World's top Go player loses second match of five against Google AI

[Trinity]

In related news, robots beat humans at solving a Rubik's Cube:

https://www.youtube.com/watch?v=ixTddQQ2Hs4

[wonderer]

Go has been mastered by AI.

Just another piece of evidence that Anselm's Ontological Argument is valid...

How is that?

You are asking why things are as they are in bizarroland?

Please don't feed the troll.  It just results in otherwise interesting threads being buried in troll excrement.

[TheBigOhMan]

The point is you haven't made something like human intelligence.
You've made something obviously and conspicuously unlike human intelligence.

And I really don't know how it is you can't see that processor speed is sheer brute force - no different than password nut crackers. How is it fair for the human that a computer can process millions of possible move combinations per minute?

It is fair because that isn't the kind of strategy the bot uses. If it used that strategy it would lose. Go cannot be won like that against good players because the number of possible moves explodes into unmanageable numbers too rapidly, far too large for the AI to search.

The bot learned the same way humans do, by watching games. It categorises various aspects of the game into higher order concepts, as humans do, and evaluates its moves based mostly on "intuition", that is, the network evaluation of the global game state, again in a similar way to humans.

You should really read more about how it works, you seem to be really failing to appreciate what has been achieved here.

The robot may be capable of doing that, but a robot will never actually understand how to play Go.

Define "actually understand". Strikes me as a distinction without a (observable) difference. BTW game 3 has just gone to alphaGo, so that's it, it has won the series. Go has been mastered by AI.

By "actually understand" I meant that a robot cannot both hold a syntaxis description of the game, and also a semantic description of said syntaxis description. And you are right, it's not something observable, since it related deeply to the nature of consciousness and awareness, which is also not observable. That doesn't mean it's not an important distinction though.

[kurros]

The point is you haven't made something like human intelligence.
You've made something obviously and conspicuously unlike human intelligence.

And I really don't know how it is you can't see that processor speed is sheer brute force - no different than password nut crackers. How is it fair for the human that a computer can process millions of possible move combinations per minute?

It is fair because that isn't the kind of strategy the bot uses. If it used that strategy it would lose. Go cannot be won like that against good players because the number of possible moves explodes into unmanageable numbers too rapidly, far too large for the AI to search.

The bot learned the same way humans do, by watching games. It categorises various aspects of the game into higher order concepts, as humans do, and evaluates its moves based mostly on "intuition", that is, the network evaluation of the global game state, again in a similar way to humans.

You should really read more about how it works, you seem to be really failing to appreciate what has been achieved here.

The robot may be capable of doing that, but a robot will never actually understand how to play Go.

Define "actually understand". Strikes me as a distinction without a (observable) difference. BTW game 3 has just gone to alphaGo, so that's it, it has won the series. Go has been mastered by AI.

By "actually understand" I meant that a robot cannot both hold a syntaxis description of the game, and also a semantic description of said syntaxis description. And you are right, it's not something observable, since it related deeply to the nature of consciousness and awareness, which is also not observable. That doesn't mean it's not an important distinction though.

Well I think what is important is that such a thing in no way affects the potential abilities of an AI.

[Lion IRC]

I still can't get my head around the supposed 'specialness' of a computer program written by a human doing what it is supposed to do - make moves in a game of Go according to a set of defined conditionals.

[Lion IRC]

...and there's too many AI skeptics in the field who aren't helping me much either.

[kurros]

I still can't get my head around the supposed 'specialness' of a computer program written by a human doing what it is supposed to do - make moves in a game of Go according to a set of defined conditionals.

Because that's just not how the AI works. At least some of the most central components of it don't work that way.

[TheBigOhMan]

The point is you haven't made something like human intelligence.
You've made something obviously and conspicuously unlike human intelligence.

And I really don't know how it is you can't see that processor speed is sheer brute force - no different than password nut crackers. How is it fair for the human that a computer can process millions of possible move combinations per minute?

It is fair because that isn't the kind of strategy the bot uses. If it used that strategy it would lose. Go cannot be won like that against good players because the number of possible moves explodes into unmanageable numbers too rapidly, far too large for the AI to search.

The bot learned the same way humans do, by watching games. It categorises various aspects of the game into higher order concepts, as humans do, and evaluates its moves based mostly on "intuition", that is, the network evaluation of the global game state, again in a similar way to humans.

You should really read more about how it works, you seem to be really failing to appreciate what has been achieved here.

The robot may be capable of doing that, but a robot will never actually understand how to play Go.

Define "actually understand". Strikes me as a distinction without a (observable) difference. BTW game 3 has just gone to alphaGo, so that's it, it has won the series. Go has been mastered by AI.

By "actually understand" I meant that a robot cannot both hold a syntaxis description of the game, and also a semantic description of said syntaxis description. And you are right, it's not something observable, since it related deeply to the nature of consciousness and awareness, which is also not observable. That doesn't mean it's not an important distinction though.

Well I think what is important is that such a thing in no way affects the potential abilities of an AI.

I think it will affect it's capacity to gain consciousness, but aside from that, it seems possible that an AI may be able to do anything else a human can do.

[Lion IRC]

Kurros,

Please tell me what I'm getting wrong. Help me.
(And don't just say...oh you don't understand....read moar..it's really special...neural networks...fuzzy logic...algorithms...it's all in the jargon.)

To what extent does the program have the 'freedom' to make its own choice of moves entirely free of compulsion?

Basic program - IF this, then do THAT. IF this, then do THAT. IF this, then do THAT....etc.  etc. And do it as fast as the (man made) hardware allows.

The (human) programmer can theoretically write a program that includes every known combination and 'best moves' based on game outcome statistics from past games.

The programmer can also write into the game software a "coin toss" random decision function for situations Wher two competing "best moves" have an equiprobable chance of leading to a win.

So in what sense does the AI game program do something which is NOT expected?

[TheBigOhMan]

Kurros,

Please tell me what I'm getting wrong. Help me.
(And don't just say...oh you don't understand....read moar..it's really special...neural networks...fuzzy logic...algorithms...it's all in the jargon.)

To what extent does the program have the 'freedom' to make its own choice of moves entirely free of compulsion?

Basic program - IF this, then do THAT. IF this, then do THAT. IF this, then do THAT....etc.  etc. And do it as fast as the (man made) hardware allows.

The (human) programmer can theoretically write a program that includes every known combination and 'best moves' based on game outcome statistics from past games.

The programmer can also write into the game software a "coin toss" random decision function for situations Wher two competing "best moves" have an equiprobable chance of leading to a win.

So in what sense does the AI game program do something which is NOT expected?

I think the key to understand it comes from the fact that this AI has the ability to change it's own program, or to add parts to it, so if left alone at the end the programmer might have a hard time figuring out the new inner workings of the AI since it has self-changed itself.

[Trinity]

The ability to change and add to its own program is part of the program. There is no 'ghost in the machine' AI.

[Lion IRC]

Right.
That's my question Trinity.
It's what the recently departed Bunyip couldn't explain.
I understand the program can store new data.

[TheBigOhMan]

The ability to change and add to its own program is part of the program. There is no 'ghost in the machine' AI.

What did you expected? a conscious AI?

[Trinity]

What I mean is that no matter how complex or 'organic' the program becomes, it still remains a program.

It may do your dishes, but not because it loves you.

[aleph naught]

Kurros,

Please tell me what I'm getting wrong. Help me.
(And don't just say...oh you don't understand....read moar..it's really special...neural networks...fuzzy logic...algorithms...it's all in the jargon.)

To what extent does the program have the 'freedom' to make its own choice of moves entirely free of compulsion?

Basic program - IF this, then do THAT. IF this, then do THAT. IF this, then do THAT....etc.  etc. And do it as fast as the (man made) hardware allows.

The (human) programmer can theoretically write a program that includes every known combination and 'best moves' based on game outcome statistics from past games.

The programmer can also write into the game software a "coin toss" random decision function for situations Wher two competing "best moves" have an equiprobable chance of leading to a win.

So in what sense does the AI game program do something which is NOT expected?

There are no if statements in the model, and the programmer does not give it any information about the game. The program learns to play the same way humans do: simply by playing over and over and seeing what works and what doesn't.