An analogy I like for powerful narrow AI is that I’d hope it could be like public transit. A dedicated driver and maintenance crew make sure it operates safely, but many people get to interact with it and benefit from it. That doesn’t mean unauthorized members of the public get to modify or control it.
I think we’re going to need a system of responsibly managed narrow AIs as public goods in the future. This seems like a much easier problem to tackle than AGI alignment, but still an interesting challenge. It seems to me like email (and its spam filter) should be a public good like the post office. I don’t really trust my current state or federal government (USA) to safely and effectively manage an official email account for me though.
I’ve been thinking lately about how a positive vision of the future for me would likely involve giving everyone access to some sort of electronic assistant who its user could trust with information about their needs and get good feedback (general life advice, medical advice, career advice, counseling, finding and bringing to attention useful information). Making a system worthy of that trust seems difficult, but very beneficial if it could be implemented well.
Mostly agree. A sophisticated user might have some great feedback on how a website ranks its products, but shouldn’t and doesn’t want to have access to the internals of the algorithm. So giving some users a slightly better surface area for interaction aside from “being part of the training data” seems like an important problem to solve. Could be useful all the way toward AGI as we would need a story of how a particular person still has some capacity for future influence.
Yeah, good point. Some balance needs to be struck in such a scenario where people are given the power to do some sort of customization, but not so much that they can warp the intended purpose of the model into being directly harmful or helping them be harmful.
Yeah defining “harm”, or more formally, a “non-beneficial modification to a human being” is a hard task and is in many ways the core problem I am pointing at. Allowing people to take part in defining what is “harmful” to themselves is both potentially helpful as it brings local information and tricky because people may have already been ensnared by a hostile narrow AI to misunderstand “harm.”
An analogy I like for powerful narrow AI is that I’d hope it could be like public transit. A dedicated driver and maintenance crew make sure it operates safely, but many people get to interact with it and benefit from it. That doesn’t mean unauthorized members of the public get to modify or control it.
I think we’re going to need a system of responsibly managed narrow AIs as public goods in the future. This seems like a much easier problem to tackle than AGI alignment, but still an interesting challenge. It seems to me like email (and its spam filter) should be a public good like the post office. I don’t really trust my current state or federal government (USA) to safely and effectively manage an official email account for me though.
I’ve been thinking lately about how a positive vision of the future for me would likely involve giving everyone access to some sort of electronic assistant who its user could trust with information about their needs and get good feedback (general life advice, medical advice, career advice, counseling, finding and bringing to attention useful information). Making a system worthy of that trust seems difficult, but very beneficial if it could be implemented well.
Mostly agree. A sophisticated user might have some great feedback on how a website ranks its products, but shouldn’t and doesn’t want to have access to the internals of the algorithm. So giving some users a slightly better surface area for interaction aside from “being part of the training data” seems like an important problem to solve. Could be useful all the way toward AGI as we would need a story of how a particular person still has some capacity for future influence.
Yeah, good point. Some balance needs to be struck in such a scenario where people are given the power to do some sort of customization, but not so much that they can warp the intended purpose of the model into being directly harmful or helping them be harmful.
Yeah defining “harm”, or more formally, a “non-beneficial modification to a human being” is a hard task and is in many ways the core problem I am pointing at. Allowing people to take part in defining what is “harmful” to themselves is both potentially helpful as it brings local information and tricky because people may have already been ensnared by a hostile narrow AI to misunderstand “harm.”