Dreaming of Utility

The causal and anatomical structure of the brain-body connection, its exploitability, the instrumental origins of causal articulation in interactive systems, and hints towards a theory of AI directability via manipulating internal causal structure.

Much of the motivation behind these ideas comes from my work on Worldspace, a cognitive framework which, among other things, functionalizes the trait (resp. faculty) of intelligence as an !c{arbitrary system's}{humans, AIs, civilizations, plants} power of (resp. approach to) optimization in some given environment. If Worldspace is analogized to medical diagnosis and treatment, this is etiology and physiology: not "how does optimizing systems look, how do they behave" but "how do optimizing systems arise, how are they structured". How is worldspace modeled by any system at all? What are the necessary features of a model that looks like reality? What are the worlds modeled by dreams? The neuroscience is just a fun case study to sketch out a general theory of causal articulation, and the notion of sensorimotor bounding as a mechanism for AI directability is an example of what ideas such a theory allows us to come up with.

The Mind's Fortress

The brain is one of the most causally isolated structures in the human body. It's got a hard outer casing (the cranium), three layers of protective soft tissue under that (meninges), rests in a shock-absorbing fluid that passively supports homeostasis (CSF), and all blood entering it is stringently filtered by the blood-brain barrier.

Why so isolated? You need a stomach to live just the same, and yet the stomach is out fighting bacteria, chemicals, and physical trauma on the front lines while the brain hides in its fortress. But consider how damage affects each:

More than isolation, though, the paths of causality to and from the brain are tightly controlled. Ignoring the fluid tractsLike cerebrospinal fluid traveling between the brain and spinal cord. As usual with biological systems, there are tons of small details like this which blur any clear boundary you might want to make. Nevertheless, consider: Individual humans hike through mountains all the time, but to a marching army, a mountain range is essentially an impenetrable barrier. You can say that any particular reification of this range as a line on a map is just a fiction, and even be more or less correct, but this won't let an army march any faster: there will always be a functional bound-ing which differentially catalyzes the formation of some fictions over others. for brevity, all causal pathways to and from the brain are organized into thirteen cranial nerves (for each hemisphere) and a spinal cord, whose connection to the medulla oblongata at the bottom of the brainstem is no wider than a dime.

Experiencing Isolation

The brain is the body's causal articulation organ; it's one of the only systems that can magnify subtle input variations (lights on a screen, spoken words) into tremendous output variations in an adaptive and non-chaotic manner. This makes it all the more vital that the body controls unwanted causal influences, influences not in the brain's "language" of neural signals. But the brain has its own methods of isolating itself as well.

Once a day, it goes to sleep, entering a state where all incoming sensory information is greatly attenuated, and outgoing motor commands that aren't autonomic (breathing, heart rate, etc.) are just blocked. What does the brain experience in this disconnected state? Fantasies, terrors, hopes, and fears; bizarre juxtapositions of past experiences and convolutions of future anticipations not dissimilar to the babbling of a foundational language model.

But these dreams are kept strictly internal. Our actions in our internal worlds don't leak out, and stimuli in the external world hardly Aside from the fact that e.g. screaming into someone's ear will wake them up, the brain can and will subtly incorporate external stimuli into its dreams. I've fallen asleep with Twitch streams playing through my headphones only to have strange dreams about playing-and-being-in video games. But for published evidence, see e.g. Nielsen (1993): Put blood pressure cuffs on both legs, and, when someone's in REM sleep, slowly inflate a randomly chosen one, rapidly deflating it just before waking them and asking them what they dreamt about. They'll usually (relative to controls) report a pressure or other bizarre feeling in either the specific leg or the legs more generally. The brain will often try to make it make sense, turning the raw feeling (a pressure on a leg) into an experience commensurate with that feeling, like something holding that leg in place. leak in. Consider how difficult a task this is:

Many sleep disorders are in fact partial failures of this mechanism: it's very tricky to decouple the brain's experience from the body's affordances for sense and movement in a controlled and reversible manner; when the decoupling comes on too strongly, you might end up totally paralyzed before you're fully asleep, hunted down by hypnagogic horrors while unable to move or even scream; when it's too weak, you might end up actually walking or talking while fully asleep.

Pathologies in this mechanism, or more generally in the sleep-wake transition, can tell us a lot about how the causal articulation of the mind-environment interaction works. Anecdote: a couple years ago, I occasionally had false awakenings—episodes where one experiences waking up and going about their morning routine, only to realize that this was a dream, and they're still in bed.

They often happen back-to-back, two or three in a row before you actually wake up—which is a very surreal experience. I would get out of bed, brush my teeth, make cereal, and often have entire conversations, both in person and through text message Yes, I can read and write in dreams. I can't point to any reason why, but it's never been a problem for me. In fact, just last night I was in some stats class where I was trying to solve a recursive probability problem. Someone uses a particular probability test, either this Bayesian one or this frequentist one, to decide whether to get on a bus—what is the probability that they're a Bayesian given that they don't get on? My scratch work (writing down various prior and conditional probability computations) stayed coherent, except for a few cases where I scratched stuff out and it disappeared the next time I looked. I couldn't find the answer, since, as I mentioned to the professor, we didn't have full information on how they'd make the decision to get on if they were a frequentist. The professor just drew some lines on a graph and told me to 'learn calculus'... , before... slowly, my experience begins to come "loose"... I'm still in bed, aren't I... yeah, these are blankets. Well, now I'm awake, and, introspecting, everything just feels coherent—it's nice how you can always be certain that you're in reality, isn't it? So I get up, find that we didn't have the type of cereal I originally picked, that I didn't actually get a text message, and... the world begins to come loose again... no, dammit, I thought it was real this time... hello again, blankets. Eventually, though, the world doesn't come loose, and I can be pretty certain that I'm awake for real.

Now, in the perhaps thirty second period between my being (falsely) confident that I'm eating cereal and my being (truly?) confident that I'm in bed, I'm experiencing two entirely different world-models. Are they running on different brain structures? Occam's razor says they're almost certainly not: given that the "tech" required to construct the experience of a cohesive external reality already exists in the brain, why would it be reinvented elsewhere? It's much more reasonable to say that the brain's construction of reality is the same whether asleep or awake, and that it's sort of like an ant: when there's no trail for it to pick up on, it just meanders in no particular direction, but, when it happens upon a trail—a causal process that generates a manifold of inputs it's capable of binding into a single coherent, enduring experience—it quickly entrains to that trail; that is its reality.

27 Connections

In general, the brain's causal control is not causal isolation. If I chuck a pebble at someone's head, I do, in fact, manage to affect their brain: the impact is transmitted into nerve signals which go to their brain, where they generate a feeling like 'ow!' and an impetus to turn around and curse at me. But note that the causal influence is not direct: it's always translated to the brain's own language, that of neural signals, in a safe way that cannot physically harm the brain. The body is a filter for real-world inputs, and it's only through this filter that the brain can be affected by the outside world in a way that allows it to model this world. You can chuck a battle axe at someone's head and thereby directly physically influence the state of their brain, but this influence will be illegible—it will be nothing that the brain (which has no nociceptors; it cannot itself feel anything) can understand and intelligently respond to.

If we were to draw a causal model representing the way in which signals to the brain are safetyproofed, part of it might look like this:

Without cerebrospinal fluid and meninges, for instance, a bumpy car ride might cause your brain to rattle around your skull; with them, your brain is insulated from the direct impact of the jerks and accelerations, only receiving them indirectly through proprioception and tactition. Causal influences which would be dangerous if absorbed directly are turned into harmless With respect to the range of phenomena in the natural environment. Photosensitive epilepsy allows flashing lights to directly harm the brain; there are probably other, more generic threats. neural signals: you have to play the brain's game if you want to affect it.

Often, real-life causal models just get arbitrarily complicated as we zoom in on details, and never really split up into discrete sets of causes or effects. But the causal control of the brain admits graphical depiction: ignoring blood and CSF movement(which mostly don't matter conceptually; if there were no influence, neither general anesthetics nor exsanguination would have any effect on the brain, but these are pathological situations), we can get a pretty good idea of what it would look like to take the above diagram and zoom in on the brain. Here it is:

Some terminology: afferent pathways conduct nerve signals that arrive at the brain with sense-information, while efferent pathways conduct nerve signals that exit the brain with motor commands.

All rectangles save the spinal cord are cranial nerves. On each hemisphere, there are:

Finally, by 'spinal cord' I mean the aforementioned dime-sized portion at the end of the medulla oblongata. These are all the places where neural signals can enter or exit the brain; counting them, we find that there are only $(13 \times 2)+1=27$ connections that you have to control if you want to completely isolate the brain from the body and external world.

Life Hacking

Imagine I perform a surgery on some poor fool in which I insert a thin disc through each of these twenty-seven connections: each disc is capable of inputting and outputting arbitrary neural signals on either of its faces, and the collection is therefore capable of performing a man-in-the-middle (MitM) attack on someone's brain. With such an apparatus, I could make their body perform any action that some motor signals would make it perform, independent of what motor signals they're actually trying to send. I could make them juggle or play the violin as though they had been practicing their whole life, make conversation while maintaining eye contact, or anything else they'd ordinarily be incapable of due to mental factors. Or—were most efferent signals to just be blocked, they'd find themselves paralyzed You have to be careful here: since efferent signals from the brainstem control functions necessary for survival like digestion, breathing, and heart rate via the autonomic nervous system, blocking all of them would be fatal pretty quickly—it would be equivalent to brain death. So I'm going to distinguish between autonomic signals and other motor signals; absent a connection to the brain, the former have to be provided by some sort of generalized pacemaker. .

At the same time as I'm puppeting the brain's motor output, I can simultaneously puppet its sensory input, making their mind experience any world that some sensory signals would make it experience, independent of what sensory signals the body's actually trying to send. The nice thing to do would be to simply replicate both sensory signals going to their brain and motor/autonomic signals leaving their brain; this "passthrough mode" would be indistinguishable from being unaltered. On the other hand, I could block all signals going either way, using a pacemaking computer to keep the body going; this "isolation mode" would be a perfection of the sleeping state.

But. The real fun starts when I get a computer capable of simulating a world. When it is able to (a) simulate a full world which one can interact with through an avatar, (b) decode all efferent signals into movements of the avatar, and (c) encode the avatar's simulated sensory stimuli into afferent signals, then by blocking all motor signals going to the body It is, by the way, necessary to divert your brain's autonomic efferent signals to the simulator and send fake, pacemaker signals to the body. If you run a marathon in the sim, you should experience an elevated heart rate, but your actual heart rate should not increase; incorporating autonomic signals into the sim and sending pacemaker signals to the real body is how you do this.
Would you die in real life if you died in the sim? No, of course not. I very strongly doubt that any afferent signals corresponding to actual fatal experiences would be able to manipulate the brain into shutting off, except perhaps for those reverse-engineered to do so—and these will be weird, just-barely-interpretable experiences drawn from the entropic sea rather than perfectly simple, legible things like being shot in the head. This is why images that kill you when you look at them, like this one, look the way they do. and all real sensation, replacing the latter with artificial stimuli, I can put the person in a fake world which they sense and interact with just as they do with the real world. If it's fast enough at the decoding-computing-encoding process to compute real physics in real time—say I downloaded another 16 gigabytes of RAM—they could actually just experience whatever world they (or I) wanted.

Many of the possibilities represented above are schematically represented as causal networks in the below diagram; the lower right network is a reality-augmentation scheme generalizing the simulation scheme.

The autonomic arrows can be ignored for most purposes; the story of each scheme is given by tracing the afferent (blue, sensory) flows upwards and the efferent (red, motor) flows downwards. In the simulation scheme, for instance, the afferent path tells us that environmental stimuli become sensory signals in the body, but are blocked by our discs ("I/O control"), which replace them with artificial sensory signals computed in a simulated world. The efferent path tells us that the motor actions the brain plans in response to its experience of the simulated world are relayed not to the body but to the computer, as is required to give them control over their avatar.

Sensorimotor Bounding

I suspect that causal articulation is a common property of the kinds of systems that we'd call intelligent, and in such a way so as to make man-in-the-middle attacks like the one depicted above for the human brain possible for general artificial intelligences.

There's a sort of "sensorimotor bounding" criterion that arbitrary optimizing systems must satisfy for us to be able to model them as optimizing systems: they act on the world in a certain way which takes into account the state of the world, and therefore have to have at any given point in time some interfaces through which effects (changes with causal import) flow "into" and "out of" them. Some systems are fuzzy and non-Cartesian enough that we can't get anything out of treating them in this way, but those we care about will be treatable in such a way. You can call globotechnosociocapitalism the thing that'll really kill us, but, however you slice it, it's too big, too fuzzy, too hard to bound. Artificial intelligences, on the other hand, can be bounded, albeit not spatially It's dangerous to think of these sensorimotor bounds as neat spatial bounds; this commonly holds for particular biological organisms, but fails just as often as it holds for most actually important situations. You can delineate certain patterns among the spatial extent of these interfaces, these patterns tending to organize themselves along the structural "creases" provided by the semantics of the implementing systems, but what statistical/ideatic forces conspire to make these patterns spatially coherent are weak at best.
Take a DAO, for instance. They're understandable as optimizing systems. But as physical systems? Even if you manage to say anything useful about the ways in which modern computers physically articulate their computations as electromagnetic field changes, there's still the fact that the (say, Ethereum) blockchain is a stochastic jumble of computations carried out by computers throughout the world. Even if you figure that out, you have to deal with the fact that the DAO itself, while governed by its smart contract, is instantiated by countless tiny changes to the state of the Ethereum Virtual Machine, stochastically jumbled across the blockchain itself. To solve the problem once is akin to unscrambling an egg; here, you're trying to unscramble an egg with a cocktail shaker in a house fire. The only good strategy is to rethink your life.
In general, you need to understand a system's semantics in order to perform any sort of sensorimotor bounding, since many of the systems we care about, including optimizing systems, don't have any meaningful physical instantiation. Obviously, since we are physical, anything that can actually affect us must do so through its physical effects, but many of the everyday systems we reason about with implementation-agnostic: they have the same qualitative behavior regardless of how they're physically instantiated. If you wire me a million dollars, all you've physically done is shuffle a few electromagnetic signals in some servers; but you'd never be able to pinpoint the changes in the ${\bf E}$ and ${\bf B}$ fields—and, in fact, you shouldn't be able to, because a wire transfer should be agnostic as to which banks we use, where the servers are located, whether they use SSDs or HDDs, or all the other things which massively change the physical instantiation of your action. This physical illegibility and indeterminacy of the transfer are kept under control by the semantics of our computational and banking systems, which allow me to shuffle more electromagnetic signals in a way that ends up producing very clear physical events, like the sudden migration of tens of thousands of beanie babies to a single apartment in the United States. — much better to think about them, then.

Any AGI, insofar as it senses and acts on a world, constructs its sensations in and dispatches its actions on an internal representation of the world. Attempts to understand what this representation directly says about the world may not prove successful, but, by understanding how to locate and bound any such representation, we can modify and safetyproof its relation to our reality. For instance, there's no necessary reason that the AGI ought to locate whatever it calls utility in our reality; if merely dreaming of utility is good enough for it, with us modifying its dreams so it ends up thinking solutions to our problems, that may be good enough for us. If it thinks about and acts on controlled realities, whether our simulated worlds or its dream worlds, and if it places its notions of self, sense, action, environment, and utility entirely in these realities, we can control it without even having to restrict it.

Addenda

The original essay was meant to be short and simple; these are parts I wrote, but cut from it. Instrumental Causality discusses the manner in which causal articulation arises out of necessity (or instrumentality) in optimizing systems, and how we are to understand and model it; Cartesian Confusion discusses the manners in which we come to believe we're directly observing and acting on reality (and therefore think ourselves to be "above" sensorimotor bounding).

Instrumental Causality

A dialogue on the a priori origins of causal articulation in physical systems.

Bob
Rephrase the question. Internal/external influence breaks down when you consider temperature, entropy, decay, and so on; what we can feasibly want is to keep an adversarial outside world from altering a physical thing in a way that we don't want. Because (a) we do not know what an adversarial agent might attempt, (b) most things we want to protect are in certain organized, or low-entropy, states that make them valuable in the first place, and (c) even random changes to a low-entropy state will almost always raise its entropy, making it less valuable, we ought to close off all relevant causal channels, putting emphasis on each in proportion to the extent to which that channel can effect the kinds of changes to the object that we count as a raising of entropy. If the low-entropy object is a brain, then we want to protect it from the things that can easily shake it up, destroy its information—chemical, physical, biological attacks—rather than from radio waves or magnetic fields, which, at the energy scales made 'canonical' by the natural environment, don't really alter the brain. (It's not impossible, see TMS, but it just doesn't really happen).

Alice
But if you cut it off from the world, what's the point of having it? Aside from sentimentality, we usually protect things because they're useful to us, which requires our keeping them open to interaction. So, take the thing in question and suppose it ought to be able to interact with the outside world—that's where it derives its use. How do you alter your protections to accommodate this?

Bob
Create controlled interfaces between the thing and the outside world. Find the smallest possible causal channels required to convey whatever effect the inside thing ought to be having on the outside thing, find the ways to implement them that interfere the least with the existing isolation mechanisms, and guard them very carefully to make sure nothing passes through them that isn't of their causal type.

Alice
There are three unstated assumptions that threaten this approach, as you've laid it out:

Bob
I don't see how this logic of partial accommodation is anything but a random buildup of congruences. You've pointed out that humans talk and eat through the same hole, but this is just a quirk of human anatomy. That we use the same physical structures for talking and eating is a random result of the fact that those structures were already in place, serving some other purpose, when the need to talk arose. Our mouths and throats are an accident of evolution.

Alice
That's exactly what I mean. Those biological structures that were already in place for eating lended themselves to co-optation for a new purpose which, while subjectively independent, could be achieved through shared objective means. Necessity, then opportunism.
It's as though you have a ring $R$ of causal effects, and a subset $S$ chosen subjectively, independently of the objective ring structure, representing those effects you want to allow. If you want to reify $S$, you need to deal with the fact that any physical structures you create will be usable in ways not dependent on $S$, because $S$ isn't a property of the physical world—it's a property of the way you've chosen to look at the world. The reification, as another subset of $R$, will consist of all possible ways that the physical structures necessary in the creation of $S$ could be used. This objectification will look like the ideal generated by $S$, $\langle S\rangle$.
The simplest case is that $R$ is a principal ideal domain: if your entire universe is sending emails to addresses, and you want to develop the ability to send five particular emails to five particular addresses, then you'll need to create the ability to send arbitrary emails to arbitrary addresses. For general $R$, things will be far more complicated, but you'll always find that the ideal is articulated by common denominators among the elements of $S$.
Now, how do you implement your protections for our thing in such a way so as to allow it to interact with the world in some particular ways?

Bob
I guess... we have to find common sets of underlying physical correlates for our causal channels and implement guarding mechanisms on this physical level? Trying to work on the physical level directly seems much more dangerous, though: we can never work out all the macro-level phenomena that come from atomic movements, and generally only see what our subjective motives guide us to see.
To use the ideal metaphor, we never really see the whole ideal $\langle S\rangle$. We have our own subjective frameworks, and don't really have the motivation, creativity, or combinatorial capacity to conceive of all the ways that our constructions can be combined. Basic computers could've been created with the telegraph relays available in the early 19th century; the ancient Chinese used heated air to float balloons for signaling purposes, but never scaled up to human flight.
An adversary that conceptualizes subjective causal channels in a slightly different manner, whether due to different motives or a different style of cognition, could see an entirely new angle to the same objective reality—could exploit parts of $\langle S\rangle$ that we'd never see. So, by trying to work at the level of physical reality, aren't we playing a much less certain game?

Alice
You're playing the exact same game: the exploits your causal channels allow don't go away upon your reframing the situation. Thinking of physical realities just lets you see how uncertain this game is. The moment we sought to both protect a thing and allow the outside world to interface with it, we let this uncertainty enter; these goals are in conflict, and this is the resulting tension.
Our subjectivity will always be our greatest vulnerability: we think of developing methods for controlling interactions, but even the notion of an "interaction" is a mental construct, a subjective fabrication on top of reality, and, like most such constructs, it breaks when we put too much weight on it. An artificial superintelligence may be able to keep track of statistical ensembles of particles in a way that allows them to formulate and place guarantees on the behavior of all macro-level phenomena A macro-level phenomenon is a schema common to a space of micro-level situations, so a Solomonoff-like utility maximizer that assigns to all possible schemas a Kolmogorov complexity-based weight and an effective utility obtained by seeing what the max-entropy microstates (or, near-equivalently, randomly sampled microstates) look like for the actual utility function should understand how to maximize expected utility through macrostate manipulation. (...i think? this is hand-wavy speculation). —to construct a map which genuinely covers the territory—but humans can only get so far; our maps fray at the edges.

Bob
Should we not then defy the question itself, saying that it just isn't right to talk about taking on causal articulation as a goal?

Alice
Evolution did not decide to causally articulate the brain. Don't think of it as a goal that one consciously has in mind. I'm emphasizing what naturally happens because I want you to think of causal articulation as an instrumental correlate. Arbitrary goals in the real world often end up having us protect some Macguffin while simultaneously keep it open for interaction, and causal articulation strategies are created to this end; because they're generally not conceived of as such, they end up evolving in the manner I described: messy, overlapping, and mixed up.

Bob
Evolution doesn't have a mental construct of interaction. So why does its causal articulation control interactions?

Alice
The notions of causal articulation and interaction exist within us, in application to things, whether we're directly instantiating them or merely referencing them. If they were puppets, you'd be manipulating both; and, after making one push the other, you'd be asking me why it did that.

Alice
No! You have mathematical proofs about conceptual constructs. Whenever you apply them to reality it will always be by assuming that reality is in a certain way, that this aspect of reality is not only clearly delineable but identifiable with this mathematical object, that this aspect of reality is genuinely coherent to speak about as a numerical parameter, and so on and so on.

Alice
Because when we can't mathematically model a situation, we don't use mathematical models. There is something about reality which allows mathematical structure to appear in a sort of fuzzy, fractal way even at the level of macroscopic conceptual delineations, that is clear, but it does not approach the level of axiomatizability. What mathematics we do see working is either the product of engineering, which does not admit axioms and proofs, computer science and statistics, which do in fact suffer from the fact that we cannot apply its proofs directly to reality, or physics, which does sometimes seem to provide true mathematical limits on reality by virtue of limiting possible experience, but pretty much arrives at these in a haphazard heuristic way, mathematicians trailing behind in an attempt to construct axioms that manage to prove what the physicists already know Have you ever tried to understand quantum field theory from a mathematical point of view? It's a total mess! A discombobulated forest of artificial structures designed to capture the things physicists do within a rule-based framework, each with a thousand little ad-hoc variations designed to make sense of all the ways physicists end up breaking these rules. Never ask what the measure ${\cal D}\phi$ of a path integral is—and if some smartass tells you it's the Wiener measure, ask them to actually compute the path integrals physicists do with that measure and count how many milliseconds it takes them to backtrack. .

The schema of causality lends a deep sense to our understanding of interactions between systems. Natural selection, for instance, is a fundamentally aleatoric force whose creation of complex modular systems such as ourselves seems almost absurd at first—but, when its designs are understood as abstract structurings of causal flows within an environment (e.g., we think of "the" human, "the" mouth -- not any particular human, any actual mouth, but abstract designs giving rise to abstract environmental interactions such as "eating", "talking"), we can understand why these designs turn out the way they do. We can make sense of the ubiquitous multicausality of our organs this way: mutation pressure is simply taking what is given to it and marginally adjusting it, and, when these adjustments happen to open up entirely new causal flows (e.g., when the breathing and eating apparatus becomes useful for 'vocalizing'), natural selection rapidly pursues the newly available fitness gradients.

Another area in which the schema of causality makes such deep sense of things is in the design of technological interfaces. In cognizing interactions between abstract systems, we can think about what causal flows are inherent in a given goal: if there's some API which has to permit access to medical records on a remote server to users with the right credentials, for instance, then it's easy to see that the goal necessarily requires the remote server to be open to arbitrary data inputs from arbitrary computers; given how networks work, this causal channel must be opened for the goal to be possible at all. The very fact that it parses credentials at all tells us about what causal flows must necessarily be allowed to pass through this channel: contingent on an input having some structure that any possible valid request could have (e.g., it encodes a well-formatted JSON request from a non-blacklisted host)(and the very fact that this sort of validity is required means that it somehow has to check for this sort of validity, which is another causal flow inherent in the goal etc. etc.), it has to do some computation which searches that request for credentials and, if it finds them, it has to do some computation which checks if these are valid credentials. All of these are causal flows with physical implementations; as such, they are all possible targets for attack by an adversary.

There is a largely mathematizable structure to interaction which is made visible to us through the schema of causality -- the lens of causal channels, causal flows One question that, after everything, still stumps me is -- to what extent does causality "really" exist? Obviously this is tied in with the question of formation of objective macroscopic patterns (ideatics), but there's a physical aspect to it given by the existence and seeming fundamentality of time (which seems to emerge from the unidirectionality of (the most basic kind of) causality). It is clearly conceivable to me how space could be dissolved into more fundamental structures in a reductive way In fact, non-relationalist theories of spacetime seem relatively non-intuitive to me; space feels like it ought to merely emerge as a macroscopic structure of latencies of causal relations between different events (with lightspeed the tradeoff between causality, the basis of time, and spatial extension). But this doesn't answer why all these relations should end up fitting into three spatial dimensions. Or twenty-six; the Kaluza-Klein theory (in which adding a fourth spatial dimension parametrized by a tiny circle causes the Maxwell equations to just pop right out of general relativity) is a good example of how force can arise solely from the structuring of what relations underlie spatial extension, but why would they be structured in such a way in the first place?
Note that there does seem to be plenty of work on this problem, related to certain causal approaches to quantum gravity (see e.g. this), but I'm not familiar enough with it to have an opinion. , but not so with time. Many theories of quantum gravity do in fact dissolve space, but they don't dissolve causality—many of them explicitly make it fundamental (e.g. "causal dynamical triangulation", the "causal sets" framework).

Cartesian Confusion

There are two related illusions conflated by this single word. The first is dualistic—that one can be cleanly divided from the world, rather than being embedded into it. The second is agentic—that one receives observations directly from the world, that one chooses actions to directly make upon the world. Both of these arise from the illusion of a self, projected by humans not just onto ourselves but onto the conceptual fabrications we call 'systems'. When we reify, say, evolution, we imagine it as a intelligent actor—an eye separate from what it sees, a hand separate from what it manipulates. When we think of this actor using the same fundamental schemas with which we think of ourselves and others (e.g., as 'wanting' certain outcomes), we imagine it seeing and acting on a reality external to itself. For now, though, it'll be the second illusion I'm concerned with here, and the way both aspects of it (the non-Cartesian nature of both perception and action) are made clear by the causal articulation of the brain.

Non-Cartesian Perception

When I see a yardstick, a clock, or any other physical object while walking around my bedroom during a false awakening, where is that object located? Even if my dream is correct about the object's location and appearance within my room, the object that appears in my dream can only physically be in my brain. But my brain isn't even a yard long—so in what sense can it contain the yardstick I see? It can only be in the sense that the yardstick is, as an experienced object, the product of neural activity. It's like -- if I'm painting a bowl of grapes that sits in front of me, then, no matter how realistic my painting, it is still fundamentally not the bowl of grapes itself, just a certain pigmenting of my canvas which I've managed to couple very finely to the appearance of the actual bowl of grapes Consider that if I were colorblind, my painting would be different and less accurate, even though the bowl of grapes would be the same. Now consider that I am 'colorblind' as regards the infrared and ultraviolet ranges. Insofar as my senses are reliable—the same inputs yield experiences which I depict via the same outputs—they're like a homomorphism which is long way away from being an isomorphism. . It exists on and is confined to the canvas, and its relation to reality is merely one of reference.

If I truly wake up, and see the yardstick right after getting out of bed, then what the object is for me has not changed; given the way false awakenings chain together, I can't even justifiably say whether it's real or not. What has changed is that the experienced yardstick is now coupled by my perceptive apparatus to the physical yardstick—but the fact of this coupling is not immediately deducible from the experience being coupled. It is defeasibly deducible in general from the hidden coherences reality has (I won't explain here), but this is neither immediate nor applicable to particular cases.

This is all to make the obvious point that the world as you experience it is necessarily in your head, and whether it's "real" or not consists only in whether the perceptive process serves to predictably couple it to the physical facts. You only ever interact with the canvas, but there happens to be a system which, when you paint an arm reaching for the painted grapes, reaches a real arm for the real grapes To be clear, the canvas is a metaphor only for the indirection of the brain's relation to reality; there isn't, actually, any one single interface through which the brain senses or acts. It's more like... a canvasciousness, uncountable and continuous, permeating the mind and self-articulating prior to the concept of a self. I'm not going to explain here, but Aella's A Mild Solipsism Party, which talks about "canvasciousness" more phenomenologically, is on the right track. .

Non-Cartesian Action

If our subject tries to pick up a ball on the ground but doesn't feel their arms moving concurrent with their sending the neural signals that move their arms, they might notice something's up! The easiest and most natural solution to this is to just let them control their simulated avatar as they'd control their real body. But in principle, our ability to prevent our subject from realizing something's off depends only on what they sense, not their actions. If we rig the environment such that they only ever decide to make actions which they expect to induce sensations that happen to correspond to the sensations we're going to provide, we don't actually have to take their actions into account at all. But this strategy requires us to 'read their minds' to an extent: since the causal network looks like $sensation \to brain \to action$, even full knowledge of P(action | brain, sensation) won't allow you to compute P(action | sensation) unless you also know their actual brain state or have full knowledge of P(brain | sensation) as well. (The latter is possible for very powerful sensations: if you give them the sensation of suddenly being stabbed in the eye, then probably nothing will go wrong if you throw away their next actions and just decide to give them the sensation of bringing their hands up to their face and screaming).

But the entire point of the experiment is to demonstrate how we can isolate the brain from the body and environment, so let's follow a "cerebral sovereignty" principle, rejecting strategies that require us to guide the brain into specific states, and a "cerebral privacy" principle, rejecting strategies that require us to gather direct knowledge about brain states. It still follows that movement can generate the percept of action, and action can generate the percept of movement, without any actual relation between them.

The next time you're sitting somewhere and find that your leg's gone into a deep enough sleep to render your foot immobile, consider trying really really hard to move it. From my own experience, this just won't work—the leg wakes up on its own terms—but there is definitely something it is like to send that motor signal: it's like holding down a mental button, or lever, which sort of impels one's experienced body to be in a certain way. This often comes with a slight feeling that the foot is in fact moving, but you can visually confirm that it's perfectly still. When the leg wakes up again, it's nearly impossible to feel this mental lever.

I would use an analogy: I've played enough Playstation games to be instinctively familiar with the layout of a Playstation controller. So when playing a new Playstation game, I can start learning the controls the moment I pick up the controller, and from then on my actions won't parse as "$\bigcirc$, $\sf{R1}$, $\square, {\sf{R2}...}, \triangle, {\sf{R3}}$" but as "heavy attack, block, light attack, charge up special, kick, activate special" One aspect to this I haven't yet made sense of is that it only works so well because the button assignments are idiomatic. Particular buttons fit into common roles, or relations, across games, and the button assignments I give are ones pretty natural in this sense. If the assignments weren't idiomatic, the way my brain would learn to parse actions would be... weird in a way I haven't specifically looked into. . I will not have any conscious awareness of pressing buttons on a controller, just as I didn't have any conscious awareness of hitting the keys on this keyboard while typing the last paragraph. But I've held an Xbox controller maybe like ten times in my life, and I have to look at the controller to remember which button is what. Were I to play the same new game, my actions would just parse as a list of illegible button presses like "$\between, \Xi, \ltimes, \dagger, \, \natural, \nabla$" which happened to very fuzzily tie to certain game states As I write this, the idea comes to me: literally pretending that the Xbox controller is just a funny Playstation controller might instantly solve this problem. But, if this would work, it's only because the two have very similar shapes, and I've never exploited this similarity except maybe subconsciously. Replace 'Xbox' with something more exotic and the point still holds.. There's a sort of 'zipping' that goes on as I learn how to play the game, where my certainty over P(controller input | hand position) increases and is increased by my certainty over P(game actions | controller input), but that's out of scope, I think.

The brunt of the analogy is this: my ability to send specific motor signals to specific parts of my body is like a Playstation controller I've been using since the day I was born. Even though the mental buttons are there in some sense, I'm never conscious of pressing them, just of the consequent movements. And if I try to isolate the experience of pressing the button without doing anything about the movements, I'll just end up confusing it with something else like a constructed kinesthetic extrapolation. But if an external factor can render a button unresponsive (in this case, a leg falling asleep), and I can do the thing that corresponds to mashing that button (trying really really hard to move my foot), I have the opportunity to recognize that there's a button here that I'm mashing.

(In general, the brain does a certain thing that's like... extrapolative development I can't word it like "predictive processing"—that phrase is taken, and philosophers feel the need to have their say. Due to their severe intellectual deficiencies, the vast majority of academic philosophers cannot engage with arguments unless they have enough points of reference—simple concepts which they have simple opinions on—to be able to construct a low-fidelity mockery of the argument which they understand and find indistinguishable from the original argument. They need to be able to match everything to a "side" so that they can regurgitate what standard counterarguments exist in the lookup tables they call 'minds'; it's no wonder that philosophy of mind is a hopeless field. Reference to the philosophical literature, or any sort of engagement with philosophers, is actively counterproductive. The best case is that they pattern-match and emit some irrelevant statement like "actually the Wubbles chandelier argument for profunctionalist $\varphi$-literalism demonstrates that subphysicalist t-werewolves can't exist without infinite referential regresses" before disengaging., both backwards and forwards in time. The brain's activity is a world model, the surface of which appears to us as the entire phaneron, and the $\frac d{dt}$ of this model is trying to reduce tension among possibly-incongruent properties. Sensations force themselves into the world model by directly affecting the brain; to feel a sensation requires some sort of higher-level coherence which can exist without the sensation itself).

Going back to the topic of misleading someone as to the body they're controlling, this gives us a perspective more compatible with the physical facts: it's as simple as unplugging their mental controller and plugging it into a port of your own creation.