I don't think it's clear. How long before our digital neural networks exceed the complexity of our human biological ones? It's possible today with a super cluster of sorts. That is to have roughly the same number of connections between neuron type thingies - it still wouldn't be able to match our brains functionally - that's another problem. And yes our biological neurons are way more sophisticated than what we use today for AI, but again that can be overcome eventually.

This suggests that it's possible one day to have computers some orders of magnitude better. If you look at it from first principles, of course it's possible. The brain is unlikely to be the most efficient design of neural network allowable in this universe. So given enough time, we'll learn how to build it better.

Then it's just a manufacturing and energy problem to match the number of human minds on the planet. So no, I don't think it's impossible at all.

Just ridiculously freaking hard, and not likely to happen in our lifetimes.

It's not a matter of it being "hard". Nor is it a matter of "complexity" (how many parts something has).

A simulation is a model which picks out a tiny subset of regularities in the target to model. There is an infinite density of such regularities to pick upon, because we are imposing the structure on the target in order to model it.

The target of the model has no "model structure" it has causal structure. That is, when light interacts with the surface of a mirror its interaction isnt "abstract", ie., some description. It is an actual photon interacting with an actual electric field, etc.

To "model to infinite density", ie., to have every single test that can possibly be applied to a model come out identical to that test of the target, the model needs to be just another example of the target.

The only thing which can be investigated in any way to behave as light hitting a mirror, is light hitting a mirror.

A digital computer is just an electric field oscillating across a silicon surface. It cannot be programmed into being a mirror, nor into being light.

Programming gives the electric field a "model structure". Chalk gives a blackboard a "model structure". Lego gives a bridge a "model structure".

Programming cannot not -- it is impossible -- give silicon the causal structure of light interacting with a mirror.

Model structure is actually just an observer-relative isomorphism: when the user of the computer (chalkboard, lego,...) looks at it, the user, is able to inform himself of the target by use of the model. To do so the user identifies certain aspects of the model with the target. The model is not at all causally alike the target.

No amount of lego will make a lego brain. No amount of oscillation in an electric field will make a thought. Neurological activity, and indeed every causal mechanism of the universe, is only described by a model.

Or, here’s another possibility:

If we knew the complete laws of physics (or perhaps just invented some self-consistent laws) we could simulate light hitting in a mirror in complete detail; and within the simulation it would be indistinguishable from reality.

But we can’t actually do that because it would take a ludicrous amount of computing power. And the structure of the laws of physics might be such that it’s never possible.

That’s very different from your strict model/causal distinction. I don’t know of any evidence to suggest which view is correct.

I think your claims are overblown. You might be right but you might be wrong.

Saying "within the simulation", begs the question and rather makes a mess of the issue.

The question is, at the outset, whether simulations are actual instances (eg., of thinking).

My claim is that they are not. Only when a human being looks at a simulation does it inform them of the target. The simulation isnt the same thing as its target.

There is no "within the simulation". A simulation is just an abbacus. A digital computer is just a fancy abbacus with a wood-to-LCD converter attached.

There is no "within the wood". It's just wood.

My claim is kind of orthogonal, that “actual instances” is a red herring and that the only thing that matters is observational equivalence.

Yes, but you're defining "observation" as some weird intra-wood process (, intra-silicon in the case of digital machines).

Observation means, for example, reflecting some light off the thing. Does a piece of silicon become transparent if I program it to model glass? No.

So, trivially, it is observed immediately not to be equivalent.

What you mean by "observation" is: can a person using the model system inform themselves of the target.

Do I get the message "lets visible light through: 98%" from the machine when I have correctly modelled the glass.

That message isn't observation, it's calculation. Calculation is what happens when I use a tool to inform myself about something.

That the machine emits the right symbols in a way that it is programmed to, so that I acquire accurate beliefs about the world, says nothing of the machine. The machine does not become transparent.

Please do explain how you would test whether you are being simulated or if you're an "actual" instance.

As in, perhaps there's an outer universe where our hairiest quantum physics are trivially solved. They are simulating a very simplified universe model, and everything we see is already simplified.

If you can prove it, you're far more clever than me. If you can't, then being "actual" seems to have no practical use.

I like how you put that. The abacus and the computer are the same in that the consume external power, and adhere to a set of rules. But there is a large difference between an abacus and a computer. A computer can modify the program that is running. It can change its own rules. An abacus cannot do that.

My argument against simulation is similar in thst certain algorithms for modeling physical processes can only run in exponential time. Protein folding is a good example. How could a computer simulation perform exponential time operations efficiently? It wouldn't work no matter how big a computer was made because the complexity would explode very quickly while reality can do it in real-time.

There is no need for the simulated time to be equally "fast" (whatever THAT means) as fhe real one. it is annoying and impractical for us observing the simulation from the outside, but it wouldn't make any difference from the view of the simulation (under assumption of a closed system simulation)

That's another good way of illustrating the problem, however this is still just a symptom of the underlying problem. Models do not have the causal structure of the thing they model.

The reason the model would require near infinite time to run is that its modelling a causal event, it isnt an instance of the same causal event.

When electrical fields oscillate they model a bridge. No amount of activity would ever make them "solid".

All "real" stuff is infinite. This form of infinity is really about saying that our descriptions cannot capture the "full depth" of the world. The world itself has no "depth". Models, rather, are simply partial descriptions of it.

To turn a model into the thing its modelling via increasing its descriptive power quickly introduces infinities: its impossible. To make a model "accurate" in this sense, you must actually just make another example of the target. Ie., build that bridge.

> How could a computer simulation perform exponential time operations efficiently?

Assume P=NP.

P currently not being NP is exactly the kind of thing someone would build into a simulation to prevent stack overflows.

(I don't believe in simulation because: why would anyone that advanced bother?)

Very interesting idea! It's not immediately clear, however, that "P!=NP" is something that you could build into a simulation.

(The demonstration that the lowest possible complexity of comparison-based sorting is O(n log n) comes to mind as a related example.)

Quantum computers can solve protein folding efficiently.

"Can" or "could in principle"?

"Can" as in quantum chemistry is one of the best applications for quantum computers.

Sorry for the naivety, I am not an expert of quantum computing, therefore I asked. I am quite interested in the topic, might you give me some references, please?

Googling for "quantum chemistry computers" yields a number of interesting results, e.g.

https://www.chemistryworld.com/feature/quantum-chemistry-on-...

https://www.technologyreview.com/s/603794/chemists-are-first...

Or if you like something by Feynman:

http://doc.cat-v.org/feynman/simulating-physics/simulating-p...

> No amount of lego will make a lego brain. No amount of oscillation in an electric field will make a thought.

We can never build a Lego brain that is indistinguishable from a physical brain because brain cells are distinguishable from Lego, but this doesn't mean that a computer can't think. We accept that distinct humans share the property of sentience even though there are observable differences between them. Why is 'thought' required to occur in a cluster of brain cells that share the physical and chemical properties of human brains in general, but not the specific properties of any one person's brain in particular? Using your terminology, what is the 'causal structure' of thought?

I don't think there are relevant difference between thinking things.

Thought is just a particular set of biochemical reactions occurring across particular kinds of biological systems (nervous systems).

When you get hungry, you start thinking of food. These thoughts are, literally, products of the innervation of your stomach.

I don't know what "thinking" is if computers are the kinds of things which can do it; I'd guess it would be nothing we are, in fact, doing.

Two pieces of glass may differ in size, but not in what makes them transparent. Two people may differ in all sorts of ways, but not in what makes them conscious.

That "consciousness" is limited to "people" results in a very limiting definition of consciousness. At some point, it's probably better to at least pretend that something you regard as a non-person is conscious, if only in order to make reasonably accurate predictions about the world.

I agree. Except I cant rule out a possibility that it may be possible to develop or evolve seemingly deeply intelligent virtual agents, and if/when I ever encounter such a thing I will not be able to be completely confident in treating its communications (perhaps it has requests) with no common respect for the living.

I've thought about what difference it could make whether such a program employs a PRNG or else reads a physically based entropy source - so it would be 'replayable' or else it would be to some degree unique and unrecoverable. That would seem to be a big philosophical difference yet there can be no noticeable difference between the performance of a PRNG and a real entropy source.

So I am partial to that step along the simulation thought experiment, but it requires a mysterious quality to be attachable to simulations which is not present in the popular account, where reality may be 'just a simulation'.

Only if simulations may somehow be a reality - the experiment becomes as mysterious as life when that issue is examined. But it almost never is examined, instead I see credence given to the idea that other people may be husks in the selfs own limited process, along with little awareness of how degenerate it would be to truly accept that -stepping stone in a sci-fi thought experiment.

> To "model to infinite density", ie., to have every single test that can possibly be applied to a model come out identical to that test of the target

So, what you are saying is that if simulated with increasing accuracy towards infinity density, as some point the models and imposed structures would break down, favouring one simulation routine over another?

Like, say, a physical phenomenon behaving like a wave in some set of circumstances, but like a particle in another...?

P.S. I'm perfectly okay in here, Elon. No need to pull me out to eat grubs, unless you promise to teach me kung-fu and French the easy way.

In the sense I mean, "simulating to infinite density" is just another way of saying alchemy, or magic.

If you can program silicon into being gold, then "programming" is alchemy.

Programming is about increasing the descriptive power of a model. The better the program, the more accurate it is.

Alchemy is about making X indistiguishable-in-everyway to Y. Lead to gold.

If you can make a model indistinguishable from its target, then you are claiming that programming can turn water in to wine. Silicon into gold.

Making models more accurate, does not turn them into what they model.

A model is just a abbacus. No movement of wood (, silicon, current,...) will turn it into a brain, which is biochemical system.

Unless the thing you are modelling is itself also a model. And by virtue of the limitations of our (extended) senses, everything is a model.

What you know as "gold" is really a model of gold. What we know about gold, the colour, specific weight, etc. in no way describes the actual thing and is only a representation as can be conveyed by our senses.

Everybody agreeing about the observable nature of a thing from a particular frame of reference does not imply that frame of reference is the one that is closest to the truth.

Anyway, as much fun as it is to think about such matters, they will, almost by definition, never be falsifiable, so yes, you are of course correct, even if only from the generally most useful frame of reference we call reality.

Sounds like (metaphysical) idealism.

The difference between idealism and realism here is meta-empirical. I would say idealism is false, and in fact, nothing is a model and everything is concrete.

To call the computer program a model is to say that when I look at it, I can use it to inform myself about the world.

It's an abstract property. The actual system in question is silicon and electrical current, etc. And thus shares nothing of interest with gold.

Given a choice between an airplane, a bird, a bat, and a dragonfly, which has the causal structure of flying and which are models or more-or-less poor imitations?

"No amount of lego will make a lego brain. No amount of oscillation in an electric field will make a thought."

No amount of oscillation in an electric field will make a thought as long as "thought" is defined solely as "the stuff that goes on in the goo that resides in the human noggin."

I don't think it's for certain, but anyone who has ever implemented Runge-Kutta on an n-body problem knows that simulating a "simple" system can be very not simple. Accurately simulating a chaotic system over a time t requires \Omega(t) memory because of Lyapunov doubling. As such I think any long-term simulation of the Universe would require new physics or be subject to error. The errors might be impossible to detect from within the system but they're still "there" in a formal sense. New physics could include a "pixellated universe" which has the same errors, or something else.

But, such a stimulated universe would be observed by its residents to have very odd behaviours caused by such hacks.

Like a maximum speed limit, minimum temperature, an uncertainty principle, and so-forth ;)

https://www.smbc-comics.com/comic/2012-02-29

Are those even hacks, though? They seem more like just how it is.

Are these limits necessary? Maybe these limits are just part of the nature of how matter functions, rather than being mere hacks.

We understand so damn little of this physical universe that we are in no position to say much, if anything, about it's true nature and origins.

You only have to simulate one mind. Who says that our world actually has billions of minds? It could just be a simulation for one person. The output of other people is much easier to simulate.

You don't have to model something as complex as our own universe. You only need to model something complex enough that the beings inside the model can't prove that it is a simulation.

EDIT: Also who says that time inside the simulation has to run the same rate as real time?

Even if you hit roadblocks from time to time, the general cosmic-scale conversion of "dumb matter" to "computation capable matter" can only be... exponential. Imagine all matter in a planet the size of Earth converted to computing infrastructure, considering we're only living on a thin thin crust on it.

When the evolution of intelligence goes beyond the stage of slowly evolving (if evolving at all) forms like humans, "life" will spread like a plague through the universe and its only purpose will be to convert all matter into "shit that can run the software that is we"... Any intelligence with a different purpose will simply be eaten alive but those with this one.

Only the second law of thermodynamics can stop that, and even that only "works" the way we imagine in a finite universe (and despite out forced rationalization and intellectual masturbations our brightest were capable of, we have no reason to believe anything is anything but infinitely infinite, whatever that could even mean). And we can't even imagine how the evolution of information in infinite space and time could unfold, even with "local containment" via the "light speed limit"...

An no, we are not unlikely to model anything such complicated, we are almost sure not to... Because "the children" will awake much sooner that this computational power will exist, probably doing the right thing of terminating bio-humanity as it's so (computationally) wasteful, and remnants of us will only endure in "historical entertainment simulations" thinggies... (And this is the optimistic scenario anyway, in which post-human life would retain some human-type characteristics by virtue of "descending" from us. If an alien superintelligence reaches Earth first it might not even care to analyze us well before restructuring matter for its own purpose, so paradoxically, developing superhuman-intelligence-that-will-terminate-bio-humans asap is probably "humanity's best bet" of "not being completely forgotten" / "transmitting our memes".)

What if it only had to model a single mind?