Forget the sci-fi version for a minute.
When you type "Write me an email to a customer who's late paying" and the AI hands back something sensible, it isn't thinking like a person, and it isn't looking things up in a database of every fact on Earth (not the way most people picture it, anyway).
It's doing something much simpler — and weirdly powerful.
Next-word prediction, at massive scale
During training, the model read an enormous pile of text — books, websites, code, forums, the lot — and learned the patterns: after these words, these words tend to follow.
When you send a prompt, it builds the reply one piece at a time, each time picking what best fits the pattern given everything so far (your message, plus what it's already written).
That's it. Pattern matching for language.
Everything else — "reasoning", coding, summarising, role-play, Deep Research, thinking modes — is layers stacked on top of that same trick. Good enough to feel like magic. Still not magic.
I wasted weeks early on because I pictured it as "searching Google in its head." Wrong model, and it cost me. The day I swapped that for autocomplete with a PhD in confidence, I started writing far better prompts — and trusting it a lot less on any fact I hadn't checked myself.
Why that matters for you
- It's brilliant at language-shaped tasks — drafts, summaries, rewrites, explanations, ideas, code-shaped text.
- It can be confidently wrong — it predicts plausible text, not verified truth. Always sanity-check anything that matters (money, health, law, safety).
- Your prompt steers the pattern — vague in, vague out; specific in, useful out. (You'll practise exactly this later in the lesson.)
- The mode picker steers cost and depth — "fast" vs "thinking" is the same brand charging you different amounts of compute for harder work. More on that shortly.
A prompt isn't a Google search
| You might think… | What's closer to the truth |
|---|---|
| It looks up the right answer | It generates text that looks like an answer |
| Same question → same answer every time | A bit of built-in randomness can change the wording |
| It "knows" your business | It only "knows" what's in the chat (until you wire in docs or tools later) |
| One subscription = one capability | In 2026 you're buying a menu — chat, research, images, agents — each with its own limits |
Continue — next we map that menu across the big platforms.