Thirty words that cover nearly every AI conversation you’ll be in this
year — defined for a smart person who has better things to do than follow AI news.
The basics
AI (artificial intelligence)
Software that can read, write, and reason well enough to help with real work. The umbrella term for all of it.
AI assistant / chatbot
The chat window where you talk to an AI. ChatGPT, Claude, Copilot, and Gemini are all assistants.
LLM (large language model)
The engine inside every AI assistant. It has read an enormous amount of text and predicts what comes next — which turns out to be astonishingly useful.
Model
One specific engine, with a name and a version — GPT-5, Claude, Llama, Qwen. They differ in size, speed, and skill, the way vehicles do.
Generative AI
AI that creates something new — text, images, audio, video — instead of just finding or sorting what already exists.
Prompt
Whatever you give the AI: the question you type, the document you upload, the instructions you set. Your side of the conversation.
How it works (just enough)
Token
How AI measures text — about three-quarters of a word each. Public AI tools count every token you use, like a taxi meter.
Context window
How much the AI can keep in mind at once. Overfill it, and the earliest parts of the conversation fall out of view.
Training
School for the model: showing it examples so it learns. The important question is always whose examples — and whether your conversations quietly become part of the lesson.
Inference
The model doing its actual job: reading your prompt and producing the answer. Training is school; inference is the workday.
Hallucination
When AI states something false with complete confidence. It isn't lying — it's guessing well. Check facts before you rely on them.
GPU
The chip that does AI's heavy math — and the reason NVIDIA is in the news constantly.
Words you'll hear in 2026
Reasoning model
A newer kind of model that thinks through a problem step by step before answering. Slower, but much better at math, logic, and multi-step work.
Deep research
A mode where the AI spends several minutes searching, reading, and writing a sourced report — instead of a quick chat answer.
AI agent
AI that doesn't just answer — it does. It can browse, click, fill things in, and work through multi-step tasks with light supervision.
Multimodal
AI that handles more than text: it can look at images, listen to audio, and read documents — sometimes watch video.
Copilot
AI built into software you already use — your documents, spreadsheets, email. Microsoft's product is named Copilot, but you'll hear the word used for the whole category.
Frontier model
The biggest, newest models from the major labs — the ones setting the pace this year.
Deepfake
AI-made audio or video of a real person saying or doing something they never did — convincing enough that seeing is no longer believing.
AI slop
The flood of low-effort, AI-generated content filling the internet. Knowing the name helps you spot it.
Shadow AI
Employees using AI tools the company doesn't know about — usually personal accounts, usually with company information. It's the norm, not the exception, and it's why businesses write AI use policies.
Where the AI runs (the part that matters for privacy)
Cloud AI
The engine runs on the vendor's computers, and your prompt travels to them. “The cloud” is nothing mystical — it's someone else's computers.
Private AI / local AI
The engine runs on hardware dedicated to you, and your prompt stays put.
Open vs. closed models
Open models are released for anyone to download and run on their own hardware (Llama, Qwen, GPT-OSS). Closed models run only on the vendor's servers (ChatGPT, Claude, Gemini). Good open models are what make a genuinely private server possible.
RAG — “AI that reads your documents first”
Instead of answering from memory, the AI looks things up in your files and answers from what it finds — with far fewer made-up facts.
VPN
An encrypted tunnel between your computer and a server. Anyone in between sees scrambled traffic. It's how an office reaches a private server safely.
API
How one piece of software talks to another. An API lets your business systems use AI directly, without a chat window.
Rate limit
A cap on how much you can use a service — messages per hour, pages per upload. Hit it and you wait, get a slower model, or pay up. A busy signal, basically.
Concurrent users
How many people are using a system at the same moment — not how many have logins. The honest way to size capacity.
Guardrails
The rules and filters built into an AI tool that stop harmful or off-limits answers.
AI moves fast — these are the terms as of mid-2026.
