June 11, 2008
The Secular Rapture
Via The Valve, the IEEE has a special issue on The Singularity. Ray Kurzweil and Neil Gershenfeld discuss 
"Two Paths to the Singularity" can be found here. The views of some prominent computer scientists, biologists and all-around smart people such as T.J Rogers, Gordon Moore and Steven Pinker can be found here. Pinker:
WHO HE IS Professor of psychology at Harvard; previously taught in the department of Brain and Cognitive Sciences at MIT, with much of his research addressing language development. Writes best sellers about the way the brain works, like The Blank Slate (2002) and The Stuff of Thought (2007).
SINGULARITY WILL OCCUR Never, ever
MACHINE CONSCIOUSNESS WILL OCCUR “In one sense—information routing—they already have. In the other sense—first-person experience—we'll never know.”
MOORE'S LAW WILL CONTINUE FOR 10 more years
THOUGHTS “There is not the slightest reason to believe in a coming singularity. The fact that you can visualize a future in your imagination is not evidence that it is likely or even possible. Look at domed cities, jet-pack commuting, underwater cities, mile-high buildings, and nuclear-powered automobiles—all staples of futuristic fantasies when I was a child that have never arrived. Sheer processing power is not a pixie dust that magically solves all your problems.”
Posted by Robin Varghese at 11:11 AM | Permalink
Comments
So it would help if you'd tell me whether you agree in principle that the brain-in-a-vat could in principle still have experiences just like those of an embodied brain, if the right pattern of nerve impulses and other discrete signals was fed into it (leaving aside the technical difficulty of simulating such a pattern).
In principle, no. I don't think these signals would be separable from the means of their transmission.
We are talking about a brain in a vat...although the body is only simulated, in order to feed information into the brain, you'd need an interface that would translate inputs to simulated sensory organs into real patterns of electrochemical impulses to feed into the brain. Are you saying that even if you sent the right pattern of real physical nerve impulses in through the brain's optic nerves, the brain would somehow be able to "know" whether or not the optic nerves were receiving impulses from a real retina taking in light or an interface which was stimulating different optic nerves based on the simulated light hitting a simulated retina? This would suggest some sort of mysterious psychic connection between the brain and the body, instead of the brain becoming aware of what was going on in the body through a pattern of discrete, localizable signals like nerve impulses. It would also go against the universal principle in physics that a system's behavior can only be affected by events on the system's boundary, that a system cannot have any nonlocal information about events outside it except to the extent that they influence events on the boundary.
The first mind uploads might find that their simulated environment was limited to something like a very sterile-looking hotel room, but the experience of that room might at least be close to indistinguishable from staying in a real version of the same room.
Here I have to return to Jaron Lanier's comment that the Turing Test can be considered two ways: either the machines are getting smarter or the humans are getting stupider. I certainly *hope* that hotel wouldn't be indistinguishable from an actual (sterile) hotel room experienced by a biological human, because this would imply that our perceptual and conceptual abilities had really atrophied.
Why do you think this exactly? Of course a modern simulation of a room wouldn't be indistinguishable from a real room, but that's because we lack the computing power to simulate all the objects in the room at the same level of fine-grained detail that human perception can resolve. If we could simulate the pattern of bumps in an object's surface down to something like the micrometer level, simulate variations in the way different parts of the surface scatter light at that scale, simulate characteristics like flexibility so the object would respond realistically to applied forces, and respond with its own forces on the simulated flesh of the upload (which would determine how the simulated nerves responsible for tactile sensations would fire), what do you think would be the visual or tactile giveaway that this wasn't real? A realistic Newtonian physics simulation of everyday objects at the micrometer level, rendered in realtime, isn't possible today because we don't have the computing power, but if we did I don't think writing the software would actually be all that much of a challenge, as I said it's basically just some Newtonian physics and optics. Even if we had the computers and the software, as non-uploads ourselves we might lack the right sort of interface to feed realistic tactile data into our bodies, but at least at a visual level I think we could produce images of a simulated hotel room which were indistinguishable from images of a real one, even if we were free to move the camera around and had something like a simulated stick that could be used to poke things to move them around and see how their shape changes without getting any tactile feedback. Again, the main difficulty here is the amount of computing power that would be needed, and beyond that a some more research into the fine-grained details of the way particular objects (say, carpet fibers) bend and twist in response to applied forces would be necessary, but it would be an extension of modern (much more coarse-grained) simulations of physical objects and substances like the ones you can see here and here.
We could spend the rest of this day brainstorming all the things are person in a hotel room might see, touch, smell, taste, hear, or otherwise be aware of
But you don't have to design all these sensations a case-by-case basis, you just need a general set of Newtonian laws governing how objects reflect light and change shape and create vibrations when they move (both in other objects and in the air), and a sufficiently detailed description of the properties of each object in the room (very small details of surface texture, light scattering, and responses to physical stress). If you want to make it an immersive environment for an upload, of course you'd also need to simulate the physics of their body at the same level of detail--the same rules that tell you how other objects deform in response to stress could tell you how skin, fat, muscle etc. deform when any other object in the room is pressed against them, and then you'd need to know the way that sensory nerves in the skin and muscles respond to such deformations. But compared to simulating an entire human brain, I'm pretty sure this would be a much simpler task that would require much less microscopic detail.
You don't think there could be great advantages to humanity if technical problems which might otherwise take a hundred years to solve could become available in a much shorter time?
I don't. What's the hurry?
There are all sorts of problems where the longer it takes us to figure out solutions, the more we irrevocably lose--environmental problems are a big one, finding ways to save human lives is another.
I think most physicists would say there is nothing fundamental that would prevent you from predicting the behavior of water molecules from basic quantum-mechanical laws, just that the computations get more complicated the more interacting particles you have.
Yes, but we already know the principles of water (at a macro level) from personal experience. This is what I mean by a posterior explanations. Given water, and a chemical understanding of its molecular structure, we can see how it is derived from oxygen and hydrogen. But that is not the same as saying that, having never experienced water, we could predict it from the properties of elemental hydrogen and oxygen.
The article seems fairly clear that they were able to simulate the behavior of water molecules using only the fundamental physical laws which are thought to govern all particles at the quantum level. I'm pretty sure no a posteriori knowledge about anything specific to water was used at all, and these fundamental quantum laws were not found by studying water. It seems to me you are really failing to understand something basic about reductionism if you don't see that the whole point is that specific properties of complex structures like water molecules are understood to be uniquely derivable solely from universal laws which govern the parts they're made of (in this case the protons, neutrons and electrons that make up both hydrogen and oxygen atoms). This is very much the mainstream belief among scientists in every field.
But I never said that observing speech would be enough to deduce the patterns in the brain, just that speech was caused by the output of pulses from the brain to the motor neurons that control muscles. The output of individual neurons would be expected to be a lot simpler to predict based on their electrochemical inputs, and experimenters can take also try to break down all the different changes in neurons (like changes in the density of neurotransmitters at the synapses) which influence this input/output relationship for an individual neuron.
What I think you're still missing is that if you reduce the subjectively experienced activity that is associated with a speech act (complex thoughts and feelings) to the objectively observable data of mouth and jaw musculature and production of auditory phonemes,
I never did anything like this. The subjectively experience presumably has to do with all sorts of brain activity that you can't deduce from the speech act itself; my only point was that the output from the motor neurons is determined by the activity of the neurons in the brain, so if you simulate the entire brain accurately, presumably the motor neuron activity will be accurate to and thus the upload will be able to carry on a normal intelligent conversation.
and then use only that data to correlate to specific neural patterns perceivable in the brain
What do you mean by using speech data to "correlate to" patterns of activity in the brain? Again, if you're suggesting that we could somehow deduce the brain activity from the speech acts themselves, that is definitely not what I was saying.
then all you have demonstrated is a correlate between physical speech activity with certain "functional" neurons. You haven't demonstrated a correlate to thought or consciousness.
Again, subjectivity is a philosophical matter--you have no way of knowing whether other people are really conscious either, but you take the fact that they use language in a way that is similar to the way you use language to express your own subjective states as evidence that they have similar subjective states of their own, so it would be reasonable to do the same for an upload who behaved like a biological human.
You write that "speech [is] caused by the output of pulses from the brain to the motor neurons that control muscles." That's misleading, in that the causal chain is highly indirect at best. Speech comes with a bundle of thoughts, feelings, motivations, and other "hidden" conceptual activity.
This sounds a little like Cartesian dualism, where some nonphysical mental entities like the "soul" are able to influence the behavior of physical things like muscles. In the normal scientific picture of the universe, physical events have physical causes (possibly including a random element). I suppose one could use "thoughts" and "feelings" as a shorthand for certain high-level patterns of physical activity in the brain, rather than using them to refer to subjective qualia, but then we'd get back to reductionism and the fact that the behavior of high-level physical entities like the brain or water is thought to be reducible to the interactions of the smaller physical entities that make it up; high-level entities are not thought to have any independent causal powers of their own.
If your entire argument against mind uploading depends on the idea that the physical world in not "causally closed" (physical events having physical causes), or that reductionism is not an accurate picture of the way causation works in the physical world, then there is probably nowhere further this discussion can go; just as I wouldn't know how to change the mind of an "intelligent design" advocate who has a strong conviction that their must be something more to the origin of living forms than natural selection operating on random variations, I also wouldn't know how to change the mind of someone with a strong conviction of the wrongness of the reductionist paradigm which has been so successful throughout the sciences. But the record is pretty clear that non-reductionist alternative ideas like vitalism and intelligent design have consistently failed to produce any useful research, and suffer from the "god of the gaps" syndrome where anything that currently isn't understood is ascribed to mysterious non-reductionist forces, and when it turns out that a reductionist physical explanation for that phenomenon is found, the non-reductionists simply point to some new phenomenon. I'm reminded of this section of Kenneth Miller's book Finding Darwin's God:
I hope you would at least be willing to acknowledge that if the reductionist picture of physical causation was correct, then the idea of simulating a brain in a computer would become at least somewhat plausible, even if you don't in fact believe in that reductionist picture (but I hope you would also be willing to acknowledge that rejecting this picture puts you out of the scientific mainstream).
If any given utterance is a result of specific neural activity, all the associated cognition must either (a) also be a result of that same activity, in which case our simulation is missing a lot of important data, or (b) be caused by some other neural activity, in which case we have nothing to map to.
I don't see how a detailed correlation of thought and neural activity could ever be established. We can't possibly report the full content of our internal dialog in real time, and even if we could this would not reveal any thoughts and feelings that were not fully conscious, which very likely outnumber our conscious mentation by several magnitudes.
In the reductionist picture, any high-level activity of the brain emerges entirely from the local interactions of the parts that make it up. So if we had found rules that let us accurately predict how any individual neuron behaves in response to local physical influences at its boundary (like incoming electrochemical impulses from neighboring neurons), then if you accurately map every neuron in the brain and simulate each one's interactions with its neighbors using these rules, all the appropriate high-level patterns would just emerge from this simulation "naturally" without our having to have any independent understanding of these high-level patterns whatsoever. Similarly, in the water simulation project as I understood it, there was no need for any high-level knowledge of how water molecules should behave, instead the researchers just took the general quantum rules governing the behavior and interactions of the basic units making up atoms such as electrons, and the appropriate water molecule behavior emerged naturally from this simulation.
For a map to be meaningful it has to be representational (that's practically redundant). If we are going to map the brain and hope to capture everything relevant to "mind" then we need to be clear about our signifiers and signifieds.
Not if you're a physical reductionist. High-level properties of the mind like "signifieds" are understood to just emerge from the collective interactions of large numbers of neurons (or molecules or whatever the simplest unit you want to look at), you wouldn't have to plug them into a simulation separately, just like you don't have to plug in high-level properties of water into a simulation which accurately simulates all the electrons and other basic particles that make up water molecules.
Posted by: Jesse M. | Jun 18, 2008 2:07:25 PM
I hope you would at least be willing to acknowledge that if the reductionist picture of physical causation was correct, then the idea of simulating a brain in a computer would become at least somewhat plausible, even if you don't in fact believe in that reductionist picture (but I hope you would also be willing to acknowledge that rejecting this picture puts you out of the scientific mainstream).
Really? Are you making an assumption that all that there is is knowable? There is nothing beyond the reach of scientific observation? Astrophysicists have been a bit confounded recently by the disparity between the rotational speeds of galaxies (including our own) and what the newtonian models insist. What to do? Well, it must be, they muse, that there is something else there that cannot be detected.
If something cannot be detected on a galactic scale, but that must be there because of the very subtle effects it has that were discovered only recently (by fluke?). Why not here? Reductionism may very well be able to in theory predict the hololgraphic patterns held in the brain that confer identity and memory but to what level? All of it (how would you ever know)? Or just the parts described by what we determine is deep enough (and what would we end up leaving out)?
This whole notion, that we can be perfectly modeled by perfect process based on perfect understanding and live on with perfect realism due to perfect imputs is, well, perfect nonsense.
The uncertainty principle is an initial limiter of how much can be known at the smallest levels. Is this an issue? Can we really know enough to be perfect in your new science? Old science has no such illusions.
Is good enough really you? Or do you want all of you to "survive."
What percentage of identity loss, memory error and sensory degradation is your hypothesis falsified by?
Posted by: Carlos | Jun 19, 2008 8:30:35 AM
Not quite. I'm saying that I don't believe this "right pattern of real physical nerve impulses" can be made to exist.
But the objects themselves still have to be programmed in! And all of their properties, imagined down to whatever level of detail the upload is likely to be curious about, or stumble upon accidentally.
Right now I'm looking at the nearest wall in my apartment. The brushstrokes in the paint finish on the molding is incredibly complex and detailed. Even the wall itself shows irregular marks from the roller, and streaks where the paint isn't quite thick enough to cover the primer coat, if you look closely.
The door hinges are painted over and there's a small glob at the bottom where the paint dripped to before drying. There are a couple of chips in the paint on the hinge, each owing its unique shape to the structure of the paint around it, which was unevenly applied by a human hand. Further over there are some scuff marks close to the baseboard, and a light layer of dust on the beveled edge. I'm also just now noticing that the painter's didn't mask off the baseboard when they painted the wall, and there is a very uneven edge separating them.
Granted, a sterile hotel room will be less, um, charming than my 85 year old urban apartment. It will be more uniform, with fewer quirks. But this is real life. This is the type of detail a person notices with a moment's conscious attention. Any simulation purporting to be realistic would have to account for a near infinite accounting of detail. To say that a shag rug or shower door is just physics is to invoke a universe of pure Platonic forms, with no history, and no contingency, and I would argue that any human upload would spot this deception a mile away.
I'm not arguing with this. But uniquely derivable and predictable are not the same things. Given an object or phenomenon, we can analyze the forces and properties that brought it to pass. But that's not the same as anticipating new objects and phenomena, since the laws of nature are ideals which never actually exist in isolation. Water freezes, for example, at O degrees C, unless it is salty. To predict (Or simulate) the world, you'd need to know they degree of "saltyness" (metaphorically speaking) of every part of it. All the contingent factors. Since the world is undergoing constant change, this is quite literally an impossible task.
But how can we know that we're "simulating the entire brain accurately" if the only empirical knowledge of the functional brain is acquired through the study of behavior? If subjective experience (thought, consciousness) is excluded from from the definition of the brain's function (which after all, must be separated from the "nonessential" properties of the brain, or no simulation can take place), then why would we expect it to reappear when the simulation goes online?
This is a very important point that I don't think you've grasped yet. It is one thing to say the mind is a product of, or the result of brain activity as a whole. (I don't totally buy this because I think mind arises out of brain-body-environment, and the subjective correlation of the self or the consciousness with the brain is a fairly recent one in human history. But I'll let that pass for the sake of this argument.)
It is entirely another thing to say that there is an essence of mental activity that can be reduced or subtracted from the biological brain, in the form of algorithms, and uploaded to a digital platform without suffering any significant loss of human identity. For this there is no precedent, and no empirical support.
The upload hypothesis is severely flawed by a confusion of symbol and reality, the latter being everything that is, and the former being the statements we have so far learned to make about it. To put Carlos's point into different language, our representations of the world can never be the world, because it is the nature of symbols to be incomplete. Their value is in being finite and manipulatable. As successful as science has been for our culture, it cannot transcend this limitation, because it too is a symbol-based enterprise, with currencies in math, human language and computer language.
I could be wrong on all counts, and we may be seeing conscious machines in my lifetime. But the theories on which these models rest are extremely hubristic about our ability to simulate nature, and no amount of appeals to "reductionism" is going to change that.
Posted by: Chris Schoen | Jun 19, 2008 12:56:21 PM
Post a comment