The AI words you keep hearing

Every article about AI assumes you already know the words. LLM, prompt, token, hallucination, agent, RAG. They get dropped into a sentence like you should nod along, and if you don't know them you feel like you missed the first day of class.

You didn't. Most of these words are simpler than they sound. Here is a plain decoder for the ones you actually hear, from someone who uses this stuff to get work done, not someone selling you a course on it. I am going to oversimplify a few of these on purpose. You do not need the precise textbook version to use the tool well. You need to stop getting tripped up by the vocabulary.

The thing itself

• AI. The umbrella term. Real talk: when someone says "AI" today, they almost always mean a chatbot like ChatGPT. The word covers a lot, but that is what people picture.
• Model. The actual AI. The trained thing that takes what you type and produces an answer. People say "model" and "the AI" to mean the same thing.
• LLM (Large Language Model). The kind of model behind ChatGPT and its cousins. "Large" because it read an enormous pile of text. "Language" because it works in words. At its core it is very, very good at predicting the next word. (More on that here.)
• GPT. A specific family of models from a company called OpenAI. The way people say Kleenex when they mean any tissue, a lot of folks say GPT when they mean any chatbot.
• Generative AI. AI that makes new things, text, images, audio, code, instead of just sorting or labeling things that already exist. That is the "generative" in GPT.

How you actually use it

• Prompt. What you type in. Your question, your instructions, the whole message. "Prompting" is just a fancy word for asking.
• Prompt engineering. Getting good at asking. Being clear, giving context, telling it what a good answer looks like. There is no real engineering involved. Anyone who briefs a coworker well can do it.
• Context window. How much the model can hold in its head at once. Paste in a long document and a small window starts forgetting the beginning by the time it reaches the end. A bigger window can keep more in view at the same time.
• Token. How the model counts text. Roughly, one token is about three quarters of a word. Limits and pricing get measured in tokens. You almost never need to think about it.
• Multimodal. It can handle more than text. You can show it a photo or play it audio, not just describe things in words.

Under the hood (you can skip this and be fine)

• Machine learning. Teaching a computer by showing it a mountain of examples instead of writing out step-by-step rules. It is the foundation under all of this.
• Neural network. The structure these models are built on, loosely inspired by how brain cells connect. You do not need to picture it to drive the car.
• Training. The long, expensive process of feeding the model all that text so it learns the patterns. It already happened, before you ever opened the app.
• Parameters. The dials the model tunes while it trains. There are billions of them. More parameters, very roughly, means more capable. It is the "billions of parameters" number people like to quote.
• Inference. The five-dollar word for the model actually running and giving you an answer. Training is the learning. Inference is the doing.

The ones that sound scarier than they are

• Hallucination. When the model says something false with total confidence. It is not lying. It has no idea it is wrong. It predicts words that fit, and sometimes the words that fit are not true. This is the whole reason you check anything important. (More on its limits in what AI can't do.)
• Agent. A model that does not just answer, it does. It can take steps, use tools, look things up, and work through a task across several moves. A chatbot answers your question. An agent goes and gets the thing done.
• RAG (Retrieval-Augmented Generation). Giving the model your own documents to pull from, so it answers about your business instead of just its general training. It is how you make it actually know your stuff. (I built one of these before I knew it had a name. That story is here.)
• Fine-tuning. Taking an existing model and training it a little more on your own examples so it picks up your style or your task. One way to customize it, not the only way.
• API. The plug that lets other software talk to the model. It is how AI gets built into an app instead of living in a chat window. (I shipped a working bot before I knew what those three letters meant.)
• Guardrails. The rules and limits put around a model so it behaves. What it will and will not do, and where it draws the line.
• AGI (Artificial General Intelligence). The hypothetical future AI that can do basically anything a person can. We are not there. When someone uses "AGI" in a sentence about what is available today, raise an eyebrow.

So what do you actually need to know?

Almost none of it, to start. I built a working system that answers underwriting questions before I knew half of these terms. I learned what to call the thing after I had already made it work.

The vocabulary is for people building the engines. You just need to drive. Keep this page handy for the next time someone drops one of these words like a test, and then get back to the actual work.