How an Intern Played a Key Role in Developing the World-Changing AI

In March 2016, Google DeepMind’s revolutionary artificial intelligence, AlphaGo, captivated the global audience by defeating world champion Lee Sedol in a historic five-game match of Go, an ancient Chinese board game. This milestone was viewed by millions, marking a pivotal advancement in AI technology.

Chris Madison, now a distinguished professor at the University of Toronto, played a critical role in AlphaGo’s creation while he was a master’s student. His journey began with a call from Ilya Sutskever, who later co-founded OpenAI.

Alex Wilkins: What inspired the AlphaGo project?

Elijah: Chris Madison and Ilya presented compelling arguments on why Go was a suitable challenge for AI, stating, “Do you think a skilled player can analyze a Go board and determine the optimal move in half a second?” This notion suggested the possibility of training neural networks to develop effective strategies for selecting the best moves.

Half a second reflects the rapid processing time of the visual cortex—a crucial insight from our previous work with ImageNET, an influential AI image-recognition competition.

Embracing this challenge, I joined the Google Brain team as an intern in summer 2014.

How did AlphaGo evolve from its inception?

Upon joining, I collaborated with a dedicated team at DeepMind, including Aja Huang and David Silver, who were already focused on Go. My primary task was to build the neural network, which felt like a dream come true.

We experimented with various approaches; many initial methods failed, leading to frustration. Eventually, I resorted to a straightforward strategy—training the network on a vast dataset of expert Go games to predict the next best move. This approach proved successful, laying the foundation for our project.

By the end of summer, we conducted a test match where my network outperformed DeepMind’s Thore Graepel, a competent Go player. This success sparked greater interest and investment in the project, allowing us to expand our team significantly.

How daunting was the challenge of defeating Lee Sedol?

I vividly recall the pressure of summer 2014, with a photo of Lee Sedol visible nearby. While I lacked Go expertise, my confidence grew with each network iteration. However, Aja kept reminding me, “Chris, Lee Sedol is an extraordinary player.”

Why did you depart from the AlphaGo team before the match?

David Silver expressed a desire for me to remain and further elevate the project, but I chose to focus on completing my PhD instead. I continued to advise on the project intermittently and take pride in my contributions, even though it took significant collaboration to create the version that ultimately faced Lee Sedol.

What was the atmosphere like during AlphaGo’s victory in Seoul?

The experience in Seoul during the match was indescribable—intense, emotional, and nerve-wracking. It felt reminiscent of a high-stakes sports event, where the outcome was uncertain despite our statistical advantage. From my hotel window, I witnessed crowds transfixed on giant screens showing our game, underscoring the massive impact this event had on East Asia.

What significance does AlphaGo hold for AI?

Though large-scale language models (LLMs) differ greatly from AlphaGo, the underlying technical principles remain unchanged. Initially, neural networks are trained to predict subsequent moves; similarly, today’s LLMs utilize pre-training to forecast the next word using vast amounts of text data.

AlphaGo’s advancement came from integrating human input into the neural network and refining its strategies through reinforcement learning—all focused on the objective of winning games.

As with LLMs, post-preparation reinforcement learning is essential for aligning networks with our intended applications.

In many aspects, the foundational principles of AI development remain consistent.

What areas in AI do you see as flourishing?

Our progress hinges on the availability of comprehensive data for training and reinforcement signals. Without these critical components, no algorithm, no matter how expertly designed, can make meaningful advances.

Did you empathize with Lee Sedol?

Throughout the summer of 2014, Lee Sedol became an idol, making the stakes incredibly high during the match. The immense pressure he faced was palpable, and while watching him realize the gravity of the competition, my heart went out to him. After losing, he apologized, claiming, “This is my failure, not yours,” which was truly heartbreaking.

In Go, players typically review matches to reflect on their performance. Lee Sedol couldn’t do that with AlphaGo and instead sought a friend for analysis, yet it lacked the same reflective quality. Despite the narrative often framing the match as human vs. machine, it was important to recognize the collaboration behind creating AlphaGo, showcasing the dedication of everyone involved in developing this groundbreaking technology.

As AI takes on more human-like tasks, is there still a role for humans?

The game of Go, which we find beautiful, can still teach us more through AI’s insights. The ultimate goal of Go may be to win, but it is also about enjoyment. Thus, the advent of AI doesn’t diminish board games; industries like chess continue to thrive with human appreciation.