Wednesday, August 20, 2014

Perpetual Paperclips

The other day we had a discussion with some friends about robots, and automation. The trope was addressed about the perpetual paperclip making robot. Basically, you have this problem solving robot that provides you with all of your needs. And one day you tell it. "Hey, you, make as many paperclips as you can."

And then you die without ever stopping this task loop. So after a while, the robot runs out of wire. But since its a problem solving robot, it gets more wire. Eventually there is no more wire. But the robot starts looking at ways to solve the problem and start converting everything it encounters into a way to get more wire for making paperclips, and then makes paperclips out of them.

Eventually, the entire earth is consumed and converted into paperclips and the means to make more. But the Robot isn't finished. It sends out probes to the asteroids and planets in the solar system and gets more materials. And then the universe.

Anyway, there was one guy that argued that you can't teach a computer to teach itself, what it determines it needs to learn to complete the task. But it really just boils down diagramming out how you discover the need for new knowledge and then get it to complete tasks.

Here is a diagram:



The problem comes in where the robot needs to grow beyond the previously developed methods of material acquisition. How does the robot go from, "Get wire." with a response of "You cannot get wire." and move to "Make wire." and then "Mine ore to make wire."

And that is where google helps the robot. The robot gets a "You cannot get wire" response from the store, and then googles, "Where does wire come from." (this is the critical step) Google will respond with pages of information about how wire is made and what it is made of.

If the programming can interpret the data to make a new task of "Make wire." that then spawns tasks of "get metal" and then "Mine metal" and so forth. Then the robot can attempt to convert the universe into paperclips.

Since we have created a never ending task, the robot never reaches, a NO to the "Do we have a task?" question. There is always a task. And if it can't do them. It has google to help it get the information to do the task. And no human to interrupt the process and say, "Stop making paperclips."

This scenario is unlikely in many ways. Mostly in that someone will eventually notice the over abundance of paperclips and cancels the process.  But if we can develop the knowledge interpretation software enough, issuing a make as many as you can command will be like turning the faucet on. Walking off and leaving it will make a mess of things.

No comments:

Post a Comment