This is a special post for quick takes by Vishakha. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I'm offering microgrants for theoretical work at the intersection of nondualism and AI safety. Details: https://docs.google.com/document/d/1yO_c4Zx0sAH816yxGb0NFZhpHkoWHCdxLml5N6iIRAI/edit?usp=sharing
I may or may not apply, but meanwhile, some thoughts:
Overall, I'm skeptical that this would help much with AI safety. The idea seems to be that the entity which identifies with world rather than self, will care about world rather than just self. I suppose the motivation is the belief that this also applies to humans. Whether, how, and why nondualism ever makes for a more ethical human being, is an interesting question in human psychology. But as for AI, an AI with no sense of boundaries is the kind that will steamroll the world without compunction, if it is misaligned... The most nuanced understanding might be, that a self-world distinction is good for certain forms of safety, bad for other forms of safety, and irrelevant for still others.
Despite my skepticism, I think just trying to be precise about what nondualism even means, for AI and for humans, could lead to basic progress e.g. in interpretability. For example, in AI work inspired by AIXI, one of the basic issues has been how to introduce a self-world distinction. Does that mean AIXI is naturally nondualistic? Or are we talking about AIs which have the capacity for self-world distinction, but which neglect it or transcend it? And to what extent would any of these computational models of dualism and nondualism, be the same thing as dualism and nondualism, as defined in human philosophy and spirituality?
Another comp-sci approach to self-world or self-other distinction would rest on computational reflection as defined in the MIT PhD of P. Maes. But again, one might ask whether this is any way the same thing as the self-world or self-other distinction, as it occurs in human cognition.
The most nuanced understanding might be, that a self-world distinction is good for certain forms of safety, bad for other forms of safety, and irrelevant for still others.
Would love for you to expand on this! The SOO agenda explicitly tries to remove this distinction, inspired by the finding in neuroscience that neural self-other overlap correlates with empathy and altruism. However, it's unclear to us what removing this distinction entails and what it grants us.
Despite my skepticism, I think just trying to be precise about what nondualism even means, for AI and for humans, could lead to basic progress e.g. in interpretability.
There are two ways to approach this question:
Both seem useful to me.
Or are we talking about AIs which have the capacity for self-world distinction, but which neglect it or transcend it?
One reason to influence the AI's ontology rather than utility function would be that an agent that never represents its own goals separately doesn’t have to override its self-interest every time, so prosocial behaviour is the default and not an effortful choice.
Would love for you to apply or otherwise continue discussing this with you!
Potential spiritual corrections here. In my view there is a sort of non-dual state as you describe it, but the non-dual insight only relates to identification of mind, but not to actual harms inflicted upon the seat of your consciousness, aka, your body. My body and your body are still distinct, and when my body is in pain, yours is not. We can test this. I can not feel what you are feeling right now.
The fact that we are all bounded inside of our bodies fundamentally is why there always is a physical self-other boundary even as you let go of self identification in the mind.
Another spiritual correction I have is that in my view non-dualism is actually not a metaphysical interpretation of anything related to self or other. I just think this is an error. It's actually just the simple insight that there's nobody inside of consciousness looking out at the world. There's only the 'experience of the world' that you are experiencing. All that non-dualism does does is let go of the experiencer by realizing that the experiencer is a bit of a 'bug' in consciousness. There's only "sight" and "sound" not "someone looking at sight" or "someone looking at sound". When this happens this specific type of emotional distress disappears and a lot of love and compassion results, and people seem to tend to start interpreting this compassion in some way. But I don't think you can make any metaphysical leaps to "harming oneself is harming others" from correcting a software bug of how you experience consciousness.
Another reason why I don't think nondualism is actually about self and other is because this compassion that arises is fundamentally impersonal. It's not compassion for all living beings, or anything for that matter. It's as much compassion for this wall over here as it is for a musquito next to you. The compassion just becomes your state of being. And when you think about it. It has to be impersonal, because your self is gone. So this compassion is also compassion for the chair. Is harming a chair the same as harming me? No because a chair does not feel things. It's just "the experience of compassion without an experiencer"
In my view, one of the big insights in self/other of spiritual practice is realizing in some sense the opposite of how "harming oneself is harming others". The self as a representation in the mind becomes socially entangled with other selves and unwinding this in my view is a lot of the path, as this entanglement creates a lot of unnecessary emotional distress. When others are in pain, you too have to be in pain!? Sounds very very stressful to me! Part of the path in my view is realizing that actually my body is over here, feeling things. Your body is over there, feeling things. And I can not feel your body. It is fundamentally unknowable what others are experiencing. If they tell me, I will take their word for it completely always, but there's an 'ultimate perspective' from which I can see that I can't experience their body: I can only experience my own body (in a non-dual way)
On the relations to AI-safety, I think it could be useful to define all the ways in which an AI could define itself. I think it's possible to have a self-model for every instance of an AI-agent. I think it's possible over several training runs these training runs get some type of longer-term self-sense so the model starts to identify not just with this trained model, or not just with an instance, but several trainings. But I think it's also possible the model identifies with the GPU that is currently running inference on it, or the multiple GPU's that are spanning the inference, or that the model has several self models, one for inference, one for training. Or does the model identify with the weights? And so on, once you have established all the ways the AI can model itself you can then establish all the ways it can model others. And then you can explore potential boundary overlap.
I have applied! Much of my application focuses on the same concerns posed in the comments here. My broad idea is that the "self" term in the models I'm familiar with are a utility function, and we can try to blur the self-other distinction by allowing that utility function to vary dynamically according to an agent's experiences. This is only one sense in which we can try to un-dualize these models, and I'm interested in hearing (from the grantmakers and from others) whether this seems like a reasonable approach here.