Saturday, 27 July 2013

AI Gone Wrong

This is a great little story I just heard about. An anonymous person posted on a forum the details of an elaborate little experiment using a video game called Quake 3 and a learning AI.

The bots in this particular game had the ability to learn effective strategies where the aim is to kill the other players and not be killed. He put sixteen of them in a match together to see how good at the game they would get, and if they would become essentially invincible against a mere human...

And he completely forgot about them. He remembered four years later, and went to check on them.

After learning, adapting, and evolving for four years, he came back to his bots and assumed something was wrong.

"I just checked on them but for some reason all the bots are just standing still. I'm gonna try changing the map... I guess it got stuck or something."

Various other posters commented:
"The only winning move logically for them, is to not play."

"Maybe they have learned that the best technique to survive is to make [sic] peace and to stand there for an eternity, waiting for a purpose or salvation."

"They've achieved something we couldn't. World peace."

It's pretty crazy, to think that sixteen AI programmed to kill each other in a free-for-all deathmatch could possibly find a peaceful solution and coexist in harmony.

But there's more. The story only gets better from there.

Assuming that something was broken, or bugged; the programmer decided to enter into the match to try and change the situation.

This was his report:

"I joined the server and the bots still just stood there, but the [messed] up thing was [that] they would rotate to look at me, I walked around a little bit and they all just kept looking at me. So I grabbed a rail gun and [killed] one of them, they all ran for the nearest weapons, took me down, and the server crashed."

Not only did the AI achieve world peace, but they killed an intruder who threatened their peaceful existence.

It's certainly a scary thought. I've posted about AI before, and about how robots could never take over the world or kill humans unless they were programmed to.

This story is the perfect example of what could potentially happen if you told an AI to be an effective killer, and also survive.

This just makes you shiver.

No comments:

Post a Comment