Why You Can’t Replace Your Brain with a Computer Simulation of a Brain
There are several related ideas that have been discussed on this forum that I’d like to address in this post – among them:
- If you produced a sufficiently accurate computer simulation of a brain (or a human), you could produce a new conscious entity in the world (as happens when a new human being is created biologically).
[*]If you produced a sufficiently accurate computer simulation of a populated world, the beings in that world could be conscious and would believe themselves to be living in the world that is being simulated (and, therefore, it’s possible that we ourselves, and our universe, are merely simulations).
[*]If you had a computer running a sufficiently accurate computer simulation of a brain, you could hook that computer up to a robot body to produce a conscious robot.
As interesting as these scenarios may be, they all suffer from fatal flaws which make them impossible.
Now before describing these flaws, I want to make it clear that just because these scenarios can’t happen,
this does not mean that it is impossible to build conscious machines.
To understand why, we need to first distinguish between simulations and replicas.
These terms have a range of meanings, of course, but
for our purposes here, we will use the term
replica for anything which mimics all the physical features of a given system at some functional scale.
So a perfect clone of an animal would be a replica of that animal, and perfect dollhouse furniture pieces are replicas of the full sized originals.
So what is the difference between that and a
simulation?
Well, for instance, I could use a computer simulation of a piece of furniture, say a chair, to help design and build a replica. I could do this by programming the computer so that it is able to display detailed images of the chair from any angle (even cross-section), allow me to view it with different stains and fabrics, tell me how much material I’d need for various sizes of replicas, how much weight they could tolerate, and so forth.
But I can’t sit in the simulated chair, or actually put it into my daughter’s doll house, or upholster it with real fabric.
That’s because only the machine running the simulation (which we’ll call the
simulator) is independently physically real (more on this later).*
And while the physical qualities of the replica are identical to those of the original, the physical qualities of the simulator are not – and this is the crucial difference between replication and simulation.
* I use the term “independently physically real” here to distinguish objectively real things and events from things which exist symbolically (like the “bear” in the emblem of the city of Madrid) or in the imagination (like my recollection of a bear). The seal is real, my brain is real, but there are no real bears in either one.
OK, so that being the case, how do simulations work?
After all, if a simulation can be so accurate that we can use it to predict the behavior of a corresponding replica (like how much weight a chair can take before breaking) then there
must be a real correspondence between the simulation and the system it’s simulating (which we’ll call the
target system or just the
target).
Doesn’t this make them, in some sense, at least in part
the same thing?
Actually, it doesn’t.
To explain this, we’ll use a bit of jargon –
physical computation and
logical (or symbolic) computation.
A computation is simply a change in the state of a system which follows a set of rules.
Because the behavior of our universe can be described in terms of rules (the so-called “laws” of physics) we can describe real events as
physical computation.
The melting of ice, or salt dissolving in water, or pool balls ricocheting off the sides of a billiard table – all of these are changes in states of systems, which can be described in terms of rules, which is what makes physics and chemistry possible in the first place... without regular “laws” no changes in our world would be predictable.
(For more on that, see the work of
Stephen Wolfram.)
Logical computations, on the other hand, are symbolic. Simulations are created by setting up a system of
physical computations in such a way that they
mimic the computations of another system, real or imagined. And if we agree on what the states of the simulator are supposed to represent, we can “read” the machine in order to understand something about the target system.
This is what allows us, for example, to set up an abacus in such a way that the movements of a few beads can represent the addition and subtraction of numbers much larger than the number of beads in the abacus.
And this is what allows us to add, subtract, multiply, and divide large numbers using mathematical symbols on paper rather than, say, using quantities of rocks and counting the groups we end up with.
And this is what allows us to set up rooms that make us think we’re flying airplanes.
Computing machines
In the early 20th century, Alan Turing developed a hypothesis regarding symbolic computation and “human computers” — that is, men and women writing out calculations by hand. His idea (or one of his many brilliant ideas, anyway) was that any calculation (symbolic computation) performed by a human computer could also be performed by a mechanical computer, even though the machine did not consciously “know” or “understand” the rules or the symbols.
It was a beautiful and wonderful and heretical and audacious idea... that a dumb and blind machine could perform such tasks. This idea is now taken for granted.
(Some folks have since misinterpreted Turing to assert that a mechanical computer can do
anything a human brain can do, but this is
not Turing’s notion and is not derivable from it.)
In short, as long as the machine is set up to move from symbol to symbol in the same way the human calculator does, then it will end up with the same symbol(s) at the end of the process as the human will, without having to know what they represent.
So like an abacus, any computing machine – including those powered by electricity rather than finger muscles – if properly set up can get you to the right symbol. And as long as you know what the symbols mean and set up some hardware to display them, you can read the machine to get your answer, or you can set up some other hardware to make something happen in the real world as a result of the configuration of machine parts you end up with (whether that’s painting a car door or driving a rover across Mars).
Modern computers can take this idea farther and produce what we can call
intuitive symbols – that is, symbols designed so as to be naturally readable by any human, such as a pattern of lights on a screen which looks like a chair to a human eye (but not to all animal eyes because we do not bother to design our screens that way) or vibrations of a speaker cone which sound like a babbling brook to human ears (but not to all animal ears).
In short, the simulation is set up by a programmer (or some other designer, as in the case of the abacus) so that the
physical computations of the simulator, which are
not identical to the physical computations of the target system, are nevertheless similar enough in certain aspects that you can judge the state of the target system by observing the state of the simulation.
The necessary components
Now one thing you might have noticed is that whenever we talk about simulations, there are always 3 elements involved – the target, the simulator, and the designer/reader.
Interestingly, this is not true for replicas, which have only 2 necessary components.
Many microscopic organisms, for instance, reproduce by a process of replication. In this case the original organism serves as target and designer (as we’ll see, this cannot be true of simulations) no reader is necessary and the replicant serves as both simulation and simulator (as we’ll see, this also cannot be true of simulations).
This difference arises necessarily from the fact that the physical qualities of the simulator, unlike the replica, must differ from the physical qualities of the target of the simulation.
This is why a simulation of a tornado won’t damage anything inside your computer, but a replica of a tornado produced in a storm box can damage objects inside the box.
But wait a minute now... doesn’t the simulation have to be real in some sense?
If it weren’t real, how could we use it?
This is true, the simulation is “real”. But
it doesn’t exist in the simulator.
It can’t exist there because the laws of physics forbid it.
All real things – all things or events which are locatable in space and have a duration in time – can be described in terms of matter and energy.
And if there are two real events A and B (all things can be described as events – even your sofa is simply the event of the molecules which make up your sofa retaining a particular relationship during a period of time... well, something like that, at least, we won’t dive all the way down that rabbit hole just now)
you cannot cause both A and B with only enough matter and energy to cause either A or B alone.
OK, but what if A and B are identical? Isn’t the simulator identical with the simulation?
The answer is no.
To illustrate why, let’s consider this claim: In my garage, there is a truck, there is a vehicle, and there is a pick-up. But if you look in my garage, you see only one automobile.
This can be true because if you ask me to describe the truck, the vehicle, and the pick-up, you’ll get an
identical answer.
On the other hand, if I were to describe one in terms of bits and voltage potentials, and another in terms of funnel clouds and wind speeds, there would be an obvious problem. And this is precisely where we end up with the simulator and simulation.
Where is the simulation?
It’s similar to the situation of watching Lawrence Olivier play Hamlet on stage. Are Olivier (a 20th century British actor) and Hamlet (a 13th century Danish prince) both on stage?
No, only Olivier is on stage.
But we agree he’s playing Hamlet, so where is Hamlet?
Hamlet is real – but
not independently real. He is only real as a state in the brain of an observer in the audience. He exists as an act of imagination, which involves matter and energy in the brain.
For the Shakespeare buff, Hamlet is real as an act of imagination... real because the behavior of imagining really happens (requiring no actual Danish prince). But if the performance is a free dress rehearsal and in the audience is a recent immigrant who speaks no English, never heard of Shakespeare, is from a country with no similar theatrical tradition, and who just wanted to get out of the rain, for him there can be no Hamlet, either real as a man or real as an activity of his brain cells.
The same is true of our simulation, which as we have noted cannot be a simulation without a programmer/reader.
Simulations begin and end with imagination
Remember, the target is not involved with the simulation. Tornadoes don’t set up computer simulations of tornadoes.
Instead, a programmer or designer sets up the system according to his/her
ideas about the target (even if that target is the designer), which may itself be either actual or imaginary. And the reader uses the simulation to change the state of his/her brain. And the simulator itself is not identical to the target (or to the brain of the programmer/reader).
But wait... can’t the simulator itself be considered the reader? Can’t we get the programmer/reader’s brain out of the picture that way?
As it turns out, no, we can’t.
That’s because
it’s impossible to design a simulation in such a way that it must represent exactly one target. Which means a simulator-as-reader could not determine which of all the possible scenarios it was supposed to be simulating.
In fact, even if the simulator were a conscious machine with perfect knowledge of its own “body”, it would be unable to determine – from its knowledge of its body alone – even the simple fact that its bodily functions were supposed to represent something to someone in the first place. (That fact exists only as states of other brains.)
The perfect simulation
But wait, what if the simulation is perfect? I mean right down to the quarks. Wouldn’t
that limit the simulation to one event?
Well, no, for a couple of reasons. (We’ll ignore problems with the idea of such a feat in the first place.)
First, let’s say we run a perfect simulation of a small river system. Unless you already know that’s what the simulation is supposed to represent (or at least that it’s supposed to represent something meaningful or useful to humans) there’s no reason why the target’s temporal and spatial dimensions should be applied to the sim. It could just as well represent a system in which height, width, depth, and time were all flip-flopped.
Sure, you could program an environment to fix these values, but then that environment would have the same problem.
The second problem is that there’s no way to determine from the simulation (without bringing in outside knowledge, which drags the observing brain back into the system) that the simulation is indeed perfect and complete, and that it’s not instead an incomplete simulation of any number of larger systems.
(It also seems intuitive to me that it should be mathematically impossible to have enough information in the system to explain the system to itself – each addition of explanatory information would require additional explanatory information to explain it, which would also require additional explanatory information, ad infinitim – but I don’t have a cipher for it.)
So no, we can’t use the simulator itself as a reader, even if the machine were perfectly conscious of the state of its parts.
The fallout
The upshot of all this is simply that
the simulation must always exist only in the imagination of designers and readers.
The simulator has its own independent existence, but the simulation cannot exist in the simulator. Matter and energy in the brain account for the reality of the simulation.
This is why even a perfect simulation of a brain will not, and cannot, produce a new conscious entity in the world, the way that creating a baby does. To do that, a mechanical
replica brain is needed.
And this is why even a perfect simulation of a populated environment will not produce conscious beings “in the world of the simulation” who believe they are living inside the target world.
Even if it were true that our universe was itself a simulator machine produced by some hyperdimensional race, we as parts of the simulator would have no way of knowing what it is that our universe is intended to simulate – only the hyperdimensional designers and readers would know that. If we’re parts of a simulator (we cannot be parts of a simulation) then what we observe is only the simulator, not the simulation.
And this is why you can’t put a computer running a simulation of a brain inside a robot and expect it to behave like a brain.
Whenever you replace a physical part with another physical part, only the physical computations matter to the system, and the physical behavior of the simulator
must be different from the physical behavior of the target system.
You might as well say you can replace your liver with a computer simulation of a liver.
But couldn’t you just take the symbolic outputs and convert them back into physical outputs, and voila?
The problem with this approach is that it means that none of the physical computation, the real work performed by the object, is getting done during the segment where you’ve made the replacement.
And we know consciousness is performed by the brain itself. It is not an impulse which exits the brain, bound for somewhere else in the body. It is a biological function of the organ, so replacing it with a machine that does something else (e.g. runs simulations of things or plays Tetrus) would mean that the function is no longer performed, which means no consciousness.
And since we know that running the simulation doesn’t make the machine conscious, we cannot get consciousness that way either.
Consciousness is a real event. However it’s done by the brain, only a functional replica of the brain will also perform that behavior.
But wait a minute... what if the brain is a computer? Wouldn’t it work then?
Well, if that were true, it could work, but only if the computer were set up to run like your brain, and since your brain isn’t a simulator machine, you’re right back to needing a replica, not a simulation.
And besides, your brain isn’t a computer – or at least, not the kind we think of when we use the term “computer”.
But that will have to wait for a later post.
