# AI JARGON IN PLAIN ENGLISH **Every confusing AI word, explained like you're a smart person who just hasn't heard it yet.** From alchemy-ai.dev — free forever, no affiliate links, no gatekeeping. Last updated: July 2026. --- Nobody is born knowing this stuff. Every developer you've ever seen looked all of these up once. Terms are grouped by topic, not alphabetically, so you can read any section top to bottom and it'll build on itself. --- ## PART 1 — THE BASICS **AI (Artificial Intelligence)** — Software that can do things we used to think required a human brain: writing, answering questions, coding, recognizing images. When people say "AI" today, they almost always mean the chat kind — ChatGPT, Claude, Gemini. **LLM (Large Language Model)** — The actual technology behind chat AIs. A "language model" is a program trained on enormous amounts of text until it gets extremely good at predicting what words should come next. That sounds simple, but at massive scale it produces something that can write essays, debug code, and explain physics. ChatGPT, Claude, and Gemini are all LLMs wearing different outfits. **Model** — One specific version of an AI brain. Companies ship many: a big smart slow one, a small fast cheap one, and generations over time (like car model years). "GPT-5.5," "Claude Sonnet 5," and "Gemini 3.1 Pro" are models. The app you chat in (ChatGPT, claude.ai) is just the storefront; the model is what's behind the counter. **Prompt** — Whatever you type to the AI. That's it. "Prompting" = asking. A "good prompt" just means a clear, specific ask with enough background. There is no secret magic prompt language, whatever YouTube says. **Training / Training data** — How a model learns: it's shown a giant slice of text (books, websites, code) and learns patterns from it. This happens *before* you ever talk to it — the model is not learning from your chats in real time (though some companies may use chats to train *future* models — see Part 5). **Knowledge cutoff** — The date the model's training data ends. Ask about anything after that date and the model genuinely doesn't know — unless it has web search bolted on. This is why AIs recommend outdated tools and old prices with total confidence. **Hallucination** — When an AI states something false as if it were fact — a fake citation, a made-up function, a wrong date — with a completely straight face. It's not lying (lying requires knowing better); it's confidently filling a gap. **The single most important habit in this hobby: verify anything that matters.** **Inference** — Fancy word for "the model running." Training is school; inference is the job. When you're charged for AI, you're paying for inference. --- ## PART 2 — WORDS ON THE PRICING PAGE **Token** — How AI measures text: chunks of about ¾ of a word ("understanding" might be 3 tokens). Every limit and price you'll ever see is in tokens. 1,000 tokens ≈ 750 words ≈ a page and a half. **Context window** — How much the AI can hold in its head *at once*: your conversation, your documents, its own replies. Measured in tokens (200,000 tokens ≈ a 500-page book). When a chat exceeds the window, the oldest stuff falls out the back — which is why an AI "forgets" what you said an hour ago. Bigger window = longer memory *within a conversation*. **Rate limit / Usage limit** — The cap on how much you can use before the service says "take a break." Free tiers: strict. Paid tiers: looser but real. When people complain about "hitting limits," it's this. **API (Application Programming Interface)** — A way for *software* to talk to the AI directly, instead of you typing into a chat window. When you build an app with AI inside it, or run tools like Cline/Aider, they use an API. Paid per token, like a utility meter. **API key** — The password for your API account. Anyone who has it can spend your money. Never share it, never paste it into random tools, never put it in code you publish. If a tool asks for one, set a spending limit in your provider's dashboard the same day. **Open source / Open weights** — The model's files are published for anyone to download, inspect, and run on their own machine (Llama, DeepSeek). Opposite: **closed/proprietary** — you can only use it through the company's service (GPT, Claude, Gemini). **Local AI** — Running a model on your own computer instead of over the internet. Total privacy, no subscription, no limits — but you need a beefy machine and the models you can fit are less capable than the cloud giants. Tools that make it easy: **Ollama** and **LM Studio**. **GPU** — The graphics card. AI's raw material — models run on GPUs, which is why local AI needs a gaming-class computer and why NVIDIA became one of the richest companies on earth. --- ## PART 3 — WORDS FROM THE BUILDER WORKFLOW **Agent / Agentic AI** — An AI that doesn't just answer, it *does*: reads files, edits code, runs commands, tries again when something fails. Claude Code, Cline, and Aider are agents. Chat answers; agents act. **Claude Code** — Anthropic's agent. Point it at a project folder (in VS Code or a terminal) and it builds — reading your files, writing code across many of them, running tests. The backbone of this site's workflow. **CLAUDE.md** — A plain text file describing your project — what you're building, for whom, with what tech — that Claude Code reads before working, like a contractor reading blueprints. The template on this site is exactly this. (The ".md" is Markdown — see below.) **Markdown (.md)** — A dead-simple way to format text: `# heading`, `**bold**`, `- bullet`. Readable by humans as-is, loved by AIs. This very guide is Markdown. You can open .md files in any text editor, including Notepad. **IDE / Code editor** — The program you write code in. **VS Code** is the free one most of the world uses, and the one this site's workflow assumes. (IDE = Integrated Development Environment, but nobody expands it.) **Terminal / Command line** — The text-only window where you type commands to your computer (`git push`, `firebase deploy`). Looks intimidating, is actually just typing. Every builder gets comfortable with about ten commands and Googles the rest forever. **Repository (repo)** — A project folder tracked by **Git**, software that snapshots your files so you can undo anything and see history. **GitHub** is the website where repos live online. **Commit** = save a snapshot. **Push** = upload snapshots to GitHub. **Deploy** — Taking the app on your computer and putting it on the internet where real people can use it. Services like **Firebase Hosting** and **Vercel** do the hosting; "deploying" is the act of shipping to them. **Frontend / Backend** — Frontend: the part users see (pages, buttons — HTML/CSS/JavaScript). Backend: the machinery behind it (databases, logins, payments). Many simple sites — including this one — are almost all frontend. **Framework** — Pre-built scaffolding for apps (React, Next.js). Powerful, and also a whole thing to learn. This site's builds skip frameworks on purpose: plain HTML/CSS/JavaScript is simpler, faster to learn, and enough. **Vibe coding** — Building software by describing what you want to an AI and iterating on what it produces, without deeply reading the code. The term is half-affectionate, half-warning: great for prototypes; risky for anything handling money or private data without a review. **System prompt** — Standing instructions a model gets before your conversation starts ("you are a helpful assistant, refuse X, format Y"). You mostly can't see it, but it shapes every answer. A CLAUDE.md works like a system prompt you write yourself. **MCP (Model Context Protocol)** — A standard plug system that lets AI tools connect to outside things — your files, databases, browsers, other services. Like USB for AI: one connector standard instead of custom cables for everything. (There's a full MCP page on the site.) **RAG (Retrieval-Augmented Generation)** — The technique behind "chat with your documents": instead of hoping the model memorized something, the system *looks it up* in your files first and hands the relevant bits to the model to answer from. When a tool says "connect your knowledge base," it's RAG. **Fine-tuning** — Taking a trained model and giving it extra training on specialized data (medical notes, legal contracts, your company's style). How the thousands of specialized Llama variants get made. You will probably never need to do this — using a good general model well beats fine-tuning for almost everyone. --- ## PART 4 — WORDS FROM THE NEWS **AGI (Artificial General Intelligence)** — The hypothetical AI as capable as a human at everything. Doesn't exist. When companies say it's close, remember they're raising money; when critics say it's impossible, remember people said that about everything AI has already done. Nobody knows. Anyone who claims certainty is selling something. **Benchmark** — A standardized test for AI models (you'll see names like SWE-bench for coding). Useful for rough rankings; ignore small differences — a model that scores 2 points higher isn't meaningfully "smarter," and companies teach to the test. Your real work is the only benchmark that matters. **Parameters** — The internal dials a model tunes during training, in the billions. Roughly correlates with capability, but efficiency improved so much that small modern models beat huge old ones. Mostly a spec-sheet flex. **Reasoning model / "Thinking" mode** — Models that work through problems step by step before answering — slower and more expensive, but better at math, logic, and tricky code. Most 2026 flagships have a fast mode and a thinking mode; some tools show the "thinking" as it happens. **Multimodal** — Handles more than text: sees images, hears audio, sometimes generates pictures or video. All current flagships are multimodal to some degree. **Compute** — Shorthand for raw processing power (all those GPUs). "The compute race" = companies building warehouse-sized AI data centers. Explains both the breathtaking spending numbers in the news and why electricity keeps coming up in AI stories. **Alignment / AI safety** — The field working on making AI actually do what humans intend — from "don't make up court cases" today to "don't be dangerous at scale" tomorrow. Why some AIs refuse certain requests. Anthropic (Claude's maker) was founded specifically with this focus. **Jailbreak** — Tricking an AI into ignoring its safety rules. Mentioned here so you know the word — sites promising "jailbroken" AI are a great way to get scammed, and API terms-of-service bans are real. **Slop** — The internet's word for low-effort AI-generated content flooding feeds and search results. Worth knowing because it names the thing this site teaches you NOT to make: use AI to build real things well, not to spray filler. --- ## PART 5 — WORDS ON THE PRIVACY SETTINGS PAGE **Training opt-out** — A setting in most AI chat apps: "don't use my conversations to train future models." If that idea bothers you, go flip it — takes two minutes. (ChatGPT: Settings → Data Controls. Claude: Settings → Privacy. Gemini: Activity controls.) **Data retention** — How long the company keeps your chats on their servers, even after you delete them from your view. Buried in privacy policies; varies by company and by whether a court order is involved. **E2E vs. cloud** — Your AI chats are generally **not** end-to-end encrypted like Signal messages — the company can technically access them. Practical rule from the guides on this site: never paste passwords, API keys, financial details, or other people's private information into any cloud AI. For truly sensitive work, that's what local AI is for. **PII (Personally Identifiable Information)** — Names, addresses, SSNs, account numbers — anything that identifies a person. The stuff you should scrub before pasting a document into an AI, especially other people's. --- ## THE REAL SECRET You now know more AI vocabulary than 95% of the people arguing about AI online. None of it was hard — it was just unexplained. Bookmark this. When some article throws a term at you like you should already know it, look it up here, roll your eyes, and keep building. --- *From alchemy-ai.dev — a free AI education platform by Stacked Alchemist LLC. No course to sell. No paywall. No gatekeeping.*