction of a large physical server. You rent a portion of a company's machines (like DigitalOcean, Hetzner, or AWS) and get full control: install whatever you want, configure as you need, access 24/7 via terminal. Ideal for running apps, bots, databases, or websites without paying for your own hardware.

Vibe coding is a way of building software where you describe your idea in natural language and an AI agent writes, tests, and fixes the code for you. You don't need to know how to code — you just need to know what you want to build. The term was coined by Andrej Karpathy (ex-Tesla, ex-OpenAI) in 2025 and quickly became mainstream. Tools like Cursor, Claude Code, Bolt, Lovable, and Replit Agent are examples of vibe coding platforms. The key is not the code itself but the intent: the developer describes the 'vibe' (the essence of what they want) and the AI executes it. This doesn't replace good engineers — it amplifies them.