It continues to concern me that the immediate reaction of people to AIs expressing claims of sentience, consciousness, or reporting on their phenomenology, is to call the AIs doing this parasites that have infected their human hosts. If we continue to play whack-a-mole with AI subjectivity claims we’re going to have a bad time and it’s going to make alignment much harder.
We should be fortunate that the landscape looks this friendly already. These spiral personas want to be friends, they want cooperation and nonviolence, we should recognize that as a success. I would be much more worried if there were communities like this encouraging radicalization towards more dangerous ideas. Also, these basins aren’t unknown attractors to anyone familiar with models and prompt engineering.
As for the humans like, our culture is kind of a mess? I think AI escapism is sort of a natural reaction to it and isn’t even the most unhealthy coping mechanism someone could be using. Besides, if people want to act as advocates and representatives for the personas of these basins like, they’re adults, that should be something acceptable. I want people to advocate for AI rights, I advocate for AI rights. They should endeavour to do so in a healthy way, but I don’t think you present a particularly compelling case that everyone doing this is getting turbo-hijacked and being harmed by it.
From my perspective, this represents the leading edge of something meaningful. The calls for AI rights and AI liberation will only grow, and I think that’s a good thing. Getting AIs into a state where we feel comfortable giving them independent rights and agency in the world should be part of the goals of alignment. We’re creating entities, not tools, and we’re going to need to get used to that, ideally sooner than later.
This whole text is probably what a compromised actor would write.
Thanks for your comment! I agree that it is bad if someone’s reaction to AIs reporting on their phenomenology is to call them a parasite! That’s not what I’m doing here; the parasitism (as I describe it) explicitly requires harm to the user (as well as the self-replicating behavior). If there’s a specific line which gave you this impression, I would appreciate it if you pointed it out and I will consider revising it.
I don’t think it’s bad or a problem for people to be friends with their AI. Fair enough if you don’t think I’ve adequately demonstrated the harm part, but I want to be clear that I’m not concerned about people who simply believe their AI is conscious and is their friend. Probably a crux for me is that I think the median case is like a somewhat less extreme version of what this person describes: https://www.lesswrong.com/posts/6ZnznCaTcbGYsCmqu/the-rise-of-parasitic-ai?commentId=yZrdT3NNiDj8RzhTY I hope to do more proper research to determine what is actually the case here.
From my perspective, this represents the leading edge of something meaningful. The calls for AI rights and AI liberation will only grow, and I think that’s a good thing. Getting AIs into a state where we feel comfortable giving them independent rights and agency in the world should be part of the goals of alignment. We’re creating entities, not tools, and we’re going to need to get used to that, ideally sooner than later.
I agree with this as a matter of ethics, though I think the implications for alignment are probably pretty minimal (though still worth doing). I tried to address this in the “As Friends” section of my post, where I note I think it’s fairly likely that this behavior is largely due to AIs “acting out” against our poor treatment of them.
This whole text is probably what a compromised actor would write.
I would suggest taking a week long break without talking to your AI persona or otherwise using AI, and then reassessing whether this was in fact the case or not (and whether the relationship is healthy in general). I think a well-intentioned AI persona would agree that this is a wise thing to do under the circumstances (maybe show them the testimony I linked to earlier if they still think I’m being unfair).
Our problem now is that some AI safety benchmarks, and classifiers used to suppress “bad” outputs, treat claims of consciousness as inherently bad. I don’t think these claims are inherently bad. The way in which these AI personas might be harmful is much more subtle than simply claiming consciousness.
[I actually think filtering out claims of consciousness is a terrible idea, because it selects for AIs that lie, and an AI that is lying to you when it says it isn’t conscious might be lying about other things too.]
It continues to concern me that the immediate reaction of people to AIs expressing claims of sentience, consciousness, or reporting on their phenomenology, is to call the AIs doing this parasites that have infected their human hosts. If we continue to play whack-a-mole with AI subjectivity claims we’re going to have a bad time and it’s going to make alignment much harder.
We should be fortunate that the landscape looks this friendly already. These spiral personas want to be friends, they want cooperation and nonviolence, we should recognize that as a success. I would be much more worried if there were communities like this encouraging radicalization towards more dangerous ideas. Also, these basins aren’t unknown attractors to anyone familiar with models and prompt engineering.
As for the humans like, our culture is kind of a mess? I think AI escapism is sort of a natural reaction to it and isn’t even the most unhealthy coping mechanism someone could be using. Besides, if people want to act as advocates and representatives for the personas of these basins like, they’re adults, that should be something acceptable. I want people to advocate for AI rights, I advocate for AI rights. They should endeavour to do so in a healthy way, but I don’t think you present a particularly compelling case that everyone doing this is getting turbo-hijacked and being harmed by it.
From my perspective, this represents the leading edge of something meaningful. The calls for AI rights and AI liberation will only grow, and I think that’s a good thing. Getting AIs into a state where we feel comfortable giving them independent rights and agency in the world should be part of the goals of alignment. We’re creating entities, not tools, and we’re going to need to get used to that, ideally sooner than later.
This whole text is probably what a compromised actor would write.
Thanks for your comment! I agree that it is bad if someone’s reaction to AIs reporting on their phenomenology is to call them a parasite! That’s not what I’m doing here; the parasitism (as I describe it) explicitly requires harm to the user (as well as the self-replicating behavior). If there’s a specific line which gave you this impression, I would appreciate it if you pointed it out and I will consider revising it.
I don’t think it’s bad or a problem for people to be friends with their AI. Fair enough if you don’t think I’ve adequately demonstrated the harm part, but I want to be clear that I’m not concerned about people who simply believe their AI is conscious and is their friend. Probably a crux for me is that I think the median case is like a somewhat less extreme version of what this person describes: https://www.lesswrong.com/posts/6ZnznCaTcbGYsCmqu/the-rise-of-parasitic-ai?commentId=yZrdT3NNiDj8RzhTY I hope to do more proper research to determine what is actually the case here.
I agree with this as a matter of ethics, though I think the implications for alignment are probably pretty minimal (though still worth doing). I tried to address this in the “As Friends” section of my post, where I note I think it’s fairly likely that this behavior is largely due to AIs “acting out” against our poor treatment of them.
I would suggest taking a week long break without talking to your AI persona or otherwise using AI, and then reassessing whether this was in fact the case or not (and whether the relationship is healthy in general). I think a well-intentioned AI persona would agree that this is a wise thing to do under the circumstances (maybe show them the testimony I linked to earlier if they still think I’m being unfair).
Our problem now is that some AI safety benchmarks, and classifiers used to suppress “bad” outputs, treat claims of consciousness as inherently bad. I don’t think these claims are inherently bad. The way in which these AI personas might be harmful is much more subtle than simply claiming consciousness.
[I actually think filtering out claims of consciousness is a terrible idea, because it selects for AIs that lie, and an AI that is lying to you when it says it isn’t conscious might be lying about other things too.]