
The Alignment Problem

Such games are said to have “dense” rewards, which makes them relatively easy to learn.
Brian Christian • The Alignment Problem
As we’ve seen, one of the small but powerful techniques that made AlexNet so successful in 2012 was an idea known as “dropout”: certain parts of the neural network would be turned off—certain neurons simply would be “unplugged”—at random during each step of training.
Brian Christian • The Alignment Problem
The third advantage of imitation is that it allows the student (be it human or machine) to learn to do things that are hard to describe.
Brian Christian • The Alignment Problem
Infants’ preference for looking at new things is so strong, in fact, that psychologists began to realize that they could use it as a test of infants’ visual discrimination, and even their memory.
Brian Christian • The Alignment Problem
also seems that babies are keenly aware of when they’re being imitated.
Brian Christian • The Alignment Problem
“You start to pin it to incentive, motivation, as a response to a stimulus that makes you move. . . .
Brian Christian • The Alignment Problem
behaving predictably presumes that the observer knows what your goal is; behaving legibly presumes they do not.
Brian Christian • The Alignment Problem
At the other side of the spectrum from boredom is addiction—not a disengagement but its dark reverse, a pathological degree of repetition or perseverance.
Brian Christian • The Alignment Problem
cooperation works best when it is framed as an interaction, not as two separate and distinct “learn, then act” phases.