Shutterstock/Emre Akkoyun
What do paperclips have to do with the end of the world? Ask any researcher trying to confirm that artificial intelligence works for our benefit, and it’s more than you might imagine. It’s from
This dates back to 2003, when Oxford University philosopher Nick Bostrom posed a thought experiment. Imagine that a super-intelligent AI has set a goal to produce as many paper-his clips as possible. Bostrom suggested that he could quickly determine that killing all of humanity would be vital to the mission. The reason for this is the possibility of switching off and the fact that it is packed with atoms that can be transformed into more paperclips.
Of course, this scenario is absurd, but it presents a thorny problem. AI doesn’t “think” the same way we do, and if we don’t take great care in explaining clearly what we want it to do, it will behave in unexpected and harmful ways. There is a possibility. . “The system will optimize what you actually tell it to, but not what you intend it to,” says author Brian Christian. Alignment problem Visiting Scholar at the University of California, Berkeley.
The issue is whether AI is concerned with long-term existential risks such as the extinction of the human race or imminent harm such as misinformation and prejudice caused by AI. It boils down to the question of how to ensure that you make informed decisions.
Either way, the challenge of AI tuning is significant, Christian says, because of the inherent difficulty in translating vague human desires into the ruthless numerical logic of computers. He thinks the most promising solution is to have a human provide feedback on her AI’s decisions, which it uses for retraining…
