Singularity

[PixyMisa]

Now, the next step is where I personally have my doubts about the Singularity.

By my calculation, in 2010 we will be able to build a computer as intelligent as a human, using only established technology (45nm chips aren't in mass production yet, but they have been made) and existing knowledge (it's a cell-function simulation of the brain; we don't need to understand anything about, say, consciousness; all we need to know is what the cells do and how they are hooked together).

Aside: Obviously it will be enormously difficult to build and program this machine, but we can start work now, because nothing much is going to change.

The next step is this: In 2010, we can build a computer as smart as a human for $200 million. So by 2012 or so, when the 35nm process is in full production, we will be able to do it for $100 million... Or we can build a computer twice as smart for the same amount of money.

One of the key ideas of the Singularity is that once you have a computer twice as smart as a human, things begin to take off by themselves. The computer itself takes a hand in the research program. If it can come up with a modified neuron-simulation algorithm which is 50% more efficient than what we normal humans managed, then suddenly it is three times smarter than us. And then the new 3x computer can come up with better ways to build the chips or whatever...

Personally, I strongly suspect that we will run into the law of diminishing returns here. I think that a computer "twice as smart" as a normal human will be less than twice as capable at solving problems, and also that the problems associated with developing ever smarter computers will become less and less tractable. I could be wrong, of course.

 

[PixyMisa]

One other thing:

In case you're wondering why we aren't building such a machine today, well, if we start now it might cost a billion dollars and take 15 years, while if we wait until 2010 it would only cost $200 million and take 10 years. That's because we're already approaching the singularity.

 

[kuroyume0161]

Certainly the same order of magnitude, so no big deal.

That's why I stated it - this is the current consensus.

I'd be interested in a link, if you have one.

Look at Wikipedia:Neuron:Challenge to the neuron doctirine for a hint whilst I attempt to find the original sources.

Um, yes. Any current server processor can handle that. No biggie.

The largest memory modules currently shipping are 4GB. We'd need 150,000 of them. Though for the rest of the calculations I was basing my numbers on the 45nm process, so divide that by 4.

Yes, let's put these together. As I said, any current server or 64-bit processor can theoretically handle 625TB of memory (1024 TB max for some Intel 64-bit cpus). But there is no such system that can hold that much. As you note, and I am aware, the largest memory modules are currently 4GB (and they are atrociously expensive - as I've looked, having server systems of my own). No computer in the world is currently capable of actually using 625TB of memory, admit it (not in an addressing sense, but in the physical available memory sense). Even rack-mounted server systems can do maybe 168GB at best. And the worlds largest seems to not be available yet and still falls short: 13TB SGI supercomputer. Even my biggest computer only has 1/2TB of disk storage. I'd need 1250 times that to reach 625TB in hdd.

I don't see why you consider networks in general to be problematic. Inter-neuron communication is not exactly fast.

They are highly undependable and if concurrency is an issue, well, just look at any online multi-user game to see how difficult it is to keep even tens of thousands of users concurrent (with very specific data compressing techniques and the fact that most of the time not all data must be transmitted to every player) without lag (that word is known to any online gamer).

ETA: I was working with a figure of 5*10^9 neurons. The actual number is 100*10^9. So multiply any of my numbers by 20 (12500 TB! or 12.5 PB). And if we are to include the glia (50:1 ratio to neurons) as unitary parts, that number increases yet again to 5*10^12 (!!!). Need I even plug home the size of that number.

 

[PixyMisa]

Look at Wikipedia:Neuron:Challenge to the neuron doctirine for a hint whilst I attempt to find the original sources.

Thanks.

Yes, let's put these together. As I said, any current server or 64-bit processor can theoretically handle 625TB of memory (1024 TB max for some Intel 64-bit cpus). But there is no such system that can hold that much.

Well, no. It would be a custom-built system, but the components would be standard. That's true for all really large computer systems anyway.

No computer in the world is currently capable of actually using 625TB of memory, admit it (not in an addressing sense, but in the physical available memory sense).

It's simply a question of plugging enough modules together. There just isn't a problem here.

They are highly undependable and if concurrency is an issue, well, just look at any online multi-user game to see how difficult it is to keep even tens of thousands of users concurrent (with very specific data compressing techniques and the fact that most of the time not all data must be transmitted to every player) without lag (that word is known to any online gamer).

That's the internet. That has no bearing on networks in general. None. Myrinet, for example, has a maximum latency of two microseconds.

ETA: I was working with a figure of 5*10^9 neurons. The actual number is 100*10^9. So multiply any of my numbers by 20 (12500 TB! or 12.5 PB). And if we are to include the glia (50:1 ratio to neurons) as unitary parts, that number increases yet again to 5*10^12 (!!!).

I was using 10¹² as the number of neurons, when it's really just 10¹¹. But assuming the glia do something but are less complex than neurons, it works out anyway.

Need I even plug home the size of that number.

It's not very big.

I have about 3 terabytes of disk in my three computers at home. Only 6 gigabytes of memory, but then I didn't spend $200 million.

If we need a megabit of memory for each neuron (let's ignore disk, because it simply isn't fast enough to be useful) that's 125,000 bytes times 10¹¹, or in round figures 10¹⁶ bytes.

Right now, that would cost less than a billion dollars at retail prices. Obviously I hadn't allocated that much memory into the initial pricing, so we need to double it - $2 billion now, $400 million in 2010, $200 million in 2012-2013.

Oh, and it works out to 2GB of memory for each processor (2005 technology) or 8GB for 2010 tech where I assume the processors are four times as powerful.

I'm not saying you'll get one on your desk any time soon, but there's a difference between "it can't be done" and "it would be of comparable cost to other major science projects".

I haven't added in the glia, but we can do that too.

 

[Darat]

...snip...

Hell, we still have problems making an operating system that doesn't crash.

But that isn't a fundamental problem, after all there is a lot of evidence that our "operating system" crashes a lot yet we seem to muddle along fine.

 

[Ziggurat]

But that isn't a fundamental problem, after all there is a lot of evidence that our "operating system" crashes a lot yet we seem to muddle along fine.

No, it's not a fundamental problem. My main point is that our software is primitive and stupid, and that applies even when we make it robust, efficient and bug-free. Microprocessors may be speeding up exponentially, but our software complexity isn't. The current Firefox browser, for example, is way better than the original Netscape, but it isn't really much smarter.

 

[PixyMisa]

No, it's not a fundamental problem. My main point is that our software is primitive and stupid, and that applies even when we make it robust, efficient and bug-free. Microprocessors may be speeding up exponentially, but our software complexity isn't. The current Firefox browser, for example, is way better than the original Netscape, but it isn't really much smarter.

Human neurons aren't much smarter than bumblebee neurons, either.

 

[Soapy Sam]

My suspicion is that intelligence is an emergent property of complex switching systems. I suspect some networks probably already display intelligence. I also think intelligence is wholly useless and totally unidentifiable unless it can act in the real world.
I don't think we are capable of recognising machine intelligence and I don't know if we ever will be. Equally, I don't know if intelligent networks will ever be aware of us.

Let's face it- we can't agree on whether a dolphin is intelligent. How will we recognise it in a machine?

As for modelling the human brain, I'm sceptical. First, to model something we must understand it- can we say that by 2012 we will understand the function of every neurotransmitter, every protein molecule, every nutrient, every sub system and so on? To the level where we can reproduce it in electronic (or photonic) analogue? And if we can do that, can we make it run slow enough to do what brains do in real time? And if we do- why would we suppose that the output of such a system would resemble mental activity?
And if it did, would it be murder to switch it off?

For me, the singularity if it occurs at all is more likely to be due to convergence of neural processing with electronics / photonics. Give a human brain a 100% accurate long term memory. Use artificial subsystems to control musculature. Take the support load off the brain . Free the systems which are currently tied up keeping us balanced, upright, in pain, whatever and channel those into data processing , conscious or sub.
Inbuilt radio to provide telepathy. Make humans the network nodes.
That way we have a proven technology- organic and tested by evolution, enhanced and linked by IT. Alien . Spooky. Hackable.

Thank ed I won't be around . Though maybe they'll be able to store a human mind by then , so you can wake me up in 2105 and tell me again how wrong I was. Heaven or hell?

 

[PixyMisa]

My suspicion is that intelligence is an emergent property of complex switching systems. I suspect some networks probably already display intelligence. I also think intelligence is wholly useless and totally unidentifiable unless it can act in the real world.

That's reasonable.

I don't think we are capable of recognising machine intelligence and I don't know if we ever will be. Equally, I don't know if intelligent networks will ever be aware of us.

I'm not so sure that is.

Let's face it- we can't agree on whether a dolphin is intelligent. How will we recognise it in a machine?

The question isn't so much whether dolphins are intelligent, as how intelligent they are. They pass they mirror test, I believe. (That is, they recognise their reflection.)

As for modelling the human brain, I'm sceptical. First, to model something we must understand it- can we say that by 2012 we will understand the function of every neurotransmitter, every protein molecule, every nutrient, every sub system and so on? To the level where we can reproduce it in electronic (or photonic) analogue?

A valid question, but if not 2012, then 2015, or 2020. It's a question of time, not of possibility.

And if we can do that, can we make it run slow enough

?

to do what brains do in real time?

Yes. That was what my back-of-the-envelope calculations showed. It would be expensive, but it can definitely be done.

And if we do- why would we suppose that the output of such a system would resemble mental activity?

Why would it be anything else?

And if it did, would it be murder to switch it off?

I think so, yes.

For me, the singularity if it occurs at all is more likely to be due to convergence of neural processing with electronics / photonics. Give a human brain a 100% accurate long term memory. Use artificial subsystems to control musculature. Take the support load off the brain . Free the systems which are currently tied up keeping us balanced, upright, in pain, whatever and channel those into data processing , conscious or sub.
Inbuilt radio to provide telepathy. Make humans the network nodes.
That way we have a proven technology- organic and tested by evolution, enhanced and linked by IT. Alien . Spooky. Hackable.

And probably inadequate. The Singularity requires an exponential increase in the rate of technological progress, and that would only provide an incremental increase. Unless the technological advances can continually be channeled back to increasing the rate of technological advance, the Singularity doesn't happen. So it's not good enough to make people effectively smarter; you have to use the increased intelligence to find ways to increase their intelligence still further. Otherwise you get The Bump rather than The Spike.

Of course, I'm far from convinced that the Singularity will happen anyway.

 

[Darat]

To extend Soapy's thoughts a bit more. It could be that what we consider is our intelligence is very dependent on what it "runs on", it may be that we can't actually simulate it/model it to a sufficient degree to reproduce it. (This stumbling block doesn't mean of course there is any special "magic essence" involved in what we are, just that the map is not the territory.)

Following on from this (and what Soapy posted) does it make sense to use the same word "intelligence" for the "intelligence" a "artificial" entity might possess? Many science fiction authors (who use the idea of the singularity - whilst avoiding it head on) have attempted to show that these types of entities will be quite "unknowable" to us in the sense that we've always said "gods" are!

 

[PixyMisa]

To extend Soapy's thoughts a bit more. It could be that what we consider is our intelligence is very dependent on what it "runs on", it may be that we can't actually simulate it/model it to a sufficient degree to reproduce it.

I don't see why this would be so; it isn't true for anything else.

(This stumbling block doesn't mean of course there is any special "magic essence" involved in what we are, just that the map is not the territory.)

Yes, it does mean there is a special "magic essence" involved. If you say that a complete physical model of a system does not produce a model of the emergent properties of that system, then you are claiming that it works by magic.

Following on from this (and what Soapy posted) does it make sense to use the same word "intelligence" for the "intelligence" a "artificial" entity might possess?

Yes. It is intelligence.

Many science fiction authors (who use the idea of the singularity - whilst avoiding it head on) have attempted to show that these types of entities will be quite "unknowable" to us in the sense that we've always said "gods" are!

Well, there's two different things there. If you completely and accurately model the workings of the human brain, the output of the model will be the same as the output of the brain being modelled.

On the other hand, if you just whack a whole lot of computing power together and teach it to make inferences from observations, there's no telling how it's going to end up.

 

[Darat]

I don't see why this would be so; it isn't true for anything else.

Isn't it? Aren't most of even our current best models only an approximation and don't always give the same results as the "real" system do?

Yes, it does mean there is a special "magic essence" involved. If you say that a complete physical model of a system does not produce a model of the emergent properties of that system, then you are claiming that it works by magic.

That statement I can agree with e.g. "a complete physical model" but that is not we are talking about; we are talking about in effect making a model from "information" that is created in a very different way from how our brains are e.g. atoms and molecules. In other words the model is not made from the same stuff (no matter how closely it models that stuff) as the "real thing". Hard not to get all metaphysical about this but all I am saying is a model is NOT the same as the real thing and whatever the differences are between the model and the real thing may be something that would prevent them sharing the same emergent properties. That doesn't involve any magical essence just the fact that different things arranged in different ways have different properties.

Yes. It is intelligence.

What definition for intelligence are you using?

Well, there's two different things there. If you completely and accurately model the workings of the human brain, the output of the model will be the same as the output of the brain being modelled.

Back to my point - a map is not the territory no matter how good the map is. (Of course it may be that we can model a system like the human barin in a different medium and find its outputs are indistinguishable from the "real" thing. It's just that I think it's something we should be careful about assuming at this early stage of our knowledge.)

On the other hand, if you just whack a whole lot of computing power together and teach it to make inferences from observations, there's no telling how it's going to end up.

The sc-fi short story that goes:

Humans had done it, after milleninia of work they had created what would be the most powerful computer posible, a computer that could answer any question conceivable.

Despite this power all that was in front of the small team ready for the today's ceremony was a small console with a speaker, a light and two buttons "on" and "off". With a trembling hand the woman pressed the small discrete "on" button. A blue light illuminated her face and a pleasant voice asked "What can I answer?", she looked at her colleagues, they nodded, she turned back to the console and with a slightly unsteady voice asked "Is there a God?"

Suddenly a flash of lightening crashed into the console melting the "off" button and the now not quite so pleasant voice said "There is NOW!".

 

[sphenisc]

Actual plans to have go at this are discussed in this thread...

http://www.internationalskeptics.com/forums/showthread.php?t=42424

 

[hammegk]

If you completely and accurately model the workings of the human brain, the output of the model will be the same as the output of the brain being modelled.

Unknown, but at least possibly true, assuming all the axioms you base your version of reality on are correct.

On the other hand, if you just whack a whole lot of computing power together and teach it to make inferences from observations, there's no telling how it's going to end up.

That would depend on the intelligence and skill of the original programmer. And I suspect you will need "outside-the-system" code revision, and human intelligence will be deciding where it ends up.

 

[Soapy Sam]

Even if Pixy's version is correct- we can build an electronic analogue of a brain and that produces an analogue of a mind- (and like Darat, I see no reason to suppose this is necessarily the case)- Is it worth it? Humans are very good at some things.( Creativity, having fun). Computers are very good at different things.(remembering and accessing data, number crunching).
Why use computers to reproduce consciousness, when we can use them to enormously augment the consciousness we already have? Let the grunts do the grunt work and permit the pigs to fly.

 

[PixyMisa]

Isn't it? Aren't most of even our current best models only an approximation and don't always give the same results as the "real" system do?

Best models of what? Some models - weather and climate models are a good example - are quite limited and serve only as an indication. Other models - QM models of simple systems - are accurate to the limits of floating-point precision.

That statement I can agree with e.g. "a complete physical model" but that is not we are talking about; we are talking about in effect making a model from "information" that is created in a very different way from how our brains are e.g. atoms and molecules. In other words the model is not made from the same stuff (no matter how closely it models that stuff) as the "real thing". Hard not to get all metaphysical about this but all I am saying is a model is NOT the same as the real thing and whatever the differences are between the model and the real thing may be something that would prevent them sharing the same emergent properties. That doesn't involve any magical essence just the fact that different things arranged in different ways have different properties.

Yes it does. If your model is accurate, and it doesn't produce the same results, then the system works by magic.

If you want to argue that the model isn't suffciently accurate, then argue that point, but the statement

In other words the model is not made from the same stuff (no matter how closely it models that stuff) as the "real thing".

Is ascribing magical properties to the real system.

What definition for intelligence are you using?

A functional one. Does it behave intelligently? If yes, it's intelligent.

Back to my point - a map is not the territory no matter how good the map is.

Magic.

(Of course it may be that we can model a system like the human barin in a different medium and find its outputs are indistinguishable from the "real" thing. It's just that I think it's something we should be careful about assuming at this early stage of our knowledge.)

Why?

The sc-fi short story that goes:

Humans had done it, after milleninia of work they had created what would be the most powerful computer posible, a computer that could answer any question conceivable.

Despite this power all that was in front of the small team ready for the today's ceremony was a small console with a speaker, a light and two buttons "on" and "off". With a trembling hand the woman pressed the small discrete "on" button. A blue light illuminated her face and a pleasant voice asked "What can I answer?", she looked at her colleagues, they nodded, she turned back to the console and with a slightly unsteady voice asked "Is there a God?"

Suddenly a flash of lightening crashed into the console melting the "off" button and the now not quite so pleasant voice said "There is NOW!".

"Answer" by Fredric Brown.

 

[PixyMisa]

Unknown, but at least possibly true, assuming all the axioms you base your version of reality on are correct.

It's necessarily true, given my axioms. It may turn out that we can't model the brain sufficiently accurately, though we currently have no reason to believe that. (Penrose doesn't know what he's talking about.)

That would depend on the intelligence and skill of the original programmer. And I suspect you will need "outside-the-system" code revision, and human intelligence will be deciding where it ends up.

Yep. At least, in the starting stages.

 

[PixyMisa]

Even if Pixy's version is correct- we can build an electronic analogue of a brain and that produces an analogue of a mind- (and like Darat, I see no reason to suppose this is necessarily the case)

I see no reason for it not to be the case.

Is it worth it? Humans are very good at some things.( Creativity, having fun). Computers are very good at different things.(remembering and accessing data, number crunching).
Why use computers to reproduce consciousness, when we can use them to enormously augment the consciousness we already have? Let the grunts do the grunt work and permit the pigs to fly.

Because we can make a computer-based intelligence arbitrarily powerful.

Okay, that may not be a good idea. But we never let that stop us in the past!

 

[PixyMisa]

Actual plans to have go at this are discussed in this thread...

http://www.internationalskeptics.com/forums/showthread.php?t=42424

Interesting link.

However, that machine is only one as powerful as the human brain, and we're talking about one powerful enough to simulate the human brain. 10 petaflops vs. 1 exaflop. Mine is 100 times faster.

But this link is very interesting. Scientists are using Blue Gene (one of the fastest computers in the world) to model part of the neo-cortex right down to the molecular level. That's the sort of research program that will lead to the full-scale brain-in-a-box that we've been discussing.

 

[Melendwyr]

But your extrapolation hinges on, at least, the idea of simulating a neuron by a 1 Mflop operation. I suppose it would be more than possible to simulate the neurons in an ant's brains with computers today, given an ant's size. Has such been done?

It's estimated that an ant's brain has roughly the computational power of an Apple IIe - so while I don't know if it's been done yet, it's certainly within the realm of possibility.