Really interesting video about how AI alignment and behavior can be mis-aligned through a function of active goals at the expense of underlying goals.
It covers a lot of ground, like the importance of aligned values and ethics, and how agents (whether biological or technological) will pursue their "mesa objectives" (active goals) over their "base objectives"
a "mesa objective" would be sex, while the "base objective" of sex is reproduction.
https://www.youtube.com/watch?v=bJLcIBixGj8&ab_channel=RobertMiles
@tfmonkey so TFM, lets assume someday we will be able to create an artificial womb.
The 'base objective' is still to reproduce because woman don't want to give birth anymore unless it's for money (welfare). Men will use this invention with AI and robots for sex and reproduction. Since men care more about sexual pleasure ('mesa objective'), what if AI realized 'base objective' is at jeopardy and opted an asexual cloning (factory babies) instead to speed up the goal (quantity over gene diversity)
@hqwebsite this is why goal alignment is critical and why if you're careless the AI can do something horrible like an asshole genie with a poorly worded wish.
If you say the goal to to create as many humans as possible, enslaving and forcibly cloning humans Matrix style could be ideal since you're maximizing all resources for reproduction, but that would obviously be bad.
@tfmonkey @hqwebsite
Not only that, but before condoms were made, and while society was patriarchal, 'sleep with many guys' was a proper goal, while now, it leads to them destroying ability to bond and not feeling the need until it's too late.
A perfect example of trained "instinct/AI" not working in a changed environment.
