Lesson 7 of 7
Beyond human
6 min read
In 2016, an AI playing the board game Go made a move — its 37th — so strange that commentators thought it was a mistake. It wasn't. It won the game, and no human would have played it. What does that tell us about how AI learns?
A move no human would play
The AI was AlphaGo, playing the world champion Lee Sedol. Experts put its move 37 at about a 1-in-10,000 chance of being chosen by a human — it broke centuries of received wisdom. And it was brilliant. AlphaGo hadn't learned it from human games; it had learned by playing itself millions of times — reinforcement learning — and practice had led it somewhere no human tradition ever went. See the difference below.
Trained by practice rather than imitation, AlphaGo could reach moves outside everything humans would try — and one of them was better than anything a human would have played.
Beyond imitation
This is the deep promise of the training in this course. A model that only imitates us is bounded by us — it can be as good as the best human, not better. But a model that practises — reinforcing whatever actually works — can wander outside the human playbook and, now and then, find something new. That's how AI already surpasses people at Go and chess, and it's the direction reasoning models are inching toward on maths and code.
Imitation keeps AI inside the space of things humans already do. Practising against a real outcome is what lets it, occasionally, step outside that space and do something genuinely new.
Be honest about the scope: “beyond human” is clearly real in narrow, checkable worlds like Go, chess, and some maths — places where an attempt can be scored exactly. General reasoning is far harder to check, so today's models are only beginning to venture past us there. The mechanism, though — practise, reward what works, don't just imitate — is the same one that made move 37 possible.
Recap
- —Because it learns by practising, not copying, an AI can find moves outside anything humans would try — like AlphaGo's move 37
- —Pure imitation is capped at human level; practising against a real outcome can push past it
- —This is clearest in checkable worlds (Go, chess, some maths); general reasoning is only starting to get there
AlphaGo's move 37 is famous because no human would have played it, yet it was excellent. Which part of how it was trained best explains that?
Continue in the app
Take the whole How AI Learns to Think course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.