In November 2019, legendary South Korean Go player Lee Sedol announced his retirement, specifically citing his success in scoring a single victory in a five-match showdown against AI.
The battle between Sedol and the technology occurred three years prior in a match which saw the 18-time world champion face down against AlphaGo – an AI system developed by Google DeepMind.
While Sedol lost the battle, the occasion marked a significant moment in the development of AI and gave the world a glimpse into the future potential of the technology.
In a recent blog post celebrating the 10th anniversary of the match, penned by Demis Hassabis, Sedol suggested that AlphaGo offered a “definitive preview of the AI era”
“It served as a ‘roadmap to the future’, sending a clear signal to humanity about how the world was about to change,” he said.
So what made AlphaGo such a monumental moment in the history of AI? Simply put, the sheer complexity of Go meant researchers were forced to pioneer new techniques in areas such as deep learning and reinforcement learning.
What is Go?
Go is an ancient, complex board game which originated in China. Players compete to control territory using stones placed on intersections which cannot be moved once placed. These stones can, however, be captured by the opposition player by surrounding them.
The rules and concept are rather simple, but it’s incredibly complex due to the huge volume of possible outcomes.
Go’s “branching factor” – which refers to the number of possible moves, decisions, or paths the player can take – stands at around 250 moves per turn. Compare that to chess, which has around 35, and the complexity becomes clear.
Go is a highly complex game, considered by AI researchers as the “perfect challenge”
(Image credit: Getty Images)
In a recent podcast reflecting on the anniversary, Thore Graepel, distinguished research scientist and member of the AlphaGo team at the time, said this reputation for complexity made it a prime target for computer scientists.
“The game of Go seemed like the perfect challenge for AI because the game has such simple rules yet it leads to such complex gameplay with tactics and strategies and complex patterns,” he said.
“Once the game of chess had been solved, as it were, or at least Deep Blue had won against the world champion. Go was this open challenge. It’s much more complex than chess by many orders of magnitude and nobody was expecting it to be solved any time soon,” Graepel added.
“It was the perfect game to tackle at the time.”
Previous programs had been developed with the aim of mastering the game, yet according to DeepMind they had only achieved “the level of human amateurs despite decades of work”.
A key factor behind this was that AI training and fine-tuning methods at the time meant systems were unable to compensate for the huge number of possible moves, and crucially, lacked the creativity of human players.
AlphaGo sets a high bar
Development of AlphaGo sought to address this directly. Researchers combined neural networks with advanced search algorithms to help the system identify and pursue a varied range of potential moves, the company explained in a 2016 blog post.
“One neural network – known as the “policy network” – selects the next move to play,” DeepMind explained. “The other neural network – the “value network” – predicts the winner of the game.”
Reinforcement learning, a method of training AI systems to help them find their own way of solving problems, was used to fine-tune AlphaGo. This process involved putting the AI through thousands of games so it could learn how human players typically acted.
Thereafter, AlphaGo played against different versions of itself thousands of times, according to DeepMind. Each time it played itself, it learned from mistakes; becoming more adept and intuitive.
Jeff Watkins, chief AI officer at Leeds-based AI consultancy, NorthStar Intelligence, told ITPro that AlphaGo’s development highlighted significant advances in areas such as reinforcement learning.
“Due to the size of the board and available moves, Go had seemed like a difficult challenge for machines to excel in, requiring more creativity in moves over a staggering number of possible combinations,” he said.
“From a technical perspective, it showed that deep learning had advanced to the point where it could now handle these highly-complex domains.”
“Surprise, awe, and some discomfort”
AlphaGo’s first match against a human player came in October 2015, where it defeated three-time European champion Fan Hui, scoring 5-0.
When the system defeated Sedol, the shockwaves it sent through both the global player base and tech industry cannot be understated. Sedol was the winner of 18 world titles and ranked among one of the best players of all time.
Reflecting on the victory, Watkins said the event produced a mixture of “surprise, awe, and some discomfort”.
“Chess was seen as difficult, but had already fallen to the machines, but Go was seen as a final frontier due to its reliance on judgement and ‘feel’,” he said.
A pivotal – and now legendary – moment in the Sedol match, known as “move 37” further fueled the flames of AI hysteria. This saw the system inflict a “shoulder hit” on Sedol, which involved placing a stone diagonally against an opponent’s aimed at reducing their territory.
This was initially thought to be a mistake by the system, but Sedol was flummoxed and took 15-minutes to think up a response. In the end, the move proved to be a decisive factor, securing victory for AlphaGo in the second game.
Sedol conducted a review of the final AlphaGo match with fellow professionals.
(Image credit: Getty Images)
“The now infamous “Move 37” genuinely shocked experts, not because it looked good on paper, but because it was original and even other-worldly, only revealing its brilliance many moves later,” Watkins explained.
“This was the moment when many people stopped seeing AI as a fast calculator, and more of a power that could be used to produce original strategies that humans had not anticipated.”
Paving the way for future AI innovation
As ITPro reported at the time, AlphaGo’s 4- 1 victory against Sedol was watched by over 200 million people globally, generating huge interest and raising questions about the long-term implications of AI.
Watkins told ITPro the victory was a “clear landmark in the movement of modern AI” and showcased the technology’s potential.
“Generative and agentic AI may dominate today’s media headlines, but AlphaGo represents a moment where we were forced to reset our expectations of what AI is capable of,” he said.
“It showed us that systems trained through self-improvement could outperform elite-level humans in extremely complex domains, and raised the bar of ambition in the field upwards a few notches.”
While AlphaGo was a generation removed from the large language model (LLM) boom witnessed in recent years, Watkins noted it was an “important precursor to it and a proof point that machines can learn in surprising ways.”
Indeed, Hassabis noted in his anniversary blog post that the latest Gemini models use “some of the techniques we pioneered with AlphaGo” and its successor, AlphaZero.
FOLLOW US ON SOCIAL MEDIA
Follow ITPro on Google News and add us as a preferred source to keep tabs on all our latest news, analysis, views, and reviews.
You can also follow ITPro on LinkedIn, X, Facebook, and BlueSky.
