The conscious tape

Or the information on the tape, or - whatever it is that's conscious, it is a static thing, not a living, dynamic thing. If computation is an output, process doesn't matter. Time doesn't enter into it.

Time is right where it should be in at: In the tape. We, the Matrix-Lords, are above the time of the tape, but it doesn't mean the consciousness within the tape isn't living and dynamic and all that.

The question presented is very intriguing, thank you for it.

This question arises when I consider the moral status of intelligent agents. If I encounter a morally-significant dormant Turing machine with no input devices, do I need to turn it on?

If yes, notice that state N of the machine can be encoded as the initial state of the machine plus the number N. Would it suffice to just start incrementing a counter and say that the machine is running?

If I do not need to turn anything on, I might as well destroy the machine, because the Turing machine will still exist in a Platonic sense, and the Platonic machine won't notice if I destroy a manifestation of it.

David Allen notes that consciousness ought to be defined relative to a context in which it can be interpreted; somewhat similarly, Jacob Cannell believes that consciousness needs some environment in order to be well-defined.

I think the answer to my moral question is that the rights of an intelligent agent can't be meaningfully decomposed into a right to exist and a right to interact with the world.

ADDED: Many people are replying by saying, "Obviously, option 2 is correct," then listing arguments for, without addressing the problems with option 2. That's cheating.

Phil, I have to say that I don't think the problems with option 2 are actually presented in you post. But that does not mean that we are allowed to dodge the question implicit in your post: how to formally distinguish between two computational processes, one conscious, the other not. Let me start my attempt with a quote:

"Consciousness is overrated. What we call consciousness now is a very imperfect summary in one part of the brain of what the rest is doing." - Marvin Minsky

I believe with Minsky that consciousness is a very anthropocentric concept, inheriting much of the complexity of its originators. I actually have no problem with an anthropocentric approach to consciousness, so I like the following intuitive "definition": X is conscious if it is not silly to ask "what is it like to be X?". The subtle source of anthropocentrism here, of course, is that it is humans who do the asking. As materialists, we just can't formalize this intuitive definition without mapping specific human brain functions to processes of X. In short, we inherently need human neuroscience. So it is not too surprising that we will not find a nice, clean decision procedure to distinguish between two computational processes, one conscious the other not.

Most probably you are not happy with this anthropocentric approach. Then you will have to distill some clean, mathematically tractable concept from the messy concept of consciousness. If you agree with Hofstadter and Minsky, then you will probably reach something related to self-reflection. This may or may not work, but I believe that you will lose the spirit of the original concept during such a formalization. Your decision procedure will probably give unexpected results for many things: various simple, very unintelligent computer programs, hive minds, and maybe even rooms full of people.

What is consciousness, and what kinds of things are conscious?

I've seen this debated many times, and I suspect that people are merely arguing over the meaning of a word that does not happen to carve reality at the joints.

What is it in the physical universe that you call consciousness and which might be present or absent in a computational device? What prediction is being made by any of the theories 1 through 3 in this post, that I can go out and test?

I endorse the first alternative; the intuition at first felt wrong (in a Chinese Room sort of way), but that feeling disappeared when I realized the following:

I was envisioning a tape (call it Tape A) which only recorded some very small end result of Turing Machine A, like the numerical output of a calculation or the move Deep Blue makes. And that seems too "small" somehow to encapsulate consciousness— I felt that I needed the moving Turing machine to make it "live" in all its detail.

But of course, it's trivial to write a different Turing machine which writes on a tape (call it Tape B) the entire history of Machine A's computation (as well as its output), and this indeed has the required richness for me to be comfortable in calling Tape B conscious.

Your question is well-posed, but I doubt that it really attacks the problem of consciousness.

I don't understand what it could possibly mean for an output or an algorithm to be consciousness. Consciousness, whatever it might be caused by or composed of, means subjective awareness of qualia.

In this sense, "3" is not even slightly conscious. Neither is "output=0; if input=3, let output=1." Neither is "output=0; if input=3, let output=1, get input by scanning source code of self and reducing it to a number." The last example will behave as if it were aware of its own algorithm, but there is no reason to think that it is actually conscious in the sense of having subjective qualia. I do not understand how any amount of complication, recursion, or nuance could take these basic elements and turn them into something that would experience qualia.

It's not that consciousness is or ought to be mysterious; on the contrary, we ought to ask questions about consciousness whenever and however we can. So far, though, I think the only fair response to those questions is "I don't know." It is part of the strength of a rationalist that she will notice when she is confused; our attempt to apply concepts like "algorithm" and "output" to the problem of consciousness ought to result in confusion, for the simple reason that the concepts already in our toolbox do not adequately correspond to the concept we are trying to investigate.

In humans today, the output is not produced all at once -but from a computationalist perspective, that isn't important.

I don't think this is valid; the complexity you're discarding here may actually be essential. In particular, interacting with some sort of universe that provides inputs and responds to outputs may be a necessary condition for consciousness. Perhaps consciousness ought to be a two-place predicate. If you have a universe U1 containing a mind M, and you simulate the physics of U1 and save the final state on a tape that exists in universe U2, then conscious(M,U1) but ~conscious(M,U2). On the other hand, if U1 interacts with U2 while it's running, then conscious(M,U2).

Suppose we view consciousness as both a specific type of computation and a specific range of computable functions. For any N, there will always be a lookup table that appears conscious for a length of time N, in particular, "the lifetime of the conscious creature being simulated". A lookup table, as Eliezer once argued, is more like a cellphone than a person - it must have been copied off of some sort of real, conscious entity.

Is the feature of a lookup table that makes it unconscious its improbability without certain types of computation, that is, the large amount of code? Is consciousness computing some function without using much code? That doesn't seem right.

A Turing machine is much more reasonable than a lookup table.

Premise 1: I am conscious. Premise 2: The physical universe I am part of may be computable. Therefore: The physical universe I am part of may be a Turing machine. Conclusion: Aspects of a Turing machine computation may be conscious.

What are aspects of a computation? I know them when I see them, or at least do occasionally. There is no reason I know of that a rigorous definition is impossible.

A powerful computer in a sealed box is about to be fired away from the Earth at the speed of light; it will never produce output and we'll never see it again. From the point of view of perspective 1, the whole program is thus equivalent to a gigantic no-op. Nonetheless, I'd rather that the program running on it simulated conscious beings in Utopia than conscious beings in Hell. This I think forces me to perspective 2: that actually doing the calculations makes a moral difference.

EDIT: the "speed of light" thing was a mistake. Make that "close to the speed of light".

I don't understand the difference between "computed" and "computation", here.

I think the best definition of consciousness I've come across is Hofstadter's, which is something like "when you are thinking, you can think about the fact that you're thinking, and incorporate that into your conclusions. You can dive down the rabbit hole of meta-thinking as many times as you like." Even there, though, it's hard to tell if it's a verb, a noun, or something else.

If we want to talk about it in computing terms, you can look at the stored-program architecture we use today. Software is data, but it's also data that can direct the hardware to 'do something' in a way that most data cannot. There is software that can introspect itself and modify its own code (this is used both for clever performance hacks and for obfuscation).

My view is that consciousness is a property of my thought processes — not every thought will have the same level of introspection, or even introspection at all. It's something that my mind is doing (or isn't doing, depending on what I'm thinking about). The property we ascribe to entities that we call 'consciousness' I would instead term 'the ability to think consciously' or 'the ability to have consciousness'. It seems to me that my thought processes are software running on the hardware that is the human brain. If my mind were uploaded, and its software state written to permanent storage and then stopped running, I would say that this recorded state still has the ability to think consciously, but it is not doing so, since it's not thinking at all, so at that time it is not conscious. (But of course it could be, if started back up.)

I understand Computationalism as a direct consequence of digital physics and materialism: all of physics is computable and all that exists is equivalent to the execution of the universal algorithm of physics (even if an exact description of said algorithm is unknowable to us).

Thus strictly speaking everything that exists in the universe is computation, and everything is computable. Your option 1 seems to make a false distinction that something non-computational could exist - that consciousness could be something that is computable but is not itself computation. This is impossible - everything that is computable and exists is necessarily some form of computation. Computation is the underlying essence of reality - another word for physics.

But the word 'consciousness', to the extent it has useful meaning, implies a particular dynamic process of computation. Consciousness is the state of being conscious, the state of being conscious of things, a state of active cognition. Thus any static system can not be conscious. A mind frozen in time would necessarily be unconscious.

If consciousness is computed, then there are no necessary dynamics. All that matters is getting the right output. It doesn't matter what algorithm you use to get that output, or what physical machinery you use to compute it.

This doesn't seem quite correct. There are necessary dynamics - out of the space of all dynamics (all potential computational or physical processes), some set of them are 'conscious'. There is no single correct output. There is a near infinite set of correct outputs - the definition of which is based on what the correct dynamics computes based on the inputs.

Consciousness is not the output anymore than a car is it's exhaust or microsoft windows is a word document.

You can use the input->output mappings to understand and define the black-box process within because the black-box process is physical and so it is governed by some algorithm. But the algorithm is not just it's output.

Computationalism means considering two computational outputs equivalent if they contain the same information, whether they're computed with neurons and represented as membrane potentials, or computed with Tinkertoys and represented by rotations of a set of wheels.

No, not quite - computationalism via functionalism means considering two processes functionally equivalent if they produce the same outputs for the same inputs. It's not just about outputs.

A key idea in functionalism is that one physical system can realize many different functional algorithms simultaneously. A computer is a computational system running physics on the most base level, but it also can have other programs running at an entirely different functional encoding level - like words written in letters that are themselves composed of sentences or entire books.

If consciousness is computation, then we have the satisfying feeling that how we do those computations matters. But then we're not computationalists anymore!

Err yes and no. Consciousness is always strictly defined in relation to an environment, so speed is always important. But beyond that, how you do the comptuations matters not at all. There are an infinite number of equivalent algorithms and computations in theory, but the set of realizable equivalent algorithms and computational processes that enact them is finite in reality because of physics.

Another way of looking at:

There are many possible patterns of matter/energy that are all automobiles, or dinosaurs, or brains.

There are many possible patterns of matter/energy that are all conscious - defined as patterns of matter/energy that enact a set of intelligence algorithms we label "conscious". The label is necessarily functional.

(Warning: I expect that the following comment has at least one major error, since this topic is well outside my usual area of knowledge. Please read it as a request for edification, not as an attempt to push forward the envelope.)

Until we can detect or explain qualia in the wild, how can we make rational claims about its computability?

To make a simple analogy, suppose we have a machine which consists of a transparent box, a switch, and a speaker. Inside the box is a lightbulb and a light sensor. The switch controls the light, and the light sensor is hooked up to the speaker and makes it emit a tone IFF light is detected.

Suppose a species with no sight or concept of light is attempting to reverse-engineer the device. A reverse-engineered simulation of this machine could achieve the same external output as the original by connecting the switch more directly to the speaker, without going through a light-emitting-and-detecting phase. That would be an equivalent algorithm to the one that the original machine is executing, from the perspective of the observers, but it wouldn't be doing the same thing.

Similarly, qualia has an effect on the world and it can be detected, but at the moment only in a tentative and indirect way. In particular, we don't have tests that can distinguish false positives from true positives very well (just as the sightless scientists in the example haven't figured out how to distinguish the tone from the light).

To put it another way, the simulated machine is kind of a zombie machine; it has all the proper external observable stuff going on, but not the correct internal process. From an evolutionary perspective a zombie is unlikely, but it seems like a naive reverse engineer could make one pretty easily, since they have no way of verifying if they've got the important part working if the external indicators for the important part can be easily accidentally faked.

Executive summary: Until we can detect naturally occurring qualia, it seems plausible that any simulations we create might accidentally be zombies and we wouldn't be able to tell.

One of my favorite Philip Dick stories ("Gur Ryrpgevp Nag") is about the consciousness tape.

Whatever it is that's conscious, you can compute it and represent it in a static form. The simplest interpretation is that the output itself is conscious. So this leads to the conclusion that, if a Turing machine computes consciousness and summarizes its output in a static representation on a tape, the tape is conscious.

I don't see the contradiction here, although the sheer scale might be throwing off intuition. I'm perfectly willing to say that if you actually had such a tape, it would be conscious. Needless to say, the tape would be inconceivably huge if you really wanted to represent the astronomical amount of inputs and outputs that make up something we would call "conscious."

Methinks the discussion suffers from lacking the distinction between consciousness-as-property: "subsystem C of system X is conscious-in-X" and consciousness-as-ontological-fact: "C exists in a conscious way" . Consciousness (of "C" in "X") is option-1-computed in the sense that it is a Platonic entity (as a property of platonically considered "subsystem C of system X"). It is option-2-computation in the sense that all such entities "C" could be said to perform various computations in "X" (and it is the ensemble of computations that the property detects). To draw moral conclusions ("C exists in a conscious way"), one needs to take X=reality.

Sort of both. I think they reconcile more easily than you think.

Consciousness entities have behavior, including internal behavior (thoughts). Behavior obviously doesn't exist in stasis, which seems to be the point that you don't like about 1.

Consciousness is not the algorithm that generates that behavior, but rather that algorithm (or any equivalent algorithm) in action. It requires inputs, so that behavior can't just be read as a series of outputs and determined to be "the consciousness"; rather, consciousness is a property that entities operating inside an environment have in relation to that environment. This environment-entity combination could be read off as a series of numbers, but a series of numbers, even ones perfectly representing a thing, are not the thing itself. That's just a way of specifying something, not duplicating it.

Lastly, remember that consciousness is an illusion, but like free will, it is an illusion both difficult not to hold and useful to play along with. Much more so than for free will, consciousness can be a very useful descriptor of how a system behaves, so good definitions and understanding of it are worthwhile.

Option 4: there is no such thing as consciousness. Its all just an elaborate hoax our minds play on us. In reality we are 'just' state machines with complicated caches that give the appearance of a conscious mind acting. The illusion is good enough to fool us all, and works well enough for real life.

The more i learn off neuro science the more I get the impression that there is no real 'person' hidden in the body, just a shell, that runs lots of software over data storage, and seems to act somewhat consistent. So the closer you look, the less consciousness is left to see.

Please note that I still have trouble understanding how qualia work.

Consciousness is a roughly defined and (leaky) abstraction.

So this leads to the conclusion that, if a Turing machine computes consciousness and summarizes its output in a static representation on a tape, the tape is conscious.

Without context the content of the tape has no meaning. So the consciousness that has been output on the tape, is only a consciousness in the context that can use it to generate the consciousness abstraction.

It is the set of "stuff" that produces the consciousness abstraction that can be called conscious. In a Turing machine, this "stuff" would be the tape plus the machine that gives the tape the necessary context.

I understand Computationalism as a direct consequence of digital physics and materialism: all of physics is computable and all that exists is equivalent to the execution of the universal algorithm of physics (even if an exact description of said algorithm is unknowable to us).

Thus strictly speaking everything that exists in the universe is computation, and everything is computable. Your option 1 seems to make a false distinction that something non-computational could exist - that consciousness could be something that is computable but is not itself computation. This is impossible - everything that is computable and exists is necessarily some form of computation. Computation is the underlying essence of reality - another word for physics.

But the word 'consciousness', to the extent it has useful meaning, implies a particular dynamic process of computation. Consciousness is the state of being conscious, the state of being conscious of things, a state of active cognition. Thus any static system can not be conscious. A mind frozen in time would necessarily be unconscious.

If consciousness is computed, then there are no necessary dynamics. All that matters is getting the right output. It doesn't matter what algorithm you use to get that output, or what physical machinery you use to compute it.

This doesn't seem quite correct. There are necessary dynamics - out of the space of all dynamics (all potential computational or physical processes), some set of them are 'conscious'. There is no single correct output. There is a near infinite set of correct outputs - the definition of which is based on what the correct dynamics computes based on the inputs.

Consciousness is not the output anymore than a car is it's exhaust or microsoft windows is a word document.

You can use the input->output mappings to understand and define the black-box process within because the black-box process is physical and so it is governed by some algorithm. But the algorithm is not just it's output.

Computationalism means considering two computational outputs equivalent if they contain the same information, whether they're computed with neurons and represented as membrane potentials, or computed with Tinkertoys and represented by rotations of a set of wheels.

No, not quite - computationalism via functionalism means considering two processes functionally equivalent if they produce the same outputs for the same inputs. It's not just about outputs.

A key idea in functionalism is that one physical system can realize many different functional algorithms simultaneously. A computer is a computational system running physics on the most base level, but it is also can have other programs running at an entirely different functional encoding level - like words written in letters that are themselves composed of sentences or entire books.

If consciousness is computation, then we have the satisfying feeling that how we do those computations matters. But then we're not computationalists anymore!

Err yes and no. Consciousness is always strictly defined in relation to an environment, so speed is always important. But beyond that, how you do the comptuations matters not at all. There are an infinite number of equivalent algorithms and computations in theory, but the set of realizable equivalent algorithms and computational processes that enact them is finite in reality because of physics.

Another way of looking at:

There are many possible patterns of matter/energy that are all automobiles, or dinosaurs, or brains.

There are many possible patterns of matter/energy that are all conscious - defined as patterns of matter/energy that enact a set of intelligence algorithms we label "conscious". The label is necessarily functional.

If I were to attempt to characterize consciousness in computational terms, I would probably start with a diagram like that for the Mealy machine in this pdf. I would label the top box simply "computation" and the lower box "short term memory. I would speculate that consciousness has something to do with that feedback loop through short term memory. I might even go so far as to claim that the information flowing through short term memory constitutes the "stream of consciousness".

If this approach is taken, there are some consequences. One is a kind of anti-anti-zombie principle. You simply cannot characterize consciousness by looking at the I/O function from inputs X to outputs Z. It is at least conceivable that the same I/O function might be implemented using a completely different feedback trace - one which encoded the information Y differently is one possibility. Another possible change might be to use something other than "short term memory" to buffer the feedback.

A second consequence is that if you want to capture consciousness on tape, you probably need to capture Y rather than Z. A third consequence is that the "seat of consciousness" is either in the computational machinery, in the short term memory, or (my intuition) in both, together with the communication paths that tie them together.