Slashdot Mirror
Google's AlphaGo AI Secretively Won More Than 50 Straight Games Against World's Top Go Players (qz.com)
An anonymous reader quotes a report from Quartz: When Google's artificial intelligence program AlphaGo made history by taking down Korea's Lee Sedol -- one of the world's best Go players -- in a landslide 4-1 victory in March, Chinese player Ke Jie was skeptical. He famously wrote on Weibo the next day, "Even if AlphaGo can defeat Lee Sedol, it can't beat me," and has since agreed to take on the AI at an undecided time. But now even Ke, the reigning top-ranked Go player, has acknowledged that human beings are no match for robots in the complex board game, after he lost three games to an AI that mysteriously popped up online in recent days. The AI turned out to be AlphaGo in disguise. On Jan. 4, after winning more than 50 games against several of the world's best Go players, Ke included, a user registered with an ID of "Master" on two Chinese board game platforms came forward to identify itself as AlphaGo. "I'm AlphaGo's Doctor Huang," the user "Master" wrote on foxwq.com, according to screenshots from Chinese media reports. Taiwanese developer Aja Huang is a member of Google's DeepMind team behind the AI. Since Dec. 29, Master has defeated a long list of top Go players including Korea's Park Jung-hwan (world No. 3), Japan's Iyama Yuta (No. 5) and Ke in fast-paced games. He won 51 games straight before his 52nd rival, Chen Yaoye, went offline, forcing the game to be recorded as a tie. By Jan. 4 when the test was completed, Master had racked up 60 wins, plus the one tie, and zero loss, according to numerous reports (link in Chinese).
He won 51 games straight before his 52nd rival, Chen Yaoye, went offline, forcing the game to be recorded as a tie.
So the only way to win is not to play.
Live today, because you never know what tomorrow brings
I'm waiting for the AI Rust players.
I should use this sig to advertise my book ISBN-13 : 978-1501515132.
Don't forget; the Master Control Program started off as a chess program. Remember Encom!
An official confirmation from Demis Hassabis, a co-founder of DeepMind.
I'll be impressed when you write an AI that can competently play Civ5.
I'm not really sure if it's a more difficult problem than Go or not (I'd think so with all of the decisions to be made), but holy hell is the shipped AI in all Civ games useless.
This is not AI, simply because the rules of Go were programmed into the computer to start with. If it had to figure out the rules and the idea of winning by itself then that would be amazing...
However it was taught what a good move is by some point or similar system, that's hardly self learning....
Are you kidding? Then only a very small %age of humanity could be considered intelligent as there are a lot of things that a vanishingly small number of children would NEVER figure out without help.
Pain is merely failure leaving the body
You seemed have to missed the fact that many of the top professional players were lining up to play against this bot. They view it as free training lession, not to beat an AI bot, but to beat their human counter parts. Since Lee Sedol played against AlphaGo, he has gained in strength, so much even that a certain point, using a certain method, AlphaGo was the strongest player, not because it had played more games, but because Lee Sedol had won so many games. Ke Jie, to be considered the strongest player at the moment, has made remarks that humans have only touched at the truth behind go, after he played against Master(P). Most go players have a very high regard for the game, as they sense that it is much deeper than human mind can consider. For this reason, I guess, many professional go players find this a very exicting time, because it will enhance their understanding of the game. In this view it is very unlikely that a professional player will use a trick to force a tie.
It is indeed true that one neural network was trained using a collection of games, but the version of AlphaGo that played against Lee Sedol last year, was using two neural networks, and the second one (for evaluating the positional strength of a board configuration) was trained by letting AlphaGo play against itself. It is not known how he current version of AlphaGo works, whether any additional neural networks were added, but if it has become stronger, it has done so by playing against itself. It should be noted that this latest version of AlphaGo was playing some surprising moves that made some people believe, it could not have been AlphaGo, or that it was a version that was trained without using a collection of games. It should be noted that DeepMind announced it wanted to experiment with training a neural network with zero additional knowledge for the game of go, just like they did with the neural networks playing old video games.
Go games typically end in a resignation, actually. Even decent players can tell when they're going to lose, and playing to the end when it's obvious you're going to lose is considered very rude.
Karma: Terrifying (mostly affected by atrocities you've committed)
As we see yet another instance of hubris in action, this time the assertion of Go players and hangers-on that "Go is so much more complex than chess, so it will never be mastered by a machine". Computational complexity or large problem space has little to do with either play-ability or ease of mastery.
Next challenge?