The Mirror Conundrum

Post Reply New Topic

Since there's a thread discussing AI, and another discussing singularities in technology, I thought I'd throw out a puzzle that is strongly related to both of these issues. You don't even have to solve the conundrum - showing that an answer (either yes or no) exists is sufficient.

The problem is this. You have volunteered to take part in an experiment. The first part of this experiment involves using all kinds of devices to examine the structure of your body, the makeup of your DNA, the signals and chemicals present in your brain, etc. This is all far enough in the future that all this is possible. This data is used to produce two clones - one physical and one virtual. The virtual clone exists in an electronically simulated universe that simulates all of the known laws of physics and models everything that is happening. However, due to the sheer volume of data, it can only simulate a single room with a computer terminal, and a smallish set of objects that can be used in experiments. These objects were selected by you before you were cloned.

Your clone will also be placed in a room. This room will be absolutely identical to the virtual one and contain the same set of objects.

You, too, will be placed in your own room, again identical to the other two and with the same set of objects.

You, your physical clone and your virtual clone are wheeled into your respective rooms when you are all asleep. The doors are locked. Necessary facilities are available in side rooms, which are present in all three cases.

Your task. Using the objects selected and the terminal (which gives you the ability to videoconference with one or both other rooms), can you find a way to determine which of you three is the computer simulation? (At least two out of the three of you would need to come to the same conclusion based on the same data.)

Now, you all start from a theoretically identical place, but all three will diverge fairly rapidly, the same way identical twins do. The factors influencing the three of you will be quite different - two because you're in a different point in space, the third because almost all of space does not exist in the simulation at all. The differences in nature between physical and simulatable time are going to be femtoseconds in places, perhaps microseconds in others.

The challenge, though, is not to establish that there's a difference between the three. That's guaranteed. Figuring out which is the simulated you is not so easy. But can you do it?

Now, I don't expect an answer on how to figure it out, though if anyone posts that it would be cool. No, this is much easier. It would also be nice if you could come up with a yes or no answer, but I'm really not even asking for that. All I want to know is whether the answer is going to be either a yes or a no.

Well, isn't that obvious? It's a yes or no question, so the answer has to be one of them, right? Wrong. There are entire classes of problem that can be described but for which it is impossible to know whether or not an answer actually exists. Other problems will have multiple valid solutions, but they are also mutually exclusive. It gets really fun when mutually exclusive solutions of this kind exist all at the same time.

Just find a way to leave the room. If you do, then you know you're not a computer simulation, since the simulation can only simulate that one room.

I side with him on that on.

The door's locked. And if you insist, bolted and made out of laminated titanium.

Ok, now what?

Yes and you don’t need any objects.
Just agree to all get lost. Since there are no temporal constraints on the sim he will have to travel back to a time to when he doesn’t exist and you and your clone won’t see him.

Wait 4-7 days. The first two to die of thirst are the real ones.

Or if you insist there is water, then throw the water on the keyboard, and the ones whose keyboard shorts out are the real ones.

One of my selected objects would be a radio receiver. (If I was feeling especially Macgyverish I might not even request this as one of my objects; I might fashion a crude antenna out of the terminal's power cord or something.) Is the simulated environment going to include radio broadcasts emanating from an "outside" which isn't part of the simulation? (If so, what if I request a transmitter too and attempt to interact with hams outside the room?)

If I can't do that -- let's say it's specified that the room is a perfect Faraday cage -- I might try to devise some physics experiments which would have easily verifiable results but be computationally infeasible to simulate.

Its occured to me he'l say the sim will hae a fake death. or whatever. However, in a morally concious society, the people locked in the room will not be permitted to die. And the sim cannot simulate anything but the room. So just have everyone refuse food/drink till they pass out, then the two who are revived by paramedics are the two real ones.

If you have a long string of particles in quantum superpositions and you collapse them all, it will result in a truly random string. Then it's just a matter of distinguishing this from the pseudorandom string generated by the virtual clone.

How would you tell which of them was truly random and which was pseudorandom? If the simulation is any good, the pseudorandom should be statistically almost indistinguishable from genuinely random and would never repeat.

However, this may be a key element of such a test.

One of the fun things about this is that the brain creates a simulation within itself. You do not directly experience the data from your senses, you only experience the updates that those senses have made to the simulation. However, the nature of the simulation must be very different from the "real world" simulator I am proposing the AI lives within.

The other fun thing is that this thought experiment places very strict limitations on how a computer could be made intelligent according to the Strong Turing Test, precisely because what you are doing is trying to break that test, Indeed, the Strong Turing Test is worded in more or less the same way - being at a terminal and trying to decide who is the person and who is the machine. The biggest difference I've added is that I've made all three people initially identical, so as to eliminate some of the easy ways of falsifying the non-human.

However, it is possible to imagine extending this further - allowing the simulation to be updated from external stuff such as radio signals and so on.

Now, there are people who argue that computers cannot ever become intelligent - that true intelligence is non-computable. If that is the case, then no matter how sophisticated this experiment got, you would ALWAYS be able to tell which was the machine. If, as seems likely, you can only infer it from a breakdown in the data available to the simulation, I would argue that true machine intelligence is not only possible but inevitable -- but quite possibly requires the same layering that human intelligence needs.

Does that sound right?

Any pseudorandom sequence, no matter how complicated, could be described as some sort of formula. So yes, it could definitely be detected given enough computing power. Of course enough might mean a googolplex times what we currently have, but it is calculable.

As for AI, that depends more on the human brain than on technology. We already know what technology's capable of, and we have a pretty good idea of what it might be capable of in the future. The human brain we know somewhat less about. We know it consists of neurons that fire on certain conditions and that the brain as a whole has a certain connection topology. If that's all there is, then we should be able to map it and reproduce it as a microprocessor. But some people think there's more, for example free will or a soul or whatever they may call it. So it really boils down to a more metaphysical question.

Which is the point of the test! If we take the neurons, connections, chemistry, etc, and duplicate them exactly, then mimic all of the possible processes and interactions that the brain may have with the external universe, you will either have something that is identical or different. It can't be both.

Now, if the brain requires truly random quantum events to function (and some people have argued that it does - "Emperor's New Mind"by Roger Penrose follows that line of thinking), or requires something metaphysical like a soul or spirit then true AI is impossible because that cannot be simulated. A direct test is then possible to establish who is the AI.

If the brain does NOT require anything outside of the purely physical and deterministic, then only an indirect test would be possible. No direct test could ever exist, no matter how good or sophisticated, that could distinguish between the person and the AI, although you may be able to test the AI's universe and see if it is simpler than your own. (One of the reasons for confining to a room was to limit the ways you could test for simplicity.)

We don't have a good idea of the way the brain works, but this thought experiment tries to fudge that by assuming that anything quantifiable is quantified and anything deterministic is duplicated. It tries to get at the heart of the issue, which is whether there is anything inherent in the concept of intelligence that is non-computable - not by the philosophical musings Roger Penrose uses but by trying to produce an actual test that will produce one result for all real intelligences (including clones) and produce the opposite result for all AIs.

If that test doesn't exist, then the simulation must contain everything needed and therefore no metaphysical or non-deterministic element needs to exist. It might, I'm not arguing it doesn't, it just isn't needed for intelligence to exist.

smoke a cigarette!

having to calculate the position and velocity of each smoke particules will very rapidly exhaust the limited computing power of the simulated room.

I think that all 3 are in a simulation. Physics tells me that all matter/energy is made up of "information in space", that the only reason I consider something to exist is my perception of information (light,sound,touch..). In order to know how my environment was created, I would need to be able to see the simulation from an outside perspective. In this experiment there is one simulation inside another (the virtual computer world inside the virtual human world), maybe the only difference would be the concept of infinity (since the computer is more limited). The "signals and chemicals" in the computer clone's brain remember an infinite world, by closing it's eyes and trying to imagine something infinite, it should be the only one of the 3 to come to a limit that it can image.

Reminds me of some stuff David Deutsch wrote about, expanding on Alan Turing's work. Ages since I read it though, so I could just be talking baloney.

No. No test can distinguish them.

=========

I think outskirt has made a good point.

One does not need to make the simulation "infinite". There is a clear upper limit on the bandwidth of our (meaningful) sensory input. It must cope with that, and must be arbitrarily forced to cope with no more than that! We can't have the simulation working faster than the "original" and the (first) clone.

I see no reason to limit the simulation to a room, particularly, except that the clone would need a fully distinct cloned world, populated by other cloned individuals, otherwise the clone could meet the original "in real life".

So within your scenario, these days, I am pretty convinced the answer is "No". That is, there is no test that can distinguish the three entities. So far as I am concerned, they would all be "me".

Once we somehow managed to stop talking simultaneously, saying the identical things, we would diverge. I guess this will happen... as there is some level at which the duplication is not utterly perfect. However, that difference has nothing to do with any feature usable to distinguish one from another.

==========

And... as a final point, I'm looking forward to the days when some of my Is will be being simulated, and the simulation speed is not throttled down, nor are any of my simulations' other features (e.g. near infinite, perfect recall) being inhibited. I wonder what a million of me would be chattering on about?