The Singularity is Near (or not)

[BenBurch]

You know best, of course. Better than we mere engineers who have implemented them.

 

[PixyMisa]

Well, apparently so.

Your definition of computers is too narrow to be of any practical use, and your statement regarding simulating a computer with a neural network is categorically incorrect.

 

[This is The End]

Yeah, I'm pretty sure you guys are having a semantics argument.

 

[soylent]

Well, I certainly expect to see a lot of cool stuff in my lifetime, but it doesn't amount to a "singularity".

I expect 3d printing take off in a big way. Just like we regularly squander more computing power than it took to design the hydrogen bomb, break the enigma code and land on the moon on porn, games, music, spellchecking and distributing kittens with funny captions across the world; I expect a similar abundance on highly precise 3d printing/manufacturing that will never be as cheap as massproduction with conventional techniques, but cheap enough that people will use it to produce all sorts of customized objects. Everything from fancy door-knobs, sex toys(silicone is a useful material for many things; I expect printers for it eventually), nose rings, you name it.

Here's what can be done right now with an expensive machine by and for experts:

Give it a couple of decades and I expect the average person in the developed world to take it for granted as something completely obvious that if you can model it in 3d, you can print it and share the design; either on your home printer if you don't need super high quality or in the printing shop if you need something big, or precise or with multiple materials or anything fancy like that. If you want a bottle opener with your main character from and an MMO as the handle, you just grab a plain bottle opener from a standard library, import your character model from that game and compose the two toghether and hit print.

I expect the iphone/ipad stuff to be only the beginning. I expect mobile phones to be as close to a real trichorder as anything well ever make. If it's only a buck extra, why wouldn't I want it to have embarassing amounts of computational power? If it's just a buck extra why wouldn't I want the universal translator that while not as good as a real person, does a pretty bang up job overall? If it's just a dollar extra I'll take the one with the ultrasonograph, a laser 3d-digitizer/scanner, a miniature scintilation counter and an IR camera.

 

[BenBurch]

Well, apparently so.

Your definition of computers is too narrow to be of any practical use, and your statement regarding simulating a computer with a neural network is categorically incorrect.

Proof? Show me a simulation of a Von Neumann architecture computer implemented entirely in neural networks, and I do not mean cheating by having a neural network designed not to learn, such that its just a transistor array, but a true neural network. Failing that, show me a Turing Machine. And the criterion is that it has to function accurately 100% of the time.

 

[BenBurch]

BTW, that paper you cited is worthless. He cannot implement that.

ETA: AND those are fixed networks, not learning. So not true neural networks. You decide the weights before you build it. This is not what NN are.

 

[PixyMisa]

Proof? Show me a simulation of a Von Neumann architecture computer implemented entirely in neural networks, and I do not mean cheating by having a neural network designed not to learn, such that its just a transistor array, but a true neural network. Failing that, show me a Turing Machine.

That's what I just showed you. The paper I linked to specifically proves that a neural network with feedback is Turing-equivalent (and that a neural network without feedback is not).

Turing equivalence. It's not just for breakfast any more.

And the criterion is that it has to function accurately 100% of the time.

Are you asking this as a mathematical proposition - in which case, it's already established - or as a physical reality? If the latter, you already know what's wrong with that.

 

[PixyMisa]

BTW, that paper you cited is worthless. He cannot implement that.

Okay, apparently you don't already know what's wrong with that.

Of COURSE you can't implement it. It's a mathematical concept that happens to involve infinities. So do Turing machines. You can, however, approximate it to an arbitrary degree of precision.

ETA: AND those are fixed networks, not learning. So not true neural networks. You decide the weights before you build it. This is not what NN are.

Sorry, that's simply irrelevant. You claimed that no-one knew how to get a neural network to simulate a computer, which is categorically untrue. You are now retreating to the position that some features of neural networks are not required to simulate a computer.

So?

I have two legs. I am not using them to type this response. That doesn't prove that people with legs can't type, which is what you are arguing here.

As with your problem with the term computer, you have narrowed the term neural network to some undefined subset of what neural networks actually do.

 

[PixyMisa]

Well, I certainly expect to see a lot of cool stuff in my lifetime, but it doesn't amount to a "singularity".

I expect 3d printing take off in a big way. (snip)

Yeah, the singularity specifically requires human-equivalent (though not necessarily human-like) plus-epsilon artificial intelligence. The premise being that if technological progress is driven by human intelligence (sounds good), and human intelligence can create slightly-greater than human-level artificial intelligence (mathematically and physically entirely plausible, but very difficult), then those slightly-greater-than-human-level artificial intelligences can not only generate technological advances in general, but can also generate a new generation of more advanced AIs that can produce more technological advancement faster and more powerful AIs sooner.

And once you tip over that line, it all snowballs until you run into basic physical limitations. The singularity is the inflection point: Standing on one side of it, you can't tell what things will be like on the other side.

Personally, I believe that Kurzweil is right in the general case and unreasonably optimistic in most of the specifics.

 

[soylent]

Well, to be fair to Kurzweil, robots have been making awesome progress in just the last few years but I still don't think were going to have human-brain level of quality software/general intelligence anytime this century.