AI Isn’t Hard. You Just Need a Cheat Sheet.
Running Markup AI — a content governance platform built on top of large language models — means I spend a lot of time talking to executives about AI, as well as friends and family who aren’t involved in the tech sector for a living.
And I’ve watched the same scene play out over and over: a smart leader sits through a vendor demo, someone drops “context window” or “RAG pipeline,” and they nod along confidently while having absolutely no idea what was just said.
I’ve been that person. You probably have too.
Here’s the thing — the concepts aren’t complicated. The inventors just gave simple ideas intimidating names.
Here’s your cheat sheet.
Large Language Model (LLM) or Foundational Model or Frontier Model
The engine under every AI writing and reasoning tool you’ve used. It was trained on a massive amount of text and learned to predict what words follow other words. When you use ChatGPT, Claude, or Gemini — you’re talking to an LLM. Frontier Model refers specifically to the most powerful ones at any given moment.
Plain English: An AI that learned to reason by reading the entire internet.
Tokens
A token is roughly three-quarters of a word. AI models process everything — your question and their answer — in tokens. That’s how costs are calculated and limits are set. A “128K Context Window” means the model can handle about 128,000 tokens — roughly a 300-page book — at once. See below for a definition of Context Window.
Plain English: AI’s unit of measurement for text. There’s a limit to how much it can handle, and you’re billed for what you use.
Prompt
Just what you type into an AI. A basic prompt is a question. A sophisticated one gives the AI a role, constraints, and instructions before it starts.
“Prompt engineering” is mostly just — be specific, give context, say what you want and don’t want. It’s not about engineering the way you think about engineering.
Plain English: What you say to the AI. Better instructions = better results.
Hallucination
When an AI produces something that sounds completely plausible but is simply wrong — a made-up citation, an invented statistic, a confident assertion that’s total fiction. LLMs are trained to produce text that sounds right based on predictive modeling, not text that is right.
Verify anything consequential.
Plain English: When the AI confidently makes something up. Trust but verify.
Context Window
How much of the conversation an AI can “remember” and work with at once. A large context window means it can read a whole contract and respond intelligently. A small one means you’re working with a goldfish.
Plain English: The AI’s working memory. More is better.
RAG (Retrieval-Augmented Generation)
AI models have a knowledge cutoff — they don’t know what happened after their training ended. RAG is the workaround: you feed the AI fresh documents at the moment it needs to answer.
Instead of answering from memory, it answers from the binder you hand it.
Plain English: A way to give an AI current or proprietary information it wasn’t trained on.
Inference
The act of actually using an AI — asking a question, getting a response. Different from training, which is the earlier process of teaching the model. “Inference cost” is what you pay per interaction.
Plain English: Using the AI. Training teaches it; inference is when you run it.
Guardrails
Rules, filters, and boundaries you put around an AI to keep it from going off the rails. Think brand guidelines, compliance policies, content standards — anything that constrains what the AI can say, how it says it, or what topics it avoids. Without guardrails, AI will happily generate content that’s off-brand, non-compliant, or just plain wrong.
This is what we do at Markup AI — we’re the guardrails for enterprise content.
Plain English: The fences that keep AI from doing something stupid or dangerous.
Multi-modal
An AI that doesn’t just work with text — it can also process images, audio, video, or some combination. GPT-4o, Gemini, and Claude can all “see” images you show them, not just read text you type.
Plain English: AI that can work with more than just words — pictures, audio, video too.
Agent / AI Agent
An AI that doesn’t just answer questions — it takes actions. Browse the web, write code, send emails, coordinate with other AI systems. The difference between an AI that tells you how to do something versus one that actually does it.
Plain English: An AI that does things, not just says things.
Vibe Coding
Describing what you want in plain English and letting the AI write the code — no programming knowledge required. You’re not writing syntax; you’re describing intent. Tools like Replit, Cursor, and Lovable have made this accessible to anyone who can articulate what they want built.
The downside: if you can’t read the code the AI produces, you can’t catch its mistakes. It’s powerful, but “trust but verify” applies here too.
Plain English: Getting AI to write software for you by describing what you want instead of knowing how to code it.
You don’t need to write code. You don’t need a PhD. You need to ask the right questions — and that starts with knowing the vocabulary.
The next time someone mentions their “fine-tuned inference layer with multimodal guardrails,” you can nod.
But this time — you’ll actually know what they’re talking about.
That’s not a small thing.