Imagine a "Sims" game produced in the far, far future. The "people" in the game are produced by sophisticated AI; they're almost human. You want them to do science; figure out how their world works.
Also, you want the game to last a long time.
One problem is obvious: make the "world" they live in too simple they'll figure it out quickly. But the other extreme may discourage them so much they give up. (A third problem is that they may invent something totally unforeseen; if the Sim doesn't have robust error-handling, it'll crash.)
What to do?
Suppose I give the Sim's physics engine AI of its own; an AI slightly more powerful than the Sims'. As the Sim's subject the Physics Engine to more sophisticated inquiries, it evolves more sophisticated responses. When the Sims discover the "black body" problem, the Physics Engine invents quantum mechanics.
The physics engine can't contradict itself; if gravity makes things fall down, it can't start making things fall up. Apparent contradictions are allowed. And the Physics Engine, being intelligent, will game the rules, even cheat, if it can.
Intelligence makes mistakes. Find something wrong, and you've found intelligence.
No comments:
Post a Comment