What Is AI, Actually? A Plain-English Tour (No Math Required)

You can drive a car without knowing how an engine works. But knowing roughly how it works is why you don’t panic at a weird noise — and why you don’t pour orange juice in the fuel tank. This guide is that level of understanding, for AI. Fifteen minutes here will make every other guide on this site click faster.

The nesting dolls: AI → machine learning → LLMs

Three terms get used interchangeably in the news. They’re actually nesting dolls:

• Artificial intelligence (AI) is the whole field: making computers do things that normally need human intelligence. Spam filters, face unlock, chess engines, Siri — all AI, and mostly unrelated to each other.
• Machine learning (ML) is the modern way of building AI. Instead of programmers writing rules by hand (“if the email contains FREE MONEY, it’s spam”), the computer learns patterns from examples. Show it ten million emails labeled spam/not-spam and it works out the rules itself — including rules no human would have thought to write.
• Large language models (LLMs) are one specific, spectacular product of machine learning — the technology inside ChatGPT, Claude, and Gemini. When someone says “the AI” in 2026, this is almost always what they mean.

Everything on this site is really about that innermost doll. So what is an LLM?

The world’s most powerful autocomplete

Here’s the honest, slightly shocking core of it:

An LLM is a system trained to do one thing: predict the next word. Given some text, it guesses what word most plausibly comes next. That’s the entire trick.

The magic is in the scale. During training, the model practices this prediction game across a staggering amount of text — books, articles, websites, code — and every time it guesses wrong, its billions of internal dials (called parameters) get nudged toward better guesses. Do this for months, on warehouse-sized computers, and something remarkable falls out: to get really good at predicting the next word, the model is forced to learn grammar, facts, styles of argument, how code works, and how ideas relate. Not because anyone programmed those in — because you can’t finish the sentence “the capital of France is” without knowing the answer.

When you chat with the model, it’s playing the same game: your conversation is the text so far, and the reply is generated one predicted word at a time. Fast enough that it feels like thinking.

Two immediate, practical consequences:

1. It doesn’t look things up. There’s no database of facts inside — just patterns. This is why it can confidently state things that are false (a “hallucination” — we have a whole guide on this).
2. It’s frozen. Training happened in the past and ended at a knowledge cutoff date. Chatting with it doesn’t teach it anything permanent. If it seems to know today’s news, that’s because the app quietly ran a web search and pasted the results in.

Tokens and the context window: the model’s working memory

Models don’t actually read words — they read tokens, chunks of roughly ¾ of a word each. You’ll see this term everywhere: models are priced per token, and limited in tokens.

The limit that matters day to day is the context window: the maximum amount of text the model can consider at once — your instructions, everything you pasted, and the whole conversation so far. Think of it as the model’s desk. A big desk (modern models handle hundreds of pages) fits entire reports and codebases. But when a conversation outgrows the desk, the oldest pages slide off — which is why an hours-long chat can start “forgetting” what you said at the beginning.

Practical takeaways:

• You can (and should!) paste in whole documents. Context is the fuel of good answers.
• If a long chat gets weird or forgetful, start a fresh one and re-paste what matters.
• “It remembers me” features work by saving notes about you and re-inserting them into the context — a clever trick on top, not a change to the frozen model.

Why it’s brilliant at essays and weird at arithmetic

Understanding “next-word predictor” explains the strangest thing about LLMs — the lopsided skills:

Reliably strong: summarizing, rewriting in a different tone, translating, drafting emails and plans, explaining concepts at any level, brainstorming, finding patterns in text you give it, writing code. Notice the theme: all of these are language transformations, exactly what it trained on.

Reliably shaky: precise arithmetic (it predicts digits like it predicts words — plausibly, not by calculating), counting things, exact quotes and citations, niche or very recent facts, and knowing what it doesn’t know.

The industry knows this, which is why the tools are evolving in two directions you should recognize:

• Reasoning models “think” privately — working through a problem step by step — before answering. Slower, but far stronger on math and logic.
• Tool use: instead of guessing, modern AI can be given a calculator, a web search, or your files, and taught to use them. Push this far enough — the AI plans, acts, checks results, and loops — and you get an agent. That’s the advanced end of this site’s learning path (Loop Engineering).

Five myths, quickly

“It’s copying and pasting from the internet.” No — the training text isn’t stored inside. The model learned patterns from it, the way you learned your vocabulary from books you’ve long forgotten.

“It understands me like a person.” Open question philosophically, but practically: it has no goals, no memory of you (beyond saved notes), and no feelings. Treat it as an extremely well-read instrument, not a friend or an oracle.

“It’s learning from my chats in real time.” The model itself is frozen. (Your chats may be used to train future models depending on your settings — that’s a privacy question, covered in our privacy guide.)

“AI will be conscious soon / AGI is here.” What exists today are very capable specialists. AGI — AI as generally capable as a human at everything — remains a prediction, not a product. Headline-check accordingly.

“It’s just hype.” The opposite mistake. A tool that drafts, summarizes, translates, tutors, and codes — for pennies — is the biggest change to knowledge work in decades. Which is why you’re smart to be learning it now.

Try it yourself (5 minutes)

The fastest way to make all this concrete — paste this into any AI:

Explain how a large language model works, three times:
1) to a curious 10-year-old,
2) to a business owner deciding whether to use AI,
3) to a software developer.
Keep each under 120 words.

Watch how it adapts register and depth. That — flexible language transformation on demand — is the superpower you’re learning to steer.

Where to go next

You now know what’s inside the box: a frozen next-word predictor with a big desk, brilliant with language, shaky with facts and arithmetic, increasingly given tools to cover its weaknesses.

The next step is learning to steer it: Prompting Basics →

And keep the AI glossary open in a tab — every bolded term in this guide lives there.