Glossary
The words everyone uses, in plain English
53 terms you will meet in client calls, tool docs and every Zertiko track. One or two sentences each, grouped so they make sense together. Taught in F1 · Working with AI.
The AI itself
What is actually answering when you type into Claude, ChatGPT, Grok or Gemini.
- Model (LLM) · also: large language model
- The AI brain: a large language model trained to predict the next word so well that it can write, summarize, code and reason. Claude Sonnet, GPT-6 Astra and Grok are models; the apps around them are products.
- Training
- How a model is made: it reads an enormous amount of text (and images, audio) and adjusts billions of internal numbers until its predictions are good. Training costs millions; you never do it. You use a model that has already been trained.
- Weights
- Those billions of internal numbers. “The weights” are the model, the way a recipe is the cake. Change the weights and you have a different model.
- Open weights · also: open-weight model, open source model
- A model whose weights are published for anyone to download and run (Llama, Qwen, Mistral). “Closed” models (Claude, GPT, Grok) are used through the company’s app or API only. Open weights can run on your own computer or on Cloudflare Workers AI.
- Frontier model
- The most capable models available at a given moment, made by a handful of labs. Frontier models cost more per use and are what you reach for when quality matters; smaller or open models are cheaper and fine for simple tasks.
- Inference
- Running the model to get an answer. Training happens once; inference happens every time you send a message. You pay for inference, by subscription or by token.
- Token
- The unit a model reads and writes in: roughly three-quarters of a word. Prices, limits and “context windows” are measured in tokens. 1,000 tokens is about 750 words.
- Context window
- How much the model can hold in mind at once: your prompt, the files you attached and the conversation so far. Exceed it and the model forgets the beginning. Long files are why we summarize first.
- Prompt / system prompt
- What you send the model. A system prompt is the standing instruction set (who the AI is, rules, voice) that sits above the conversation; your Claude Project instructions or a Bot’s skill are system prompts.
- Hallucination
- A confident, fluent answer that is wrong: an invented statistic, a fake citation, a product that does not exist. The reason every Zertiko track has a fact-check step.
- Fine-tuning
- Taking a trained model and training it a little more on your own examples so it adopts a style or a specialty. Rarely needed for service work; a good prompt and a few examples get you most of the way.
- Temperature
- A setting for how adventurous the model’s word choices are. Low is predictable and consistent (good for data and code); high is more varied (good for brainstorming).
- Multimodal / vision model
- A model that can read images, hear audio or watch video as well as text. Why you can paste a screenshot of a spreadsheet and ask what is wrong with it.
- Image model (diffusion) · also: diffusion model, text-to-image
- A different kind of model that makes pictures from text, by starting with noise and “denoising” toward your description. Flux, Stable Diffusion, GPT Image and Midjourney are image models.
- TTS / STT
- Text-to-speech (a synthetic voice reads text aloud: ElevenLabs) and speech-to-text (audio becomes a transcript: Whisper, Otter). Voice agents chain STT → model → TTS.
- Embeddings / vector database
- A way to turn text into a list of numbers so similar meanings sit near each other; a vector database stores those numbers so you can find “the paragraph most like this question”. The machinery under RAG.
- RAG · also: retrieval-augmented generation
- Retrieval-Augmented Generation: before answering, the system looks up relevant documents (your FAQ, your policies, a client’s files) and hands them to the model with the question. It is how a chatbot answers from your knowledge base instead of guessing. “Knowledge base” features in Chatbase, Retell or Grok Bot are RAG.
- GPU
- The graphics chip that does the math for training and inference. “Needs a GPU” means a model or render will crawl on an ordinary laptop. Cloud GPUs are rented by the hour.
Things that act
The difference between a chat, a bot, an agent and a swarm is how much it does on its own.
- Chatbot
- A program that answers in conversation. Can be a simple FAQ bot or a full model behind a chat window. It talks; it does not usually act.
- Bot
- Any program that runs tasks automatically, usually inside another service: a Slack bot, a Discord bot, Grok Bot. Bots can be dumb (fixed rules) or smart (a model inside).
- Agent
- A model given tools and a goal, that decides its own steps: read the inbox, draft replies, check the calendar, ask you when unsure. Claude Code, Codex, Grok Bot and the ranking-watch agent in A4 are agents. The word “agentic” just means “works like an agent”.
- Agentic flow / workflow
- A multi-step job an agent runs: research → draft → design → approve → schedule. Some steps are fixed (a workflow), some are decided by the model (agentic). Routines in Grok Bot and Cron Triggers on Workers are how flows run on a schedule.
- Tool / function calling
- How a model does things beyond text: it asks to call a named function (check_availability, send_email, run_sql) with arguments, your code runs it, and the result goes back to the model. Every agent is built on this.
- Computer use
- An agent that sees the screen and controls mouse and keyboard, so it can operate apps that have no API (Xcode, Blender, a legacy portal). Slower than tools; essential for the long tail.
- Swarm / multi-agent · also: multi-agent system, squad
- Several agents working on one job, each with a role (researcher, writer, reviewer), passing work between them. Powerful for big tasks, expensive and harder to control; most service work needs one good agent with a checklist.
- Routine / cron / scheduled task
- A job that runs on a schedule without you: every Thursday at 9, the first of the month at 6. Cron is the old Unix way of writing schedules (0 6 * * 1 = Mondays 06:00); Grok Bot calls them routines; Claude calls them scheduled tasks.
- Skill
- A reusable instruction file an agent loads by name (“load the client voice skill”). Think of it as a saved procedure.
- Guardrails / approvals
- Rules that limit what an agent may do alone: ask before sending, never delete, only read this folder. Grok Bot’s Auto-review and Claude Code’s permission prompts are guardrails. The reason an agent with your inbox is safe.
How software talks
The plumbing words that come up the moment you connect two tools.
- API · also: REST API, endpoint
- Application Programming Interface: a set of URLs a service exposes so programs (not people) can use it. “Does it have an API?” means “can an agent or script talk to it without clicking around?”. You usually get an API key to use it.
- API key / token / secret
- The password a program uses to call an API. Treated like cash: stored in a secrets vault or .env file, never pasted in chat, rotated if leaked.
- Webhook
- The reverse of an API call: a service calls *your* URL when something happens (a payment succeeded, a form was submitted, a call ended). Your code receives the event and reacts. “Send me a webhook” = “notify my program automatically”.
- JSON
- The plain-text format programs use to exchange data: curly braces, keys and values. When an API or a webhook sends you something, it is almost always JSON.
- MCP · also: MCP server, connector, plugin
- Model Context Protocol: a standard way to plug tools and data sources into an AI app. A “connector”, “plugin” or “MCP server” gives Claude, ChatGPT or Grok Bot a new set of tools (Gmail, Blender, a scheduler, your own database).
- SDK
- Software Development Kit: a library in a programming language that wraps an API so you can call it with normal code instead of raw web requests.
- Script
- A short program that does one job, run from the terminal: rename 400 files, pull a report, build a scene in Blender. Python, JavaScript and PowerShell are the usual script languages. Agents write scripts constantly; you just need to read them.
- CLI / terminal
- Command Line Interface: typing commands into a text window (Terminal on Mac, PowerShell on Windows) instead of clicking. Claude Code, Codex CLI, Wrangler and Git are all CLIs. Desktop Commander gives an AI a terminal.
- Headless
- Running a program without its window: Blender rendering from a script, Claude Code answering a prompt from the command line, a browser controlled by code. Agents love headless because it is scriptable.
- Tunnel
- A secure pipe from a machine behind a firewall to the public internet, so a cloud service can reach it without opening ports. Cloudflare Tunnel is how your Worker calls a program on your home computer.
Building and shipping
Words from the build-and-deploy tracks.
- Repository (repo) / Git / GitHub
- Git tracks every change to a project’s files; a repository is the project with its history; GitHub hosts repositories online. “Commit” is a saved checkpoint; “push” uploads it. Git is the undo button that makes AI coding safe.
- Deploy
- Putting a build on the internet so people can use it. “Deploy to Cloudflare” = the Wrangler command that uploads your site or Worker.
- Cloudflare Workers / serverless / edge
- Code that runs in Cloudflare’s data centers around the world, started on demand, with no server you maintain. “Serverless” means you never manage a machine; “edge” means it runs near the visitor. D1 is their database, KV their cache, R2 their file storage, Workers AI their models.
- Astro
- A web framework for building fast sites: pages are HTML with as little JavaScript as possible. Zertiko itself and the A4 platforms are Astro sites on Workers.
- WordPress
- The most-used website system in the world (about 40% of sites): themes, plugins, an editor non-technical owners can use. Easy to start, easy to bloat; A3 builds it lean.
- Xcode
- Apple’s app for building iPhone, iPad and Mac apps. The only way to compile and submit iOS apps, which is why A5 needs a Mac. SwiftUI is the language-and-toolkit you write the app in; the Simulator is the fake iPhone on your screen.
- TestFlight / App Store Connect
- TestFlight is Apple’s beta app: testers install your app before it is public. App Store Connect is the website where you manage the app, its listing and its review.
- n8n · also: Zapier, Make
- A visual automation tool: drag boxes (a trigger, a Gmail node, a Google Sheets node, an AI node), connect them, and the flow runs. Open source; you can self-host it. The no-code cousin of writing a script, and a common “automation freelancer” deliverable.
- ComfyUI
- A node-based desktop app for running image and video models: you wire boxes (load model, prompt, sampler, upscale) into a graph and run it. The professional way to drive Stable Diffusion and Flux locally, with full control; the opposite of a one-box web generator.
- LoRA / checkpoint
- In image models: a checkpoint is a full set of weights you load; a LoRA is a small add-on trained on a style or a face that steers the checkpoint. “Train a LoRA on the client’s product photos” makes consistent product images.
- Blender
- Free, professional 3D software: modeling, animation, lighting, rendering, video editing. Drives films and game cinematics; A1 has an AI operate it through scripts and the Blender MCP.
- DNS / domain / SSL
- A domain is the name (zertiko.com); DNS is the phone book that points it at a server; SSL is the certificate that makes it https. Moving DNS to Cloudflare is how a domain gets attached to a Worker.
- Core Web Vitals / PageSpeed
- Google’s speed and stability scores for a page, measured on real phones (LCP, INP, CLS). PageSpeed Insights is the free report; 90+ on mobile is the Zertiko target.
- llms.txt
- A plain-text map of a website written for AI assistants to read, like robots.txt is for crawlers. Part of “LLM optimization”: being findable in AI answers, not just Google.
- No-code / low-code
- Building software by configuring tools (Canva, Durable, n8n, Chatbase) instead of writing code (no-code), or with a little code where needed (low-code). Most Level 1 tracks are no-code; the Advanced tracks are low-code with an AI writing the code.
Want to use these for real?
Every Zertiko track teaches one AI-powered service end to end, free, with a certification at the end.
Browse the Academy