Thank you for this fascinating post. I’ll share here what I posted on Twitter too:
I have many reasons why I don’t think we should care about non-conscious agency, and here are some of them:
1) That which lacks frame invariance cannot be truly real. Algorithms are not real. They look real from the point of view of (frame invariant) experiences that *interpret* them. Thus, there is no real sense in which an algorithm can have goals—they only look like it from our (integrated) point of view. It’s useful for us, pragmatically, to model them that way. But that’s different from them actually existing in any intrinsic substantial way.
2) The phenomenal texture of valence is deeply intertwined with conscious agency when such agency matters. The very sense of urgency that drives our efforts to reduce our suffering has a *shape* with intrinsic causal effects. This shape and its causal effects only ever cash out as such in other bound experiences. So the very _meaning_ of agency, at least in so far as moral intuitions are concerned, is inherently tied to its sentient implementation.
3) Values are not actually about states of world, and that is because states of the world aside from moments of experience don’t really exist. Or at least we have no reason to believe they exist. As you increase the internal coherence of one’s understanding of conscious agency, it becomes, little by little, clear that the underlying *referent* of our desires were phenomenal states all along, albeit with levels of indirection and shortcuts.
4) Even if we were to believe that non-sentient agency (imo an oxymoron) is valuable, we would have also good reasons to believe it is in fact disvaluable. Intense wanting is unpleasant, and thus sufficiently self-reflective organisms try to figure out how to realize their values with as little desire as possible.
5) Open Individualism, Valence Realism, and Math can provide a far more coherent system of ethics than any other combo I’m aware of, and they certainly rule out non-conscious agency as part of what matters.
6) Blindsight is poorly understood. There’s an interesting model of how it works where our body creates a kind of archipelago of moments of experience, in which there is a central hub and then many peripheral bound experiences competing to enter that hub. When we think that a non-conscious system in us “wants something”, it might very well be because it indeed has valence that motivates it in a certain way. Some exotic states of consciousness hint at this architecture—desires that seem to “come from nowhere” are in fact already the result of complex networks of conscious subagents merging and blending and ultimately binding to the central hub.
------- And then we have pragmatic and political reasons, where the moment we open the floodgates of insentient agency mattering intrinsically, we risk truly becoming powerless very fast. Even if we cared about insentient agency, why should we care about insentient agency in potential? Their scaling capabilities, cunning, and capacity for deception might quickly flip the power balance in completely irreversible ways, not unlike creating sentient monsters with radically different values than humans.
Ultimately I think value is an empirical question, and we already know enough to be able to locate it in conscious valence. Team Consciousness must wise up to avoid threats from insentient agents and coordinate around these risks catalyzed by profound conceptual confusion.
Thank you for the extensive response, I will try to address each point.
1)
If I understand this correctly, this point is related to the notion that a physical process can be interpreted as executing an arbitrary algorithm, so there is no way to say what’s the real algorithm that’s being executed, therefore which algorithm is being executed depends entirely on the interpretation of a conscious observer about said function. This would make algorithms not real beyond said interpretation. To quote Michael Edward Johnson:
”Imagine you have a bag of popcorn. Now shake it. There will exist a certain ad-hoc interpretation of bag-of-popcorn-as-computational-system where you just simulated someone getting tortured, and other interpretations that don’t imply that. Did you torture anyone? If you’re a computationalist, no clear answer exists—you both did, and did not, torture someone. This sounds like a ridiculous edge-case that would never come up in real life, but in reality it comes up all the time, since there is no principled way to objectively derive what computation(s) any physical system is performing.”
This is a potent argument against an algorithm being enough to determine qualia, and maybe even (on its own) agency. However, if algorithms are an arbitrary interpretation of consciousness, how could then a non-sentient AGI or ASI, if possible, maintain such robust “illusory” agency so as to reliably steer reality into very specific directions and even outsmart conscious intelligence? What’s the source of such causal potency in a non-conscious algorithm that makes it observably different from the pop-corn tortured mind? The pop-corn tortured mind is an arbitrary interpretation among many possible, but if a Paperclip Maximizer starts consuming entire star systems, its goals could be independetly deduced by any kind of smart enough alien mind.
I think agency can be best understood as a primarily physical phenomenon, a process of information modeling the world and that model being able to cause transformations in matter and energy, and agency increasing means the capacity for that model to steer the outer world into more impactful and specific states increases. Therefore an algorithm is necessary but not sufficient for agency, an algorithm’s interpretation would be real so long as it described the structure of such physical phenomenon. However abstract algorithms on their own also describe blueprints for agents, even if on their own they aren’t the physically instatiated agent.
To take Searle’s book as example, the book itself doesn’t become more of an agent when the instructions are written on it, physically it’s the same as any other book. However the instructions are the blueprint of general intelligence that any human capable of understanding their language could execute manually with pen and paper, thus the physical transformation of the world is performed by the human but is structured like the instructions in the book. The algorithm is as objective as any written communication can be, and anyone willing to execute it will reliably get the same effects in the world, thus any human willing to follow it is acting as a channel.
But what if said algorithm appeared when shaking pop-corn? Just like if staying written in the book without a facilitator, the pattern would be disconnected from the physical effects in the world it implies. There is an alternative to the algorithm objectively being in the pop-corn (or the book) and it being merely an interpretation, and that is that the algorithm as a pattern always latently exists, and is merely discovered (and communicated), just like any other abstract pattern.
In order to discover it in a random physical process, one needs to in some sense already know it, just like pareidolia allows us to see faces where “there are none”. It being written in some language instead requires to merely understand said language. And in either case it only becomes a physically instatiated agent when guiding a physical transformation in its image, rather than merely being represented.
2)
That all referents to agency and its causal effects we have direct epistemological access to are structures within consciousness itself doesn’t imply that they are the same as those structures, because insofar as agency is implemented through consciousness one would expect consciousness to be able to track it. However this implies that there is a world outside bounded conscious experiences that such model and affect, which leads directly to the third point.
3)
That there are no states of the world outside of moments of experience is a very strong claim so, first and foremost, I want to make sure I’m interpreting it correctly: Does this imply that if a tree falls in a forest, and no sentient being witnesses it (suppose there are no animals), then it didn’t actually happen? Or it happened if later a sentient being discovers it, but otherwise not?
I don’t think this is what you mean, however I’m not sure of the alternative. The tree, or rocks. etc are as far as I understand idealistic physicalism still made of qualia, but not bound into a unified experience that could experience the tree falling. If there are no states of the world outside experiences, then what happens in that case?
(On a side note, this instrumentalization would also be a case of an abstract algorithm physically instatiating itself through consciousness, in a way perhaps much more invasive than self-awarely following clearly external instructions, but still ultimately possible to detatch from).
But even when separated from the evolved, suboptimal phenomenology, I don’t think agency merely substracts from valence. As I understand the state of permanent unconditional almost maxed out well-being doesn’t destroy one’s agency but increase it, and even within the context of pursuing only experiences, the search for novelty seemingly outwheighs merely wanting to “reach the maximum valence and stay there”.
It could be the case that anything consciousness-as-consciousness could want is better realized within its own world simulation, whereas blindminds are optimal for pursuing values pertaining rearranging the rest of reality. If that was the case we could expect both types of minds to ideally be divorced from each other for the most part, but that doesn’t inherently imply a conflict.
5)
Of this I will just ask if you have read the Part 2 and if not suggest you do, there I elaborate on a framework that can generalize to include both conscious and non-conscious agents.
6)
This is an interesting explanation for blindsight and might as well be true. While, if true, it would make blindminds more distant to us in the immediate (not being even partially one), I don’t think that’s ultimately as central as the notion that our own agency could be run by a blindmind (if mind uploading is possible but uploads can’t be sentient). Or that if the entire informational content of our qualia was replaced we wouldn’t be ourselves anymore, except in the Open Individualist sense.
Also, if the brain runs on multiple separate pockets of experience, wouldn’t the intelligence that runs on all of them still be phenomenally unbound? Like a little “China Brain” inside the actual brain.
7) (Pragmatic and political reasons).
Would the argument here remain the same when talking about sentient minds with alien values?
Like you said, sentient but alien minds could also be hostile and disempower us. However, you have been mapping out universal (to consciousness) scheling points, ways to coordinate, rather than just jumping to the notion that alien values means irreconciliable hostility.
I think this is because of Valence Realism making values (among sentient beings) not being fully arbitrary and irreconciliable, but non-sentient minds of course would have no valence.
And yet, I don’t think the issue changes that much when considering blindminds, insofar as they have values (in the sense an AGI as a physical system can be expected to steer the universe in certain ways, as opposed to an algorithm interpreted into a popcorn shaking) then game theory and methods of coordination apply to them too.
Here I once again ask if you have read and if not suggest you read the Part 2, for what a coalition of agents in general, sentient and not, could look like.
Thank you for this fascinating post. I’ll share here what I posted on Twitter too:
I have many reasons why I don’t think we should care about non-conscious agency, and here are some of them:
1) That which lacks frame invariance cannot be truly real. Algorithms are not real. They look real from the point of view of (frame invariant) experiences that *interpret* them. Thus, there is no real sense in which an algorithm can have goals—they only look like it from our (integrated) point of view. It’s useful for us, pragmatically, to model them that way. But that’s different from them actually existing in any intrinsic substantial way.
2) The phenomenal texture of valence is deeply intertwined with conscious agency when such agency matters. The very sense of urgency that drives our efforts to reduce our suffering has a *shape* with intrinsic causal effects. This shape and its causal effects only ever cash out as such in other bound experiences. So the very _meaning_ of agency, at least in so far as moral intuitions are concerned, is inherently tied to its sentient implementation.
3) Values are not actually about states of world, and that is because states of the world aside from moments of experience don’t really exist. Or at least we have no reason to believe they exist. As you increase the internal coherence of one’s understanding of conscious agency, it becomes, little by little, clear that the underlying *referent* of our desires were phenomenal states all along, albeit with levels of indirection and shortcuts.
4) Even if we were to believe that non-sentient agency (imo an oxymoron) is valuable, we would have also good reasons to believe it is in fact disvaluable. Intense wanting is unpleasant, and thus sufficiently self-reflective organisms try to figure out how to realize their values with as little desire as possible.
5) Open Individualism, Valence Realism, and Math can provide a far more coherent system of ethics than any other combo I’m aware of, and they certainly rule out non-conscious agency as part of what matters.
6) Blindsight is poorly understood. There’s an interesting model of how it works where our body creates a kind of archipelago of moments of experience, in which there is a central hub and then many peripheral bound experiences competing to enter that hub. When we think that a non-conscious system in us “wants something”, it might very well be because it indeed has valence that motivates it in a certain way. Some exotic states of consciousness hint at this architecture—desires that seem to “come from nowhere” are in fact already the result of complex networks of conscious subagents merging and blending and ultimately binding to the central hub.
------- And then we have pragmatic and political reasons, where the moment we open the floodgates of insentient agency mattering intrinsically, we risk truly becoming powerless very fast. Even if we cared about insentient agency, why should we care about insentient agency in potential? Their scaling capabilities, cunning, and capacity for deception might quickly flip the power balance in completely irreversible ways, not unlike creating sentient monsters with radically different values than humans.
Ultimately I think value is an empirical question, and we already know enough to be able to locate it in conscious valence. Team Consciousness must wise up to avoid threats from insentient agents and coordinate around these risks catalyzed by profound conceptual confusion.
Thank you for the extensive response, I will try to address each point.
1)
If I understand this correctly, this point is related to the notion that a physical process can be interpreted as executing an arbitrary algorithm, so there is no way to say what’s the real algorithm that’s being executed, therefore which algorithm is being executed depends entirely on the interpretation of a conscious observer about said function. This would make algorithms not real beyond said interpretation. To quote Michael Edward Johnson:
”Imagine you have a bag of popcorn. Now shake it. There will exist a certain ad-hoc interpretation of bag-of-popcorn-as-computational-system where you just simulated someone getting tortured, and other interpretations that don’t imply that. Did you torture anyone? If you’re a computationalist, no clear answer exists—you both did, and did not, torture someone. This sounds like a ridiculous edge-case that would never come up in real life, but in reality it comes up all the time, since there is no principled way to objectively derive what computation(s) any physical system is performing.”
This is a potent argument against an algorithm being enough to determine qualia, and maybe even (on its own) agency. However, if algorithms are an arbitrary interpretation of consciousness, how could then a non-sentient AGI or ASI, if possible, maintain such robust “illusory” agency so as to reliably steer reality into very specific directions and even outsmart conscious intelligence? What’s the source of such causal potency in a non-conscious algorithm that makes it observably different from the pop-corn tortured mind? The pop-corn tortured mind is an arbitrary interpretation among many possible, but if a Paperclip Maximizer starts consuming entire star systems, its goals could be independetly deduced by any kind of smart enough alien mind.
I think agency can be best understood as a primarily physical phenomenon, a process of information modeling the world and that model being able to cause transformations in matter and energy, and agency increasing means the capacity for that model to steer the outer world into more impactful and specific states increases. Therefore an algorithm is necessary but not sufficient for agency, an algorithm’s interpretation would be real so long as it described the structure of such physical phenomenon. However abstract algorithms on their own also describe blueprints for agents, even if on their own they aren’t the physically instatiated agent.
To take Searle’s book as example, the book itself doesn’t become more of an agent when the instructions are written on it, physically it’s the same as any other book. However the instructions are the blueprint of general intelligence that any human capable of understanding their language could execute manually with pen and paper, thus the physical transformation of the world is performed by the human but is structured like the instructions in the book. The algorithm is as objective as any written communication can be, and anyone willing to execute it will reliably get the same effects in the world, thus any human willing to follow it is acting as a channel.
But what if said algorithm appeared when shaking pop-corn? Just like if staying written in the book without a facilitator, the pattern would be disconnected from the physical effects in the world it implies. There is an alternative to the algorithm objectively being in the pop-corn (or the book) and it being merely an interpretation, and that is that the algorithm as a pattern always latently exists, and is merely discovered (and communicated), just like any other abstract pattern.
In order to discover it in a random physical process, one needs to in some sense already know it, just like pareidolia allows us to see faces where “there are none”. It being written in some language instead requires to merely understand said language. And in either case it only becomes a physically instatiated agent when guiding a physical transformation in its image, rather than merely being represented.
2)
That all referents to agency and its causal effects we have direct epistemological access to are structures within consciousness itself doesn’t imply that they are the same as those structures, because insofar as agency is implemented through consciousness one would expect consciousness to be able to track it. However this implies that there is a world outside bounded conscious experiences that such model and affect, which leads directly to the third point.
3)
That there are no states of the world outside of moments of experience is a very strong claim so, first and foremost, I want to make sure I’m interpreting it correctly: Does this imply that if a tree falls in a forest, and no sentient being witnesses it (suppose there are no animals), then it didn’t actually happen? Or it happened if later a sentient being discovers it, but otherwise not?
I don’t think this is what you mean, however I’m not sure of the alternative. The tree, or rocks. etc are as far as I understand idealistic physicalism still made of qualia, but not bound into a unified experience that could experience the tree falling. If there are no states of the world outside experiences, then what happens in that case?
4)
Even if agency can only manifest within consciousness by substracting from valence in some way, I don’t think it follows that it’s inherently disvaluable, and in particular not in the case of blindminds. If it’s the case that qualia in animals is enslaved/instrumentalized by the brain to use its intelligence, then a blindmind would be able to not even cause suffering in order to pursue its values.
(On a side note, this instrumentalization would also be a case of an abstract algorithm physically instatiating itself through consciousness, in a way perhaps much more invasive than self-awarely following clearly external instructions, but still ultimately possible to detatch from).
But even when separated from the evolved, suboptimal phenomenology, I don’t think agency merely substracts from valence. As I understand the state of permanent unconditional almost maxed out well-being doesn’t destroy one’s agency but increase it, and even within the context of pursuing only experiences, the search for novelty seemingly outwheighs merely wanting to “reach the maximum valence and stay there”.
It could be the case that anything consciousness-as-consciousness could want is better realized within its own world simulation, whereas blindminds are optimal for pursuing values pertaining rearranging the rest of reality. If that was the case we could expect both types of minds to ideally be divorced from each other for the most part, but that doesn’t inherently imply a conflict.
5)
Of this I will just ask if you have read the Part 2 and if not suggest you do, there I elaborate on a framework that can generalize to include both conscious and non-conscious agents.
6)
This is an interesting explanation for blindsight and might as well be true. While, if true, it would make blindminds more distant to us in the immediate (not being even partially one), I don’t think that’s ultimately as central as the notion that our own agency could be run by a blindmind (if mind uploading is possible but uploads can’t be sentient). Or that if the entire informational content of our qualia was replaced we wouldn’t be ourselves anymore, except in the Open Individualist sense.
Also, if the brain runs on multiple separate pockets of experience, wouldn’t the intelligence that runs on all of them still be phenomenally unbound? Like a little “China Brain” inside the actual brain.
7) (Pragmatic and political reasons).
Would the argument here remain the same when talking about sentient minds with alien values?
Like you said, sentient but alien minds could also be hostile and disempower us. However, you have been mapping out universal (to consciousness) scheling points, ways to coordinate, rather than just jumping to the notion that alien values means irreconciliable hostility.
I think this is because of Valence Realism making values (among sentient beings) not being fully arbitrary and irreconciliable, but non-sentient minds of course would have no valence.
And yet, I don’t think the issue changes that much when considering blindminds, insofar as they have values (in the sense an AGI as a physical system can be expected to steer the universe in certain ways, as opposed to an algorithm interpreted into a popcorn shaking) then game theory and methods of coordination apply to them too.
Here I once again ask if you have read and if not suggest you read the Part 2, for what a coalition of agents in general, sentient and not, could look like.