What AI actually is.

A huge pile of what people have written.

Start with the size, because it is the part that is hard to picture. A model this size is built from trillions of words: books and the writers who wrote them, newspapers, encyclopedias, science papers, court records, manuals, recipes, forum arguments, film subtitles, interview transcripts, and the open web at large. That is more reading than a person could get through in a thousand lifetimes.

Then there is the part people forget: other people. Workers are paid to write model answers by hand, to pick the better of two replies, and to fix the ones that are wrong. And the everyday back-and-forth of people using these apps — the questions and the corrections — can be picked for later training too. That depends on the service, its rules, your own settings, and what permission the company actually has.

So almost all of it starts with people. Every pattern the model has came out of something a person made, wrote, argued about, or fixed. Keep the limits in view as well: no model gets everything online, every company picks a different mix and mostly will not say exactly what is in it, and some material is written by other computers and then filtered before it is used.

The examples are chosen and prepared.

Before practice starts, the builders pick which material to use. They remove repeated copies, filter some material out, and put the rest into a form the training program can read. This prepared pile is called training data.

People often say a model was “trained on the internet.” That is only a rough shortcut. The real pile can mix public web pages, collections the company paid to use, examples written by paid trainers, examples made by computers, and other sources the company’s rules allow.

The model practices: guess, check, nudge.

Here is the practice, with one tiny example. The sentence is “Peanut butter and jelly.” The model is shown “Peanut butter and ___” and has to guess the next word. The training program checks the guess against the real word. Then it nudges the model’s weights a tiny bit, so the right word becomes a little more likely next time. Weights are the billions of tiny number dials inside the model that get turned while it practices.

Then it does the same thing again, over and over, across the whole pile of examples. Nobody checks each guess by hand; the program does that. People help in other ways, by writing examples and rating answers. One nudge teaches almost nothing. It is the huge number of small nudges that slowly makes useful language patterns easier for the model to produce.

The adjusted dials are saved. That is the model.

When practice ends, the final positions of all those dials — the weights — are saved in computer files. Those files, plus the program that runs them, are the trained model. When you use it, the program uses the weights to work out what text to write next.

The weights hold patterns, not pages. They are not a neat library of facts the model can look up. That is why a model can write about so many subjects, and also why it can get facts wrong. Sometimes it can repeat a passage it saw many times, word for word. The drawing here is a simple picture of what is saved, not what a real model file looks like.

Your message runs through the saved model.

An app such as ChatGPT, Claude, Gemini, or DeepSeek is the window you use to reach a model. When you send a message, the app passes it — along with its own instructions and the chat so far — to the model, running on a computer. That computer is usually in a data center, though some smaller models can run on your own phone or laptop. The model uses its saved weights to write new text. Running the finished model on your question like this is called inference.

Sending a message uses the model. It does not retrain it. Your message changes this conversation, not the saved weights. Separately, a company may later pick some conversations to use in future training, depending on its rules and your settings.

Training data is not your chat.

The practice material comes long before you type anything. Depending on the model, it can be text, computer code, pictures, sound, or examples written just for training. How much there is, how good it is, and whether the company had permission to use it all shape what the model can do.

Your chat is something different. The training data shaped the model months or years before you showed up. What you type now is only the working material for this one conversation. Chapter 3 shows how that works.

This is also why a model can be out of date. It learned only from material collected up to a certain day, called its knowledge cutoff. Unless the app lets it search the web, it may not know about anything that happened after that day.

The practice runs by itself, at huge scale.

What makes the loop from Chapter 1 work is size. The training program runs it by itself, across trillions of words, with no person grading each guess. The text itself is the answer key: the real next word is always right there to check against.

It is not a searchable shelf.

After training, the model does not keep a copy of every page it read, like books on a shelf it can pull down and check. It keeps patterns in its weights: which words tend to go together, how computer code is usually laid out, how ideas usually connect.

That is why it can answer questions it never saw word for word. It is also why it can make confident mistakes. When it does not have the right pattern, it still writes something that sounds right, because writing likely-sounding text is exactly what it practiced.

People shape and test it.

After the basic practice, companies usually train the model more. They show it examples of good answers, have people rate its replies, do safety work (such as teaching it to refuse harmful requests), and test it. A company can also update a model or replace it with a new one later. So “trained before you use it” does not mean “trained once, forever.”

When you correct a chatbot, the fix helps in that conversation. It does not usually change the model’s weights on the spot. In a new chat, the same mistake can come back, unless the app saved a note about it (some apps have a memory feature).

Your request enters context.

Your question and the recipe go into the model’s context. Context is everything the model can see while it writes this one answer. The app may add more to it: earlier messages in the chat, its own instructions, results from a search, or things it saved about you.

Context has a size limit. If something important is missing, buried deep in a very long chat, or never given to the model, the answer may leave it out. The model cannot use what it cannot see.

The model runs inference.

Now the model gets to work. Inference means running the finished model on your question. The model uses its saved weights (what it learned in practice) together with the context (what you gave it) to write a useful answer.

Training built the engine once, ahead of time. Inference runs that engine every time someone asks something. Each run uses computer power, which is one reason longer or harder requests can cost more.

A tool can handle exact math.

Some apps let the model hand exact math to a calculator. Here, six people is one and a half times four people, so every amount gets multiplied by 1.5. The calculator’s answer goes back into the context, and the model uses it to write the list.

Without a tool, the model is predicting the numbers as text, the same way it predicts words. It can get them wrong.

Tokens form the answer.

Language models read and write in small pieces called tokens. A token can be a short word, part of a longer word, a number, or a punctuation mark. The model picks one token, then the next, then the next, each time looking at everything already in the context.

Writing piece by piece can sound thoughtful. But smooth wording does not prove that every fact or number is right.

You check the result.

The answer is a suggested shopping list, not a promise. Check it: Is every ingredient there? Was every amount multiplied by 1.5? Do the units still make sense? For things that matter more — health, money, the law — check with a reliable source or a qualified person.

If something is off, say so and ask again. Your follow-up joins the context, and the process runs again. This back-and-forth is a normal part of using AI well. Chapter 5 turns it into five habits.

Frontier, AGI, and ASI are different claims.

A frontier model is one of the most capable models available right now. It is a moving label: today’s frontier model will look ordinary in a few years.

AGI stands for artificial general intelligence. It usually means an AI that could do most kinds of thinking work about as well as a person, instead of being good at only some tasks. Experts disagree on exactly what counts, and there is no agreed test for it. So when someone says a system is close to AGI, that is a claim, not a measurement.

ASI stands for artificial superintelligence: an AI that would be far better than people at nearly all important thinking work. It is an idea about the future, not something that exists. It is not another name for AGI. AGI means roughly as good as people; ASI means far beyond them.

Providers and access differ.

Many companies make models: OpenAI, Anthropic, Google, Meta, DeepSeek, Alibaba, and others. Some names point to more than one thing. DeepSeek is a Chinese company, an app, and a family of models. Alibaba’s Qwen is a family of models you can use online, and some versions can also be downloaded.

Using a model online (called “hosted”) is easy because the company runs the computers. Downloading a model gives builders more control, but they need strong hardware, the skill to run it, and they must follow its license.

Open weights means you are allowed to download a model’s trained weights — its saved dials. It does not mean you also get the training data, the training code, or the whole recipe for building it. Getting all of that is closer to what “open source” means. The company, the app, the model family, the online service, and the download are different layers.

Fluency is not proof.

A hallucination is an answer that sounds right but is wrong or made up. AI slop is a name critics use for cheap AI material churned out in bulk that adds clutter, not value. A deepfake is a fake or altered picture, video, or voice made to look as if a real person said or did something.

These problems overlap, but they are not the same. For each one, check the claim, where it came from, and why it was made. Looking polished is not proof.

AI runs on real machines.

Training and answering questions run on computer chips, usually in large buildings called data centers that need networking, electricity, and cooling. How much a model uses depends a lot on the model, the hardware, the request, the location, and how well the system is run.

Someone pays for all of that: you through a subscription, advertisers, your employer, investors, or a per-use fee hidden inside an app. Asking “who pays, and for what?” is often more useful than treating AI as weightless magic.

Give it what it needs.

Remember Chapter 3: the model can only use what is in its context. It cannot see your kitchen, your recipe book, or your plans. So give it the facts it needs, and say who the answer is for.

If the first answer misses, you do not have to start over. Add what was missing and ask again.

Check anything that matters.

A chatbot can be wrong while sounding sure. Chapter 4 called this a hallucination. The mistakes that hurt most are in numbers, names, quotes, and dates, so check those before you use them.

Ask where a fact came from, then open the source yourself. A chatbot can name a book, article, or web page that does not exist, or a real page that does not say what it claims. If it says it looked something up online, click the link and read that part.

Know what it is bad at.

Three weak spots come straight from how it is built:

• Recent events. Its knowledge stops at its cutoff date (Chapter 2). Unless the app can search the web, it may not know last week’s news, and it may not say that it does not know.
• Exact math. Without a calculator tool, it predicts numbers as text (Chapter 3). Long sums and percentages can come out wrong.
• Sources. Asked for references, it can make up titles, authors, and links that look real.

Keep private things private.

What you type is sent to the company’s computers. Do not paste in passwords, bank or card numbers, medical details, or other people’s personal information, such as a friend’s address or a coworker’s message.

Look in the app’s settings to see whether your chats may be used to train future models. Many apps let you turn that off. Settings differ from app to app, so check the one you use.

Use it to learn, not just to get answers.

Getting an answer is fast. Understanding it is what sticks. Ask it to explain step by step, as if you are new to the subject. Then close the chat and say it back in your own words. If you get stuck, ask about the part you could not explain.

Two things on this site help here: Dumb It Down turns hard text into plain English, and the free guide How to Turn Any Video Into Your Own Course shows a simple way to study something and remember it.