The biggest weakness of AlphaGo is that inherent badness is only available in humans

Machines are not invincible. AlphaGo has made a mistake that should not be worth it: Overconfidence.

Korean player Lee Sedol won his first victory over Google's AlphaGo artificial intelligence system , after losing three previous games in a row. Although, after winning the first three games of the five-game match, AlphaGo won the final, but the remaining two games still took place even as a matter of procedure. Anyway, Lee Sedol has regained a bit of pride for himself and for millions of people watching this match online.

Reporters and photographers applauded when AlphaGo declared a loss after a 5-hour game. And they screamed even more loudly as they entered the press room after the match.

The dominance of AlphaGo during the previous three games shows the superior power of modern technologies within it - technologies that can be used to reinvent everything from scientific research to robots. Meanwhile, Lee Sedol's victory this time shows that machines do not mean that they cannot make mistakes.

Moreover, Lee Sedol's victory in the fourth game shows that even the most talented AI machine has a long way to go before it can truly repeat human thoughts. A machine that can beat the top player on the Go, does not mean that it will pass a science test in the 8th grade, chat like a human or express common emotions.

World champion Lee Sedol.

What is AlphaGo's weakness?

As commentator Chris Garlock said before the game started, a big question: Does AlphaGo have any weaknesses?

It was a question in the press room after the third game, when Lee Sedol apologized to the Korean public and the Go Go community."I don't know what to say today, but I think I must apologize first." I said. "I should have a better result, a better match than the previous game." The Korean player admits to being under the influence of public pressure. The match is the focus of attention in Korea, where there are about 8 million go players. But now that the pressure of victory has ceased, he declares that he will continue to search for the weakness of the machine.

"Even though AlphaGo is a very powerful program, I don't say it's a perfect program." I said. "Yes, compared to humans, its moves are very different and there are times that are superior. But I think AlphaGo still has some weaknesses."

He was unhappy with the second game, when he felt he had made a serious mistake and did not take advantage of AlphaGo's errors . "There are some opportunities that I missed ," he said.

Play the second game again

The fourth game started very similar to the second game, when Lee Sedol tried to compensate for past mistakes. As in the second game, he played white, meaning that he was behind the opponent. And he responded to the opening step of AlphaGo similar to the second game."It's like the same game." Commentator Michael Redmond said after six steps of the game.

Commentator Chris Garlock asks if Lee Sedol wants to play the most unusual moves to find AlphaGo's weakness. But as Redmond points out, that doesn't really work in the first game. From AlphaGo's operating mechanism, it seems unlikely that unusual, or even exotic moves could work for this machine.

The fourth game started very similar to the second game, when Lee Sedol tried to compensate for past mistakes.

By using neural network technology - simulating the neural neural network of the human brain - AlphaGo learns to play the Go flag by analyzing thousands of thousands of players' previous moves. Then, combined with intensive learning technology , it plays chess itself to push its level to higher levels. In essence, these self-help games help create new moves that computers can use to retrain themselves. Therefore, this is not the human move.

In other words, the AlphaGo system does not operate to play in the usual way, it plays in a way that people will never do.

Outside the match

When the match took place, Lee Sedol took a lot more time for each move than his opponent. The same thing happened in Monday and Tuesday, so when his watch runs out of time, the Korean player must increase the speed of placing troops.

Meanwhile, on the contrary, AlphaGo manages time quite well, and it hardly makes any mistakes. Before the match, Demis Hassabis, who oversaw the AlphaGo team, said a neural network was added to AlphaGo to control the time.

Some commentators have come up with their "strategy" : "I just need to unplug the plug". Redmond talks about AlphaGo. " It depends on the Internet connection, right? All we need is someone with scissors." Indeed, AlphaGo relies on an Internet connection , to link up with a vast network of Google data center facilities worldwide. But a scissors is not enough to cut that rope. Before the match, Google ran a separate fiber optic cable in the Four Seasons Hotel to ensure connection quality.

AlphaGo manages time quite well, and it has almost no mistakes.

When first building AlphaGo, Hassabis and his team train and run the system on a single computer. But by October, before battling the three-time European champion, Fan Hui, the team was upgraded with a higher processing power. Typical neural networks run on a large number of interconnected computers, each equipped with GPU graphics processing chips, for machine learning algorithms. In October, Hassabis said AlphaGo ran on a network with 170 GPU cards and 1,200 standard processors.

A dangerous battle

When the clock clocked for 2 hours, Redmond called the match a typical Lee Sedol game, seeing it growing "into a very dangerous battle". Lee Sedol likes to walk on the blade and he is good at it. But as Redmond points out, it's AlphaGo.

Lee Sedol now seems to have a better position in the third game. He was also calmer. But after about 25 minutes of playing, Redmond, also a good goer, felt AlphaGo was in a tough position. And Lee Sedol's time is only 25 minutes on the clock, almost an hour less than AlphaGo. This difference is very important, because when your watch runs out of time, the player must make the move go in less than 60 seconds.

At this point, AlphaGo started playing in a way that Redmond and Garlock found unimpressive or "sluggish" moves . This sign shows that the machine is too confident about its victory. Because AlphaGo is programmed to maximize the likelihood of victory , not the score of victory. "This is like AlphaGo saying: I think I'm in the lead. I'm going to end this game." Garlock said. "And Lee Sedol needs to do something special, even if it doesn't work." Meanwhile, Lee Sedol slightly leaned forward, focusing his thoughts, taking a few minutes for his next step.

Lee Sedol's move 78 and AlphaGo's 79: the turning point of the game.

And after a few minutes, a big move appeared. It was the seventh move, when Lee Sedol placed one of his troops - a spike - in the middle of the board."It is the most complicated country to play." Andrew Jackson, an online commentator for the American Go Association, said. And this complexity changed the machine's gameplay. Demis Hassabis later posted on Twitter that AlphaGo made a serious mistake in the 79th move, right after Lee Sedol's big move. And about 10 moves later, AlphaGo's calculations show that its chances of winning are gone.

Surrender?

After that, almost at the end of the three-hour clock, Lee Sedol's watch ran out of time. But AlphaGo's play continues to confuse commenters."I have the impression that AlphaGo has lost its route." Redmond said. Again, this shows that AlphaGo is having a hard time, but it reminds Garlock to question whether the machine will ever surrender.

According to David Silver, another researcher in the team that created AlphaGo, the machine will surrender, not when its ability to win is zero, but when the ability to win is below 20%."We feel this is more respect than the way people play this game." Silver said. "It is disrespectful to continue playing in a position that is clearly close to failure."

At this point, the machine still hasn't decided its chance has fallen below that threshold. But according to commentators, its position has started to deteriorate."I have a feeling that AlphaGo is running out of water to win." Redmond said, adding in the past, AlphaGo used to walk poorly when it found itself ahead.

In the end, both Redmond and Garlock agreed that the game was going well with Lee Sedol. But Korean players continue to face difficulties in time. He made two mistakes when he made the move for more than 60 seconds. Then he went to make the move in just a few milliseconds before the clock ran out. If he made another mistake, he would have only 30 seconds to make the move.

When the game was 3 hours 40 minutes, Lee Sedol got up and left the room to take a break. The drama of the game was pushed to the highest level from the first game."Lee Sedol is having a chance," Redmond said.

The moment AlphaGo surrendered, the cheers rang out from the Korean commentary room and the crowd of reporters.

End of the match

But the Korean player still faces the trouble of time. And AlphaGo hasn't surrendered yet. The machine still has 15 minutes left on the watch. After that, AlphaGo went to a country that Redmond called "another meaningless move". Soon after, the machine surrendered.

After the match, Redmond and the others pointed out the turning point of the game is Lee Sedol's move 78. In the middle of the game, Lee Sedol is still behind. But he reinforced that move into the center of the board and redirected the game."Lee Sedol went to a great country. It surprised me. I'm sure it will surprise most opponents. I think it surprised AlphaGo , " Redmond said.

The moment AlphaGo surrendered, the cheers rang out from the Korean commentary room and the crowd of reporters. Then came the applause from the comment room in English. Contrary to the atmosphere of a day ago here. On Saturday, Lee Sedol found a weakness of AlphaGo. And now that the emotions have changed, people can finally win.