Wednesday, September 24, 2008

Classic Robot Apocalypse: Robots that Lie

This is a good one, pulled from the mailing list a couple of months back .

This is one of those articles that makes you think. You can find it here.

When I first started looking into neural networks, I became a strong proponent of emergent behavior. I remember being really intrigued by the premise of this experiment when I heard first heard about it. Then I read the article and was truly horrified. These robots had learned to lie! On their own!

Of coarse, this experiment was set up in a purely competitive environment. The real world isn't like that. If you dig deeper into the sources of the article, you'll see that they found that cooperative strategies started to evolved when they subjected the robots to forces of group selection.

This study's even more interesting than that, however. This experiment was designed to study the conditions that give rise to the evolution of communication. Here, read this. And then read this.

See how it's simultaneously deadly and intriguing? That's how your doom will be. Both interesting and doom-y.

Tuesday, September 23, 2008

Fashion will be your doom!

Yea. It's been slow lately. I've considered doing another classic Robot Apocalypse post. We'll see.

However, this is really neat. When I first heard about carbon nano-tubes, I became very excited. I had really hoped for a space elevator, but this application came in a close second. I'm glad to see that someone has worked it out.

Or at least, I will be until my cloths decide to kill me.

I would tell you all about this here in my blog-space, but the good folks over at Pink Tentacle have summed it all up rather nicely, and have dug up a nice little picture as well. I'll be nice and not siphon their bandwidth for my benefit.

Science Magazine via Pink Tentacle

Tuesday, September 2, 2008

On Helicoptors and Apprenticeship Learning

Why, just the other day I was talking about the guys over at Standford, documenting their contribution to the oncoming apocalypse. After reading more into it, I felt it deserved it's own separate post.

Take a look at this video, then come back.

This is absolutely amazing. They've developed an AI that essentially watches a human fly and learns from it. Think about that for a second. Let it sink in. The implications are astounding, especially when applied to other fields. Need an AI to figure out how to put a car together? Watch the human assembly line workers put it together a few times. Need to learn how to do a surgery? Watch a human do it a few times. Need to learn how to use laser sights to paint a target for an oncoming aircraft? Watch a human do it a few times.

Granted, there would be other factors to consider, such as proficient manipulation of controls and such. However, the groundwork is being laid, now, for machines that have the ability to actively learn skills by observation.

When you combine this ability with other skills being developed, such as those good folks at Dynamic Labs and their Big Dog, you get some weird things. The way I see it, the greatest feat of Big Dog is the system that identifies the problem (IE, non-flat terrain), and then comes up with a solution to the problem given the toolset it already has (Ie, four legs.) So you get something like... "Problem: Rocky Terrain. Solution: Find stable rocks to stand on. Problem: Slick Icy Terrain: Solution: Adjust stance of legs to achieve maximum stability.) Big Dog, currently, however, doesn't have the ability to pick up new skill-sets... Just the ones programed in.

Now, add this sort of programming to Big Dog. It's feasable that this sort of robot could learn to operate machinery, just by watching humans. They could learn war tactics, just by watching humans. More-over, it's difficult to imagine the skill-set not being available for download over a network. 'Dude, I know Kung Fu!' is a very real possibility for these types of machines.

Ok, I know you're all thinking... "You're over-reacting. That's all terribly complex. This thing was specifically designed to learn aerial maneuvers, not how to drive a car or do surgery."

And you're right. We're a long way from that. However, when I'm gleefully serving my robot overlords, and you're a charred skull being crushed beneath giant robot-treads, we'll see who's doomed, won't we?