AlphaGo

Artificial intelligence that plays Go

This article dissects a computer program. For the cinematic rendition, see AlphaGo (film).

| Developer | Google DeepMind AlphaGo is a computer program developed by DeepMind Technologies, a subsidiary of Google, specifically designed to play the ancient board game of Go. Its creation marked a significant milestone in the field of artificial intelligence (AI), demonstrating capabilities that many experts had previously considered decades away.

Initially, AlphaGo emerged onto the scene with a remarkable ability to challenge and eventually overcome human professional players. Subsequent iterations of the program saw its power and strategic understanding deepen considerably. One prominent evolution was the version that competed under the name Master, which further solidified AlphaGo's dominance. After its tenure in competitive play, AlphaGo Master was succeeded by an even more formidable version, AlphaGo Zero. This particular variant represented a paradigm shift, as it was entirely self-taught, learning the intricacies of Go without any prior exposure to human games or expert strategies. The principles underlying AlphaGo Zero were then generalized into a broader program known as AlphaZero, which extended its mastery beyond Go to encompass other complex strategic games, including chess and shogi. The lineage continued with MuZero, a program that pushed the boundaries further by learning optimal strategies without even being explicitly taught the rules of the games it played.

The core methodology employed by AlphaGo and its successors is rooted in a sophisticated combination of algorithms. These systems leverage a Monte Carlo tree search to meticulously explore potential moves and their outcomes. This search is intelligently guided by knowledge previously acquired through extensive machine learning. Specifically, AlphaGo utilizes an artificial neural network, a sophisticated form of deep learning, which is rigorously trained through vast datasets of both human and computer-generated games. This neural network is not merely a static repository of knowledge; it is continuously refined to identify the most advantageous moves and to accurately predict the winning probabilities associated with those moves. This iterative process of training and self-improvement means that as the neural network becomes more adept, it enhances the effectiveness of the tree search, leading to progressively stronger and more nuanced move selections in subsequent iterations.

In a landmark event in October 2015, the original AlphaGo achieved a historic feat by becoming the first computer Go program to defeat a human professional Go player, Fan Hui, without any handicap on a standard full-sized 19×19 board. This achievement, initially kept under wraps, was formally announced in January 2016, coinciding with the publication of a detailed paper in the prestigious journal Nature outlining the algorithms that made such a breakthrough possible.

The subsequent year, March 2016, witnessed an even more widely publicized and impactful challenge: AlphaGo faced off against Lee Sedol, a legendary 9-dan professional Go player from South Korea, in a five-game match. This was the first instance where a computer Go program managed to defeat a 9-dan professional without handicap. While Lee Sedol notably secured a victory in the fourth game, demonstrating a glimmer of human ingenuity against the machine, he ultimately resigned in the decisive fifth game, resulting in a final score of 4 games to 1 in favor of AlphaGo. In recognition of this unprecedented triumph, AlphaGo was bestowed with an honorary 9-dan by the esteemed Korea Baduk Association. The dramatic lead-up to and the actual challenge match with Lee Sedol were meticulously documented in a compelling documentary film also titled AlphaGo, directed by Greg Kohs. The profound significance of AlphaGo's victory was further underscored when it was selected by the journal Science as one of the "Breakthrough of the Year" runners-up on 22 December 2016.

The program's journey of dominance continued into 2017. At the Future of Go Summit, the enhanced Master version of AlphaGo confronted and defeated Ke Jie, who was then universally recognized as the world's number one ranked player, in a three-game match. Following this decisive victory, AlphaGo was once again honored, this time with a professional 9-dan rank by the Chinese Weiqi Association.

Remarkably, after its resounding success against Ke Jie, DeepMind made the decision to retire AlphaGo from competitive play. This wasn't a sign of weakness, but rather a strategic pivot, allowing the research team to dedicate its efforts to exploring other frontiers within AI research. The legacy of AlphaGo, however, continued to evolve. The entirely self-taught AlphaGo Zero not only achieved a stunning 100–0 victory against the earlier competitive version of AlphaGo, but its successor, AlphaZero, was widely considered the world's preeminent player in Go by the close of the 2010s, a testament to the continuous and rapid advancement in AI capabilities.

Part of a series on Artificial intelligence (AI)

Major goals

Approaches

Applications

Philosophy

History

Controversies

Glossary

Glossary
v
t
e

History

For nearly two decades, the intellectual landscape of artificial intelligence had been marked by a peculiar challenge: the ancient board game of Go. It was, to put it mildly, considered a significantly more formidable adversary for computers than games like chess. The sheer combinatorial complexity of Go, with its staggering branching factor and the nuanced, aesthetic nature of its strategy, made it incredibly difficult to construct a direct evaluation function or to employ traditional AI methods such as alpha–beta pruning, tree traversal, and simple heuristic searches. These conventional approaches, while effective in games with more constrained move sets, simply buckled under the vast possibilities of a Go board.

Even after IBM's formidable Deep Blue had famously vanquished world chess champion Garry Kasparov in the 1997 match, the strongest Go programs available could only aspire to an amateur 5-dan level. They consistently failed to defeat a professional Go player without a significant handicap. There were glimmers of progress, of course; in 2012, the software program Zen, utilizing a four-PC cluster, managed to beat Masaki Takemiya (a 9-dan professional) twice, albeit with five- and four-stone handicaps. A year later, in 2013, Crazy Stone similarly bested Yoshio Ishida (another 9-dan) with a four-stone handicap. These were notable achievements, but the elusive goal of defeating a top professional without handicap remained stubbornly out of reach, a testament to Go's profound depth.

According to David Silver of DeepMind, the AlphaGo research project was initiated around 2014. Its primary objective was to rigorously test the limits of what a neural network powered by deep learning could achieve in the game of Go. The results were, to say the least, compelling. AlphaGo represented a monumental leap over its predecessors. In a series of 500 games against other leading Go programs of the era, including the aforementioned Crazy Stone and Zen, AlphaGo, running on a single computer, achieved a near-perfect record, winning all but one match. When deployed in a distributed configuration across multiple machines, AlphaGo's dominance was absolute, winning all 500 games against other programs and even securing 77% of its games against its single-computer counterpart. The distributed version, as of October 2015, was a beast of computational power, harnessing 1,202 CPUs and 176 GPUs to fuel its strategic prowess.

Match against Fan Hui

In October 2015, the distributed incarnation of AlphaGo achieved its inaugural, publicly recognized triumph against a human professional. It decisively defeated Fan Hui, the then European Go champion and a 2-dan professional (out of a possible 9 dan), with a clean sweep of five games to zero. This wasn't just a win; it was a historic moment, marking the first time a computer Go program had ever beaten a professional human player on a full-sized board without the aid of a handicap. The announcement of this groundbreaking achievement was deliberately withheld until 27 January 2016, a strategic delay timed to coincide with the publication of a comprehensive paper in the journal Nature, which meticulously detailed the advanced algorithms and methodologies that underpinned AlphaGo's success.

Match against Lee Sedol

The world watched with bated breath as AlphaGo prepared to face its greatest challenge yet: South Korean professional Go player Lee Sedol. A prodigious 9-dan player, Lee Sedol was widely regarded as one of the best Go players in the world at the time. The highly anticipated five-game match unfolded at the luxurious Four Seasons Hotel in Seoul, South Korea, across five intense days: March 9, 10, 12, 13, and 15, 2016. Each game was meticulously video-streamed live, drawing a global audience.

In a stunning display of machine intelligence, AlphaGo claimed victory in four of the five games. Lee Sedol, however, etched his name into history by securing a win in the fourth game, becoming the only human player to defeat AlphaGo in any of its 74 official competitive games. The computational infrastructure supporting AlphaGo during this match was formidable, running on Google's cloud computing platform with servers located in the United States. The match adhered to Chinese rules, incorporating a 7.5-point komi, and allotted each side two hours of primary thinking time, followed by three 60-second byoyomi periods. Notably, the version of AlphaGo deployed against Lee Sedol utilized a similar computational footprint to the one that had triumphed over Fan Hui, reportedly employing 1,920 CPUs and 280 GPUs, as detailed by The Economist.

At the time of this monumental encounter, Lee Sedol held the second-highest number of Go international championship victories globally, trailing only his compatriot, Lee Chang-ho, who had maintained the world championship title for an astonishing 16 years. While the absence of a single, universally recognized international ranking in Go meant his precise global standing could fluctuate across different sources, Lee Sedol was consistently ranked among the top players, with some even placing him as high as fourth. Crucially, AlphaGo's training was not tailored to specifically counter Lee Sedol or any other individual human player; its development focused on general mastery of the game.

The initial three games were swiftly claimed by AlphaGo, with Lee Sedol resigning in each instance, a stark demonstration of the AI's unexpected strength. However, in a moment that captivated the world, Lee Sedol managed to secure a victory in the fourth game, forcing AlphaGo's resignation at move 180. This brief respite for human pride was short-lived, as AlphaGo swiftly regained its footing, winning the fifth and final game by resignation, thus clinching the series 4-1.

The stakes for this match were substantial: a US $1 million prize. As AlphaGo emerged victorious, the prize money was charitably donated to various organizations, including [UNICEF](/UNICEF). Lee Sedol, for his part, received a participation fee of$ 150,000 for competing in all five games, along with an additional $20,000 for his singular victory in Game 4.

In June 2016, at a university presentation in the Netherlands, Aja Huang, a key member of the DeepMind team, offered a fascinating post-mortem. He revealed that the team had successfully identified and patched the specific logical weakness that AlphaGo exhibited during the pivotal Game 4 against Lee Sedol. Huang explained that after Lee's now-famous move 78—dubbed the "divine move" by many professionals—AlphaGo's sophisticated policy network, designed to identify the most accurate move order, failed to precisely guide the program toward the correct continuation. Its value network, which assesses positional strength, had not adequately accounted for the likelihood or impact of Lee's 78th move. Consequently, when the "divine move" was made, AlphaGo struggled to make the necessary logical adjustments, leading to a diversion and confusion in its computing powers that ultimately contributed to its defeat in that game. With the patch, AlphaGo would have maintained Black's advantage as originally intended.

Sixty online games

The digital Go world experienced a subtle tremor on 29 December 2016. A new account, "Magister" (or "Magist" in its Chinese server incarnation), surfaced on the Tygem server, originating from South Korea, and began challenging professional players. This enigmatic player swiftly changed its name to "Master" on December 30th before migrating to the FoxGo server on 1 January 2017. The speculation reached a fever pitch, and on 4 January, DeepMind officially confirmed what many had already suspected: "Magister" and "Master" were both controlled by an updated, more potent version of AlphaGo, now designated AlphaGo Master.

By 5 January 2017, AlphaGo Master's online record stood at an unblemished 60 wins and 0 losses. This included three decisive victories over Ke Jie, who was then the top-ranked Go player in the world, and who had, in fact, been discreetly informed beforehand that Master was an AlphaGo variant. The scale of this dominance was such that Gu Li, another prominent professional, went so far as to offer a bounty of 100,000 yuan (approximately US$14,400) to any human player who could manage to defeat Master. Playing at an astonishing pace of 10 games per day, with little to no discernible rest between matches, Master's superhuman efficiency quickly solidified suspicions that it was an AI.

Its list of vanquished adversaries read like a roll call of Go's elite: world champions such as Ke Jie, Park Jeong-hwan, Yuta Iyama, Tuo Jiaxi, Mi Yuting, Shi Yue, Chen Yaoye, Li Qincheng, Gu Li, Chang Hao, Tang Weixing, Fan Tingyu, Zhou Ruiyang, Jiang Weijie, Chou Chun-hsun, Kim Ji-seok, Kang Dong-yun, Park Yeong-hun, and Won Seong-jin. The roster also included national champions and world championship runners-up like Lian Xiao, Tan Xiao, Meng Tailing, Dang Yifei, Huang Yunsong, Yang Dingxin, Gu Zihao, Shin Jinseo, Cho Han-seung, and An Sungjoon. Almost all of these 60 games were fast-paced affairs, utilizing three 20 or 30-second byo-yomi periods. In a rare concession, Master offered to extend the byo-yomi to one minute when playing against Nie Weiping, a gesture of respect for his age. After securing its 59th victory, Master finally revealed its true identity in the chatroom, confirming it was controlled by Dr. Aja Huang of the DeepMind team, subsequently changing its nationality designation to the United Kingdom. Upon the completion of these extraordinary games, DeepMind co-founder Demis Hassabis tweeted, "we're looking forward to playing some official, full-length games later [2017] in collaboration with Go organizations and experts," hinting at future public engagements.

The Go community was, understandably, profoundly impressed, and somewhat unnerved, by the program's performance and its distinctly nonhuman playing style. Ke Jie famously remarked, "After humanity spent thousands of years improving our tactics, computers tell us that humans are completely wrong... I would go as far as to say not a single human has touched the edge of the truth of Go." A rather definitive statement, if you ask me.

Future of Go Summit

The Future of Go Summit, held in Wuzhen in May 2017, served as the ultimate proving ground for AlphaGo Master. Here, it engaged in a highly anticipated three-game match against Ke Jie, the world's then-No.1 ranked player. Beyond this marquee event, AlphaGo Master also participated in two games against a selection of top Chinese professionals, a unique pair Go game, and a collaborative match against a team of five human players.

Google DeepMind sweetened the pot with a substantial 1.5 million dollar winner's prize for the three-game match between Ke Jie and Master, with the losing side still receiving a respectable 300,000 dollars. As it turned out, Master demonstrated its unwavering superiority by winning all three games against Ke Jie. This resounding triumph led to AlphaGo being formally awarded a professional 9-dan rank by the Chinese Weiqi Association, a recognition that solidified its place in the game's history.

Following its decisive victory against Ke Jie, the reigning world Go champion, AlphaGo was officially retired from competitive play. DeepMind further announced the disbandment of the specialized team that had worked on the game, signaling a strategic shift towards focusing on broader AI research initiatives. As a parting gift to the global Go community, DeepMind released 50 full-length AlphaGo vs. AlphaGo matches after the Summit, offering unprecedented insight into the machine's internal strategies.

AlphaGo Zero and AlphaZero

The relentless march of progress in AI continued beyond AlphaGo's retirement. On 19 October 2017, AlphaGo's team published another seminal article in the journal Nature, introducing AlphaGo Zero. This groundbreaking version distinguished itself by learning entirely without human data, relying solely on self-play. Yet, it proved to be significantly stronger than any of its human-champion-defeating predecessors. Through this process of playing countless games against itself, AlphaGo Zero astonishingly surpassed the strength of AlphaGo Lee (the version that defeated Lee Sedol) in a mere three days, achieving a perfect 100-0 record against it. It then reached the formidable level of AlphaGo Master in just 21 days, and within 40 days, it had eclipsed all previous AlphaGo versions.

Further expanding on this self-play paradigm, DeepMind released a paper on arXiv on 5 December 2017, detailing the generalization of AlphaGo Zero's approach into a single, unified algorithm named AlphaZero. This algorithm achieved superhuman levels of play across multiple games—chess, shogi, and Go—within a mere 24 hours of training. It accomplished this by defeating the respective world-champion programs: Stockfish in chess, Elmo in shogi, and even the 3-day version of AlphaGo Zero itself. The implications were profound, signaling a generalizable approach to mastering complex strategic domains.

Teaching tool

On 11 December 2017, DeepMind, in a gesture of goodwill to the Go community (or perhaps just to show off), launched an AlphaGo teaching tool on its website. This invaluable resource meticulously analyzes the winning rates of various Go openings, all calculated by the formidable AlphaGo Master. The tool compiles a staggering 6,000 Go openings derived from 230,000 human games, each subjected to an exhaustive 10,000,000 simulations by AlphaGo Master. Many of these openings are thoughtfully augmented with human move suggestions, offering a bridge between the machine's profound insights and human understanding.

Versions

Early iterations of AlphaGo underwent rigorous testing across a spectrum of hardware configurations, varying the number of CPUs and GPUs, and operating in both asynchronous and distributed modes. Each move was allocated a mere two seconds of thinking time during these trials. The resulting Elo ratings provide a clear progression of its strength. It's worth noting that in competitive matches where more thinking time per move was permitted, AlphaGo consistently achieved even higher ratings, showcasing the direct correlation between computational resources and strategic depth.

Configuration	Search threads	No. of CPU	No. of GPU	Elo rating
Single	40	48	1	2,181
Single	40	48	2	2,738
Single	40	48	4	2,850
Single	40	48	8	2,890
Distributed	12	428	64	2,937
Distributed	24	764	112	3,079
Distributed	40	1,202	176	3,140
Distributed	64	1,920	280	3,168

In May 2016, Google unveiled its proprietary hardware innovation: "tensor processing units (TPUs)." The company revealed that these specialized chips had already been strategically deployed in various internal projects, including the AlphaGo match against Lee Sedol, demonstrating their critical role in accelerating machine learning tasks.

During the Future of Go Summit in May 2017, DeepMind transparently disclosed that the version of AlphaGo deployed in this high-profile event was indeed AlphaGo Master. They also provided a fascinating insight into the relative strengths of the different software versions. AlphaGo Lee, the iteration that famously challenged Lee Sedol, was capable of giving AlphaGo Fan (the version that defeated Fan Hui) a three-stone handicap and still winning. Even more impressively, AlphaGo Master was estimated to be another three stones stronger than AlphaGo Lee, highlighting the rapid and profound advancements in its capabilities.

Versions	Hardware	Elo rating	Date	Results
AlphaGo Fan	176 GPUs, distributed	3,144	Oct 2015	5:0 against Fan Hui
AlphaGo Lee	48 TPUs, distributed	3,739	Mar 2016	4:1 against Lee Sedol
AlphaGo Master	4 TPUs, single machine	4,858	May 2017	60:0 against professional players; Future of Go Summit
AlphaGo Zero (40 block)	4 TPUs, single machine	5,185	Oct 2017	100:0 against AlphaGo Lee; 89:11 against AlphaGo Master
AlphaZero (20 block)	4 TPUs, single machine	5,018	Dec 2017	60:40 against AlphaGo Zero (20 block)

Algorithm

As of 2016, AlphaGo's formidable capabilities were underpinned by a sophisticated algorithm that seamlessly integrated a blend of machine learning and tree search techniques. This complex system was honed through extensive training, drawing knowledge from both human gameplay and self-generated computer play. At its heart lay the Monte Carlo tree search, a probabilistic search algorithm, which was intelligently guided by two specialized deep neural networks: a "value network" and a "policy network." These networks, crucial to its performance, were implemented using cutting-edge deep learning technology.

Before raw game data was fed into these neural networks, a limited amount of game-specific feature detection pre-processing was applied. This involved, for instance, highlighting whether a particular move corresponded to a known nakade pattern, providing the networks with more structured input. The networks themselves were architecturally convolutional neural networks, typically comprising 12 layers, and were rigorously trained through a process known as reinforcement learning.

The initial "bootstrapping" of AlphaGo's neural networks was achieved by leveraging human gameplay expertise. It was first trained to emulate human play by attempting to predict and match the moves of expert players sourced from a vast database of approximately 30 million recorded historical games. Once this initial phase had endowed it with a baseline level of proficiency, AlphaGo embarked on a more advanced training regimen. It was pitted against other instances of itself in an immense number of self-play games, continuously refining its strategies through reinforcement learning to enhance its overall playing strength. In a rather polite, yet calculated, measure to avoid "disrespectfully" prolonging a hopeless game for its human opponent, the program was specifically coded to resign if its internal assessment of its win probability dipped below a predetermined threshold. For the high-stakes match against Lee Sedol, this resignation threshold was set at a pragmatic 20%.

Style of play

Toby Manning, who served as the match referee for AlphaGo's historic encounter with Fan Hui, characterized the program's playing style as notably "conservative." This seemingly paradoxical description for a revolutionary AI stems from AlphaGo's fundamental strategic directive: it strongly favors maximizing its probability of winning, even if that means winning by a smaller margin, over pursuing a lower probability of winning by a larger margin. This distinct approach to maximizing winning probability stands in stark contrast to the tendencies of many human players, who often aim to maximize territorial gains, sometimes at greater risk. This fundamental difference helps to elucidate some of AlphaGo's more "odd-looking" or unconventional moves that initially perplexed human observers. It developed a penchant for opening moves that were either rarely or never seen in human professional play. Furthermore, AlphaGo displayed a particular fondness for employing shoulder hits, especially when an opponent seemed to be overly concentrated in a particular area, demonstrating a dynamic and adaptable strategic repertoire.

Responses to 2016 victory

AI community

AlphaGo's resounding victory in March 2016 wasn't just another notch on the belt; it was a watershed moment, a genuine major milestone in artificial intelligence research. Go had long been held as the "hard problem" of machine learning, a challenge widely believed to be far beyond the technological grasp of the time. Most experts had conservatively estimated that a Go program of AlphaGo's caliber was at least five years away, with some even positing it would take another decade before computers could realistically challenge, let alone defeat, Go champions. This widespread skepticism meant that at the outset of the 2016 matches, the overwhelming consensus among observers was that Lee Sedol would undoubtedly prevail over AlphaGo.

With games like checkers, famously "solved" by the Chinook computer engine, and chess, conquered by Deep Blue, now consigned to the annals of AI history, AlphaGo's triumph over Go signaled a profound shift. The victory effectively marked "the end of an era," as Deep Blue's Murray Campbell succinctly put it, declaring that "board games are more or less done and it's time to move on." The traditional benchmarks for AI progress, once defined by mastery over popular board games, were now obsolete.

A crucial distinction often drawn between AlphaGo and its predecessors like Deep Blue or Watson lies in the general-purpose nature of its underlying algorithms. Many commentators viewed AlphaGo's success as compelling evidence that the scientific community was making tangible progress toward achieving artificial general intelligence (AGI). This potential shift sparked broader discussions about society's preparedness for the future impact of machines with general purpose intelligence. Entrepreneur Guy Suter, however, offered a pragmatic counterpoint, reminding everyone that AlphaGo, for all its brilliance, was narrowly focused; "[It] couldn't just wake up one morning and decide it wants to learn how to use firearms." AI researcher Stuart Russell voiced concerns that AI systems like AlphaGo had advanced far more rapidly and become significantly more powerful than anticipated, underscoring the urgent need to develop robust methods to ensure they "remain under human control."

The philosophical debate surrounding AI's future intensified. Scholars such as Stephen Hawking had, even prior to the matches in May 2015, warned of the potential for future self-improving AI to achieve genuine general intelligence, potentially leading to an unforeseen AI takeover. However, not all scholars shared this dire outlook. AI expert Jean-Gabriel Ganascia, for example, expressed skepticism, believing that "Things like 'common sense'... may never be reproducible," and instead highlighted the immense potential for AI to raise hopes in diverse fields such as healthcare and space exploration. Computer scientist Richard Sutton offered a more balanced perspective, stating, "I don't think people should be scared... but I do think people should be paying attention."

In China, AlphaGo's victory resonated with particular intensity, serving as a veritable "Sputnik moment." It undeniably played a pivotal role in convincing the Chinese government to prioritize and dramatically escalate funding for artificial intelligence research and development, setting the stage for a national strategic focus on AI.

In 2017, the DeepMind AlphaGo team was deservedly honored with the inaugural IJCAI Marvin Minsky medal for Outstanding Achievements in AI. Professor Michael Wooldridge, Chair of the IJCAI Awards Committee, lauded AlphaGo as "a wonderful achievement, and a perfect example of what the Minsky Medal was initiated to recognise." He further emphasized that what particularly impressed the IJCAI was AlphaGo's ability to achieve its breakthroughs through a brilliant synthesis of classic AI techniques alongside the state-of-the-art machine learning methods for which DeepMind had become synonymous. It was, as he concluded, "a breathtaking demonstration of contemporary AI."

Go community

Go holds a cherished place in the cultural fabric of China, Japan, and Korea. The 2016 matches between AlphaGo and Lee Sedol were not merely a technical demonstration; they were a global spectacle, captivating an estimated hundred million viewers worldwide. Many of the world's top Go players, initially bewildered by AlphaGo's unconventional plays, came to characterize them as moves that seemed questionable at first glance but revealed their profound strategic genius in hindsight. As one observer noted, "All but the very best Go players craft their style by imitating top players. AlphaGo seems to have totally original moves it creates itself." This bespoke, almost alien, style of play indicated that AlphaGo had unexpectedly become significantly stronger, even compared to its October 2015 match against Fan Hui, which had already marked the first time a computer had ever beaten a Go professional without handicap.

The impact on the human Go community was palpable. The day after Lee Sedol's initial defeat, Jeong Ahram, a leading Go correspondent for a major South Korean daily, somberly remarked, "Last night was very gloomy... Many people drank alcohol." In a profound gesture of recognition, the Korea Baduk Association, the governing body for Go professionals in South Korea, bestowed upon AlphaGo an honorary 9-dan title, acknowledging its creative skills and its undeniable contribution to advancing the game's understanding.

China's Ke Jie, then an 18-year-old widely considered the world's preeminent Go player, initially exuded confidence, claiming he would be able to defeat AlphaGo. However, he then notably declined to play against it, expressing a peculiar fear that the AI might "copy my style." As the matches against Lee Sedol progressed, Ke Jie's stance wavered. After analyzing the first three games, he conceded that "it is highly likely that I (could) lose," only to regain a measure of confidence after AlphaGo exhibited its rare flaws in the fourth match.

Both Toby Manning, the referee for AlphaGo's match against Fan Hui, and Hajin Lee, the secretary general of the International Go Federation, offered forward-looking perspectives. They reasoned that in the future, Go players would increasingly leverage computers to analyze their games, identify errors, and ultimately refine their skills, transforming AI from an opponent into a powerful teaching tool.

After the second game, a visibly shaken Lee Sedol admitted to feeling "speechless": "From the very beginning of the match, I could never manage an upper hand for one single move. It was AlphaGo's total victory." Following his third defeat, Lee expressed regret for his misjudgment of AlphaGo's capabilities, stating, "I misjudged the capabilities of AlphaGo and felt powerless." He was quick to emphasize that the defeat was "Lee Se-dol's defeat" and emphatically "not a defeat of mankind." Lee Sedol acknowledged that his eventual loss to a machine was "inevitable" but maintained that "robots will never understand the beauty of the game the same way that we humans do." His hard-won victory in game four, he declared, was a "priceless win that I (would) not exchange for anything."

AlphaGo documentary film (2016)

The impact of AlphaGo extended beyond the realm of algorithms and professional play, culminating in a compelling documentary film that chronicled its journey.

Reception

The documentary's critical reception was overwhelmingly positive. On Rotten Tomatoes, it achieved a perfect average rating of 100% based on 10 reviews, a rare feat for any film, let alone one about artificial intelligence and a niche board game.

Michael Rechtshaffen of the Los Angeles Times lauded the documentary with a positive review, highlighting its engaging human elements: "It helps matters when you have a group of engaging human subjects like soft-spoken Sedol, who's as intensively contemplative as the game itself, contrasted by the spirited, personable Fan Hui, the Paris-based European champ who accepts an offer to serve as an advisor for the DeepMind team after suffering a demoralizing AI trouncing." Rechtshaffen also commended the film's unexpected sequences, which, fueled by the passion of producer Hauschka's Volker Bertelmann, skillfully wove together strategic and philosophical components.

John Defore, writing for The Hollywood Reporter, described the documentary as "an involving sports-rivalry doc with an AI twist." He mused on the film's broader implications: "In the end, observers wonder if AlphaGo's odd variety of intuition might not kill Go as an intellectual pursuit but shift its course, forcing the game's scholars to consider it from new angles. So maybe it isn't time to welcome our computer overlords, and won't be for a while - maybe they'll teach us to be better thinkers before turning us into their slaves." A rather comforting thought, if you manage to ignore the underlying dread.

Greg Kohs, the film's director, shared his approach to tackling such a complex subject matter. He admitted that "The complexity of the game of Go, combined with the technical depth of an emerging technology like artificial intelligence seemed like it might create an insurmountable barrier for a film like this." However, he found his initial "innocent unawareness of Go and AlphaGo actually proved to be beneficial. It allowed me to approach the action and interviews with pure curiosity, the kind that helps make any subject matter emotionally accessible." Kohs articulated his deeper aspiration for the film: "Unlike the film's human characters – who turn their curious quest for knowledge into an epic spectacle with great existential implications, who dare to risk their reputation and pride to contest that curiosity – AI might not yet possess the ability to empathize. But it can teach us profound things about our humanness – the way we play board games, the way we think and feel and grow. It's a deep, vast premise, but my hope is, by sharing it, we can discover something within ourselves we never saw before."

Professional Go player perspective

Hajin Lee, a former professional Go player, offered a glowing assessment, describing the documentary as being "beautifully filmed." She noted that beyond the raw narrative of the matches, the film effectively conveyed the emotional resonance and atmosphere through its artful scene arrangements. She cited examples such as the intense close-up shots of Lee Sedol as he grappled with the realization of AlphaGo AI's intelligence, the palpable distress and affliction captured in the atmospheric scene of the Korean commentator following the initial defeat, and the pervasive tension that permeated the room during the games. Lee also praised the documentary for providing crucial context by detailing the background of AlphaGo's technology and shedding light on the customs of the Korean Go community. While complimentary, she suggested areas for further enhancement, such as delving into the specifics of AI predecessors to AlphaGo, offering more insight into the lives and pride of professional Go players to better contextualize Lee Sedol's pre-match confidence, and exploring the evolving perception of Go AI among players both during and after the monumental match.

Fan Hui, the professional Go player who famously became AlphaGo's first human professional opponent, provided insights into the AI's training methodology. He explained that "DeepMind had trained AlphaGo by showing it many strong amateur games of Go to develop its understanding of how a human plays before challenging it to play versions of itself thousands of times, a novel form of reinforcement learning which had given it the ability to rival an expert human. History had been made, and centuries of received learning overturned in the process. The program was free to learn the game for itself." A rather understated way to describe a revolution, wouldn't you agree?

Technology and AI-related fields

James Vincent, a reporter from The Verge, offered a more critical, yet insightful, perspective on the documentary's narrative approach. He observed that "It prods and pokes viewers with unsubtle emotional cues, like a reality TV show would. 'Now, you should be nervous; now you should feel relieved'." Vincent astutely pointed out how the AlphaGo footage meticulously captured the moment Lee Sedol truly acknowledged the profound power of the AlphaGo AI. In the initial game, Lee, drawing on his extensive experience, likely underestimated the AI, assuming an easy victory. However, the early game dynamics quickly deviated from his expectations. After suffering the first loss, a discernible nervousness and erosion of confidence began to set in. His reactions to AlphaGo's relentless attacks, particularly his declaration of merely wanting to win the match, inadvertently betrayed his frustration and led to uncharacteristic behavior. The stark contrast was evident when Lee deliberated for 12 minutes on a single move, while AlphaGo responded in a mere minute and a half, treating each alternative with unwavering consistency and devoid of emotional reaction, continuing the game as if its opponent's struggle was nonexistent.

Vincent further concluded that "suffice to say that humanity does land at least one blow on the machines, through Lee's so-called 'divine move'." He then broadened his contemplation to the future implications: "More likely, the forces of automation we'll face will be impersonal and incomprehensible. They'll come in the form of star ratings we can't object to, and algorithms we can't fully understand. Dealing with the problems of AI will take a perspective that looks beyond individual battles. AlphaGo is worth seeing because it raises these questions."

Murray Shanahan, a distinguished professor of cognitive robotics at Imperial College London and a senior research scientist at DeepMind, offered a broader view of Go's significance within the AI landscape. "Go is an extraordinary game but it represents what we can do with AI in all kinds of other spheres," Shanahan stated. "In just the same way there are all kinds of realms of possibility within Go that have not been discovered, we could never have imagined the potential for discovering drugs and other materials." A rather hopeful outlook, considering the universe's general disinterest in our affairs.

Similar systems

The pursuit of AI mastery in Go wasn't confined solely to DeepMind. Facebook had also been diligently developing its own Go-playing system, known as Darkforest. This system, much like AlphaGo, was predicated on the powerful combination of machine learning and Monte Carlo tree search. While Darkforest demonstrated considerable strength against other computer Go programs, as of early 2016, it had not yet achieved the milestone of defeating a professional human player. It was generally estimated to be of similar strength to other established programs like CrazyStone and Zen, having notably lost to both.

Another significant contender emerged in DeepZenGo, a system developed with robust support from the video-sharing website Dwango and the venerable University of Tokyo. In November 2016, DeepZenGo faced off against the legendary Go master Cho Chikun, who holds the record for the most Go title wins in Japan. In a closely watched match, DeepZenGo lost 2–1, showcasing considerable strength but ultimately falling short of a complete victory against a top human professional.

The far-reaching influence of AlphaGo's methodological innovations extended beyond the realm of games. A 2018 paper published in the journal Nature explicitly cited AlphaGo's approach as the foundational basis for a novel method of computing potential pharmaceutical drug molecules. This groundbreaking application highlighted how the principles of Monte Carlo tree search guided by neural networks, originally honed in the abstract world of Go, could be generalized and applied to a wide array of complex, real-world scientific and industrial challenges. Indeed, systems leveraging this potent combination have since been explored for an increasingly diverse range of applications, proving the unexpected utility of mastering a seemingly esoteric board game.

Example game

Here, we observe a snapshot of the tactical prowess of AlphaGo Master (playing as white) against Tang Weixing on 31 December 2016. AlphaGo secured a victory by resignation, but not before delivering a move that garnered widespread acclaim: White 36. A subtle, yet decisive, stroke of genius, apparently.

File:AlphaGo Master - Tang Weixing (game 1, 2016-12-31) - moves 1-99.png|First 99 moves File:AlphaGo Master - Tang Weixing (game 1, 2016-12-31) - moves 100-186.png|Moves 100–186 (149 at 131, 150 at 130)

Impacts on Go

The highly publicized documentary film about AlphaGo, which garnered a stellar 100% rating on Rotten Tomatoes, naturally fostered a widespread hope that professional players like Lee Sedol and Fan Hui would significantly benefit from their unprecedented experience of playing against AlphaGo. The expectation was that these encounters would catalyze a new era of human insight and strategic development within the game.

However, as of May 2018, the reality painted a somewhat different picture. Their respective ratings had shown little substantial change. Lee Sedol was ranked 11th in the world, a respectable position, but not the return to peak dominance some might have anticipated. Fan Hui's ranking stood at 545th, indicating a less dramatic shift. This perhaps highlights the profound, and perhaps unsettling, truth that while AlphaGo offered new perspectives, integrating such alien "divine moves" into human intuition and competitive strategy proved to be a far more complex undertaking than simply observing them.

The ultimate testament to AlphaGo's enduring impact, and perhaps a somber reflection on the future of human professional Go, came on 19 November 2019. On this date, Lee Sedol, the only human to ever defeat AlphaGo in an official game, announced his retirement from professional play. His reasoning was stark and unequivocal: he argued that he could never again aspire to be the top overall player of Go due to the increasing, and seemingly insurmountable, dominance of AI. Lee poignantly referred to these artificial intelligences as "an entity that cannot be defeated," a statement that resonates with the cosmic weariness of knowing one's peak has been surpassed by something fundamentally different.