Finally finished "Thinking Fast and Slow", highly recommend it, great survey of behavioral psychology and economics. But what really caught my attention is the "two systems" discussion and it's implications for AI.

Trying to summarize, Kahneman describes experiments that lead to his conclusion that there are two distinct systems in the brain.  The first system is fast and cheap and tries to match the current situation to pattern seen before and then just uses whatever worked last time.  The second system fires whenever you willfully force it to or when the first system hits an unexpected situation (like not being able to find a pattern, or starting to execute on a plan that worked last time and finding that it's not doing what was expected).  This second system boots up your full cognitive processing to solve hard problems and is very slow.

The second system is also very expensive (in terms of calorie consumption, you can see it in increased heart rate and respiration) and triggers feelings of displeasure when it fires (probably to discourage its use, evolutionarily, laziness has its advantages). This expensive second system also can only do one thing at a time (leading to, for example, the need to stop talking when you are driving a car if you see an unexpected or potentially dangerous situation ahead).  Because it is so expensive, it is also hard to fire up when people are physically tired or under cognitive load (people will just give up on solving hard problems under these circumstances, instead choosing to guess or not answer).

I'm not aware of work in AI that explicitly tries to mimic the two systems Kahneman describes seeing in his behavioral experiments.  Closest I know of is explore-exploit in machine learning and various systems caching partial results from computation, neither of which isn't the same.  If you know of work like this, please let me know.  And, if you are interested in AI or an AI researcher, might consider picking up this book, I suspect you'd enjoy it, I certainly did.
Shared publiclyView activity