Say you want to make somebody in a neighboring country smile, but: - you can’t travel or send a message to their country - some people do interact with their country, but you’re not sure who - you have a map of everyone in their country you bought - you have a powerful AI A naive approach might be to train your AI to make people smile at a distance by influencing the people in between. Unfortunately, when only the smiling of the distant person is in the AI’s reward function, it may learn to enslave the people in between so as to: - report on the person’s smiling - form groups discussing how much they smile and what influences it - stimulate them smiling in a way totally unknown to you, excessively and perversely, and - be left confused and helpless, wondering where their god went, when you pull the AI out. I was briefly thinking on this challenge because it is similar to the challenge of controlling your body when you have a severe dissociative palsy from items similar to the ones listed above. [It also overlaps with other extreme puzzles like teleportation] Basically, the challenge of rapidly influencing something at a distance _without_ harming anything between you and it simplifies to the challenge of even more rapidly influencing things nearby in a harmless way, such that they participate in the more distant influence in a way that is to their forthright benefit. Ideally, of course, we communicate instead of influence, and we hope that influence without harm simplifies to communication and beg people to make requests of us rather than throw us to the runaway paperclip factories. Anyway, simple ideas start out with rapidly comprehending communities and individuals and preparing for them “information and experience packages” that aid them, help them understand what is going on, and make it forthrightly clear that is to the benefit of their values, selves, and communities to help the more distant relation happen. In a situation like this, if a model of somebody’s values is wrong, it’s important to pull out quickly and make things right, because the impact is so large. Given the goal is rapid distant effect, it’s hard to quickly figure out how to hold the meaning and importance of two-way communication, and I suspect theories similar to faster-than-light-discovery and attempts at time travel could be a little relevant — but I’m thinking providing living beings or robots that can learn independently how to resolve things in forthright caring, respectful, and contributive ways could be helpful if back-and-forth communication remains a challenge.