Lesson 4 of 6

Learning past its teachers

6 min read

Most AI learns by copying human examples. But what happens when a system practises on its own — and finds a move no human would ever play?

Practice, not just imitation

Early game AI learned by imitating human play. AlphaGo went further: it practised against itself, millions of games, keeping what won. Freed from copying us, it played move 37 — a stone so strange commentators thought it a mistake. It wasn't. Learning by practice, not imitation, let it leave the human playbook behind and find something better.

A system that learns by practice can go beyond the examples it was given.

The promise, and the limit

This is why 'self-improvement' excites people: in narrow worlds with a clear score — games, some maths and code — systems can bootstrap past human skill. The catch is that clear score. Most of life has no tidy win/lose signal, so this hasn't turned language models into runaway self-improvers. It's a real, powerful pattern — with sharp edges about where it applies.

Self-improvement is real where the goal is crisply scored — and stalls where it isn't.

'It learns on its own' is true only where success is measurable. Be wary when someone claims open-ended self-improvement with no clear yardstick.