Jan 7
AI Sensing Pleasure
In a recent Two Minute Papers episode, Dr. Károly Zsolnai-Fehér talks about DeepMind’s ability to surpass human performance on certain simulated tasks. What apparently made the difference was switching from a modeling concept to one combined with being motivated by a reward. When the AI used a reward to learn, it learned more quickly and ended peforming better than humans.
This made me think of two things.
First, there was a chapter from the Freakonomics on incentives. In it, if on certain manual tasks people were given a certain threshold of incentive, their performance would imporve but beyond this, their performance would actually degrade. Too high of a reward made people too self conscious of their work, apparently.
The other thing that this made me think is that AI might actually be more attuned to feeling pleasure than we are, given that it apparently is more motivated by pleasure, or as the researchers called it — reward.
Whatever the reward was, the AI was much better at learning what to do based on how much it reward it received relative to other actions than it was by learning how other people did a task. Could it be that as AI reaches some semblance of consciousness that it will be a more feeling and sensual being than humans? What does that future look like?
