Five players with seven world titles ganged up on AlphaGo. They still lost.

Tomorrow is Ke Jie's last game against AlphaGo.

Today Ke Jie rested and AlphaGo kept playing: a pair go match in the morning, a team match in the afternoon.

The pair go match put "Gu Li + AlphaGo" against "Lian Xiao + AlphaGo." Each side had one hour of thinking time, shared between the human and the machine. Once the hour was gone, each side had one minute per move.

Before the match, Gu Li said he would have to play as a student, working to understand AlphaGo's moves and see whether the two of them could cooperate.

The pair go match started at 8:30. The playing area had been moved from the second-floor game room down to the ground-floor hall where Ke Jie's press conference was held; the referee, the humans and the computer sat in the middle of the stage.

Unlike Ke Jie's games, the pair go and team matches drew far less attention. The media room seats 80 and held fewer than 20 people. The hall had about 30% as many spectators as it did for Ke Jie's games.

DeepMind sent two amateur players to place stones for AlphaGo. Teammates were not allowed to talk during play (not that a player could discuss anything with AlphaGo anyway). The order of moves was Gu Li (black) — Lian Xiao (white) — AlphaGo (black) — AlphaGo (white).

At 10:59, 2 hours and 29 minutes in, Gu Li raised his card and resigned.

Half an hour earlier he had been winning. Lian Xiao had already used up his thinking time and was down to the one-minute countdown, white was behind on the board, and Gu Li still had 17 minutes. "The game is almost over," said Hua Yigang, vice chairman of the Chinese Weiqi Association.

"In the middle game, Gu Li's lead was at one point very large," said Liu Jing, 8-dan.

Then AlphaGo (black) and Gu Li each played a bad move and the whole position turned over. In the analysis room, Nie Weiping and Ke Jie thought Gu Li's black stones were in trouble. On stage, Gu Li shook his head.

Before he raised the card, AlphaGo (black) had already moved to resign, and Gu Li refused. He wanted to see how AlphaGo behaves when its odds are bad.

In pair go, a resignation needs both players on a side to agree.

"AlphaGo resigns when it puts its own chance of winning below 30%," said Wang Lei, 8-dan, who commentated on the pair go match.

In the end Gu Li could not repair the damage, and finally raised the blue card. His partner entered the command into AlphaGo and confirmed. Lian Xiao's side, playing white, had won by resignation in the middle game.

"I was wondering whether AlphaGo had blundered. I turned around and looked at the computer, thinking it had played a mistake, and I had no idea what it was thinking. Lian Xiao and I both went 'ah,'" Gu Li said afterwards. "I should have saved more time for later — I often miscount the score in the countdown. I had wanted to play this one all the way into the endgame to see how AlphaGo handles it..."

At the press conference, Lian Xiao said he had come to admire the DeepMind team after AlphaGo beat Lee Sedol, and that Master's appearance made people realize how much room Go still has for new ways of playing. "AlphaGo may be a player hundreds of years ahead of us, but it is not completely invincible. It does back down, which is very human. I think we will improve alongside it."

Five players gang up on AlphaGo in the team match, and still lose

The match where humans ganged up on AlphaGo is over. On a tangled board, five Chinese players joined forces and still lost, after exactly four hours of play.

The team match kept the Chinese rules: black moves first, and black pays komi to white at the count. Thinking time per side was cut from the three hours of the individual games to 2.5 hours. China's five players were:

• Zhou Ruiyang: 9-dan, best international result the 2012 Bailing Cup;
• Mi Yuting: 9-dan, best international result the 2013 Mlily Cup;
• Shi Yue: 9-dan, best international result the 2012 LG Cup;
• Tang Weixing: 9-dan, best international results the 2013 Samsung Cup and the 2016 Ing Cup;
• Chen Yaoye: 9-dan, best international results the 2012 Chunlan Cup and the 2016 Bailing Cup;

Dr. Aja Huang again served as AlphaGo's "human arm." At yesterday's press conference after the second game, Ke Jie called Dr. Huang the key figure in AlphaGo, and said sitting across the board from him was like playing a robot.

"You can see it on the broadcast, I move around a lot. Dr. Huang is the opposite, very calm. He sits there without moving, doesn't go to the toilet, doesn't drink water. I really admire it. I don't know how he does it," Ke Jie said, laughing.

Zhou Ruiyang sat across from Dr. Huang and played the stones for the Chinese team; the other four sat behind him working out the position. After each AlphaGo move, Zhou Ruiyang turned around to discuss it with them. Ke Jie had wanted to be the team's human arm, but the others talked it over and worried he would play moves of his own, so they said no.

Mi Yuting thought playing as a group had advantages and disadvantages, and that it came down to how the team performed on the day. Gu Li thought a team of five should normally hold a small edge, "but one problem is that the team match only has 2.5 hours. If the players don't agree and start arguing, there won't be enough time."

The Chinese team looked relaxed at the start. "Zhou Ruiyang smiled — he may be quite happy with how the game is going," Gu Li said. Whether he was playing a stone or turning around to talk, the smile stayed on Zhou Ruiyang's face.

The oldest of the five on stage was Chen Yaoye, born in 1989. Gu Li rated them highly. "Tang Weixing is very active and plays especially well in big tournaments. Mi Yuting won his first world title against me — I made a bad, basic mistake and lost my confidence on the spot, and felt there was already a gap between me and the young players born in the 1990s."

As the game went on, the Chinese team fell into long deliberations again and again. "Changing your move after you have decided is usually bad. It breaks your train of thought; better to follow one path through... I think they still lost a move on the lower side, and used a bit too much time," Gu Li said.

By then AlphaGo had used 20 minutes of thinking time and the Chinese team more than an hour.

"AlphaGo takes under a minute a move, basically," Gu Li said. Most of the time Zhou Ruiyang was the only one smiling; the four behind him sat with their chins in their hands, looking grim.

"Zhou Ruiyang seems to be the opposition. He thinks from black's side and decides none of his teammates' moves work — not this one, not that one."

"Time is tight" was the line Gu Li repeated most often in his commentary. When he came off shift, he said the team's position did not look good — slightly worse than Ke Jie's yesterday.

Away from the board, Sina reported that Ke Jie played a move from AlphaGo's side of the position. Everyone else thought it was poor, and while they were still arguing about it, AlphaGo's stone landed on exactly the spot Ke Jie had chosen. The arguing stopped.

"According to AlphaGo's evaluation, for the first 50 moves, or the first 100, the two sides were very, very close." That was DeepMind CEO Demis Hassabis on Ke Jie's second game yesterday.

The game began at 12:30. By 15:14 the Chinese team had 25 minutes of thinking time left and AlphaGo had 1 hour 52 minutes. "Zhou Ruiyang is going to be racing the clock," Gu Li said. "Smiling like that, it must be bad." On camera Zhou Ruiyang had his chin in one hand, turned sideways and leaning forward, smiling as he propped himself on the table.

Ke Jie and Nie Weiping now put the team's chances at around 30%, after losing too much in the opening.

In the endgame the Chinese team clawed back some of the deficit. "Mr. Liu, the position looks a lot better than when we sat down," said Wang Lei, who had taken over the commentary from Gu Li, to his partner Liu Jing.

Part of the reason is that AlphaGo plays every stone to maximize its chance of winning, nothing else. Put simply: given a 99% chance of winning by 5 points and a 100% chance of winning by 1 point, AlphaGo takes the second.

Human players can use this to narrow the margin. They cannot change the result.

"Now it's only a question of whether AlphaGo wants a big win or a small one," Wang Lei said. "It loves beating you by half a point." Half a point is the smallest margin in Go. Ke Jie lost the first game by exactly that.

When the Chinese team used up its 2.5 hours and entered the countdown, Tang Weixing took over from Zhou Ruiyang. Wang Lei explained why: in an earlier practice team match, the five Chinese players had played Ke Jie, and there too it was Tang Weixing who came in during the countdown, with the position slipping away, and turned the game around.

No miracle this time. "At this point the game looks hopeless. Five players shouldn't be coming out of the opening like this," Liu Jing said.

Only Tang Weixing was still fighting. The teammates behind him had started chatting.

"Playing on now doesn't mean much," Liu Jing said.

At 16:29 the five conferred again and found a good move, which drew a reply from AlphaGo that no human would play and narrowed the gap to "half a point to one and a half points." That move made Dr. Aja Huang, a strong amateur himself, shake his head.

At 16:30 the Chinese team resigned in the middle game.

"I was very excited about the chance to play a team game against AlphaGo, and very grateful to my teammates for trusting me to play the stones. But my performance today was probably not satisfying," Zhou Ruiyang said at the press conference. "Our preparation for playing black may not have been thorough enough. AlphaGo has a lot of moves with white — its fourth move already surprised us... As people expected, we never really had a good chance in the whole game."

What is the point of humans ganging up on AlphaGo?

An earlier article raised this question.

The point is clear enough. If an artificial-intelligence algorithm can beat one human, can human collective intelligence and teamwork beat it? And can a human-plus-machine pairing produce a better result than either alone?

Nie Weiping thinks so. He wrote on Weibo: "Had a very good talk with Hassabis yesterday. Human players partnered with AlphaGo could beyond any doubt beat AlphaGo on its own. We were both completely sure of that."

Pairings like this have been tried before. In 2005, Playchess.com ran a "freestyle" chess tournament in which entrants could team up with anyone or anything, human or computer. Normally chess sites run "anti-cheating" algorithms to stop entrants from using chess programs.

With a large prize on offer, several grandmasters entered with multiple computers in tow. At first the results were what everyone expected: a human teamed with a computer beat the strongest computer by a wide margin.

Hydra, like Deep Blue, was a dedicated chess supercomputer, yet it lost to a fairly strong human player with a fairly weak laptop. Human strategic direction plus the computer's exact tactical calculation.

The eventual winners were not a "human + high-end computer" team, but two American amateurs with three computers.

Their trick was to intervene and "coach," pushing the computers to search particular positions more deeply. That approach beat the grandmasters' deep understanding of chess, and it also beat the raw computing power of the other entrants.

"Weaker human + machine + a better process" beat the powerful computer on its own, and beat "stronger human + machine + a worse process."

"The pair go and the Chinese team game were both fun and fascinating. They hint at what humans and machines can do together. Congratulations to everyone involved." Hassabis posted that tweet after today's two matches.

David Silver, AlphaGo's lead programmer, also praised the games and the seven Chinese players: "Today's games were not about winning or losing, but about exploring the many variations of Go and how we can learn from them... Maybe this gives us some insight into the future of Go and into how humans and machines can work together."

Tomorrow is the last game of Ke Jie's three-game series against AlphaGo, with Ke Jie again playing white. It still starts at 10:30, and DeepMind will stream it in Chinese and English on its YouTube channel.

Cover image/Weibo user Tianlelu Xiaozuzhang