How AI Learns to Think

From rambler to assistant

A freshly trained model won't answer your question — it autocompletes it. Here's the training that fixes that.

Talking to a simulated expert

The assistant isn't a mind or a database. It's a character — a statistical impression of the helpful people it was shown, with a confident voice that is separate from whether it's right.

Thinking takes tokens

Why an AI forced to answer instantly fails — and the same model, allowed to think out loud, gets it right.

Practice makes perfect

Copying good answers only gets a model so far. The leap comes when it starts to practise — and keeps whatever works.

Watching it reason

Given room to practise, models learn to think out loud — and, strikingly, to catch and fix their own mistakes.

Learning what we prefer

Most questions have no answer key. So how do you train a model to be helpful or tasteful — and what goes wrong when you try?

Beyond human

Because it learns by practising, not copying, an AI can find moves no human would make — sometimes better ones.