Skip to main content

Key Matches Review

The history of AlphaGo is written through matches that shocked the world. From the secret matches in London in October 2015 to the farewell performance in Wuzhen in May 2017, each game rewrote humanity's understanding of Go and artificial intelligence.

This article provides a complete review of the background, process, and significance of these key matches.

Fan Hui Matches (October 2015): The Secret 5:0​

Background: Why Fan Hui?​

Before AlphaGo challenged the world's top players, DeepMind needed a "testing ground." They needed a professional player to verify AlphaGo's true strength, but this player had to meet several conditions:

  1. True professional level: Amateur players couldn't accurately test AI strength
  2. Willing to keep confidential: No leaks before the paper was published
  3. Convenient location: Easy to conduct multiple official games
  4. Open mindset: Willing to take an AI opponent seriously

Fan Hui perfectly met these conditions. Born in Xi'an, China, this professional player achieved 1-dan in 1996, rose to 2-dan in 2000, and later moved to France where he became the European Go Champion. He was the strongest professional player in Europe at the time and also held an open attitude toward artificial intelligence.

Match Arrangements​

In October 2015, Fan Hui was invited to DeepMind headquarters in London. After signing a confidentiality agreement, he played 5 official games against AlphaGo.

Match conditions:

  • Time control: 1 hour per side, with 30-second byo-yomi
  • Rules: Chinese rules, 7.5 komi
  • Environment: DeepMind office, with Aja Huang placing stones on behalf of AlphaGo

The Shocking 5:0​

The result shocked everyone: AlphaGo won 5:0.

GameDateResultNotes
Game 1October 5AlphaGo wins by resignationFan Hui plays Black
Game 2October 6AlphaGo wins by resignationFan Hui plays White
Game 3October 7AlphaGo wins by resignationFan Hui plays Black
Game 4October 8AlphaGo wins by 1.5 pointsFan Hui plays White
Game 5October 9AlphaGo wins by resignationFan Hui plays Black

Animation E1: These 5 games demonstrated how Policy Network guides search direction

Fan Hui later recalled:

"I lost the first game, and I thought I must have been careless. I lost the second game and started taking it seriously. After losing the third, fourth, and fifth games, I knew it wasn't my problem—Go had changed."

Why Keep It Secret?​

DeepMind chose to keep it secret for several reasons:

  1. Academic publication: The paper needed peer review before publication
  2. Verification time: Needed time to confirm reproducibility of results
  3. Business strategy: Choosing the optimal time to announce
  4. Protecting Fan Hui: Avoiding pressure on him before the news went public

This secret was kept for exactly three months until the Nature paper was published in January 2016.

Fan Hui's Transformation​

After losing these 5 games, Fan Hui didn't feel dejected. Instead, he became a member of the AlphaGo team, responsible for testing and improving the system.

"I wasn't defeated by AI—I became part of AI's development. This is an honor, not a disgrace."

This open attitude later became a model for the Go community facing AI.

Lee Sedol Matches (March 2016): Five Games That Changed the World​

Preparing for the Match of the Century​

On January 27, 2016, after the Nature paper was published, DeepMind announced they would challenge the world's top player. Target: Lee Sedol.

Why Lee Sedol?

  • 18 world championship titles: One of the most successful players of the past decade
  • The "Divine Calculator" nickname: Known for precise calculation
  • Fighting style: Enjoys complex, intense games
  • 35 years old at his peak: Optimal balance of experience and stamina

Animation E3: Lee Sedol's style was perfect for testing the limits of MCTS

Match Setup​

  • Location: Four Seasons Hotel, Seoul, Korea
  • Dates: March 9-15, 2016
  • Prize: $1 million USD (winner takes all, or split/donated to charity)
  • Rules: Chinese rules, 7.5 komi
  • Time control: 2 hours per side, 1-minute byo-yomi with 3 periods

Broadcast in over 200 countries and regions worldwide, with an estimated audience exceeding 200 million.

Game 1: A Shocking Beginning​

March 9, 2016

Lee Sedol played Black. The opening was conventional. But in the middle game, AlphaGo demonstrated astonishing whole-board vision.

On move 102, AlphaGo played what seemed like a retreat, giving up territory on the right side. Professional commentators expressed confusion. But 20 moves later, the brilliance of this move became clear—AlphaGo used the sacrificed stones to build central thickness, ultimately gaining an advantage across the entire board.

Result: AlphaGo wins by resignation

After the game, Lee Sedol said:

"I was shocked. I didn't expect to lose, and I certainly didn't expect to lose so thoroughly."

Animation E5: This game demonstrated Value Network's ability to evaluate the whole board

Game 2: Birth of the "Divine Move"​

March 10, 2016

This game gave birth to move 37, known as the "Divine Move." (See next article: In-Depth Analysis of "Move 37")

AlphaGo played a "fifth-line shoulder hit" in the upper right—a position humans would almost never consider. The commentator immediately called it a "mistake," but 50 moves later, this move proved to be the key to victory.

Result: AlphaGo wins by resignation

Korean commentator Kim Seong-ryong 9-dan said after the game:

"I've been playing Go for 50 years and have never seen anything like this. AlphaGo made me rethink what Go is."

Animation E7: Move 37 demonstrated how AI discovers strategies unknown to humans

Game 3: Desperate 3:0​

March 12, 2016

In this game, Lee Sedol attempted an unconventional opening, hoping to lead AlphaGo into unknown territory. He adopted a variation of the "Kobayashi" opening, attempting to win through complex fighting.

But AlphaGo's response remained calm. It demonstrated remarkable adaptability—no matter what humans threw at it, it could find the optimal response.

Result: AlphaGo wins by resignation

The score reached 3:0, removing any suspense from the match. But everyone was wondering: Could humans win even one game?

Game 4: Humanity's Counterattack​

March 13, 2016

This game will be forever remembered in history—not for AI's magic, but for humanity's counterattack.

At move 78, Lee Sedol, in byo-yomi, played a stunning move: a brilliant fifth-line tesuji.

This was a "wedge" tesuji that appeared ordinary but threw AlphaGo into confusion. In the following moves, AlphaGo's win-rate evaluation fluctuated wildly, and it played several obvious mistakes.

Animation E9: This game exposed MCTS weaknesses in certain positions

The DeepMind team later analyzed that AlphaGo's win-rate evaluation had errors in that position. It underestimated the power of Lee Sedol's move, leading to poor subsequent responses.

Result: Lee Sedol wins by resignation

This was AlphaGo's only defeat in official competition. Lee Sedol said emotionally:

"This victory is priceless. It proves that human players can still defeat AI—at least in certain positions."

Google DeepMind CEO Demis Hassabis tweeted:

"Lee Sedol is a true legend. He found AlphaGo's weakness and exploited it precisely."

Game 5: The Final Outcome​

March 15, 2016

After securing a precious victory, Lee Sedol entered Game 5 with a lighter mindset. He adopted a more aggressive strategy, attempting to find AlphaGo's weaknesses again.

But the DeepMind team had made emergency adjustments after Game 4. This version of AlphaGo seemed more robust, no longer making the evaluation errors seen before.

Result: AlphaGo wins by resignation

Final score: AlphaGo 4:1 Lee Sedol

Historical Significance of the Match​

The impact of this match extended far beyond the Go community:

For Artificial Intelligence​

  • Proved the power of deep learning: AI can surpass humans in complex decision-making tasks
  • Milestone for reinforcement learning: Self-play training was proven effective
  • Sparked subsequent research: Triggered a wave of AI investment

For the Go Community​

  • Traditional theories challenged: Many "joseki" were proven suboptimal
  • Training methods changed: Professional players began using AI-assisted training
  • New moves emerged: AI introduced many innovative techniques

For the Public​

  • AI awareness awakening: Ordinary people began paying attention to artificial intelligence
  • Increased tech coverage: Mainstream media extensively covered AI progress
  • Movies and documentaries: Led to the AlphaGo documentary

Animation E11: This match marked a "phase transition" moment in AI capabilities

Master's 60-Game Win Streak (January 2017): Online Speed Go Shock​

The Mysterious "Master" Account​

On December 29, 2016, an account named "Master" appeared on Chinese Go servers Yike and Tencent's Fox Weiqi.

This account's performance was unbelievable:

  • Won against all opponents: Not a single loss
  • All opponents were top players: Including world champions and 9-dan professionals
  • Extremely short thinking time: Almost instant moves

Soon, the entire Go community was discussing: Who exactly is "Master"?

The 60-Win Streak​

From December 29 to January 4, 2017, "Master" played 60 speed games, winning all of them.

The list of defeated players reads like a Go Hall of Fame:

RankPlayerRecord
World #1Ke Jie (China)0-3
World #2Park Junghwan (Korea)0-2
World #3Iyama Yuta (Japan)0-1
LegendNie Weiping (China)0-1
LegendGu Li (China)0-2
.........

In total, over 50 professional 9-dan players were included, covering top players from China, Japan, and Korea.

Animation E13: Speed games demonstrated Policy Network's real-time decision-making ability

Identity Revealed​

On January 4, 2017, after completing the 60th victory, "Master" revealed its identity in the chat room:

"I am Dr. Huang of AlphaGo."

Dr. Huang is Aja Huang, a core member of the AlphaGo team.

DeepMind subsequently officially confirmed: "Master" was a new version of AlphaGo, and this test's purpose was to verify the system's stability in an online environment.

Professional Players' Reactions​

The impact of the 60-game win streak was even more profound than the Lee Sedol match. This time there were more opponents across a wider range.

Ke Jie (lost three times to Master):

"The gap between humans and AI is larger than we imagined. We always thought we understood Go, but Master made me feel like we haven't even begun to understand it."

Nie Weiping (Chinese Go Saint):

"I've been playing Go for 60 years, and this is the first time I've felt so powerless. This isn't a gap in technique—it's a gap in dimensions."

Gu Li (eight-time world champion):

"After losing to Master, I began thinking about what value human players have. Do we still need professional competitions?"

Technical Analysis​

This version of AlphaGo (later called AlphaGo Master) showed significant improvements over the Lee Sedol match version:

MetricLee VersionMaster VersionImprovement
Elo rating~3,600~4,800+1,200
Self-play win rate-99%+-
Policy accuracy~57%~62%+5%
Training timeSeveral monthsAdditional months-

Animation E15: Elo improvement demonstrated exponential progress through self-play

Ke Jie Match (May 2017): The King's Curtain Call​

The Last Challenger​

After Master's 60-game win streak, few believed humans had any chance against AlphaGo. But one person still yearned for a match—Ke Jie.

At 19 years old, Ke Jie was the world's number one ranked player. He had repeatedly stated publicly:

"I don't think AlphaGo can defeat me. Even if Master beat me three times in speed games, official matches are different."

Google accepted the challenge.

Wuzhen Go Summit​

In May 2017, the "Future of Go Summit" was held in Wuzhen, Zhejiang Province, China. This was a grand event centered around AlphaGo, including:

  1. Ke Jie three-game series: The strongest human vs the strongest AI
  2. Pair Go: Human + AlphaGo vs Human + AlphaGo
  3. Team Go: Five top Chinese players together against AlphaGo

Three-Game Series: The 3:0 Outcome​

Game 1 (May 23)

Ke Jie played Black and adopted a relatively steady "Chinese Opening." This was a deliberate choice—Ke Jie hoped to avoid being defeated by AlphaGo's whole-board vision and instead fight for opportunities in the details.

But AlphaGo's responses were flawless. It found the most accurate move at every critical moment, gradually accumulating advantages.

Result: AlphaGo wins by 1/4 point (0.5 point)

This is the smallest possible margin of victory in Go. Ke Jie shed tears after the game:

"I gave it everything, but I still fell short by just a little bit."

Animation E17: The 1/4 point margin demonstrated AI's precise control

Game 2 (May 25)

Ke Jie changed strategy, adopting an opening that imitated AlphaGo. He used the "3-3 point invasion" new technique—exactly the innovation AlphaGo had brought to the Go world.

"If your way of playing is better, I'll learn your way of playing."

But AlphaGo remained unfazed. It continued at its own pace, demonstrating amazing calculation ability in the middle-game fighting.

Result: AlphaGo wins by resignation

Game 3 (May 27)

In the final game, Ke Jie went all-in. He adopted an extremely aggressive fighting style, attempting to drag AlphaGo into chaos.

In the opening stage, Ke Jie did create some complex situations. But AlphaGo's responses remained precise, giving Ke Jie no chance to turn the game around.

Result: AlphaGo wins by resignation

Final score: AlphaGo 3:0 Ke Jie

Animation E19: The three-game series demonstrated AlphaGo's absolute dominance

Pair Go and Team Go​

Besides the Ke Jie series, the summit also featured two innovative formats:

Pair Go (May 26)

Lian Xiao + AlphaGo vs Gu Li + AlphaGo

The interesting aspect of this match was: What happens when human players disagree with AlphaGo?

Results showed: The side that fully followed AlphaGo's suggestions performed better. When human players tried to "correct" AlphaGo's moves, it often worsened the position.

Result: Lian Xiao + AlphaGo wins

Team Go (May 26)

Team China (Zhou Ruiyang, Shi Yue, Tang Weixing, Chen Yaoye, Mi Yuting) vs AlphaGo

Five top Chinese players cooperated against one AI. They could discuss fully and jointly decide each move.

But the result was predictable: AlphaGo wins by resignation.

This match proved: Even with top human players working together, they couldn't defeat AlphaGo.

AlphaGo's Retirement Announcement​

On May 27, 2017, after the Ke Jie series ended, DeepMind issued an important statement:

"This is AlphaGo's final public match. We believe AlphaGo has completed its mission—proving that AI can reach superhuman levels in Go, a pinnacle of human intelligence.

Going forward, we will apply the techniques learned from AlphaGo to more important problems: healthcare, energy, materials science. This is the true value of artificial intelligence."

They also announced:

  1. AlphaGo teaching tool: Will release AlphaGo's game analysis for players to study
  2. 50 self-play game records: Publishing AlphaGo vs AlphaGo game records
  3. Technical paper: Will publish AlphaGo Zero research in Nature

Animation E21: AlphaGo's retirement marked the end of an era

Historical Position of the Matches​

Technical Milestones​

AlphaGo's matches have milestone significance in AI history:

YearEventSignificance
1997Deep Blue defeats KasparovVictory of brute-force search
2011Watson wins Jeopardy!Breakthrough in natural language processing
2016AlphaGo defeats Lee SedolVictory of deep learning + reinforcement learning
2017AlphaGo Zero 100:0Victory of pure self-learning

Animation E23: Each milestone represents evolution in AI methodology

Impact on the Go Community​

Changes in Game Record Study

Traditionally, professional players mainly studied human game records. But after AlphaGo, AI game records became required study.

  • 3-3 point opening: AlphaGo proved direct corner invasion is an effective strategy
  • The power of shoulder hits: Move 37 changed understanding of this tesuji
  • Value of thickness: AI showed new ways to convert thickness

Revolution in Training Methods

Professional players' training methods fundamentally changed:

Traditional MethodAI Era Method
Study human game recordsStudy AI game records
Rely on teacher guidanceUse AI analysis tools
Memorize josekiUnderstand AI evaluation logic
Practice gamesAI post-game review

Rise of New Generation Players

Players who grew up after 2016 are called "AI natives." Their playing style is clearly influenced by AI:

  • More focused on efficiency than traditional aesthetics
  • More willing to try unconventional moves
  • More reliant on precise calculation than intuition

Philosophical Reflections​

AlphaGo's victory sparked profound philosophical discussions:

What is the nature of intelligence?

Does AlphaGo "understand" Go? Or is it merely performing precise calculations? This question remains unresolved.

Where is human value?

When AI surpasses humans in Go, do Go competitions still have meaning? Many players reconsidered the meaning of their profession.

Interestingly, after AlphaGo, global attention to Go actually increased. People realized: Go is not just competition, but also art and philosophy.

Direction of AI Development

AlphaGo's success made people both excited and worried about AI. DeepMind's choice to retire AlphaGo and turn to solving "truly important problems" was itself an ethical choice.

Animation E25: AlphaGo sparked widespread discussion about AI ethics

Bonus: Other Important Matches​

Matches Against Other AIs​

Beyond public matches, AlphaGo also played numerous games against other Go AIs:

OpponentVersionResult
Crazy StoneStrongest Go program in 2015All wins
ZenJapan's strongest Go AIAll wins
Older AlphaGoSelf-play between versions-

Internal Tests​

The DeepMind team conducted extensive internal tests:

  • AlphaGo Lee vs AlphaGo Master: Master version win rate exceeded 99%
  • AlphaGo Master vs AlphaGo Zero: Zero version win rate exceeded 89%
  • Matches between versions with different training times: Observing learning curves

This test data was later published in papers, becoming important material for studying AI learning.

Animation Mapping​

Core concepts covered in this article and their animation numbers:

NumberConceptPhysics/Math Correspondence
Animation E1Policy Network guiding searchProbability distribution
Animation E3Testing MCTS limitsTree search depth
Animation E5Value Network whole-board evaluationValue function
Animation E7Discovering unknown strategiesExploration vs exploitation
Animation E9MCTS weaknessesBoundary conditions
Animation E11"Phase transition" in capabilitiesCritical phenomena
Animation E13Real-time decision abilityInference speed
Animation E15Exponential progress from self-playIterative optimization
Animation E17Precise control abilityNumerical stability
Animation E19Absolute dominanceConvergence to optimum
Animation E21End of an eraTask completion
Animation E23Methodology evolutionParadigm shift
Animation E25AI ethics discussionSocial impact

Further Reading​

References​

  1. Silver, D., et al. (2016). "Mastering the game of Go with deep neural networks and tree search." Nature, 529, 484-489.
  2. Silver, D., et al. (2017). "Mastering the game of Go without human knowledge." Nature, 550, 354-359.
  3. AlphaGo Documentary (2017), directed by Greg Kohs.
  4. DeepMind Official Blog: AlphaGo Series Articles
  5. Lee Sedol Match Official Game Records and Commentary (Korea Baduk Association)
  6. Wuzhen Go Summit Official Records