How Artificial Intelligence Actually Works

Start with the keyboard on your phone. You type "I'll be home in a" and it offers minute, bit, few. Nobody thinks your phone understands your commute. It has seen billions of sentences and it knows which word tends to come next.

A large language model is that idea, scaled until it stops looking like the same idea. Same core job: predict what comes next. The difference is the size of the pattern it's matching, and the size turns out to matter enormously.

That's the whole trick. Everything else is engineering.

Prediction, at a scale that gets strange

The systems behind ChatGPT, Claude, Gemini and the rest are built on an architecture called a transformer, introduced by Google researchers in a 2017 paper with the excellent title "Attention Is All You Need." The key move was a mechanism that lets the system weigh every word in a passage against every other word, all at once, instead of shuffling through them in order. That's what "attention" means in this context. It's how the model figures out that in "the trophy didn't fit in the suitcase because it was too big," the word it points at the trophy.

Text gets chopped into tokens first. A token is a chunk of characters, often a short word or a piece of a longer one, averaging somewhere around three-quarters of a word in English. The model takes your tokens, runs them through layer after layer of arithmetic, and produces a probability for every possible next token. Then it picks one, adds it to the end, and does the whole thing again.

One word at a time. Every time. There's no plan, no outline, no sentence sitting finished in a buffer waiting to be released. The essay assembles itself at the tip.

The weights are where everything lives

The arithmetic is governed by numbers called parameters, or weights. Modern models have hundreds of billions of them. Training is the process of setting those numbers.

It works by brute force and correction. Feed the system a passage with the next word hidden, let it guess, measure how wrong it was, nudge every relevant weight a hair in the direction that would have been less wrong. Repeat across an enormous corpus of text. Do it trillions of times.

This runs on racks of specialized chips, mostly Nvidia's, in buildings that draw power measured in megawatts and need serious cooling to survive their own waste heat. Training a frontier model is a capital project on the scale of a factory, which is why only a handful of companies do it. Running the finished model to answer your question is far cheaper, though at hundreds of millions of questions a day the electricity bill stops being a rounding error.

After the raw training comes the shaping. Humans rate outputs, the model gets tuned toward the responses people judged helpful and honest, and the polite, cooperative thing you talk to is the product of that second pass. The base model underneath is a much stranger animal.

Why it invents things

Here's the part people get backwards. The model isn't looking anything up.

There's no database inside it. Nothing retrieves a fact and hands it over. When you ask what year the Hoover Dam was finished, the correct answer comes out because the pattern "Hoover Dam" and "1936" appeared together across the training data often enough to leave a deep groove in the weights. Confidence and correctness are two different things, and the model only has access to one of them.

So when you ask about something thin in the training data, the machinery doesn't stop. It can't. It produces the most plausible-sounding continuation, which is how you get a court filing citing cases that never existed, or a perfectly formatted footnote to a book nobody wrote. People call this hallucination. It isn't a malfunction. It's the same process that gets the dam right, running on worse material.

Understand that and the fear and the hype both deflate a little. You're not dealing with an oracle or a liar. You're dealing with a system that has no concept of truth as a category, only of what usually follows what.

What it's genuinely good at

Plenty, and dismissing it's as silly as worshipping it.

• Transformation. Turning a messy transcript into clean notes, a long document into a summary, English into Spanish. The information is already in front of it, so the invention risk drops hard.
• First drafts. Anything where the blank page is the obstacle and you'll edit heavily anyway.
• Code scaffolding. Boilerplate, test cases, explaining what an unfamiliar function does. Programmers who use it well treat it like a fast junior who needs review.
• Pattern-heavy perception. Speech recognition has gotten remarkable. So has flagging anomalies in medical imaging, though that runs as a second set of eyes for a physician, not a replacement for one.

What all of those share: a human checks the output, and the cost of a mistake is a correction rather than a catastrophe.

What it isn't

It has never seen a chair. It has read about chairs in a hundred million sentences and can tell you a great deal about chairs. But there's no body, no room, no afternoon light, no memory of your grandfather's chair. Whatever the models are doing internally, it grew entirely out of text and images, and the disconnect shows up in physical reasoning and in anything requiring common sense about the world.

It also can't be accountable. Accountability requires somebody who can be wrong, know it, and bear the weight. A model can produce an apology-shaped sentence instantly and mean nothing by it. When a system's output causes harm, the responsibility runs to the people who deployed it, and any arrangement where that responsibility evaporates is a bad arrangement.

And it has no stake in your family, your marriage, your church, or your kid's character. Don't outsource the things that require love to a thing that can't have any.

Using one without getting burned

Four rules that will save you most of the trouble.

Verify anything with a number, a name, a date, or a citation in it. Those are exactly the details the model generates most fluently and most unreliably. Thirty seconds of checking.

Give it the source material. Paste in the contract, the article, the data. A model working from text you supplied is dramatically more accurate than one working from memory.

Don't paste anything you wouldn't email to a stranger. Medical records, Social Security numbers, client files, the details of a private family situation. Consumer chatbots aren't a confessional.

Ask it to argue against itself. "What's the strongest case that this answer is wrong?" It will often find the hole, because finding holes is also a pattern it has seen a lot of.

The jobs question, straight

Nobody honest knows the shape of it yet. What's visible so far is pressure on entry-level work in writing, translation, basic coding and customer support. Not elimination, mostly compression, the same way spreadsheets compressed the work of bookkeepers without ending accounting.

The work that looks most durable is work requiring hands, presence, trust, or a person willing to sign their name to a decision. Electricians are fine. Plumbers are fine. So is anyone whose value is that a customer looks them in the eye.

If you want a hedge, it's the oldest one there is. Learn a skill the world can't do without and be the kind of man whose word settles the matter.

The machine will hand you an answer every single time you ask. It will never once care whether it's true. That job is still open, and it's yours.