Seriously though: I just read an article describing efforts to create a robot/software that allows the robot/software to improve its own operation. Much like Natural Selection, the system had to go through an enormous amount of attempts to generate anything, and the vast majority of these were utter failures. But over time, the system got better and better, till it was outperforming the scientist's own programing of the same function. But, even more disturbing, the manner in which the robot/software did this was not understood - for instance, on the computer chip was the subject of the repogramming, there were sections entirely disconnected from the rest. We normally would never program like that. The scientist could not figure out what these isolated sections did, but lo! As soon at these sections were disabled, the entire process stopped.
So, robots designing better robots - that will be our Darwinistic doom. Maybe.
No comments:
Post a Comment