Breath of the Wild is a physics engine you can think in
The new Zelda replaced scripted puzzles with a world that follows consistent rules. It's the best example I've seen of what a training environment for intelligence could look like.
I bought a Switch for The Legend of Zelda: Breath of the Wild two weeks ago and have been losing sleep over it since. Plenty of people have already written about how good it is. What interests me is how it’s built, because it’s different from almost every other game I’ve played.
In most games, puzzles are scripted. There’s one intended solution, the designer put the right object in the right place, and the game checks whether you did the thing. Breath of the Wild mostly doesn’t work that way. It has a small set of consistent rules and lets you combine them however you want.
Nintendo’s director talked at GDC this month about the game having a “chemistry engine” alongside the physics engine. Fire spreads through dry grass and creates an updraft you can ride with a glider. Metal conducts electricity, so in a thunderstorm you want to drop your metal sword. Wood floats, and you can freeze water into platforms. Wind pushes sails and also pushes flames. None of this is specific to one puzzle. It’s true everywhere in the world, and the puzzles are just places where the rules become useful.
So players solve the same problem in completely different ways, including ways the designers didn’t plan. I’ve seen people cross a chasm by building a bridge out of logs, by freezing a block of ice under a pillar to launch themselves, by setting fire to grass and riding the rising air. The game supports all of them because it never scripted the solution, just the physics.
What I keep thinking is that this is a much better environment for learning than the ones AI labs are using. Atari games teach reflexes. Mazes teach navigation. A world like this teaches something closer to intuitive physics and chemistry: if you learn once that metal attracts lightning, that fact is useful everywhere, and you can reason with it in situations you’ve never seen. The rules are simple enough to learn and rich enough to combine.
A child figures out the world in a similar way. There’s no scripted puzzle, just consistent rules about how things fall, break, float and burn, discovered by poking at them. My guess is that the agents that eventually learn general physical common sense will be trained in environments like this, systemic worlds with consistent rules and lots of freedom, not in environments with one correct answer.
Also, if anyone at Nintendo reads this: the rain climbing mechanic is cruel.