The initial unveiling of AlphaGo’s full capabilities elicited a strong, visceral response. Lee Sedol, then globally recognized as the preeminent player of the ancient Chinese board game Go, exhibited clear signs of agitation at the artificial intelligence’s demonstrable skill. A palpable tension filled the hall in Seoul, South Korea, as the hushed audience struggled to contain their surprise. It rapidly became evident to Lee, and to the tens of millions observing worldwide, that this AI represented a significant departure from its predecessors.
AlphaGo’s success transcended mere victory over Lee. Its performance exhibited an almost human intuition. “AlphaGo actually does have an intuition,” noted Google co-founder Sergey Brin in 2016, shortly after AlphaGo secured a 3-0 lead in the series. “It makes beautiful moves. It even creates more beautiful moves than most of us could think of.”
The match concluded with Google DeepMind’s AlphaGo system achieving a 4-1 victory. Lee Sedol conveyed his state of “shock” following the outcome.
A decade has now passed since this pivotal event for AlphaGo and artificial intelligence as a whole. While the widespread success of large language models like ChatGPT has made marveling at AI commonplace, AlphaGo served as our initial preview of what was to come. Ten years later, the question arises: what is AlphaGo’s enduring legacy, and has the underlying technology fulfilled its initial promise?
“Large language models are now quite different in some ways from AlphaGo, but there’s actually an underlying technological thread that really hasn’t changed,” observes Chris Maddison from the University of Toronto, who was an integral part of the original AlphaGo development team.
The Neural Network Foundation
This foundational technology is neural networks. These are complex mathematical structures that draw inspiration from the biological brain and are implemented through computer code. Traditionally, programming a machine to play a game involved a human explicitly detailing the rules it should follow in various scenarios. In contrast, a neural network enables the machine to learn autonomously.
However, even with the aid of a neural network, mastering the game of Go presented a formidable challenge. This ancient Chinese strategy game, where two players strategically place black and white stones to control territory on a 19×19 board, allows for an astronomical number of potential configurations—estimated at 10171 possible positions. For perspective, the entire observable universe contains approximately 1080 atoms.
The significant advancement stemmed from Maddison and his colleagues’ efforts to replicate human intuition. They achieved this by training a neural network to predict the next most advantageous move based on analyzing millions of recorded professional games. While human players acquire intuition through extensive experience, their capacity is inherently limited compared to an AI that can process vast datasets, offering a distinct advantage.
Furthermore, AlphaGo was not constrained to learning solely from human play. It could engage in millions of games against itself, thereby refining its strategic capabilities. “By learning through these games, it could discover new knowledge and could go beyond human-level players,” explains Pushmeet Kohli at Google DeepMind.
The final iteration of the system that competed against Lee was considerably more sophisticated than Maddison’s initial prototypes. Nevertheless, the core message remained clear: neural networks were exceptionally effective. “AlphaGo definitively showed that neural nets can do pattern recognition better than humans. They can essentially have intuition that surpasses humans,” states Noam Brown, formerly of OpenAI.
Expanding Capabilities Beyond Games
Following AlphaGo’s success, Google DeepMind and the broader AI research community shifted their focus to applying this fundamental understanding to real-world challenges, particularly within the fields of mathematics and biology. A prominent illustration of this transition is AlphaFold, an AI developed to predict the three-dimensional structure of proteins from their constituent chemical components with unprecedented accuracy, surpassing any human-designed program. The development of AlphaFold earned its creators the Nobel Prize in Chemistry.
More recently, AlphaProof, another AI based on neural networks, demonstrated exceptional performance at the International Mathematical Olympiad, a highly competitive mathematics contest for students. Its proficiency left mathematicians astonished. “Not only can you get this beyond-human-level intelligence in a game, but you can get that experience in important scientific applications,” notes Kohli.
The underlying principles guiding both AlphaGo-style AI and the systems powering large language models (LLMs) such as ChatGPT share significant commonalities. The initial phase, known as pretraining, involves exposing a neural network to a massive corpus of human-generated data. For AlphaGo, this meant complete Go game records; for LLMs, it encompasses vast swathes of the internet. The subsequent phase, termed post-training or reinforcement learning, further refines the network’s capabilities by defining success and allowing the AI to autonomously identify pathways to achieve it.
In AlphaGo’s case, this involved self-play for millions of rounds to ascertain optimal winning strategies. For AlphaFold, the AI was guided by examples of successfully folded proteins, enabling it to deduce the underlying principles. ChatGPT employs a process called reinforcement learning from human feedback, where it learns to favor responses that human users prefer. Alternatively, it can be presented with a defined problem, such as in mathematics or coding, and then iteratively refine its approach by feeding its own output back into itself, mirroring a form of internal deliberation.
The Challenge of Explainability
Despite these advancements, significant challenges persist. Neural networks often function as opaque “black boxes.” Even with concerted efforts to understand their internal workings, many models are simply too large and complex to be comprehended at a fundamental level.
During AlphaGo’s now-famous move 37, initial reactions suggested a potential error. However, as the game unfolded, it became apparent that the move was a strategic masterstroke. Crucially, even Google DeepMind’s engineers could not ascertain AlphaGo’s rationale for this specific play. It also highlights the possibility that such a move could have been a genuine mistake, with no clear way to ascertain the reasoning behind it.
“These models will come up with answers and we will not know whether they are genius insights or hallucinations,” states Kohli. “We are still all actively working on trying to resolve those sorts of questions.”
A crucial element of AlphaGo’s success was the availability of abundant data for initial training and a clear, unambiguous definition of success. Consequently, it is logical that AI is currently achieving its most significant breakthroughs in domains where these conditions are also met. Maddison points to areas like mathematics and programming as prime examples, where defining and verifying correctness is straightforward. “The similarities between these approaches are telling us something, and it’s telling us what are the raw necessary ingredients for progress.”