Artificial Intelligence & the Downfall

[Brian]

Apparently they have a big red button that turns off all the computer servers at Google in case they get, ummm, carried away / too clever.

 

There are already programs that are analysing the internet in order to learn. For example, shortly after Brexit, the French finance minister gave a long speech about how the pound was going to devalue loads, it had already done so by 20% by then anyway. Well, there are financial computers that read all the news paper headlines and then make predictions as to what effect this will have on the stock markets and then shift money accordingly in microseconds.

 

This one computer read the headlines based on what the French finance minister said and shifted lots of currency out of sterling. When this happened, then a load of other similar programs saw the fall in the pound and also sold sterling. The currency then took another 10 to 20% dip for a few minutes before equilibrium kicked in and the pound then went back to what it was.

 

This 'flash' crash is an example of computers learning from what we all put on the internet. If that is a computers main source of learning, its a bit worrying.

So like most people, then?

 

 

Bateson's levels of learning framework comes to mind...

[MooNiNite]

Obama's assertion that developing high levels of AI is becoming more available and wide spread is important. 

 

It is the same with programming, individuals in their room with no money can create websites or databases stronger than million dollar corporations. 

 

The accessibility is becoming more and more widespread. 

[MooNiNite]

Saying that Specialized AI is the same as AI strongly undermines what true AI is. We are talking about a level of consciousness not a dead formula. 

 

However, on the other hand, it is important to understand that we ourselves can't define consciousness or AI itself. 

Edited January 10, 2017 by MooNiNite

[Brian]

https://www.extremetech.com/extreme/240695-scientists-discover-natures-algorithm-intelligence

[dust]

First off, I'd just like to urge anyone who hasn't to watch the video from the OP, or even just respond to the actual words I wrote in the OP. The question in hand is the potential future dangers of AI, and general AI in particular.
 
Not that I'm suggesting everyone must agree with what Deutsch says, but that if you have an idea of the view being put forward, have considered AI from this slightly different perspective than is usually found in mainstream media/culture, and perhaps most importantly we are all starting from the same point of discussion, the thread would be a bit more coherent.
 
Dissecting Obama and Musk and their personal take on the issue, and talking about job loss to machines and other stuff is all fine, but it's not the point. There is a recent topic in Off Topic regarding ever-increasing automation and job loss etc.
 

 

Obama seems to understand some basics. For example he make a difference between specialized IA (that is a real IA) and general IA (the sience fiction IA). Unfortunately, he also makes mistakes like claiming that an IA can produce a cure an unknown disease (maybe he meant researchers using IA to optimise certain molecular structures to find a cure, or that it possible to train an IA to recognise a certain known disease by its symptoms). His advisors are also quite correct in claiming we are still a long way away from a general IA. The other guy confirm it by stating that a general IA won't happen without a major breakthrough (and, at the same time, he spread more clichés about geeks).
 
It must also be noted that there were no breakthrough in IA for at least 20 years. AlphaGo is a deep neural network, which exists since the 90s. The difference between now and then is that we have warehouse sized data-centers and we can analyse a large enough subset of the billions of possible moves in a go game to make a good guess about beneficial moves.

 
He didn't claim that AI can cure diseases right now, but that if harnessed properly it 'can' (will be able to) cure diseases in the future.

 

And yes, neural networks have existed since the '60s, and major developments were made in the '80s and '90s, but new developments within the area are certainly still being made.

 

I won't pretend to understand anything beyond a very simple idea of how these various neural networks work, but as I understand it, AlphaGo combines the Monte Carlo tree search (random search of possible moves (more than billions!)) with a convolutional neural network, and is able to use this CNN alone (without the tree search) to beat some other Go programs (e.g. Patchi!). So yes, processing power is much higher than it was when Deep Blue beat Kasparov at chess, for example, but it is not so simple as just speed/computing power being improved.

 

 

There are games an IA can beat a human player, like chess, backgammon, cards against humanity, etc... For these games the cases are clearly defines, there is a situation before a move, then an other situation after the move.
 
In the case of games where IA are still weak against a human opponent are soccer, basketball, boxing, etc... For these games the cases are ill defined and beneficial situation difficult to determine.

 
Also the machine would need a body fit for task...

Edited January 10, 2017 by dust

[Jox]

Perdix micro-drones ...

 

https://www.youtube.com/watch?v=B-rZHbc4YTE

[Taomeow]

[Eques Peregrinus]

He didn't claim that AI can cure diseases right now, but that if harnessed properly it 'can' (will be able to) cure diseases in the future.

 

And yes, neural networks have existed since the '60s, and major developments were made in the '80s and '90s, but new developments within the area are certainly still being made.

 

I won't pretend to understand anything beyond a very simple idea of how these various neural networks work, but as I understand it, AlphaGo combines the Monte Carlo tree search (random search of possible moves (more than billions!)) with a convolutional neural network, and is able to use this CNN alone (without the tree search) to beat some other Go programs (e.g. Patchi!). So yes, processing power is much higher than it was when Deep Blue beat Kasparov at chess, for example, but it is not so simple as just speed/computing power being improved.

OK, I misunderstood it, my English is not perfect. (And now, I notice than I reversed AI into IA, this is because in English the adjective precede the name, while in French it is the opposite.) They really are speaking about a (yet) non existing AI. Current AI can still be used to determine the shape of some proteins with a good accuracy.

 

I am not a specialist in AI either. Still I did not learn of any major discovery made in the last 20 years. Of course multiple variants of neural network or strategies were invented since then, like the Monte Carlo Tree Search you mentioned (which was invented in 2006). They are useful in certain situations, but not really ground-breaking.

 

Because we have way more computation power now, heuristics which can technically be made are also better. It was already relevant in the progress on AI even in 1996 when DeepBlue beat Kasparov (DeepBlue was simulating 200 millions moves per second), but as you mentioned, the heuristic was build differently. I did not find the number of moves computed by AlphaGo, but, according to this article, it ran on 1202 cores to beat the world 2nd best player. The computing power needed to train the neural network, however is not mentioned.

 

 

Also the machine would need a body fit for task...

I was quite serious in this statement (well almost, the comparison between Muhammad Ali and Kasparov was a joke). It is not really a problem of having or not having a body, an AI analyse a model, in which the data is obtained from a real game. For example, at the moment, an AI is able to determine what is a pick-and-rolls and can be used to help to determine what was a good or a bad throw (but we are still far from placing and directing players on the pitch with an AI).

 

 

 

To go back to the dangerous AI, like SkyNet in Terminator, which is able to hack and take control of other systems automatically. Such AI does not yet exists, but there already is a competition to make an ersatz. Currently the binaries hacked are much simplified compared to real programs, but, frankly speaking, this is quite frightening.

http://arstechnica.com/information-technology/2016/08/the-world-series-of-hacking-without-humans/

 

edit: typo

Edited January 11, 2017 by Eques Peregrinus

[Henchman21]

Why would a bank make a press release that we might be living in a computer program?

http://www.independent.co.uk/life-style/gadgets-and-tech/news/bank-of-america-the-matrix-50-per-cent-virtual-reality-elon-musk-nick-bostrom-a7287471.html

[Brian]

Why would a bank make a press release that we might be living in a computer program?

http://www.independent.co.uk/life-style/gadgets-and-tech/news/bank-of-america-the-matrix-50-per-cent-virtual-reality-elon-musk-nick-bostrom-a7287471.html

So that they can embezzle money from their customers and blame it on a bug in reality.

[Zhongyongdaoist]

The people who are espousing ideas like this do not understand the limits of computability. Simply put, neither neither consciousness, nor intelligence are computable functions, and if they are not computable they cannot be written into programs that can be run on computers, even the "brain as computer" model fails for this reason. While it may be a necessary presupposition of reductive materialism to think that "machines" can think, it is not one necessary to any "spiritual" approach to life, so why anyone is bothering with this is beyond me. This presupposition is shared by both commentators in the original interview, and thus it is largely a waste of time to listen to it, because neither of them really questions it as they proceed along their merry way.
 
Here is the abstract of a nice technical paper that summarizes some aspects of this:
 

In this article we review Tononi's (2008) theory of consciousness as integrated information. We argue that previous formalizations of integrated information (e.g. Griffith, 2014) depend on information loss. Since lossy integration would necessitate continuous damage to existing memories, we propose it is more natural to frame consciousness as a lossless integrative process and provide a formalization of this idea using algorithmic information theory. We prove that complete lossless integration requires noncomputable functions. This result implies that if unitary consciousness exists, it cannot be modelled computationally. ("Is Consciousness Computable? Quantifying Integrated Information Using Algorithmic Information Theory", arVix, Emphasis mine ZYD)

 
This and articles like it, just scratch the surface.
 
At best, what is touted as artificial intelligence is imitation intelligence, something that can be passed off as intelligence, like cubic zirconium can be passed off as a diamond, to the unwary, but while there can be, and are real, artificial diamonds, there is not now, and cannot be, artificial intelligence.
 
For a less technical discussion, you might want to start here:
 
The Brain is not Computable, MIT Technology Review
 
I am sorry to be so short about this, but I really don't have time to enter into a longer discussion now, or in the near term.

[dust]

 

OK, I misunderstood it, my English is not perfect. (And now, I notice than I reversed AI into IA, this is because in English the adjective precede the name, while in French it is the opposite.) They really are speaking about a (yet) non existing AI. Current AI can still be used to determine the shape of some proteins with a good accuracy.

 

I am not a specialist in AI either. Still I did not learn of any major discovery made in the last 20 years. Of course multiple variants of neural network or strategies were invented since then, like the Monte Carlo Tree Search you mentioned (which was invented in 2006). They are useful in certain situations, but not really ground-breaking.

 

Because we have way more computation power now, heuristics which can technically be made are also better. It was already relevant in the progress on AI even in 1996 when DeepBlue beat Kasparov (DeepBlue was simulating 200 millions moves per second), but as you mentioned, the heuristic was build differently. I did not find the number of moves computed by AlphaGo, but, according to this article, it ran on 1202 cores to beat the world 2nd best player. The computing power needed to train the neural network, however is not mentioned.

 

Nobody's English is perfect..that was an easy misunderstanding to make. He should have used 'will' instead of 'can'.

 

I suppose 'ground-breaking' is a subjective term.. perhaps there hasn't been a complete reexamination of how to approach building these machines, but who's to say that a single groundbreaking invention will be the key? Maybe it's just a matter of time, of consistent development. I dunno.

 

Anyway, the speed of discovery doesn't deny that the discovery will likely be made, eventually...

 

 

 

To go back to the dangerous AI, like SkyNet in Terminator, which is able to hack and take control of other systems automatically. Such AI does not yet exists, but there already is a competition to make an ersatz. Currently the binaries hacked are much simplified compared to real programs, but, frankly speaking, this is quite frightening.

http://arstechnica.com/information-technology/2016/08/the-world-series-of-hacking-without-humans/

 

Yah

[dust]

The people who are espousing ideas like this do not understand the limits of computability. Simply put, neither neither consciousness, nor intelligence are computable functions, and if they are not computable they cannot be written into programs that can be run on computers, even the "brain as computer" model fails for this reason. While it may be a necessary presupposition of reductive materialism to think that "machines" can think, it is not one necessary to any "spiritual" approach to life, so why anyone is bothering with this is beyond me.

 

Well... there are people sincerely trying to create more and more complex machines, and even if the limit is an imitational intelligence, there is still the question of what that might entail. And some just find it interesting!

 

 

This and articles like it, just scratch the surface.

 

At best, what is touted as artificial intelligence is imitation intelligence, something that can be passed off as intelligence, like cubic zirconium can be passed off as a diamond, to the unwary, but while there can be, and are real, artificial diamonds, there is not now, and cannot be, artificial intelligence.

 

Thanks for the article. People have believed lots of things to be impossible...

 

There is another thing in this article mentioned in the interview, and something I'd hoped to discuss -- the assimilation of machines. Any thoughts on the ethics of this?

[silent thunder]

I have carried within me, the abiding and unshakable sense that there is sentience on all levels of what I can perceive and what is beyond my perceptions. 

 

Stones, plants, dirt, water, light, sound, planets, stars, books, pens, computers... aware all.

 

At least, that's how it's always seemed to me.  For what that's worth

 

Thanks Bums for another great thread!

[Eques Peregrinus]

The people who are espousing ideas like this do not understand the limits of computability. Simply put, neither neither consciousness, nor intelligence are computable functions, and if they are not computable they cannot be written into programs that can be run on computers, even the "brain as computer" model fails for this reason. While it may be a necessary presupposition of reductive materialism to think that "machines" can think, it is not one necessary to any "spiritual" approach to life, so why anyone is bothering with this is beyond me. This presupposition is shared by both commentators in the original interview, and thus it is largely a waste of time to listen to it, because neither of them really questions it as they proceed along their merry way.

 

Here is the abstract of a nice technical paper that summarizes some aspects of this:

 

 

This and articles like it, just scratch the surface.

 

At best, what is touted as artificial intelligence is imitation intelligence, something that can be passed off as intelligence, like cubic zirconium can be passed off as a diamond, to the unwary, but while there can be, and are real, artificial diamonds, there is not now, and cannot be, artificial intelligence.

 

For a less technical discussion, you might want to start here:

 

The Brain is not Computable, MIT Technology Review

 

I am sorry to be so short about this, but I really don't have time to enter into a longer discussion now, or in the near term.

 

Consciousness is not computable, but an AI does not needs to be sentient to be dangerous. This is why I took the example of a competition to build "hacking AI".

 

When an attack occurs on a system, it will, in general, be detected by a program (which uses sometimes an AI) called an IDS (intrusion detection system), which is going to identify, and block the attack then log it for the administrator. The problem is that the IDS, being a program, is subject to the same limitations you mentioned. Therefore, in this situation, we can have a fight between equal opponents (as long as the admin is not doing any intervention).

 

A few years ago, even a simple program, attacking a very specific weakness, could spread worldwide in minutes (this is what is called a worm). Today malwares appears by thousands each day, and sometimes they are programmed to hide from the anti-virus by changing some parts of itself to change its signature. Given their increasing complexity it would not be surprising to see malwares using machine learning techniques to search for certain common vulnerabilities in the future.

 

On the good side, such techniques could also be used by developers to test their software.

[dust]

I have carried within me, the abiding and unshakable sense that there is sentience on all levels of what I can perceive and what is beyond my perceptions. 

 

Stones, plants, dirt, water, light, sound, planets, stars, books, pens, computers... aware all.

 

At least, that's how it's always seemed to me.  For what that's worth

 

Thanks Bums for another great thread!

 

Perhaps I know something of this sense. Not sentience, but certainly a sense of awe and connectedness. And not really 'artificial'

[Taomeow]

The people who are espousing ideas like this do not understand the limits of computability. Simply put, neither neither consciousness, nor intelligence are computable functions, and if they are not computable they cannot be written into programs that can be run on computers, even the "brain as computer" model fails for this reason. While it may be a necessary presupposition of reductive materialism to think that "machines" can think, it is not one necessary to any "spiritual" approach to life, so why anyone is bothering with this is beyond me. This presupposition is shared by both commentators in the original interview, and thus it is largely a waste of time to listen to it, because neither of them really questions it as they proceed along their merry way.

 

Here is the abstract of a nice technical paper that summarizes some aspects of this:

 

 

This and articles like it, just scratch the surface.

 

At best, what is touted as artificial intelligence is imitation intelligence, something that can be passed off as intelligence, like cubic zirconium can be passed off as a diamond, to the unwary, but while there can be, and are real, artificial diamonds, there is not now, and cannot be, artificial intelligence.

 

For a less technical discussion, you might want to start here:

 

The Brain is not Computable, MIT Technology Review

 

I am sorry to be so short about this, but I really don't have time to enter into a longer discussion now, or in the near term.

 

I don't have a longer discussion in mind either, but if you have a moment, defining the particular kind of "intelligence" you are talking about would be helpful.  Unless we agree on what it is, what is it that we deem possible or impossible to create artificially?  What is "intelligence?" 

[Marblehead]

http://www.mediapost.com/publications/article/292780/

[dust]

I'm glad to see some of you obviously know your stuff (and a damn sight more than me!) and in differing areas (and with differing opinions!). All I know of AI tech itself is what I've read online in the last few days, so nothing at all really. I'm about to read about memristors but have a feeling it'll be as beyond me as when I tried reading about convolutional neural networks.

 

Quantum computing in this area is also promising things, no?

[Eques Peregrinus]

This online book is nice if you want to learn about neural networks.

 

At the moment, I don't think researchers have as a priority to develop a "quantum AI". In fact, a mathematically unsolvable problem is not solvable whether in a classical or in a quantum framework. But a few problems hard to solve with a classical computer can be solved faster with a quantum computer. This is the case for the discrete logarithm and the integer factorization problems, which are used, respectively, in the El Gamal and RSA algorithms. (These algorithms are used in cryptography, for key exchange.)

 

Also: http://www.smbc-comics.com/comic/the-talk-3

[Miffymog]

This online book is nice if you want to learn about neural networks.

 

At the moment, I don't think researchers have as a priority to develop a "quantum AI". In fact, a mathematically unsolvable problem is not solvable whether in a classical or in a quantum framework. But a few problems hard to solve with a classical computer can be solved faster with a quantum computer. This is the case for the discrete logarithm and the integer factorization problems, which are used, respectively, in the El Gamal and RSA algorithms. (These algorithms are used in cryptography, for key exchange.)

 

Also: http://www.smbc-comics.com/comic/the-talk-3

 

 

Yes, yes yes - but of all the numerous algebraic problems there are, those that are solved more efficiently by a Quantum computer is still, at the moment, relatively small. Currently they need keeping at liquid Helium temperatures and they just cant be that complex yet. This, of course, may well change in time.

 

But its a bit like nuclear fusion. A great concept that has so much potential. However, how many decades before it has a serious social impact?

[dust]

This online book is nice if you want to learn about neural networks.

 

At the moment, I don't think researchers have as a priority to develop a "quantum AI". In fact, a mathematically unsolvable problem is not solvable whether in a classical or in a quantum framework. But a few problems hard to solve with a classical computer can be solved faster with a quantum computer. This is the case for the discrete logarithm and the integer factorization problems, which are used, respectively, in the El Gamal and RSA algorithms. (These algorithms are used in cryptography, for key exchange.)

 

Also: http://www.smbc-comics.com/comic/the-talk-3

 

The comic is brilliant

 

Will have a look at the book

 

 

But its a bit like nuclear fusion. A great concept that has so much potential. However, how many decades before it has a serious social impact?

 

What social impact exactly?