cube_flipper of smoothbrains.net recently made something resembling the following argument in a talk. I like the argument because it uses tools of computationalism to argue against computationalism: it argues within the Solomonoff Induction framework, against the computationalist position on phenomenal consciousness.
This is my own interpretation of the argument; if you don’t like something about it, please don’t blame cube_flipper.
First, I should explain what is at stake in this argument.
The computationalist position views the brain as a computer, and claims that “what it is like” to be a brain (or anything else) depends entirely on what computations are implemented by that brain (or other device). This implies, for example, that if you could simulate a brain perfectly on a digital computer, then it would have exactly the same experiences as its non-digital analogue.
The physicalist position instead identifies conscious experience as some specific (but relatively simple) physical phenomena in the brain, such as the activity of the EM field or some quantum phenomena. Here’s cube_flipper’s write-up of the EM theory in particular.
The argument considered here argues in favor of physicalist theories, but does so from premises which I think will seem plausible to many computationalists.
Premise 1: Solomonoff Induction (SI) is a good normative theory of epistemic rationality; IE, if SI would believe something, so should we.
Premise 2: In order to make predictions about observations, we should (in practice, for our universe) make a model which consists of three parts:
Premise 3: Computationalist phenomenal bridges are complex relative to physicalist phenomenal bridges.[1]
Premise 4: Physical phenomenal bridges are at least as compatible with the data of experience as computationalist phenomenal bridges.
Premise 5: Our theory of phenomenal consciousness should be identified with our theory of phenomenal bridges.
Conclusion: We should prefer physicalist theories of consciousness to computationalist theories of consciousness.
Argument:
We should judge theories of consciousness in the same way that we judge theories of physics, IE, by balancing predictive accuracy with simplicity of the theory, as stipulated by SI.
When we do this, we come up with theories which describe physical reality (EG the Schrodinger Equation), plus theories which read off experiences from physical reality (EG neural correlates of consciousness), plus some probability distribution over who we are (EG the Born rule).
It might possibly be that computationalist theories are about equally good at fitting with the data of experience, but they aren’t better, at least not with respect to observations so far.
However, computational ways of building phenomenal bridges are going to be very complex compared to physicalist ways of doing so. Therefore, we should prefer the physicalist theories
I have difficulty accepting the conclusion, because my thinking on phenomenal consciousness frames it as a problem of mapping between our first-person perspectives (our direct experience) and a shared third-person perspective (the shared understanding of an objective world). A high-quality upload of a human could be a participant in that sense-making process, which suggests to me that a theory of phenomenal consciousness should be invariant to (some sort of) computational equivalence.
The assumptions run deep, so they are difficult to question. It feels clear to me, for example, that agentic desire is inherent to pain and suffering (there has to be a thing that wants not-that for it to count as pain/suffering). The physicalist view refuses to depend on such things.
Nonetheless, it is important to question one’s assumptions. So, what do I think of the premises of the arguments?
I don’t believe that SI is the correct normative theory precisely, but it does seem hard to get away from something roughly like SI. I prefer Garrabrant Induction as a normative theory (IE, I think it is closer to normative for us, since we are computationally bounded agents). This does have a somewhat different character (more inclined to invoke specialist theories for specific topics, for example, rather than one big unified theory).
Nonetheless, it seems difficult to deny the part of SI which is important for this argument, namely that simpler theories should be preferred. We can quibble over which notion of simplicity (in SI terms, which universal machine to use), and we can worry about malign prior arguments. I'd certainly prefer to have some stronger reason for a position. Still, simplicity is a very important heuristic.
One might wish to argue that simplicity is somehow being misapplied here. Perhaps physical theories should be simple, but anthropic theories (theories determining the probability that you are you as opposed to someone else) need not be? (EG, because there is no need for anthropic theories in one’s epistemology?) I think the grounding in a common theory of rationality (SI) addresses this complaint, however: differing opinions can either show how it does not work out as described within SI, or state their disagreement with SI (and ideally give an alternate rationality framework that escapes the argument).
The argument for premise 2 is an empirical one: it seems hard to credibly model the world in other ways than this.
I will divide the discussion of premise 2 into two parts:
I think this is essentially a question of physical reductionism. (Not the physicalist theory of consciousness we've been discussing; rather, physical reductionism being the claim that all things can be defined in terms of physical things.) One might argue: SI will split up the world into physics plus self-locator because physics is all there is, and SI is smart, so SI will figure out that physics is all there is.
I don’t think this is necessarily true, even if one believes physical reductionism. SI might squish things together for reasons of compression. The theories which compress most effectively are not necessarily easily interpreted into comprehensible parts.
However, this isn't an objection to the spirit of the claim, I think.
We might more plausibly invoke condensation, or some other theory which similarly does a better job of representing the normative pressure to separate out concepts in a comprehensible way. The spirit of the claim is that the physical world would pop out in such a theory. SI might not cleanly separate its physical hypothesis from its phenomenal hypothesis (the bridge) and anthropic uncertainty (the pointer), but if you picked apart its code, that division might still be what's going on in some sense.
So, as with premise 1, I think we can quibble about the theory of rationality, but doing so doesn’t undermine the plausibility of the position being expressed. I suspect many readers will find it very plausible that a superintelligence will model the world in this way.
One possible objection is that modeling physics and then finding ourselves in it is not a practical way to model the world for computationally bounded agents such as ourselves. It might be normative for agents with unbounded computing power, but we are not such agents.
I think physical reductionist will not be dissuaded by this argument. Non-physical-reductionists, on the other hand, probably should not buy premise 2a.
SI only cares about predicting observations well. We can accept 2a, yet only divide our world-model into two parts: the physical world calculation, plus the function which looks at the world and tells us our observation.
2b postulates that we can split the question into our location plus the function that turns a location into observations.
This bakes in two important ideas: that we have a location, and that other locations can also contain experiences (not directly accessible to us).
There’s a confusing question of what format the locations should be given in. Are they like single points? Are they areas? This doesn’t seem like a big obstacle to me, though.
It does feel significant to assume that the function which decodes our consciousness out of physics can be meaningfully applied to other locations. It is like a physical law: translation-invariant. I'm told people interested in physicalist theories spend time thinking about what theories fit with physical invariances such as frame-invariance.
2b seems closely related to premise 4: one might argue that the phenomenal bridge should be a function of location because one wants to use it as a general theory of phenomenal consciousness. In other words, this is another argument from realism, but this time realism about consciousness: “I do think other people’s experiences are real, not just my own, therefore I think ideal rational reasoning would recognize that.”
Both 2a and 2b can also be framed in terms of what you want a theory of phenomenal consciousness to do for you. The point is to bridge between physics and qualia! The point is to tell us about the consciousness of other people!
Why should we expect computationalist bridges to be more complex than physicalist bridges?
A physicalist bridge needs to be able to pick out some physical phenomenon, such as patterns in the EM field.
A computational bridge needs to do that as well, to parse the physical model, but it also needs to contain the complexity of several layers of interpretation. For example, to interpret computers, we need to interpret floating-point arithmetic. This extra complexity penalizes the hypothesis.
You might want to argue that floating-point arithmetic isn’t really so complex. Ok. But computationalism supports arbitrary layers: logic gates are implemented out of electrical components; machine code is implemented out of logic gates; low-level languages are implemented out of machine code; high-level languages are implemented out of low-level languages; application programs are implemented from low- or high-level languages.
The computationalist theory of phenomenal consciousness doesn’t care about how many implementation layers are stacked on top of each other. There’s not supposed to be a penalty for that. Yet, if we have to represent each of those layers in the phenomenal bridge, description-length piles up.
If you deny premise 3, you must either think that physical phenomenal bridges are going to be very complex, or there is a simple specification of the computationalist phenomenal bridge. For example, I have a strong intuition that there is a simple theory of implementation (the question of whether a physical structure implements a computation). If true, then computationalist phenomenal bridges become easy to compute.
However, as far as I know, implementation is still an unsolved technical problem, even though I personally feel like there should be a solution. If there isn't, we shouldn't even believe computationalism for computers.[2] It seems quite possible that there is a simple notion of implementation, but that the definition is not computable (and therefore infinitely complex by the standards of SI).
Without a specific proposal, I am just flatly denying premise 3, which suggests I should be open to the alternative.
(Even with a specific proposal, one still has to believe that it is shorter than physicalist phenomenal bridges, which is hard to believe when we’re assuming physics is the basic data fed into the function. Otherwise premise 3 still holds.)
Differences in how well a theory of consciousness accounts for the data of experience can easily overpower differences in prior probability, so if computationalist theories had an edge in that respect, the debate about description length could easily be irrelevant.
A computationalist might argue: “It is easy to construct a counterexample. If we modify the brain in a way that keeps it computationally equivalent (typically easy to find in the computational realm) but which modifies the prediction of the physicalist theory, then of course the person will experience no change in their experience, disproving the physicalist theory.”
If I understand correctly, cube_flipper welcomes such an experiment (save for the fact that it seems far beyond our current technology), and anticipates having a different experience due to the modified physical field. I make the opposite prediction, myself, siding with the computationalist. However, I am not sure this would be as good a test as it sounds. A physicalist, if I understand correctly, could consistently claim that such an experiment is deluding the subject, essentially doing something like modifying the memory of the experience so that they inaccurately feel the same, when in fact there was a difference.
I do think it is true that we can be in better or worse positions to report our own experiences accurately, and modifying someone’s brain may put them in a worse position. (Still, the test seems important. Fitting people’s self-reports of conscious experience is an important feature of a theory of phenomenal consciousness, even if we must concede that there should be some failure cases.)
An earlier version of this assumption was “If it is rational for each of us individually to think our personal qualia work a specific way, then we should also think each other’s work that way.”
Granting premises 1-4, a conscious being should believe in a physicalist phenomenal bridge. You should anticipate things such as: changes to the EM field should change your conscious experience, even in cases where computational theories would not anticipate this.
Without premise 5, this doesn’t yet extend to other people. The fact that the phenomenal bridge can look at other people (premise 2b) doesn’t yet mean that it should. This is an anti-solipsism assumption, as well as an assumption against theories such as “physicalism for me, but computationalism for everyone else”.
I have nothing against premise 5. I think there might be more lurking under the surface if we were to dig into this premise more, but I don’t have much to say about it now.
I think this is a fun argument. For mental computationalists who are also physical reductionists, it highlights the importance of solving the implementation problem; without a simple theory of implementation, premise 3 seems difficult to deny. For those who find themselves in an uncomfortable position of believing the assumptions but denying the conclusion, remember that computationalism and physicalism are not the only two options when it comes to theories of consciousness.
cube_flipper also claims that computationalist theories have some related defects which don’t quite fit into the argument I’m articulating.
The computationalist picture isn’t just complex; it is also arbitrary, meaning it has a lot of degrees of freedom. For example, do we interpret a specific signal as a 1 or a 0?
When it comes to our own personal phenomenal bridges, of course, these choices are constrained by experience. More precisely: in order to apply the SI analogy, we have to choose some way of encoding our experience into bits. This then constrains all the arbitrary choices, since we are reasoning about our rationality as if we’re experiencing bits. This feature is shared with the physicalist theories, since we’re encoding everything into bits in order to apply SI as a theory of rationality.
However, when it comes to other people, 1 could just as well be 0, amongst other arbitrary mappings (110 could become 000, etc). The computationalist picture is committed to a sort of invariance, where the same abstract computation should have the same experience. Yet, if qualia are modeled via binary encodings (the SI picture), there’s got to be arbitrary choices here, which means you don’t necessarily trust other people’s reports of their qualia (they could be swapping 1 and 0 compared to your interpretation, or other such mappings).
This arbitrariness implies a lack of predictive power. This is the old “Is my red the same as your red?” problem.
Physicalist theories are (in my limited understanding) supposed to avoid this problem, because unlike computationalist theories, they’re happy to make specific claims like thus-and-sech configuration of the EM field corresponding universally to the experience of red.
This pithy observation is due to Mahmoud Ghanem.