How AI Actually Works
Someone asked me how AI works. Then someone else. Then someone else. Each time I started to answer, and each time I realized I didn’t actually know — not well enough to say out loud.
This is the explanation I wish I’d had. Six short chapters. The first one is the version that fits in thirty seconds. Each next chapter goes a little deeper. Stop wherever you’ve got enough; the next chapter is always one click below. Read all six and you’ll know more about AI than most adults.
Chapter 1 — AI is a really good guesser
Imagine someone who has read every book ever written. Not just read — memorized the patterns. Sentences. Paragraphs. How ideas connect. How questions get answered. How stories begin and end.
Now play a game with them. Say the first few words of a sentence and ask them to guess the rest.
“Once upon a time, there was a…” They’d guess “princess.” Or “dragon.” Or “small village in the mountains.” They’d guess well, because they’ve read so many stories that they know what usually comes next.
That’s AI. It’s a spectacularly good guesser at what word comes next, based on having read just about everything humans have written down.
When you type something into ChatGPT — a question, a half-finished sentence, an instruction — it isn’t searching a database. It’s guessing what would come next. Then guessing again. And again. One word at a time. The result feels like an answer or an essay or a poem. Underneath, it’s a guessing machine.
If that’s enough, you can stop here. Most of what people mean when they say “AI” in 2026 fits in that one paragraph.
If you want to know how it got so good at guessing — and why everyone is so excited (and a little worried) about it right now — keep going.
Chapter 2 — How we got here
People have been trying to make computers think for about seventy years. The story has three big chapters of its own.
1950s–1980s: rules
The first idea: if we just write enough rules — IF the sentence contains “capital of France,” THEN answer “Paris” — computers can fake intelligence. It worked for narrow tasks. It mostly didn’t work for anything interesting. Language is too messy. There are too many rules. The world refused to fit in IF/THEN.
1980s–2010s: learning from examples
A different idea: instead of giving computers rules, give them examples and let them figure out the rules themselves. This is called machine learning. Show a computer ten thousand pictures of cats labeled cat and ten thousand pictures of dogs labeled dog, and it can learn to tell them apart, even on pictures it’s never seen. Nobody wrote a “cat detection rule.” The computer found one.
The technology behind this is a neural network. The name comes from a loose analogy to brain cells (neurons). It’s layers of math that pass signals to each other and gradually learn to recognize patterns.
It’s a fair question whether this is also how your brain learns to tell cats from dogs. At a high level — see many examples, gradually build up patterns you can apply to new cases — yes, roughly. That part of the analogy is real. What’s different is the mechanism. Real neurons are chemical, fire in complex patterns over time, and have thousands of connections each. The “neurons” in code are just weighted sums of numbers passed through a math function. So the what (learn from examples) rhymes between brains and AI. The how is very different — and honestly, we don’t fully understand how brains do it either.
You already lived through the first wave of this. Amazon’s “people who bought this also bought…” Netflix’s movie recommendations. Spotify’s Discover Weekly. YouTube’s “what to watch next.” By the mid-2000s, all of these were running on machine learning — the same general idea, predicting what you’d like based on patterns in what millions of other people had liked. It was AI. Nobody called it AI. You just called it “how Amazon knows what I like.”
2010s: it suddenly worked for language and pictures too
Recommendation engines were a real success, but ML still stumbled at the harder stuff — understanding what’s in a picture, translating a sentence properly, holding a conversation. Then in the 2010s, two things changed: we got a lot more data (the whole internet, basically), and we got way more computing power.
The compute came from an unlikely place — graphics cards (GPUs) built by a company called NVIDIA for video games. It turned out the kind of math that lets a video game render fast — millions of tiny calculations happening at once — was exactly the kind of math neural networks need. NVIDIA’s software layer, CUDA, let researchers point these chips at AI problems instead of games. That accident is most of why AI took off when it did. (It’s also why NVIDIA, a graphics-card company, is now one of the most valuable businesses in the world.)
Suddenly neural networks started working. Image recognition got really good. Voice recognition got really good. Translation got really good. By 2016, AI was beating humans at the world’s hardest board game (Go). Things were accelerating.
And then, in 2017, a single research paper changed everything.
Want to know what was in that paper? → Chapter 3.
Chapter 3 — The 2017 breakthrough: attention
In June 2017, eight researchers at Google published a paper with one of the cheekiest titles in computer science: “Attention Is All You Need.”
The paper introduced something called the Transformer. It’s the engine that powers ChatGPT, Claude, Gemini, and basically every AI you’ve heard of. The “T” in “GPT” literally stands for Transformer.
Here’s what it changed.
Before 2017, AI read text one word at a time, like someone with very poor short-term memory. By the end of a long sentence, it had mostly forgotten the beginning. So it was bad at anything long — translation of full paragraphs, answering questions about a page of text, writing more than a sentence or two that hung together.
The Transformer did something different. It looks at all the words in a chunk of text at once, and it learns to pay attention to the words that matter most for whatever it’s trying to figure out.
Think about how you read a mystery novel. You don’t remember every word. But your brain pays extra attention to clues — a strange noise, a missing key, a name dropped twice. Background details fade; clues stand out. That’s attention.
The Transformer learned to do that for text. Reading “The cat sat on the mat because it was warm,” it figures out that “it” refers to the mat, not the cat, by attending to the relevant earlier words. Reading a long story, it can keep track of characters and what they’re doing because it’s attending to them across the whole text.
It sounds modest. It is, in a way. But it turned out to be the missing piece. With attention, AI could finally handle real, long, complex text without falling apart.
Want to know how an obscure research paper became the chatbot you can talk to? → Chapter 4.
Chapter 4 — From a paper to ChatGPT
After 2017, a small company called OpenAI started building on the Transformer. They called their model GPT — Generative Pre-trained Transformer.
- GPT-1 (2018) was a proof of concept. Most people didn’t notice.
- GPT-2 (2019) was good enough that OpenAI initially declined to release the full version, worried it could be misused.
- GPT-3 (2020) was a leap. It could write essays, code, jokes, emails. Researchers were stunned.
- GPT-3.5 and GPT-4 (2022, 2023) kept improving. So did competitors — Anthropic’s Claude, Google’s Gemini, Meta’s Llama.
The pattern across all of these: more data, more computing power, more parameters (a parameter is just a number inside the neural network that gets tuned during training — modern models have hundreds of billions of them). At each step, the model could do more. Sometimes it could do things nobody specifically taught it — solve word problems, debug code, translate poetry. Researchers started calling these emergent abilities. We don’t fully understand why they emerge. They just do.
But there was one more piece needed to turn a giant text-prediction engine into something you can actually have a useful conversation with.
A raw GPT model, trained on the internet, will write you anything — including factual nonsense, biased takes, harmful instructions, and strange non-answers. The internet contains all of those; the model learned all of those. To make it useful and (mostly) safe, OpenAI added a training step called RLHF: Reinforcement Learning from Human Feedback.
The idea is simple. Show the model two possible responses to a question. Have a human pick the better one. Do this millions of times. The model learns to prefer the kinds of answers humans actually want — helpful, honest, harmless, polite.
The combination — a giant pre-trained Transformer + RLHF — became ChatGPT. OpenAI launched it on November 30, 2022. It hit one million users in five days. A hundred million in two months. It’s the fastest-growing consumer product in history.
Claude, Gemini, and the other modern AIs all use the same basic recipe. Different details, similar story.
Want to know what’s happening right now? → Chapter 5.
Chapter 5 — What’s happening right now
Three big shifts are happening in 2026, all on top of the same Transformer foundation.
Multimodal AI — sees, hears, makes things
Early AIs only handled text. Now they handle images, audio, video. You can show ChatGPT a photo of your fridge and ask what to cook. You can have a real-time voice conversation with Claude. You can ask AI to generate a picture, a song, a video clip. The line between “AI that reads” and “AI that sees and hears and makes things” is dissolving.
Agents — they take actions, not just talk
A regular chatbot answers questions. An agent does things. An AI receptionist that answers your phone and books appointments. An AI that drafts and sends your emails. An AI that compares twenty flights and books the cheapest one. Most agents are still rough — they get stuck, make weird mistakes, occasionally book the wrong flight. But the trajectory is unmistakable. More and more “AI tools” will actually do things, not just suggest things.
Smaller, faster, cheaper
The biggest models cost tens or hundreds of millions of dollars to train. But researchers have figured out how to make smaller models that are almost as good and run on a phone. A model that would have required a supercomputer five years ago now runs on your laptop. Open-source AI — free for anyone to download — has caught up to the leading commercial models on many tasks.
Underneath all three, the same Transformer engine from 2017. Just bigger, broader, and aimed at more problems.
Want the big open questions even experts can’t answer? → Final chapter.
Chapter 6 — The big questions
This is the part nobody fully knows. Smart people disagree, sometimes loudly. Here are the four big ones.
You’re done
You now know more about AI than most adults.
You know the basic mechanism (pattern prediction). You know roughly how we got here (rules → machine learning → neural networks → Transformers → ChatGPT). You know what’s happening now (multimodal, agents, smaller models). And you know the questions even the experts can’t answer (Is it thinking? Is it conscious? Where is this going?).
One last suggestion. If you want to feel what was just described, open chat.openai.com or claude.ai — both free — and try the prediction game from Chapter 1. Type “Once upon a time, there was a” and see what comes back. Or paste a long article and ask for a summary. Or describe a problem and ask for ideas.
The first time you try it, you’ll feel the strangeness. After a few tries the strangeness fades and the usefulness takes over. That’s how everyone starts.
Welcome to it.
Did this explanation work for you? Did a chapter lose you, or did something not quite click? Email help@aiforyourday.com — this article gets rewritten when readers tell me where it went wrong.