Via BB from a couple of days ago: this Wired article on how Google works.

You need three things for a genetic algorithm: a problem, a set of recombinable solutions, and a scoring-breeding mechanism.

This, for example, is all you need to make a relatively-small field-programmable gate array understand simple speech.

Google’s “signals” are it’s recombinable solutions, and its users — essentially everyone on the Internet — are its breeding and scoring system. The problem? Understanding language — AI-completeness itself.

If I was a Googlon my 20% time would be devoted to random generation, mutation, testing, and breeding, of new signals.

Perhaps Google really is SkyNet — no foolin’!

Quick! Someone ask it how the net amount of entropy in the universe can be massively decreased? (thx, Don)