Digital Economy Dispatch #198 -- What the Use of AI for Go Teaches Us about Delivering AI-at-Scale

Digital Economy Dispatch #198 -- What the Use of AI for Go Teaches Us about Delivering AI-at-Scale
25th August 2024

"Now, this is not the end. It is not even the beginning of the end.
But it is, perhaps, the end of the beginning”
Winston Churchill

As we reflect on the rapid advances in AI over the past few months, it’s easy to get carried away. Are we are witnessing a key moment in the digital transformation of business and society, one that will likely be remembered as a turning point in how we interact with machines and how they, in turn, shape our world?

The latest wave of AI developments, exemplified by large language models and generative AI, feels like a big step forward. While we have seen previous “false dawns” for AI, perhaps we can say that it signifies the end of the beginning phase for AI – the conclusion of AI's infancy and the dawn of a new era of possibilities and challenges.

To better understand where we are and where we might be heading with AI at scale, I have found it useful to examine the journey of AI in the game of Go. Like many people, I have struggled with playing Go over many years. It is such as simple game to describe in terms of its basic rules with a very straightforward objective: A set of black and white stones are placed on a board with the aim of surrounding territory. The game takes 5 minutes to learn, but a lifetime to master. There is a bewildering set of choices and complex strategies to consider to play the game well. These have raised Go into an artform that has been studied and revered for hundreds of years in many parts of the world.

Due to this combination of simple rules and complex strategies, academics and scientists have found Go to be an ideal testcase to gauge the progress of AI. And conversely, the story of how Go has been impacted by AI serves as a powerful use case for the broader adoption of AI and offers valuable lessons for executives, managers, and professionals grappling with addressing AI-at-Scale and the integration of AI in their organizations. What can we learn from reviewing this journey?

The Early Days of AI for Go: Brute Force and Human Knowledge

When AI researchers first tackled Go, they approached it much like they had approached chess – with brute force computation applied to extensive databases of human knowledge. Early Go programs relied on processing vast numbers of previous games, extracting patterns and strategies that human players had developed over centuries. These systems then used this knowledge, combined with complex algorithms, to calculate optimal moves.

This approach yielded some success but had clear limitations. The game of Go, with its 19x19 board and virtually infinite possible game states, proved far more complex than chess (some say that there are more possible Go moves than there are atoms in the universe!). Even the most powerful computers couldn't process enough moves to consistently outperform top human players. Moreover, these early AI systems were fundamentally limited by the human knowledge they were built upon – they couldn't surpass the collective wisdom of the Go masters whose games they had analyzed, and couldn’t process the vast number of alternative moves to plan very far ahead. As a result, adapting these learned patterns in new contexts was far from straightforward.

In many ways, this early stage of AI for Go mirrors the state of AI in many sectors today. We see AI systems that are impressive within narrow domains, leveraging vast amounts of human-generated data and knowledge. These systems can often outperform humans in specific tasks but struggle with generalization, complex decision making, and adapting to novel situations. Out of context their error rates increase and their behaviour can become unreliable, unstable, and fragile.

The AlphaGo Revolution: Learning from Scratch

The landscape of AI for Go changed dramatically with the introduction of DeepMind's AlphaGo and its successors. Instead of relying solely on human knowledge, these systems used deep learning and reinforcement learning to teach themselves the game from scratch. AlphaGo started with the basic rules of Go and played millions of games against itself, gradually improving its understanding and strategy. This led to AlphaGo Zero, an improved version of AlphaGo that doesn't need human data to learn. AlphaGo Zero uses one combined system to learn both how to make moves and judge board positions. This approach, combined with self-play, made it much better than the original AlphaGo.

Applying these techniques has led to a quantum leap in the performance of AI in Go playing. Most dramatically, in 2016, AlphaGo defeated Lee Sedol, one of the world's top Go players, in a highly publicized five-game match. The victory was widely heralded as a watershed moment in AI history, demonstrating that machines could now outperform humans in a domain long considered too intuitive and complex for artificial intelligence.

One of the most striking moments of this match occurred in game two, with what became known as "Move 37". On the 37th move, AlphaGo placed a stone in a position that initially baffled both Lee Sedol and the human commentators. It was a move that defied conventional Go wisdom, yet it proved to be a brilliant play, ultimately contributing to AlphaGo's victory in that game.

Move 37 was significant not just for its effectiveness, but for what it represented. Here was an AI system making a move that no human expert would have considered, yet it was undeniably strong. It was interpreted as highlighting AI's potential to uncover new strategies and approaches that lie beyond human intuition or established knowledge.

This phase of AI in Go mirrors the current state of cutting-edge AI research and development. We're seeing AI systems that can learn and improve on their own, often surpassing human performance in specific domains. These systems are not just regurgitating human knowledge but generating novel insights and approaches. While still largely based on compute-intensive deep search techniques, the results being obtained are very impressive through a combination of super-fast processors, extensive data-driven training, and sophisticated neural network algorithms.

The Human Response: Lee Sedol's "God Move"

Yet, the story of AI in Go offers much more than a lesson in the power of computing. Despite losing the overall match, Lee Sedol managed to secure a victory against AlphaGo in the fourth game. His winning strategy hinged on what fans later dubbed the "God Move" – a brilliant and unexpected play that exploited a weakness in AlphaGo's understanding of the game. When asked later about that move, Lee Sedol didn’t provide a set of probabilities and statistics. He simply said “it was the only move that made sense”.

This moment serves as a powerful reminder of human ingenuity and adaptability. Even when faced with a seemingly invincible AI opponent, Lee Sedol found a way to innovate and overcome. It underscores the unique strengths that humans bring to the table – creativity, intuition, and the ability to think outside established patterns.

In our broader AI landscape, Lee Sedol's “God Move” represents the ongoing importance of human insight and creativity. As Margret Bowden defined it, creativity can be achieved in three ways: Through combinatorial, exploratory, or transformational means. Machines and humans have different strengths and weaknesses in these 3 areas. The extensive compute capabilities now available make AI particularly effective in combinatorial and exploratory scenarios. Humans excel in transformational situations. As AI systems become more powerful, our role shifts from competing directly with machines to finding innovative ways to work alongside and complement them, identifying their blind spots and leveraging their strengths.

The Aftermath: Lee Sedol's Retirement

Perhaps the most poignant part of this story is its conclusion. In 2019, just three years after his match with AlphaGo, Lee Sedol announced his retirement from professional Go. His reason was telling: "Even if I become the number one, there is an entity that cannot be defeated."

Lee Sedol's decision reflects a profound shift in how we perceive AI and its capabilities. For a master of his calibre to feel that the pinnacle of his field had been irreversibly claimed by AI is a stark illustration of how quickly the landscape can change. It's a reminder that AI's impact isn't just about efficiency or performance metrics – it has human consequences for individuals and can fundamentally alter the nature of entire fields and professions. There are profound implications and deep human consequences to increased use of digital technologies such as AI.

Go Lessons for Leaders in the Age of AI-at-Scale

With AI poised to transform large parts of business and society at an unprecedented scale, what lessons can we draw from the story of AI in Go? Here are a few thoughts:

• Embrace the power of self-learning systems: The leap from knowledge-based AI to self-learning systems like AlphaGo was transformative. In our organizations, we should look for opportunities to implement AI systems that can learn and improve autonomously, rather than relying solely on pre-programmed knowledge.

• Expect the unexpected: Move 37 showed us that AI can generate solutions that lie outside human intuition or established best practices. As leaders, we need to be prepared for AI to challenge our assumptions and be open to radically new approaches.

• Cultivate human creativity: Lee Sedol's “God Move” reminds us of the unique value of human creativity and lateral thinking, particularly in transformational creativity scenarios. As AI takes over more routine tasks, we should focus on nurturing these distinctly human capabilities in our teams.

• Expect profound changes: Lee Sedol's retirement illustrates how quickly AI can reshape entire fields. We must be proactive in anticipating how AI might transform our industries and help our workforce adapt.

• Ethics and governance are crucial: The story of AI in Go played out in the contained world of a board game. As we deploy AI at scale in the real world, with real stakes, we must prioritize ethical considerations and robust governance frameworks. AI’s impacts extend far beyond the logic and statistics of computer games.

• Collaboration is key: The most exciting possibilities lie not in AI replacing humans, but in finding new ways for humans and AI to work together, each leveraging their unique strengths. Encouraging and supporting the right balance will be challenges facing all those in leadership positions.

• Continuous learning is non-negotiable: The rapid progress from early Go programs to AlphaGo underscores the breakneck pace of AI development. Fostering a culture of continuous learning and adaptation in your organization to keep pace with these changes is essential.

To Boldly Go

As we navigate this new era of AI adoption to deliver AI-at-Scale, it's crucial to remember that we are still in the early stages. The story of AI for Go, from its early beginnings to its current dominance, took place over just a few decades. The broader AI revolution is unfolding even more rapidly.

We stand at an critical point, where the decisions we make today about how to develop, deploy, and govern AI will shape the future of our organizations and society at large. It's a responsibility we must approach with both excitement for the possibilities and a clear understanding of the challenges to ensure humans and technology work effectively together.

The shockwave that Go masters faced with AlphaGo had a deep effect, altering the perception of Go forever. By learning from the past, staying adaptable, and maintaining our human creativity and ethical principles, we can ensure that the age of AI-at-Scale is one that augments and enhances human potential rather than diminishing it.