I first provide resistance to Seth’s push towards weak and strong biological naturalism. Seth appeals to electromagnetic field theories here, supported by suggestions that anaesthetics disrupt how mitochondria generate the electromagnetic fields that may systematically influence neuronal activity; he also here alludes to findings of self-organising agential properties of biological material. These theories and findings are intriguing, however they are not yet sufficiently specific to consciousness: electromagnetic fields and cell assemblages might be more generically involved in maintaining life, agency, or other functions upstream or downstream of consciousness. Additional ordinary consciousness science is required before we should consider them specifically informative about consciousness, and thus for pushing towards weak/strong biological naturalism.
If I am right that it is premature to accept weak/strong biological naturalism, then our best option – while still doing justice to Seth’s overall line of reasoning against current computational functionalism – is to stick to the level of relatively substrate-dependent yet computational biological naturalism. If we do this, then I see mortal computation as most promising.
When discussing mortal computation, Seth approvingly invokes the following type of argument against conscious AI: Since immortal computation is extremely energy hungry, energy-efficient brains cannot be doing immortal computation. Therefore, the case for conscious (immortal) AI is undermined.
The thought behind this argument is that we have ample evidence that brains sustain consciousness, because we can extrapolate from our own conscious experience and from our knowledge of neuroscience and neurology. That extrapolation does not extend to current AI, because immortal computation makes AI decidedly different from mortal, neural computation. Hence, in the absence of other positive reasons to believe current AI is conscious, we should be agnostic about conscious AI.
The argument is powerful but notice that it rests on the assumption that the difference between mortal and immortal computation is relevant to consciousness. This assumption should be substantiated, because it might just be that consciousness depends on something else than computation, whether mortal or immortal – the fact that we are conscious and that our brains are mortal computers could be a coincidence, after all. For example, my scepticism notwithstanding, non-computational electromagnetic field theories of consciousness might in the fullness of time win the day, making distinctions in computation otiose.
Responding to this challenge requires looking at the functions of mortal and immortal computation – is there anything about the function of either that suggests relevance for consciousness? Raising the gaze from substrate to function thus emerges as an important task for Seth’s overall biological naturalism; his case will be strengthened if it speaks to consciousness with more specificity.
Seth appeals convincingly to the free energy principle for the idea that substrate-dependent self-evidencing matters for consciousness, because it underpins multimodal survival-relevant integration and valenced allostatic regulation. This criterion is consistent with the claim that the difference between mortal and immortal computation provides a watershed for consciousness. But, as a criterion for consciousness, it is still too broad since it threatens to imply that too many things on the mortal side of the watershed could be conscious, such as simple allostatic, sensing organisms, or a group of people with a division of cognitive labour. To begin to narrow down the criteria, consider first what may be distinctive of mortal computing function.
Immortal computation rests on the assumption that “the only way to get a general purpose computer to perform a specific task was to write a program that specified exactly what to do in excruciating detail” (Hinton, Reference Hinton2022:13). This program complexity ensures error-free function but, in addition to consuming much energy, comes at the cost of rigidity and inability to generalise to new contexts.
Mortal computation, in contrast, is less concerned with avoiding errors and more concerned with being flexible and resilient in the face of uncertainty and volatility. It is not the energy consumption or the substrate-dependence per se that is distinctive, but this difference in functionality: mortal computers can solve tasks in uncertain, changing environments (see also Ororbia & Friston, Reference Ororbia and Friston2023).
The key question then is if any specific types of mortal computation – ways of operating close to the edge of uncertainty and change – evoke key characteristics of consciousness?
I think we can begin to discern key characteristics of consciousness in some kinds of mortal computation, if we start from our own undoubtedly conscious, self-supervised ways of self-evidencing in an uncertain world (rather than starting from life, survival, or sensing). Our consciousness is characterised by an ability to detach from sensory input (in dreaming, mind-wandering and counterfactual deliberation) and to selectively re-attach (in sustained perceptual attention); and our conscious experience is infused with our ability to (actually and counterfactually) model ourselves, optimistically, as relatively persistent agents selectively sampling observations in an uncertain and volatile world. This is consciousness-relevant functionality for which mortal computation is well-suited. It allows not just existence in an uncertain world, but self-aware existence that ‘exploits’ volatility and uncertainty (for more along these lines, see (Hohwy, Reference HohwyIn press); for active inference modelling of consciousness-specific phenomena, see (Whyte et al., Reference Whyte, Corcoran, Robinson, Smith, Moran, Parr, Friston, Seth and HohwyIn press).
I think Seth’s case for biological naturalism would be bolstered by some considerations in this vein, bringing the functioning of substrate-dependent systems closer to consciousness. But my hunch is that this is most promising for mortal computational biological naturalism, rather than the weak/strong versions.
Since computationalism is then not unequivocally off the table, this would (as Seth notes) leave the door open for future conscious AI. Concretely, we could begin by building artificial systems not by copying parameter values from one system to another but rather by ‘educating’ systems to distil information (cf. Hinton, Reference Hinton2022). Then we could perhaps get to self-supervised systems whose mortal self-evidencing is optimistically ‘existence-dependent’, rather than strictly substrate-dependent (cf. Fisher & Hohwy, Reference Fisher and Hohwy2024). Combining this with active inference under the free energy principle, the substrate of mortal computation comes out as precision control of the causal dependencies of nodes within a Markov blanket, working as a generative model. Artificial agents could potentially have such inner structure, in ways specific to consciousness.
You have
Access
I first provide resistance to Seth’s push towards weak and strong biological naturalism. Seth appeals to electromagnetic field theories here, supported by suggestions that anaesthetics disrupt how mitochondria generate the electromagnetic fields that may systematically influence neuronal activity; he also here alludes to findings of self-organising agential properties of biological material. These theories and findings are intriguing, however they are not yet sufficiently specific to consciousness: electromagnetic fields and cell assemblages might be more generically involved in maintaining life, agency, or other functions upstream or downstream of consciousness. Additional ordinary consciousness science is required before we should consider them specifically informative about consciousness, and thus for pushing towards weak/strong biological naturalism.
If I am right that it is premature to accept weak/strong biological naturalism, then our best option – while still doing justice to Seth’s overall line of reasoning against current computational functionalism – is to stick to the level of relatively substrate-dependent yet computational biological naturalism. If we do this, then I see mortal computation as most promising.
When discussing mortal computation, Seth approvingly invokes the following type of argument against conscious AI: Since immortal computation is extremely energy hungry, energy-efficient brains cannot be doing immortal computation. Therefore, the case for conscious (immortal) AI is undermined.
The thought behind this argument is that we have ample evidence that brains sustain consciousness, because we can extrapolate from our own conscious experience and from our knowledge of neuroscience and neurology. That extrapolation does not extend to current AI, because immortal computation makes AI decidedly different from mortal, neural computation. Hence, in the absence of other positive reasons to believe current AI is conscious, we should be agnostic about conscious AI.
The argument is powerful but notice that it rests on the assumption that the difference between mortal and immortal computation is relevant to consciousness. This assumption should be substantiated, because it might just be that consciousness depends on something else than computation, whether mortal or immortal – the fact that we are conscious and that our brains are mortal computers could be a coincidence, after all. For example, my scepticism notwithstanding, non-computational electromagnetic field theories of consciousness might in the fullness of time win the day, making distinctions in computation otiose.
Responding to this challenge requires looking at the functions of mortal and immortal computation – is there anything about the function of either that suggests relevance for consciousness? Raising the gaze from substrate to function thus emerges as an important task for Seth’s overall biological naturalism; his case will be strengthened if it speaks to consciousness with more specificity.
Seth appeals convincingly to the free energy principle for the idea that substrate-dependent self-evidencing matters for consciousness, because it underpins multimodal survival-relevant integration and valenced allostatic regulation. This criterion is consistent with the claim that the difference between mortal and immortal computation provides a watershed for consciousness. But, as a criterion for consciousness, it is still too broad since it threatens to imply that too many things on the mortal side of the watershed could be conscious, such as simple allostatic, sensing organisms, or a group of people with a division of cognitive labour. To begin to narrow down the criteria, consider first what may be distinctive of mortal computing function.
Immortal computation rests on the assumption that “the only way to get a general purpose computer to perform a specific task was to write a program that specified exactly what to do in excruciating detail” (Hinton, Reference Hinton2022:13). This program complexity ensures error-free function but, in addition to consuming much energy, comes at the cost of rigidity and inability to generalise to new contexts.
Mortal computation, in contrast, is less concerned with avoiding errors and more concerned with being flexible and resilient in the face of uncertainty and volatility. It is not the energy consumption or the substrate-dependence per se that is distinctive, but this difference in functionality: mortal computers can solve tasks in uncertain, changing environments (see also Ororbia & Friston, Reference Ororbia and Friston2023).
The key question then is if any specific types of mortal computation – ways of operating close to the edge of uncertainty and change – evoke key characteristics of consciousness?
I think we can begin to discern key characteristics of consciousness in some kinds of mortal computation, if we start from our own undoubtedly conscious, self-supervised ways of self-evidencing in an uncertain world (rather than starting from life, survival, or sensing). Our consciousness is characterised by an ability to detach from sensory input (in dreaming, mind-wandering and counterfactual deliberation) and to selectively re-attach (in sustained perceptual attention); and our conscious experience is infused with our ability to (actually and counterfactually) model ourselves, optimistically, as relatively persistent agents selectively sampling observations in an uncertain and volatile world. This is consciousness-relevant functionality for which mortal computation is well-suited. It allows not just existence in an uncertain world, but self-aware existence that ‘exploits’ volatility and uncertainty (for more along these lines, see (Hohwy, Reference HohwyIn press); for active inference modelling of consciousness-specific phenomena, see (Whyte et al., Reference Whyte, Corcoran, Robinson, Smith, Moran, Parr, Friston, Seth and HohwyIn press).
I think Seth’s case for biological naturalism would be bolstered by some considerations in this vein, bringing the functioning of substrate-dependent systems closer to consciousness. But my hunch is that this is most promising for mortal computational biological naturalism, rather than the weak/strong versions.
Since computationalism is then not unequivocally off the table, this would (as Seth notes) leave the door open for future conscious AI. Concretely, we could begin by building artificial systems not by copying parameter values from one system to another but rather by ‘educating’ systems to distil information (cf. Hinton, Reference Hinton2022). Then we could perhaps get to self-supervised systems whose mortal self-evidencing is optimistically ‘existence-dependent’, rather than strictly substrate-dependent (cf. Fisher & Hohwy, Reference Fisher and Hohwy2024). Combining this with active inference under the free energy principle, the substrate of mortal computation comes out as precision control of the causal dependencies of nodes within a Markov blanket, working as a generative model. Artificial agents could potentially have such inner structure, in ways specific to consciousness.
Acknowledgements
Thank you to Tim Bayne for discussion.
Financial support
JH receives funding from the Three Springs Foundation.
Competing interests
None.