Every AI word you will hear, in plain language

A machine that does things we used to need a person for.

That is the whole definition. Reading a letter and explaining it, writing a message, recognising a face in a photo, suggesting the next word in a text. None of it is magic and none of it is a robot with feelings. It is software that got good at tasks that used to require a human. Some of it has been in your phone for ten years.

Nobody wrote the rules. It was shown millions of examples until it got good at guessing.

The old way of building software was to write every rule by hand: if this, then that. Machine learning threw that out. Instead you show the machine millions of examples and let it work out the pattern by itself. That is why nobody, including the people who built it, can point at a line and say this is where it decided that. It learned the way a child learns a language, by soaking in examples, not by memorising grammar rules.

The finished result of all that learning. One very large file of patterns.

When the training is over, everything the machine learned is stored in a file. That file is the model. Claude, ChatGPT, Gemini and Copilot are each a company's model plus a nice screen to type into. When you hear someone say a new model came out, it means the training was done again, longer or better, and the result is sharper than the last one.

A model trained on text, whose entire job is guessing the next word.

Large because it read an enormous amount. Language because what it read was text. And here is the part that explains everything else on this page: its only skill is picking the next word that most likely follows. Do that over and over, very fast, and you get sentences, paragraphs, letters, whole reports. It is your phone's next-word suggestion, a million times stronger.

Training happened once, months ago. Your question uses the finished result, in a second.

These two things get confused constantly, and the confusion causes real mistakes. Training is the long expensive part: thousands of machines reading for months, done once, finished, over. Using it is what you do when you type a question, and it takes a fraction of a second. Your conversation does not train it. Your question does not teach it anything permanent. It is reading from what it already learned.

The date the reading stopped. It knows nothing after that, unless it looks it up.

Every model has a date after which it simply was not there. Ask about something from last week and it will either tell you it does not know, or, worse, confidently tell you something out of date. Many assistants can now search the web to fill the gap, and when they do they usually show you the links. The habit to build: for anything recent, ask where did you get that, and check the link.

It makes something new instead of finding something that already exists.

This is the real difference between Google and an assistant. Google finds pages that somebody already wrote. An assistant writes a new one, for you, that did not exist before you asked. That is why it can draft your letter and why it cannot give you a source for every sentence: there often is no source, because it just made the sentence. Both things are true at once, and knowing which one you need is half the skill.

It can look at a photo you take.

Point your phone at a letter from the school, a medical form, a bill, the back of a medicine box, a handwritten note, and ask what does this say and what do I do about it. For anyone dealing with paperwork in a second language, this is the single most useful thing on the list. It reads photographs of documents well, and handwriting less well. Session 7 is built entirely around this.

You can talk to it, and it talks back.

Press the microphone and speak normally, in French or in English. No typing at all. If you type slowly, or your eyes tire, or your hands are busy with the cooking, this changes what the tool is for completely. It also handles an accent far better than a keyboard handles a spelling.

Text, photo, voice and documents, all in the same conversation.

A fancy word for an ordinary thing. You can photograph a letter, ask a question about it out loud, and get a written reply you can copy, without ever leaving the conversation. When you see multimodal on a product page, that is all it is promising.

The chip that made all of this possible. Not one fast worker, but ten thousand slow ones working at the same time.

A normal computer processor is one brilliant worker doing one thing at a time, very fast. A GPU is the opposite shape: thousands of simple workers all doing the same small sum at once. It was built to draw video game graphics, where every pixel needs the same small calculation. Then somebody noticed that the arithmetic behind a model is exactly that shape too. That accident is why the AI boom happened when it did, and why these chips are now the most fought-over objects on earth.

The model does not live on your phone. Your question travels to a building full of those chips, and the answer travels back.

When you press send, your words leave your phone and land in a warehouse somewhere with thousands of GPUs and a serious electricity bill. The answer comes back a second later. Two consequences worth knowing: it needs an internet connection to work, and whatever you type leaves your device. Which is why Session 15 spends its time on what never to put in there.

Every answer costs real electricity and real hardware. That is why free runs out and why the best models are paid.

It is worth saying once, plainly, so nobody feels cheated when they hit a wall on a Tuesday afternoon. Each answer costs the company money. So the free plan gives you a good model with a cap, and when you reach the cap it either slows down or asks you to wait a few hours. That is normal, it is not a punishment, and you did nothing wrong. For all sixteen of our sessions the free plan is enough. Nobody has to pay to keep up with this series.