[Ummon]

In case you guys haven't been following AI stuff google created an artificial intelligence that is revolutionary because it has aspects that are able to be generalized to many different tasks. It has been a huge leap in artificial intelligence. Now the program is showing off its stuff in the Game Go (the Chinese game with the white and black stones). AlphaGo has defeated Lee Sedol the current second ranked player internaltionally 5 to 0.

You can watch the matches here:

Match 1

[youtube]vFr3K2DORc8[/youtube]


Match 2

[youtube]l-GsfyVCBu0[/youtube]

Match 3

[youtube]qUAmTYHEyM8[/youtube]

Match 4

[youtube]yCALyQRN3hw[/youtube]

Match 5

[youtube]mzpW10DPHeQ[/youtube]

[Besoeker]

In case you guys haven't been following AI stuff google created an artificial intelligence that is revolutionary because it has aspects that are able to be generalized to many different tasks. It has been a huge leap in artificial intelligence. Now the program is showing off its stuff in the Game Go (the Chinese game with the white and black stones). AlphaGo has defeated Lee Sedol the current second ranked player internaltionally 5 to 0.

You can watch the matches here:

Do you think anyone on a forum will watch over 20 hours of video to get the point?
Did you?

[Prosthetic Conscience]

Your post comes from the future of an alternative reality. In the real world, the score is 3-1, as your Wikipedia link notes, with the last game not yet played (starts about 29 hours from now, according to your own post).

[Harmattan]

In case you guys haven't been following AI stuff google created an artificial intelligence that is revolutionary because it has aspects that are able to be generalized to many different tasks. It has been a huge leap in artificial intelligence.

Not really. This is a mix of old ideas, coupled with a huge computing power like few researchers ever had access to. It is not even close to be intelligent and it has about nothing to do with what intelligent machines will be.

However it is indeed useful to tackle a fair number of complex and specialized problems, provided that you can train the machine with a few millions of samples, and a clear metric to rate every state or outcome. But it is still a brainless number crunching beast, only less trivial than some previous attempts.

The advances of the past decade in AI are certainly impressive in the restricted cases they are used, but they look like evolutionary dead-ends when it comes to producing a human-level AI. At least there is now money on this topic but I find the directions and advances disappointing.

[Potemkin]

The advances of the past decade in AI are certainly impressive in the restricted cases they are used, but they look like evolutionary dead-ends when it comes to producing a human-level AI. At least there is now money on this topic but I find the directions and advances disappointing.

It could be argued that they have already achieved a "human-level AI", in the sense that the computer can beat a top-level human player at the game of Go. And Deep Blue beat Kasparov at chess a few years back. What more do they need to do? The problem, of course, is the issue of precisely what we mean by the word "intelligence". Is the ability to crunch large amounts of numbers very quickly "intelligence"? If so, then computers have already surpassed human levels of intelligence. But these "intelligent" machines are not about to take over the world any time soon. AlphaGo doesn't care that it just beat Lee Sedol at Go; in fact, it isn't even aware that it has done so. There is, in short, something missing. We can call that 'something' subjectivity. This is actually more than just self-awareness - it involves things like emotional affect, motivation, self-image, reflexivity, and so forth. In the field of psychoanalysis, thinkers like Jacques Lacan have tried to define what human subjectivity is and how it is formed and maintained from moment to moment. Their work has demonstrated one thing above all else - how highly non-trivial the concept of subjectivity actually is, and how little we understand it. We do know one thing though - it has little to do with number-crunching. It seems to me that researchers in AI have paid little attention to exactly what it is they are trying to replicate; after all, how can we hope to create an artificial version of something if we still have little or no understanding of the original? This is why AlphaGo seems like an evolutionary dead-end to you. If AI just means the ability to crunch numbers very quickly and defeat humans at chess or Go, then not only is it not an evolutionary dead-end, it has already achieved its goal. But there is something deeply dissatisfying about this - and that is the absence of any theory of subjectivity. AI - artificial intelligence - should in fact be renamed to AS - artificial subjectivity. Then we might actually start to get somewhere with trying to create an artificial simulacrum of the human mind.

[Tomato_Cake]

I hope the first true AI machine solves the problems of world peace and feeding the masses. Goodness knows we're too dumb to figure it out.

[Potemkin]

I hope the first true AI machine solves the problems of world peace and feeding the masses. Goodness knows we're too dumb to figure it out.

Actually, that was figured out long ago, by a guy named Karl Marx. Putting his ideas into practice, however, has proved to be... somewhat problematic.

[Harmattan]

We can call that 'something' subjectivity. This is actually more than just self-awareness - it involves things like emotional affect, motivation, self-image, reflexivity, and so forth.

No, what machines currently lack are the very basics: the ability to efficiently analyze their senses through induction, abduction and deduction, in order to form an understanding of the world upon which they can act. And so far all the trails pursued by Google & co are evolutionary dead-ends for this purpose. Any bird is more intelligent, adaptable, than AlphaGo would be with an infinite computing power. The very algorithms currently in use are the problem.

Once we have those basics, then we will focus on the aspects you mentioned. But so far our car only runs provided there is a slope or horses pulling it. We need an engine more than a dashboard.


As for what inelligence is, my car is not intelligent even though it runs faster than me. Show me a machine that can perform as well as a bird with everyday's challenges and we will finally have an embryo of intelligence. But such a thing does not exist today. The job will be done when a machine will be able to completely take over an intellectual profession (not just its repetitive components). AlphaGo and its successors will never do this.

[noemon]

It could be argued that they have already achieved a "human-level AI", in the sense that the computer can beat a top-level human player at the game of Go. And Deep Blue beat Kasparov at chess a few years back. What more do they need to do? The problem, of course, is the issue of precisely what we mean by the word "intelligence". Is the ability to crunch large amounts of numbers very quickly "intelligence"? If so, then computers have already surpassed human levels of intelligence. But these "intelligent" machines are not about to take over the world any time soon. AlphaGo doesn't care that it just beat Lee Sedol at Go; in fact, it isn't even aware that it has done so. There is, in short, something missing. We can call that 'something' subjectivity. This is actually more than just self-awareness - it involves things like emotional affect, motivation, self-image, reflexivity, and so forth. In the field of psychoanalysis, thinkers like Jacques Lacan have tried to define what human subjectivity is and how it is formed and maintained from moment to moment. Their work has demonstrated one thing above all else - how highly non-trivial the concept of subjectivity actually is, and how little we understand it. We do know one thing though - it has little to do with number-crunching. It seems to me that researchers in AI have paid little attention to exactly what it is they are trying to replicate; after all, how can we hope to create an artificial version of something if we still have little or no understanding of the original? This is why AlphaGo seems like an evolutionary dead-end to you. If AI just means the ability to crunch numbers very quickly and defeat humans at chess or Go, then not only is it not an evolutionary dead-end, it has already achieved its goal. But there is something deeply dissatisfying about this - and that is the absence of any theory of subjectivity. AI - artificial intelligence - should in fact be renamed to AS - artificial subjectivity. Then we might actually start to get somewhere with trying to create an artificial simulacrum of the human mind.

Nice.

[Tomato_Cake]

Actually, that was figured out long ago, by a guy named Karl Marx. Putting his ideas into practice, however, has proved to be... somewhat problematic.

At least a true AI machine would be able to learn from past mistakes, without repeating them like an insane Che zealot teenager, or sommat.

[AFAIK]

How complicated is Go? I've never played but since all the information is openly available I assume it's the same process as playing chess or naughts and crosses. All the information is there in front of you and if you allow the computer enough time it can analyse all possible moves and select the best course of action. Do they play against a clock requiring the computer to complete its calculations quickly?

Computers struggle with games where some information is hidden. Online poker players use software to help them analyse their opponents but no one has developed a computer that can act on this information in a successful manner. There's too much psychology and room to maneuver for a computer to keep up let alone excel.

[Potemkin]

How complicated is Go?

The rules are very simple, but the strategy is insanely complex. Even top-level players often have to rely on intuition and 'feel'. Even calculating who's winning at any given moment is sometimes almost impossible.

I've never played but since all the information is openly available I assume it's the same process as playing chess or naughts and crosses. All the information is there in front of you and if you allow the computer enough time it can analyse all possible moves and select the best course of action.

In theory, yes; in practice, no. There are more possible positions in a game of Go than there are atoms in the entire known universe. This is why Go was a more difficult game to solve than chess for the AI researchers. In any given position, a chess player has about 20 possible moves; in Go, the player has about 200 possible moves. Strategically, it's on a whole 'nother level. In fact, it could be argued that humans are terrible at playing Go; even top-level players are winging it half the time. It's actually too difficult a game for the human mind.

Computers struggle with games where some information is hidden. Online poker players use software to help them analyse their opponents but no one has developed a computer that can act on this information in a successful manner. There's too much psychology and room to maneuver for a computer to keep up let alone excel.

The basic problem is that computers have no theory of mind. They cannot ascribe motivations or subjectivity to other players, since they have none themselves.

[Ganeshas Rat]

How complicated is Go? I've never played but since all the information is openly available I assume it's the same process as playing chess or naughts and crosses. All the information is there in front of you and if you allow the computer enough time it can analyse all possible moves and select the best course of action. Do they play against a clock requiring the computer to complete its calculations quickly?

As I remember, the tree of all possible Go moves is about 10^200 when Chess is 10^80. So, to analyze all possible moves you need about 1 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 times more computational power.

Actually, Go is so complex, nobody plays it direct way with counting and analyzing moves: it's more about art, with controlling composition, filling, tension, balance between attack and defense, "strongholds" and "expansions". Only some simple patterns in the local surrounding can be remembered. It's very orient game that has nothing common with Chess experience.

[Rugoz]

This is actually more than just self-awareness - it involves things like emotional affect, motivation, self-image, reflexivity, and so forth. In the field of psychoanalysis, thinkers like Jacques Lacan have tried to define what human subjectivity is and how it is formed and maintained from moment to moment. Their work has demonstrated one thing above all else - how highly non-trivial the concept of subjectivity actually is, and how little we understand it. We do know one thing though - it has little to do with number-crunching.

It has all to do with number crunching. You've been trained for billions of years for spreading your genes in a very challenging environment. Naturally the outcome is very complex. If self-awareness is not useful for winning in GO why do we expect AlphaGo to develop such behavior?

Also its a fallacy to think that AI researchers actually have a full understanding of what AlphaGo is doing. I guess they can analyze the weights of the network and give an intuitive explanation but that's about it. A brain is infinitely more complex.

[Rancid]

Question about Go.

Can you create the illusion of retreat and then turn the tables and destroy your opponent? I always love games that allow for a rope-a-dope strategy.

[Potemkin]

Can you create the illusion of retreat and then turn the tables and destroy your opponent? I always love games that allow for a rope-a-dope strategy.

Yes. If you are losing a local tactical battle, you can 'tenuki' (ie, play in some other part of the board). After a few moves in that other part of the board, the situation will have changed in the board as a whole (since all the local battles are strategically interconnected), and you can then return to the original battle and play a winning sequence which would not have been possible before you 'retreated'.

Once you get the hang of it, Go is the most awesome board game ever.

[Truth To Power]

I hope the first true AI machine solves the problems of world peace and feeding the masses. Goodness knows we're too dumb to figure it out.

Actually, that was figured out long ago, by a guy named Karl Marx. Putting his ideas into practice, however, has proved to be... somewhat problematic.

No, actually Marx was nothing but a brash, abrasive, self-important nincompoop who succeeded only in infecting millions of other people with his own delusions and confusions.

How complicated is Go? I've never played but since all the information is openly available I assume it's the same process as playing chess or naughts and crosses. All the information is there in front of you and if you allow the computer enough time it can analyse all possible moves and select the best course of action. Do they play against a clock requiring the computer to complete its calculations quickly?

As I remember, the tree of all possible Go moves is about 10^200 when Chess is 10^80. So, to analyze all possible moves you need about 1 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 000 times more computational power.

It may be more informative to consider the number of "good" games -- i.e., the number of different games that might be played by strong players. This depends on the length of the game and the "branching factor": the number of distinct moves that strong players would consider playable at each decision point. In chess, the length of a game between strong players averages about 40 moves (i.e., 80 half-move decision points), and the branching factor around 2. This gives a number of possible good games in the neighborhood of 2^80 or about 10^24.

Go games are longer (~200 half-moves), and the branching factor for good moves is much higher, probably around 6. So the number of possible good games is about 6^200, or about 10^52 (edit: sorry, it was 10^152). So for each good chess game that could be played, there are about 10^128 good go games.

Actually, Go is so complex, nobody plays it direct way with counting and analyzing moves: it's more about art, with controlling composition, filling, tension, balance between attack and defense, "strongholds" and "expansions". Only some simple patterns in the local surrounding can be remembered. It's very orient game that has nothing common with Chess experience.

Reading (predicting the outcomes of tactical engagements) is important in go, but much less so than in chess. Simultaneous play is much easier for a strong go player than a strong chess player, though blindfold play is much harder because the board is so much larger and perception of overall positions is more important.

Computers struggle with games where some information is hidden. Online poker players use software to help them analyse their opponents but no one has developed a computer that can act on this information in a successful manner. There's too much psychology and room to maneuver for a computer to keep up let alone excel.

This is incorrect. Online poker is actually dominated by computer programs, which easily take the money of all but the very best human players. In a sense, this is a natural outcome: human players are emotionally involved with their money, which makes them commit errors in betting. Computer programs don't care about money at all, so they can simply play to win. Often this means very aggressive betting, which is also characteristic of top human players.

[Potemkin]

No, actually Marx was nothing but a brash, abrasive, self-important nincompoop who succeeded only in infecting millions of other people with his own delusions and confusions.

So sorta like every important thinker who's ever lived then?

[Truth To Power]

No, actually Marx was nothing but a brash, abrasive, self-important nincompoop who succeeded only in infecting millions of other people with his own delusions and confusions.

So sorta like every important thinker who's ever lived then?

No, many important thinkers have been reasonable people, and have actually contributed to human understanding rather than undermining it as Marx did. People like Newton (who was certainly abrasive, by all accounts, but not brash or self-important), who basically made modern civilization possible; Aristotle, who revolutionized human understanding in almost every field he investigated; or Darwin, who provided the clear, factual answer to the eternal question, "Why are we here?" Marx, by contrast, contributed nothing of significance but confusion and misunderstanding, and even went out of his way to derogate his literary, economic, moral and intellectual superiors, like Jacques Turgot, Adam Smith, JS Mill and Henry George.

BTW, Wikipedia has some information on go combinatorics:

https://en.wikipedia.org/wiki/Go_and_mathematics

Turns out that the number of "possible" go games is somewhere between 10^500 and 10^1000, if you assume play results in a normal game termination after a few hundred moves. However, if you allow nonsensical play where players fill in their own territories and eyes, resulting in repeated captures over large areas of the board, the game can be much, much, MUUUUCH longer -- so long that ladder exponents are required to express the total number of possible games. Wikipedia gives an absolute upper limit to the length of such "games" as 10^48 moves, and the number of possible games of this type as 10^10^171. This is a number that dwarfs not merely the number of atoms in the universe (a relatively tiny 10^80, which by comparison you could practically count on your fingers) but the highest number considered to express anything physically real: the number of cubic Planck lengths in the universe, which is probably around 10^186. The length of chess games is bounded by the 50-move rule (if no piece is captured or pawn moved in 50 moves, the game is a draw), so the maximum possible game length is far less than a million moves, and the number of possible games correspondingly modest.

[lucky]

But it is still a brainless number crunching beast

Given that brains can also be described as "number crunching beasts", the adjective "brainless" seems inappropriate. In fact, the neural network topology of AlphaGo is rather reminiscent of what we know about how human visual perception works.