The AI words, in plain English

AI

A computer that writes, draws or talks back instead of just storing things. That is the whole idea, and everything else here is a name for part of it.

You type “write a thank-you message for a customer who just collected her order” and it writes one, in about four seconds. Nothing was looked up; it was written for you.

A prompt

What you type. That is all it means. When somebody says write a better prompt, they mean tell it more about what you actually want.

“Write a message” gets you something generic. “Write a short WhatsApp message to a customer whose order is two days late, friendly, apologise once, offer 10% off her next order” gets you something you can send. Same tool, same minute.

A model

The thing that answers you. ChatGPT, Claude and Gemini are different models, the way Toyota and Honda are different cars, and they get new versions every few months.

Ask all three to write the same advert and you get three different adverts, none of them wrong. That is why the course has you try more than one instead of marrying the first.

A chat

One conversation. Inside a chat it remembers what you both said; start a new one and it remembers nothing, which is why you sometimes have to say it all again.

Tell it about your shop on Monday and keep asking in the same chat all week, and it still knows. Open a fresh chat on Friday and it asks who you are, because to it you are new.

Context

Everything it can see right now, your words, and anything you pasted or uploaded in this chat. It knows nothing else about you or your work.

Paste your price list in, then ask “which three are under 500 taka?” and it answers correctly. Ask the same question in a new chat and it has never seen the list.

Making things up

It will answer even when it does not know, in exactly the same confident voice. That is not lying, it genuinely cannot tell the difference, and we spend a whole day on how to catch it.

Ask for the phone number of a shop it has never heard of and it may hand you one that looks perfectly real. Check anything you would act on, a number, a price, a date, a law.

Generative AI

The family name for all of this: a machine that makes something new instead of finding something that already exists. When somebody says gen AI, this is the thing they mean.

Google finds you a photo of a red sofa that somebody already took. Generative AI draws you a red sofa that has never existed. That difference is the whole word.

An LLM

The letters you will see printed next to model names. It stands for large language model, and it means a model whose one real skill is words.

When an advert says “built on the latest LLM”, it is saying “it is good with words” in a way that sounds more expensive. Now it cannot impress you.

Training

How it learned, by reading an enormous amount of writing long before you met it. That reading stopped on a date, which is why last week's news is often missing.

Ask about something that happened last week and it may not know, or may guess. Ask it to write a polite refund message and it is excellent, because that has been written for years.

A hallucination

The name the industry gives to making things up. Same thing, grander word, and now you will know what an article means when it uses it.

A headline reading “the chatbot hallucinated a court case” means it invented one and said it confidently. The same thing your own chat does with a phone number.

Text to image, text to video

You type a sentence and it draws or films it. Several of the tools in this course do exactly that, and the sentence you type is still just a prompt.

Type “a bar of pandan soap on a banana leaf, morning light, photographed from above” and you have a product picture, without a camera and without a studio.

An AI agent

A model allowed to take steps on its own instead of giving you one answer: search, open, write, check. Useful, and worth watching, because every step can be wrong.

Instead of answering “here is how to find suppliers”, it goes and searches, opens ten pages and hands you a list. Faster, and worth reading before you trust it.

A chatbot

The box you type into, on somebody's website or inside an app. It is not a different kind of AI, it is the doorway to one.

The little window on a courier's website that answers “where is my parcel” at two in the morning. What you are learning to use is what answers on their side.

Machine learning

The older, wider name for teaching a computer from examples instead of writing out every rule. All of this sits inside it.

Nobody wrote a rule for every way a customer can misspell your shop's name. It learned from seeing enough of them, the same way your phone learned to finish your sentences.

Prompt engineering

A job-advert word for the thing you will spend this course practising: saying clearly what you want. There is no engineering in it.

When you write “for a Bangladeshi customer, in Bangla, under 40 words, polite” instead of “write something nice”, you are doing the whole job.

A deepfake

A video or a voice of a real person, made by a machine, saying something they never said. The same kind of tool that makes an advert can make this.

A voice note that sounds exactly like a family member asking you to send money. Agree a question only the real person could answer, and ring them back on the number you know.

Automation

Getting the work to happen without you pressing anything, by connecting the tools you already use. The last day of the course is this.

A customer fills in your order form, and a thank-you message goes out and a row appears in your sheet, while you are asleep. You set it up once.

Your data

What happens to what you type. It is the first question most people ask, and it has a plain answer: it depends on a setting, and you can change it.

Every tool in this course has a switch that stops your chats being used to train it, and we show you where it is. Until you find it, do not paste a customer's phone number or NID.

That is enough to begin. Everything else you will pick up by doing it, which is the only way any of this sticks.