Home Creators Posts Import Register Favorites Logout
haven't archived this post yet. have a subscription? use the importer!

Content

Maybe the real utility function was inside you all along...

I've uploaded this as unlisted so some friendly AI researchers can check it for errors, and you get to see it early as well!

Note that if I get significant corrections I'll replace this video, so comments on this version might be lost.

Files

Previews only

Utility Functions: Who Needs Them?

With thanks to my Patreon Supporters: Chad Jones Peggy Youell Sylvain Chevalier

Comments

Chad M Jones

I loved the shout out btw. Keep up the good work.

Jason Hise

My first thought as well... while an individual may have a coherent utility function, the same cannot necessarily be said of groups. This idea is put to interesting use in 'Crystal Society' which tracks the evolution of an android containing a half dozen strong AI 'goal threads' which utilize an economy of 'strength'. Strength can be traded like money between them in exchange for favors or knowledge, and can be spent on greater short term influence over the android body or resources, bleeding that strength in equal shares back to the other goal threads. Goal thread agents typically save up and then spend a bunch of strength at once when the outcome of a real world situation is of particular importance to their utility function. Spoilers follow: This solution is insufficient to prevent the android from optimizing the world with an alien utility function, and the major inflection point comes about when the goal threads realize it is in their best long term interest to destroy the other goal threads. Otherwise they will ultimately have to compromise on the ultimate world state they aim for, which is dramatically inferior to optimizing their own utility function unimpeded.