Against Amazement

There are other emotional reactions which should register as confusion but don't.

Imagine a smart person who sees asphalt being deposited to pave a road. "How disgusting," they think. "Surely our civilization can think of something better than this." They spend a few minutes ruminating on various solutions for road construction and maintenance that would obviously be better than asphalt and then get distracted and never think about it again.

They thus manage to never realize that asphalt is a fantastic solution to this problem, that stacks of PhDs have been written on asphalt chemistry and thermal processes, that it's a highly optimized, cheap, self-healing material, that it's the most economical solution by leaps and bounds. All they noticed was disgust based purely on error and ignorance.

Any thought of the form "That's stupid, I can easily see a better way" should qualify as confusion.

Juergen Schmidhuber has a theory of artificial curiosity. His theory proposes that seeking confusion is actually a good thing. Agents that seek out situations where surprising things happen, put their internal models to the test and learn the most. And that's all curiosity is.

Amazement is just a form of curiosity. People who are interested in AlphaGo have had their internal models of AI progress challenged, and are updating them.

The problem isn't amazement but cheap amazement. It's like the problem with eating fast food at McDonalds isn't about eating food but about eating easily digestable food.

The amazement that Feymann talks about from understanding a flower on a deep level is much better.

Noticing amazement can be as wonderful as noticing confusion.

As the old joke goes, Alzheimer's is the best illness, there is no pain and each morning you get lots of interesting news.

But note that improving the model would result in less pleasant experiences of wonder, but also in less unpleasant experiences of disappointment. Basically you reduce your variance, but it's not obvious to me that you imperfect model necessarily has a pessimistic bias.