Slashdot Mirror
Google AI AlphaGo Wins Again, Leaves Humans In the Dust (cnet.com)
Google's AlphaGo has defeated the world's best Go player in the second out of three games, scoring an overall win for the artificial intelligence algorithm in the fiendishly complex board game. CNET adds: The human gave it his all. "Incredible," wrote DeepMind founder and CEO Demis Hassabis on Twitter while the match was underway. "According to AlphaGo evaluations Ke Jie is playing perfectly at the moment." The match took place over a year after AlphaGo bested Lee Sedol, one of the world's top Go players, in four out of five matches in March 2016. It also beat European champion Fan Hui 5-0 in October 2015. The match was being played in China, the place where the abstract and intuitive board game was born. The government, however, isn't a big fan of letting its citizens know about the battle and has censored all the livestreams in the country.
Being able to read deeper than a human in go is insufficient. There are too many positions to calculate even close to all of them, and a given position is very difficult to determine who is even winning. The news this time is that AlphaGo is beating the player who is considered to be the best player in the world. This match in particular is like when Deep Blue beating Gary Kasparov. If you're wondering why it winning a second game against Ke Jie is news, it's because it sort of proves that the first game wasn't a fluke. If AlphaGo could only beat Ke Jie in 1 in 7 games, we wouldn't say that it is better than him. So that it is continuing to win is impressive.
These are the first couple of games at long time controls with the new machine, so it's not beating people over and over. And this is a new challenge, Ke Jie is considered the best player in the world, and has been able to study AlphaGo's style, so arguably a tougher challenge than Lee Sedol.
AlphaGo is training the Go players to become better players with every defeat. When I had the Sargon II chess game for the Commodore 64, it took me five years to learn how to consistently beat the game on the hardest difficulty level. I remembered the last the game I played because I had to defeat nine Queens to win the game.
Let's consider one thing as well... How much electricity is Alpha Go's hardware using vs that poor human brain? Efficiency wise we still have the computer beat, even if we don't win and just come close.
Like your quoted text says: AlphaGo optimizes for the winning probability, not stone difference. A win is a win. If it wins 99% of its games with minimal differences, you could still say it leaves humans in the dust. Of course, we've only seen 2 games, so it's a bit too early to call that yet.
On BBC news front page: http://www.bbc.com/news/techno...
Just because China censors something doesn't mean it didn't happen.
Don't be fooled by humans ability to play Go.
Humans are capable of basic rudimentary communication.
Although humans aren't truly intelligent, their behavior can at times mislead you into believing they are. However humans simply don't have a complex enough brain to have real thoughts.
All leading experts agree that on the evolutionary scale, space flight is a pre-requisite to intelligence. Like many species, humans are incapable of flight. Humans can construct very rudimentary machines enabling them to achieve atmospheric or space flight. Having no natural ability to fly demonstrates a true lack of any real intelligence in humans.
Humans can be trained to do tricks in exchange for a reward. Almost all humans will do various tricks in exchange for money, travel or entertainment. Some humans can even be coaxed to perform scientific experiments, or even publish a new scientific theory in exchange for scientific knowledge or gadgets. Don't be fooled by this. Their inability to instantly master new languages is one of many obvious indicators that humans are not truly intelligent. Humans actually believe that they created the machines.
Give humans some raw materials of high quality, and they can construct things demonstrating a rudimentary ability to create order and structure. Their adaptability allows them to use a wide variety of building materials such as steel, wood, plastic and stone to make their own nests.
Well maintained humans are completely safe in society. They are commonly seen nowdays in public. They are generally not known to hurt or attack. Get them vaccinated. Keep your license up to date. It is good to have a tag on your human in case it gets lost.
Humans are content with very little. They are inexpensive to keep and maintain. Provide fresh food and water. Play with your human every day. A human will be perfectly happy with an old, nearly worn out, obsolete Intra Glactica Net connection. They aren't demanding. Bring your human a new high tech toy every day and they will always be very happy to see you and greet you with great affection.
Humans are great at watching over the living unit when you are gone. They are instantly house-trained, and can be left indoors. Treated well, a human can almost seem as if it could genuinely love you.
I'll see your senator, and I'll raise you two judges.
Total aside - I once was playing someone I would normally win against, but I screwed up and he completely overran me. He decided to show off by getting about six queens on the board.
The look on his face when he got the last one and realised I had placed myself in a stalemate position ready for when he did so...
You beat the computer equivalent of an average club-level player.
It was the only chess game I had for the Commodore 64. I never got around to getting Battle Chess.
Here you go, they are working on it:
DeepMind has already begun working with the UK's national health service to develop apps and other tools for diagnosis.
That's not how games and sports work. No matter how slim the margin, no matter how much chance may have been involved, if one side wins, that's Destiny.
Look at last year's World Series. The Chicago Cubs' 109-year championship drought finally came to an end. It didn't matter one bit that the fate of their century-long losing streak ended up hinging on just a couple of plays in the 10th inning of the final game. Nobody went around saying "Yes, but they barely won."
You know, just maybe, the point of making an AI the best player in the world at Go, is to learn -how- to approach the 'real' problems afterwards. Just sayin'
It's not about the man-vs-machine showdown that the media sells. Go is about our quest to understand the game, individually and collectively.
AlphaGo is exciting because it is a breakthrough in our understanding of the game. Playing against a stronger opponent has always been a great method to improve your own game. You lose, you study your loss, you repeat. That's something every go player appreciates, but top professionals can't do that. They are the strongest so they have trouble finding stronger opponents... until now. In about 70 public games played AlphaGo has already made an impact on top level play and on our collective understanding of the game. What's to come is even more exciting.
I think I speak for every go player when I say we'd love to get beat by AlphaGo every single day.
For those of us watching from the sidelines, it was a fantastic match, an outstanding performance by Ke Jie (the human). Up to a certain point in the middle-game he played perfectly by AlphaGo's own evaluation. See here https://twitter.com/demishassa...
Seriously, why does China feel the need to censor this? Aren't they all in with technology - this isn't some religion coming up from below to overrun the great Communist State.
Posted from my Android phone. Oh, I can change this? There, that's better...
It doesn't really say anything about humans being obsolete when an AI trounces a human at tic-tac-toe.
Last time that came up there was the threat of Global Thermonuclear War. The computer was even innocent about it "Want to play a game?"
In many games and sports the final score can be misleading. If you want to know how close it was you have to watch (and understand) the match.
In match 1 AlphaGo won by half a point, the smallest margin of victory possible, but it was not a close game. AlphaGo was leading since relatively early. It's AlphaGo style to "bleed" points away when its leading to make the game simpler, safer, and ultimately still win. AlphaGo is very good at calculating the would-be final score.
In match 2 well into the mid-game it was still dead even. Both players kept raising the stakes again and again, so when the bottom finally came off the difference in points was huge. Yet it was a very close match and a superb performance by Ke Jie.
In their calculation, a Chinese national got beat by an American AI machine. Now drop out "national" and "AI machine", that's all the Chinese censors needed to know.
There is a difference between a narrow victory and 'leaving someone in the dust'. Don't pretend there isn't.
Nobody went around saying "Yes, but they barely won."
We aren't talking about the ways society celebrates victory, we're talking about the accuracy of the claim that the AI left the human player 'in the dust'.
Marketing. Of course it has no real purpose. We all know that computers are good at running programs with strict rulesets.
Uh... duh? GP was being sarcastic and making a joke...
Yeah I don't think we're anywhere near *real* AI. I mean we don't have any idea how the human brain really works, the only example of an intelligence we know about, and neural nets don't work anything at all like a human brain does, the design is merely "inspired" by the concept of neurons connected to together in same fashion with a feedback loop.
We all know that computers are good at running programs with strict rulesets.
Except computers were lousy at playing Go just a few years ago.
And now they are good at it. They were lousy at playing Chess at one point too. What game should they be good at next? Who cares? Computers can do ANY logic game better than any human. That is the one thing they are particularly good at.
It's not that simple. Partially, it's not blinded by 3,000 years of what the "right" move is and it makes the actual right move. Thus- human go players will be able to learn from it.
But mainly, it has much more experience playing than human players. Picture a go "savant" who was able to play 2,500 games a day... now stretch your imagination to 25,000 games a day. Now make that 250,000 games a day. Alphago plays more games than that per day against the best player in the world (itself).
It still learns playing against itself all the time.
She was like chocolate when she drank... semi-sweet at first and then increasingly bitter.
Don't mention that here. The AI nutters will eat you alive!
Nope. If you lose, you're simply a loser. It doesn't matter whether it was by half a point, by a random bounce, or by a bad referee call.
But you're wrong in another way as well: For many years, there have been people posting here who claimed "Go is so much harder than chess! Computers will never be able to beat humans at go!." Those claims are most definitely left in the dust now.
Moreover, after centuries of practice, humans aren't going to get significantly better at playing Go. Computing systems are just getting started. Humans will undoubtedly be left hopelessly behind very soon, just like they have been with chess.
Except that it isn't all they need to know... the fact that this computer beat the world's best Go player doesn't lessen the worth of the player or devalue China's reputation in the slightest. The point of the exercise was to demonstrate that the game of Go is sufficiently computable that it is possible to design software that will never lose to any human player. If the program can consistently beat the best player in the world, then its victories can probably be discounted as mere luck or caused by mistakes that the human player made, and the goal has been achieved. That the best player happens to be Chinese is superfluous to this, and while I understand that there can be some sense of national pride in a country having the world's best player at some sport or other event, this match doesn't even change that fact about their player. He is still the best player in the world, and until some other person beats him, that fact will remain. Politically, the outcome of this human vs computer match is actually entirely neutral, and I am greatly concerned that a government that would want to place such political significance on the results of a game such as this that they felt the need to censor it have a sufficiently misplaced set of priorities that will invariably be quite detrimental to their country.
File under 'M' for 'Manic ranting'
You don't understand Go.
It's is DOZENS of ORDERS OF MAGNITUDE more complex than even chess, which we only "beat" in the last, what? 20 years?
It's some many DOZENS of ORDERS OF MAGNITUDE more complex, that nobody was realistically expecting computers to outclass humans this century.
Look at my previous posts and you'll see that I *HATE* the term AI, deep learning or anything else, because those things that claim to be that categorically are not.
But Google beating Go is way more impressive than "we built a big computer". It's even more than "we built millions of big computers and joined them together". It's honestly something slightly different - towards thinking and away from brute-force.
I'm not claiming it's anything NEAR being actual AI. But it's been surprising to anyone who understands this field.
My professor in university (professor means something in the UK, way beyond "doctor" or "teacher") was writing these kinds of Go-playing programs for his research. He was a mathematician that dedicated his life to analysing the game and game theory around it, and would travel to Go conferences the world over. And he would never have expected this result to be possible in his lifetime.
I mean we don't have any idea how the human brain really works
We don't have to. We know enough that we can create a self learning machine that beats a human at a highly intuitive game. From here, it's just a matter of scaling and improvement, and we can even use a computer to help with the design.
Our own brains have evolved from simple primate brains in less than a million generations, without anybody around to understand how they work.
The real intelligence isn't in the people playing the game, but the people who made the game in the first place. This is why AI is a misnomer here.
What they SHOULD do, is limit the power the computer playing the human can use. Since the Brain uses around 20W, the computer should be limited to the same. It's not just to keep things kinda even, it's an additional constraint that could drive development of the system.
What game should they be good at next? Who cares? Computers can do ANY logic game better than any human.
Realistic first person shooter games, driving cars, and designing better computers.
Humans set up the methods for teaching alphago but the actual learning is so complex that humans can't understand it.
In your analogy, it's a bit more like the brilliant grandson beats others. The grandfather knows how he instructed his son and a bit of how his son instructed his grandson, but the grandson is so far advanced from the grandfather that his methods can't be understood well.
She was like chocolate when she drank... semi-sweet at first and then increasingly bitter.
your 'learning machines' are not going to 'evolve' into fully conscious minds,
-- anonymous coward on Slashdot
heavier-than-air flying machines are impossible. I have not the smallest molecule of faith in aerial navigation other than ballooning.
-- Lord Kelvin, not an anonymous coward.
I wonder how many people will get the reference.
Sorry, fresh out of mod points.
Funny, I could beat it very easily at kickboxing...
Non-Linux Penguins ?
It's a matter of complexity. Tic-tac-toe is simple enough that it is trivial to write a program in a procedural fashion such that the computer will always either win or come to a draw, but never lose. I did this myself in high school for a class project. That was possible because the board is very small and the rules only allow for a maximum of 9 moves between the two players. The computer doesn't have to do any prediction about future game states but can just refer to a small lookup table of game states and make the appropriate best move.
With Go the board is relatively large at 19x19 and the rules allow for pieces to be removed allowing for play on the same position again at a later time. The result is that the number of possible game states is immense, well beyond our ability to catalogue them. Chess faced the same problem though it has a much smaller set of possible game states. The solution in chess was to predict what the board would look like a few moves in the future and pick moves that would result in a stronger board state based on parameters specified by the programming team. Alpha Go is doing the same kind of thing but with a huge caveat. In the case of Deep Blue all of it's decisions were weighted by the programmers so that it would know which board states were stronger. In the case of Alpha Go the computer was pitted against its self in thousands of games so it could learn and assess it's own values for future board states.
The benefits of the Alpha Go approach have been made apparent by its use of moves that had never been seen in tournament play. Whereas Deep Blue seemed to play much more traditionally.
Of course, we've only seen 2 games, so it's a bit too early to call that yet.
Against the greatest human Go player. Those two games tell us all we need to know.
The human can dedicate their life to Go, and focus on it to exclusion. But at the end of the day, that human needs to eat, sleep, and take time off to rest. That human is one brain, and has one set of inputs, and can chew on one situation at a time.
AlphaGo can add in more processors, memory, and suck in hundreds of thousands of matches to learn from. AlphaGo can be cloned an infinite number of times, and play against itself in parallel, learning from every match.
Two wins against the top human in the world is really all that we need to know about how well it plays Go. Humans just can't learn this particular task as fast as the AI can, and can't likely learn it as deeply either. How long did AlphaGo take to learn to compete at the highest level compared to how long the humans did? A couple of years? Unless you think this is the absolute best the computer can do, another couple of years and its abilities will look nearly magical compared to human players.
Velociraptor = Distiraptor / Timeraptor
Realistic first person shooter games, driving cars, and designing better computers.
Going along with GP's confusion between 'computers' and 'software', there are already people who use neural nets to optimize the design of other neural nets. So that last part is already beginning to be covered.
Going along with GP's confusion between 'computers' and 'software',
The confusion doesn't matter, because both hardware and software can be improved by neural nets. It's nothing but a complicated logic game, and "Computers can do ANY logic game better than any human".
Yes but computers are getting faster, so at some point they were going to be fast and powerful enough to be good at Go.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
Ok but what real life application does this have?
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
I for one welcome our.....screwit--the bots are coming, run!
Table-ized A.I.
Not be just being faster, they wouldn't. Take a bad algorithm, and make it a million times faster, and it's still bad. A factor of a million in Go only gets you a few extra levels of depth, which is useless if you're still using the same bad evaluation of the board.
Maybe you should read about the technology of AlphaGo first before you make such a claim. It uses not just computing power but also a combination of fancy algorithms such as neural networks. Neural networks are motivated by our brain. And you have to admit that our brain is some physical device - it can be simulated to some point by a "searching and processing data" device.
I wonder how many people will get the reference.
I know only one reference to jumping queen.
https://www.youtube.com/watch?v=vrFOr8oF95A
https://www.youtube.com/watch?...
The original lyrics were "I get my kicks above the waistline, sunshine".
Well played by the AC.
I *HATE* the term AI, deep learning or anything else, because those things that claim to be that categorically are not.
So what term should we use for a bot that plays Go?
Ok but what real life application does this have?
Saving lives, actually. https://deepmind.com/applied/
The original lyrics were "I get my kicks above the waistline, sunshine".
I've never heard that song before. My father's truck only had two radio stations: talk and country. ;)
For a non-club-level player, wouldn't that be exciting?
I'd find that exciting.
You just used it.
"Kill" Go? No. Much like Deep Blue did with Chess, AlphaGo will bring a renaissance of innovative play to Go, and it has already started. Much conventional wisdom is being re-evaluated already in light of AlphaGo, and moves that were once considered "correct" in certain situations are already being reconsidered as sub-optimal.
Yeah, think of it like sports. Encouraging kids to go outside and do something to get fit is laudable. To that extent some people are... icons of a particular sport and are rewarded with fame and cash and a career.
Chess, Go, and games in general are mental equivalents. Instead of stretching muscles and getting physically fit, it stretches you mind and makes you more mentally fit. Thinking and deep thought are skillsets that people need to exercise. And Ke Jie is an icon at the top of the game, presumably as a star figure encouraging kids to go play the game.
And to that extent, chess and go programs help people learn the game. ...That said, we take this shit WAY too far. Especially shit like football where most of the fans don't even play the damn game themselves. Way too many kids are treating a sports scholarship like the goal they're working towards rather than the means to a real education. Some are even deluded into thinking they can go pro. I've heard that if a kid in China is good enough at Go, they get pulled out of regular school and go to Go school. Way to take your best and brightest and put them on a path that doesn't actually directly help anyone. Well, I imagine most drop off of that at some point and go chase a more productive career, and that probably works out fine.
The only thing this says is that AI is a threat.
How about showing cases where AI adds good (and instantly available) jobs for the displaced, especially the long-term jobless? Not service/staffing jobs, but actual lines of work with an actual future.
Twitter supports and protects racists - by smearing their critics with the "Hate Speech" label.
So the next step is that it needs to play itself?
If it beat a human that it determined was playing "perfectly", that's the only next opponent.
Perhaps they will teach even go masters something?
-Styopa
AlphaGo's playing style seems similar to that of some great chessplayers:
Karpov: ... Karpov's "boa constrictor" playing style is solidly positional, taking no risks but reacting mercilessly to any tiny errors made by his opponents. As a result, he is often compared to his idol, the famous José Raúl Capablanca, the third World Champion. Karpov himself describes his style as follows:
...
Let us say the game may be continued in two ways: one of them is a beautiful tactical blow that gives rise to variations that don't yield to precise calculations; the other is clear positional pressure that leads to an endgame with microscopic chances of victory.... I would choose [the latter] without thinking twice. If the opponent offers keen play I don't object; but in such cases I get less satisfaction, even if I win, than from a game conducted according to all the rules of strategy with its ruthless logic.
Capablanca: ... Capablanca excelled in simple positions and endgames, and his positional judgment was outstanding, so much so that most attempts to attack him came to grief without any apparent defensive efforts on his part. However, he could play great tactical chess when necessary ... He was also capable of using aggressive tactical play to drive home a positional advantage, provided he considered it safe and the most efficient way to win, for example against Spielmann in the 1927 New York tournament. ...
or maybe Petrosian: Petrosian was a conservative, cautious, and highly defensive chess player who was strongly influenced by Aron Nimzowitsch's idea of prophylaxis. He made more effort to prevent his opponent's offensive capabilities than he did to make use of his own. He very rarely went on the offensive unless he felt his position was completely secure. He usually won by playing consistently until his aggressive opponent made a mistake, securing the win by capitalizing upon this mistake without revealing any weaknesses of his own. ... Petrosian was known for his use of the "positional exchange sacrifice", where one side sacrifices a rook for the opponent's bishop or knight. Kasparov discussed Petrosian's use of this motif: "Petrosian introduced the exchange sacrifice for the sake of 'quality of position', where the time factor, which is so important in the play of Alekhine and Tal, plays hardly any role. Even today, very few players can operate confidently at the board with such abstract concepts. Before Petrosian no one had studied this. By sacrificing the exchange 'just like that', for certain long term advantages, in positions with disrupted material balance, he discovered latent resources that few were capable of seeing and properly evaluating."
"Moreover, after centuries of practice, humans aren't going to get significantly better at playing Go."
If one of the best 20th or 21st century chessplayers went back in time to play the best 19th century players, the 19th century players would lose, and I think it unlikely you could justify making a similar comment about chess, which makes me suspect that it's not true for Go.
Examples: in 1873 Steinitz" introduced a major improvent in playing style: "...All of Steinitz's successes up to 1872 inclusive were achieved in the attack-at-all-costs "Romantic" style exemplified by Anderssen. But in the Vienna 1873 chess tournament, Steinitz unveiled a new "positional" style of play which was to become the basis of modern chess..."
Later, in the 20th century, the Russian (or more accurately Soviet) school of chess made a major improvement by emphasising the dynamic possibilities of positions.
If you wish: to cite only a few, in 1845 I gave the first mathematical development of Faraday's idea that electric induction takes place through an intervening medium, or "dielectric", and not by some incomprehensible "action at a distance", and it was partly in response to my encouragement that Faraday undertook the research in September 1845 that led to the discovery of the Faraday effect, which established that light and magnetic (and thus electric) phenomena were related; in 1846, at the age of 22, I became the Professor of Natural Philosophy at Glasgow University, and in 1900 I gave a lecture on the two major problems not solved by 19th century physics - how matter moves through the aether (including the puzzling results of the Michelson–Morley experiment), which was solved by Einstein's Special Theory of Relativity, and indications that the Law of Equipartition in statistical mechanics might break down, which led to the developent of quantum mechanics.
What, dear AC, have you achieved?
Yours etc, Lord Kelvin
Nevertheless, people have given up trying to beat computers at chess. People simply aren't going to catch up by making incremental improvements over the span of centuries, and the same thing is going to happen with go.
You think the brain is some magical device, seat of the soul or some such? It's a wetware computer, nothing more. Grow up bud, time to leave the fairy tales behind.
I do not want your cheap brainburning drugs. They are useless for work. And I am a working man today.
Am I missing something?
Everything, by the looks of it. Are you a computer? This is about getting a computer to do something that typically only humans can do (learn/improve a skill).
I do not want your cheap brainburning drugs. They are useless for work. And I am a working man today.
Go home everyone, Freeze128 isn't impressed.
Figure out a health-care bill that is equitable for all.
You think such a bill would get voted in? lol.
I do not want your cheap brainburning drugs. They are useless for work. And I am a working man today.
Nope. If you lose, you're simply a loser.
Denying the existence of numerical subtraction is a novel line of argumentation, I'll grant you that much.
It doesn't matter whether it was by half a point, by a random bounce, or by a bad referee call.
Yet again: I'm not concerned with the intensity of your emotional response. Still you refuse to grasp what leaving someone in the dust actually means: defeat by a significant margin. I really don't care whether you find the score-difference to be interesting, it remains that this is what the phrase means.
Looking at discussion elsewhere in this thread, it may have been Alpha Go's 'deliberate' strategy not to pull that far ahead: it may have made the 'choice' to play it safe and keep slightly ahead for the remainder of the game.
But you're wrong in another way as well: For many years, there have been people posting here who claimed "Go is so much harder than chess! Computers will never be able to beat humans at go!." Those claims are most definitely left in the dust now.
Uh... did I ever make that claim? No. As you imply, that's a stupid claim to make. In principle, there's no problem-solving task at which humans will necessarily outperform computers.
Moreover, after centuries of practice, humans aren't going to get significantly better at playing Go. Computing systems are just getting started. Humans will undoubtedly be left hopelessly behind very soon, just like they have been with chess.
Agreed, even acknowledging breakthroughs in 'human-play strategy' which colinwb points out.
It's clear that the headline is in fact about "emotional response", and it is completely appropriate. Everyone knows that the computers will only get better, and that human go players are in fact being left in the dust, starting today.
You need to stop focusing on literal interpretation and look at the context of the situation before getting all agitated about word choices.
Go more like a language by which you communicate abstract ideas
This idea of "points" and "winning" is really a tacked-on thing. You can increase your abstract literacy by increasing your 'strength' in Go and finding more experienced players to converse with.
But the point is not to win, it's to find more experienced people to talk with.
A machine has absolutely no use for this, and anyone controlling the machine has no use for this either. This whole show is simply about propaganda and the triumph of might over right by brainless brute-force data processing. It's purely about demoralizing people, not about showcasing meaningful technological advances.
Google is basically bringing a flamethrower to a carpentry contest and saying they won because they were 'done' with their wood first.
How Go is supposed to be played is this: people build their strength to gain literacy, then they play games to construct simulated situations that are abstractly parallel to their real-life problems and search for insights in the matter.
Tokugawa and Toyotomi for example used to play Go. Neither was a master of the game, yet both were masters of their domains in reality. Their strength in the game didn't matter, they were literate enough, and they were able to communicate through the game to their immense profit and the strength of Japan, and so took the time to play.
It's too bad most people really have no idea what the point of games in general is. Too much low self esteem, too much blind ambition.