A very intriguing AI development

Reinforcement learning has been a hot tech in AI for a while now. The basic idea is to use rewards to steer an AI toward a goal you want it to reach. It's one of the key techniques that Google's AlphaGo uses, and that program just won the first game (of several) against the highest-ranked human Go player. A problem for this approach, however, is that often there just aren't a lot of rewards on the way to a goal.

A new and fascinating paper, which you can read about here in the MIT Technology Review, offers another way to keep the AI improving: give it curiosity. Mind you, this is not curiosity as we experience it, but rather a very self-focused desire to explore whatever in the world can directly affect the AI. This artificial curiosity can encourage the AI to keep busy improving itself even when no rewards are in sight.

I find this both exciting and a bit frightening. Imagine any entity that explores the world and learns from it but cares only about itself. Phrased that way, this sounds more dangerous than it is--right now. I'm also ignoring the likely safeguards programmers would embed in the AI. Still, it's a fascinating and a bit scary prospect, and one I hope to follow.

Interesting times.